rock physics and seismic signatures of sub …jz636wn5735/dejtrakulwong_p_dissertation...this...
TRANSCRIPT
ROCK PHYSICS AND SEISMIC SIGNATURES OF SUB-RESOLUTION
SAND-SHALE SYSTEMS
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF GEOPHYSICS
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE
OF DOCTOR OF PHILOSOPHY
Piyapa Dejtrakulwong
December 2012
http://creativecommons.org/licenses/by-nc-sa/3.0/us/
This dissertation is online at: http://purl.stanford.edu/jz636wn5735
© 2012 by Piyapa Dejtrakulwong. All Rights Reserved.
Re-distributed by Stanford University under license with the author.
This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 United States License.
ii
I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.
Gerald Mavko, Primary Adviser
I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.
Tapan Mukerji, Co-Adviser
I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.
Stephan Graham
Approved for the Stanford University Committee on Graduate Studies.
Patricia J. Gumport, Vice Provost Graduate Education
This signature page was generated electronically upon submission of this dissertation in electronic format. An original signed hard copy of the signature page is on file inUniversity Archives.
iii
iv
v
Abstract
This dissertation aims to improve the interpretation of thinly bedded sand-shale
systems that are below resolution of conventional well-log and seismic data by using
rock physics and quantitative seismic analysis. The key contributions of this
dissertation are (1) incorporation of parameter uncertainties into existing models that
are used for estimating petrophysical properties of sub-resolution sand-shale systems,
(2) a new method for approximating fluid substitution in thinly bedded sand-shale
reservoirs that is applicable at the measurement scale without the need to downscale
the measurements, and (3) an application of rock physics, spatial statistics, and
feature-extraction based attributes to quantitatively interpret seismic data for sub-
resolution reservoir properties such as net-to-gross ratios, saturations, and stacking
patterns.
Most of rock physics relations are derived for rocks that are considered
homogeneous at particular scales. If these relations are applied to measurements at
other scales, the relations often fail when the measurements represent average
properties of heterogeneous rocks (for example, a stack of interbedded sand and shale
layers). First, we investigate the Thomas-Stieber model, a model commonly used for
estimating volume fraction of sand and its porosity in thinly bedded sand-shale
sequences. We present sensitivity and uncertainty analyses of this model under various
scenarios, especially when the model assumptions are violated. We also extend the
model by incorporating uncertainties into the model parameters using Monte Carlo
simulations in a Bayesian framework.
vi
Next, we propose a simple graphical mesh interpretation and accompanying
equations for approximating fluid substitution in sub-resolution interbedded sand-
shale sequences. The advantages of our method are as follows. Even when it is applied
to the measurements at their original scales, our method appropriately changes fluid in
the sands only, without the need to downscale the measurements. The interbedded
sand layers can be either clean or shaly (i.e., sand with dispersed clay). We illustrate
the performance of the model using both synthetic and real well log data and present
sensitivity analysis of the model parameters.
Estimating reservoir properties of sub-resolution sand-shale reservoirs from
seismic data is not straightforward because the relation between seismic signatures and
rock properties are not unique. This relation is even further complicated by the spatial
arrangement of the sub-resolution layers. This dissertation presents a workflow for
seismic interpretation of such thin reservoirs. The workflow consists of four main
steps: (1) estimate transition matrices (Markov chain model) at the well location from
log data, (2) use the matrices to create various sand-shale sequences with varying
reservoir properties such as net-to-gross ratios, saturations, and stacking patterns, (3)
generate synthetic seismograms corresponding to the sequences and from these
seismograms extract attributes which will be used as a training set, (4) and finally use
the training set to estimate reservoir properties of the area away from the well. Most of
seismic attributes discussed here are obtained using feature-extraction techniques,
which compare amplitudes of seismogram segments and find new representations of
these seismograms in a new, smaller set of features. We apply the workflow to both
synthetic data and real data from channelized turbidite deposits in West Africa.
vii
Acknowledgement
My PhD journey at Stanford University is one of the most meaningful experiences
in my life. I would not have come this far on my journey without help and support
from several people. First of all, I am deeply grateful to Gary Mavko, my adviser, for
his teaching and guidance throughout my time at Stanford. His insightful ideas have
tremendous impact on my research. I would like to thank Tapan Mukerji, my co-
adviser, for his advice and inspirational ideas. I would like to thank both Gary and
Tapan for helping me improve my presentation skills.
I am grateful to Jack Dvorkin for his continuous support. I am very impressed by
his kindness and willingness to help students. I am grateful to Steve Graham for his
wonderful teaching and advice on geology concepts. I am indebted to Claude Reichard
for his teaching on technical writing. His advice helps improve the writing of this
dissertation tremendously. I would like to thank Tiziana Vanorio for her support
throughout these years.
I would like to thank Fuad Nijim, Tara Illich, and all administrative staffs for their
help. I also would like to thank the SRB program and its affiliates and thank the SEG
scholarship for financial support.
I would like to thank all SRB colleagues for their support, advice, and friendship:
Kyle Spikes, Kevin Wolf, Tanima Dutta, Kaushik Bandyopadhyay, Richa, Carmen
Gomez, Franklin Ruiz, Ratnanabha Sain, Ramil Ahmadov, Danica Dralus, Nishank
Saxena, Adam Tew, Yu Xia, Kenichi Akama, Dario Grana, Adam Allan, Amrita Sen,
Sabrina Aliyeva, Yuki Kobayashi, Ammar El Husseiny, Priyanka Dutta, Humberto
viii
Samuel Arevalo-Lopez Sr, Chisato Konishi, Huyen Bui, Cinzia Scotellaro, Stephanie
Vialle, and Fabian Krzikalla. I am also grateful to Ezequiel Gonzalez and Juan-
Mauricio Florez for useful discussion and suggestion on my research.
I would like to thank Peerapong Dhangwatnotai for always being by my side.
Lastly, I dedicate this dissertation to my family in Thailand. Their unconditional love
and support always encourage me to keep on walking.
ix
Contents
Abstract……………………………………………………………………………….v
Acknowledgements…………………………………………………………………vii
Contents……………………………………………………………………………...ix
List of Tables……………………………………………………………………….xiii
List of Figures……………………………………………………………………….xv
Chapter 1 Introduction ................................................................................................ 1
1.1 Objective ........................................................................................................ 1
1.2 Background and motivation ........................................................................... 2
1.2.1 Rock physics relations and sub-resolution heterogeneity ...................... 3
1.2.2 Seismic property estimation of sub-resolution heterogeneous systems . 4
1.3 Definitions of terms, symbols, and abbreviations .......................................... 5
1.4 Chapter description ........................................................................................ 6
1.5 References ...................................................................................................... 7
Chapter 2 Sensitivity and uncertainty analysis of the Thomas-Stieber model for
property estimation of thin sand-shale reservoirs ..................................................... 9 2.1 Abstract .......................................................................................................... 9
2.2 Introduction .................................................................................................. 10
2.3 Volumetric properties of sand-shale mixtures ............................................. 12
2.3.1 Total porosity and volume of shale ...................................................... 12
2.3.2 Important assumptions of the Thomas-Stieber model .......................... 16
2.4 Sensitivity and uncertainty analysis ............................................................. 19
2.4.1 Model with correct input parameters .................................................... 19
2.4.2 Model with uncertain input parameters ................................................ 24
x
2.4.3 Example with well log data .................................................................. 29
2.5 The Thomas-Stieber model on rock-physics cross-plots ............................. 31
2.5.1 Density and volume fraction of shale ................................................... 32
2.5.2 Neutron porosity and density ................................................................ 33
2.5.3 Velocity and total porosity ................................................................... 34
2.5.4 Vp/Vs ratio and acoustic impedance ................................................... 36
2.5.5 Application to real data ........................................................................ 37
2.6 Discussion .................................................................................................... 38
2.7 Conclusions .................................................................................................. 42
2.8 Acknowledgements ...................................................................................... 42
2.9 References .................................................................................................... 42
Chapter 3 Fluid substitution for sub-resolution interbedded sand-shale sequences
using the mesh method ............................................................................................... 47 3.1 Abstract ........................................................................................................ 47
3.2 Introduction .................................................................................................. 49
3.3 Elastic properties of interbedded sands ........................................................ 52
3.3.1 Models .................................................................................................. 52
3.3.2 Modeling the V-point using the Voigt-Reuss-Hill average .................. 56
3.4 Fluid substitution for interbedded sands ...................................................... 60
3.4.1 Approximate fluid substitution in shaly sands ..................................... 60
3.4.2 Graphical interpretation and equation derivations for fluid substitution
in interbedded sands ............................................................................. 62
3.4.3 Important note ...................................................................................... 66
3.5 Synthetic examples ...................................................................................... 66
3.5.1 Case 1: Shaly sand (sand with dispersed clay) with fully-oil-saturated
effective porosity (𝑆𝑊𝑒=0) .................................................................. 67
3.5.2 Case 2: Clean sand interbedded with shale, varying 𝑉𝑠𝑎𝑛𝑑, and fully-
oil-saturated effective porosity (𝑆𝑊𝑒=0) ............................................. 69
3.5.3 Case 3: Shaly sand (𝑉𝑑𝑖𝑠𝑝= 0.15) interbedded with shale, varying
𝑉𝑠𝑎𝑛𝑑, and fully-oil-saturated effective porosity (𝑆𝑊𝑒=0) ................ 70
3.5.4 Case 4: Shaly sand (normal distribution of 𝑉𝑑𝑖𝑠𝑝 with a mean of 0.1
and a standard deviation of 0.05) interbedded with shale, varying
𝑉𝑠𝑎𝑛𝑑, and fully-oil-saturated effective porosity (𝑆𝑊𝑒=0) ................ 71
3.5.5 Case 5: Interbedded sand-shale sequences with systematic changes of
both the 𝑉𝑑𝑖𝑠𝑝 in the sand layers and the 𝑉𝑠𝑎𝑛𝑑 ................................ 73
3.6 Pitfalls in interpretation ................................................................................ 74
3.7 Sensitivity analysis ....................................................................................... 75
3.8 Real data example ........................................................................................ 78
3.9 Discussion .................................................................................................... 84
3.9.1 The mesh method using both bulk and shear moduli ........................... 84
3.9.2 Possible modification of the mesh method when key assumptions are
relaxed .................................................................................................. 85
3.9.3 Using rock-physics trends to constrain clean-sand properties ............. 86
xi
3.9.4 Limitation of the mesh method ............................................................. 87
3.9.5 Comparison with alternative methods .................................................. 87
3.9.6 Upscaled Gassmann’s equations .......................................................... 89
3.10 Conclusions ................................................................................................ 102
3.11 Acknowledgements .................................................................................... 103
3.12 References .................................................................................................. 103
Chapter 4 Seismic signature and uncertainty in petrophysical property
estimation of thin sand-shale reservoirs ................................................................. 107
4.1 Abstract ...................................................................................................... 107
4.2 Introduction ................................................................................................ 109
4.3 Forward modeling for seismic response and attributes .............................. 111
4.3.1 Markov chain models in stratigraphic sequences ............................... 112
4.3.2 Rock-physics models for sand-shale mixtures ................................... 116
4.3.3 Seismic attributes ................................................................................ 118
4.4 Seismic signatures for 1-D Synthetic example .......................................... 124
4.4.1 Model setup ........................................................................................ 124
4.4.2 Scenario 1: Effect of net-to-gross ratios ............................................. 125
4.4.3 Scenario 2: Effect of saturation .......................................................... 138
4.4.4 Scenario 3: Effect of stacking patterns ............................................... 150
4.4.5 Discussions ......................................................................................... 157
4.5 Net-to-gross estimation from 2-D sections ................................................ 160
4.5.1 Model setup ........................................................................................ 161
4.5.2 Results ................................................................................................ 162
4.5.3 Net-to-gross estimation using a Bayesian framework ........................ 163
4.6 Local net-to-gross estimation in non-stationary sequences ....................... 164
4.6.1 Model setup ........................................................................................ 165
4.6.2 Results and discussion ........................................................................ 167
4.7 Discussion .................................................................................................. 170
4.7.1 Comparisons with amplitude attributes .............................................. 170
4.7.2 Notes on the feature-extraction based attributes ................................ 171
4.8 Conclusions ................................................................................................ 173
4.9 Acknowledgements .................................................................................... 173
4.10 References .................................................................................................. 174
Chapter 5 Seismic signature and uncertainty in petrophysical property
estimation of thin sand-shale reservoirs: Case studies .......................................... 179
5.1 Abstract ...................................................................................................... 179
5.2 Introduction ................................................................................................ 181
5.3 Geological background .............................................................................. 181
5.4 Case study 1: Effect of reservoir properties and stacking pattern on seismic
signatures at well location .......................................................................... 183
5.4.1 Net-to-gross estimation from well log data ........................................ 183
xii
5.4.2 Seismograms at well location ............................................................. 186
5.4.3 Transition matrices and rock property calibration .............................. 187
5.4.4 Synthetic seismograms ....................................................................... 191
5.4.5 Effect of reservoir properties and stacking pattern on seismic signatures
............................................................................................................ 192
5.4.6 Discussion ........................................................................................... 193
5.5 Case study 2: Estimating sub-resolution reservoir properties from a 2-D
seismic section ................................................................................................ 203
5.5.1 2-D seismic section ............................................................................. 203
5.5.2 Rock property calibration at well location ......................................... 203
5.5.3 Non-stationarity when moving away from the well ........................... 205
5.5.4 Synthetic seismogram generation and seismic attribute extraction .... 207
5.5.5 Sensitivity analysis ............................................................................. 211
5.5.6 Property estimation of the 2-D seismic section .................................. 214
5.5.7 Discussion ........................................................................................... 220
5.6 Conclusions ................................................................................................ 221
5.7 Acknowledgements .................................................................................... 222
5.8 References .................................................................................................. 222
xiii
List of Tables
Table 1.1: Definitions of terms, symbols, and abbreviations used throughout this
dissertation. Note that Vi is the volume of the ith
layer illustrated in Figure 1.1, and
i = 1, 2,…, 6. ........................................................................................................... 5
Table 3.1: Models used to generate elastic moduli of brine-saturated shaly-sand. ...... 57
Table 3.2: Models used to estimate effective solid moduli for fluid substitution. ....... 57
Table 3.3: Summary of methods usually used in fluid substitution and comments for
when these methods are applied to sub-resolution interbedded sand-shale
sequences. ............................................................................................................. 88
Table 4.1: Kernel functions used for extracting seismic attributes. Note that xi is the ith
seismogram. ........................................................................................................ 122
Table 4.2: Form of transition matrices for generating sequences used in investigating
net-to-gross effects on seismic signatures. Values of parameter k range from 0.45
to 0.95. The four lithologic states are sand (s), shaly-sand (sh-s), sandy-shale (s-
sh), and shale (sh). ............................................................................................... 126
Table 4.3: Summary of the methods used to compute seismic attributes for net-to-
gross estimation. Choices of parameters for each method are also included.
Performance of each method is shown as a success rate in classifying a data point
into three net-to-gross classes: <0.38, 0.38-0.465, and >0.465 .......................... 137
Table 4.4: Summary of the methods used to compute seismic attributes for water-
saturation (in sand layers) effect. Choices of parameters for each method are
similar to those listed in Table 4.2, unless otherwise specified. Performance of
each method is shown as a success rate in classifying a data point into three
water-saturation categories: Sw=0.1, 0.5, and 1. Results are shown in the columns
corresponding to the transition matrices used in simulations. The actual transition
matrices and sample sequences are shown in Figure 4.23. ................................. 149
xiv
Table 4.5: Summary of the methods used to compute seismic attributes for stacking-
pattern effects using sequences generated from the fixed-sampling transition
matrices. Choices of parameters for each method are also included. Performance
of each method is shown as a success rate in classifying a data point into three
stacking patterns: retrogradational, progradational, and aggradational patterns. 152
Table 4.6: for stacking-pattern effects using sequences generated from the fixed-
sampling transition matrices. Choices of parameters for each method are also
included. Performance of each method is shown as a success rate in classifying a
data point into three stacking patterns: retrogradational, progradational, and
aggradational patterns. ........................................................................................ 156
Table 5.1: Specifications for synthetic earth models with various net-to-gross ratios,
stacking patterns, saturating fluids, and layer thicknesses. ................................. 191
xv
List of Figures
Figure 1.1: Different types of clay/shale distribution in sand. The volume of the ith
layer and the volume of dispersed clay are denoted by Vi and Vdisp, respectively. . 6
Figure 2.1: Shale distribution in sand (Modified after Mavko et al., 2009). ................ 12
Figure 2.2: Relation between total porosity and shale volume. Point A and E represent
clean sand and pure shale point, respectively. Points B and C represent sand with
dispersed clay, and point D is where the original sand pore space is completely
filled with dispersed clay. ..................................................................................... 13
Figure 2.3: Graphical determination of the laminated shale and the dispersed clay
volumes (𝑉𝑙𝑎𝑚and 𝑉𝑑𝑖𝑠𝑝, respectively) for a measurement at point F. In each
case, the shale volume is determined by drawing a line from the shale point (E),
passing through point F, and intersecting line AD. The intersection point (C)
represents shaly sand that is the sandy end-member of the lamination, and this
shaly sand point has 𝑉𝑑𝑖𝑠𝑝equal to its x-coordinate (𝑉𝑑𝑖𝑠𝑝= 0.2). Then, 𝑉𝑙𝑎𝑚 is
simply the ratio between the lengths of line CF and CE, which is equal to 0.4. .. 15
Figure 2.4: Distorting total porosity diagram (blue) for computing effective porosity
(red). ...................................................................................................................... 16
Figure 2.5: Total porosity and shale volume after cement is added into the remaining
sand pore space of the sands on the dispersed sand line. The blue line is the
resulting dispersed sand line after adding cement with a volume of 25% of the
porosity of the sand end-point into the pore space (Modified after Juhasz, 1986).
............................................................................................................................... 17
Figure 2.6: Thomas-Stieber diagram (left) with plotted properties from the
corresponding earth model which is represented by an interbedded sand-shale
sequence (right). The properties of the individual sand types (A, B, and C) and
shale are shown in circles. The average total porosity and shale volume fraction
of the earth model is marked with an “X.” The volume fractions of laminated
shale and dispersed clay can be determined graphically by first drawing a line that
xvi
originates from the shale point, passes through point X, and intersects the
dispersed line. The volume fraction of dispersed clay is simply equal to the x-
coordinate of the intersection point, and the volume fraction of laminated shale
(𝑉𝑙𝑎𝑚 = 1 – 𝑉𝑠𝑎𝑛𝑑) is the ratio between the distance from point X to the
intersection point and the distance from the shale point to the intersection point.20
Figure 2.7: Estimated sand fraction and sand property when applying Thomas-Stieber
diagram to point “X” in Figure 2.6 (left). The estimated values are compared with
the true values from the earth model (Figure 2.6, right). ...................................... 20
Figure 2.8: Thomas-Stieber diagram (left) with plotted properties from the
corresponding earth model which is represented by an interbedded (shaly) sand-
(sandy) shale sequence (right). The properties of sand and shale layers within the
earth model are shown in circles. The average total porosity and shale volume
fraction of the earth model is marked with an “X.” The volume fractions of
laminated shale and dispersed clay can be determined graphically by first drawing
a line that originates from the shale point, passes through point X, and intersects
the dispersed line. The volume fraction of dispersed clay is simply equal to the x-
coordinate of the intersection point, and the volume fraction of laminated shale
(𝑉𝑙𝑎𝑚 = 1 – 𝑉𝑠𝑎𝑛𝑑) is the ratio between the distance from point X to the
intersection point and the distance from the shale point to the intersection point.21
Figure 2.9: Estimated sand fraction and sand property when applying Thomas-Stieber
diagram to point “X” in Figure 2.8 (left). The estimated values are compared with
the true values from the earth model (Figure 2.8, right). ...................................... 22
Figure 2.10: Relations between total porosity and shale volume fraction of bimodal
mixtures using the fractional packing model by Kolterman and Gorelick (1995)
and the Thomas-Stieber model (i.e., ideal mixing model). Data points
representing sand/shaly-sand interbedded with shale (solid circles) are generated
using the fractional packing model with varying sand fractions and volume
fractions of dispersed clay..................................................................................... 24
Figure 2.11: Percentage differences between the estimated sand fractions (left), the
estimated volume fractions of dispersed clay in the sand (right) and the true
values. The ideal mixing model and the fractional packing model (i.e., non-ideal
mixing) are outlined. ............................................................................................. 24
Figure 2.12: Multiple realizations of Thomas-Stieber diagrams generated from a set of
sand and shale end-points. Three data points are labeled. ................................... 26
Figure 2.13: Posterior distributions for estimated sand fractions of the three data points
shown in Figure 2.12. ............................................................................................ 27
Figure 2.14: Posterior distributions of volume fraction of dispersed clay in sand (left)
and sand fraction (right) for the three data points shown in Figure 2.12. ............. 27
Figure 2.15: Standard deviations of estimated properties at different locations on the
Thomas-Stieber diagram. The two estimated values are (left) volume fraction of
dispersed clay and (right) sand fraction. The three data points are the same points
as in Figure 2.12. ................................................................................................... 29
Figure 2.16: Three lithofacies from a detailed core analysis by Lowe (2004).
Lithofacies 1 represents thick-bedded to massive sandstone. Lithofacies 2 and 3
represent interbedded, thin-bedded sandstone and mudstone. In lithofacies 2, the
xvii
sandstone beds are 2 – 20 cm thick and in lithofacies 3 the beds are less than 2 cm
thick (Modified from Dutta, 2009). ...................................................................... 30
Figure 2.17: Petrophysical analysis of the selected well-log interval using the Thomas-
Stieber model. (Left) Total porosity and gamma ray values for three lithofacies in
the selected interval. The median total porosity and gamma ray values for each
lithofacies are shown in solid circles. From these median points, the up-down or
left-right bars indicate the interquartile ranges (i.e., from 1st to 3
rd quartiles) of
each property. A Thomas-Stieber diagram is also superimposed on the data.
(Right) Variation of gamma ray values with depth. Data points are color-coded by
lithofacies similar to the left panel. The gamma ray log shows an upward-fining
trend. ..................................................................................................................... 30
Figure 2.18: Probability density functions of properties of the sand end-point and the
shale end-point. (Left) total porosity and (right) gamma ray value. ..................... 31
Figure 2.19: Estimated sand fractions for the selected well-log interval. (Left) gamma
ray log of the selected interval and (right) posterior distributions of estimated
sand fractions. ....................................................................................................... 31
Figure 2.20: Relationship between density and volume fraction of shale in shaly sand
lamination. Point A and E represent clean sand point and pure shale point,
respectively. Points B and C represent sand with dispersed clay, and point D is
where the original sand pore space is completely filled with dispersed clay
(Modified after Mavko et al., 2009). ..................................................................... 32
Figure 2.21: Density-neutron plot for the dispersed and laminated sand-shale systems.
............................................................................................................................... 34
Figure 2.22: Velocity to total porosity curves for the dispersed sand-shale system. The
model used here follows Dvorkin and Gutierrez, 2002. ....................................... 35
Figure 2.23: Velocity to total porosity curves for the dispersed sand-shale system. The
model used here follows Dvorkin and Gutierrez, 2002. Each black line represents
lamination between sand (or shaly sand) and shale with volume fraction of
laminated shale ranging from 0 to 1. Each blue line represents lamination between
sand with a volume fraction of dispersed clay ranging from 0 to 𝜙𝑐𝑙𝑒𝑎𝑛 𝑠𝑎𝑛𝑑
and shale with a constant volume fraction of laminated shale. ............................. 35
Figure 2.24: Rock-physics template shown as a cross-plot between Vp/Vs and AI with
superimposed rock-physics trends. The green and magenta curves represent the
shale and wet-sand lines, respectively. Along these lines, the change in porosity is
due to packing or grain sorting. At each porosity value, the red curve which is
connected to the wet-sand line represents the corresponding gas-saturated sand
with varying saturations. The dispersed and laminated sand-shale system is
constructed using the sand and shale points shown in red circles. Each blue line
represents lamination between sand with a volume fraction of dispersed clay
ranging from 0 to 𝜙𝑐𝑙𝑒𝑎𝑛 𝑠𝑎𝑛𝑑 and shale with a constant volume fraction of
laminated shale. ..................................................................................................... 37
Figure 2.25: Selected well-log data on different cross-plots for laminated and
dispersed sand-shale systems. The data points are color-coded by their
corresponding gamma ray values. ......................................................................... 38
xviii
Figure 2.26: The effect of vertical resolution on cross-plots between total porosity and
shale volume of a synthetic earth model between shaly sand and shale. The left
panel shows a cross-plot between the two measurements with two different
resolutions; whereas, the middle panel shows a cross-plot when both
measurements share the same resolution. The right panel is a short section of the
corresponding synthetic earth model. Sand and shale layers are shown in white
and black, respectively. ......................................................................................... 40
Figure 2.27: Estimated sand fraction and sand property when applying Thomas-Stieber
diagram to points in Figure 2.26 to investigate the effect of vertical resolution on
interpretation using cross-plots between total porosity and shale volume............ 40
Figure 2.28: The effect of vertical resolution on cross-plots between total porosity and
shale volume of a synthetic (non-stationary) earth model between shaly sand and
shale. The left panel shows a cross-plot between the two measurements with two
different resolutions; whereas, the middle panel shows a cross-plot when both
measurements share the same resolution. The right panel is a short section of the
corresponding synthetic earth model. Sand and shale layers are shown in white
and black, respectively. ......................................................................................... 41
Figure 2.29: Estimated sand fraction and sand property when applying Thomas-Stieber
diagram to points in Figure 2.28 to investigate the effect of vertical resolution on
interpretation using cross-plots between total porosity and shale volume............ 41
Figure 3.1: Inverted-V relation between P-wave velocity and total porosity for a
dispersed sand-shale system, following the model of Dvorkin and Gutierrez
(2002). In this case, curves are computed using the Hashin-Shtrikman lower
bound (HSLB). The solid magenta and blue curves represent the sandy-shale and
shaly-sand legs of the dispersed sand-shale system. ............................................. 53
Figure 3.2: Relation between velocity and total porosity for a dispersed sand-shale
system, following the model of Dvorkin and Gutierrez (2002) with a slight
modification. Each red line represents interbedding of sand (or shaly sand) and
shale with volume fraction of interbedded shale ranging from 0 to 1. Each blue
line represents interbedding of sand with a volume fraction of dispersed clay
ranging from 𝑉𝑑𝑖𝑠𝑝 = 0 to 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑆 and shale with a constant volume
fraction of interbedded shale. ................................................................................ 54
Figure 3.3: P-wave compliance (C=1/M) versus effective porosity curves for a
interbedded sand-shale system. Each red line represents interbedding of sand (or
shaly sand) and shale with volume fraction of laminated (interbedded) shale
ranging from 𝑉𝑙𝑎𝑚 = 0 to 𝑉𝑙𝑎𝑚 = 1. Each blue line represents interbedding of
sand with a volume fraction of dispersed clay ranging from 𝑉𝑑𝑖𝑠𝑝 = 0 to
𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑆 and shale with a constant volume fraction of laminated (interbedded)
shale. ..................................................................................................................... 55
Figure 3.4: Shaly sand lines before and after fluid substitution when only P-wave
moduli are used in calculations. The model used to generate wet shaly sand is
listed in the lower right corner ((a) – (d) and Table 3.1). Starting with the brine-
saturated shaly sand (black dash-line), oil is substituted using the Gassmann
xix
shaly-sand equation (Dvorkin et al., 2007) and three different models for effective
solid moduli (Table 3.2). ....................................................................................... 59
Figure 3.5: Shaly sand lines before and after fluid substitution when both bulk and
shear moduli are used in calculations. The model used to generate wet shaly sand
is listed in the lower right corner ((a) – (d) and Table 3.1). Starting with the brine-
saturated shaly sand (black dash-line), oil is substituted using the Gassmann
shaly-sand equation and three different models for effective solid moduli (Table
3.2). ....................................................................................................................... 60
Figure 3.6: Approximate fluid substitution for a dispersed shaly-sand line. Applying
fluid substitution to this shaly-sand line is approximately equivalent to moving
the clean sand point up or down, following the usual Gassmann’s equation, while
the V-point at the other end of this line is fixed. Then, the new clean sand point is
connected to the fixed V-point by another straight line. Here, we show an
example when the clean sand point is moved down after fluid substitution. ........ 62
Figure 3.7: Graphical interpretation of fluid substitution by our mesh method.
Applying approximate fluid substitution to any clean sand (or sand with dispersed
clay) interbedded with shale is simply equivalent to distorting the mesh. Here the
distortion is shown in the P-compliance (C=1/M) versus effective porosity plane.
The distortion moves the mesh accordingly with the change in the clean sand
compliance after fluid substitution. The blue arrow shows how a data point inside
the triangular diagram moves after the distortion. ................................................ 63
Figure 3.8: Schematic diagram and terminology for the laminated (interbedded) sand-
shale system. ......................................................................................................... 64
Figure 3.9: Fluid substitution results from four different procedures for shaly sands
with no interbedding. The procedures are Gassmann’s equation applied to sand
layers only + upscaling, Gassmann shaly-sand equation (Dvorkin et al., 2007)
applied to each sand layers only + upscaling, Gassmann’s equation applied at the
measurement scale, and our mesh method. For this synthetic case, volume
fractions of dispersed clay in the sand range from 𝑉𝑑𝑖𝑠𝑝= 0 to 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑠. ..... 68
Figure 3.10: Fluid substitution results from four different procedures for interbedded
clean sand-shale sequences. The procedures are Gassmann’s equation applied to
sand layers only + upscaling, Gassmann shaly-sand equation (Dvorkin et al., 2007)
applied to each sand layers only + upscaling, Gassmann’s equation applied at the
measurement scale, and our mesh method. Note that results of Gassmann’s
equation + upscaling, Gassmann shaly-sand equation + upscaling, and the mesh
are on top of each other. The x-axis represents pseudo-depth. ............................. 69
Figure 3.11: Sand fractions of the synthetic model for interbedded clean sand-shale
sequences. The x-axis represents pseudo-depth. ................................................... 69
Figure 3.12: Percentage differences between velocities after fluid substitution by
Gassmann’s equation and Gassmann’s equation applied to only sand layers
followed by upscaling. These differences are plotted against sand fraction and
volume fraction of shale (𝑉𝑠ℎ). In this clean sand case, sand fraction is simply
equivalent to 1 – 𝑉𝑠ℎ. ........................................................................................... 70
Figure 3.13: Fluid substitution results from four different procedures for interbedded
shaly sand-shale sequences, with a fixed volume fraction of dispersed clay. The
xx
procedures are Gassmann’s equation applied to sand layers only + upscaling,
Gassmann shaly-sand equation (Dvorkin et al., 2007) applied to each sand layers
only + upscaling, Gassmann’s equation applied at the measurement scale, and our
mesh method. The x-axis represents pseudo-depth. ............................................. 70
Figure 3.14: Sand fractions of the synthetic model for interbedded shaly sand-shale
sequences, with a fixed volume fraction of dispersed clay. The x-axis represents
pseudo- depth. ....................................................................................................... 71
Figure 3.15: Fluid substitution results from four different procedures for interbedded
shaly sand-shale sequences, with varying volume fractions of dispersed clay. The
procedures are Gassmann’s equation applied to sand layers only + upscaling,
Gassmann shaly-sand equation (Dvorkin et al., 2007) applied to each sand layers
only + upscaling, Gassmann’s equation applied at the measurement scale, and our
mesh method. The x-axis represents pseudo-depth. ............................................. 71
Figure 3.16: Sand fractions of the synthetic model for interbedded shaly sand-shale
sequences, with varying volume fractions of dispersed clay. The x-axis represents
pseudo-depth. ........................................................................................................ 72
Figure 3.17: Changes in P-wave velocity after fluid substitution from four different
procedures: Gassmann’s equation applied to sand layers only + upscaling,
Gassmann shaly-sand equation (Dvorkin et al., 2007) applied to sand layers only
+ upscaling, Gassmann’s equation applied at the measurement scale, and our
mesh method. Each location on the triangular diagram represents a interbedded
sand-shale sequence with a unique pair of sand fraction and volume fraction of
dispersed clay values. ............................................................................................ 73
Figure 3.18: Pitfalls in interpretation of fluid substitution results when thin
interbedding exists. (Left) Synthetic gas-saturated data with sand fractions greater
than or equal to 0.5. (Middle) Results after fluid substitution from gas to brine
using our mesh approach. (Right) Results after fluid substitution from gas to brine
using Gassmann’s equation. Data points are color-coded by water saturation
values before fluid substitution. ............................................................................ 75
Figure 3.19: Sensitivity analysis results of the mesh method for five input parameters:
P-wave velocity and total porosity of clean sand, effective porosity and effective
water saturation of the interbedded package, and elastic modulus of fluid. The
thick lines on the top of each subplot are the oil-saturated velocity and the
reference velocities after fluid substitution from oil to brine using four procedures
with all correct input parameters: Gassmann ignoring interbedding, Gassmann
applied to each sand layer + upscaling, Gassmann shaly-sand equation (Dvorkin
et al., 2007) applied to each sand layer + upscaling, and our mesh method. Results
of sensitivity analysis for each parameter are shown as a velocity distribution,
which is normalized to one. Sand fractions and volume fractions of dispersed clay
for each synthetic model are shown on above each subplot. ................................ 78
Figure 3.20: Selected dataset represents an interbedded sand-shale sequence. The data
is color-coded by Gamma ray values. The Thomas-Stieber-Yin-Marion model is
superimposed onto the data. The set of parallel lines labeled as 𝑉𝑠𝑎𝑛𝑑 represent
volume fractions of laminated (interbedded) sand according to the Thomas-
Stieber-Yin-Marion model. ................................................................................... 81
xxi
Figure 3.21: Comparisons of fluid substitution results which are color-coded by sand
fractions estimated from the Thomas-Stieber-Yin-Marion model. Four fluid
substitution procedures are used: Gassmann’s equation (i.e., ignoring the effect of
sand-shale interbedding), Gassmann shaly-sand equation (i.e., assuming that shale
is all dispersed; Dvorkin et al., 2007), the downscaling-upscaling procedure, and
the mesh method. Note that points with very low sand fractions are excluded from
the plot. Lines X = Y are super-imposed for comparison purpose. ...................... 81
Figure 3.22: Comparisons of fluid substitution results which are color-coded by sand
fractions estimated from the P-compliance versus effective porosity plane. Four
fluid substitution procedures are used: Gassmann’s equation (i.e., ignoring the
effect of sand-shale interbedding), Gassmann shaly-sand equation (i.e., assuming
that shale is all dispersed; Dvorkin et al., 2007), the downscaling-upscaling
procedure, and the mesh method. Note that points with very low sand fractions
are excluded from the plot. Lines X = Y are super-imposed for comparison
purpose. ................................................................................................................. 83
Figure 3.23: Instability of the inversion step in the downscaling-upscaling procedure
leading to spikes in P-wave velocity after fluid substitution. ............................... 84
Figure 3.24: Fluid substitution results by four different procedures for shaly sands,
with no interbedding. The procedures are Gassmann using total porosity and
ignoring interbedding, Gassmann using total porosity (applied to sand layers only)
+ upscaling, Gassmann using effective porosity by Dvorkin et al., 2007 (applied
to each sand layers only) + upscaling and our method (i.e., the mesh). For this
synthetic case, volume fractions of dispersed clay in the sand range from
𝑉𝑑𝑖𝑠𝑝 =0 to 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑠. Both bulk and shear moduli are used for this synthetic
example. ................................................................................................................ 85
Figure 3.25: Modified upper bounds for modeling the clean-sand diagenetic trend.
This example represents a quartz-water system. ................................................... 87
Figure 3.26: Percentage differences between the baselines and the predicted velocities
by the upscaled Gassmann’s equation (left) and the Gassmann shaly-sand
equation (Dvorkin et al., 2007) (right) for the first scenario, where P-wave moduli
and approximate Gassmann’s equation are used for the upscaled Gassmann. Each
location on the triangular diagram represents an interbedded sand-shale sequence
with a unique pair of sand fraction and volume fraction of dispersed clay values.
Note that the color scales of the two panels are different. .................................... 96
Figure 3.27: Percentage differences between the baselines and the predicted velocities
by the upscaled Gassmann’s equation (left) and the Gassmann shaly-sand
equation (Dvorkin et al., 2007) (right) for the second scenario, where both bulk
and shear moduli, and approximate Gassmann’s equation are used in calculations.
Each location on the triangular diagram represents an interbedded sand-shale
sequence with a unique pair of sand fraction and volume fraction of dispersed
clay values. Note that the color scales of the two panels are different. ................ 96
Figure 3.28: Percentage differences between the baselines and the predicted velocities
by the upscaled Gassmann’s equation (left) and the Gassmann shaly-sand
equation (Dvorkin et al., 2007) (right) for the second scenario, where both bulk
and shear moduli, and the actual Gassmann’s equation are used for the upscaled
xxii
Gassmann. Each location on the triangular diagram represents an interbedded
sand-shale sequence with a unique pair of sand fraction and volume fraction of
dispersed clay values. Note that the color scales of the two panels are different.
The color scale on the left panel does not cover either the actual maximum or
minimum values of the percentage differences. The color scale is adjusted to
show values of the majority of the results. Those results indicated by a magenta
ellipse are unreliable due to instability of the method at low effective porosity. . 97
Figure 3.29: Comparisons of five sets of fluid substitution results which are color-
coded by sand fractions estimated from the Thomas-Stieber-Yin-Marion model.
Five fluid substitution procedures are used: Gassmann’s equation (i.e., ignoring
the effect of sand-shale interbedding), Gassmann shaly-sand equation (i.e.,
assuming that shale is all dispersed; Dvorkin et al., 2007), the downscaling-
upscaling procedure, the mesh method, and the upscaled Gassmann’s equation.
Note: points with very low sand fractions are excluded from the plot. Lines X = Y
are superimposed for comparison purpose. ......................................................... 100
Figure 3.30: Sensitivity analysis results of the upscaled Gassmann for six input
parameters: P-wave velocity of wet porous shale, volume fraction of dispersed
clay, sand fraction, effective water saturation, bulk modulus of fluid, and effective
porosity of the interbedded package. The thick lines on the top of each subplot are
the oil-saturated velocity and the reference velocities after fluid substitution from
oil to brine using the downscaling-upscaling procedure and the upscaled
Gassmann with all correct input parameters. Results of sensitivity analysis for
each parameter are shown as a velocity distribution, which is normalized to one.
Sand fractions and volume fractions of dispersed clay for each synthetic model
are shown on the top of each subplot. ................................................................. 102
Figure 4.1: Overall workflow for seismic-signature study and property estimation.
First, thin sand-shale sequences are generated using Markov-chain models and
rock physics relations. Then, seismic responses for these sequences are modeled,
and seismic attributes are extracted. Finally, the attributes are related to reservoir
properties, which can be used for reservoir characterization of target areas. ..... 112
Figure 4.2: Three examples of transition matrices with fixed sampling intervals:
retrogradational, progradational, and aggradational sequences. The lithologic
states in the transition matrices are sand (s), shaly sand (sh-s) sandy shale (s-sh)
and shale (sh). The off diagonal elements marked by arrows control the
directionality of the sequences. ........................................................................... 114
Figure 4.3: Examples of an embedded-form transition matrix with realizations of
sequences. The lithologic states in the transition matrix are sand (s), shaly sand
(sh-s) sandy shale (s-sh) and shale (sh). An example of thickness distributions
used is shown in the lower left corner................................................................. 115
Figure 4.4: Examples of SP log responses showing stacking patterns in parasequence
sets. (Left) Retrogradational, (middle) progradational and (right) aggradational
patterns (Modified after Van Wagoner et al., 1990). .......................................... 116
Figure 4.5: Illustrations of sand-shale mixtures, with their porosity and velocity values
related to clay content (Modified after Marion et al., 1992). Porosity versus clay
xxiii
content shows a V-shaped trend, where the two end points are the pure sand and
pure shale porosity. Selected clay fractions corresponding to the four lithologic
states are marked. The states are sand (s), shaly-sand (sh-s), sandy-shale (s-sh),
and shale (sh). ..................................................................................................... 117
Figure 4.6: Wavelet-transform analysis for extracting attributes from a seismogram.
(Left) modulus of wavelet-transform coefficients for the seismogram. The red
arrow indicates the modulus along a particular scale. (Right) a log-log plot of
scale versus variance of modulus of wavelet-transform coefficients. The plot is
shown in open circles which are fit by a straight line. The slope and intercept of
this line are used as seismic attributes. ............................................................... 119
Figure 4.7: Modulus of wavelet coefficients (shown in the middle two boxes) of two
seismogram segments. Dissimilarity between these seismograms is defined as
distance between their modulus maps. ................................................................ 120
Figure 4.8: Workflow for generating seismic responses of multiple realizations of thin
sand-shale sequences. The four lithologic states in the transition matrix are sand
(s), shaly-sand (sh-s), sandy-shale (s-sh), and shale (sh). ................................... 125
Figure 4.9: Distributions of net-to-gross ratios from sequence realizations generated
using the transition matrices in Table 4.2. Red lines indicate values of limiting
sand distributions. From top left to bottom right, values of the parameter k defined
in the transition matrices are 0.45, 0.55, 0.65, 0.75, 0.85 and 0.95, respectively.
............................................................................................................................. 126
Figure 4.10: Slope and intercept of wavelet transforms of seismic responses for 2
different standard deviations of velocity distributions: (left) v = 0.1 km/s and
(right) v = 0.2 km/s. Points correspond to 6 sets of 200 realizations generated
from different transition matrices with various net-to-gross ratios (𝝅sand). Sw is
0.1 for sand layers and 1 for the others. .............................................................. 127
Figure 4.11: Probability distributions of slope and intercept values corresponding to
data points in Figure 4.10. (Left column) v = 0.1 km/s and (right column) v =
0.2 km/s. The direction of increasing net-to-gross ratios (𝝅sand) is to the left for all
plots. .................................................................................................................... 127
Figure 4.12: Classical MDS results for studying effect of net-to-gross ratios on seismic
signature. (Top left) Dissimilarity matrix showing pairwise Euclidean distances
between any two seismograms from realizations with standard deviations of
velocities equal to 0.1 km/s. (Bottom left) Correlation coefficient between
dissimilarity and distance (between points in new coordinates resulting from
classical MDS) versus numbers of included coordinates. (Top right) Classical
MDS results with the first two coordinates. Each point represents one seismogram,
color-coded by its true net-to-gross value from its corresponding sand-shale
sequence. (Bottom right) Classical MDS results with the first two coordinates
converted into polar coordinates. ........................................................................ 131
Figure 4.13: Metric MDS results for studying effect of net-to-gross ratios on seismic
signature. (Top left) Dissimilarity matrix showing pairwise (Euclidean) distances
between any two seismograms from realizations with standard deviations of
velocities equal to 0.1 km/s. (Bottom left) Correlation coefficient between
dissimilarity and distance (between points in new coordinates resulting from
xxiv
metric MDS) versus numbers of included coordinates. (Top right) Metric MDS
results with the first two coordinates. Each point represents one seismogram,
color-coded by its true net-to-gross value from its corresponding sand-shale
sequence. (Bottom right) Metric MDS results with the first two coordinates
converted into polar coordinates. ........................................................................ 132
Figure 4.14: Non-metric MDS results for studying effect of net-to-gross ratios on
seismic signature. (Top left) Dissimilarity matrix: pairwise (Euclidean) distances
between any two seismograms from realizations with standard deviations of
velocities equal to 0.1 km/s. (Bottom left) Correlation coefficient between
dissimilarity and distance (between points in new coordinates resulting from non-
metric MDS) versus numbers of included coordinates. (Top right) Non-metric-
MDS results with the first two coordinates. Each point represents one seismogram,
color-coded by its true net-to-gross value from its corresponding sand-shale
sequence. (Bottom right) Non-metric MDS results with the first two coordinates
converted into polar coordinates. ........................................................................ 133
Figure 4.15: Gaussian KPCA results for studying effect of net-to-gross ratios on
seismic signature. (Left) Kernel matrix using a Gaussian kernel. Each element in
the matrix (K(xi,xj)) corresponds to the Gaussian kernel function evaluated using
a pair of seismograms (xi,xj) from realizations with standard deviations of
velocities equal to 0.1 km/s. (Right) Projections of seismograms, which
correspond to sequences with different net-to-gross ratios, onto the first two
principal components from the Gaussian KPCA. Each point is color-coded by the
net-to-gross ratio from its corresponding sand-shale sequence. ......................... 134
Figure 4.16: Dynamic similarity KPCA results for studying effect of net-to-gross
ratios on seismic signature. (Left) Kernel matrix using a dynamic similarity
kernel. Each element in the matrix (K(xi,xj)) corresponds to the dynamic
similarity kernel function evaluated using a pair of seismograms (xi,xj) from
realizations with standard deviations of velocities equal to 0.1 km/s. (Right)
Projections of seismograms, which correspond to sequences with different net-to-
gross ratios, onto the first two principal components from the dynamic similarity
KPCA. Each point is color-coded by the net-to-gross ratio from its corresponding
sand-shale sequence. ........................................................................................... 134
Figure 4.17: Inverse multi-quadric KPCA results for studying effect of net-to-gross
ratios on seismic signature. (Left) Kernel matrix using an inverse multi-quadric
kernel. Each element in the matrix (K(xi,xj)) corresponds to the inverse multi-
quadric kernel function evaluated using a pair of seismograms (xi,xj) from
realizations with standard deviations of velocities equal to 0.1 km/s. (Right)
Projections of seismograms, which correspond to sequences with different net-to-
gross ratios, onto the first two principal components from the inverse multi-
quadric KPCA. Each point is color-coded by the net-to-gross ratio from its
corresponding sand-shale sequence .................................................................... 135
Figure 4.18: Polynomial KPCA results for studying effect of net-to-gross ratios on
seismic signature. (Left) Kernel matrix using a polynomial kernel. Each element
in the matrix (K(xi,xj)) corresponds to the polynomial kernel function evaluated
using a pair of seismograms (xi,xj) from realizations with standard deviations of
xxv
velocities equal to 0.1 km/s. (Right) Projections of seismograms, which
correspond to sequences with different net-to-gross ratios, onto the first two
principal components from the polynomial KPCA. Each point is color-coded by
the net-to-gross ratio from its corresponding sand-shale sequence. ................... 135
Figure 4.19: Linear KPCA results for studying effect of net-to-gross ratios on seismic
signature. (Left) Kernel matrix using a linear kernel. Each element in the matrix
(K(xi,xj)) corresponds to the linear kernel function evaluated using a pair of
seismograms (xi,xj) from realizations with standard deviations of velocities equal
to 0.1 km/s. (Right) Projections of seismograms, which correspond to sequences
with different net-to-gross ratios, onto the first two principal components from the
linear KPCA. Each point is color-coded by the net-to-gross ratio from its
corresponding sand-shale sequence. ................................................................... 136
Figure 4.20: Change in MDS classification success rate when the number of
coordinates included as net-to-gross attributes increases. Results from three MDS
methods are shown. ............................................................................................. 136
Figure 4.21: Change in KPCA classification success rate when the number of principal
components included as net-to-gross attributes increases. .................................. 138
Figure 4.22: Parallel coordinate plot. (Left) Polylines of the first five principal
components from the Gaussian KPCA. Each line is color-coded by its
corresponding net-to-gross class. Three classes are <0.38, 0.38-0.465, and >0.465.
(Right) The solid lines are the median (i.e., the 0.5-quantile) of the component
values. The dash lines surrounding the median are the 0.45- and 0.55- quantiles.
............................................................................................................................. 138
Figure 4.23: Three selected matrices with the same limiting distribution and their
sample sequences. From left to right, sand layers become more blocky (i.e.,
groups of sand layers become thicker). The lithologic states are sand (s), shaly
sand (sh-s) sandy shale (s-sh) and shale (sh). Note that each row of matrix B is
equal to a fixed probability vector, and thus the lithologic states generated using
this matrix are considered as independent random events (i.e., the current state
has no dependency on the previous states). ........................................................ 139
Figure 4.24: Slope and intercept attributes for varying water saturation within the sand
layers. Results in each plot are from sequences generated by using the transition
matrix shown in the lower-right corner of each plot. From left to right, the
transition matrices correspond to A, B, and C in Figure 4.23............................ 143
Figure 4.25: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the first two coordinates from classical
MDS. Each point is color-coded by the water-saturation value of the sand layers.
From left to right, subplots correspond to transition matrices A, B, and C,
respectively. ........................................................................................................ 143
Figure 4.26: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the two coordinates from metric MDS.
Each point is color-coded by the water-saturation value of the sand layers. From
left to right, subplots correspond to transition matrices A, B, and C, respectively.
............................................................................................................................. 143
xxvi
Figure 4.27: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the two coordinates from non-metric MDS.
Each point is color-coded by the water-saturation value of the sand layers. From
left to right, subplots correspond to transition matrices A, B, and C, respectively.
............................................................................................................................. 144
Figure 4.28: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the first two principal components from the
Gaussian KPCA. Each point is color-coded by the water-saturation value of the
sand layers. From left to right, subplots correspond to transition matrices A, B,
and C, respectively. ............................................................................................. 144
Figure 4.29: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the first two principal components from the
dynamic similarity KPCA. Each point is color-coded by the water-saturation
value of the sand layers. From left to right, subplots correspond to transition
matrices A, B, and C, respectively. ..................................................................... 144
Figure 4.30: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the first two principal components from the
inverse multi-quadric KPCA. Each point is color-coded by the water-saturation
value of the sand layers. From left to right, subplots correspond to transition
matrices A, B, and C, respectively. ..................................................................... 145
Figure 4.31: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the first two principal components from the
polynomial KPCA. Each point is color-coded by the water-saturation value of the
sand layers. From left to right, subplots correspond to transition matrices A, B,
and C, respectively. ............................................................................................. 145
Figure 4.32: First 21 eigenvalues from classical MDS, color-coded by the transition
matrices used in generating sand-shale sequences for investigating the effect of
saturations. .......................................................................................................... 145
Figure 4.33: Correlation coefficient between dissimilarity and distance (between
points in new coordinates resulting from an MDS algorithm) versus numbers of
included coordinates from MDS results for investigating the effect of saturations.
(Left) metric MDS and (right) non-metric MDS. ............................................... 146
Figure 4.34: Change in MDS classification success rate when the number of
coordinates included as saturation attributes increases. Three MDS algorithms are
compared. Results when using sequences generated by transition matrices A (left),
B (middle) and C (right) are shown. ................................................................... 146
Figure 4.35: Change in KPCA classification success rate as the number of principal
components included as attributes increases, when using sequences generated by
transition matrices A. .......................................................................................... 146
Figure 4.36: Change in KPCA classification success rate as the number of principal
components included as attributes increases, when using sequences generated by
transition matrices B. .......................................................................................... 147
Figure 4.37: Change in KPCA classification success rate as the number of principal
components included as attributes increases, when using sequences generated by
transition matrices C. .......................................................................................... 147
xxvii
Figure 4.38: Component-wise analysis for the Gaussian KPCA. (Left) Parallel
coordinate plot of the first ten principal components of the Gaussian KPCA for
matrix B. The solid lines are the median (the 0.5 quantile). The dash lines
surrounding the median are the 0.45 and 0.55 quantiles. (Right) Projections of
seismograms onto the first and sixth principal components of the Gaussian KPCA.
............................................................................................................................. 148
Figure 4.39: Component-wise analysis for the inverse multi-quadric KPCA. (Left)
Parallel coordinate plot of the first ten principal components of the inverse multi-
quadric KPCA for matrix C. The solid lines are the median (the 0.5 quantile). The
dash lines surrounding the median are the 0.45 and 0.55 quantiles. (Right)
Projections of seismograms onto the first and sixth principal components of the
inverse multi-quadric KPCA. .............................................................................. 148
Figure 4.40: Sample sequences from three selected transition matrices with the same
limiting distribution: [.25 .25 .25 .25]. From left to right, columns represent
retrogradational, progradational, and aggradational stacking patterns, respectively.
Red arrows schematically indicate transitions from coarse to fine grains, and vice
versa. Note that other interpretations of transitional patterns are possible. The
lithologic states are sand (s), shaly sand (sh-s) sandy shale (s-sh) and shale (sh).
............................................................................................................................. 150
Figure 4.41: Parallel coordinate plot of the first ten components of the Gaussian
KPCA result for investigating the effect of stacking pattern on seismic signatures.
The solid lines are the median (i.e., the 0.5-quantile) of the component values.
The dash lines surrounding the median are the 0.45- and 0.55- quantiles. All lines
are color-coded by the types of stacking patterns. .............................................. 151
Figure 4.42: Sample sequences using the embedded-form transition matrices shown on
the left with varying averages of exponentially-distributed lithologic thicknesses
along the sequences. From (A) to (C), the three stacking patterns are
retrogradational (overall thinning- and fining-upward), progradational (overall
thickening- and coarsening-upward), and aggradational stacking patterns in
parasequence sets, respectively. Red arrows schematically show a series of
progradational parasequences within retrogradational, progradational, and
aggradational parasequence sets. Note that other interpretations of such
parasequence patterns are possible The lithologic states are sand (s), shaly sand
(sh-s) sandy shale (s-sh) and shale (sh). .............................................................. 154
Figure 4.43: Change in KPCA classification success rate when the number of principal
components included as attributes increases. ...................................................... 157
Figure 4.44: Parallel coordinate plot of the first ten coordinates of the classical MDS
result. The solid lines are the median (i.e., the 0.5-quantile) of the component . 157
Figure 4.45: Projections of seismograms with added noise onto the first two principal
components from the dynamic similarity KPCA. Each projected point is color-
coded by the water-saturation value of the sand layers within the corresponding
sequence. Percentages of noise added are specified in each panel. The success
rate of classifying a projected seismogram into its corresponding saturation class
is shown in the lower left corner of each panel................................................... 159
xxviii
Figure 4.46: A realization of one 2-D geologic section used in the numerical example.
The area at the left end marked with the red box corresponds to the well location.
Sand and shale are colored in yellow and blue respectively. The thicknesses of
sand layers decrease linearly away from the well. The total thickness of reservoir
is 150 m. An example of the thickness distribution used to simulate the sequence
at the well location is also shown........................................................................ 161
Figure 4.47: Results from realizations of 1,000 2-D sections show how slope and
intercept vary with varying net-to-gross ratios. .................................................. 162
Figure 4.48: Contour plots and marginal distributions for the high (i.e., equal to or
greater than 0.6) and low (i.e., equal to or less than 0.35) net-to-gross values. .. 162
Figure 4.49: (Lower left corner) Posterior distributions of net-to-gross ratios for three
selected locations on the unknown seismic section labeled as (1), (2) and (3). The
true values for each location are shown in the table at the lower right corner. ... 164
Figure 4.50: Two transition matrices used to generate non-stationary sequences and
examples of sequence realizations. For simulation steps in the top half of the
reservoir interval, the probability of using matrix A is 0.8. The same is true for
the bottom half of the reservoir with matrix B. The lithologic states considered
are sand (S) and shale (Sh). ................................................................................. 166
Figure 4.51: Segmentation of seismograms and sequences. (Left) application of an 80-
ms moving window to a seismogram (Right) seismograms segments and their
corresponding sequence segments. ..................................................................... 166
Figure 4.52: The first two components of KPCA with Gaussian kernel. A total of 1200
seismogram segments are represented as points which are color-coded by the net-
to-gross values of the corresponding sequence segments. .................................. 167
Figure 4.53: The first two components of KPCA with Gaussian kernel. A total of 2070
seismogram segments generated from the same forms of transition matrices as
described in the text are represented as points which are color-coded by the net-
to-gross values of the corresponding sequence segments. Nine seismogram
segments are plotted in black and treated as unknowns. ..................................... 169
Figure 4.54: Probability distributions of local net-to-gross estimation for the three
(unknown) seismograms. Each row represents the results for each unknown
seismogram. Each column represents local net-to-gross estimations at a specified
position of the sequences which correspond to the unknown seismograms. The
red dash lines represent the true net-to-gross values. .......................................... 169
Figure 4.55: Amplitude attributes extracted from seismograms which correspond to
sequences with different net-to-gross ratios. ....................................................... 171
Figure 4.56: Success rates in classifying a data point into three net-to-gross classes
(<0.38, 0.38-0.465, and >0.465) by using the attributes specified on the horizontal
axis. The rates are shown on top of each bar. Att1: RMS and absolute amplitude at
the top of the reservoir, Att 2: the first two principal components of Gaussain
KPCA, Att3: the first two principal components of dynamic similarity KPCA, and
Att4: the first 20 principal components of linear KPCA. .................................... 171
Figure 5.1: Six identified lithofacies from the study area offshore Equatorial Guinea,
West Africa (Lowe, 2004; Dutta, 2009). ............................................................ 182
xxix
Figure 5.2: Well A from deep-water turbidite deposits, offshore Equatorial Guinea,
West Africa. The zone of interest is highlighted. From left to right, the curves are
gamma ray, bulk density, density-derived porosity, P-wave velocity, and water
saturation (SW), respectively. .............................................................................. 184
Figure 5.3: Petrophysical analysis of the selected well-log interval using the Thomas-
Stieber model. (Left) Total porosity and gamma ray values for three lithofacies in
the selected interval. The median total porosity and gamma ray values for each
lithofacies are shown in solid circles. From these median points, the up-down or
left-right bars indicate the interquartile ranges (i.e., from 1st to 3
rd quartiles) of
each property. A Thomas-Stieber diagram is also superimposed on the data.
(Right) Variation of gamma ray values with depth. Data points are color-coded by
lithofacies similar to the left panel. The gamma ray log shows an upward-fining
trend. ................................................................................................................... 185
Figure 5.4: Estimated net-to-gross ratios for the selected well-log interval. (Left)
gamma ray log of the selected interval and (right) posterior distributions of
estimated net-to-gross ratios. .............................................................................. 185
Figure 5.5: Acoustic impedance and the corresponding synthetic seismogram of the
target zone. (Left) Acoustic impedance of the target zone shown in magenta,
together with 200 data points above and 200 data points below the target zone.
(Right) Synthetic seismogram generated using a 30-Hz zero-phase Ricker wavelet.
............................................................................................................................. 187
Figure 5.6: Synthetic-seismogram test set. Note that each panel shows four repetitions
of one seismogram. From left to right, the panels show seismograms that are
generated from the original well log (i.e., fining-upward), the inverted log (i.e.,
coarsening-upward), the shuffled log (i.e., serrated), the brine-saturated log, and
the log with smaller layer thicknesses within the target zone. ............................ 187
Figure 5.7: Sample realizations of synthetic earth models which cover a range of net-
to-gross ratios and various stacking patterns. The sequence realizations are
generated from the three categories of transition matrices: with an upward-
increasing trend (left), with an upward-decreasing trend (middle), and with no
trend in net-to-gross ratios (right). ...................................................................... 190
Figure 5.8: Probability density functions of acoustic impedance for all lithologic states:
oil sand (blue), shale (red), and wet sand (black). .............................................. 190
Figure 5.9: Synthetic seismograms corresponding to five sets of sand-shale sequences.
Model I to V correspond to sequence realizations generated from various
transition matrices, sand properties, and layer thicknesses. Refer to Table 5.1 for
detailed specifications. ........................................................................................ 192
Figure 5.10: Projections of the noise-free seismograms from the training set and the
test set onto the first two principal components after the application of the linear
KPCA. Each point from the training set is color-coded by the net-to-gross ratio
from its corresponding sand-shale sequence. The test set, shown in black symbols,
is plotted on top of the projections of the training set. ........................................ 193
Figure 5.11: Cross-plot between the RMS and the maximum amplitude of noise-free
seismograms from both the training set and the test set. The points from the
training set are color-coded by the net-to-gross ratios from their corresponding
xxx
sand-shale sequences. The points from the test set are shown in black symbols.
Two distinct trends corresponding to changes in stacking pattern and net-to-gross
ratio are marked by the blue and the red arrows, respectively. ........................... 194
Figure 5.12: Projections of the noisy seismograms from the test set onto the first two
principal components from the linear KPCA. The test set shown in black symbols
is plotted on top of the projections of the training set, which are color-coded by
the net-to-gross ratios from their corresponding sand-shale sequences. The
percentages of noise added to the seismograms are shown in the lower left corner
of each plot. ......................................................................................................... 196
Figure 5.13: Cross-plot between the RMS and the maximum amplitude of noisy
seismograms from both the training set and the test set. The points from the
training set are color-coded by the net-to-gross ratios from their corresponding
sand-shale sequences, and the points from the test set are shown in black symbols.
The percentages of noise added to the seismograms are shown in the lower left
corner of each plot. .............................................................................................. 197
Figure 5.14: Linear KPCA results (0% noise) with net-to-gross (NTG)/stacking-
pattern class labels. NTG1, NTG2, and NTG3 represent NTG≤0.46, NTG>0.46
and NTG≤0.54, and NTG>0.54, respectively. The three stacking patterns are
fining-upward (F), coarsening-upward (C), and serrated (S). ............................. 197
Figure 5.15: Classification success rate when using linear KPCA attributes, amplitude
attributes, and dynamic similarity KPCA attributes at various noise levels. ...... 198
Figure 5.16: Effect of layer thickness on linear KPCA results. Each plot represents the
projections of seismograms onto the first two principal components of linear
KPCA when layer thickness is 1.5 cm (left) or 2.5 cm (right). Points are color-
coded by the net-to-gross ratios from their corresponding sand-shale sequences.
............................................................................................................................. 199
Figure 5.17: Effect of layer thickness on amplitude results (RMS versus maximum
amplitude). Layer thickness used for modeling sequences is 1.5 cm for the left
plot and 2.5 cm for the right plot. Points are color-coded by the net-to-gross ratios
from their corresponding sand-shale sequences. ................................................. 199
Figure 5.18: Probability density functions of two sets of acoustic impedance showing
small and large contrast (shown in red and green, respectively) between the sand
and the shale (shown in solid and dash lines, respectively). ............................... 200
Figure 5.19: Effect of impedance contrast on linear KPCA results. Each plot
represents the projections of seismograms onto the first two principal components
of linear KPCA when the overlap between the sand and the shale impedances is
small (left) or large (right). Points are color-coded by the net-to-gross ratios from
their corresponding sand-shale sequences. ......................................................... 201
Figure 5.20: Effect of impedance contrast on amplitude results (RMS versus
maximum amplitude). The overlap between the sand and the shale impedances is
small in the left panel and large in the right panel. Points are color-coded by the
net-to-gross ratios from their corresponding sand-shale sequences. ................... 201
Figure 5.21: Effect of phase rotation on linear KPCA results. Points represent
projections of the seismograms from the Model I training set and the well location
onto the first two principal components from the linear KPCA. Each point is
xxxi
color-coded by the net-to-gross ratio from its corresponding sand-shale sequence.
The three seismograms at the well with three different phases shown in black
symbols are plotted on top of the projections of the training set. The projection of
the zero trace is used as a reference point. .......................................................... 202
Figure 5.22: A 2-D seismic section extracted along the turbidite channel from
proximal (left) to distal (right) directions. The location of well A is marked. .... 203
Figure 5.23: Prior distributions for thickness (left) and net-to-gross ratio (right),
assigned to sequence models when generating a training set. ............................ 204
Figure 5.24: Relative locations of two existing wells (well A and well B). Well A
(black symbol) is located in the distal direction along the channel, and well B (red
symbol) is located in the proximal direction along the channel. Distance is
measured from a reference point. The selected 2-D seismic line is shown in blue.
............................................................................................................................. 205
Figure 5.25: Average properties of the overburden and underburden of the interval of
interest from proximal and distal wells. Three properties are shown: P-wave
velocity (Vp), bulk density (RHOB), and acoustic impedance (AI). .................. 206
Figure 5.26: Probability density functions of acoustic impedance for the sand (blue)
and the shale (red) lithologic states used in generating a training set. ................ 207
Figure 5.27: Projections of the training set and the test seismograms onto the first two
principal components of KPCA with a linear kernel. The training points are color-
coded by the reservoir thicknesses of their corresponding sand-shale sequences.
The test points are labeled by their trace numbers. ............................................. 208
Figure 5.28: Projections of the training set onto the first two principal components of
KPCA with a linear kernel. In each panel, the training points are color-coded by
the reservoir parameter specified at the top of each panel. ................................. 209
Figure 5.29: Projections of the training set onto the first two principal components of
KPCA with a Gaussian kernel. In each panel, the training points are color-coded
by the reservoir parameter specified at the top of each panel. ............................ 209
Figure 5.30: Projections of the training set onto the first two principal components of
KPCA with a linear kernel, when the values of underburden multiplier,
overburden multiplier, and reservoir thickness are fixed. Points are color-coded
by net-to-gross ratio. ........................................................................................... 210
Figure 5.31: The first 20 eigenvalues of the linear KPCA results. ............................. 211
Figure 5.32: Comparison of two empirical cumulative distributions for the overburden
multiplier values in class 3. The prior distribution and the class-conditional
distribution are shown in blue and red, respectively. The shaded area represents
area between the two curves. .............................................................................. 213
Figure 5.33: Normalized difference measure between a class-conditional empirical
cumulative distribution and a prior empirical distribution for a parameter. The red
line represents a threshold used in determining whether a parameter has a
significant impact on the output response (i.e., seismic signature). .................... 214
Figure 5.34: Pareto plot of estimated sensitivity values. Parameters are ranked
according to their sensitivity values. ................................................................... 214
Figure 5.35: Estimated probability density functions for reservoir thickness for all
seismic traces in the test set. The last three traces (#55-57) correspond to the
xxxii
additional three traces which are added for evaluating the performance of our
property estimation. Each density function is shown as a column-wise color scale.
............................................................................................................................. 216
Figure 5.36: The mean values of reservoir thickness for all seismic traces in the test set.
The last three points are added for evaluating the performance of our property
estimation. ........................................................................................................... 216
Figure 5.37: Estimated probability density functions for the underburden multiplier for
all seismic traces in the test set. The last three traces (#55-57) correspond to the
additional three traces which are added for evaluating the performance of our
property estimation. Each density function is shown as a column-wise color scale.
............................................................................................................................. 216
Figure 5.38: The mean values of underburden multiplier for all seismic traces in the
test set. The last three points are added for evaluating the performance of our
property estimation. ............................................................................................ 216
Figure 5.39: Estimated probability density functions for the overburden multiplier for
all seismic traces in the test set. The last three traces (#55-57) correspond to the
additional three traces which are added for evaluating the performance of our
property estimation. Each density function is shown as a column-wise color scale.
............................................................................................................................. 217
Figure 5.40: The mean values of overburden multiplier for all seismic traces in the test
set. The last three points are added for evaluating the performance of our property
estimation. ........................................................................................................... 217
Figure 5.41: Estimated probability density functions for net-to-gross ratio for all
seismic traces in the test set. The last three traces (#55-57) correspond to the
additional three traces which are added for evaluating the performance of our
property estimation. Each density function is shown as a column-wise color scale.
............................................................................................................................. 217
Figure 5.42: The mean values of net-to-gross ratio for all seismic traces in the test set.
The last three traces (#55-57) correspond to the additional three traces which are
added for evaluating the performance of our property estimation. ..................... 217
Figure 5.43: The estimated probability densities of reservoir thickness for seismic
traces at the well and trace #55-57, which are added for evaluating the
performance of our property estimation. The true thicknesses for all four test
traces are marked by the magenta line. ............................................................... 218
Figure 5.44: The estimated probability densities of underburden multiplier for seismic
traces at the well and trace #55-57, which are added for evaluating the
performance of our property estimation. The true multipliers for three test traces
are marked by the lines shown in the same colors as their densities. Note that
since trace #56 is associated with well B, the “true” multiplier of well A is
irrelevant and thus not shown here. .................................................................... 218
Figure 5.45: The estimated probability densities of overburden multiplier for seismic
traces at the well and trace #55-57, which are added for evaluating the
performance of our property estimation. The true multipliers for three test traces
are marked by the lines shown in the same colors as their densities. Note that
xxxiii
since trace #56 is associated with well B, the “true” multiplier of well A is
irrelevant and thus not shown here. .................................................................... 219
Figure 5.46: The estimated probability densities of net-to-gross ratio for seismic traces
at the well and trace #55-57, which are added for evaluating the performance of
our property estimation. The true net-to-gross ratios for all four test traces are
marked by the magenta line. ............................................................................... 219
Figure 5.47: The estimated probability densities of net-to-gross ratio for seismic trace
# 57 for both when the non-stationarity is taken into account (blue curve) and
when the non-stationarity is not taken into account (red curve). The true net-to-
gross ratio for trace #57 is marked by the blue line. ........................................... 220
xxxiv
1
Chapter 1
Introduction
1.1 Objective
The overall goal of this dissertation is to improve the interpretation and property
estimation of thinly bedded sand-shale systems that cannot be resolved with
conventional well log and seismic data by modifying existing rock physics relations
and establishing new ones to account for heterogeneity of rocks at the measurement
scales, and by using quantitative seismic analysis which combines rock physics,
spatial statistics of the geology, and seismic attributes to infer rock properties of the
sub-resolution sand-shale systems. The specific objectives are as follows:
To incorporate natural variations and uncertainties in real observations into
existing models for estimating petrophysical properties of thinly bedded
sand-shale reservoirs;
CHAPTER 1: Introduction 2
To establish rock physics relations that can be directly applied to
heterogeneous rocks at the measurement scale, without the need to
downscale the measurements;
To investigate the effects of rock properties and stacking patterns on
seismic signatures of thinly bedded sand-shale reservoirs;
To estimate reservoir properties (e.g., volume fraction of sand, saturation,
and stacking patterns) of thinly bedded, sub-resolution sand-shale
reservoirs from seismic data.
1.2 Background and motivation
Definitions of thin beds vary depending on the context. In both formation
evaluation analysis and seismic study, thin beds generally refer to beds with
thicknesses below the vertical resolution of measuring tools. This is the definition we
follow in this dissertation. The resolution varies depending on measuring tools. While
conventional logging tools have typical vertical resolutions of a few feet, the vertical
resolution of seismic data, generally defined as a quarter of the dominant wavelength,
is typically in the range of tens of meters. Geologically, thin beds vary in thickness
from 3 to 10 cm. Layers with thicknesses less than 1 cm are referred to as laminae
(Campbell, 1967; Boggs, 2001; Passey et al., 2006).
The problem of interpreting thin beds has received much study. For example,
thinly interbedded sand-shale sequences can cause low-resistivity pay, a well-known
phenomenon in formation evaluation. The thin hydrocarbon-bearing sand layers have
high resistivity, while the thin wet shale layers have low resistivity. When resistivity
logging devices cannot resolve each individual layer, the resistivity measurements are
averages of the thin sand and shale layers (arithmetic or harmonic average, depending
on the direction of the current flow). Due to this averaging, the resistivity
measurements of the hydrocarbon interval can appear to be lower than they actually
are (Boyd et al., 1995; Passey et al., 2006). Core samples provide direct indicators of
thin beds (Passey et al., 2006), but their use is often limited because of their high
operating cost and limited availability. Other logs, such as high-resolution fullbore
CHAPTER 1: Introduction 3
formation microimager (FMI) and nuclear magnetic resonance log (NMR), provide
indirect indicators of thin beds. However, these logs are still subject to environmental
effects, inaccurate calibration, and improper processing (Passey et al., 2006).
Moreover, even though well log data provide high-resolution information about
reservoirs, they lack spatial coverage. Therefore, the need remains for better
interpretation of conventional well log and seismic data in thin sub-resolution bedded
reservoirs. This dissertation attempts to improve such interpretation by applying rock
physics relations and quantitative seismic interpretation.
1.2.1 Rock physics relations and sub-resolution heterogeneity
Rock physics relations are often obtained from controlled laboratory experiments
or derived for rocks that are homogeneous at particular scales. However, in practice,
these relations are often applied to measurements at scales where the rocks are
heterogeneous. When the scale of heterogeneity is smaller than what geophysical tools
can resolve, the measurements represent average values of the heterogeneous rocks
(for example, a stack of interbedded sand-shale layers that cannot be individually
resolved by well logging tools). Applying rock physics relations to such measurements
directly at the measurement scale can result in erroneous predictions. For example,
Dvorkin and Uden (2006) demonstrated that when the rock physics transform between
well-log-scale impedance and hydrate saturation is applied to seismic-scale
impedances of a 1-D synthetic earth model which contains three sand layers with
methane hydrate, the predicted hydrate saturations can significantly differ from the
true saturations. Another case of when rock physic relations fail to account for
heterogeneity is the application of the isotropic Gassmann’s fluid substitution equation
(Gassmann, 1951) to the measurements at their original scales, without accounting for
the sub-resolution sand-shale interbedding. In this case, the predicted changes in
elastic properties can be erroneous because Gassmann’s equation is not appropriate for
fluid substitution in the low-permeability shale layers (e.g., Katahara, 2004; Skelt,
2004a; Skelt, 2004b; Chopra, 2005). This dissertation aims to investigate existing rock
physics relations to account for heterogeneity, and establish new relations that can be
CHAPTER 1: Introduction 4
directly applied to heterogeneous rocks without the need to downscale the
measurements.
1.2.2 Seismic property estimation of sub-resolution heterogeneous systems
Estimating reservoir properties from seismic response can be challenging when
thicknesses of sedimentary beds are below seismic resolution. Examples of previous
studies on seismic property estimation of thin layered reservoirs are as follows. Vernik
et al. (2002) estimated seismic-scale sand volume of deepwater turbidite deposits by
cross-plotting P- and S-impedance inverted from seismic data. Dvorkin (2005)
introduced cumulative attributes (CATTS), which can be computed by integrating the
seismic trace repeatedly, for estimating total pore volume of the sub-resolution
reservoir. These studies provide useful attributes for estimating properties of sub-
resolution reservoirs, based on rock physics transforms (i.e., from elastic to reservoir
properties). However, because these transforms are applied to seismic-scale attributes,
for example inverted P-impedance, which is the upscale value of the fine-scale
heterogeneity, these attributes cannot be used to infer the fine-scale patterns of the
reservoir. Stright (2011) proposed a methodology for predicting sub-seismic
lithofacies using P-impedance and Vp/Vs. The strategy behind this method is to
generate a seismic-scale attribute template with lithofacies information by
incorporating fine-scale heterogeneity from well log and core data. This template can
then be applied to the seismic volume to infer sub-seismic lithofacies proportions.
Because this method transforms the attributes into lithofacies on a point-by-point basis,
the predicted lithofacies of all adjacent points along the vertical direction may not
reflect a spatial pattern of the actual lithofacies arrangement. Xie et al. (2004)
predicted seismic facies of thin reservoirs by comparing seismic traces away from the
well with those traces at the wells whose facies are known. The correlation coefficient
is used as a similarity measure among the seismic traces. This method is based solely
on pattern recognition of seismic traces without accounting for the facts that different
sets of rock properties and geometry can result in similar seismic signatures and that
similar reservoir properties but different geometry can result in different seismic
signatures (Takahashi, 2000; Spikes, 2008). This dissertation aims to investigate both
CHAPTER 1: Introduction 5
the effects of rock properties (e.g., sand volume, saturation) and the stacking patterns
of thinly bedded sand-shale reservoirs on seismic signatures and ultimately to estimate
these properties from seismic data.
1.3 Definitions of terms, symbols, and abbreviations
Important terms, symbols, and definitions which will be used throughout this
dissertation are defined in Table 1.1. Several volumetric parameters such as volume
fraction of laminated shale and dispersed clay are illustrated by the clay/shale
distribution diagram in Figure 1.1.
Table 1.1: Definitions of terms, symbols, and abbreviations used throughout this dissertation. Note that Vi is the volume of the ith layer illustrated in Figure 1.1, and i = 1, 2,…, 6.
Term Symbol or
abbreviation
Definition
Effective porosity 𝜙𝐸𝑓𝑓 volume fraction of effective pore space which excludes
volume of clay-bound water
Effective
saturation 𝑆𝑊𝑒 volume fraction of water in the effective pore space
Net-to-gross
(reservoir) ratio
NTG volume fraction of reservoir sand layers
𝑁𝑇𝐺 =𝑉2 + 𝑉4 𝑉𝑖6𝑖=1
Sand fraction 𝑉𝑠𝑎𝑛𝑑 volume fraction of all sandy layers including shaly sand
𝑉𝑠𝑎𝑛𝑑 =𝑉2 + 𝑉4 + 𝑉6 𝑉𝑖6𝑖=1
Shale volume 𝑉𝑠ℎ volume fraction of shale including dispersed clay and
laminated shale
𝑉𝑠ℎ =𝑉1 + 𝑉3 + 𝑉5 + 𝑉4 ∗ 𝑉𝑑𝑖𝑠𝑝 1 + 𝑉6 ∗ 𝑉𝑑𝑖𝑠𝑝 2
𝑉𝑖6𝑖=1
Total porosity 𝜙𝑇 volume fraction of void space in the rock
Volume of
dispersed clay 𝑉𝑑𝑖𝑠𝑝 volume fraction of clay dispersed in the sand pore space
𝑉𝑑𝑖𝑠𝑝 =𝑉4 ∗ 𝑉𝑑𝑖𝑠𝑝 1 + 𝑉6 ∗ 𝑉𝑑𝑖𝑠𝑝 2
𝑉𝑖6𝑖=1
Volume of
laminated shale 𝑉𝑙𝑎𝑚 volume fraction of total shale layers
𝑉𝑙𝑎𝑚 = 1 − 𝑉𝑠𝑎𝑛𝑑 =𝑉1 + 𝑉3 + 𝑉5 𝑉𝑖6𝑖=1
CHAPTER 1: Introduction 6
Figure 1.1: Different types of clay/shale distribution in sand. The volume of the ith layer and the volume of dispersed clay are denoted by Vi and Vdisp, respectively.
In rock physics, a laminated sand-shale sequence generally refers to a sequence of
alternating sand and shale units. However, in sedimentology this same alternating
sequence is generally described as interbedded sandstone and shale. This dissertation
uses the terms laminated and interbedded interchangeably.
1.4 Chapter description
The remainder of this dissertation is organized into four chapters, each of which
discusses about either existing or new tools for interpreting sub-resolution sand-shale
reservoirs.
Chapter 2 presents the sensitivity and uncertainty analyses of the Thomas-Stieber
model (Thomas and Stieber, 1975). This model is commonly used in thinly-bedded
sand-shale reservoirs to estimate volumetric fractions of various types of clay/shale
distribution within the sand-shale mixtures. We investigate various scenarios when
certain model assumptions are not met and show their effects on the estimated
properties from the model. We also show how uncertainty in the input parameters can
be incorporated into the model.
Chapter 3 proposes a simple graphical mesh interpretation and accompanying
equations for fluid substitution in sub-resolution interbedded sand-shale sequences.
The sand layers can be either clean or shaly (i.e., sand with dispersed clay). Our mesh
method can be applied to measurements directly at their original scale, without the
need to downscale the measurements, while still changing fluid in the sand layers only.
Laminated shale
Clean sand
Sand with dispersed clay ( V disp1 )
Sand with pore space completely filled with clay ( V disp2 )
CHAPTER 1: Introduction 7
Chapter 4 investigates seismic signatures of thin sand-shale reservoirs using
wavelet-transform based attributes and feature-extraction based attributes. Note that
feature extraction techniques transform input data into new representations in lower-
dimensional space (Cunningham, 2008). We use the standard statistical rock-physics
approach to explore what-if scenarios of how seismic signatures of the thin reservoirs
change with varying reservoir properties, such as net-to-gross ratio, saturation, and
stacking pattern. We present synthetic examples of how to apply these seismic
attributes to estimate reservoir properties in real-data applications.
Chapter 5 applies the seismic attributes introduced in Chapter 4 to real well log
and seismic data from deep-water deposits in West Africa.
1.5 References
Boggs, S., 2001, Principles of sedimentology and Stratigraphy, 3rd
ed.
Boyd, A., Darling, H., Tabanou, J., Davis, B., Lyon, B., Flaum, C., Klein, J., Sneider,
R.M., Sibbit, A., and Singer, J., 1995, The lowdown on low-resistivity pay:
Schlumberger Oilfield Review, 7(3), 4-18.
Campbell, C.V., 1967, Lamina, laminaset, bed and bedset: Sedimentology, 8(1), 7-26.
Chopra, S., 2005, Expert Answers: Gassmann’s equation: CSEG Recorder in May, 8-
12.
Cunningham, P., 2008, Dimension reduction: Machine learning techniques for
multimedia: case studies on organization and retrieval, Springer, Eds. M. Cord and
P. Cunningham.
Dvorkin, J., 2005, Cumulative seismic attribute for ϕh determination: SEG Expanded
Abstracts, 24, 1550-1553.
Dvorkin, J. and Uden, R., 2006, The challenge of scale in seismic mapping of hydrate
and solutions: The leading edge, 25(5), 637-642.
Gassmann, F., 1951, Uber die elastizitat poroser medien: Vier Natur Gesellschaft, 96,
1-23.
Katahara, K., 2004, Fluid Substitution in Laminated Shaly Sands: SEG Expanded
Abstracts, 23, 1718-1721.
Passey, Q. R., Dahlberg, K. E., Sullivan, K. B., Yin, H., Brackett, R. A., Xiao, Y. H.,
and Guzmán-Garcia, A. G., 2006, Petrophysical evaluation of hydrocarbon pore-
CHAPTER 1: Introduction 8
thickness in thinly bedded clastic reservoirs: AAPG Archie series, 1, American
Association of Petroleum Geologists, Tulsa, Oklahoma, U.S.A.
Skelt, C., 2004a, Fluid substitution in laminated sands: The Leading Edge, 23, 485–
488.
Skelt, C., 2004b, The influence of shale distribution on the sensitivity of
compressional slowness to reservoir fluid changes: SPWLA 45th
Annual Logging
Symposium, June 6-9.
Spikes, K.T., 2008, Probabilistic seismic inversion based on rock-physics models for
reservoir characterization, Ph.D. Thesis, Stanford University.
Stright, L.E, 2011, Multiscale modeling of deep-water channel deposits: An
interdisciplinary study integrating geostatistics, geology and geophysics The
cretaceous Cerro Toro formation, southern Chile, the Eocene Ardath & Scripps
formations, southern California, and the Oligocene Puchkirchen formation, upper
Austria, Ph.D. Thesis, Stanford University.
Takahashi, I., 2000, Quantifying Information and Uncertainty of Rock Property
Estimation from Seismic Data, Ph.D. Thesis, Stanford University.
Thomas, E. C. and Stieber, S. J., 1975, The distribution of shale in sandstones and its
effect upon porosity: 16th Annual Logging Symposium, SPWLA, Paper T.
Vernik, L., Fisher, D., and Bahret, S., 2002, Estimation of net-to-gross from P and S
impedance in deepwater turbidites: The leading edge, 21(4), 380-387.
Xie, D., Wood, J.R., and Pennington, W.D., 2004, Quantitative seismic facies analysis
for thin-bed reservoirs: a case study of the central Boonsville field, Fort Worth
Basin, north-central Texas: SEG Expanded Abstracts, 23, 1484-1487.
9
Chapter 2
Sensitivity and uncertainty analysis
of the Thomas-Stieber model for
property estimation of thin sand-
shale reservoirs
2.1 Abstract
This chapter presents sensitivity and uncertainty analyses of the Thomas-Stieber
model and its application for estimating petrophysical properties in thinly bedded
sand-shale reservoirs. This model describes how total porosity relates to different
types of clay/shale distribution within the sand-shale mixture. By inputting the clean
sand and shale properties, this model can be used deterministically to estimate both the
volume fraction of the sandy layers within the lamination and the volume fraction of
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 10
dispersed clay within the sandy layers. Even though applying this model to a dataset is
straightforward, several factors including simplistic assumptions and uncertainties of
estimated input parameters can affect the data interpretation.
The objective of this study is to incorporate natural variations and uncertainties
which can be found in real observations into the model and to point out some pitfalls
in the interpretation when these variations are neglected. For a case of bimodal mixing
(i.e., sand and pure shale), this model can be used to infer volume fraction of sand and
its porosity. However, when the sand beds within the interval of interest have a range
of properties (e.g., varying clay content), the model cannot deterministically estimate
property of each individual sand layer. Instead, it gives an average property of all the
sand beds within the stack. The uncertainties of the input parameters of the model can
be used in stochastic methods to quantify uncertainty of our interpretation. We use
Monte Carlo simulations propagated through the Thomas-Stieber model to estimate
the posterior distributions of interpreted volume fractions of sand in a Bayesian
framework. We apply this stochastic approach to well log data from deep-water
turbidite deposits in West Africa. We also present the analogs of the Thomas-Stieber
model on other cross-plots.
2.2 Introduction
Geophysical measurements at various scales are used to infer the properties and
structure of the subsurface. These data often represent average properties of multiple
thin layers when the heterogeneities are below the resolution of the measuring tools.
The lack of knowledge about the existence and properties of thin lamination could
mislead the interpretation and thus affect the evaluation of the thin reservoirs. For
example, a direct application of isotropic Gassmann’s equation (Gassmann, 1951)
without accounting for the sub-resolution, interbedded sand-shale layers would be
erroneous because fluid substitution is likely to be applicable only in permeable sand
beds (e.g., Katahara, 2004; Skelt, 2004a; Skelt, 2004b; Chopra, 2005). Thus, a more
accurate description of the thin reservoirs is needed for better reservoir
characterization.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 11
The Thomas-Stieber model (Thomas and Stieber, 1975) is a useful tool for
evaluating thinly bedded sand-shale reservoirs because it provides not only the
evidence of fine-scale lamination but also the petrophysical properties of rocks within
the lamination (e.g., Dutta, 2009; Yadav et al., 2010; Voleti et al., 2012). This model
describes the relations between total porosity and shale volume of the sand-shale
mixtures. These relations are not unique, but dependent on how clay/shale is
distributed in the mixtures. Since any rocks that can be described using the Thomas-
Stieber model are assumed to be mixtures between the two end-member rocks (i.e.,
sand and shale), by inputting the properties of the sand and the shale end-points the
model can be constructed and used to infer volume fractions of different types of
clay/shale distributions. These inferred shale volumes are crucial for improving
reservoir evaluation. Even though the Thomas-Stieber model can be constructed and
applied easily, several assumptions of this model can be too simplistic for real data.
For example, by using this model the interpretation of any data point is limited to
lamination between sandy layers and shale layers, and all the sandy layers share the
same petrophysical properties. This interpretation may not always be true. Therefore,
it is important to be aware of the limitations of the model and to explore how data
interpretation is affected under more realistic scenarios or under uncertainty in the
model’s input parameters.
In this chapter, we generate synthetic rocks to test the model under various
scenarios which do not fit the model assumptions. Then, we incorporate uncertainties
in input parameters into the interpretation using Monte Carlo simulations. The
remainder of this chapter is organized as follows. Section 2.3 briefly reviews the
Thomas-Stieber volumetric model for a sand-shale mixture system, including
dispersed and laminar mixes, and discusses important assumptions of the model. In
Section 2.4, we perform sensitivity and uncertainty analyses of the model by
incorporating natural variations into the model in order to quantify uncertainty in our
interpretation. We present results of the analyses through both synthetic examples and
real well log data. Section 2.5 shows analogs of the Thomas-Stieber model on other
cross-plots, and finally Section 2.6 discusses other sources of uncertainties that can be
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 12
found in real data application and provides possible constraints on choices of input
parameters.
2.3 Volumetric properties of sand-shale mixtures
Here a volumetric property refers to a petrophysical property whose representative
value for a rock sample depends on the volume fractions and the individual properties
of the rock’s constituents, but not how they are arranged or distributed. The volumetric
properties covered in this section are total porosity, effective porosity, and shale
volume.
2.3.1 Total porosity and volume of shale
Thomas and Stieber1 (1975) provided a simple model for how total porosity of a
sand-shale mixture varies with shale volume, depending on the distribution of
clay/shale in the mixture. There are three main types of distribution: dispersed clay,
laminated shale, and structural clay (Figure 2.2). Here we focus on the dispersed and
laminar types. As noted in Chapter 1, in rock physics, a laminated sand-shale sequence
generally refers to a sequence of alternating sand and shale units. However, in
sedimentology this same alternating sequence is generally described as interbedded
sandstone and shale. We use the terms laminated and interbedded interchangeably.
Figure 2.1: Shale distribution in sand (Modified after Mavko et al., 2009).
1 Thomas-Stieber model was originally derived using gamma ray response.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 13
Figure 2.2: Relation between total porosity and shale volume. Point A and E represent clean sand and pure shale point, respectively. Points B and C represent sand with dispersed clay, and point D is where the original sand pore space is completely filled with dispersed clay.
The variation of total porosity with shale volume in a sand-shale mixture is
schematically shown in Figure 2.2. The dispersed lines AD and DE describe the
topology of bimodal mixtures of sand grains and clay, and a V-shaped relation
between the shale volume and the mixture porosity (Marion, 1990). When clay is
dispersed into the original sand pore space without disturbing the sand pack, total
porosity linearly decreases with volume of shale (the shaly-sand line AD) as
(0 ≤ 𝑉𝑠ℎ ≤ 𝜙𝑠): 𝜙𝑇 = 𝜙𝑠 − 1 − 𝜙𝑠ℎ 𝑉𝑠ℎ .
(2.1)
In equation 2.1, 𝜙𝑇, 𝜙𝑠, and 𝜙𝑠ℎ are the total, sand, and shale porosity respectively,
and 𝑉𝑠ℎ is the shale volume.
At this stage, sand grains provide the load-bearing matrix of the mixture. When
clay is continuously added until its volume is equal to sand porosity (D), the original
sand pore space is completely filled with clay:
𝑉𝑠ℎ = 𝜙𝑠 : 𝜙𝑇 = 𝜙𝑠𝜙𝑠ℎ .
(2.2)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Shale volume
Tota
l poro
sity
Dispersed line
AV
lam = 0.2
Vlam
= 0.4
Vlam
= 0.6
Vlam
= 0.8
EF
B
C
D
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 14
When shale volume is greater than sand porosity, sand grains are displaced and
disconnected; as a result, the mixture changes from grain-supported to clay-supported
sediments. At this stage, porosity increases linearly with increasing shale volume (the
sandy-shale line DE) because the solid sand grains are replaced by porous clay (Yin,
1992; Avseth et al., 2005):
𝜙𝑠 < 𝑉𝑠ℎ ≤ 1 : 𝜙𝑇 = 𝜙𝑠ℎ𝑉𝑠ℎ .
(2.3)
We refer to the volume of dispersed clay as 𝑉𝑑𝑖𝑠𝑝 .
In the laminar distribution, the total porosity is simply a weighted average of the
porosities of all the end-members. For example, in the case of clean sand at point A
laminated (interbedded) with pure shale at point E, the total porosity of the mixture
can be expressed as
𝜙𝑇 = 1 − 𝑉𝑙𝑎𝑚 𝜙𝑠 + 𝑉𝑙𝑎𝑚 𝜙𝑠ℎ ,
𝑉𝑙𝑎𝑚 = 1 − 𝑉𝑠𝑎𝑛𝑑 ,
(2.4)
where 𝑉𝑙𝑎𝑚 is the volume of laminated shale and 𝑉𝑠𝑎𝑛𝑑 is the volume fraction of
laminated sand which can be clean or shaly. We refer to this volume as the sand
fraction.
The other line segments (BE, CE and DE) correspond to lamination between shaly
sands (i.e., sand with dispersed clay) and pure shale. Note that these line segments
radiate from the shale point (E), because the Thomas-Stieber model assumes that one
of end-members in a lamination is pure shale. Different volumes of laminated shale
within the mixture are marked by the iso-laminar lines, represented by the set of blue
lines in Figure 2.2 (Thomas and Stieber, 1975).
In order to construct the diagram in Figure 2.2, we need a minimum of two input
parameters: porosities of the clean sand end-point (at 𝑉𝑠ℎ = 0) and the pure shale end-
point (at 𝑉𝑠ℎ = 1). If volume of shale is derived from gamma ray logs, we need two
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 15
additional parameters: the gamma ray values for clean sand and shale. Assuming that
there are only two types of rocks alternating within the lamination—clean sand (or
shaly sand) and pure shale—the diagram, which combines both dispersed and
laminated models, can be used graphically or algebraically to determine the existence
of thin lamination even when measurements cannot resolve each layer, and to estimate
the amount and properties of sand in the lamination. For example in Figure 2.2, point
F (𝑉𝑠ℎ = 0.52 and total porosity = 0.138) represents lamination between shaly sand (C)
and shale (E), with 𝑉𝑙𝑎𝑚 = 0.4. The shaly sand has a total porosity equal to 0.138, and
thus this shaly sand has 𝑉𝑑𝑖𝑠𝑝 = 0.2 (Equations 2.1 and 2.4). Alternatively, both
𝑉𝑙𝑎𝑚 and 𝑉𝑑𝑖𝑠𝑝 of point F can be obtained graphically (Figure 2.3). Since pure shale is
assumed to be one of the end-members in the lamination, a line is drawn by always
starting from the shale point (E) and then passing through point F until the line
intersects the dispersed line (AD). This intersection point (C) represents the other end-
member of the lamination. Being on the dispersed line, point C corresponds to shaly
sand whose 𝑉𝑑𝑖𝑠𝑝 is simply obtained by reading off the x-coordinate of the point. 𝑉𝑙𝑎𝑚
is then equivalent to the ratio between the lengths of line CF and line CE.
Figure 2.3: Graphical determination of the laminated shale and the dispersed clay volumes (𝑉𝑙𝑎𝑚 and 𝑉𝑑𝑖𝑠𝑝 , respectively) for a measurement at point F. In each case, the shale volume is determined by drawing a line from the shale point (E), passing through point F, and intersecting line AD. The intersection point (C) represents shaly sand that is the sandy end-member of the lamination, and this shaly sand point has 𝑉𝑑𝑖𝑠𝑝 equal to its x-coordinate (𝑉𝑑𝑖𝑠𝑝 = 0.2). Then, 𝑉𝑙𝑎𝑚 is simply the ratio between the lengths of line CF and CE, which is equal to 0.4.
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Shale volume
Tota
l poro
sity
Dispersed line
AV
lam = 0.2
Vlam
= 0.4
Vlam
= 0.6
Vlam
= 0.8
EF
B
C
D
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 16
2.3.1.1 Computing effective porosity by distorting grids
From total porosity (Equations 2.1 and 2.3), we can easily write an expression for
effective porosity:
0 ≤ 𝑉𝑠ℎ ≤ 1 𝜙𝐸𝑓𝑓 = 𝜙𝑇 − 𝜙𝑠ℎ𝑉𝑠ℎ ,
(2.5)
where 𝜙𝑠 < 𝑉𝑠ℎ ≤ 1. Note that Equation 2.5 is applicable under the assumption that
both dispersed clay and laminated shale have the same mineralogy and properties (e.g.,
porosity).
According to Equation 2.5, the transformation from total porosity to effective
porosity is equivalent to moving the pure shale corner and the sandy-shale line DE in
Figure 2.2 down to zero; any data point inside the triangular diagram is distorted
proportionally (Figure 2.4, Mavko et al., 2009).
Figure 2.4: Distorting total porosity diagram (blue) for computing effective porosity (red).
2.3.2 Important assumptions of the Thomas-Stieber model
The Thomas-Stieber model assumes that there are only two types of rocks
alternating within the lamination. Consequently, the interpretation of any data point in
the grid in Figure 2.2 is limited to mixing between sand and shale or between shaly
sand (sand with dispersed clay) and shale. However, laminations consisting of only
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Shale volume
Tota
l poro
sity
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 17
two alternating elements may not always be present. For example, both the sand and
shale properties within the interval of interest can vary due to variations in amount of
clay present. Dawson et al. (2008) characterized shale samples from deep-water
depositional settings and reported different types of shale due to differences in
compositions, fabric and sealing capacity. Since this assumption of two rocks
alternating in the lamination is only valid for very simple rocks, in the next section we
will study synthetic examples to understand the limitations in interpretation of more
complex rocks when using this model.
The model assumes that any porosity reduction is caused only by clay filling the
pore space, not by cementation or grain sorting (Thomas and Stieber, 1975; Ball et al.,
2004; Mavko et al., 2009). For the cross-plot of total porosity versus shale volume,
adding cement to the remaining sand pore space reduces total porosity but does not
affect shale volume. For example if cement (25% of the sand porosity) is added into
the sands on the dispersed sand line, the resulting total porosity and shale volume can
be plotted as in Figure 2.5. The blue line is the line of 25%-cement volume, which is
expressed as a percentage of the porosity of the sand end-point.
Figure 2.5: Total porosity and shale volume after cement is added into the remaining sand pore space of the sands on the dispersed sand line. The blue line is the resulting dispersed sand line after adding cement with a volume of 25% of the porosity of the sand end-point into the pore space (Modified after Juhasz, 1986).
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Shale volume
Tota
l poro
sity Shale
Sand
cement volume added = 25% of the
sand porosity
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 18
The dispersed distribution in the Thomas-Stieber model represents ideal mixing
and assumes that one component does not disturb the packing of the other. This
assumption holds under a necessary but not sufficient condition that there is a large
difference between the mean sizes of two components mixed together (Cumberland
and Crawford, 1987; Koltermann and Gorelick, 1995; Dvorkin and Gutierrez, 2002;
Revil et al., 2002). Even though the size of clay particles is much smaller than the
mean size of sand grains, other factors that affect real sedimentary rocks (e.g., range of
grain size, grain shape, geological process) can easily lead to deviation from the ideal
packing (Koltermann and Gorelick, 1995; Boggs, 2001; Revil et al., 2002). Similar to
the Thomas-Stieber model, a fractional packing model developed by Koltermann and
Gorelick (1995) describes mixtures, where clay and sand are dispersedly mixed, in
terms of relations between total porosity and shale volume fraction. However, this
fractional packing model relaxes the assumption that one component does not disturb
the other by allowing the shaly-sand domain and the sandy-shale domain to coexist in
the dispersed mixing. This model is shown to fit laboratory measurements of sand-clay
mixtures better than the ideal mixing model (Koltermann and Gorelick, 1995). In the
next section using synthetic examples, we will explore how interpretation is affected
by applying the ideal mixing model to the non-ideal mixtures.
The Thomas-Stieber model assumes that both the dispersed clay in the sand and
the laminated clay/shale have the same properties. This assumption is not always true
because the compositions and properties of clay/shale can vary depending on the
origin of the clay/shale within the sand-shale mixture. For example, while detrital clay
is transported and deposited at the deposition site, authigenic clay is diagenetically
formed after deposition (Thomas and Stieber, 1975; Asquith, 1990; Worden and
Morad, 2003; Worthington, 2003).
Even assuming that all of the model assumptions hold, in real applications of this
model, the input parameters (i.e., properties of the end-points used to construct the
diagram) have to be estimated. Uncertainties in this parameter estimation can affect
the interpretation of data. To understand the effects, we perform sensitivity and
uncertainty analysis of the Thomas-Stieber model in the next section.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 19
2.4 Sensitivity and uncertainty analysis
In the previous section, we have introduced the Thomas-Stieber model and related
cross-plots, and we showed that once the cross-plots are constructed, they can be used
to deterministically infer the volume fraction and property of sand within thin sand-
shale lamination. In this section, we perform sensitivity and uncertainty analyses on
the model, particularly the cross-plot between total porosity and volume fraction of
shale (or gamma ray), and study how uncertainties in various aspects of this model
affect the estimated sand fractions and sand properties. Two sources of uncertainties
are considered: uncertainties from violations of the model assumptions and
uncertainties from the model parameters. In section 2.4.1, we investigate three
scenarios, where the model parameters are correct but the earth models do not follow
the model assumptions, and analyze the effects on data interpretation. In section 2.4.2,
we investigate how uncertainties in the model parameters propagate through the model.
2.4.1 Model with correct input parameters
2.4.1.1 Mixing various sands with shale
In Figure 2.6 (right), the earth model is composed of shale layers interbedded with
sand types A, B, and C, each of which has a different amount of dispersed clay in its
pore space and thus a different total porosity. Assuming that the layers in this earth
model are too thin to be resolved by logging tools, the measured properties of the earth
model therefore represent average properties of the layers. The average total porosity
and volume fraction of shale of the model is shown in Figure 2.6 (left) and marked
with an ―X,‖ with the properties of the individual sand layers being shown in circles.
By constructing the Thomas-Stieber diagram with the correct clean sand and pure
shale end-points, the sand fraction and property of the sand can be estimated. Figure
2.7 shows a comparison of the estimated sand fraction and the estimated sand property
with the true values. While the sand fraction is estimated correctly, the volume
fraction of dispersed clay in sand (i.e., the sand property) is either over- or
underestimated, compared to the actual properties of sand A, B, or C. The estimated
sand property is only the volumetric average of all the sand layers. Because the
Thomas-Stieber model assumes that only one sand type is laminated (interbedded)
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 20
with shale, given only the average (i.e., upscaled) properties at point ―X,‖ it is
impossible to estimate (i.e., downscale) properties of individual sand types, or even to
know how many sand types are present within the interval.
Figure 2.6: Thomas-Stieber diagram (left) with plotted properties from the corresponding earth model which is represented by an interbedded sand-shale sequence (right). The properties of the individual sand types (A, B, and C) and shale are shown in circles. The average total porosity and shale volume fraction of the earth model is marked with an ―X.‖ The volume fractions of laminated shale and dispersed clay can be determined graphically by first drawing a line that originates from the shale point, passes through point X, and intersects the dispersed line. The volume fraction of dispersed clay is simply equal to the x-coordinate of the intersection point, and the volume fraction of laminated shale (𝑉𝑙𝑎𝑚 = 1 – 𝑉𝑠𝑎𝑛𝑑 ) is the ratio between the distance from point X to the intersection point and the distance from the shale point to the intersection point.
Figure 2.7: Estimated sand fraction and sand property when applying Thomas-Stieber diagram to point ―X‖ in Figure 2.6 (left). The estimated values are compared with the true values from the earth model (Figure 2.6, right).
2.4.1.2 Mixing shaly sand and sandy shale
The earth model composed of shaly sand interbedded with sandy shale is shown in
Figure 2.8, right. The measured properties of this earth model, which are the average
True V sand True V disp in sand
Estimated Estimated
V disp
in sand
A B C
E
stim
ate
d V
dis
p in
sa
nd
Total
porosity
Vshale
B
C
B
A
C
A
A
estimated volume of laminated shale
estimated volume of dispersed clay
X
To
tal p
oro
sity
E
stim
ate
d V
sa
nd
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 21
total porosity and volume fraction of shale, are shown in Figure 2.8 (left) and marked
with an ―X,‖ with the actual shaly sand and sandy shale properties being shown in
circles. By constructing Thomas-Stieber diagram with the correct clean sand and pure
shale end-points, the sand fraction and sand property can be estimated. Figure 2.9
shows a comparison of the estimated sand fraction and the estimated sand property
with the true values. Even though both end-points used for the Thomas-Stieber
diagram are correct, the sand fraction and sand property are not estimated correctly
because the diagram always assumes that the shale end-point (or pure shale) is one of
two end-members in the sand-shale lamination, whereas in this example the
lamination does not contain pure shale, but the sandy shale. Figure 2.9 shows the
possible errors which can result when one of the end-members is chosen incorrectly.
Figure 2.8: Thomas-Stieber diagram (left) with plotted properties from the corresponding earth model which is represented by an interbedded (shaly) sand-(sandy) shale sequence (right). The properties of sand and shale layers within the earth model are shown in circles. The average total porosity and shale volume fraction of the earth model is marked with an ―X.‖ The volume fractions of laminated shale and dispersed clay can be determined graphically by first drawing a line that originates from the shale point, passes through point X, and intersects the dispersed line. The volume fraction of dispersed clay is simply equal to the x-coordinate of the intersection point, and the volume fraction of laminated shale (𝑉𝑙𝑎𝑚 = 1 – 𝑉𝑠𝑎𝑛𝑑 ) is the ratio between the distance from point X to the intersection point and the distance from the shale point to the intersection point.
Total
X porosity
Vshale
estimated volume of laminated shale
estimated volume of dispersed clay
To
tal p
oro
sity
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 22
Figure 2.9: Estimated sand fraction and sand property when applying Thomas-Stieber diagram to point ―X‖ in Figure 2.8 (left). The estimated values are compared with the true values from the earth model (Figure 2.8, right).
2.4.1.3 Deviation from ideal mixing
Assuming that the shaly sand layers in the earth model represent non-ideal
mixtures, we generate total porosities of the shaly sand layers using the fractional
packing model of Koltermann and Gorelick (1995). For shale volume fractions
ranging from 0 to 1, the total porosities predicted by the fractional packing model are
represented as points on the blue curve in Figure 2.10. The model’s input parameters
used in this example are 𝜙𝑠= 0.34, 𝜙𝑠ℎ= 0.3, and 𝜙𝑚𝑖𝑛 =0.18. These parameters are
estimated by Kolterman and Gorelick (1995) to fit laboratory measurements for sand-
clay mixtures with diameter ratios greater than 100 and under 30 MPa confining
pressure. Note that 𝜙𝑚𝑖𝑛 is the minimum porosity observed when the shale volume
fraction is equal to the sand porosity. At this point, the Thomas-Stieber model assumes
an ideal mixture where the sand pore space is completely filled with clay and the
model predict that 𝜙𝑚𝑖𝑛 is equal to 𝜙𝑠𝜙𝑠ℎ (Equation 2.2). In contrast, the fractional
packing model assumes a non-ideal mixture where clay disturbs the packing of the
sand, prohibiting the complete filling of the sand pore space by clay. Therefore, in the
fractional packing model 𝜙𝑚𝑖𝑛 is greater than the corresponding value for the ideal
packing model, 𝜙𝑠𝜙𝑠ℎ (Koltermann and Gorelick, 1995).
Using the fractional packing model, we generate a set of data points (solid circles
in Figure 2.10) by varying sand fractions and volume fraction of dispersed clay to
create various sand-shale sequences. Then we estimate sand fractions and sand
properties of these points using the Thomas-Stieber model (the red curve in Figure
True V disp in sand Estimated
V disp
in sand
True V sand Estimated
sand fraction
E
stim
ate
d V
sa
nd
E
stim
ate
d V
dis
p in
sa
nd
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 23
2.10) to show how our data interpretation is affected when applying the ideal mixing
model to the non-ideal mixtures. Results are shown in Figure 2.11 as percentage
differences between the estimated and true value using the relation
% 𝑑𝑖𝑓𝑓 = 100 ∗𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑒 − 𝑡𝑟𝑢𝑒
𝑡𝑟𝑢𝑒 .
(2.6)
In this analysis of the ideal-mixing assumption, both the sand fractions and the
volume fractions of dispersed clay are underestimated (i.e., negative percentage
differences). Consequently, the sand in the reservoir is interpreted to have less volume
fraction but be cleaner than the actual properties. The largest underestimations for the
sand fraction and the volume fraction of dispersed clay are approximately 12% and
26%, respectively. Therefore, the estimated volume fractions of dispersed clay are
affected more severely from using the ideal mixing model than the estimated sand
fractions are. Applying the ideal mixing model to the non-ideal mixtures also affects
the interpretation of sand-shale sequences at different locations on the triangular
diagram differently. Fortunately, the better reservoir quality sequences (i.e., those
located near the upper left corner of the diagram, where the sand fractions within the
sequences are greater and the volume fractions of dispersed clay in the sand are
smaller) are least affected by the assumption of ideal mixing. The overall magnitude
of the effect depends on the value of 𝜙𝑚𝑖𝑛 . In general, if 𝜙𝑚𝑖𝑛 increases, the mixtures
deviate more and more from the ideal mixing, and thus the errors in the interpretation
resulting from using the ideal mixing model on the non-ideal mixtures will get larger.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 24
Figure 2.10: Relations between total porosity and shale volume fraction of bimodal mixtures using the fractional packing model by Kolterman and Gorelick (1995) and the Thomas-Stieber model (i.e., ideal mixing model). Data points representing sand/shaly-sand interbedded with shale (solid circles) are generated using the fractional packing model with varying sand fractions and volume fractions of dispersed clay.
Figure 2.11: Percentage differences between the estimated sand fractions (left), the estimated volume fractions of dispersed clay in the sand (right) and the true values. The ideal mixing model and the fractional packing model (i.e., non-ideal mixing) are outlined.
2.4.2 Model with uncertain input parameters
We have previously shown that even when the correct sand and shale end-points
are used in the Thomas-Stieber model, other complications (e.g., multiple sand types)
0 0.2 0.4 0.6 0.8 1
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Shale volume
Tota
l poro
sity
Fractional packing model
Ideal mixing model
Shale
Sand
0.2 0.4 0.6 0.8
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Shale volume
Tota
l poro
sity
0.2 0.4 0.6 0.8
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Shale volume
Tota
l poro
sity
-10
-8
-6
-4
-2
-25
-20
-15
-10
-5
Fractional
packing
model
Ideal mixing
model
Color-coded by % difference between
the estimated
and the true sand fractions
Color-coded by % difference
between the estimated and the
true volumes of dispersed clay
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 25
can lead to errors in the estimation of sand fraction and sand properties. Moreover, in
practice it is almost impossible to know what the correct end-points are, without
detailed core analyses. Therefore, there are always uncertainties associated with the
model’s input parameters (i.e., the properties of the sand and the shale end-points). In
this section, we perform sensitivity analyses of how uncertainties in the input
parameters affect the estimated sand fractions and sand properties using Monte Carlo
simulations. Note that there is also uncertainty in the data. We will not present this
issue here; however, the same analysis shown in this section can be extended to
incorporate data uncertainties into the model.
Using the decomposition of conditional probability, we have the following
relations:
𝑃 𝐴 𝐷 = 𝑃 𝐴,𝐸 𝐷 𝑑𝐸 = 𝑃 𝐴 𝐸,𝐷 𝑃 𝐸 𝐷 𝑑𝐸 ,
𝑃 𝐴 𝐷 = 1
𝑃(𝐷) 𝑃 𝐴 𝐸,𝐷 𝑃 𝐷 𝐸 𝑃 𝐸 𝑑𝐸 ,
(2.7)
where A is the property of interest (e.g., sand fraction), D is a data point, and E
represents the two input properties (e.g., clean sand and pure shale total porosities).
P(D) is a normalization constant.
For simplicity, P(E) is assumed to be known. For a set of acquired data points,
information such as the location, geological setting, and depth of the collection site
can be used to constrain and estimate the distributions of the properties of the
endpoints. In our examples, the gamma ray and total porosity values of both the clean
sand and pure shale points are assumed to be normally distributed, and all parameters
are assumed to be independent. Note that the distributions of the properties of the
endpoints need not be normal. Also note that we compute shale volume fraction (to be
used in Equations 2.1 – 2.4) from gamma ray values by using a linear transformation.
We also assume that for a pair of endpoints any point inside the diagram that is
constructed using these endpoints is equally likely to occur. Thus the probability
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 26
density function is equal to a constant value (i.e., the inverse of the area of the
triangular diagram) that is defined over the entire the triangle. The probability that a
small area surrounding a data point will be chosen or P(a small area surrounding D|E)
is equal to that small area divided by the area of the triangle.
In our example, both the gamma ray and porosity distributions of the shale
endpoint have larger variances than those of the sand endpoint. To run a Monte Carlo
simulation, we first separately draw the properties of the two endpoints, each of which
is a pair of gamma ray and total porosity values, from the selected normal distributions
and input the two end-points into the Thomas-Stieber model. Then we use the model
to infer the values of sand fractions and volume fractions of dispersed clay in the sand
for three selected data points. By drawing multiple realizations of the endpoints
(Figure 2.12), we finally obtain both the joint posterior distributions between sand
fraction and volume fraction of dispersed clay (Figure 2.13) and the marginal
distributions of each variable (Figure 2.14) for the three data points.
Figure 2.12: Multiple realizations of Thomas-Stieber diagrams generated from a set of sand and shale end-points. Three data points are labeled.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 27
Figure 2.13: Posterior distributions for estimated sand fractions of the three data points shown in Figure 2.12.
Figure 2.14: Posterior distributions of volume fraction of dispersed clay in sand (left) and sand fraction (right) for the three data points shown in Figure 2.12.
All three data points have quite similar ranges of uncertainty in the estimated
volume fraction of dispersed clay, with the range of point 1 being slightly smaller
point 2 and point 3. Furthermore, point 1, which has the highest sand fraction, has the
smallest uncertainty in the estimated sand fraction partially due to its location on the
triangular diagram; this point is closer to the sand end-point that is assumed to be less
uncertain than the other endpoint. Due to shale-distribution-dependent relations
between total porosity and gamma ray of the Thomas-Stieber diagram (i.e.,
geometrical interpretation), the variations in the input parameters can influence
estimated properties of points with high and low sand fractions differently.
Sand fraction
Volu
me o
f dis
pers
ed c
lay in s
and
0.4 0.5 0.6 0.7 0.8 0.9
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
(1)
(2)
(3)
0 0.1 0.2 0.3 0.40
5
10
15
20
25
30
Volume of dispersed clay in sand
(1)
(2)
(3)
0.4 0.6 0.8 10
5
10
15
20
Sand fraction
(1)
(2)
(3)
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 28
To calculate standard deviations of estimated properties at different locations on
the diagram, we use a simple error propagation technique:
𝜎𝑦2 =
𝜕𝑦
𝜕𝑥𝑖
2𝑛𝑖=1 ∙ 𝜎𝑥𝑖
2 ,
(2.8)
where y is a function of x1, x2, …, xn and 𝜎𝑘 is the standard deviation of the variable k.
Note that all parameters are assumed to be independent. Since this equation is based
on a linear approximation of the actual model, its result is reliable only for small
parameter uncertainties (Gujer, 2008). We take variable y to be either the estimated
sand fraction or the estimated volume fraction of dispersed clay, each of which is a
function of six parameters: gamma ray measurement, total porosity measurement,
gamma ray and total porosity values of the sand end-point, and gamma ray and total
porosity values of the shale end-point. We use the mean gamma ray and total porosity
values of the end-points from Figure 2.12 to construct the diagrams in Figure 2.15.
Then, we assign non-zero standard deviations to both the gamma ray and the total
porosity values of both endpoints, while assigning zero standard deviations to the
other parameters. Finally, by using Equation 2.8 we propagate uncertainties in the
properties of the endpoints through the Thomas-Stieber model.
Figure 2.15 shows how uncertainty in the estimated properties varies for different
locations on the Thomas-Stieber diagram. Laminated sequences with cleaner sand and
higher sand fractions, which are represented in the upper left corner of the diagram,
are least affected by the uncertainties of the endpoints. However, the resulting patterns
of uncertainty shown in Figure 2.15 are not universal, but they depend on the setup of
the analysis.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 29
Figure 2.15: Standard deviations of estimated properties at different locations on the Thomas-Stieber diagram. The two estimated values are (left) volume fraction of dispersed clay and (right) sand fraction. The three data points are the same points as in Figure 2.12.
2.4.3 Example with well log data
In this section, we show examples of estimating sand fractions and volume fraction
of dispersed clay with well log data from deep-water turbidite deposits, offshore
Equatorial Guinea, West Africa. An evidence of fine-scale lamination is provided by a
detailed core analysis of this study area (Lowe, 2004). Six lithofacies are identified,
but in this example we focus on only three lithofacies (Figure 2.16). The first
lithofacies is characterized by thick-bedded to massive sandstone, whose bed
thicknesses are greater than 20 cm. Both the second and the third lithofacies are
characterized by interbedded, thin-bedded sandstone and mudstone. In the second
lithofacies, the sandstone bed thicknesses range from 2 to 20 cm and the rock units in
this lithofacies have volume fractions of sandstone greater than 20%. In the third
lithofacies, the sandstone thicknesses are less than 2 cm (Lowe, 2004; Dutta, 2009).
For the selected interval, the gamma log shows an upward-fining trend (Figure
2.17, right). In Figure 2.17 (left), total porosity values are plotted against
corresponding gamma ray values. Suppose that some geological information provides
us the estimates of total porosity and gamma ray distributions for both the sand and the
shale end-points (Figure 2.18). Then Equation 2.7 can be applied to each data point in
Figure 2.17 (left) to estimate its corresponding posterior distributions of sand fraction
(Figure 2.19).
20 40 60 80
0.1
0.15
0.2
0.25
0.3
Gamma ray
Tota
l poro
sity
Color-coded by standard deviation
of estimated sand fraction
20 40 60 80
0.1
0.15
0.2
0.25
0.3
Gamma ray
Tota
l poro
sity
Color-coded by standard deviation
of volume of dispersed clay
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.02
0.03
0.04
0.05
0.06
0.07
0.08
>0.084
(1)
(2)
(3)
(1)
(2)
(3)
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 30
Figure 2.16: Three lithofacies from a detailed core analysis by Lowe (2004).
Lithofacies 1 represents thick-bedded to massive sandstone. Lithofacies 2 and 3 represent interbedded, thin-bedded sandstone and mudstone. In lithofacies 2, the sandstone beds are 2 – 20 cm thick and in lithofacies 3 the beds are less than 2 cm thick (Modified from Dutta, 2009).
Figure 2.17: Petrophysical analysis of the selected well-log interval using the Thomas-Stieber model. (Left) Total porosity and gamma ray values for three lithofacies in the selected interval. The median total porosity and gamma ray values for each lithofacies are shown in solid circles. From these median points, the up-down or left-right bars indicate the interquartile ranges (i.e., from 1st to 3rd quartiles) of each property. A Thomas-Stieber diagram is also superimposed on the data. (Right) Variation of gamma ray values with depth. Data points are color-coded by lithofacies similar to the left panel. The gamma ray log shows an upward-fining trend.
20 30 40 50 60 70 80 900.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Gamma ray
Tota
l poro
sity
Lithofacies 1
Lithofacies 2
Lithofacies 3
40 60 80
1070
1075
1080
1085
1090
1095
1100
Gamma ray
Depth
(m
)
0.25 m
Lithofacies 1 Lithofacies 2 Lithofacies 3
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 31
Figure 2.18: Probability density functions of properties of the sand end-point and the shale end-point. (Left) total porosity and (right) gamma ray value.
Figure 2.19: Estimated sand fractions for the selected well-log interval. (Left) gamma ray log of the selected interval and (right) posterior distributions of estimated sand fractions.
2.5 The Thomas-Stieber model on rock-physics cross-plots
In sections 2.3 and 2.4, we have discussed two types of clay/shale distribution (i.e.,
dispersed and laminar) in a sand-shale mixture and how the volume fraction of each
type can be estimated using the Thomas-Stieber model, which can be represented as a
plot between total porosity and shale volume fraction (or gamma ray). In this section,
we present the extension of this model to other rock properties on various examples of
0 0.2 0.4 0.60
2
4
6
8
10
Total porosity
0 50 1000
0.02
0.04
0.06
0.08
0.1
Gamma ray
Sand
Shale
40 60 80
1075
1080
1085
1090
1095
Gamma ray
Depth
(m
)
Lithof acies 1
Lithof acies 2
Lithof acies 3
0 0.5 1
1075
1080
1085
1090
1095
Sand fraction
Depth
(m
)
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 32
rock-physics cross-plots (Sections 2.5.1 – 2.5.4). Then in Section 2.5.5, we apply all
the cross-plots to the same dataset previously described in Section 2.4.3.
2.5.1 Density and volume fraction of shale
Although the Thomas-Stieber model is developed for porosity and shale volume
fraction, a similar model can be derived analogously for other volumetric properties
(e.g., density versus shale volume fraction). Density of any point on the diagram in
Figure 2.20 can be written as follows:
𝜌 = 1 − 𝑉𝑙𝑎𝑚 𝜌𝑠ℎ𝑎𝑙𝑦 𝑠𝑎𝑛𝑑 + 𝑉𝑙𝑎𝑚 𝜌𝑠ℎ ,
𝜌𝑠ℎ𝑎𝑙𝑦 𝑠𝑎𝑛𝑑 = 𝜌𝑠 + 𝑉𝑑𝑖𝑠𝑝 𝜌𝑠ℎ − 𝜌𝑓𝑙 ,
(2.9)
where 𝜌𝑠 ,𝜌𝑠ℎ and 𝜌𝑓𝑙 are the densities of clean sand, shale, and saturating fluid in the
effective pore space, respectively. Note that 0 ≤ 𝑉𝑑𝑖𝑠𝑝 ≤ 𝜙𝑠 . Thus, to construct the
diagram, we need not only the two end-points ( 𝜌𝑠 and 𝜌𝑠ℎ ) but also 𝜙𝑠.
Figure 2.20: Relationship between density and volume fraction of shale in shaly sand lamination. Point A and E represent clean sand point and pure shale point, respectively. Points B and C represent sand with dispersed clay, and point D is where the original sand pore space is completely filled with dispersed clay (Modified after Mavko et al., 2009).
If we assume that the mineral compositions in clean sand and pure shale are
known, an alternative parameterization can be derived. For example, when sand and
0 0.2 0.4 0.6 0.8 11.9
2
2.1
2.2
2.3
2.4
2.5
Vshale
De
nsity
Vlam = 0.2
Vlam = 0.6
Vlam = 0.8
Vlam = 0.4
B
C
D
A
E
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 33
shale are only composed of quartz and clay respectively, we can replace 𝜌𝑠 and 𝜌𝑠ℎ in
Equation 2.9 using the following expressions:
𝜌𝑠 = 𝜌𝑞𝑡𝑧 1 − 𝜙𝑠 + 𝜌𝑐𝑙𝑎𝑦 𝑉𝑑𝑖𝑠𝑝 1 − 𝜙𝑠ℎ + 𝜌𝑤𝑉𝑑𝑖𝑠𝑝 𝜙𝑠ℎ + 𝜌𝑓𝑙 𝜙𝑠 − 𝑉𝑑𝑖𝑠𝑝 ,
𝜌𝑠ℎ = 𝜌𝑐𝑙𝑎𝑦 + 𝜌𝑐𝑙𝑎𝑦 𝜙𝑠ℎ + 𝜌𝑤𝜙𝑠ℎ ,
(2.10)
where 𝜌𝑞𝑡𝑧 ,𝜌𝑐𝑙𝑎𝑦 ,𝜌𝑤𝑎𝑛𝑑 𝜌𝑓𝑙 are the densities of quartz, clay, water, and saturating
fluid in the effective pore space, respectively. In this alternative parameterization, we
write any density on the diagram in Figure 2.20 in terms of the two ―hidden‖
endpoints: 𝜙𝑠 and 𝜙𝑠ℎ . Note that while the Thomas-Stieber model, which relates total
porosity and shale volume fraction (Figure 2.2), is totally independent of saturating
fluids, the relation between density and shale volume fraction (Figure 2.20) does
depend on properties of the saturating fluids. Consequently when superimposing a
density model onto any dataset of sand-shale sequences, the fluid inside the effective
pore space of all the rock data must be the same.
2.5.2 Neutron porosity and density
Neutron-density cross-plots are commonly used in the shaly-sand analysis to
determine volume fraction of clay minerals (Asquith, 1990). Katahara (2008)
suggested that the contrast between neutron porosity and density porosity is often
directly related to clay content more than gamma ray values because radioactivity
measured by the gamma ray log may originate from other non-clay minerals (e.g.,
feldspar), or some clay minerals may be considered non-radioactive (e.g., kaolinite,
chlorite) (Guest, 1990). Figure 2.21 illustrates a neutron-density cross-plot for both the
dispersed and the laminated sand-shale systems.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 34
Figure 2.21: Density-neutron plot for the dispersed and laminated sand-shale systems.
2.5.3 Velocity and total porosity
Analogous to the Thomas-Stieber volumetric model, there are several rock-physics
models for elastic properties of the dispersed and laminated sand-shale systems
(Avseth et al., 2005). The dispersed system consists of two different domains: sandy
shale and shaly sand. The sandy shale describes a mixture where sand grains are
suspended in clay-rich matrix, whereas the shaly sand describes a mixture where
porous clay is added into the sand pore space. The transition between the two domains
occurs when the original sand pore space is completely filled with clay (Marion, 1990;
Yin, 1992; Dvorkin and Gutierrez, 2002). We refer to this transition point as the V-
point.
Dvorkin and Gutierrez (2002) used the Hashin-Shtrikman lower bound (HSLB) to
model sandy shale and the V-point, since the bound is realized by a mixture where
elastically stiffer material (i.e., quartz grains) is enveloped by softer material (i.e.,
porous shale). At the same time, they treated shaly sand as a HSLB mixture between
clean sand and the V-point. In this case, HSLB is also used; the bound is realized by a
mixture where stiffer material (i.e., sand filled with clay) is enveloped by softer
material (i.e., clean sand). When plotting elastic properties against total porosity for
the dispersed sand-shale system, an inverted-V trend is observed (Figure 2.22) and has
been shown by several authors to be good approximation for data from this type of
0 0.2 0.4 0.6 0.8 1
1
1.5
2
2.5
3
water
wet clay
clean sand
Neutron porosity
Density (
g/c
c)
direction of increasing
sand fractions
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 35
sand-shale system (e.g., Dvorkin and Gutierrez, 2002; Avseth et al., 2005; Flórez,
2005).
Figure 2.22: Velocity to total porosity curves for the dispersed sand-shale system. The model used here follows Dvorkin and Gutierrez, 2002.
Figure 2.23: Velocity to total porosity curves for the dispersed sand-shale system. The model used here follows Dvorkin and Gutierrez, 2002. Each black line represents lamination between sand (or shaly sand) and shale with volume fraction of laminated shale ranging from 0 to 1. Each blue line represents lamination between sand with a volume fraction of dispersed clay ranging from 0 to 𝜙𝑐𝑙𝑒𝑎𝑛 𝑠𝑎𝑛𝑑 and shale with a constant volume fraction of laminated shale.
For a wave traveling perpendicular to the layers in thin laminations, the effective
P-wave modulus (M) is computed using the Backus average (Mavko et al., 1998):
0 0.1 0.2 0.3 0.4 0.5 0.6
2
2.5
3
3.5
4
4.5
5
5.5
6
6.5
Total porosity
Vp
Shaly-sand line
Sandy-shale line
Shale
Sand
Quartz
V-point: Sand completelyfilled with clay
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.42
2.5
3
3.5
Shale
Sand
Total porosity
Vp (
km
/s)
V-point
Shaly-sand line
increasing
sand fractions
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 36
𝑀𝑙𝑎𝑚𝑖𝑛𝑎𝑟−1 = (1 − 𝑉𝑙𝑎𝑚 )𝑀𝑠𝑎𝑛𝑑 (𝑜𝑟 𝑠ℎ𝑎𝑙𝑦 𝑠𝑎𝑛𝑑 )
−1 + 𝑉𝑙𝑎𝑚𝑀𝑠ℎ−1 ,
𝑉𝑙𝑎𝑚 = 1 − 𝑉𝑠𝑎𝑛𝑑 .
(2.11)
In Figure 2.23, each black line represents lamination between sand (or shaly sand)
and shale with a volume fraction of laminated shale ranging from 0 to 1. Each blue
line represents lamination between sand with a volume fraction of dispersed clay
ranging from 0 to 𝜙𝑐𝑙𝑒𝑎𝑛 𝑠𝑎𝑛𝑑 and shale with a constant volume fraction of laminated
shale.
2.5.4 Vp/Vs ratio and acoustic impedance
A cross-plot between Vp/Vs ratio and acoustic impedance (AI) is a useful rock-
physics template (RPT) for lithology and pore fluid interpretations. The cross-plot is
constructed by first computing the dry rock moduli at the high-porosity end member
using Hertz-Mindlin contact theory. Then this high-porosity point is connected to the
zero-porosity mineral point with either the modified lower or upper Hashin-Shtrikman
bounds, depending on the choice of sedimentological trends (e.g., using the lower
bound for the sorting trend). The corresponding saturated rock moduli are computed
using Gassmann’s equations (Avseth et al., 2005). Figure 2.24 illustrates an example
of Vp/Vs versus AI cross-plots with incorporated dispersed and laminated sand-shale
models.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 37
Figure 2.24: Rock-physics template shown as a cross-plot between Vp/Vs and AI with superimposed rock-physics trends. The green and magenta curves represent the shale and wet-sand lines, respectively. Along these lines, the change in porosity is due to packing or grain sorting. At each porosity value, the red curve which is connected to the wet-sand line represents the corresponding gas-saturated sand with varying saturations. The dispersed and laminated sand-shale system is constructed using the sand and shale points shown in red circles. Each blue line represents lamination between sand with a volume fraction of dispersed clay ranging from 0 to 𝜙𝑐𝑙𝑒𝑎𝑛 𝑠𝑎𝑛𝑑 and shale with a constant volume fraction of laminated shale.
2.5.5 Application to real data
In Sections 2.5.1 – 2.5.4, we explored the analogs of the Thomas-Stieber model on
various cross-plots. In this section, we integrate different cross-plots and apply them to
the same dataset used in Section 2.4.3; before cross-plotting the data, we perform fluid
substitution on the data using our mesh method2
(Chapter 3) so that they are
completely saturated with brine. Fluid substitution results are shown on various cross-
plots in Figure 2.25. The laminated trend fits the data points on all cross-plots quite
well.
2 The mesh method refers to our proposed fluid-substitution method for interbedded sand-shale
sequences. Refer to Chapter 3 for more details about this method.
2 3 4 5 6 7 8 9
1.5
2
2.5
3
3.5
4
10% porosity
20%
30%
40%
50%
60%
10% porosity
20%
30%
0% gas
100%
AI (g/cm3 x km/s)
Vp/V
s
Shale
Sand
70%
40%
V-pointincreasing
sand fractions
shaly-sand line
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 38
Figure 2.25: Selected well-log data on different cross-plots for laminated and dispersed sand-shale systems. The data points are color-coded by their corresponding gamma ray values.
2.6 Discussion
2.6.1 Extension of the uncertainty analysis
We have shown the sensitivity and uncertainty analyses on the Thomas-Stieber
model represented as cross-plots between total porosity and gamma ray (or shale
volume fraction). The same analyses (e.g., uncertainty in endpoints) can be applied to
cross-plots of other volumetric or elastic properties. Note that this chapter does not
incorporate all available data for evaluating thinly bedded sand-shale reservoirs. Other
useful data such as resistivity, elastic anisotropy, NMR-T2 distribution can also be
incorporated into the analyses (e.g., Georgi and Schӧn, 2005; Passey et al., 2006).
Furthermore, as noted earlier, even though we do not account for uncertainty in the
measurements in this chapter, this uncertainty can be incorporated into the analyses
easily by following similar procedures as described in Section 2.4.2.
0 0.1 0.2 0.3 0.42
2.5
3
3.5
Shale
Sand
Total porosity
Vp (
km
/s)
0 0.5 10
0.1
0.2
0.3
0.4
Shale
Sand
Shale volume
Tota
l poro
sity
0 0.5 12
2.2
2.4
2.6
2.8
ShaleSand
Shale volume
Density (
g/c
c)
0.1 0.2 0.3 0.4
2.1
2.2
2.3
2.4
2.5
water
Shale
Sand
Neutron porosity
Density (
g/c
c)
30
40
50
60
70
gamma ray
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 39
2.6.2 Constraint on properties of the end-points
By integrating elastic properties into the analysis of thinly bedded sand-shale
reservoirs, rock-physics models can help constrain the choices of the end-point
properties and reduce uncertainty in these input parameters into the Thomas-Stieber
model. For example, elastic properties of well-sorted, clean sands can be computed
using Hertz-Mindlin contact theory and Gassmann’s fluid substitution equations.
Since elastic properties are combined with volumetric properties, the cross-plot
analyses are applicable not only to well log data for investigating the sub-resolution
properties, but also to data at other scales, for example, to seismic inversion results
(e.g., Vp/Vs, AI).
2.6.3 Effect of resolution on consistency in interpretation of different cross-plots
Because geophysical or well logging tools have their own vertical resolutions,
their measurements may represent properties of different volumes of rocks. Therefore,
combining various measurements can create inconsistency in the interpretation. For
example, when cross-plotting measurements from a lower-resolution gamma ray log
against those from a higher-resolution density log, if the rocks are very heterogeneous
at a short scale, the cross-plot may illustrate a misleading pattern or trend, which
affects the interpretation. Resolution of well logging tools can be affected by many
factors including logging speed, tool geometry (Serra, 1984).
We generate two synthetic examples to illustrate the effect of resolution on
interpretation from cross-plots. The first example is represented by an earth model
composed of shaly sand interbedded with shale (Figure 2.26, right). All layers have
the same thickness equal to one unit. The shaly sand and shale layers have fixed
volume fractions of dispersed clay equal to 0.1 and 1, respectively. Their
corresponding total porosity values are 0.225 and 0.25, respectively. We compute
arithmetic averages of both the shale volume and total porosity using window lengths
of 45 and 15 units to simulate a lower-resolution gamma ray log (i.e., shale volume)
and a higher-resolution density log (i.e., density porosity), respectively. The cross-plot
between total porosity and shale volume with different resolutions is shown in Figure
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 40
2.26 (left). For comparison, we also compute and cross-plot total porosity and shale
volume when both measurements have the same resolutions (Figure 2.26, middle).
When cross-plotting the two properties that have different vertical resolutions, the data
points are slightly misplaced. As a result, the interpretation of sand fractions (V𝑠𝑎𝑛𝑑 )
and volume fractions of dispersed clay ( V𝑑𝑖𝑠𝑝 ) using this cross-plot is slightly
erroneous (Figure 2.27). Note that the magnitude of the error also depends on the
contrast in layer properties that are averaged and the geometric arrangement of these
layers.
Figure 2.26: The effect of vertical resolution on cross-plots between total porosity and shale volume of a synthetic earth model between shaly sand and shale. The left panel shows a cross-plot between the two measurements with two different resolutions; whereas, the middle panel shows a cross-plot when both measurements share the same resolution. The right panel is a short section of the corresponding synthetic earth model. Sand and shale layers are shown in white and black, respectively.
Figure 2.27: Estimated sand fraction and sand property when applying Thomas-Stieber diagram to points in Figure 2.26 to investigate the effect of vertical resolution on interpretation using cross-plots between total porosity and shale volume.
0 0.5 10.05
0.1
0.15
0.2
0.25
0.3
Shale volume
Tota
l poro
sity
Different vertical resolutions
0 0.5 10.05
0.1
0.15
0.2
0.25
0.3
Shale volume
Tota
l poro
sity
Same vertical resolution
sand point sand point
shale pointshale point
0 0.5 10
0.2
0.4
0.6
0.8
1
True Vsand
Estim
ate
d V
sand
0 0.2 0.40
0.1
0.2
0.3
0.4
True Vdisp
Estim
ate
d V
dis
p
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 41
The second example of the effect of resolution has a setup similar to the first
example, but the volume fractions of dispersed clay in the shaly-sand layers are not
fixed and the layer arrangement is non-stationary along the sequence. When cross-
plotting the two properties that have different vertical resolutions, the data points are
clearly misplaced (Figure 2.28, left). Interestingly, the interpretation of sand fractions
( V𝑠𝑎𝑛𝑑 ) is only slightly affected by the difference in resolutions; whereas, the
interpretation of volume fractions of dispersed clay (V𝑑𝑖𝑠𝑝 ) is more affected (Figure
2.29).
Figure 2.28: The effect of vertical resolution on cross-plots between total porosity and shale volume of a synthetic (non-stationary) earth model between shaly sand and shale. The left panel shows a cross-plot between the two measurements with two different resolutions; whereas, the middle panel shows a cross-plot when both measurements share the same resolution. The right panel is a short section of the corresponding synthetic earth model. Sand and shale layers are shown in white and black, respectively.
Figure 2.29: Estimated sand fraction and sand property when applying Thomas-Stieber diagram to points in Figure 2.28 to investigate the effect of vertical resolution on interpretation using cross-plots between total porosity and shale volume.
0 0.5 10.05
0.1
0.15
0.2
0.25
0.3
0.35
Shale volume
Tota
l poro
sity
Different vertical resolutions
0 0.5 10.05
0.1
0.15
0.2
0.25
0.3
0.35
Shale volume
Tota
l poro
sity
Same vertical resolution
shale point
sand pointsand point
shale point
0 0.5 10
0.2
0.4
0.6
0.8
1
True Vsand
Estim
ate
d V
sand
0 0.1 0.2 0.3 0.40
0.1
0.2
0.3
0.4
True Vdisp
Estim
ate
d V
dis
p
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 42
2.7 Conclusions
The Thomas-Stieber model can be used for evaluating thinly bedded sand-shale
reservoirs as it can help determine both the sand fractions within the lamination and
the property of the sand. We present the sensitivity and uncertainty analyses of the
model, when natural variations are integrated into the model in order to quantify
uncertainty associated with reservoir evaluations. Using synthetic examples, we
illustrate that even when the correct end-points are used in the Thomas-Stieber model,
rocks in nature can be so complex that they violate several assumptions of the model,
resulting in erroneous interpretation of reservoir properties. When the properties of the
end-points are uncertain, this uncertainty can be incorporated into the model using
Monte Carlo simulations in a Bayesian framework. The Thomas-Stieber model is not
only limited to a cross-plot between total porosity and gamma ray (or shale volume),
but also other rock properties including elastic properties. Rock-physics models can
then be used to constrain the properties of the end-points. If measurements come from
tools with different vertical resolutions, there can be inconsistency among the cross-
plots. Through synthetic examples, we show that this resolution problem can lead to
erroneous interpretation and that the estimated sand fractions seem to be affected by
the resolution less severe than the estimated volume fractions of dispersed clay.
2.8 Acknowledgements
This work was supported by the SRB Project, the American Chemical Society
grant number 46350-AC8 and the Stanford Center for Reservoir Forecasting. We
would like to thank Hess Corporation for providing the data.
2.9 References
Asquith, G. B., 1990, Log evaluation of shaly sandstones: A practical guide: American
Association of Petroleum Geologists, Course note series # 31.
Avseth, P., Mukerji, T. and Mavko, G., 2005. Quantitative Seismic Interpretation,
Cambridge.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 43
Ball, V., Erickson, S., Brown, L., 2004. A Model-centric Approach to Seismic
Petrophysics: SEG Expanded Abstracts, 23, 1730.
Boggs, S., 2001, Principles of sedimentology and Stratigraphy, 3rd
ed.
Chopra, S., 2005, Expert Answers: Gassmann’s equation: CSEG Recorder in May, 8-
12.
Cumberland, D. J. and Crawford, R. J., 1987, The packing of particles: Handbook of
powder technology, v.6.
Dawson, W. C., W. R. Almon, K. Dempster, and Sutton, S. J., 2008, Shale variability
in deep-marine depositional systems: implications for seal character – subsurface
and outcrop examples: American Association of Petroleum Geologists, Search and
Discovery Article #50128, accessed 3 March 2009;
http://www.searchanddiscovery.com/documents/2008/08144dawson/ndx_dawson.
Dutta, T., 2009. Integrating Sequence Stratigraphy and Rock-physics to Interpret
Seismic Amplitudes and Predict Reservoir Quality, Ph.D. Thesis, Stanford
University.
Dvorkin, J. and Gutierrez, M. A., 2002, Grain sorting, porosity and elasticity:
Petrophysics, 43(3), 185-196.
Flórez, J., 2005. Integrating Geology, Rock physics, and Seismology for Reservoir-
quality Prediction, Ph.D. Thesis, Stanford University.
Gassmann, F., 1951. Uber die elastizitat poroser medien: Vier Natur Gesellschaft, 96,
1-23.
Georgi, D. and Schӧn, J., 2005, Elastic wave anisotropy and shale distribution:
SPWLA 46th
Annual Logging Symposium, June 26-29, Paper Q.
Guest, K., 1990, The use of core derived quantitative mineralogical data to improve
formation evaluation: Advances in core evaluation: Accuracy and precision in
reserves estimation, Ed. Paul F. Worthington, Pennsylvania: Gordon and Breach
Science Publishers, 187-210.
Gujer, W., 2008, Systems analysis for water technology, Springer.
Juhasz, I., 1986, Assessment of the distribution of shale, porosity and hydrocarbon
saturation in shaly sands: Transactions of the SPWLA 10th European Formation
Evaluation Symposium, Aberdeen, 22 April, AA1-15.
Katahara, K., 2004, Fluid substitution in laminated shaly sands: 74th Annual Meeting,
SEG Expanded Abstracts, 1718-1721.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 44
Katahara, K., 2008, What is Shale to a Petrophysicist?: The Leading Edge, 27, 738-
741.
Koltermann, C. E. and Gorelick, S. V., 1995, Fractional packing model for hydraulic
conductivity derived from sediment mixtures: Water Resources Research, 31(12),
3283 – 3297.
Lowe, D. R., 2004, Report on core logging, lithofacies, and basic sedimentology of
Equatorial Guinea: Hess internal report.
Marion, D., 1990, Acoustical, mechanical and transport properties of sediments and
granular materials: Ph.D. Thesis, Stanford University.
Mavko, G., Mukerji, T., and Dvorkin, J., 1998. The Rock Physics Handbook, 1st
edition, Cambridge.
Mavko, G., Mukerji, T., and Dvorkin, J., 2009, The Rock Physics Handbook, 2nd
edition, Cambridge.
Passey, Q. R., Dahlberg, K. E., Sullivan, K. B., Yin, H., Brackett, R. A., Xiao, Y. H.,
and Guzmán-Garcia, A. G., 2006, Petrophysical evaluation of hydrocarbon pore-
thickness in thinly bedded clastic reservoirs: AAPG Archie series, 1, American
Association of Petroleum Geologists, Tulsa, Oklahoma, U.S.A.
Revil, A., Grauls, D., and Brévart, O., 2002, Mechanical compaction of sand/clay
mixtures: Journal of Geophysical Research, 107(B11).
Serra, O., 1984, Fundamentals of well-log interpretation: The acquisition of logging
data, 1st edition, Elsevier.
Skelt, C., 2004a, Fluid substitution in laminated sands: The Leading Edge, 23, 485–
488.
Skelt, C., 2004b, The influence of shale distribution on the sensitivity of
compressional slowness to reservoir fluid changes: SPWLA 45th
Annual Logging
Symposium, June 6-9.
Thomas, E. C. and Stieber, S. J., 1975, The distribution of shale in sandstones and its
effect upon porosity: 16th Annual Logging Symposium, SPWLA, paper T.
Voleti, D., Tyagi, A., Singh, A., Mundayat, V., Pandey, V., and Saxena, K., 2012, A
new petrophysical interpretation approach to characterize the thinly laminated
reservoir using conventional tools: SPE Oil and Gas India conference and
exhibition, SPE 152912.
Worden, R. H., and Morad, S., 2003, Clay minerals in sandstones: controls on
formation distribution and evolution: Int. Assoc. Sedimentol. Spec. Publ., 34, 3 –
41.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 45
Worthington, P. F., 2003, Effect of clay content upon some physical properties of
sandstone reservoirs: Int. Assoc. Sedimentol. Spec. Publ., 34, 191 – 211.
Yadav, L., Dutta, T., Kundu, A., and Sinha, N., 2010, A new approach for the realistic
evaluation of very thin reservoirs of Krishna Godavari basin, East coast, India:
SPE Asia Pacific Oil & Gas conference and exhibition, SPE 132970.
Yin, H., 1992, Acoustic Velocity and Attenuation of Rocks: Isotropy, Intrinsic
Anisotropy, and Stress Induced Anisotropy, Ph.D. Thesis, Stanford University.
CHAPTER 2: Sensitivity and uncertainty analysis of the Thomas-Stieber model 46
47
Chapter 3
Fluid substitution for sub-resolution
interbedded sand-shale sequences
using the mesh method
3.1 Abstract
This chapter provides a simple graphical mesh interpretation and accompanying
equations for approximating fluid substitution in sub-resolution interbedded sand-
shale sequences. The sand layers can be either clean or shaly (i.e., sand with dispersed
clay). Because geophysical logging tools have limited vertical resolution, their
measurements often represent average properties of multiple sedimentary layers. If
this sub-resolution sand-shale interbedding is not properly accounted for, applying
fluid substitution at the measurement scale can lead to erroneous predictions. Given
that all the properties and input parameters are known, physically the most accurate
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 48
way to perform fluid substitution in these interbedded sequences is to first downscale
(i.e., invert) the measurements for the properties of the two end-members, which are
sand and shale. Gassmann’s equation is then applied only to the sand end-member,
and finally the properties of the sand with a new fluid and the shale are upscaled (i.e.,
averaged) back to the measurement scale. However, in practice the required input
parameters are only estimates. Therefore, fluid substitution by the downscaling-
upscaling procedure can be unstable and non-robust.
We propose a mesh method which combines rock-physics models for dispersed
and interbedded sand-shale systems, the Thomas-Stieber model, and Gassmann’s fluid
substitution equation so that our method can be used for fluid substitution in
interbedded sand-shale sequences directly at the measurement scale, without the need
to downscale the measurements, while still changing fluid in the sand layers only. We
apply our method to both synthetic and real well log data, and compare the results
with those predicted by simply using Gassmann’s equation (which ignores the effect
of thin sand-shale interbedding) and by the downscaling-upscaling procedure. Since
ideally the downscaling-upscaling procedure is the appropriate way to perform fluid
substitution in interbedded sand-shale sequences, we use fluid substitution results from
this procedure as baselines for both synthetic and real data. These baselines are
considered correct in synthetic data, because all the parameters to be estimated are
known. However, as previously noted the downscaling-upscaling procedure may yield
unreliable results for real data, especially in data with very low sand fractions.
Therefore, we select only a portion of the results from the downscaling-upscaling
procedure as the baselines for real data.
The results of our comparison are the following. Since Gassmann’s equations do
not account for the effect of sub-resolution sand-shale interbedding, the changes in
elastic moduli after fluid substitution are overpredicted when compared with the
baselines. In contrast, the results from the mesh method agree well with the baselines.
Thus even when our mesh method is applied at the measurement scale, the method
appropriately accounts for the effect of sub-resolution sand-shale interbedding. We
also perform a sensitivity analysis of five input parameters needed for the mesh
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 49
method which are in decreasing order of sensitivity clean sand velocity, clean sand
porosity, effective porosity, effective water saturation and fluid elastic modulus.
3.2 Introduction
Fluid substitution is an important rock physics technique used to quantitatively
predict elastic properties of rocks when one saturating pore fluid replaces another. One
tool commonly used for fluid substitution is the isotropic Gassmann’s equation
(Gassmann, 1951). Fluid substitution using this equation involves two steps. First, the
rock’s dry-frame bulk modulus (𝐾𝑑𝑟𝑦 ) is estimated from the initial measurements of
rocks saturated with the original “fluid 1” (e.g., P- and S-wave velocities, density)
using the following relation (Mavko et al., 2009):
𝐾𝑑𝑟𝑦 =𝐾𝑠𝑎𝑡 ∗ 𝜙𝑇𝐾𝑚𝑖𝑛 𝐾𝑓𝑙∗ + 1 − 𝜙𝑇 − 𝐾𝑚𝑖𝑛𝜙𝑇𝐾𝑚𝑖𝑛 𝐾𝑓𝑙∗ + 𝐾𝑠𝑎𝑡 ∗ 𝐾𝑚𝑖𝑛 − 1 − 𝜙𝑇
.
(3.1)
In equation 3.1, 𝐾𝑠𝑎𝑡 ∗ and 𝐾𝑓𝑙∗ are the bulk modulus of the saturated rock and the bulk
modulus of the original saturating fluid (fluid 1). 𝐾𝑚𝑖𝑛 is the bulk modulus of the solid
mineral, and 𝜙𝑇 is the total porosity. Second, this dry-frame modulus is used to predict
the rock’s bulk modulus when pore fluid 1 is replaced by “fluid 2” using the following
relation:
𝐾𝑠𝑎𝑡 =𝜙𝑇 1 𝐾𝑚𝑖𝑛 − 1 𝐾𝑓𝑙 + 1 𝐾𝑚𝑖𝑛 − 1 𝐾𝑑𝑟𝑦
𝜙𝑇 𝐾𝑑𝑟𝑦 1 𝐾𝑚𝑖𝑛 − 1 𝐾𝑓𝑙 + 1 𝐾𝑚𝑖𝑛 1 𝐾𝑚𝑖𝑛 − 1 𝐾𝑑𝑟𝑦 ,
(3.2)
where 𝐾𝑓𝑙 is the bulk modulus of fluid 2. Although the bulk modulus of the saturated
rock depends on the pore fluid, Gassmann’s equations predict that the shear modulus
remains unchanged under fluid substitution (e.g., Berryman, 1999; Smith et al., 2003).
Even though Gassmann’s fluid substitution equation seems simple and easy to
implement, applying it to real rocks often yields erroneous results, because real rocks
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 50
can be so complex that several assumptions in Gassmann’s theory are violated (e.g.,
Smith et al., 2003; Han and Batzle, 2004; Chopra, 2005). For example, Gassmann’s
theory assumes that rocks are homogeneous. However, since geophysical tools have
limited vertical resolution, their measurements always represent averages of the
intrinsically heterogeneous rocks. Consider, for example, a stack of sub-resolution
interbedded sand and shale layers. Shaly portions of the stack have low permeability
and contain bound water associated with their clay-mineral component; these
characteristics violate Gassmann’s assumption regarding pore connectivity and perfect
pore-fluid communication. As a result, the traditional Gassmann’s equation may not
be appropriate for fluid substitution in the shale layers (Chopra, 2005; Dvorkin et al.,
2007; Katahara, 2008). Therefore, if the isotropic Gassmann’s fluid substitution
equation is applied to the measurements at their original scales, without accounting for
the sub-resolution sand-shale interbedding, the resulting predictions may be erroneous
(e.g., Katahara, 2004; Skelt, 2004a; Skelt, 2004b; Chopra, 2005).
The physically correct way to perform fluid substitution in the thinly-layered
system is to first downscale the measurements for the separate sand and shale end-
members’ properties, then apply Gassmann’s equation to the sand layers only, and
finally upscale the layers back to the measurement scale by using the Backus average
(Katahara, 2004; Skelt, 2004a; Chopra, 2005; Singleton and Keirstead, 2011). This
downscaling-upscaling procedure is an ideal solution when the shale properties and
sand fraction are known. However, in practice uncertainties in estimated parameters
and uncertainties in measurements can affect fluid substitution results, especially
during the downscaling step. Several authors (e.g., Katahara, 2004; Skelt, 2004a;
Singleton and Keirstead, 2011) have proposed methods for fluid substitution in
interbedded sand-shale sequences in which the sand is assumed to be clean and its
volume fraction is equivalent to simply (1 − 𝑉𝑠), where 𝑉𝑠 is shale volume fraction.
Thus these methods are not intended for the case where the sandy layers are shaly (i.e.,
sand having dispersed clay inside its pore space).
The presence of clay dispersed in the pore space of sand complicates the fluid
substitution problem, because as previously noted, immobile bound-water in clay may
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 51
block communication of the pore fluids, which violates a basic assumption of
Gassmann’s equation (Smith et al., 2003; Mavko et al., 2006; Dvorkin et al., 2007;
Simm, 2007). Several proposed methods for fluid substitution in shaly sand are simply
modifications of the traditional Gassmann’s equation. Simm (2007) discussed
workflows for fluid substitution in shaly sand and suggested an alternative approach
which adjusts the dry-rock modulus parameters in Gassmann’s equation using
empirical trends. Dvorkin et al (2007) proposed a method for fluid substitution in
shaly sand which uses effective porosity instead of total porosity in the traditional
Gassmann’s equation, and also assumes a new composite mineral phase by combining
wet porous shale and solid minerals together. Both of these proposed methods are
intended for fluid substitution in shaly sand, but not for interbedded sequences. Skelt
(2004b) showed that when laminated (interbedded) shale is modeled as dispersed,
large errors may be introduced into the predicted compressional velocities. Thus, none
of the above methods provides a complete solution for the problem of fluid
substitution in interbedded (shaly) sand-shale sequences.
In this chapter, we propose a simple graphical “mesh” method for fluid
substitution in sub-resolution interbedded sand-shale sequences, which appropriately
changes fluid in the sands only, without the need to downscale the measurements.
Unlike previous methods which are limited to fluid substitution for interbedded clean
sand and shale or for simply shaly sand, our mesh method is applicable also to
interbedded sand-shale sequences, in which the sand layers can be either clean or
shaly. The remainder of this chapter is organized as follows. Section 3.3 discusses
rock-physics models for a sand-shale mixture system, including dispersed and laminar
(interbedded) mixes. In Section 3.4, we integrate fluid substitution into the sand-shale
models, leading to the graphical mesh interpretation and accompanying equations for
fluid substitution in interbedded sand-shale sequences. Section 3.5 shows a
comparison between fluid substitution results by the mesh approach and by other
methods using synthetic data. Section 3.6 discusses some pitfalls in interpretation
when sub-resolution is not accounted for. Then, in Section 3.7, we run a sensitivity
analysis of input parameters needed for the mesh method. In Section 3.8, we test our
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 52
method on real data, and finally in Section 3.9 we discuss limitations and possible
extensions of the mesh approach.
3.3 Elastic properties of interbedded sands
3.3.1 Models
While the Thomas-Stieber volumetric model1 (Thomas and Stieber, 1975) relates
total porosity to shale distribution in the sand-shale mixtures, several rock-physics
models have explored elastic properties of the dispersed and interbedded sand-shale
systems (Avseth et al., 2005). The dispersed portion of the sand-shale system consists
of two different domains: “sandy shale” and “shaly sand” (Figure 3.1). The sandy
shale describes a mixture where sand grains are suspended in clay-rich matrix,
whereas the shaly sand describes a mixture where porous clay is added into the sand
pore space. The transition between the two domains occurs when the original sand
pore space is completely filled with clay (Marion, 1990; Yin, 1992; Dvorkin and
Gutierrez, 2002). We call this transition point the V-point.
Dvorkin and Gutierrez (2002) used the Hashin-Shtrikman lower bound (HSLB) to
model sandy shale and the V-point, since the bound is realized by a mixture where
elastically stiffer material (i.e., quartz grains) is enveloped by softer material (i.e.,
porous shale). At the same time, they treated shaly sand as a HSLB mixture between
clean sand and the V-point. In this case, the HSLB is also used; the bound is realized
by a mixture where stiffer material (i.e., sand filled with clay) is enveloped by softer
material (i.e., clean sand). When plotting elastic properties against total porosity, the
shaly-sand and sandy-shale legs of the dispersed sand-shale system form an “inverted-
V” trend (Figure 3.1), and this trend has been shown by several authors to be good
approximation for data from this type of sand-shale system (e.g., Dvorkin and
Gutierrez, 2002; Avseth et al., 2005; Flórez, 2005).
1 Refer to Chapter 2 for more details.
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 53
Figure 3.1: Inverted-V relation between P-wave velocity and total porosity for a
dispersed sand-shale system, following the model of Dvorkin and Gutierrez (2002). In this case, curves are computed using the Hashin-Shtrikman lower bound (HSLB). The solid magenta and blue curves represent the sandy-shale and shaly-sand legs of the dispersed sand-shale system.
We follow an approach similar to that of Dvorkin and Guiterrez (2002) to model
elastic properties of shaly sand with a slight modification in the calculation of the V-
point. Instead of using HSLB for the V-point, we use the Voigt-Reuss-Hill average
(VRH) between properties of quartz at zero total porosity (𝜙𝑇= 0) and porous shale at
𝜙𝑇= 𝜙𝑠 . This achieves better consistency in fluid substitution results for synthetic
data. The elastic modulus at the V-point corresponds to the value of the VRH average,
evaluated at the point where 𝜙𝑇= 𝜙𝑠𝜙𝑠 . The shaly-sand line can then be modeled by
a lower elastic bound that connects the V-point and the clean sand point (𝜙𝑇= 𝜙𝑠).
When the elastic contrast between the two materials is relatively small, using either
HSLB or the Reuss lower bound (with P-modulus only) gives similar results (Dvorkin
and Gutierrez, 2002). In this study, we model the shaly-sand line using the Reuss
average of P-wave moduli rather than the bulk and shear modulus, since S-wave
information is often unavailable or unreliable. Using only the P-wave moduli,
equations for elastic properties of the V-point and the shaly-sand line are as follows:
𝑉𝑠 = 𝜙𝑠 : 𝑀𝑉𝑜𝑖𝑔𝑡 ≈ (1 − 𝜙𝑠)𝑀𝑞𝑡𝑧 + 𝜙𝑠𝑀𝑠 ,
𝑉𝑠 = 𝜙𝑠 : 𝑀𝑅𝑒𝑢𝑠𝑠−1 ≈ (1 − 𝜙𝑠)𝑀𝑞𝑡𝑧
−1 + 𝜙𝑠𝑀𝑠−1 ,
𝑉𝑠 = 𝜙𝑠 : 𝑀𝑠𝑎𝑛𝑑 −𝑓𝑖𝑙𝑙𝑒𝑑 −𝑤𝑖𝑡 −𝑠𝑎𝑙𝑒 ≈ 0.5 ∗ 𝑀𝑉𝑜𝑖𝑔𝑡 + 𝑀𝑅𝑒𝑢𝑠𝑠 |𝑣𝑠= 𝜙𝑠 ,
0 ≤ 𝑉𝑠 ≤ 𝜙𝑠 : 𝑀𝑠𝑎𝑙𝑦 −𝑠𝑎𝑛𝑑−1 ≈ (1 − 𝐹)𝑀𝑠
−1 + 𝐹𝑀𝑠𝑎𝑛𝑑 −𝑓𝑖𝑙𝑙𝑒𝑑 −𝑤𝑖𝑡 −𝑠𝑎𝑙𝑒−1 ,
0 0.1 0.2 0.3 0.4 0.5 0.6
2
2.5
3
3.5
4
4.5
5
5.5
6
6.5
Total porosity
Vp
Shaly-sand line
Sandy-shale line
Shale
Sand
Quartz
V-point: Sand completelyfilled with clay
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 54
𝑤𝑖𝑡 𝐹 =𝑉𝑠𝜙𝑠
.
(3.3)
In equation 3.3, 𝑀 is the P-wave modulus, and subscripts s, sh, and qtz stand for clean
sand, pure shale and quartz, respectively.
Figure 3.2: Relation between velocity and total porosity for a dispersed sand-shale system, following the model of Dvorkin and Gutierrez (2002) with a slight modification. Each red line represents interbedding of sand (or shaly sand) and shale with volume fraction of interbedded shale ranging from 0 to 1. Each blue line represents interbedding of sand with a volume fraction of dispersed clay ranging from 𝑉𝑑𝑖𝑠𝑝 = 0 to 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑆 and shale with a constant volume fraction of interbedded shale.
For a wave traveling perpendicular to the layers in thin sand-shale sequences, the
effective P-wave modulus is computed using the Backus average (Backus, 1962;
Mavko et al., 1998):
𝑀𝑙𝑎𝑚𝑖𝑛𝑎𝑟−1 = (1 − 𝑉𝑙𝑎𝑚 )𝑀𝑠𝑎𝑛𝑑 (𝑜𝑟 𝑠𝑎𝑙𝑦 𝑠𝑎𝑛𝑑 )
−1 + 𝑉𝑙𝑎𝑚𝑀𝑠−1 ,
𝑉𝑙𝑎𝑚 = 1 − 𝑉𝑠𝑎𝑛𝑑 ,
(3.4)
where 𝑉𝑙𝑎𝑚 is the volume fraction of laminated (interbedded) shale and 𝑉𝑠𝑎𝑛𝑑 is the
volume fraction of laminated (interbedded) sand which can be clean or shaly. We refer
to this volume as the sand fraction. As noted in Chapter 1, in this dissertation we use
0 0.1 0.2 0.3 0.4
2
2.5
3
3.5
4
4.5
5
Total porosity
Vp (
km
/s)
Shale
Sand
V-point
(HSLB)
V-point
(VRH)
increasing
sand fractions
Shaly-sand line
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 55
the terms laminated and interbedded interchangeably. Both the laminated sand-shale
sequence and the interbedded sand-shale sequence here refer to a sequence of
alternating sand and shale units.
Both Equations 3.3 and 3.4 are simply harmonic averages of the P-wave moduli of
all components. These equations are equivalent to the volume-weighted arithmetic
averages of the P-wave compliances or compressional compliances (C = 1/M). Figure
3.3 shows a cross-plot between the compressional compliance and effective porosity.
In this domain, all the curves become straight lines. In both Figure 3.2 and Figure 3.3,
each red line represents interbedding of sand (or shaly sand) and shale. Along the line,
the volume fraction of laminated (interbedded) shale (𝑉𝑙𝑎𝑚 ) changes from 𝑉𝑙𝑎𝑚 = 0 to
𝑉𝑙𝑎𝑚 = 1 but 𝑉𝑑𝑖𝑠𝑝 is constant. Each blue line represents interbedding of sand and shale,
where 𝑉𝑙𝑎𝑚 is held constant and 𝑉𝑑𝑖𝑠𝑝 changes from 𝑉𝑑𝑖𝑠𝑝 = 0 to 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑠. We refer to
the set of red and blue lines as the mesh. (Note that non-linear parameterization of the
shaly-sand curves can be chosen by using alternative models for the curves instead of
using the Reuss average.)
Figure 3.3: P-wave compliance (C=1/M) versus effective porosity curves for a interbedded sand-shale system. Each red line represents interbedding of sand (or shaly sand) and shale with volume fraction of laminated (interbedded) shale ranging from 𝑉𝑙𝑎𝑚 = 0 to 𝑉𝑙𝑎𝑚 = 1. Each blue line represents interbedding of sand with a volume fraction of dispersed clay ranging from 𝑉𝑑𝑖𝑠𝑝 = 0 to 𝑉𝑑𝑖𝑠𝑝 =𝜙𝑆 and shale with a constant volume fraction of laminated (interbedded) shale.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.40.02
0.04
0.06
0.08
0.1
Effective porosity
1/M
(1/G
Pa)
Increasing sand fractions
Clean sand
Pure shale
V point: sand completely
filled with clay
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 56
In this work, we assume that fluid substitution only occurs in the effective porosity.
Since the shale point and the points along the sandy-shale line have zero effective
porosity, their properties remain unchanged under fluid substitution.
3.3.2 Modeling the V-point using the Voigt-Reuss-Hill average
Even though the P-wave velocity of the V-point from VRH is much larger than
that estimated from the HSLB (Figure 3.2), the VRH average is chosen here to ensure
consistency with the fluid substitution calculations. In the case of non-monomineralic
rocks, the effective mineral bulk modulus is needed as an input in Gassmann’s
equation. Since we will follow common practice and use the VRH average to estimate
this effective modulus (Dvorkin and Guiterrez, 2002; Smith et al., 2003; Kumar, 2006),
it is essential to also model shaly-sand in our synthetic data using the VRH to calculate
the V-point.
To illustrate possibly inconsistent results during fluid substitution, we test various
scenarios where existing models are mixed and matched to generate three elements
needed in fluid substitution for synthetic data: the shaly-sand line, V-point, and
effective mineral modulus. First, we model the brine-saturated shaly sand line and the
V-point with the specifications listed in Table 3.1. In every case, brine will be replaced
by oil. The method of Dvorkin et al (2007) for fluid substitution in shaly sand uses
effective porosity instead of total porosity in Gassmann’s equation (Equations 3.1 and
3.2) and also assumes a composite mineral phase consisting of wet porous shale and
solid minerals together. Hereafter we refer to this method as the Gassmann shaly-sand
model. The effective porosity (𝜙𝐸𝑓𝑓 ) can be calculated by following the dispersed
clay-porosity relation in the Thomas-Stieber model (Equation 3.5; refer to Chapter 2
for more details).
0 ≤ 𝑉𝑠 ≤ 1 𝜙𝐸𝑓𝑓 = 𝜙𝑇 − 𝜙𝑠𝑉𝑠 .
(3.5)
Various methods of computing the effective mineral modulus are listed in Table 3.2.
Results of fluid substitution are shown in Figure 3.4 and Figure 3.5. Note that in both
figures all the lines labeled as “oil-model…” are fluid substitution predictions. The
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 57
“oil- shaly-sand” lines are constructed by connecting the V-points and the clean sand
points that are saturated with oil.
Table 3.1: Models used to generate elastic moduli of brine-saturated shaly-sand.
Case V-point
(𝝓𝑻 = 𝝓𝒔𝝓𝒔𝒉)
Models for shaly-sand lines
(mixing between V-point and clean
sand)
MHH VRH*
VRH
MHR VRH* Reuss
MRR Reuss**
Reuss
MLR HSLB***
Reuss *Voigt-Reuss-Hill average of quartz and wet-shale moduli **Reuss average of quartz and wet-shale moduli ***Hashin-Shtrikman lower bound of quartz and wet-shale moduli
Table 3.2: Models used to estimate effective solid moduli for fluid substitution.
Model # Models for effective mineral (mixing between quartz and wet
shale)
1 VRH
2 Reuss
3 HSLB
Figure 3.4 shows results when only P-wave moduli are used in all calculations
(assuming S-wave information is unavailable or unreliable). When the models used to
compute the V-point modulus and the effective mineral modulus for the Gassmann
shaly-sand equation (Dvorkin et al., 2007) are not the same, Dvorkin’s fluid
substitution (from brine to oil) near zero effective porosity can result in either
inconsistent or negative moduli.
First, we discuss cases where inconsistent results are observed. Since only fluids in
the effective porosity are replaced, at zero effective porosity we expect no fluid-
substitution effect. Thus, as effective porosity decreases and approaches zero, the
fluid-substituted P-wave compliance values should approach the P-wave compliance
of the wet-shaly sand at zero effective porosity. However, this is not the case when the
V-point modulus is larger than the effective mineral modulus used in the Gassmann
shaly-sand equation (Note that V-point modulus from VRH > HSLB > Reuss). For
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 58
example, in Figure 3.4(a) and (b), the V-point modulus is computed using the VRH
average which is larger than the modulus computed using either HSLB or the Reuss
average; as a result, after fluid substitution is applied, there are non-zero changes in
compliance at zero effective porosity (i.e., the cyan and red curves), creating
inconsistent results. This change in compliance at zero effective porosity after fluid
substitution is also observed in Figure 3.4(d) when HSLB is used to compute the V-
point modulus, and the Reuss average is used to compute the effective mineral
modulus. However, the changes in compliance at zero effective porosity in Figure
3.4(d) are smaller than those in Figure 3.4(a) and (b).
Second, we discuss cases where the fluid-substituted P-wave compliance near zero
effective porosity becomes negative and exhibits a mathematical singularity. This
implausible value is observed when the V-point modulus is smaller than the effective
mineral modulus used in the Gassmann shaly-sand equation. For example, in Figure
3.4(c), the V-point modulus is computed using the Reuss average which is the smallest
modulus among those computed using all the considered models. When either VRH or
HSLB is used to compute the effective mineral modulus, both negative compliance
and singularity can be observed (the green and cyan curves in Figure 3.4(c)).
When the same model is used to compute the V-point modulus and the effective
mineral modulus for the Gassmann shaly-sand equation, the resulting compliances
after fluid substitution are very similar to the compliances along the oil- shaly sand
lines.
All of the above observations are the results of the inconsistency test when only P-
wave moduli are used in the calculations (Figure 3.4). When both bulk and shear
moduli are used (Figure 3.5), the results are similar, except that the magnitude of the
inconsistency becomes smaller (Figure 3.5(a) and (b)).
In summary, choices of models used for constructing the shaly-sand lines, the V-
point, and for computing effective mineral moduli used in fluid substitution equations
can vary; however, consistency among models need to be taken into consideration to
ensure that fluid substitution results are consistent with all the assumptions (e.g., no
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 59
change in compliance at zero effective porosity) and that the results fall within a range
of plausible values (e.g., no singularity). Based on our inconsistency test results, when
applying fluid substitution using the Gassmann shaly-sand equation, the model for
computing the V-point modulus and the model for computing the effective mineral
modulus should be the same. In the next section, we introduce fluid substitution for
sub-resolution interbedded sand-shale sequences with the mesh, and we will see that
our mesh method is much less sensitive to the modeling choices for computing the V-
point.
Figure 3.4: Shaly sand lines before and after fluid substitution when only P-wave moduli are used in calculations. The model used to generate wet shaly sand is listed in the lower right corner ((a) – (d) and Table 3.1). Starting with the brine-saturated shaly sand (black dash-line), oil is substituted using the Gassmann shaly-sand equation (Dvorkin et al., 2007) and three different models for effective solid moduli (Table 3.2).
0 0.1 0.2 0.3 0.40
0.05
0.1
0.15
0.2
Effective porosity
1/M
(1/G
Pa)
0 0.1 0.2 0.3 0.40
0.05
0.1
0.15
0.2
Effective porosity
1/M
(1/G
Pa)
0 0.1 0.2 0.3 0.40
0.05
0.1
0.15
0.2
1/M
(1/G
Pa)
Effective porosity
0 0.1 0.2 0.3 0.40
0.05
0.1
0.15
0.2
Effective porosity
1/M
(1/G
Pa)
wet-shaly sand
oil-model1 (VRH)
oil-model2 (Reuss)
oil-model3 (HSLB)
oil-shaly sand
MHH MHR
MRR MLR
(a) (b)
(c) (d)
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 60
Figure 3.5: Shaly sand lines before and after fluid substitution when both bulk and shear moduli are used in calculations. The model used to generate wet shaly sand is listed in the lower right corner ((a) – (d) and Table 3.1). Starting with the brine-saturated shaly sand (black dash-line), oil is substituted using the Gassmann shaly-sand equation and three different models for effective solid moduli (Table 3.2).
3.4 Fluid substitution for interbedded sands
In Section 3.3.1, we have shown the diagram between P-compliance and effective
porosity for the dispersed and interbedded sand-shale system. In this section, we use
this diagram to show how to do approximate fluid substitution in shaly sands,
introduce a simple graphical interpretation, and derive equations for fluid substitution
in interbedded sands including sand with dispersed clay.
3.4.1 Approximate fluid substitution in shaly sands
Based on the results of the inconsistency test in Section 3.3.2, from this point on
we will only discuss cases where the same model is used for computing both the V-
point moduli and the effective mineral moduli. When the shaly-sand lines are modeled
0 0.1 0.2 0.3 0.40
0.05
0.1
0.15
0.2
1/M
(1/G
Pa)
Effective porosity
0 0.1 0.2 0.30
0.05
0.1
0.15
Effective porosity
1/M
(1/G
Pa)
0 0.1 0.2 0.3 0.40
0.05
0.1
0.15
0.2
Effective porosity
1/M
(1/G
Pa)
0 0.1 0.2 0.3 0.40
0.05
0.1
0.15
0.2
Effective porosity
1/M
(1/G
Pa)
wet-shaly sand
oil-model1 (VRH)
oil-model2 (Reuss
oil-model3 (HSLB)
oil-shaly sand
MHR
MRR MLR
(d)
(a) (b)
(c)
MHH
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 61
by the Reuss average, these lines become straight in the plot of P-compliance versus
effective porosity (Figure 3.4(b)-(d) and Figure 3.5(b)-(d)). If we apply the Gassmann
shaly-sand equation (Dvorkin et al., 2007) to each point along these (dispersed) shaly-
sand lines, the points with a new fluid form curves that are approximately straight
lines. In our synthetic example, we consistently use VRH to compute both the V-point
moduli and the effective mineral moduli for the Gassmann shaly-sand equation, and
we approximate that the shaly-sand line remains straight under fluid substitution.
Because of this linear approximation, applying fluid substitution to the entire shaly-
sand line can be done by following a few simple steps. First, fluid substitution (using
the traditional Gassmann’s equation) is applied to the clean sand point. Then this clean
sand point is moved to the new fluid modulus. Finally, the new shaly-sand line (with
the new fluid) is constructed by drawing a straight line connecting the new clean sand
point and the V-point (Figure 3.6). As a result of this linearity in fluid substitution of
the shaly-sand line, the changes in P-compliance (after fluid substitution) for sands
with different amounts of dispersed clay are directly proportional to the change in P-
compliance of the clean sand point.
When the linearity of the shaly-sand line is approximately preserved under fluid
substitution (from fluid 1 to fluid 2), it is equivalent to say that the Reuss average
(used for the shaly-sand line) is approximately consistent with the fluid substitution
equation. In other words, if we use the Reuss average to model the shaly-sand line
with an original fluid and apply fluid substitution to this line, we would obtain the
same results as if we were to model the shaly-line with the new fluid directly. The
same consistency with fluid substitution equations can be seen when using the
Raymer-Hunt-Garder models (Spike and Dvorkin, 2005).
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 62
Figure 3.6: Approximate fluid substitution for a dispersed shaly-sand line. Applying fluid substitution to this shaly-sand line is approximately equivalent to moving the clean sand point up or down, following the usual Gassmann’s equation, while the V-point at the other end of this line is fixed. Then, the new clean sand point is connected to the fixed V-point by another straight line. Here, we show an example when the clean sand point is moved down after fluid substitution.
3.4.2 Graphical interpretation and equation derivations for fluid substitution in
interbedded sands
Using results in Figure 3.6, here we introduce a simple graphical interpretation for
fluid substitution in interbedded sands. A sedimentary package of sand (or shaly sand)
interbedded with shale can be represented as a point inside the triangular diagram in
Figure 3.3. The effective porosity of the package is fully-saturated with an original
fluid. Instead of estimating the effect of fluid substitution on a point inside the diagram
one by one, we investigate how the whole triangular diagram is changed when the
original fluid in the effective porosity is replaced by a new fluid.
The triangular diagram in Figure 3.3 is composed of two main elements: the end-
points (i.e., the clean sand, V-point, and pure shale), and the dispersed lines (i.e., the
shaly-sand line connecting the clean sand to the V-point, and the sandy-shale line
connecting the V-point to the pure shale). The mesh is simply filled in by modeling
the interbedding of points along the shaly-sand line and the pure-shale point. Under
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.40.02
0.04
0.06
0.08
0.1
Effective porosity
1/M
(1/G
Pa)
Pure shaleClean sand
(fluid 1)
Clean sand
(fluid 2)
Increasing sand fractions
Shaly-sand
line
V point: sand completely
filled with clay
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 63
fluid substitution, the sandy-shale line, which extends from the pure shale and V-point,
remains unchanged due to their zero effective porosity. Therefore, both the pure shale
and V-point corners are fixed. In Section 3.4.1, we approximate that the shaly-sand
line remains straight and moves accordingly with the clean sand point under fluid
substitution. Since we know how the two main elements of the triangular diagram
change under fluid substitution to a new fluid, we can draw a new triangular diagram
and fill in the mesh. In summary, the change of the whole triangular diagram under
fluid substitution simply depends on how the clean sand point moves. Therefore,
applying approximate fluid substitution to any data point (e.g., either sand or shaly
sand interbedded with shale) inside the triangular diagram can be done by simply
distorting the mesh (Figure 3.7).
Figure 3.7: Graphical interpretation of fluid substitution by our mesh method. Applying approximate fluid substitution to any clean sand (or sand with dispersed clay) interbedded with shale is simply equivalent to distorting the mesh. Here the distortion is shown in the P-compliance (C=1/M) versus effective porosity plane. The distortion moves the mesh accordingly with the change in the clean sand compliance after fluid substitution. The blue arrow shows how a data point inside the triangular diagram moves after the distortion.
To derive the corresponding equations, we let an arbitrary point X in Figure 3.8
represent a sedimentary sequence where shaly sand is interbedded with shale. The
volume fraction of dispersed clay in the sand layers is equal to 𝑉𝑑𝑖𝑠𝑝 . We assume that
the starting effective porosity in shaly sand is fully saturated with fluid 1. In the
derivations, we use a set of variables defined as follows:
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 64
𝐶𝑠𝑎𝑡1 is the P-wave compliance of point X with its effective porosity
saturated with fluid 1
𝐶𝑠 is the P-wave compliance of shale
𝐶𝑑𝑖𝑠𝑝 𝑓𝑙𝑢𝑖𝑑 1 is the P-wave compliance of the shaly-sand layer whose
effective porosity is saturated with fluid 1
∆𝐶𝑠 and ∆𝐶𝑠𝑎𝑡 are the changes in the P-wave compliance for clean sand and
point X when fluid 1 is substituted by fluid 2, respectively.
Figure 3.8: Schematic diagram and terminology for the laminated (interbedded) sand-shale system.
With these definitions, we can write P-wave compliances of point X when
saturated with fluid 1 and 2 as follows:
𝐶𝑠𝑎𝑡1 = 𝑉𝑠𝑎𝑛𝑑 ∗ 𝐶𝑑𝑖𝑠𝑝 𝑓𝑙𝑢𝑖𝑑 1 − 1 − 𝑉𝑠𝑎𝑛𝑑 ∗ 𝐶𝑠 ,
(3.6)
𝐶𝑠𝑎𝑡2 = 𝑉𝑠𝑎𝑛𝑑 ∗ 𝐶𝑑𝑖𝑠𝑝 𝑓𝑙𝑢𝑖𝑑 2 − 1 − 𝑉𝑠𝑎𝑛𝑑 ∗ 𝐶𝑠 .
(3.7)
0Effective porosity
1/M
(1/G
Pa)
}
shale clean sand point
Csat1
Csat2
{ Point X
Cdisp
f luid1
}
Cdisp
Csand
Csat
Cdisp
fluid2
sandsand - vdisp
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 65
We are interested in the change in the compliances of point X when fluid 1 is
substituted fluid 2. This change can be expressed as
∆𝐶𝑠𝑎𝑡 = 𝐶𝑠𝑎𝑡1 − 𝐶𝑠𝑎𝑡2,
= 𝑉𝑠𝑎𝑛𝑑 ∗ ∆𝐶𝑑𝑖𝑠𝑝 (by substituting Equations 3.6 and 3.7).
(3.8)
From Figure 3.8, we have
∆𝐶𝑠/ ∆𝐶𝑑𝑖𝑠𝑝 = 𝜙𝑠/(𝜙𝑠 − 𝑉𝑑𝑖𝑠𝑝 ) ,
or ∆𝐶𝑑𝑖𝑠𝑝 = (𝜙𝑒𝑓𝑓 𝑠𝑎𝑛𝑑 /𝜙𝑠)∆𝐶𝑠 (by substituting 𝜙𝑒𝑓𝑓 𝑠𝑎𝑛𝑑 = 𝜙𝑠 − 𝑉𝑑𝑖𝑠𝑝 ),
(3.9)
where 𝜙𝑒𝑓𝑓 𝑠𝑎𝑛𝑑 is the effective porosity in the sand layers.
We substitute Equation 3.9 into Equation 3.8 to yield the approximate equation for
the change in the P-compliance after performing fluid substitution at point X:
∆𝐶𝑠𝑎𝑡 = 𝑉𝑠𝑎𝑛𝑑 ∗ (𝜙𝑒𝑓𝑓 𝑠𝑎𝑛𝑑 /𝜙𝑠)∆𝐶𝑠 ,
or ∆𝐶𝑠𝑎𝑡 = (𝜙𝐸𝑓𝑓 /𝜙𝑠)∆𝐶𝑠 ,
(3.10)
where 𝜙𝐸𝑓𝑓 is the effective porosity of the interbedded sand-shale sequence at point X.
Finally, we obtain an expression for compliance of the interbedded sand-shale
sequence at point X with new fluid (𝐶𝑠𝑎𝑡2):
𝐶𝑠𝑎𝑡2 = 𝐶𝑠𝑎𝑡1 − (𝜙𝐸𝑓𝑓 /𝜙𝑠)(𝐶𝑠𝑎𝑛𝑑 𝑓𝑙𝑢𝑖𝑑 1 − 𝐶𝑠𝑎𝑛𝑑 𝑓𝑙𝑢𝑖𝑑 2) .
(3.11)
In summary, the change in P-compliance after fluid substitution for an interbedded
sand-shale sequence is directly proportional to the change in P-compliance of the
clean sand point. The magnitude of this change is approximately equal to the ratio of
the effective porosity of the interbedded sand-shale sequence to the porosity of clean
sand. It is important to note that the starting fluid in the clean sand must be consistent
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 66
with the fluid in the effective pore space of the laminated (interbedded) sequence. For
example, let’s assume that a data point represents interbedding of 50% shale and 50%
shaly sand, whose effective pore space is only half saturated with oil. In this case, the
effective water saturation in the shaly sand layer is 0.5. If we want to apply fluid
substitution to this interbedded sequence by replacing oil with brine, we need to
compute the change in P-compliance of the clean sand point by going from (effective)
water saturation of 0.5 to water saturation of 1.
3.4.3 Important note
Our mesh method for fluid substitution in interbedded sand-shale sequences is
guaranteed to be robust for any rock constructed from the sand and shale end-members.
The method approximates changes in P-compliance for interbedded sand-shale
sequences to be proportional to the change in P-compliance of the clean sand point
after fluid substitution. The change of compliance at the clean sand point can be
estimated using the traditional Gassmann’s equation. Since clean (quartz-) sandstone
is a type of rock where Gassmann’s assumptions are most valid, applying fluid
substitution to this rock should result in a reasonable value. Based on our method,
applying fluid substitution to any interbedded sand-shale sequence is equivalent to
distorting the triangular diagram by moving only the clean sand’s corner, whose
change resulting from the traditional Gassmann’s equation is in a reasonable range
(Figure 3.8). Therefore, if a point X representing an interbedded sand-shale sequence
lies inside the triangular diagram, it is guaranteed that the fluid substitution result of
this point X is robust. Interestingly, the assurance for robust results holds regardless
of model choices for the V-point.
3.5 Synthetic examples
In this section we use synthetic examples to compare fluid substitution results
from our mesh method with three other procedures. We create synthetic examples by
first generating interbedded sand-shale sequences according to the specifications in
each case (e.g., clean sand interbedded with shale). Properties such as density, porosity
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 67
and velocity are assigned to each individual sedimentary layer. Then we upscale
volumetric and elastic properties using volumetric average and Backus average,
respectively. These upscaled values (i.e., the average properties of multiple layers)
simulate traditional well-logging measurements when the logging tools cannot resolve
individual layers.
We then apply fluid substitution (from oil to brine) to the interbedded sand-shale
sequences. First our mesh method is applied to the upscaled properties (i.e., the
equation is applied at the measurement scale). Then we apply three other fluid-
substitution procedures:
Gassmann’s equation applied to each sand layer, followed by upscaling
Gassmann shaly-sand equation (Dvorkin et al., 2007) applied to each sand
layer, followed by upscaling
Gassmann’s equation applied to upscaled properties (i.e., the equation is
applied at the measurement scale)
Note that the accurate way to perform fluid substitution in sub-resolution interbedded
sand-shale sequences is to first downscale the measurements for properties of sand and
shale end-members, apply Gassmann’s equation to the sand layers, and then upscale
the layers back to the measurement scale by Backus averaging. However, in real
applications of this downscaling-upscaling procedure, a number of parameters need to
be estimated, which can lead to non-robust and unreliable results. From the above list,
the first two procedures which apply fluid substitution in the sand layers only followed
by upscaling are equivalent to performing the downscaling-upscaling procedure, when
we know the actual properties of sand and shale end-members. These fluid substitution
results from the downscaling-upscaling procedure are considered our baselines.
3.5.1 Case 1: Shaly sand (sand with dispersed clay) with fully-oil-saturated
effective porosity (𝑺𝑾𝒆=0)
In case 1, we generate synthetic shaly sands with volume fractions of dispersed
clay in the sands ranging from 𝑉𝑑𝑖𝑠𝑝 = 0 to 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑠 . The starting effective pore
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 68
space in the shaly sands is fully-saturated with oil (𝑆𝑊𝑒=0). In this example, all clay is
dispersed, and thus there is no laminated (interbedded) shale. Oil is replaced by brine
using the four selected procedures. Results are shown in Figure 3.9.
Since the rocks in this example are simply shaly sands with no interbedding of
sand and shale, either applying Gassmann’s equation to the upscaled properties or
applying the equation to the sand layers followed by upscaling will give the same
results (the green dashed line and the black solid line in Figure 3.9). After fluid
substitution, the mesh method yields slightly higher velocities than those of the
Gassmann shaly-sand equation. However, results from both methods are similar.
Figure 3.9: Fluid substitution results from four different procedures for shaly sands with no interbedding. The procedures are Gassmann’s equation applied to sand layers only + upscaling, Gassmann shaly-sand equation (Dvorkin et al., 2007) applied to each sand layers only + upscaling, Gassmann’s equation applied at the measurement scale, and our mesh method. For this synthetic case, volume fractions of dispersed clay in the sand range from 𝑉𝑑𝑖𝑠𝑝 = 0 to 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑠.
0.1 0.15 0.2 0.25 0.3 0.35 0.4
2.6
2.8
3
3.2
3.4
3.6
3.8
4
4.2
4.4
Vp (
km
/s)
Total porosity
G. + upscaling
G. shaly-sand + upscaling
G. ignoring lamination
Mesh
Oil-saturated model
V-point (Vdisp
= clean sand porosity)
Clean sand (Vdisp
= 0)
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 69
3.5.2 Case 2: Clean sand interbedded with shale, varying 𝑽𝒔𝒂𝒏𝒅, and fully-oil-
saturated effective porosity (𝑺𝑾𝒆=0)
Figure 3.10: Fluid substitution results from four different procedures for interbedded clean sand-shale sequences. The procedures are Gassmann’s equation applied to sand layers only + upscaling, Gassmann shaly-sand equation (Dvorkin et al., 2007) applied to each sand layers only + upscaling, Gassmann’s equation applied at the measurement scale, and our mesh method. Note that results of Gassmann’s equation + upscaling, Gassmann shaly-sand equation + upscaling, and the mesh are on top of each other. The x-axis represents pseudo-depth.
Figure 3.11: Sand fractions of the synthetic model for interbedded clean sand-shale sequences. The x-axis represents pseudo-depth.
In case 2, we generate an interbedded (clean) sand-shale sequence. Sand fractions
vary with pseudo-depth (Figure 3.11). The starting effective pore space in the clean
sand is fully-saturated with oil. Oil is replaced by brine using the four selected
procedures. Results are shown in Figure 3.10. When both Gassmann’s equation and
the mesh method are applied at the measurement scale, Gassmann’s equation yields
higher velocities than the baselines, while the mesh yields exactly the same results as
the baselines. Thus, the mesh correctly accounts for the effect of sub-resolution sand-
shale interbedding, but Gassmann’s equation applied to the upscaled values
2.2
2.4
2.6
2.8V
p (
km
/s)
Depth
G. + upscaling
G. shaly-sand + upscaling
G. ignoring lamination
Mesh
Oil-saturated model
0
0.2
0.4
0.6
0.8
1
Volu
me f
raction o
f
sand in lam
ination
Depth
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 70
overpredicts the changes in velocities after fluid substitution. This overprediction
becomes smaller when sand fractions are either very small or very large (Figure 3.12).
For this example of clean sand interbedded with shale, there is no difference between
the traditional Gassmann and the Gassmann shaly-sand equations, because the sand
layers are clean (i.e., no dispersed clay).
Figure 3.12: Percentage differences between velocities after fluid substitution by Gassmann’s equation and Gassmann’s equation applied to only sand layers followed by upscaling. These differences are plotted against sand fraction and volume fraction of shale (𝑉𝑠 ). In this clean sand case, sand fraction is simply equivalent to 1 – 𝑉𝑠 .
3.5.3 Case 3: Shaly sand (𝑽𝒅𝒊𝒔𝒑= 0.15) interbedded with shale, varying 𝑽𝒔𝒂𝒏𝒅, and
fully-oil-saturated effective porosity (𝑺𝑾𝒆=0)
Figure 3.13: Fluid substitution results from four different procedures for interbedded shaly sand-shale sequences, with a fixed volume fraction of dispersed clay. The procedures are Gassmann’s equation applied to sand layers only + upscaling, Gassmann shaly-sand equation (Dvorkin et al., 2007) applied to each sand layers only + upscaling, Gassmann’s equation applied at the measurement scale, and our mesh method. The x-axis represents pseudo-depth.
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
9
10
11
Vshale, Sand fraction
%D
iff
:= 1
00*(
Vp G
. no lam
ination -
Vp G
.&up)/
(Vp G
.&up)
Vshale
Sand fraction
2.2
2.4
2.6
2.8
3
3.2
Vp (
km
/s)
Depth
G. + upscaling
G. shaly-sand+ upscaling
G. ignoring lamination
Mesh
Oil-saturated model
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 71
Figure 3.14: Sand fractions of the synthetic model for interbedded shaly sand-shale sequences, with a fixed volume fraction of dispersed clay. The x-axis represents pseudo- depth.
In case 3, we generate an interbedded (shaly) sand-shale sequence. The volume
fraction of dispersed clay in shaly sand is kept constant at 𝑉𝑑𝑖𝑠𝑝 =0.15. Sand fractions
vary with pseudo-depth (Figure 3.14). The starting effective pore space in the shaly
sand layers is fully-saturated with oil. Oil is replaced by brine using the four selected
procedures. Results are shown in Figure 3.13. Since our mesh method yields velocity
results very similar to the baselines, it accounts for both the interbedding part and the
shaly-sand (sand with dispersed clay) part during fluid substitution. Note that even
when effective water saturation changes to other values, we observe that results of the
mesh method agree well with baselines.
3.5.4 Case 4: Shaly sand (normal distribution of 𝑽𝒅𝒊𝒔𝒑 with a mean of 0.1 and a
standard deviation of 0.05) interbedded with shale, varying 𝑽𝒔𝒂𝒏𝒅, and fully-oil-
saturated effective porosity (𝑺𝑾𝒆=0)
Figure 3.15: Fluid substitution results from four different procedures for interbedded shaly sand-shale sequences, with varying volume fractions of dispersed clay. The procedures are Gassmann’s equation applied to sand layers only + upscaling, Gassmann shaly-sand equation (Dvorkin et al., 2007) applied to each sand layers only + upscaling, Gassmann’s equation applied at the measurement scale, and our mesh method. The x-axis represents pseudo-depth.
0
0.2
0.4
0.6
0.8
1
Volu
me f
raction o
f
sand in lam
ination
depth
2.2
2.4
2.6
2.8
3
3.2
Vp (
km
/s)
Depth
G. + upscaling
G.shaly-sand + upscaling
G. ignoring lamination
Mesh
Oil-saturated model
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 72
Figure 3.16: Sand fractions of the synthetic model for interbedded shaly sand-shale sequences, with varying volume fractions of dispersed clay. The x-axis represents pseudo-depth.
In case 4, we generate interbedding of sedimentary layers consisting of shaly sand
and shale. The volume fractions of dispersed clay assigned to the shaly-sand layers are
drawn from a normal distribution with a mean of 0.1 and a standard deviation of 0.05.
Sand fractions vary with pseudo-depth (Figure 3.16). The starting effective pore space
in the shaly sand layers is fully-saturated with oil. Oil is replaced by brine using the
four selected procedures. Results are shown in Figure 3.15.
Note that this case is more complicated than the previous cases because each
shaly-sand layer has different volume fractions of dispersed clay. If we follow the
normal routine of downscaling the measurements to the end-members’ properties
before applying fluid substitution, it is impossible in this case to extract all the end-
members from their average values alone. However, this is not a problem for the mesh
because we can apply our method directly to the measurements even when they
represent averages of multiple layers.
Although there are variations in properties of the interbedded sand layers (i.e.,
volume fractions of dispersed clay), the mesh still yields fluid substitution results
which agree well with those of the Gassmann shaly-sand equation (or Gassmann’s
equation) applied to each sand layers only followed by upscaling (Figure 3.15).
0
0.2
0.4
0.6
0.8
1
Volu
me f
raction o
f
sand in lam
ination
depth
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 73
3.5.5 Case 5: Interbedded sand-shale sequences with systematic changes of both
the 𝑽𝒅𝒊𝒔𝒑 in the sand layers and the 𝑽𝒔𝒂𝒏𝒅
In case 5, we generate interbedding of sedimentary layers between sand and shale.
We systematically vary volume fractions of dispersed clay in the sand layers (i.e.,
from 𝑉𝑑𝑖𝑠𝑝 = 0 to 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑆) for a constant sand fraction. Then we repeat the same
process for other sand fractions. As a result, we roughly cover all possible
interbedding scenarios allowed by the triangular diagram (Figure 3.3). The starting
effective pore space in the sand or shaly sand layers is fully-saturated with oil (𝑆𝑊𝑒=0).
Oil is then replaced by brine using the four selected procedures. Results are shown in
Figure 3.17.
Figure 3.17: Changes in P-wave velocity after fluid substitution from four different procedures: Gassmann’s equation applied to sand layers only + upscaling, Gassmann shaly-sand equation (Dvorkin et al., 2007) applied to sand layers only + upscaling, Gassmann’s equation applied at the measurement scale, and our mesh method. Each location on the triangular diagram represents a interbedded sand-shale sequence with a unique pair of sand fraction and volume fraction of dispersed clay values.
0 0.1 0.2 0.3
0.04
0.06
0.08
Effective porosity
1/M
(1/G
Pa)
Gassmann ignoring lamination
0
50
100
150
200
250
0 0.1 0.2 0.3
0.04
0.06
0.08
Effective porosity
1/M
(1/G
Pa)
Gassmann + upscaling
0
50
100
150
200
250
0 0.1 0.2 0.3
0.04
0.06
0.08
Effective porosity
1/M
(1/G
Pa)
Gassmann shaly-sand equation
+ upscaling
0
50
100
150
200
250
0 0.1 0.2 0.3
0.04
0.06
0.08
Effective porosity
1/M
(1/G
Pa)
Mesh
0
50
100
150
200
250
change in
velocity (m/s)
change in
velocity (m/s)
change in
velocity (m/s)change in
velocity (m/s)
Velocity changes after fluid substitution (from oil to brine) using
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 74
By applying Gassmann’s equation at the measurement scale and thus ignoring the
effect of thin sand-shale interbedding, the changes in velocity after fluid substitution
(from oil to brine) are much larger than the changes from the other three procedures.
These large velocity changes are reflected in a distinct color pattern in the triangular
diagram (Figure 3.17, upper left corner). Our mesh method yields fluid substitution
results similar to those obtained by applying fluid substitution to the shaly-sand layers
only, followed by upscaling.
3.6 Pitfalls in interpretation
We show two examples of pitfalls in interpretation of fluid substitution results
when sub-resolution interbedding exists. First, we generate a synthetic dataset that
represents interbedding of shaly-sand and shale, with sand fractions ranging between
𝑉𝑠𝑎𝑛𝑑 = 0.5 and 𝑉𝑠𝑎𝑛𝑑 = 1, and volume fractions of dispersed clay ranging between
𝑉𝑑𝑖𝑠𝑝 =0 and 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑆. Effective porosity is fully-saturated with gas. We perform
fluid substitution (gas to brine) using the mesh and Gassmann’s equation (ignoring
interbedding).
In the first example, suppose that thin interbedding is not accounted for. Then
changes in velocities after fluid substitution are overpredicted. Because of this
overprediction, data with small sand fractions can look as if they have larger sand
fractions. Figure 3.18 (Right) shows fluid substitution results when the effect of thin
sand-shale interbedding is ignored. In this panel, all data points are located between
the lines of 𝑉𝑠𝑎𝑛𝑑 = 0.75 and 1. However, the actual sand fractions range from 𝑉𝑠𝑎𝑛𝑑 =
0.5 to 𝑉𝑠𝑎𝑛𝑑 = 1. Thus, due to inappropriate fluid substitution the overpredicted
velocities move the data points to higher positions, where they can be misinterpreted
as having sand fractions higher than the actual values. In contrast, our mesh approach
preserves the sand-fraction properties in the process of fluid substitution (Figure 3.18,
middle).
In the second example, suppose that an interbedded sand-shale sequence with a
small sand fraction is saturated with gas. The change in velocities after fluid
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 75
substitution from gas to brine is generally expected to be high. However, because the
sequence has low sand fraction, its velocity change after fluid substitution is small,
which can be misinterpreted as if the fluid changes from oil to brine, as illustrated in
Figure 3.18 (middle).
Figure 3.18: Pitfalls in interpretation of fluid substitution results when thin interbedding exists. (Left) Synthetic gas-saturated data with sand fractions greater than or equal to 0.5. (Middle) Results after fluid substitution from gas to brine using our mesh approach. (Right) Results after fluid substitution from gas to brine using Gassmann’s equation. Data points are color-coded by water saturation values before fluid substitution.
3.7 Sensitivity analysis
In the synthetic examples, when performing fluid substitution using the
downscaling-upscaling procedures, we use all correct parameters. However, in
practice the downscaling process involves an inversion or estimations for volume
fractions and properties of the end-members (e.g., shale) in the interbedding, possibly
leading to non-robust results. In contrast, our mesh method guarantees robustness. The
method also does not depend on the shale elastic property. It only requires the
estimations of effective porosity, effective saturation, fluid properties, and the clean
sand porosity and velocity.
In this section, we show how uncertainties in input parameters affect fluid
substitution results of our mesh method (i.e., using Equation 3.11). The five input
parameters considered are clean sand’s P-wave velocity (𝑉𝑝𝑠𝑎𝑛𝑑 ), clean sand porosity
0.1 0.2 0.3 0.41.5
2
2.5
3
3.5
4
4.5
Shale Sand
Total porosity
Vp (
km
/s)
Sand
original water saturation
before fluid substitution
0.2
0.4
0.6
0.8
0.1 0.2 0.3 0.41.5
2
2.5
3
3.5
4
4.5
Shale Sand
Total porosity
Vp (
km
/s)
Sand
0.2
0.4
0.6
0.8
0.1 0.2 0.3 0.41.5
2
2.5
3
3.5
4
4.5
Shale Sand
Total porosity
Vp (
km
/s)
Sand
original water saturation
before fluid substitution
0.2
0.4
0.6
0.8
GAS BRINE - MESH BRINE –GASSMANN
Sand fraction = 0.5 Sand fraction = 0.75
Color-coded by original Sw before fluid substitution
Vsand = 0.5 Vsand = 0.75
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 76
(𝜙𝑠), effective porosity of the laminate package (𝜙𝐸𝑓𝑓 ), effective water saturation in
the sand layers (𝑆𝑊𝑒 ), and fluid properties (i.e., bulk modulus, 𝐾𝑓𝑙 ) .
To run this analysis, we generate synthetic stacks of interbedded sand-shale
sequences, where the sand can be either clean or dirty (i.e., dispersed clay exists in the
sand pore space). We explore cases where the sand fractions range from 𝑉𝑠𝑎𝑛𝑑 = 0 to
𝑉𝑠𝑎𝑛𝑑 = 1, and the volume fractions of dispersed clay range from 𝑉𝑑𝑖𝑠𝑝 = 0 to 𝑉𝑑𝑖𝑠𝑝 =
𝜙𝑆. The starting effective pore space is half filled with oil. Then, we upscale properties
of multiple layers to mimic well logging measurements. All the four procedures
described in Section 3.5 are applied to the synthetic measurements to substitute oil in
the effective porosity with brine, using all correct parameters. The results from these
procedures are used as references.
In real data application, we would not know the true parameters; instead, we would
only have distributions of possible values. We analyze the effect of these uncertainties
in the input parameters on fluid-substitution results, one parameter at a time. For each
input parameter, we randomly draw a value from its assigned distribution while
keeping all the other parameters at their true values, use these parameters in our
method, and keep the result. The process is repeated multiple times. Finally, we have a
distribution of outcomes resulting from uncertainties in that one particular parameter.
We repeat the same process for all parameters on the list.
We assume that all the input parameters previously listed are independent and
normally distributed with means (𝜇) and variances 𝜎2 . The means of the distributions
are set to be the true values used in the forward-modeling process. The standard
deviations 𝜎 are set to be 10% of the means. We test the effect of uncertainties in the
input parameters one by one. Note that a similar Monte Carlo simulation was used in
Artola and Alvarado (2006) for sensitivity analyses of Gassmann’s equations.
Results of the sensitivity analysis for the mesh method are shown in Figure 3.19 as
distributions of velocities after fluid substitution. Each subplot is indicated by a
volume fraction of laminated (interbedded) sand and a volume fraction of dispersed
clay in the sand layers. The thick lines on the top of each subplot are the oil-saturated
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 77
velocity and the reference velocities after fluid substitution from oil to brine using the
four procedures with all correct parameters.
For specific values of sand fraction and volume fraction of dispersed clay, all
distributions of velocities after fluid substitution using our method are generally
centered around the thick magenta-dashed line, which is the reference result using our
method with all correct input parameters. This centering is expected since we set the
correct values as the means of distributions for the input parameters that we want to
test.
For specific values of sand fraction and volume fraction of dispersed clay,
uncertainties in the clean sand velocity result in the largest spread of velocities after
fluid substitution using our method. Thus, our method is most sensitive to
uncertainties in the clean sand velocity. The subsequent parameters, in decreasing
order of sensitivity, are the clean sand porosity, effective porosity, effective water
saturation and fluid elastic modulus.
For the results of each tested parameter, when a volume fraction of laminated
(interbedded) sand increases with a fixed volume fraction of dispersed clay, the spread
of velocities after fluid substitution using our method becomes larger.
For the results of each tested parameter, when a volume fraction of dispersed clay
in the sand layers increases with a fixed volume fraction of laminated (interbedded)
sand, the spread of velocities after fluid substitution using our method becomes
smaller.
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 78
Figure 3.19: Sensitivity analysis results of the mesh method for five input parameters: P-wave velocity and total porosity of clean sand, effective porosity and effective water saturation of the interbedded package, and elastic modulus of fluid. The thick lines on the top of each subplot are the oil-saturated velocity and the reference velocities after fluid substitution from oil to brine using four procedures with all correct input parameters: Gassmann ignoring interbedding, Gassmann applied to each sand layer + upscaling, Gassmann shaly-sand equation (Dvorkin et al., 2007) applied to each sand layer + upscaling, and our mesh method. Results of sensitivity analysis for each parameter are shown as a velocity distribution, which is normalized to one. Sand fractions and volume fractions of dispersed clay for each synthetic model are shown on above each subplot.
3.8 Real data example
In this section, we show fluid substitution results from real log data. We compare
four procedures: Gassmann’s equation (i.e., ignoring the effect of thin sand-shale
interbedding), Gassmann shaly-sand equation (i.e., assuming that shale is all dispersed;
Dvorkin et al., 2007), the downscaling-upscaling procedure, and our mesh method.
The downscaling-upscaling procedure here refers to the most appropriate
procedure for fluid substitution in thinly interbedded sand-shale sequences, given that
all parameters are correct. The procedure includes steps as follows:
2.2 2.4 2.6 2.8 3 3.2 3.40
1
Sand fraction = 0.33 Vdis = 0
2.2 2.4 2.6 2.8 3 3.2 3.40
1
Sand fraction = 0.33 Vdis = 0.12
2.2 2.4 2.6 2.8 3 3.2 3.40
1
Sand fraction = 0.33 Vdis = 0.24
2.2 2.4 2.6 2.8 3 3.2 3.40
1
Sand fraction = 0.67 Vdis = 0
Norm
aliz
ed d
ensity f
unction
2.2 2.4 2.6 2.8 3 3.2 3.40
1
Sand fraction = 0.67 Vdis = 0.12
2.2 2.4 2.6 2.8 3 3.2 3.40
1
Sand fraction = 0.67 Vdis = 0.24
2.2 2.4 2.6 2.8 3 3.2 3.40
1
Sand fraction = 1 Vdis = 0
P-velocity (km/s)2.2 2.4 2.6 2.8 3 3.2 3.4
0
1
Sand fraction = 1 Vdis = 0.12
P-velocity (km/s)
2.2 2.4 2.6 2.8 3 3.2 3.40
1
Sand fraction = 1 Vdis = 0.24
P-velocity (km/s)
V
p sand
Sand porosity
Effective porosity
Sw e
Koil
G. ignoring lamination
G.+upscaling
G. shaly-sand +upscaling
Mesh
Oil-saturated model
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 79
1. Determine parameters including P-wave velocity, density, and total porosity of
the shale end-member, volume fraction of laminated (interbedded) shale, and
volume fraction of dispersed clay present in the sand layers. These end-
member parameters can be estimated by cross-plotting measurements with
superimposed theoretical model curves. For example, the Thomas-Stieber
triangular diagram (Chapter 2) in the total-porosity and shale-volume-fraction
plane helps break down volume fraction of laminated (interbedded) shale from
dispersed clay. Note that the Thomas-Stieber model also requires some
additional input parameters (e.g., clean sand porosity).
2. Downscale for the elastic properties of the sandy end-member by inverting the
Backus average in Equation 3.4 for the inverse P-wave modulus of the sandy
layers.
From 𝑀𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑−1 = (1 − 𝑉𝑙𝑎𝑚 )𝑀𝑠𝑎𝑛𝑑 (𝑜𝑟 𝑠𝑎𝑙𝑦 𝑠𝑎𝑛𝑑 )
−1 + 𝑉𝑙𝑎𝑚𝑀𝑠−1 , we can write
𝑀𝑠𝑎𝑛𝑑 (𝑜𝑟 𝑠𝑎𝑙𝑦 𝑠𝑎𝑛𝑑 )−1 = (𝑀𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
−1 − 𝑉𝑙𝑎𝑚𝑀𝑠−1)/(1 − 𝑉𝑙𝑎𝑚 ),
(3.12)
where 𝑀𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑−1 is the inverse P-wave modulus of the measurements. Using
equation 3.12 requires estimates of 𝑀𝑠 .
3. Apply fluid substitution to the sandy end-member using Gassmann shaly-sand
equation (Dvorkin et al., 2007).
4. Upscale the sandy and shale layers back using Equation 3.4. At the end of this
step, only the sandy layers are saturated with the new fluid.
Even though all steps in the downscaling-upscaling procedure are straightforward,
fluid substitution results from this procedure can be unstable and non-robust in
practice due to uncertainties in real measurements and errors in parameter estimation.
When these uncertainties and errors are passed on through the inversion and the fluid
substitution steps, unreliable results may be observed (e.g., negative values, velocity
spikes).
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 80
In contrast, even though the mesh method also requires parameter estimation steps,
it skips the actual Backus downscaling-upscaling steps. Since the mesh method is
based on the Thomas-Stieber-Yin-Marion model which describes sand-shale mixtures,
the mesh implicitly performs downscaling and upscaling without any actual
computation. For the same reason, we do not have to actually compute volume
fraction of laminated (interbedded) shale and dispersed clay as this shale distribution
information is embedded into the mesh method. As a result, the mesh is more robust
than the downscaling-upscaling process in practice. One could say that the mesh
method is an upscaled version of Gassmann for their type of heterogeneous system.
The well log data comes from a sand-shale system. A section representing wet
sand interbedded with shale is selected and shown in Figure 3.20. The Thomas-
Stieber-Yin-Marion model is superimposed onto the cross-plot to estimate the volume
fractions of laminated (interbedded) shale and dispersed clay, which will be used later
in the downscaling-upscaling procedure.
Starting with all brine-saturated data, we substitute oil for brine using four
procedures: Gassmann’s equation (i.e., ignoring the effect of sub-resolution sand-shale
interbedding), Gassmann shaly-sand equation (i.e., assuming that shale is all
dispersed), the downscaling-upscaling procedure, and our mesh method. Note that
here both bulk and shear moduli are used in all calculations. Since neither the
Gassmann’s equation nor the Gassmann shaly-sand equation correctly account for
sub-resolution interbedding, we expect that using these procedures the changes in P-
wave velocities after fluid substitution are overpredicted with respect to the true
velocity change. For example, if fluid substitution from brine to oil decreases P-wave
velocities, an overpredicted result will give a lower velocity value. Since the
downscaling-upscaling procedure is the appropriate way to perform fluid substitution
in interbedded sand-shale sequences, we use results from this procedure as baselines.
However, these results show some velocity spikes, especially at low sand-fraction
points. Thus, to avoid any erratic behavior, the baselines from the downscaling-
upscaling procedure are chosen to include only those points with gamma ray values
less than 85.
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 81
Figure 3.20: Selected dataset represents an interbedded sand-shale sequence. The data
is color-coded by Gamma ray values. The Thomas-Stieber-Yin-Marion model is superimposed onto the data. The set of parallel lines labeled as 𝑉𝑠𝑎𝑛𝑑 represent volume fractions of laminated (interbedded) sand according to the Thomas-Stieber-Yin-Marion model.
Figure 3.21: Comparisons of fluid substitution results which are color-coded by sand
fractions estimated from the Thomas-Stieber-Yin-Marion model. Four fluid substitution procedures are used: Gassmann’s equation (i.e., ignoring the effect of sand-shale interbedding), Gassmann shaly-sand equation (i.e., assuming that shale is all dispersed; Dvorkin et al., 2007), the downscaling-upscaling procedure, and the mesh method. Note that points with very low sand fractions are excluded from the plot. Lines X = Y are super-imposed for comparison purpose.
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Shale
Sand
Vshale
Tota
l poro
sity
Gamma ray
20
30
40
50
60
70
80
90
100
Vsand
= 1
Vsand
= 0.75
Vsand
= 0.5
Vsand
= 0.25
Vsand
= 0
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
2.8
Vp downscaling-upscaling
(km/s)
Vp m
esh (
km
/s)
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
2.8
Vp downscaling-upscaling
(km/s)
Vp G
assm
ann s
haly
-sand
(k
m/s
)
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
2.8
Vp downscaling-upscaling
(km/s)
Vp G
assm
ann (
km
/s)
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
2.8
Vp downscaling-upscaling
(km/s)
Vp d
ow
nscalin
g-u
pscalin
g
(k
m/s
)
sand fraction
0.2
0.4
0.6
0.8
1
Line X = Y
X > Y region
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 82
P-wave velocities after fluid substitution from the downscaling-upscaling
procedure (i.e., baselines) are compared with results from other procedures (Figure
3.21). The data points in Figure 3.21 are color-coded by sand fractions (𝑉𝑠𝑎𝑛𝑑 )
estimated from the Thomas-Stieber-Yin-Marion model in Figure 3.20. Black lines (X
= Y) are super-imposed onto the plots for comparison purpose. For clean sand or very
high sand-fractions data, we see similar P-wave velocities after fluid substitution
regardless of procedures used, because these data represent no or almost no sub-
resolution interbedding. For data with intermediate sand fractions, P-wave velocities
after fluid substitution using both the Gassmann’s equation and the Gassmann shaly-
sand equation are lower than velocity results from the downscaling-upscaling
procedure, because Gassmann’s equation do not account for sub-resolution
interbedding. These lower velocity values (i.e., larger changes in velocities during
fluid substitution) are clearly seen in the top row of Figure 3.21 as data points deviate
from the lines X = Y and stay in the X > Y region.
P-wave velocities after fluid substitution using the mesh method are much closer
to those of the downscaling-upscaling procedure; however, their values are not
perfectly the same (Figure 3.21, lower left corner). One possible explanation for their
small mismatches is inconsistency of sand fractions used in the downscaling-upscaling
process. We use the Thomas-Stieber-Yin-Marion model, specifically in the volumetric
(i.e., total porosity VS volume fraction of shale) space, to estimate sand fractions.
Then, we input these values into Equation 3.12 to invert for the P-compliance of sandy
layers. This inversion step operates in the elastic (i.e., velocity) space. The sand
fractions estimated from the volumetric space may not be consistent with those from
the elastic space. Since the mesh method internally operates in the elastic space, the
downscaling-upscaling procedure, that uses (potentially) different sand fractions
estimated from the volumetric space, can cause a mismatch between fluid-substitution
results of the two methods. To solve this problem, we should select a set of end-points
(e.g., clean-sand porosity, velocity) such that sand-fraction values are as consistent as
possible across all spaces. An example of fluid-substitution results when sand-
fractions values are estimated in the elastic space are shown in Figure 3.22. Note that
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 83
the color scale is the same in both Figure 3.21 and Figure 3.22. Using sand fractions
estimated from the elastic space, we achieve a slightly better match of P-wave
velocities after fluid substitution using the mesh method and the downscaling-
upscaling procedure.
Figure 3.22: Comparisons of fluid substitution results which are color-coded by sand
fractions estimated from the P-compliance versus effective porosity plane. Four fluid substitution procedures are used: Gassmann’s equation (i.e., ignoring the effect of sand-shale interbedding), Gassmann shaly-sand equation (i.e., assuming that shale is all dispersed; Dvorkin et al., 2007), the downscaling-upscaling procedure, and the mesh method. Note that points with very low sand fractions are excluded from the plot. Lines X = Y are super-imposed for comparison purpose.
As noted earlier in Section 3.8, one drawback of the downscaling-upscaling
procedure is non-robustness during the Backus-inversion step, potentially resulting in
velocity spikes and/or negative velocity. Using the downscaling-upscaling procedure,
we observe P-wave velocity spikes after applying fluid substitution to the well log data
(Figure 3.23). In contrast, velocity spikes are not observed in the fluid substitution
results from the mesh method.
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
2.8
Vp downscaling-upscaling
(km/s)
Vp m
esh (
km
/s)
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
2.8
Vp downscaling-upscaling
(km/s) V
p G
assm
ann s
haly
-sand
(k
m/s
)
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
2.8
Vp downscaling-upscaling
(km/s)
Vp G
assm
ann (
km
/s)
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
2.8
Vp downscaling-upscaling
(km/s)
Vp d
ow
nscalin
g-u
pscalin
g
(km
/s)
Sand fraction
0.2
0.4
0.6
0.8
1
Line X = Y
X > Y region
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 84
Figure 3.23: Instability of the inversion step in the downscaling-upscaling procedure leading to spikes in P-wave velocity after fluid substitution.
3.9 Discussion
3.9.1 The mesh method using both bulk and shear moduli
Most of results shown in this chapter are computed using P-wave moduli only.
However, the method is also applicable if both bulk and shear moduli are used. This
applicability is supported by results and discussion from the consistency test in
Section 3.3.2 (Figure 3.5). For example, we rerun the synthetic shaly-sand case
(Section 3.5.1) using both bulk and shear moduli in all fluid-substitution procedures.
The mesh results are still in good agreement with other procedures (Figure 3.24).
1 1.5 2 2.5 3
P-wave velocity after fluid substitution (km/s)
Depth
Downscaling-upscaling
process
the mesh
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 85
Figure 3.24: Fluid substitution results by four different procedures for shaly sands, with no interbedding. The procedures are Gassmann using total porosity and ignoring interbedding, Gassmann using total porosity (applied to sand layers only) + upscaling, Gassmann using effective porosity by Dvorkin et al., 2007 (applied to each sand layers only) + upscaling and our method (i.e., the mesh). For this synthetic case, volume fractions of dispersed clay in the sand range from 𝑉𝑑𝑖𝑠𝑝 =0 to 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑠. Both bulk and shear moduli are used for this synthetic example.
3.9.2 Possible modification of the mesh method when key assumptions are
relaxed
Our mesh method is based on two key assumptions. First, the shaly-sand line is
assumed to remain straight under fluid substitution. This assumption is only an
approximation of the actual fluid substitution results, which show that the shaly-sand
(straight) line undergoing fluid substitution turns into a non-linear curve (Section
3.4.1). If a non-linear functional form is chosen, we need to also adjust equation
derivations for fluid substitution in interbedded sand-shale sequences. Since all the
derivations are done in the compliance-effective porosity space, interbedding of any
two end-members in this space are simply equivalent to weighted averages of their
compliance and effective porosity, regardless of whether the shaly-sand line is straight
or not. This computational advantage in the compliance-effective porosity space
makes equation derivations easy even for non-linear shaly-sand curves.
The second key assumption of our mesh method is that fluid substitution is
assumed to occur only in the sand, but not in the shale (i.e., fluid substitution occurs in
0.1 0.15 0.2 0.25 0.3 0.35 0.4
2.6
2.8
3
3.2
3.4
3.6
3.8
4
4.2
4.4
Vp (
km
/s)
Total porosity
G. + upscaling
G. (eff) + upscaling
G. ignoring lamination
Mesh
Oil-saturated model
V-point (Vdisp = clean sand porosity)
Clean sand (Vdisp = 0)
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 86
effective porosity). As a result, performing fluid substitution using our method is
equivalent to distorting the mesh according to how the clean sand end-point moves,
while the shale end-point is fixed (i.e., no change in elastic property). However, it is
possible to modify our method to account for fluid substitution in shale by simply
moving the shale end-point according to the change in its elastic property under fluid
substitution, which may be computed using Gassmann’s equation (Lucier et al., 2011)
or using effective medium models (Ruiz and Azizov, 2011). If fluid substitution is
allowed in the shale layers, it is necessary to model fluid substitution also in the
dispersed clay (i.e., the shaly-sand line) for consistency. Once we are able to
approximate the elastic change of the shaly-sand line, we can modify the mesh model
accordingly by following a procedure similar to that shown in Section 3.4.2. As
previously mentioned, since all the derivations are done in the compliance-effective
porosity space, the modification of our mesh method to account for fluid substitution
in shale should not be difficult.
3.9.3 Using rock-physics trends to constrain clean-sand properties
The key parameters needed for the mesh method are the velocity and porosity of
the clean sand point. Since the results from our sensitivity analysis showed that the
mesh method is most sensitive to the clean-sand velocity (relative to the other
parameters tested), clean-sand properties must be chosen carefully. We can estimate
such properties from data with small gamma ray values as they represent cleaner sand
sections. Additionally, we can use the clean-sand diagenetic trend to constrain the
estimation of the clean-sand properties. The diagenetic trend describes the relationship
between the clean-sand velocity and porosity while clean sandstones undergoing
diagenetic processes (Figure 3.25). These processes (e.g., compaction, cementation)
cause porosity reduction, creating steep velocity-porosity curves often seen in
sandstone data which covers a wide depth range. These trends can be modeled using
the modified Hashin-Shtrikman upper bound (or the modified Voigt bound) which
connects the newly deposited sediments at their critical porosities with the
corresponding mineral points (Nur et al., 1998; Avseth et al., 2005). The clean sand
trend line also establishes a guideline for selecting clean-sand properties along depth.
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 87
Figure 3.25: Modified upper bounds for modeling the clean-sand diagenetic trend. This example represents a quartz-water system.
3.9.4 Limitation of the mesh method
Since our mesh method is the elastic analog of the Thomas-Stieber-Yin-Marion
model, our method also carries over one important assumption from that model, which
states that all the sand-shale mixtures (e.g., sand interbedded with shale, or sand with
dispersed clay) are generated by mixing two rock types: the clean sand and pure shale.
When shale is dispersedly added into the pore space of sand, the sand becomes shaly
and its total porosity decreases. The Thomas-Stieber-Yin-Marion model assumes that
the only cause of any porosity reduction is the filling of shale in the pore space. The
model does not account for porosity reduction caused by cementation or changes in
grain sorting (Thomas and Stieber, 1975; Ball et al., 2004; Mavko et al., 2009). This
assumption is therefore passed to our mesh method. As a result, the mesh method does
not properly account for the effects of cementation or grain sorting.
3.9.5 Comparison with alternative methods
The problem of fluid substitution for thin interbedded sand-shale sequences is not
new. In practice, this sub-resolution interbedding problem in fluid substitution is either
completely ignored, or is handled by one of the other fluid substitution methods. When
sub-resolution interbedding is not accounted for, fluid substitution results are often
mispredicted (e.g., overestimation when going from oil- to brine-saturated
measurements) as shown in our synthetic examples (Section 3.5) and literature (e.g.,
Skelt, 2004a). Each fluid substitution method has specific assumptions, and it requires
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
total porosityV
p (
km
/s)
critical porosity
clean
sand
trend
Modified
Voigt bound
Reuss bound
Modified upper
Hashin-Shtrikman
bound
Voigt bound
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 88
a set of parameters and possibly other evidence of sub-resolution interbedding. We
want to point out that this chapter provides the mesh method as another alternative to
deal with this fluid substitution problem in interbedded sand-shale sequence. The
mesh method is a new invention that integrates existing models to handle both
interbedded and dispersed sand-shale systems at the same time. The method is based
on assumptions and requires a set of estimated parameters. However, it reduces some
additional procedures needed to be done in other methods. We summarize different
methods used in fluid substitution and comment on the performance of each method
when it is applied to thinly interbedded sand-shale sequences (Table 3.3).
Table 3.3: Summary of methods usually used in fluid substitution and comments for when these methods are applied to sub-resolution interbedded sand-shale sequences.
Method Inputs/ parameters2 Comments
Gassmann’s equation
(using total porosity) 𝜙𝑇 ,𝑉𝑝 ,𝑉𝑠 ,𝜌, 𝑆𝑊 ,
𝐾𝑚𝑖𝑛 ,𝐾𝑓𝑙 , 𝜌𝑓𝑙
Ignoring interbedding and shale
effect
(inappropriate procedure)
Gassmann shaly-sand
equation by Dovrkin et
al., 2007 (using
effective porosity and
newly- defined
composite mineral)
𝜙𝐸𝑓𝑓 ,𝜙𝑇 ,𝑉𝑝 ,𝑉𝑠 ,
𝜌, 𝑆𝑊𝑒 ,𝑉𝑠
𝐾𝑠𝑎𝑙𝑒 ,𝜙𝑠 ,𝐾𝑚𝑖𝑛 , 𝐾𝑓𝑙 , 𝜌𝑓𝑙
Ignoring interbedding
Assuming all shale is dispersed
Downscaling-
upscaling procedure 𝜙𝐸𝑓𝑓 ,𝑉𝑝 ,𝑉𝑠 , 𝜌, 𝑆𝑊𝑒 ,
𝑉𝑙𝑎𝑚 ,𝑉𝑑𝑖𝑠𝑝 ,𝑀𝑠𝑎𝑙𝑒 ,
𝐾𝑓𝑙 , 𝜌𝑓𝑙
Correct if inputs are correct
Non-robust procedure
Mesh method 𝜙𝐸𝑓𝑓 ,𝑉𝑝 ,𝑉𝑠 , 𝜌, 𝑆𝑊𝑒 ,
𝑀𝑠𝑎𝑛𝑑 ,𝜙𝑠 ,𝐾𝑓𝑙 , 𝜌𝑓𝑙
Correct if inputs are correct
2 𝜙𝑇 ,𝜙𝐸𝑓𝑓 ,𝑉𝑝 ,𝑉𝑠 ,𝜌, 𝑆𝑊 , 𝑆𝑊𝑒 ,𝑉𝑠 : Total porosity, effective porosity, P-wave velocity, S-wave velocity,
density, water saturation, effective water saturation, and shale volume, respectively. These inputs are
measured data, or they can be derived from other measurements.
𝐾𝑚𝑖𝑛 ,𝐾𝑓𝑙 ,𝜌𝑓𝑙 : Mineral bulk modulus, and bulk moduli and densities of fluids, and respectively.
𝑀𝑠𝑎𝑙𝑒 ,𝐾𝑠𝑎𝑙𝑒 ,𝜙𝑠 : P-wave modulus, bulk modulus, and total porosity of the shale end point,
respectively
𝑀𝑠𝑎𝑛𝑑 ,𝜙𝑠: P-wave modulus and total porosity of the clean sand end point, respectively.
𝑉𝑙𝑎𝑚 ,𝑉𝑑𝑖𝑠𝑝 : Volume of laminated (interbedded) shale and volume of dispersed clay, respectively.
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 89
3.9.6 Upscaled Gassmann’s equations
From both synthetic and real data examples in Sections 3.5 and 3.8, when applying
Gassmann’s equation (Table 3.3) directly to sub-resolution interbedded sand-shale
sequences at the measurement scale, the effect of interbedding is not correctly
accounted for, resulting in erroneous predictions. In this section, we show that it is
possible to apply Gassmann’s equation at the measurement scale and account for the
effect of sand-shale interbedding at the same time by simply adjusting several input
parameters in Gassmann’s equation. We refer to this adjustment as an upscaled
Gassmann’s equation for interbedded sand-shale sequences.
We algebraically determine the required input adjustment for the case where clean
sand is interbedded with shale. Note that all the derivations will be done using
approximate Gassmann’s equation, which can be expressed in terms of P-wave moduli
(Mavko et al., 1995):
𝑀𝑠𝑎𝑡2
𝑀𝑚𝑖𝑛 −𝑀𝑠𝑎𝑡2−
𝑀𝑓𝑙2
𝜙 𝑀𝑚𝑖𝑛 −𝑀𝑓𝑙2 =
𝑀𝑠𝑎𝑡1
𝑀𝑚𝑖𝑛 −𝑀𝑠𝑎𝑡1−
𝑀𝑓𝑙1
𝜙 𝑀𝑚𝑖𝑛 −𝑀𝑓𝑙1 ,
(3.13)
where 𝑀 is the P-wave modulus, and subscripts min, fl1, and sat1 are mineral, fluid 1,
and state of being saturated with fluid 1, respectively. 𝜙 is the total porosity. First, we
apply fluid substitution (i.e., replace fluid 1 by fluid 2) to the clean sand layers, whose
P-wave moduli can be written as
𝑀𝑆 𝑠𝑎𝑡2 =
𝑀𝑆 𝑠𝑎𝑡1
𝑀𝑞𝑡𝑧 −𝑀𝑆 𝑠𝑎𝑡1−
𝑀𝑓𝑙1
𝜙𝑠 𝑀𝑞𝑡𝑧 −𝑀𝑓𝑙1 +
𝑀𝑓𝑙2
𝜙𝑠 𝑀𝑞𝑡𝑧 −𝑀𝑓𝑙2 𝑀𝑞𝑡𝑧
𝑀𝑆 𝑠𝑎𝑡1
𝑀𝑞𝑡𝑧 −𝑀𝑆 𝑠𝑎𝑡1−
𝑀𝑓𝑙1
𝜙𝑠 𝑀𝑞𝑡𝑧 −𝑀𝑓𝑙1 +
𝑀𝑓𝑙2
𝜙𝑠 𝑀𝑞𝑡𝑧 −𝑀𝑓𝑙2 + 1
,
(3.14)
where 𝑀𝑆 and 𝜙𝑠 are the P-wave modulus and total porosity of clean sand, respectively.
The mineral used here is quartz. If we let
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 90
𝐴 =𝑀𝑆 𝑠𝑎𝑡1
𝑀𝑞𝑡𝑧 −𝑀𝑆 𝑠𝑎𝑡1−
𝑀𝑓𝑙1
𝜙𝑠 𝑀𝑞𝑡𝑧 −𝑀𝑓𝑙1 +
𝑀𝑓𝑙2
𝜙𝑠 𝑀𝑞𝑡𝑧 −𝑀𝑓𝑙2 ,
(3.15)
then Equation 3.14 is equivalent to simply
𝑀𝑆 𝑠𝑎𝑡2 =𝐴 ∗ 𝑀𝑞𝑡𝑧
𝐴 + 1 .
(3.16)
For the interbedded sand-shale sequence, in which the clean sand and shale layers are
saturated with fluid 2 and brine respectively, we obtain the P-wave modulus of the
sequence by Backus averaging:
𝑀𝑠𝑎𝑡2 = 𝑉𝑠𝑎𝑛𝑑𝑀𝑆 𝑠𝑎𝑡2
+1 − 𝑉𝑠𝑎𝑛𝑑
𝑀𝑆 −1
,
(3.17)
where 𝑀𝑆 is the P-wave modulus of wet shale, and 𝑉𝑠𝑎𝑛𝑑 is the sand fraction of the
interbedded sequence. By replacing 𝑀𝑆 𝑠𝑎𝑡2 in Equation 3.17 by Equation 3.16, we
have
𝑀𝑠𝑎𝑡2 = 𝑉𝑠𝑎𝑛𝑑
𝐴 ∗ 𝑀𝑞𝑡𝑧 +𝑉𝑠𝑎𝑛𝑑𝑀𝑞𝑡𝑧
+1 − 𝑉𝑠𝑎𝑛𝑑
𝑀𝑆
−1
,
(3.18)
and if we let X = 𝑉𝑠𝑎𝑛𝑑
𝑀𝑞𝑡𝑧 +
1−𝑉𝑠𝑎𝑛𝑑
𝑀𝑆, then we obtain
𝑀𝑠𝑎𝑡2 = 𝑉𝑠𝑎𝑛𝑑
𝐴 ∗ 𝑀𝑞𝑡𝑧 + 𝑋
−1
𝑀𝑠𝑎𝑡2 =𝐴𝑀𝑞𝑡𝑧
𝑉𝑠𝑎𝑛𝑑
𝐴 ∗ 𝑋𝑀𝑞𝑡𝑧
𝑉𝑠𝑎𝑛𝑑+ 1
.
(3.19)
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 91
Now, if we directly apply Equation 3.13 to measurements which represent average
properties of clean sand and shale layers, we obtain
𝑀𝑠𝑎𝑡2∗ =
𝑀𝑠𝑎𝑡1
∗
𝑀𝑚𝑖𝑛∗ −𝑀𝑠𝑎𝑡1
∗ −𝑀𝑓𝑙1
∗
𝜙𝐸𝑓𝑓 𝑀𝑚𝑖𝑛∗ −𝑀𝑓𝑙1
∗ +
𝑀𝑓𝑙2∗
𝜙𝐸𝑓𝑓 𝑀𝑚𝑖𝑛∗ −𝑀𝑓𝑙2
∗ 𝑀𝑚𝑖𝑛
∗
𝑀𝑠𝑎𝑡1
∗
𝑀𝑚𝑖𝑛∗ −𝑀𝑠𝑎𝑡1
∗ −𝑀𝑓𝑙1
∗
𝜙𝐸𝑓𝑓 𝑀𝑚𝑖𝑛∗ −𝑀𝑓𝑙1
∗ +
𝑀𝑓𝑙2∗
𝜙𝐸𝑓𝑓 𝑀𝑚𝑖𝑛∗ −𝑀𝑓𝑙2
∗ + 1
.
(3.20)
Note that effective porosity is used instead of total porosity. If we let
𝐵 =𝑀𝑠𝑎𝑡1
∗
𝑀𝑚𝑖𝑛∗ −𝑀𝑠𝑎𝑡1
∗ −𝑀𝑓𝑙1
∗
𝜙𝐸𝑓𝑓 𝑀𝑚𝑖𝑛∗ −𝑀𝑓𝑙1
∗ +
𝑀𝑓𝑙2∗
𝜙𝐸𝑓𝑓 𝑀𝑚𝑖𝑛∗ −𝑀𝑓𝑙2
∗ ,
(3.21)
then Equation 3.20 is equivalent to simply
𝑀𝑠𝑎𝑡2∗ =
𝐵 ∗ 𝑀𝑚𝑖𝑛∗
𝐵 + 1 .
(3.22)
When Gassmann’s equation is applied at the measurement scale, in order to
correctly account for sub-resolution interbedding, we let
𝑀𝑠𝑎𝑡2∗ = 𝑀𝑠𝑎𝑡2
𝐵 ∗ 𝑀𝑚𝑖𝑛∗
𝐵 + 1=
𝐴𝑀𝑞𝑡𝑧
𝑉𝑠𝑎𝑛𝑑
𝐴 ∗ 𝑋𝑀𝑞𝑡𝑧
𝑉𝑠𝑎𝑛𝑑+ 1
.
(3.23)
One way to make Equation 3.23 hold is to equate the left side and the right side on a
term-by-term basis, which results in the following six equations:
𝑀𝑠𝑎𝑡1
∗
𝑀𝑚𝑖𝑛∗ −𝑀𝑠𝑎𝑡1
∗ 𝑀𝑚𝑖𝑛∗ =
𝑀𝑆 𝑠𝑎𝑡1
𝑀𝑞𝑡𝑧 −𝑀𝑆 𝑠𝑎𝑡1 𝑀𝑞𝑡𝑧
𝑉𝑠𝑎𝑛𝑑 ,
(3.24)
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 92
𝑀𝑠𝑎𝑡1
∗
𝑀𝑚𝑖𝑛∗ −𝑀𝑠𝑎𝑡1
∗ = 𝑀𝑆 𝑠𝑎𝑡1
𝑀𝑞𝑡𝑧 −𝑀𝑆 𝑠𝑎𝑡1 𝑋 ∗ 𝑀𝑞𝑡𝑧
𝑉𝑠𝑎𝑛𝑑 ,
(3.25)
𝑀𝑓𝑙1∗
𝜙𝐸𝑓𝑓 𝑀𝑚𝑖𝑛∗ −𝑀𝑓𝑙1
∗ 𝑀𝑚𝑖𝑛
∗ =𝑀𝑓𝑙1
𝜙𝑠 𝑀𝑞𝑡𝑧 −𝑀𝑓𝑙1
𝑀𝑞𝑡𝑧
𝑉𝑠𝑎𝑛𝑑 ,
(3.26)
𝑀𝑓𝑙1∗
𝜙𝐸𝑓𝑓 𝑀𝑚𝑖𝑛∗ −𝑀𝑓𝑙1
∗ =
𝑀𝑓𝑙1
𝜙𝑠 𝑀𝑞𝑡𝑧 −𝑀𝑓𝑙1
𝑋 ∗ 𝑀𝑞𝑡𝑧
𝑉𝑠𝑎𝑛𝑑 ,
(3.27)
𝑀𝑓𝑙2
∗
𝜙𝐸𝑓𝑓 𝑀𝑚𝑖𝑛∗ −𝑀𝑓𝑙2
∗ 𝑀𝑚𝑖𝑛
∗ =𝑀𝑓𝑙2
𝜙𝑠 𝑀𝑞𝑡𝑧 −𝑀𝑓𝑙2
𝑀𝑞𝑡𝑧
𝑉𝑠𝑎𝑛𝑑 ,
(3.28)
and 𝑀𝑓𝑙2
∗
𝜙𝐸𝑓𝑓 𝑀𝑚𝑖𝑛∗ −𝑀𝑓𝑙2
∗ =
𝑀𝑓𝑙2
𝜙𝑠 𝑀𝑞𝑡𝑧 −𝑀𝑓𝑙2
𝑋 ∗ 𝑀𝑞𝑡𝑧
𝑉𝑠𝑎𝑛𝑑.
(3.29)
From the above set of equations, we can solve for
𝑀𝑚𝑖𝑛∗ =
1
𝑋=
𝑉𝑠𝑎𝑛𝑑𝑀𝑞𝑡𝑧
+1 − 𝑉𝑠𝑎𝑛𝑑
𝑀𝑆
−1
,
𝑀𝑓𝑙1∗ =
1
𝑀𝑓𝑙1 −
1
𝑀𝑞𝑡𝑧 +
1
𝑀𝑚𝑖𝑛∗
−1
= 1
𝑀𝑓𝑙1 −
1
𝑀𝑞𝑡𝑧 +𝑉𝑠𝑎𝑛𝑑𝑀𝑞𝑡𝑧
+1 − 𝑉𝑠𝑎𝑛𝑑𝑀𝑆
−1
,
and 𝑀𝑓𝑙2∗ =
1
𝑀𝑓𝑙2 −
1
𝑀𝑞𝑡𝑧 +
1
𝑀𝑚𝑖𝑛∗
−1
= 1
𝑀𝑓𝑙2 −
1
𝑀𝑞𝑡𝑧 +𝑉𝑠𝑎𝑛𝑑𝑀𝑞𝑡𝑧
+1 − 𝑉𝑠𝑎𝑛𝑑𝑀𝑆
−1
.
(3.30)
The expressions shown in Equation 3.30 are the input adjustment required so that
when Gassmann’s equation is applied at the measurement scale, it also accounts for
the effect of sand-shale interbedding at the same time. In summary, the upscaled
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 93
Gassmann’s equation for a interbedded (clean) sand-shale sequence can be applied by
simply using Gassmann’s equation (Equation 3.13) with the following data/ input
parameters: P-wave modulus of a data point which represents interbedding of clean
sand saturated with fluid 1 and wet shale 𝑀𝑠𝑎𝑡1∗ , effective porosity at the
measurement scale (𝜙𝐸𝑓𝑓 ) , P-wave modulus of the effective mineral (𝑀𝑚𝑖𝑛∗ ) in
Equation 3.30, and P-wave moduli of the original and final saturating fluids
(𝑀𝑓𝑙1∗ 𝑎𝑛𝑑 𝑀𝑓𝑙2
∗ ) in Equation 3.30.
The exact input adjustments (Equation 3.30) are derived only for the case where
clean sand is interbedded with shale. However, we heuristically extend the derivation
results to the case where shaly-sand is interbedded with shale. P-wave moduli of the
effective mineral, the original, and the final saturating fluids required for the upscaled
Gassmann’s equation for shaly-sand interbedded with shale are as follows:
𝑚 =1
2 1 − 𝑉𝑑𝑖𝑠𝑝 𝑀𝑞𝑡𝑧 + 𝑉𝑑𝑖𝑠𝑝 𝑀𝑆 +
1
2
1 − 𝑉𝑑𝑖𝑠𝑝
𝑀𝑞𝑡𝑧 +𝑉𝑑𝑖𝑠𝑝
𝑀𝑆
−1
,
𝑀𝑚𝑖𝑛∗ =
𝑉𝑠𝑎𝑛𝑑𝑚
+1 − 𝑉𝑠𝑎𝑛 𝑑𝑀𝑆
−1
𝑀𝑓𝑙1∗ =
1
𝑀𝑓𝑙1 −
1 − 𝑉𝑠𝑎𝑛𝑑𝑀𝑚𝑖𝑛
∗ +1 − 𝑉𝑠𝑎𝑛𝑑𝑀𝑆
−1
,
and 𝑀𝑓𝑙2∗ =
1
𝑀𝑓𝑙2 −
1 − 𝑉𝑠𝑎𝑛𝑑𝑀𝑚𝑖𝑛
∗ +1 − 𝑉𝑠𝑎𝑛𝑑𝑀𝑆
−1
.
(3.31)
Note that we use the Voigt-Reuss-Hill average to compute the effective mineral (𝑚) of
the composite between quartz and wet shale (Dvorkin et al., 2007) to represent the
solid part of the shaly-sand layers as an additional step to the clean sand case. When
the volume fraction of dispersed clay (𝑉𝑑𝑖𝑠𝑝 ) is equal to zero (i.e., clean sand), 𝑀𝑚𝑖𝑛∗
in Equation 3.31 is reduced to 𝑀𝑞𝑡𝑧 , which is consistent with Equation 3.30. Also note
that 1 − 𝑉𝑑𝑖𝑠𝑝 and 𝑉𝑑𝑖𝑠𝑝 are not the actual volumetric fractions of quartz and wet shale
in the solid part of the shaly-sand layers, but rather a heuristic approximation.
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 94
3.9.6.1 Synthetic data examples
We test our upscaled Gassmann’s equation on synthetic data. Similar to Section
3.5.5, we generate interbedded sequences between sand and shale by systematically
varying both the volume fractions of dispersed clay in the sand layers (i.e., from
𝑉𝑑𝑖𝑠𝑝 = 0 to 𝑉𝑑𝑖𝑠𝑝 = 𝜙𝑆 ) and the sand fractions (i.e., from 𝑉𝑠𝑎𝑛𝑑 = 0 to 𝑉𝑠𝑎𝑛𝑑 = 1).
Thus, we provide a rough coverage for all possible interbedding scenarios allowed by
the triangular diagram (Figure 3.3). We assign a starting effective water-saturation of
0.5 (i.e., 𝑆𝑊𝑒=0.5). Then, we apply both the upscaled Gassmann and the Gassmann
shaly-sand equation (Dvorkin et al., 2007) to the synthetic data directly at the
measurement scale, and compare the results with the baselines, which are the results of
a two-step process –applying the Gassmann shaly-sand equation to the sand layers
only, and then upscaling these sand layers and shale by Backus averaging. This two-
step process is equivalent to the downscaling-upscaling procedure, and these baselines
are considered correct in synthetic data because all the parameters needed are known
(more detailed discussion in Section 3.5). Results are shown as percentage differences
between P-wave velocities after fluid substitution by the selected methods and the
baselines:
% 𝑑𝑖𝑓𝑓 = 100 ∗𝑉𝑝𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑚𝑒𝑡 𝑜𝑑 − 𝑉𝑝𝑏𝑎𝑠𝑒𝑙𝑖𝑛𝑒
𝑉𝑝𝑏𝑎𝑠𝑒𝑙𝑖𝑛𝑒 ,
(3.32)
where Vp is the P-wave velocity after fluid substitution.
We test the upscaled Gassmann in three different scenarios. The first scenario is to
compute parameter adjustments by following Equation 3.31, and then all the adjusted
parameters are input into approximate Gassmann’s equation (Mavko et al., 1995). The
next two scenarios are to test the performance of the upscaled Gassmann in the case
where both bulk and shear moduli are used. The second scenario is to compute
parameter adjustments by following steps similar to Equation 3.31, except that the
effective mineral (𝑚) of the composite between quartz and wet shale is now computed
by
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 95
𝑚𝐾 =1
2 1 − 𝑉𝑑𝑖𝑠𝑝 𝐾𝑞𝑡𝑧 + 𝑉𝑑𝑖𝑠𝑝 𝐾𝑆 +
1
2
1 − 𝑉𝑑𝑖𝑠𝑝
𝐾𝑞𝑡𝑧 +𝑉𝑑𝑖𝑠𝑝
𝐾𝑆
−1
,
𝑚𝜇 =1
2 1 − 𝑉𝑑𝑖𝑠𝑝 𝜇𝑞𝑡𝑧 + 𝑉𝑑𝑖𝑠𝑝 𝜇𝑆 +
1
2
1 − 𝑉𝑑𝑖𝑠𝑝
𝜇𝑞𝑡𝑧 +𝑉𝑑𝑖𝑠𝑝
𝜇𝑆
−1
,
and 𝑚 = 𝑚𝐾 +4
3𝑚𝜇 .
(3.33)
Then, all the adjusted parameters are input into approximate Gassmann’s equation
(Equation 3.13; Mavko et al., 1995). Finally, the third scenario is to compute
parameter adjustments by following steps similar to Equation 3.31, except that all the
P-wave moduli in the equation are replaced by bulk moduli
𝑘 =1
2 1 − 𝑉𝑑𝑖𝑠𝑝 𝐾𝑞𝑡𝑧 + 𝑉𝑑𝑖𝑠𝑝 𝐾𝑆 +
1
2
1 − 𝑉𝑑𝑖𝑠𝑝
𝐾𝑞𝑡𝑧 +𝑉𝑑𝑖𝑠𝑝
𝐾𝑆
−1
,
𝐾𝑚𝑖𝑛∗ =
𝑉𝑠𝑎𝑛𝑑𝑘
+1 − 𝑉𝑠𝑎𝑛𝑑
𝐾𝑆 −1
𝐾𝑓𝑙1∗ =
1
𝐾𝑓𝑙1 −
1 − 𝑉𝑠𝑎𝑛𝑑𝐾𝑚𝑖𝑛∗ +
1 − 𝑉𝑠𝑎𝑛𝑑𝐾𝑆
−1
,
and 𝐾𝑓𝑙2∗ =
1
𝐾𝑓𝑙2 −
1 − 𝑉𝑠𝑎𝑛𝑑𝐾𝑚𝑖𝑛∗ +
1 − 𝑉𝑠𝑎𝑛𝑑𝐾𝑆
−1
.
(3.34)
Then, all adjusted parameters are input into the actual Gassmann’s equation. We
compare results of the upscaled Gassmann for all three scenarios with those of the
Gassmann shaly-sand equation and the baselines. To assure consistency of fluid
substitution results among the selected methods, only P-wave moduli are used in all
calculations in the Gassmann shaly-sand equation and the baselines for the first
scenario, but both bulk and shear moduli are used in these two methods for the second
and the third scenarios. Results of the first, second, and third scenarios are shown in
Figure 3.26, Figure 3.27 and Figure 3.28, respectively.
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 96
Figure 3.26: Percentage differences between the baselines and the predicted velocities
by the upscaled Gassmann’s equation (left) and the Gassmann shaly-sand equation (Dvorkin et al., 2007) (right) for the first scenario, where P-wave moduli and approximate Gassmann’s equation are used for the upscaled Gassmann. Each location on the triangular diagram represents an interbedded sand-shale sequence with a unique pair of sand fraction and volume fraction of dispersed clay values. Note that the color scales of the two panels are different.
Figure 3.27: Percentage differences between the baselines and the predicted velocities by the upscaled Gassmann’s equation (left) and the Gassmann shaly-sand equation (Dvorkin et al., 2007) (right) for the second scenario, where both bulk and shear moduli, and approximate Gassmann’s equation are used in calculations. Each location on the triangular diagram represents an interbedded sand-shale sequence with a unique pair of sand fraction and volume fraction of dispersed clay values. Note that the color scales of the two panels are different.
0 0.1 0.2 0.3
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Effective porosity
1/M
(1/G
Pa)
Upscaled Gassmann
0
0.1
0.2
0.3
0.4
0.5
0 0.1 0.2 0.3
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Effective porosity1/M
(1/G
Pa)
Gassmann shaly-sand equation
0
5
10
15
20
% diff
Sand with
dispersed
shale
% diff
Clean
sand
Shale
Increasing
sand fractions
0 0.1 0.2 0.3
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Effective porosity
1/M
(1/G
Pa)
Upscaled Gassmann
0
0.5
1
1.5
2
0 0.1 0.2 0.3
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Effective porosity
1/M
(1/G
Pa)
Gassmann shaly-sand equation
0
5
10
15
20
Shale
Clean
sand
Sand with
dispersed
shale
% diff% diff
Increasing
sand fractions
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 97
Figure 3.28: Percentage differences between the baselines and the predicted velocities by the upscaled Gassmann’s equation (left) and the Gassmann shaly-sand equation (Dvorkin et al., 2007) (right) for the second scenario, where both bulk and shear moduli, and the actual Gassmann’s equation are used for the upscaled Gassmann. Each location on the triangular diagram represents an interbedded sand-shale sequence with a unique pair of sand fraction and volume fraction of dispersed clay values. Note that the color scales of the two panels are different. The color scale on the left panel does not cover either the actual maximum or minimum values of the percentage differences. The color scale is adjusted to show values of the majority of the results. Those results indicated by a magenta ellipse are unreliable due to instability of the method at low effective porosity.
In the first scenario (Equation 3.31 & approximate Gassmann), both the upscaled
Gassmann and the Gassmann shaly-sand equation (Dvorkin et al., 2007) show only
positive percentage differences in P-wave velocities relative to the baselines.
Therefore, both fluid substitution methods overpredict velocities. However, while the
prediction errors of the upscaled Gassmann are less than 0.6%, the errors of the
Gassmann shaly-sand equation can reach 20%. This very large difference in errors is
because our method accounts for the effect of sand-shale interbedding, whereas the
Gassmann shaly-sand equation ignores the effect of sand-shale interbedding by
assuming that all shale is dispersed.
Since in the upscaled Gassmann the parameter adjustment for the case where
shaly-sand interbedded with shale is only a heuristic approximation, the prediction
errors of this method are higher in the region near the shaly-sand line (i.e., sand with
dispersed clay line in Figure 3.26). In contrast, since the Gassmann shaly-sand
equation is most valid for shaly sands, the errors of this method are lower in the region
0 0.1 0.2 0.3
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Effective porosity
1/M
(1/G
Pa)
Upscaled Gassmann
-1.5
-1
-0.5
0
0.5
1
0 0.1 0.2 0.3
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Effective porosity
1/M
(1/G
Pa)
Gassmann shaly-sand equation
0
5
10
15
20
Shale
Clean
sand
Sand with
dispersed
shale
% diff% diff
Increasing
sand fractions
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 98
near the shaly-sand line. However, from this shaly-sand line, when moving inward in
the diagram, the errors of the Gassmann shaly-sand equation grow larger, especially in
the middle of the diagram where the volume fractions of laminated (interbedded) shale
are high, because all this laminated (interbedded) shale is misinterpreted as being
dispersed (Figure 3.26, right).
In the second scenario (Equation 3.33; approximate Gassmann), while the
prediction errors by the Gassmann shaly-sand equation reach 20%, the errors by the
upscaled Gassmann are less than 3% (Figure 3.27). The overall distributions of
prediction errors of both methods are similar to the first scenario (Equation 3.31;
approximate Gassmann) in Figure 3.26. While the high errors of the upscaled
Gassmann are more concentrated in the region near the shaly-sand line, the errors of
the Gassmann shaly-sand in the same region are relatively low, but they grow larger
toward the middle of the diagram (Figure 3.27). However, in this scenario even though
we use both bulk and shear moduli in computing the adjustment for the upscaled
Gassmann, we input these adjusted parameters into the approximate Gassmann’s
equation (Equation 3.13), which involves only P-wave moduli. However, we use the
actual Gassmann’s equation, which involves both bulk and shear moduli, to compute
the baselines. The difference in the approximate and the actual Gassmann’s equations
increases prediction errors of the upscaled Gassmann. Therefore, results of the
upscaled Gassmann show that there are non-zero errors near the clean sand point
(Figure 3.27, left), and that the error range in Figure 3.27(left) is larger than that of
Figure 3.26(left).
Instead of comparing results from the approximate and actual Gassmann’s
equations as in the second scenario, in the third scenario (Equation 3.34; actual
Gassmann) the actual Gassmann’s equation is used in all calculations. Therefore, the
overall absolute errors of the upscaled Gassmann in this scenario are smaller than in
the second scenario (Figure 3.27 and Figure 3.28, left). However, because of the
overgeneralization of the parameter adjustment in this scenario (i.e., simple
replacement of P-wave moduli by bulk moduli), a portion of the results corresponding
to low effective-porosity sequences becomes unreliable (those marked by a magenta
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 99
ellipse in Figure 3.28, left). Note that Gassmann’s equation is also reported to often
yield unreliable results for low porosity rocks (e.g., Smith et al., 2003).
3.9.6.2 Real data example
We test our upscaled Gassmann’s equation on the same dataset used in Section 3.8.
Starting with all brine-saturated data, we substitute oil for brine using five procedures:
Gassmann’s equation (i.e., ignoring the effect of sand-shale interbedding), the
Gassmann shaly-sand equation (i.e., assuming that shale is all dispersed; Dvorkin et al.,
2007), downscaling-upscaling, our mesh method, and our upscaled Gassmann
(Equation 3.34). Note that here both bulk and shear moduli are used in all calculations.
Sand fractions and volume fractions of dispersed clay, which are needed for the
downscaling-upscaling procedure and the upscaled Gassmann, are estimated using the
Thomas-Stieber-Yin-Marion model in the P-compliance vs. effective porosity space.
Results are shown in Figure 3.29 as P-wave velocities after fluid substitution by the
five selected procedures. As in the previous performance test on this dataset (Figure
3.22), the results of the downscaling-upscaling procedure are used as baselines.
However, to avoid any erratic behavior at very low porosity, the baselines from the
downscaling-upscaling procedure are chosen to include only those points with gamma
ray values less than 85.
In Figure 3.29, results of Gassmann’s equation, the Gassmann shaly-sand
equation, the baselines, and the mesh are exactly the same as the results for this dataset
in the previous performance test (Figure 3.22). Both Gassmann’s equation and the
Gassmann shaly-sand equation overestimate the fluid effect because sub-resolution
interbedding is not correctly accounted for. In contrast, the results from the mesh
method agree well with the baselines, and the results from the upscaled Gassmann
show the best agreement with the baselines. However, when using the upscaled
Gassmann, unreliable results (e.g., velocity spikes) are observed, especially in low
porosity data. Analogous to the synthetic case in Figure 3.28 (left), this instability of
the upscaled Gassmann may come from the overgeneralization in parameter
adjustment, where P-wave moduli are simply replaced by bulk moduli (Equation 3.34).
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 100
Figure 3.29: Comparisons of five sets of fluid substitution results which are color-coded by sand fractions estimated from the Thomas-Stieber-Yin-Marion model. Five fluid substitution procedures are used: Gassmann’s equation (i.e., ignoring the effect of sand-shale interbedding), Gassmann shaly-sand equation (i.e., assuming that shale is all dispersed; Dvorkin et al., 2007), the downscaling-upscaling procedure, the mesh method, and the upscaled Gassmann’s equation. Note: points with very low sand fractions are excluded from the plot. Lines X = Y are superimposed for comparison purpose.
3.9.6.3 Sensitivity analysis
In this section, we show how uncertainties in input parameters affect fluid
substitution results of the upscaled Gassmann. For this analysis, we use only P-wave
moduli information (Equation 3.31). The six input parameters are P-wave velocity of
wet porous shale (𝑉𝑝𝑠), volume fraction of dispersed clay (𝑉𝑑𝑖𝑠𝑝 ), sand fraction
(𝑁𝑇𝐺), effective water saturation in the sand layers (𝑆𝑊𝑒 ), fluid properties (i.e., bulk
modulus, 𝐾𝑓𝑙 ), and effective porosity of the interbedded sequence (𝜙𝐸𝑓𝑓 ).
The procedure and setting used here are similar to those of Section 3.7. The
starting effective pore space is half filled with oil. Then, we apply the upscaled
Gassmann and the downscaling-upscaling procedure to substitute brine for oil using
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
Vp downscaling-upscaling
(km/s)
Vp m
esh (
km
/s)
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
Vp downscaling-upscaling
(km/s)
Vp G
assm
ann (
km
/s)
(Dvork
in e
t al.,
2007)
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
Vp downscaling-upscaling
(km/s)
Vp G
assm
ann (
Tota
l) (
km
/s)
1.8 2 2.2 2.4 2.61.8
2
2.2
2.4
2.6
Vp downscaling-upscaling
(km/s)
Vp u
pscale
d G
assm
ann (
km
/s)
0.2
0.4
0.6
0.8
1Sand fraction
X > Y region
Line X = Y
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 101
all correct parameters, and these results are used as references. We assume that all five
input parameters are independent and normally distributed with means ( 𝜇 ) and
variances 𝜎2 . The means of the distributions are set to be the true values used in the
forward-modeling process. The standard deviations 𝜎 are set to be 10% of the means.
We test the effect of uncertainties in the input parameters one by one.
Results of the sensitivity analysis for the upscaled Gassmann are shown in Figure
3.30. Each subplot is indicated by a sand fraction (𝑉𝑠𝑎𝑛𝑑 ) and a volume fraction of
dispersed clay in the sand layers (𝑉𝑑𝑖𝑠𝑝 ). The thick lines on the top of each subplot are
the oil-saturated velocity and the reference velocities after fluid substitution from oil
to brine using the downscaling-upscaling procedure and the upscaled Gassmann with
all correct parameters. During the sensitivity analysis, one by one each input
parameter is assumed to be uncertain while the other parameters are assumed to be
correct, and the sensitivity analysis results are shown as a distribution of velocities
after fluid substitution.
When applying the upscaled Gassmann to the data, fluid substitution results are
only slightly affected by volume fraction of dispersed clay, effective water saturation,
and fluid bulk modulus. Uncertainty in effective porosity has a greater effect on the
results when the sand fraction in an interbedded sequence increases. For low sand-
fraction sequences, while the uncertainty in P-wave velocity of shale leads to a large
uncertainty in fluid substitution results, the uncertainty in sand fraction has only a
moderate effect on the results. However, when sand fraction increases, fluid
substitution results become less sensitive to shale velocity and become more sensitive
to the sand fraction. Overall, results show that the upscaled Gassmann is most
sensitive to P-wave velocity of wet porous shale and sand fraction, but this sensitivity
also depends on properties of interbedded sequences (e.g., sand fractions).
It is interesting to note that for our mesh method, fluid substitution results are most
sensitive to P-wave velocity of clean sand (Figure 3.19). The effect of uncertainty in
clean sand velocity on the results is more pronounced for high sand-fraction sequences
than for low sand-fraction sequences. Therefore, an alternative procedure for fluid
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 102
substitution in interbedded sand-shale sequences is to use the mesh method in low
sand-fraction sequences and to use the upscaled Gassmann in high sand-fraction
sequences. By doing so, fluid substitution results should be less sensitive to
uncertainty in input parameters.
Figure 3.30: Sensitivity analysis results of the upscaled Gassmann for six input parameters: P-wave velocity of wet porous shale, volume fraction of dispersed clay, sand fraction, effective water saturation, bulk modulus of fluid, and effective porosity of the interbedded package. The thick lines on the top of each subplot are the oil-saturated velocity and the reference velocities after fluid substitution from oil to brine using the downscaling-upscaling procedure and the upscaled Gassmann with all correct input parameters. Results of sensitivity analysis for each parameter are shown as a velocity distribution, which is normalized to one. Sand fractions and volume fractions of dispersed clay for each synthetic model are shown on the top of each subplot.
3.10 Conclusions
We provide a simple mesh method for approximating fluid substitution in sub-
resolution interbedded sand-shale sequences. The method is done by simply applying
Gassmann’s fluid substitution to clean sand and distorting the mesh. As a result, for
any sequence representing sand (or shaly-sand) interbedded with shale, the elastic
property after fluid substitution is scaled to the behavior of the clean sand that has
2.5 3 3.50
1
Sand fraction = 0.33 Vdisp = 0
2.5 3 3.50
1
Sand fraction = 0.33 Vdisp = 0.12
2.5 3 3.50
1
Sand fraction = 0.33 Vdisp = 0.24
2.5 3 3.50
1
Sand fraction = 0.67 Vdisp = 0
Norm
aliz
ed d
ensity f
unction
2.5 3 3.50
1
Sand fraction = 0.67 Vdisp = 0.12
2.5 3 3.50
1
Sand fraction = 0.67 Vdisp = 0.24
2.5 3 3.50
1
Sand fraction = 1 Vdisp = 0
P-velocity (km/s)
2.5 3 3.50
1
Sand fraction = 1 Vdisp = 0.12
P-velocity (km/s)
2.5 3 3.50
1
Sand fraction = 1 Vdisp = 0.24
P-velocity (km/s)
V
p shale
Vdisp
Sand fraction
Sw e
Koil
Effective porosity
Downscaling-upscaling
Upscaled Gassmann
Oil-saturated model
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 103
undergone fluid substitution. The mesh method is directly applicable at the
measurement scale, without the need to downscale the measurements, but its fluid
substitution results still agree well with those predicted using the actual downscaling-
upscaling procedure. The sensitivity analysis of five input parameters needed for the
mesh method shows that this method is most sensitive to the clean-sand P-wave
velocities, whose values can be constrained by rock physics trends. The method is
only robust for rocks that can be represented as sand-shale mixtures, in which any
porosity reduction is caused only by filling of shale in the pore space. Therefore, the
method does not account for effects of sorting and cementation. Because the mesh
method does not cover all possible variations in rocks, this method should not be
universally applied to a whole dataset without checking for its applicability. To
alleviate the problem of non-universality of the mesh method, we recommend
applying both the mesh method and the traditional Gassmann’s equation to the data.
The traditional Gassmann’s equation tends to overestimate the change of elastic
moduli after fluid substitution. The more accurate fluid substitution result is the one
that has a smaller change in elastic moduli, but how accurate is this observation should
be addressed in future analyses.
3.11 Acknowledgements
This work was supported by the Stanford Rock Physics and Borehole Geophysics
project.
3.12 References
Artola, F.A.V. and Alvarado, V., 2006, Sensitivity Analysis of Gassmann’s Fluid
Substitution Equations: Some Implications in Feasibility Studies of Time-lapse
Seismic Reservoir Monitoring: Journal of Applied Geophysics, 59, 47 – 62.
Avseth, P., Mukerji, T., and Mavko, G., 2005, Quantitative Seismic Interpretation,
Cambridge.
Backus, G. E., 1962, Long-wave elastic anisotropy produced by horizontal layering:
Journal of Geophysical Research, 67, 4427–4440.
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 104
Ball, V., Erickson, S., Brown, L., 2004, A Model-centric Approach to Seismic
Petrophysics: SEG Expanded Abstracts, 23, 1730.
Berryman, J. G., 1999, Tutorial: Origin of Gassmann’s equations: Geophysics, 64(5),
1627 – 1629.
Chopra, S., 2005, Expert Answers: Gassmann’s equation: CSEG Recorder in May, 8-
12.
Dvorkin, J. and Gutierrez, M.A., 2002, Grain Sorting, Porosity, and Elasticity:
Petrophysics, 43(3), a3.
Dvorkin, J., Mavko, G., and Gurevich, B., 2007, Fluid Substitution in Shaley
Sediment Using Effective Porosity: Geophysics, 72(3), o1 – o8.
Flórez, J., 2005, Integrating Geology, Rock physics, and Seismology for Reservoir-
quality Prediction, Ph.D. Thesis, Stanford University.
Gassmann, F., 1951, Uber die elastizitat poroser medien: Vier Natur Gesellschaft, 96,
1-23.
Han, D. and Batzle, M., 2004, Gassmann’s equation and fluid-saturation effects on
seismic velocities: Geophysics 69(2), 398-405.
Katahara, K., 2004, Fluid Substitution in Laminated Shaly Sands: SEG Expanded
Abstracts, 23, 1718-1721.
Katahara, K., 2008, What is Shale to a Petrophysicist?: The Leading Edge, 27, 738-
741.
Kumar, D., 2006, A tutorial on Gassmann fluid substitution: formulation, algorithm
and Matlab code: Geohorizon, 11, 4-12.
Lucier, A. M., Hoffmann, R., and Bryndzia, T., 2011, Evaluation of Variable Gas
Saturation on Acoustic Log Data from the Haynesville Shale Gas Play, NW
Louisiana, USA: The Leading Edge, March 2011, 300 – 311.
Marion, D., 1990, Acoustical, Mechanical and Transport Properties of Sediments and
Granular Materials, Ph.D. Thesis, Stanford University.
Mavko, G., Chan, C., and Mukerji, T., 1995, Fluid substitution: Estimating changes in
Vp without knowing Vs: Geophysics, 60(6), 1750-1755.
Mavko, G., Mukerji, T., and Dvorkin, J., 1998, The Rock Physics Handbook, 1st
edition, Cambridge.
Mavko, G., Mukerji, T., and Dvorkin, J., 2009, The Rock Physics Handbook, 2nd
edition, Cambridge.
Mavko, G., Avseth, P., and van Wijngaarden, J., 2006, Fluid substitution in laminated
rock intervals: SRB Meeting Volume, C1-C14.
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 105
Nur, A., Mavko. G., Dvorkin, J., and Galmudi., D., 1998, Critical porosity: a key to
relating physical properties to porosity in rocks: The leading edge, March 1998,
357 – 362.
Ruiz, F. and Azizov, I., 2011, Fluid substitution in tight shale using the Soft-porosity
model: SEG Expanded Abstracts, 30, 2272-2276.
Simm, R., 2007, Practical Gassmann Fluid Substitution in Sand/shale Sequences: First
Break, 25, 61-68.
Singleton, S. and Keirstead, R., 2011, Calibration of prestack simultaneous impedance
inversion using rock physics: The leading edge, 30(1), 70-78.
Skelt, C., 2004a, Fluid substitution in laminated sands: The Leading Edge, 23, 485–
488.
Skelt, C., 2004b, The influence of shale distribution on the sensitivity of
compressional slowness to reservoir fluid changes: SPWLA 45th
Annual Logging
Symposium, June 6-9.
Smith, T. M., Condergeld, C. H., and Rai, C. S., 2003, Gassmann Fluid Substitutions:
A Tutorial: Geophysics, 68(2), 430-440.
Spike K. T. and J. P. Dvorkin, 2005, Gassmann-consistency of velocity-porosity
transforms: The Leading Edge, 24, 581-583.
Thomas, E. C. and Stieber, S. J., 1975, The distribution of shale in sandstones and its
effect upon porosity: 16th Annual Logging Symposium, SPWLA, Paper T.
Yin, H., 1992, Acoustic Velocity and Attenuation of Rocks: Isotropy, Intrinsic
Anisotropy, and Stress Induced Anisotropy, Ph.D. Thesis, Stanford University.
CHAPTER 3: Fluid substitution for interbedded sand-shale sequences 106
107
Chapter 4
Seismic signature and uncertainty in
petrophysical property estimation of
thin sand-shale reservoirs
4.1 Abstract
Property estimation of thin sand-shale reservoirs using seismic response is often
challenging due to limited seismic resolvability. This chapter investigates seismic
signatures of sub-resolution sand-shale sequences and shows numerical examples of
using wavelet-transform based attributes and feature-extraction1 based attributes for
estimating reservoir properties which are net-to-gross ratio, saturation, and stacking
1 Feature extraction transforms the original features in high-dimensional data to create new features
which provide new representations of the data in a lower-dimensional space, while preserving as much
information in the data as possible. An example of linear feature extraction techniques is principal
component analysis (PCA) (Cunningham, 2008).
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 108
patterns. To investigate seismic signature, we generate thin sand-shale sequences (i.e.,
vertical arrangements of lithology) using 1-D Markov chain models. Layer properties
assigned to the sequences come from established binary-mixture models from rock
physics relations, in particular porosity and velocity versus sand-shale ratios. Forward
computation of the seismic response of the sequences is then used to extract attributes
and relate them to the spatial patterns and properties of thin sand-shale reservoirs. The
wavelet-transform based attributes are the slope and intercept of the log-log plot
between wavelet-coefficient modulus and scale. The feature-extraction techniques
compare the entire seismograms, capture variations in the amplitudes, and provide
new representations of the seismograms in a new feature space, which are used as
attributes. The two techniques considered here are multidimensional scaling
techniques (MDS) and kernel principal component analysis (KPCA). To quantify
attribute performance, we use the attributes for classifying seismograms with different
reservoir properties into classes and measure the accuracy of this classification by
computing the classification success rate. 1-D synthetic examples show that some of
the attributes, in particular KPCA, effectively discriminate different net-to-gross,
saturation, or stacking-pattern classes. A workflow similar to our synthetic study can
be applied to real seismic data to characterize thin sand-shale reservoirs. To illustrate
this application, we present two additional numerical examples. In the first example,
we assume that the transition matrix or thickness distributions at the well location are
known. In reality, this information can be directly extracted from the well data. Then,
we create synthetic 2-D spatial models describing geology away from the well point,
explore statistically how attributes vary with a change of net-to-gross ratios, and apply
these statistics to obtain the posterior distributions of net-to-gross ratios at three
selected locations from the unknown seismic section. In the second example, we
investigate seismic signatures of non-stationary thin sand-shale sequences, whose
layering patterns change vertically. The sequences are modeled as a first-order,
discrete, 1-D Markov chain. We use two different transition matrices to generate a
sequence in order to create non-stationarity. We analyze seismograms using moving
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 109
windows to obtain seismic signatures locally along the reservoir and estimate posterior
distributions of net-to-gross ratios along the reservoir following a Bayesian framework.
4.2 Introduction
Seismic estimation of reservoir properties is common in petroleum exploration.
However, estimating petrophysical properties of thinly layered reservoirs with layers
below seismic resolvability is often challenging. Previous studies (e.g., Khattri and Gir,
1976; Marion et al., 1994; Mukerji, 1995; Mukerji et al., 1995; Takahashi, 2000; Hart
and Chen, 2004; Hart, 2008) have shown that both the scales of heterogeneity (e.g.,
layer thickness, d) relative to the scale of seismic measurement (i.e., wavelength, λ)
and the difference in sediment stacking patterns affect seismic signatures and their
relations with rock properties. When λ/d is small, waves traveling perpendicular to the
layered medium can be described by the ray theory or short-wavelength limit in which
the total travel time through the media is simply the sum of the travel time of each
layer. The average velocity for the short-wavelength limit is the weighted average of
slowness. In contrast, when λ/d is large, waves travel through the layered medium
which behaves as a homogeneous effective medium with an effective velocity in the
effective medium theory or long-wavelength limit equal to the square root of the ratio
of the effective modulus and the average density. The effective modulus is calculated
using the Backus average (Backus, 1962). In the long-wavelength limit, when the
layered medium is perfectly periodic (e.g., alternating sand-shale system), normally
incident reflection seismograms exhibit no internal reflection between the top and the
base reflections, and the reflected amplitudes directly relate to both the properties and
the proportions of the periodic materials (Stovas et al., 2006). For a non-periodic
layered medium, internal reflections can be observed, and the relation between the top
reflectivity and the proportions of the layered materials become non-deterministic and
non-unique due to not only the natural variation in material properties but also how the
materials are arranged (Takahashi, 2000; Mukerji and Mavko, 2008).
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 110
Most stratigraphic sequences in nature reflect non-random stacking of sedimentary
patterns. Depending on depositional environments, the main characteristics of such
patterns include lateral extents, vertical arrangements of lithologies, and layer
thickness distributions (Harms and Tackenberg, 1972). Markov chains have been used
as a tool to simulate bedded sequence to capture these preferred directionality and
asymmetric facies associations signaling depositional process (e.g., Krumbein and
Dacey, 1969; Harbaugh and Bonham-Carter, 1970; Schwarzacher, 1975; Xu and
MacCarthy, 1996; Parks et al., 2000). Velzeboer (1981) modeled sequences by a first-
order Markov chain with distributions of physical properties and theoretically derived
an expression for power spectrum of the seismic reflection response. This expression
is related to sedimentary parameters such as sand-shale ratio. Sinvhal and Sinvhal
(1992) simulated multiple realizations of lithologic models from well log information
using first-order Markov chains and generate the corresponding synthetic seismograms.
Then, several features were extracted from the autocorrelation and the power spectrum
of these seismograms, and these features were used in linear discriminant analysis for
identifying and predicting lithology from seismic data. In this chapter, we build on this
work with the goal of extracting attributes from seismograms statistically and use
them for quantitative seismic interpretation of thin sand-shale reservoirs. Even though
our general workflow is similar to that from Sinvhal and Sinvhal (1992), there are
several important differences. We extend the analysis to investigate the effects of other
properties such as saturation and stacking patterns on seismic signatures of thin
reservoirs. We use rock-physics models to create what-if scenarios with varying
reservoir properties. Instead of using fixed properties for each lithology (Sinvhal and
Sinvhal, 1992), we also incorporate property variations into the models. Most of the
attributes in our study are extracted directly by comparing the amplitudes of the entire
seismogram segments.
In this chapter, we investigate seismic signatures of thin (sub-seismic resolution)
sand-shale sequences using wavelet-transform based attributes and feature-extraction
based attributes. This chapter also presents a practical workflow for seismic
interpretation of reservoir properties of thin sand-shale sequences. The rest of this
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 111
chapter is organized as follows. Section 4.3 describes the forward-modeling process
for generating multiple realizations of thin sand-shale sequences and their
corresponding seismic responses. The section also discusses both types of seismic
attributes. Section 4.4 illustrates the effects of net-to-gross ratio, saturation, and
stacking patterns on seismic signatures through 1-D synthetic examples which
demonstrate the performance of various seismic attributes. Section 4.5 provides a
practical workflow through a synthetic example for estimating net-to-gross ratios from
2-D seismic sections using a Bayesian framework. Section 4.6 shows another synthetic
example for estimating net-to-gross ratios along non-stationary sequences whose
layering patterns vary from top to bottom, and finally Section 4.7 discusses the
benefits, limitations, and possible extension of the workflow and the attributes
presented in this chapter.
4.3 Forward modeling for seismic response and attributes
In this section, we briefly review the methodology and tools used in forward
modeling to generate seismic responses and their attributes for thin sand-shale
sequences. Our forward-modeling process has three steps: (1) using Markov chain
models to generate sequences of sedimentary layers, (2) using rock-physics models to
assign physical properties to the layers and simulating seismic responses of the
sequences, and (3) extracting seismic attributes which will be later used in estimating
reservoir properties. An overall workflow of the forward-modeling steps is shown in
Figure 4.1.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 112
Figure 4.1: Overall workflow for seismic-signature study and property estimation.
First, thin sand-shale sequences are generated using Markov-chain models and rock physics relations. Then, seismic responses for these sequences are modeled, and seismic attributes are extracted. Finally, the attributes are related to reservoir properties, which can be used for reservoir characterization of target areas.
4.3.1 Markov chain models in stratigraphic sequences
We generate vertical arrangements of sand and shale layers in thin sand-shale
sequences using Markov chain models. In stratigraphic analysis, a column of
sediments can be described as a chain or spatial arrangement of a finite number of
discrete states (i.e., lithology). Markov chains use conditional probabilities to describe
the dependency of the current state on the previous states. If the transition from one
state to the next depends only on the immediately preceding state, the chain is said to
be first-order (Harbaugh and Bonham-Carter, 1970; Sinvhal and Khattri, 1983).
A Markov chain model is commonly represented by a transition matrix, whose
element pij (at the ith
row and jth
column) represents the probability of a transition from
state i to state j, or the probability of going to state j, given that i is the current state. In
a stratigraphic study, the transition matrix is usually obtained from real geological
observations and is typically constructed in one of two ways: either by counting states
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 113
using a fixed sampling interval, or by counting states only when a transition occurs (an
embedded form).
When counting states using a fixed sampling interval, the lithologic state is
determined and considered only at discrete points equally spaced along a stratigraphic
column. This allows successive points to have similar lithology, which implies that the
diagonal element (i.e., the probability that a state has a transition to itself) can be non-
zero (Krumbein and Dacey, 1969). In practice, however, selecting a proper sampling
interval for this method can be problematic. Choosing an interval that is too small
relative to the overall average bed thickness can increase the counts of transitions of a
state to itself. Consequently, the diagonal elements become very large, and
probabilities of the state transiting into the others become unreasonably small. In
contrast, using a sampling interval that is too large can miss very fine-layered
characteristics of the sequences (Sinvhal and Sinvhal, 1992). A sequence simulated by
this fixed-sampling type always yields lithologic states with thicknesses that are
geometrically distributed (Krumbein and Dacey, 1969).
In contrast, in an embedded-form transition matrix, all diagonal elements are zero,
since transitions are considered only when lithologic states change. In this case, the
step size between two consecutive states is not a fixed interval, but the actual observed
bed thicknesses (Parks et al., 2000). Thus, a sequence of states generated from an
embedded transition matrix does not contain information about thicknesses of each
layer. Thicknesses are simulated separately assuming that the layer thickness of each
lithologic state is distributed according to some distributions (i.e., semi Markov
process). A case when the layer thickness is exponentially distributed is called a
continuous-time Markov chain model. Examples of transition matrices and sequences
obtained from both the fixed-sampling and the embedded-form transition matrices are
shown in Figure 4.2 and Figure 4.3.
The probability of transition from state i to state j in n steps (denoted as p(n)
ij) can
be determined by raising the transition matrix to the power n. If a limiting distribution
exists, a successive multiplication leads to a row vector of fixed probabilities
representing proportions of each state in the long-term behavior. In terms of
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 114
stratigraphic application, this limiting distribution (or stationary distribution π) implies
proportions of each lithology in the entire sequence provided that sufficient numbers
of transition steps occur (Harbaugh and Bonham-Carter, 1970).
Figure 4.2: Three examples of transition matrices with fixed sampling intervals:
retrogradational, progradational, and aggradational sequences. The lithologic states in the transition matrices are sand (s), shaly sand (sh-s) sandy shale (s-sh) and shale (sh). The off diagonal elements marked by arrows control the directionality of the sequences.
S Sh S ShS Sh
Fixed
Retrograding Prograding Aggrading
(fining-upwards) (coarsening-upwards)
S Sh S ShS Sh
Fixed
S Sh S ShS Sh
FixedFixed
Retrograding Prograding Aggrading
(fining-upwards) (coarsening-upwards)
Fixed
Sh SSh S Sh S
Retrogradational Progradational Aggradational
(fining-upwards) (coarsening -upwards)
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 115
Figure 4.3: Examples of an embedded-form transition matrix with realizations of
sequences. The lithologic states in the transition matrix are sand (s), shaly sand (sh-s) sandy shale (s-sh) and shale (sh). An example of thickness distributions used is shown in the lower left corner.
Using either the fixed-sampling or the embedded-form transition matrices, it is
possible to generate sequences that mimic stratigraphic stacking patterns, for example,
by manipulating the fixed-sampling matrices with appropriate off-diagonal patterns, as
shown in Figure 4.2.
Parasequences and parasequence sets are two fundamental stratal units in sequence
stratigraphy. Parasequences contains a series of beds or bedsets that are genetically
related, in which younger beds are deposited and built up basinward (Van Wagoner, et
al., 1988; Van Wagoner, et al., 1990). Most siliciclastic parasequences show
progradational stacking pattern and all exhibit shoaling-upward trends in their vertical
sections with sediments which are generally coarsening-upward (e.g., beach and
deltaic parasequences) but sometimes fining-upward (e.g., tidal-flat parasequences).
Some characteristics of coarsening-upward parasequences include upward increases in
grain size, thickness of sandstone beds, and ratio of sandstone to mudstone. In contrast,
upward decreases in similar elements are observed in fining-upward parasequences
(Van Wagoner, et al., 1988; Van Wagoner, et al., 1990; Kamola and Van Wagoner,
1995; Mulholland, 1998). In response to the relation between depositional rates and
accommodation rates, lateral shifts of successive parasequences result in parasequence
02.02.06.0
2.001.07.0
3.03.004.0
5.03.02.00
S Sh-S S-Sh Sh
S
Sh-S
S-Sh
Sh
0 1 2 3 4 5 60
50
100
150
200
250
Thickness (m)S Sh S Sh S Sh
02.02.06.0
2.001.07.0
3.03.004.0
5.03.02.00
S Sh-S S-Sh Sh
S
Sh-S
S-Sh
Sh
0 1 2 3 4 5 60
50
100
150
200
250
Thickness (m)S Sh S Sh S Sh
02.02.06.0
2.001.07.0
3.03.004.0
5.03.02.00
S Sh-S S-Sh Sh
S
Sh-S
S-Sh
Sh
02.02.06.0
2.001.07.0
3.03.004.0
5.03.02.00
S Sh-S S-Sh Sh
S
Sh-S
S-Sh
Sh
0 1 2 3 4 5 60
50
100
150
200
250
Thickness (m)0 1 2 3 4 5 6
0
50
100
150
200
250
Thickness (m)S Sh S Sh S Sh
Sh SSh SSh SSh S
Sh SSh S
0 1 2 3 4 50
50
100
150
Thickness (m)
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 116
sets with predictable stacking patterns: retrogradational, progradational or
aggradational. When the depositional rate exceeds the accommodation rate, younger
parasequences are progressively deposited basinward, yielding a progradational
parasequence set. When the depositional rate is smaller than the accommodation rate,
younger parasequences are deposited landward, yielding a backstepping,
retrogradational parasequence set whose individual parasequences are progradational.
When the depositional rate is approximately equal to the accommodation rate, there is
no significant lateral shift in deposition of successive parasequences, resulting in an
aggradational parasequence set whose characteristics in terms of facies, thickness, and
sandstone to mudstone ratios stay similar throughout its vertical section (Van Wagoner,
et al., 1990; Mitchum and Van Wagoner, 1991; Kamola and Van Wagoner, 1995;
Boggs, 2001). Examples of SP log responses of parasequence sets with the three
stacking patterns are shown in Figure 4.4.
Figure 4.4: Examples of SP log responses showing stacking patterns in parasequence
sets. (Left) Retrogradational, (middle) progradational and (right) aggradational patterns (Modified after Van Wagoner et al., 1990).
4.3.2 Rock-physics models for sand-shale mixtures
After obtaining multiple sedimentary sequences, we assign petrophysical
properties to those sand-shale layers with rock-physics models and equations. The four
lithologic states (i.e., sand, shaly-sand, sandy-shale and shale) can be represented by
mixtures of sand and clay, where each state is assigned a unique clay fraction. Porosity
values corresponding to the mixtures of sand with the specified clay fractions are
determined by using the Yin-Marion dispersed-mixing model (Marion et al., 1992;
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 117
Yin, 1992). This model describes the topology of the bimodal mixtures and a V-
shaped relation between the volume fraction of clay and the mixture porosity.
When a small volume of clay (i.e., less than the sand porosity) is added to an
original packing of sand, clay starts filling the sand pore space without disturbing the
sand packing. Sand grains provide the load-bearing matrix of the mixture. At this stage,
porosity decreases, because clay particles replace some portions of the original sand
pore space. When the clay content is greater than the sand porosity, sand grains are
displaced and disconnected. The mixture changes from grain-supported to clay-
supported sediments. At this stage, porosity increases linearly with increasing clay
content, because the solid sand grains are replaced by porous clay. As a result, the plot
of the volume fraction of clay versus the total porosity of the mixture shows a V-
shaped pattern (Figure 4.5) (Marion et al., 1992; Yin, 1992).
Figure 4.5: Illustrations of sand-shale mixtures, with their porosity and velocity values related to clay content (Modified after Marion et al., 1992). Porosity versus clay content shows a V-shaped trend, where the two end points are the pure sand and pure shale porosity. Selected clay fractions corresponding to the four lithologic states are marked. The states are sand (s), shaly-sand (sh-s), sandy-shale (s-sh), and shale (sh).
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
Volume fraction of clay
Tota
l poro
sity
0.1 0.2 0.3
2
2.5
3
3.5
Total porosity
P-w
ave v
elo
city (
km
/s)
clay
0
0.5
1
S
Sh-sS-sh
Sh
S-sh
Sh-s
S
Sh
sand, shaly sand sandy-shale, shale
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 118
Given our sand-clay mixtures, we obtain the corresponding P-wave velocities
using the soft-sand model. This model uses the lower Hashin-Shtrikman bound to
construct velocity-porosity trends for sand mixing with a specified clay volume
(Avseth et al., 2005). The plot of P-wave velocity versus porosity, as clay fractions in
the mixture increase from 0 to 1, shows an inverted-V trend (Figure 4.5). In order to
explore various saturation scenarios, Gassmann’s equation (Gassmann, 1951) is used
to compute properties of rocks saturated with pore-fluid mixtures of water and oil.
Water saturation values (Sw) for each lithologic state are varied depending on case
studies. The density assigned to each layer is simply a weighted average of the
densities of all components in that particular layer. With the rock-physics models,
velocity and density values are now assigned to each sedimentary layer in the
sequences. Layer properties (i.e., velocity, density and thickness) are then input into
the Kennett algorithm (Kennett, 1983) to simulate full-waveform, normally-incident,
reflected seismograms using a zero-phase Ricker wavelet with a central frequency of
30 Hz. The simulations are performed for all reverberations.
4.3.3 Seismic attributes
In our study of seismic signatures for thin sand-shale sequences, we extract two
main types of attributes from seismograms and relate these attributes to the reservoir
properties of interest (i.e., net-to-gross ratios, saturations, and stacking patterns). The
two attributes introduced in this chapter are wavelet-transform based attribute and
feature-extraction based attributes.
4.3.3.1 Wavelet-transform-based analysis
Wavelet transform decomposes a time series signal into a set of scaled and
translated versions of a selected wavelet function. The transform has been used, for
example, to study fractal behavior of seismic data and well logs to characterize
lithofacies (e.g., Álvarez et al., 2003; López and Aldana, 2007). By using well logs,
López and Aldana (2007) showed a possible relation of lithofacies and parameters
including fractal dimension, intercept and slope obtained from linear fits to log-log
plots of variance of wavelet-transform coefficients versus scale.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 119
Using a complex Gaussian wavelet, we perform the transform on simulated
seismograms yielding wavelet-transform coefficients at various scales. We calculate
variance of the modulus of these coefficients for every scale and make a log-log
(base2) plot of the variance versus scale (Figure 4.6). Then, we extract the slope and
intercept of a linear fit as statistical attributes for each realization of the sand-shale
sequences. Wavelet-transform attributes shown in this work are derived using complex
Gaussian wavelets. Note that we have also tried a few other wavelets and found that
behavior of log-log plots between wavelet-coefficient modulus and scale changes
depending on the wavelets and the range of scale. The intercept attribute from this
wavelet-transform analysis relates to the total variance of the seismogram.
Figure 4.6: Wavelet-transform analysis for extracting attributes from a seismogram.
(Left) modulus of wavelet-transform coefficients for the seismogram. The red arrow indicates the modulus along a particular scale. (Right) a log-log plot of scale versus variance of modulus of wavelet-transform coefficients. The plot is shown in open circles which are fit by a straight line. The slope and intercept of this line are used as seismic attributes.
4.3.3.2 Multidimensional scaling and Kernel principal component analysis
Even though the thin sand-shale layers are below seismic resolution, seismograms
from sequences with similar underlying geology and properties potentially share more
statistical similarities than those from sequences with different properties.
Seismograms of length n can be considered as n-dimensional vectors in the space ℝ𝑛 .
To compare the simulated seismograms, we obtain new representations of these
seismograms in a lower dimensional space using two feature extraction techniques:
multidimensional scaling (MDS) and kernel principal component analysis (KPCA).
0 1 2 3 4 5-24
-22
-20
-18
-16
-14
-12
-10
-8
5 10 15 20 25 30
0.05
0.1
0.15
0.2Log
2(v
aria
nce)
Log2(scale)scale-0.1 0 0.1
0
500
1000
1500
2000
2500
Complex Gaussian
Wavelet
0 1 2 3 4 5-24
-22
-20
-18
-16
-14
-12
-10
-8
5 10 15 20 25 30
0.05
0.1
0.15
0.2Log
2(v
aria
nce)
Log2(scale)scale-0.1 0 0.1
0
500
1000
1500
2000
2500
Complex Gaussian
Wavelet
0 1 2 3 4 5-24
-22
-20
-18
-16
-14
-12
-10
-8
5 10 15 20 25 30
0.05
0.1
0.15
0.2Log
2(v
aria
nce)
Log2(scale)scale-0.1 0 0.1
0
500
1000
1500
2000
2500
Complex Gaussian
Wavelet
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 120
These new representations are used as attributes, which are then related to reservoir
properties.
The first technique used for extracting attributes is multidimensional scaling. To
compare p seismograms of length n using MDS, we first construct a p-by-p
dissimilarity matrix, whose element at the ith
row and jth
column (𝛿ij) is a dissimilarity
measurement between the ith
and jth
seismograms. The dissimilarity values can be
assigned from subjective judgment or objective measure (Cox and Cox, 2001). For
seismogram comparison, dissimilarity need not always be the Euclidean distance
between any two seismograms; we can define dissimilarity in many different ways.
For example, we apply wavelet transform to seismograms and obtain wavelet
coefficients for specified scales. Dissimilarity can be defined as pairwise Euclidean
distance between their modulus of wavelet coefficients at selected scales (Figure 4.7).
Figure 4.7: Modulus of wavelet coefficients (shown in the middle two boxes) of two
seismogram segments. Dissimilarity between these seismograms is defined as distance between their modulus maps.
To obtain attributes, this dissimilarity matrix is then input into an MDS algorithm
which transforms the matrix into points in lower-dimensional Euclidean space. Points
corresponding to their original objects (i.e., seismograms) are configured so that the
inter-point Euclidean distances (dij) match the input dissimilarity values (𝛿ij) as much
as possible. MDS is classified into various categories: classical (metric), metric, and
non-metric MDS. While classical MDS solves an eigendecomposition problem for
coordinates of points in lower-dimensional space, both metric and non-metric MDS
use iterative optimization techniques. Metric MDS minimizes difference between dij
and f(𝛿ij), where f is a continuous monotonic function. Examples of such functions
include identity transformation, logarithmic transformation, etc. (Cox and Cox, 2001;
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 121
Zhang, 2008). Non-metric MDS aims to find a configuration of points so that the rank
order of the dissimilarity matrix is preserved (Cox and Cox, 2001; Webb, 2002). The
main result of MDS algorithms is the relative positions of points in lower-dimensional
space; thus, absolute positions do not matter (Scheidt and Caers, 2009). Note that
classical MDS with Euclidean distances is equivalent to principal component analysis
(PCA) (Cox and Cox, 2001; Williams, 2002). We use these new coordinates obtained
by MDS as seismic attributes for reservoir characterization.
The other technique used for extracting seismic attributes is KPCA, which also
provides new representations of seismograms in lower-dimensional space. We first
briefly review the concept of linear principal component analysis (PCA). PCA is
commonly used for dimensionality reduction and linear feature extraction. This
method performs eigenvalue decomposition of a data covariance matrix to compute a
set of orthogonal directions (principal components) which capture maximum
variability of the data. The projection of each data point onto selected principal
components creates a new representation of each point in a low-dimensional space.
While PCA does not work well with non-linear features, kernel principal component
analysis (KPCA), introduced by Schӧlkopf et al. (1997), provides an alternative
nonlinear PCA method. The idea behind KPCA is to map input data nonlinearly into
some high dimensional feature space, in which the data become linearly separable;
then linear PCA is performed (Scheidt and Caers, 2009). Instead of solving an
eigenvalue problem of the covariance matrix, linear PCA in the feature space is done
by solving an equivalent system involving dot products in the feature space. These dot
products can be obtained using kernel functions without an explicit mapping from the
original space to the feature space. Results from KPCA are the projections of input
data onto the principal components.
To obtain attributes, we apply Gaussian, dynamic similarity, inverse multi-quadric,
polynomial and linear kernels (Table 4.1) to the input seismograms and obtain kernel
matrices. Then after centering these matrices, we compute eigenvalues and
eigenvectors of the matrices, and obtain coordinates of points, each of which
corresponds to a seismogram that is projected onto selected principal components.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 122
Hereafter we refer to KPCA with Gaussian, dynamic similarity, inverse multi-quadric,
polynomial, or linear kernels as Gaussian, dynamic similarity, inverse multi-quadric,
polynomial, or linear KPCA, respectively. Note that performing KPCA with a linear
kernel is equivalent to performing linear PCA (e.g., Tenenbaum et al., 2000; Rathi et
al., 2006; van der Maaten, et al., 2009). Likewise, performing KPCA with kernel
functions that depend only on 𝑥𝑖 − 𝑥𝑗 can be interpreted as solving a metric MDS as
an eigenproblem (Williams, 2002). Parameters such as 𝜎, c, or d that are used for our
study will be shown in the note column of result tables for each case scenario.
Table 4.1: Kernel functions used for extracting seismic attributes. Note that xi is the ith seismogram.
Kernel matrix Note
Gaussian kernel
𝑘 𝑥𝑖 , 𝑥𝑗 = 𝑒− 𝑥𝑖−𝑥𝑗
2
2𝜎2
σ > 0; a suggested value of 𝜎 is
about 20% of the range of the
distances (Scheidt and Caers,
2009)
𝑥𝑖 − 𝑥𝑗 is the Euclidean
distance between the ith
and jth
seismograms
Dynamic
similarity kernel
(Yan et al., 2006)
𝑘 𝑥𝑖 , 𝑥𝑗 = 𝑒−𝐷𝑃𝐹2 𝑥𝑖 ,𝑥𝑗
𝜎2 Dynamic partial function (DPF)
by Li et al. (2003)
(see below for more details)
Inverse multi-
quadric kernel 𝑘 𝑥𝑖 , 𝑥𝑗 =
1
𝑐2 + 𝑥𝑖 − 𝑥𝑗 2
𝑥𝑖 − 𝑥𝑗 is the Euclidean
distance between the ith
and jth
seismograms
Polynomial kernel 𝑘 𝑥𝑖 , 𝑥𝑗 = 𝑥𝑖 ∙ 𝑥𝑗 + 𝑐 𝑑
c ≥ 0 and d is the degree of the
polynomial (d ∈ ℕ)
Linear kernel 𝑘 𝑥𝑖 , 𝑥𝑗 = 𝑥𝑖 ∙ 𝑥𝑗 equivalent to performing PCA
The Euclidean distance ( 𝑥𝑖 − 𝑥𝑗 ) used in both the Gaussian kernel and the
inverse multi-quadric kernel is a special case of a more general Minkowski metric,
which is defined as
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 123
𝑑(𝐴,𝐵) = ak − bk r
n
k=1
1r
,
(4.1)
where d is the distance or similarity measure between objects A and B, each of which
is represented by a vector of length n (e.g., A = [a1, a2, a3, …, an]). The Euclidean
distance is the Minkowski metric with the parameter r equal to 2. Using the
Minkowski metric to measure similarity between two objects is to compare each and
every element in the objects.
Based on the idea similar to human perceptual similarity in cognitive science, Li et
al. (2003) defined a distance function called dynamic partial function (DPF) as
𝐷𝑃𝐹(𝐴,𝐵) = δkr
δk ϵ∆m
1r
,
(4.2)
where δk = |ak – bk| for k = 1,…, n and Δm is the set of the m smallest values of δ
from {δ1,…, δn}. Thus parameter m ranges from 1 to n. If m = n, Equation 4.2 is
equivalent to the Minkowski metric.
Using DPF, Yan et al. (2006) proposed a new kernel for image classification and
showed that this kernel yields higher classification accuracy than other kernels,
including Gaussian and polynomial kernels. As an extension of the Gaussian kernel,
the new kernel is defined as
𝑘 𝑥𝑖 , 𝑥𝑗 = 𝑒−𝐷𝑃𝐹 2(𝑥𝑖 ,𝑥𝑗 )
𝜎2 .
(4.3)
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 124
4.4 Seismic signatures for 1-D Synthetic example
In the previous section, we have described how we generate seismic responses and
extract seismic attributes for thin sand-shale sequences. In this section, we illustrate 1-
D synthetic examples where seismic attributes are related to net-to-gross ratios,
saturations, and various sedimentary stacking patterns.
4.4.1 Model setup
We characterize the lithology into four states: sand, shaly sand, sandy shale, and
shale. Then we use a set of 4x4 fixed-sampling transition matrices, in particular the
aggrading type, to generate multiple realizations of lithology arrangements by keeping
the layer thickness fixed at 0.5 m. Each sand-shale sequence has a total thickness of
100 m, and the sequence is embedded between two 200-m thick shale layers. The four
lithologic states are represented by mixtures of sand and clay with clay fraction values
of 0.1, 0.3, 0.6 and 0.9 in the order of increasing shaliness. Using clay-fraction
information, porosity, velocity and density values are computed following methods
provided in Section 4.3.2. We introduce uncertainties by assuming each lithologic
state having a distribution of velocities with a mean equal to the calculated velocities
and standard deviations (v) of 0.1 and 0.2 km/s. Full-waveform, normally-incident,
reflected seismograms are simulated using the Kennett algorithm (Kennett, 1983) with
a zero-phase Ricker wavelet. The central frequency is 30 Hz and the simulations are
performed for all reverberations. A summary workflow from a transition matrix to
multiple realizations of seismograms is shown in Figure 4.8.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 125
Figure 4.8: Workflow for generating seismic responses of multiple realizations of thin
sand-shale sequences. The four lithologic states in the transition matrix are sand (s), shaly-sand (sh-s), sandy-shale (s-sh), and shale (sh).
We explore three main scenarios to investigate the effect of net-to-gross ratios,
saturations, and stacking patterns on seismic signatures of thin sand-shale sequences.
We define the net-to-gross ratio as the proportion of the first lithologic state (sand) in
the entire sequence.
4.4.2 Scenario 1: Effect of net-to-gross ratios
We study three transition matrices which generate aggrading-type sequences
(Table 4.2). Each transition matrix has a different limiting distribution (𝜋), a row
vector of fixed probabilities representing proportions of each state in the long-term
behavior. Thus in long sequences (i.e., large numbers of layers per sequence) the
proportions of sand in the sequences should converge to a distinct value: 𝜋sand. In our
simulations, all sequences consist of 200 sedimentary layers. We simulate 200
sequence realizations from each transition matrix and compute the true net-to-gross
ratios of these sequences. Figure 4.9 shows that the computed net-to-gross values
scatter around 𝜋sand and that their distributions get narrower when the values of
parameter k in the transition matrices (Table 4.2) approach 1. Since the probabilities of
going from sand (or shaly-sand) to shale and the probabilities of going from sandy-
shale (or shale) to sand are equal to k, as k gets larger the simulations will eventually
turn into sequences with only sand and shale in alternating layers; as a result, the
Sh S
Sh S
Sh S
Transition matrix Thin sand-shale sequenceswith petrophysical properties
Synthetic seismograms
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 126
distributions of net-to-gross ratios become very narrow around the value of 0.5. Water
saturation (Sw) is set to be the same in all simulations: Sw=0.1 for sand layers and
Sw=1 for the other lithology states. For the effect of net-to-gross ratios, we run
simulations for velocity distributions with standard deviations (v) of 0.1 and 0.2 km/s.
Table 4.2: Form of transition matrices for generating sequences used in investigating net-to-gross effects on seismic signatures. Values of parameter k range from 0.45 to 0.95. The four lithologic states are sand (s), shaly-sand (sh-s), sandy-shale (s-sh), and shale (sh).
Sand Shaly-sand Sandy-shale Shale
Sand (1−𝑘)
3
(1−𝑘)
3
(1−𝑘)
3 𝑘
Shaly-sand (1−𝑘)
3
(1−𝑘)
3
(1−𝑘)
3 𝑘
Sandy-shale 𝑘 (1−𝑘)
3
(1−𝑘)
3
(1−𝑘)
3
Shale 𝑘 (1−𝑘)
3
(1−𝑘)
3
(1−𝑘)
3
Figure 4.9: Distributions of net-to-gross ratios from sequence realizations generated
using the transition matrices in Table 4.2. Red lines indicate values of limiting sand distributions. From top left to bottom right, values of the parameter k defined in the transition matrices are 0.45, 0.55, 0.65, 0.75, 0.85 and 0.95, respectively.
Using the wavelet transform analysis, the values of slopes and intercepts decrease
when the proportions of sand in the total sequence (𝜋sand) increase. Cross-plots and
probability distributions of both attributes, color-coded by 𝜋sand for each transition
0.2 0.3 0.4 0.50
20
40
60
Net-to-gross ratios
0.2 0.3 0.4 0.50
20
40
60
Net-to-gross ratios
0.2 0.3 0.4 0.50
20
40
60
Net-to-gross ratios
0.2 0.3 0.4 0.50
20
40
60
Net-to-gross ratios0.2 0.3 0.4 0.50
20
40
60
Net-to-gross ratios
0.2 0.3 0.4 0.50
20
40
60
Net-to-gross ratios
True net-
to-gross
ratios
sand
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 127
matrix, are shown in Figure 4.10 and Figure 4.11, respectively. For a larger standard
deviation (v) in velocity distributions, both slope and intercept values for different
𝜋sand are less spread out. One explanation for the smaller spread could be that using
larger v results in larger overlaps in velocities drawn for each state. Thus, even when
𝜋sand increases (i.e., when there are more sand layers in the sequences), there is not
much change in the overall velocities drawn for each sequence because all four
lithology states yield velocity values that are very close. Larger overlaps in the slope-
intercept attributes can lead to larger uncertainties in property estimations.
Figure 4.10: Slope and intercept of wavelet transforms of seismic responses for 2
different standard deviations of velocity distributions: (left) v = 0.1 km/s and (right) v = 0.2 km/s. Points correspond to 6 sets of 200 realizations generated from different transition matrices with various net-to-gross ratios (𝝅sand). Sw is 0.1 for sand layers and 1 for the others.
Figure 4.11: Probability distributions of slope and intercept values corresponding to
data points in Figure 4.10. (Left column) v = 0.1 km/s and (right column) v = 0.2 km/s. The direction of increasing net-to-gross ratios (𝝅sand) is to the left for all plots.
2.9 2.95 3-27
-26
-25
-24
-23
-22
-21
-20
Slope
Inte
rce
pt
2.9 2.95 3-27
-26
-25
-24
-23
-22
-21
-20
Slope
Inte
rce
pt
0.32
0.35
0.38
0.42
0.45
0.48
directionof increasing sand fraction
directionof increasing sand fraction
(sand)sand fraction
2.88 2.92 2.96 30
20
40
60
80
Slope
2.88 2.92 2.96 30
20
40
60
80
Slope
-28 -26 -24 -22 -200
0.2
0.4
0.6
0.8
Intercept
-28 -26 -24 -22 -200
0.2
0.4
0.6
0.8
Intercept
0.32
0.35
0.38
0.42
0.45
0.48
Sand fraction (sand)
direction of
increasing
sand fractions
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 128
Next, we apply wavelet transform, MDS, and KPCA algorithms to a subset of 600
seismograms corresponding to sequences generated from transition matrices (Table
4.2) with values of k equal to 0.45, 0.75, and 0.95, and v = 0.1 km/s. Then, each
algorithm outputs 600 seismic-attribute vectors as new representations for the
seismograms. For example, using wavelet-transform-based analysis, the attribute
vector consists of two elements, the slope and intercept. When using MDS algorithms,
the attribute vector simply contains coordinates of a point (i.e., seismogram) in the
lower-dimensional space.
To compare how well the attributes characterize net-to-gross ratios, we count the
number of times the linear discriminant analysis algorithm (LDA) successfully sorts
the attribute vectors (i.e., the seismograms) into their corresponding net-to-gross
classes, which are defined to be less than 0.38, between 0.38 and 0.465, and greater
than 0.465. The classification success rate is estimated using a stratified 10-fold cross
validation. The validation process first partitions the attributes into ten approximately
equal-sized subsamples (i.e., folds) such that the proportions of each class in the
subsamples are approximately equal to those in the whole sample (i.e., stratified).
Then, a subsample is held as a validation set, while treating the rest as a training set in
the classification step, where a success rate is obtained. This step is repeated ten times
so that each subsample is used once as a validation set. The overall success rate is
simply an average of the ten classification experiments. To improve the accuracy of
the estimates, this stratified 10-fold cross validation is repeated ten times. The ten
success rates are then averaged to obtain the final success rate (Witten et al., 2011).
Results from MDS and KPCA algorithms are shown in Figure 4.12 – Figure 4.17
as cross-plots between just two attributes, either the first two coordinates or the first
two principal components. The classification results are summarized in Table 4.3.
Note that three numbers are listed under the success rate column, representing the
success rate when classifying samples into classes using only the first coordinate, the
second coordinate, and both coordinates, respectively.
The metric and non-metric MDS yield better classification success rates (74% and
73% respectively) than the classical MDS (56%) when considering results in the polar
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 129
coordinate system (i.e., theta and rho values; Figure 4.12 – Figure 4.14, lower right
corner). Note that in the classical MDS, the first two coordinates correspond to the two
largest eigenvalues. Among 2-attribute results, there is not much difference in success
rates among the results from the wavelet transform, metric, and non-metric MDS.
However, the success rate of each individual attribute clearly shows which attribute is
more sensitive to net-to-gross changes. For example, the intercept attribute gives a
success rate of 73% while the slope yields a success rate of 55%. The first principal
components of the Gaussian, dynamic similarity, and inverse multi-quadric KPCA
effectively capture variations in net-to-gross ratios. Using the first component alone,
classification success rates of KPCA can be as high as 85%. When using only the first
two principal components, the dynamic similarity kernel best differentiates the three
net-to-gross classes while both the polynomial and the linear KPCA poorly separate
the net-to-gross classes (Figure 4.15 – Figure 4.18).
As previously noted, since both the classical MDS with Euclidean distances and
the linear KPCA are equivalent to linear PCA, their results are similar (Figure 4.12
(upper left corner) and Figure 4.19 (right)). When plotting only the first two
coordinates or principal components, both the low and high net-to-gross classes are
shown in circular structures which are largely overlapped. As a result, both the
classical MDS and the linear KPCA (i.e., linear PCA) yield only about 50%
classification success rate. These circular structures are also observed in the results
from metric and non-metric MDS (top left corner of Figure 4.13 and Figure 4.14).
Thus, both the linear KPCA (i.e., PCA) and the Euclidean MDS fail to detect the
intrinsic circular structures in the data (Tenenbaum et al., 2000).
The inter-point distances (dij) of MDS results do not reproduce the original
dissimilarity (𝛿ij) of seismograms well if only two coordinates (i.e., two-dimensional
space) are considered, as evidenced by the correlation coefficients between the
distances and dissimilarity for all three MDS algorithms being less than 0.9 (Figure
4.12 – Figure 4.14, lower left corner). However, the correlation increases as the
number of coordinates increases. The classical MDS results have ten large eigenvalues,
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 130
which further indicates that more coordinates may need to be included in the results to
better reproduce the dissimilarity matrix and improve classification performance.
The overall classification success rate increases as the number of coordinates or
principal components included as attributes increases (MDS in Figure 4.20 and KPCA
in Figure 4.21). Among the selected kernels for KPCA, the classification success rate
of the dynamic similarity KPCA reaches 90% with just only two principal components
included; whereas, the success rates of the polynomial and the linear KPCA (i.e., PCA)
show a gradual increase over a larger number of components. The linear KPCA yields
98% success rate when using 20 principal components. Note that to achieve this
success rate of 98%, the number of components to be included as attributes may be
less than 20 because we include components in order of decreasing eigenvalues, not
their contributions to the success rate. If only those components that significantly
improve the success rate are selected and used, the rate can rapidly increase and reach
the desired level.
Using more coordinates or components as attributes involves higher-dimensional
vectors; if these attributes are later used in other tools such as neural networks, a larger
number of attributes leads to more computation and more complex models. Moreover,
it is more difficult to visualize the multivariate results. One way to visualize high-
dimensional data is to use a parallel coordinate plot, where an n-dimensional vector is
represented by a polyline connecting the value of each element in the vector. The
horizontal and vertical axes of the plot are the non-negative integer line (e.g., from 1
to n) and the real number line, respectively. The polylines, each of which corresponds
to one attribute vector (i.e., seismogram), can be color-coded by their net-to-gross
classes. Instead of plotting individual polylines, the specified quantiles of each
element can be plotted to better see the distributions of each class at a particular
element. Figure 4.22 is an example of the parallel coordinate plots, using the first five
components of the results from the Gaussian KPCA. While the large overlaps prevent
the polylines from providing much information on how each component is distributed
in each class, the plot using quantile values clearly shows that the three net-to-gross
classes have quite distinct distributions of the first component. This observation
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 131
supports the above mentioned KPCA (Gaussian kernel) result, which achieves a
classification success rate of 80% by using only the first component. Note that the
0.45- and 0.55- quantiles are selected only to get a better picture of how each
component may contribute to the classification results. However, using 0.25- and 0.75-
quantiles gives a better statistical summary of each component.
Figure 4.12: Classical MDS results for studying effect of net-to-gross ratios on
seismic signature. (Top left) Dissimilarity matrix showing pairwise Euclidean distances between any two seismograms from realizations with standard deviations of velocities equal to 0.1 km/s. (Bottom left) Correlation coefficient between dissimilarity and distance (between points in new coordinates resulting from classical MDS) versus numbers of included coordinates. (Top right) Classical MDS results with the first two coordinates. Each point represents one seismogram, color-coded by its true net-to-gross value from its corresponding sand-shale sequence. (Bottom right) Classical MDS results with the first two coordinates converted into polar coordinates.
Seismogram #
Seis
mogra
m #
200 400 600
200
400
600
Distance
0
1
2
3
4
-2 -1 0 1 2-3
-2
-1
0
1
2
First coordinate
Second c
oord
inate
0.25
0.3
0.35
0.4
0.45
0.5
Net-to-gross ratios
-4 -2 0 2 40
0.5
1
1.5
2
2.5
Rho
Theta
0.25
0.3
0.35
0.4
0.45
0.5Net-to-gross ratios
5 10 150.5
0.6
0.7
0.8
0.9
1
# of coordinates
Corr
ela
tion c
oeff
icie
nt
betw
een
dis
sim
ilarity
and d
ista
nce
-2 -1 0 1 2
-2
-1
0
1
2
3
First coordinate
Second c
oord
inate
Net-to-gross ratios
0.25
0.3
0.35
0.4
0.45
0.5
-4 -2 0 2 4
0
0.5
1
1.5
2
2.5
Theta
Rho
Net-to-gross ratios
0.25
0.3
0.35
0.4
0.45
0.5
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 132
Figure 4.13: Metric MDS results for studying effect of net-to-gross ratios on seismic
signature. (Top left) Dissimilarity matrix showing pairwise (Euclidean) distances between any two seismograms from realizations with standard deviations of velocities equal to 0.1 km/s. (Bottom left) Correlation coefficient between dissimilarity and distance (between points in new coordinates resulting from metric MDS) versus numbers of included coordinates. (Top right) Metric MDS results with the first two coordinates. Each point represents one seismogram, color-coded by its true net-to-gross value from its corresponding sand-shale sequence. (Bottom right) Metric MDS results with the first two coordinates converted into polar coordinates.
Seismogram #
Seis
mogra
m #
200 400 600
200
400
600
Distance
0
1
2
3
4
-2 -1 0 1 2-3
-2
-1
0
1
2
3
Second c
oord
inate
First coordinate
0.25
0.3
0.35
0.4
0.45
0.5Net-to-gross ratios
2 4 6 80.7
0.8
0.9
1
# of coordinates C
orr
ela
tion c
oeff
icie
nt
betw
een
dis
sim
ilarity
and d
ista
nce
-4 -2 0 2 40
0.5
1
1.5
2
2.5
3
Theta
Rho
0.25
0.3
0.35
0.4
0.45
0.5
Net-to-gross ratios
-2 0 2-3
-2
-1
0
1
2
3
First coordinate
Second c
oord
inate
Net-to-gross ratios
0.25
0.3
0.35
0.4
0.45
0.5
-4 -2 0 2 40
0.5
1
1.5
2
2.5
3
Theta
Rho
Net-to-gross ratios
0.25
0.3
0.35
0.4
0.45
0.5
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 133
Figure 4.14: Non-metric MDS results for studying effect of net-to-gross ratios on
seismic signature. (Top left) Dissimilarity matrix: pairwise (Euclidean) distances between any two seismograms from realizations with standard deviations of velocities equal to 0.1 km/s. (Bottom left) Correlation coefficient between dissimilarity and distance (between points in new coordinates resulting from non-metric MDS) versus numbers of included coordinates. (Top right) Non-metric-MDS results with the first two coordinates. Each point represents one seismogram, color-coded by its true net-to-gross value from its corresponding sand-shale sequence. (Bottom right) Non-metric MDS results with the first two coordinates converted into polar coordinates.
Seismogram #
Seis
mogra
m #
200 400 600
200
400
600
Distance
0
1
2
3
4
-2 -1 0 1 2-2
-1
0
1
2
3
Second c
oord
inate
First coordinate
0.25
0.3
0.35
0.4
0.45
0.5Net-to-gross ratios
2 4 6 80.7
0.8
0.9
1
# of coordinatesCorr
ela
tion c
oeff
icie
nt
betw
een
dis
sim
ilarity
and d
ista
nce
-4 -2 0 2 40
0.5
1
1.5
2
2.5
3
Theta
Rho
0.25
0.3
0.35
0.4
0.45
0.5
Net-to-gross ratios
-2 -1 0 1 2-2
-1
0
1
2
3
First coordinate
Second c
oord
inate
Net-to-gross ratios
0.25
0.3
0.35
0.4
0.45
0.5
-4 -2 0 2 4
0
0.5
1
1.5
2
2.5
3
Theta
Rho
Net-to-gross ratios
0.25
0.3
0.35
0.4
0.45
0.5
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 134
Figure 4.15: Gaussian KPCA results for studying effect of net-to-gross ratios on
seismic signature. (Left) Kernel matrix using a Gaussian kernel. Each element in the matrix (K(xi,xj)) corresponds to the Gaussian kernel function evaluated using a pair of seismograms (xi,xj) from realizations with standard deviations of velocities equal to 0.1 km/s. (Right) Projections of seismograms, which correspond to sequences with different net-to-gross ratios, onto the first two principal components from the Gaussian KPCA. Each point is color-coded by the net-to-gross ratio from its corresponding sand-shale sequence.
Figure 4.16: Dynamic similarity KPCA results for studying effect of net-to-gross
ratios on seismic signature. (Left) Kernel matrix using a dynamic similarity kernel. Each element in the matrix (K(xi,xj)) corresponds to the dynamic similarity kernel function evaluated using a pair of seismograms (xi,xj) from realizations with standard deviations of velocities equal to 0.1 km/s. (Right) Projections of seismograms, which correspond to sequences with different net-to-gross ratios, onto the first two principal components from the dynamic similarity KPCA. Each point is color-coded by the net-to-gross ratio from its corresponding sand-shale sequence.
Seismogram #
Seis
mogra
m #
200 400 600
100
200
300
400
500
600
K(xi,xj)
0.2
0.4
0.6
0.8
1
-0.4 -0.2 0 0.2 0.4 0.6-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
First principal component
Second p
rincip
al com
ponent
0.25
0.3
0.35
0.4
0.45
0.5Net-to-gross ratios
Seismogram #
Seis
mogra
m #
K (xi,xj)
200 400 600
100
200
300
400
500
600
-0.5 0 0.5
-0.4
-0.2
0
0.2
0.4
0.6
First principal component
Second p
rincip
al com
ponent
0.2
0.4
0.6
0.8
1
0.25
0.3
0.35
0.4
0.45
0.5Net-to-gross ratios
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 135
Figure 4.17: Inverse multi-quadric KPCA results for studying effect of net-to-gross
ratios on seismic signature. (Left) Kernel matrix using an inverse multi-quadric kernel. Each element in the matrix (K(xi,xj)) corresponds to the inverse multi-quadric kernel function evaluated using a pair of seismograms (xi,xj) from realizations with standard deviations of velocities equal to 0.1 km/s. (Right) Projections of seismograms, which correspond to sequences with different net-to-gross ratios, onto the first two principal components from the inverse multi-quadric KPCA. Each point is color-coded by the net-to-gross ratio from its corresponding sand-shale sequence
Figure 4.18: Polynomial KPCA results for studying effect of net-to-gross ratios on
seismic signature. (Left) Kernel matrix using a polynomial kernel. Each element in the matrix (K(xi,xj)) corresponds to the polynomial kernel function evaluated using a pair of seismograms (xi,xj) from realizations with standard deviations of velocities equal to 0.1 km/s. (Right) Projections of seismograms, which correspond to sequences with different net-to-gross ratios, onto the first two principal components from the polynomial KPCA. Each point is color-coded by the net-to-gross ratio from its corresponding sand-shale sequence.
Seismogram #
Seis
mogra
m #
200 400 600
200
400
600
-0.5 0 0.5-0.5
0
0.5
First principal component
Second p
rincip
al com
ponent
0
0.5
1
0.25
0.3
0.35
0.4
0.45
0.5
K(xi,xj)
Net-to-gross ratios
Seismogram #
Seis
mogra
m #
K (xi,xj)
200 400 600
200
400
600
-2 0 2 4-4
-3
-2
-1
0
1
2
First principal component
Second p
rincip
al com
ponent
2
4
6
8
10
0.25
0.3
0.35
0.4
0.45
0.5
Net-to-gross ratios
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 136
Figure 4.19: Linear KPCA results for studying effect of net-to-gross ratios on seismic
signature. (Left) Kernel matrix using a linear kernel. Each element in the matrix (K(xi,xj)) corresponds to the linear kernel function evaluated using a pair of seismograms (xi,xj) from realizations with standard deviations of velocities equal to 0.1 km/s. (Right) Projections of seismograms, which correspond to sequences with different net-to-gross ratios, onto the first two principal components from the linear KPCA. Each point is color-coded by the net-to-gross ratio from its corresponding sand-shale sequence.
Figure 4.20: Change in MDS classification success rate when the number of
coordinates included as net-to-gross attributes increases. Results from three MDS methods are shown.
seismogram #
seis
mogra
m #
200 400 600
100
200
300
400
500
600-2
0
2
4
6
8
-2 -1 0 1 2-3
-2
-1
0
1
2
First principal component
Second p
rincip
al com
ponent
0.25
0.3
0.35
0.4
0.45
0.5
K(xi,xj)
Net-to-gross ratios
2 4 6 8 10 12
0.4
0.5
0.6
0.7
0.8
0.9
1
# of coordinates included
Cla
ssific
ation s
uccess r
ate
classical MDS
Metric MDS
Non-metric MDS
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 137
Table 4.3: Summary of the methods used to compute seismic attributes for net-to-
gross estimation. Choices of parameters for each method are also included. Performance of each method is shown as a success rate in classifying a data point into three net-to-gross classes: <0.38, 0.38-0.465, and >0.465
Method Compared
objects
Distance/ or
Kernel
Classification
success rate
(using 1st, 2nd,
both
coordinates/com
ponents)
Notes
Wavelet-
transform
– – 0.55, 0.73, 0.72
Slope/intercept/both
MDS
(classical)
Seismograms Euclidean 0.37, 0.49, 0.53
0.50, 0.49, 0.55
(theta/rho/both)
> ten large eigenvalues;
the algorithm cannot
reproduce dissimilarity
well by using only the
first two coordinates
MDS
(metric)
Seismograms Euclidean 0.38,0.43,0.45
0.45, 0.69, 0.73
(theta/rho/both)
Correlation coefficient
between distance and
dissimilarity is 0.83 when
using the first two
coordinates; the value
becomes greater than
0.95 when there are more
than 4 coordinates.
MDS
(non-
metric)
Seismograms Euclidean 0.38,0.43,0.45
0.47, 0.68, 0.72
(theta/rho/both)
Correlation coefficient
between distance and
dissimilarity is 0.88 when
using the first two
coordinates; the value
becomes greater than
0.95 when there are more
than 3 coordinates.
KPCA Seismograms Gaussian 0.81, 0.35, 0.81 σ2 = 0.5
KPCA Seismograms Dynamic
similarity
0.85, 0.58, 0.90 σ = 0.1265, m = 80% of
total seismogram length,
r = 2
KPCA Seismograms Inverse
multi-
quadric
0.79, 0.35, 0.79 c2 = 1
KPCA Seismograms Polynomial 0.50, 0.51,
0.59
c = 1, d = 2
KPCA Seismograms Linear 0.37, 0.49, 0.53 equivalent to PCA
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 138
Figure 4.21: Change in KPCA classification success rate when the number of
principal components included as net-to-gross attributes increases.
Figure 4.22: Parallel coordinate plot. (Left) Polylines of the first five principal
components from the Gaussian KPCA. Each line is color-coded by its corresponding net-to-gross class. Three classes are <0.38, 0.38-0.465, and >0.465. (Right) The solid lines are the median (i.e., the 0.5-quantile) of the component values. The dash lines surrounding the median are the 0.45- and 0.55- quantiles.
4.4.3 Scenario 2: Effect of saturation
We use sequences generated from a similar transition matrix, but with varying
saturation values. In this scenario, we show only simulations for velocity distributions
with standard deviation of 0.1 km/s. We use Gassmann’s equation to substitute
mixtures of water and oil with the desired saturations (Sw = 0.1, 0.5, and 1) into only
the sand layers in the sequences. All other lithologies always have Sw equal to 1.
0 10 20 30 40 50
0.4
0.5
0.6
0.7
0.8
0.9
1
# of principal components
Cla
ssific
ation s
uccess r
ate
Gaussian KPCA
dynamic similarity KPCA
inverse multi-quadric KPCA
polynomial KPCA
linear KPCA
1 2 3 4 5
-0.4
-0.2
0
0.2
0.4
0.6
Coordinate
Coord
inate
Valu
e
1 2 3 4 5
-0.4
-0.2
0
0.2
0.4
0.6
Coordinate
Coord
inate
Valu
e
Class 1
Class 2
Class 3
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 139
We consider three matrices (A, B, and C) which have the same limiting
distribution: [0.45 0.05 0.05 0.45] (Figure 4.23). Long sequences generated using the
three matrices should have their net-to-gross ratios within the sequence equal to 0.45;
however, in our simulation the values spread around 0.45. Therefore, to separate the
net-to-gross effect from the saturation effect, we select only those sequences with net-
to-gross ratios ranging from 0.4 – 0.5. Thus, the main difference among sequences
generated from the three transition matrices is how the sand layers are distributed. For
example, sand layers in sequences from matrix C are more clustered and blocky than
those from the other two matrices. This blocky pattern is expected since matrix C has
large probabilities of going to the sand and shale states (pss = psh-s,s = ps-sh ,sh = pshsh =
0.85).
Figure 4.23: Three selected matrices with the same limiting distribution and their
sample sequences. From left to right, sand layers become more blocky (i.e., groups of sand layers become thicker). The lithologic states are sand (s), shaly sand (sh-s) sandy shale (s-sh) and shale (sh). Note that each row of matrix B is equal to a fixed probability vector, and thus the lithologic states generated using this matrix are considered as independent random events (i.e., the current state has no dependency on the previous states).
Results from wavelet-transform, MDS, and KPCA algorithms are color-coded by
their corresponding saturation values (i.e., 0.1, 0.5, and 1) and shown in Figure 4.24 –
Figure 4.30 as cross-plots between just two attributes, either the first two coordinates
Sh S Sh S Sh S
(A) .05 .05 .05 .85
.05 .05 .05 .85
.85 .05 .05 .05
.85 .05 .05 .05
(B) .45 .05 .05 .45
.45 .05 .05 .45
.45 .05 .05 .45
.45 .05 .05 .45
(C) .85 .05 .05 .05
.85 .05 .05 .05
.05 .05 .05 .85
.05 .05 .05 .85
(A) (B) (C)
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 140
or the first two components. Results are then used in the same cross-validation scheme
described in Section 4.4.2. The classification success rates are summarized in Table
4.4. Note that since the linear KPCA (i.e., PCA) are equivalent to classical MDS with
Euclidean distances. Thereafter, when Euclidean distances are used, we only show the
results from classical MDS.
For wavelet-transform-based attributes, slope and intercept of sequences obtained
from the same transition matrix do not show any clear trend with increasing water
saturation. Their distributions largely overlap, resulting in classification rates less than
65%. However, for the same water saturation slope and intercepts values increase as
the transition matrices change from A to C, i.e., as the sands in the sequences become
more clustered and blocky (Figure 4.24). Thus, spatial statistics at the sub-seismic
scale also have an influence on seismic signatures of sub-resolution systems.
For the same transition matrix, plots of the first two coordinates from classical
MDS and the two coordinates from metric and non-metric MDS in Figure 4.25 –
Figure 4.27 show similar patterns, resulting in similar classification success rates with
the rates of classical MDS being slightly smaller than metric and non-metric MDS. In
matrix A results, all MDS algorithms effectively separate the sequences which have
sand layers fully saturated with water (Sw = 1) from the other sequences relatively well.
The other two saturation classes are almost inseparable, because their distributions
largely overlap. In matrix B results, the water-saturated sequences are still well
separated by all MDS algorithms; however, in this case the saturation class of 0.1
starts to separate out and surround the saturation class of 0.5 with less overlap than the
results in matrix A. In matrix C, both the saturation classes of 0.1 and 1 surround the
saturation class of 0.5. The overall classification success rates using two-coordinate
MDS results are less than 70%.
As with the net-to-gross case, if only two coordinates are considered, the inter-
point distances of MDS results do not reproduce the original dissimilarity of
seismograms well, with the correlation coefficients between the distances and
dissimilarity for metric and non-metric MDS algorithms being less than 0.9 (Figure
4.33). The correlation increases as the number of coordinates increases, with
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 141
correlations of matrix C being the lowest for a given number of coordinates. For
classical MDS results, there are more than two large eigenvalues, especially the
eigenvalues in matrix C whose values are almost twice as large as those in matrix A
and B (Figure 4.32). More coordinates can be included as attributes to better reproduce
the dissimilarity matrix and improve classification performance. When the number of
coordinates increases, the overall classification success rate also increases. For the
same transition matrix and same number of coordinates, the success rates from
classical, metric, and non-metric MDS turn out to be very similar (Figure 4.34).
By using the first two principal components, the dynamic similarity KPCA best
differentiates the three saturation classes for all transition matrices (Figure 4.29). Its
classification success rates are greater than 80%, whereas the other kernels have
poorer performance. From Figure 4.28, the first principal components of the Gaussian
KPCA for both the transition matrices A and B separate the sequences which have
sand layers fully saturated with water (Sw = 1) from the other sequences relatively well.
Matrix C shows a similar result, except that the class which is effectively separated by
using the first component is that with Sw equal to 0.5. For all three transition matrices,
the second component does not contribute much in differentiating the three saturation
classes, and the overall classification success rates of the Gaussian KPCA are only in
the 60-70 percent range.
Both the inverse multi-quadric KPCA in Figure 4.30 and the polynomial KPCA in
Figure 4.31 show patterns quite similar to the Gaussian KPCA, in which the water-
saturated sequences (Sw = 1) seem distinguishable from the other classes by just the
first component for both matrices A and B; however, in matrix C this saturation class
(Sw = 1) is less separable from the rest. The classification success rates using the first
two principal components of the inverse multi-quadric KPCA are generally in the 60-
70 percent range, which is slightly higher than the 50-65 percent range of the
polynomial KPCA.
One way to improve the performance of KPCA attributes is to increase the
numbers of components included as attributes. Figure 4.35 – Figure 4.37 illustrate the
changes in classification success rates with increasing numbers of included
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 142
components. Among the selected kernels for KPCA, the classification success rate of
the dynamic similarity KPCA reaches 80% with just only two principal components
included; whereas, the success rates of the classical MDS (i.e., linear KPCA or PCA)
show a gradual increase over a larger number of components. The linear KPCA yields
almost 100% success rate when using 20 or more principal components. Note that the
components are included as attributes in order of decreasing eigenvalues, not their
contributions to the success rate. If only those components that can effectively
distinguish different saturation classes are selected, with a smaller number of
components the rate can sometimes be very close to the success rate obtained when
using a larger numbers of components. A small number of components also make it
possible to visualize the results. Selecting those components with higher
discriminating power can be done by using parallel coordinate plots (Figure 4.38 –
Figure 4.39). For example, from the parallel coordinate plots of the first ten
components of Gaussian KPCA results for matrix B, we select the first and sixth
components and they yield a classification success rate of 80%, while the success rate
when using all ten components is 82% (Figure 4.38). As another example, the success
rate when using the first and sixth components of inverse multi-quadric KPCA results
for matrix C is 73%, while the success rate when using the first ten components is 82%
(Figure 4.39). Here our selection process is heuristic-based, selecting any two
components that exhibit large separation among the three classes. However, this
selection scheme does not necessarily work in all cases, especially when several
components behave similarly. Therefore, a systematic comparison of success rates
may be needed in order to select the best possible combination of components.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 143
Figure 4.24: Slope and intercept attributes for varying water saturation within the
sand layers. Results in each plot are from sequences generated by using the transition matrix shown in the lower-right corner of each plot. From left to right, the transition matrices correspond to A, B, and C in Figure 4.23.
Figure 4.25: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the first two coordinates from classical MDS. Each point is color-coded by the water-saturation value of the sand layers. From left to right, subplots correspond to transition matrices A, B, and C, respectively.
Figure 4.26: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the two coordinates from metric MDS. Each point is color-coded by the water-saturation value of the sand layers. From left to right, subplots correspond to transition matrices A, B, and C, respectively.
2.9 2.92 2.94 2.96 2.98 3-27
-26
-25
-24
-23
-22
-21
-20
Slope
Inte
rcept
2.9 2.92 2.94 2.96 2.98 3-27
-26
-25
-24
-23
-22
-21
-20
Slope
Inte
rcept
2.9 2.92 2.94 2.96 2.98 3-27
-26
-25
-24
-23
-22
-21
-20
Slope
Inte
rcept
sw=0.1
sw=0.5
sw=1
.05 .05 .05 .85
.05 .05 .05 .85
.85 .05 .05 .05
.85 .05 .05 .05
.45 .05 .05 .45
.45 .05 .05 .45
.45 .05 .05 .45
.45 .05 .05 .45
.85 .05 .05 .05
.85 .05 .05 .05
.05 .05 .05 .85
.05 .05 .05 .85
-2 -1 0 1 2-1.5
-1
-0.5
0
0.5
1
1.5
Second c
oord
inate
First coordinate
-2 -1 0 1 2-1.5
-1
-0.5
0
0.5
1
1.5
Second c
oord
inate
First coordinate-4 -2 0 2 4
-2
-1
0
1
2
Second c
oord
inate
First coordinate
sw = 0.1
sw = 0.5
sw = 1
-2 -1 0 1 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Second c
oord
inate
First coordinate-3 -2 -1 0 1 2
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Second c
oord
inate
First coordinate-4 -2 0 2 4
-3
-2
-1
0
1
2
3
Second c
oord
inate
First coordinate
sw=0.1
sw=0.5
sw=1
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 144
Figure 4.27: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the two coordinates from non-metric MDS. Each point is color-coded by the water-saturation value of the sand layers. From left to right, subplots correspond to transition matrices A, B, and C, respectively.
Figure 4.28: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the first two principal components from the Gaussian KPCA. Each point is color-coded by the water-saturation value of the sand layers. From left to right, subplots correspond to transition matrices A, B, and C, respectively.
Figure 4.29: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the first two principal components from the dynamic similarity KPCA. Each point is color-coded by the water-saturation value of the sand layers. From left to right, subplots correspond to transition matrices A, B, and C, respectively.
-2 -1 0 1 2-1.5
-1
-0.5
0
0.5
1
1.5S
econd c
oord
inate
First coordinate
-2 -1 0 1 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Second c
oord
inate
First coordinate-4 -2 0 2 4
-3
-2
-1
0
1
2
3
Second c
oord
inate
First coordinate
sw=0.1
sw=0.5
sw=1
-0.5 0 0.5-0.5
0
0.5
Second p
rincip
al com
ponent
First principal component
-0.5 0 0.5-0.5
0
0.5
Second p
rincip
al com
ponent
First principal component
-0.6 -0.4 -0.2 0 0.2 0.4-0.5
0
0.5
Second p
rincip
al com
ponent
First principal component
sw=0.1
sw=0.5
sw=1
-1 -0.5 0 0.5 1-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Second p
rincip
al com
ponent
First principal component
-1 -0.5 0 0.5 1-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Second p
rincip
al com
ponent
First principal component-1 -0.5 0 0.5 1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Second p
rincip
al com
ponent
First principal component
sw=0.1
sw=0.5
sw=1
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 145
Figure 4.30: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the first two principal components from the inverse multi-quadric KPCA. Each point is color-coded by the water-saturation value of the sand layers. From left to right, subplots correspond to transition matrices A, B, and C, respectively.
Figure 4.31: Projections of seismograms, which correspond to sequences with
different water-saturation values, onto the first two principal components from the polynomial KPCA. Each point is color-coded by the water-saturation value of the sand layers. From left to right, subplots correspond to transition matrices A, B, and C, respectively.
Figure 4.32: First 21 eigenvalues from classical MDS, color-coded by the transition
matrices used in generating sand-shale sequences for investigating the effect of saturations.
-0.6 -0.4 -0.2 0 0.2 0.4-0.5
0
0.5
Second p
rincip
al com
ponent
First principal component-0.4 -0.2 0 0.2 0.4 0.6
-0.5
0
0.5
Second p
rincip
al com
ponent
First principal component-0.5 0 0.5
-0.5
0
0.5
Second p
rincip
al com
ponent
First principal component
sw=0.1
sw=0.5
sw=1
-2 -1 0 1 2 3-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
Second p
rincip
al com
ponent
First principal component-4 -2 0 2
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
Second p
rincip
al com
ponent
First principal component-5 0 5
-5
-4
-3
-2
-1
0
1
2
3
4
Second p
rincip
al com
ponent
First principal component
sw=0.1
sw=0.5
sw=1
1 5 9 13 17 210
100
200
300
400
Eigenvalue
Matrix A
Matrix B
Matrix C
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 146
Figure 4.33: Correlation coefficient between dissimilarity and distance (between
points in new coordinates resulting from an MDS algorithm) versus numbers of included coordinates from MDS results for investigating the effect of saturations. (Left) metric MDS and (right) non-metric MDS.
Figure 4.34: Change in MDS classification success rate when the number of
coordinates included as saturation attributes increases. Three MDS algorithms are compared. Results when using sequences generated by transition matrices A (left), B (middle) and C (right) are shown.
Figure 4.35: Change in KPCA classification success rate as the number of principal
components included as attributes increases, when using sequences generated by transition matrices A.
1 2 3 4 5 6 7
0.7
0.8
0.9
1
# of coordinates
Corr
ela
tion c
oeff
icie
nt
betw
een
dis
sim
ilarity
and d
ista
nce
Matrix A
Matrix B
Matrix C
1 2 3 4 5 6 7
0.7
0.8
0.9
1
# of coordinatesCorr
ela
tion c
oeff
icie
nt
betw
een
dis
sim
ilarity
and d
ista
nce
Matrix A
Matrix B
Matrix C
2 4 6 8 10 12
0.5
0.6
0.7
0.8
0.9
1
# of coordinates included
Cla
ssific
ation s
uccess r
ate
2 4 6 8 10 12
0.5
0.6
0.7
0.8
0.9
1
# of coordinates included
Cla
ssific
ation s
uccess r
ate
2 4 6 8 10 12
0.5
0.6
0.7
0.8
0.9
1
# of coordinates included
Cla
ssific
ation s
uccess r
ate
Classical MDS
Metric MDS
Non-metric MDS
Matrix A Matrix B Matrix C
0 10 20 30 40 50 60 70 80 90
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
# of principal components or coordinates
Cla
ssific
ation s
uccess r
ate
Transition matrix A
Gaussian KPCA
dynamic similarity KPCA
inverse multi-quadric KPCA
polynomial KPCA
classical MDS (or linear KPCA)
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 147
Figure 4.36: Change in KPCA classification success rate as the number of principal
components included as attributes increases, when using sequences generated by transition matrices B.
Figure 4.37: Change in KPCA classification success rate as the number of principal
components included as attributes increases, when using sequences generated by transition matrices C.
0 10 20 30 40 50 60 70 80 90
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
# of principal components or coordinates
Cla
ssific
ation s
uccess r
ate
Transition matrix B
Gaussian KPCA
dynamic similarity KPCA
inverse multi-quadric KPCA
polynomial KPCA
classical MDS (or linear KPCA)
0 10 20 30 40 50 60 70 80 900.4
0.5
0.6
0.7
0.8
0.9
1
# of principal components or coordinates
Cla
ssific
ation s
uccess r
ate
Transition matrix C
Gaussian KPCA
dynamic similarity KPCA
inverse multi-quadric KPCA
polynomial KPCA
classical MDS (or linear KPCA)
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 148
Figure 4.38: Component-wise analysis for the Gaussian KPCA. (Left) Parallel
coordinate plot of the first ten principal components of the Gaussian KPCA for matrix B. The solid lines are the median (the 0.5 quantile). The dash lines surrounding the median are the 0.45 and 0.55 quantiles. (Right) Projections of seismograms onto the first and sixth principal components of the Gaussian KPCA.
Figure 4.39: Component-wise analysis for the inverse multi-quadric KPCA. (Left)
Parallel coordinate plot of the first ten principal components of the inverse multi-quadric KPCA for matrix C. The solid lines are the median (the 0.5 quantile). The dash lines surrounding the median are the 0.45 and 0.55 quantiles. (Right) Projections of seismograms onto the first and sixth principal components of the inverse multi-quadric KPCA.
1 2 3 4 5 6 7 8 9 10-0.6
-0.4
-0.2
0
0.2
0.4
Coordinate
Coord
inate
Valu
e
sw=0.1
sw=0.5
sw=1
-0.5 0 0.5
-0.5
0
0.5
First principal component
Six
th p
rincip
al com
ponent
Matrix B
1 2 3 4 5 6 7 8 9 10-0.3
-0.2
-0.1
0
0.1
0.2
Coordinate
Coord
inate
Valu
e
-0.4 -0.2 0 0.2 0.4 0.6
-0.4
-0.2
0
0.2
0.4
0.6
First principal component
Six
th p
rincip
al com
ponent
sw=0.1
sw=0.5
sw=1 Matrix C
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 149
Table 4.4: Summary of the methods used to compute seismic attributes for water-saturation (in sand layers) effect. Choices of parameters for each method are similar to those listed in Table 4.2, unless otherwise specified. Performance of each method is shown as a success rate in classifying a data point into three water-saturation categories: Sw=0.1, 0.5, and 1. Results are shown in the columns corresponding to the transition matrices used in simulations. The actual transition matrices and sample sequences are shown in Figure 4.23.
Method Classification success rate
(using 1st, 2nd, both coordinates/components)
Notes
Transition
matrix A
(Aggrading)
Transition
matrix B
Transition
matrix C
(Blocky-sand)
Wavelet-
transform
0.54, 0.49, 0.57 0.59, 0.52,
0.65
0.42, 0.54, 0.53 Slope/intercept/both
MDS
(classical)
0.66, 0.33, 0.65 0.66, 0.33,
0.66
0.49, 0.42, 0.53 See Figure 4.32 for
eigenvalues
equivalent to KPCA
with a linear kernel or
PCA
MDS
(metric)
0.65, 0.36, 0.65 0.67, 0.31,
0.67
0.52, 0.43, 0.57
See Figure 4.33 for
correlation between
dissimilarity and
distance
MDS
(non-metric)
0.66, 0.35, 0.66 0.68, 0.33,
0.69
0.53, 0.44, 0.58 See Figure 4.33 for
correlation between
dissimilarity and
distance
KPCA
(Gaussian)
0.64, 0.32, 0.62
(1&3) 0.73
(1&3&8) 0.80
(1to10) 0.80
0.73, 0.35,
0.73
(1&6) 0.80
(1to10) 0.82
0.60, 0.36, 0.61
(1&5) 0.69
(1to10) 0.74
𝜎2 = 0.5
Trials with other
components
KPCA
(Dynamic
similarity)
0.71, 0.76, 0.88
(1to10) 0.89
0.70, 0.73,
0.87
(1to10) 0.89
0.81, 0.54, 0.84
(1to10) 0.90
𝜎 = 0.0398, m = 80%
of total seismogram
length, r = 4
KPCA
(Inverse multi-
quadric)
0.65, 0.32, 0.64
(1&3) 0.71
(1&3&7) 0.77
(1to10) 0.81
0.68, 0.34,
0.67
(1&7) 0.78
(1to10) 0.82
0.54, 0.42, 0.60
(1&6) 0.73
(1to10) 0.82
c2 = 1
Trials with other
components
KPCA
(Polynomial)
0.68, 0.32, 0.65
(1&7) 0.79
(1&3&7) 0.80
(1to10) 0.80
0.66, 0.35,
0.65
(1&8) 0.76
(1to10) 0.77
0.49, 0.43,
0.52
(1&8) 0.60
(1&10) 0.67
c = 1 and d = 2
Trials with other
components
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 150
4.4.4 Scenario 3: Effect of stacking patterns
We select transition matrices to mimic three stacking patterns within
retrogradational, progradational or aggradational parasequence sets while keeping net-
to-gross ratios and saturations the same. The three transition matrices shown in Figure
4.40 have the same limiting distributions: [0.25 0.25 0.25 0.25]. Note that each row of
matrix C is equal to a fixed probability vector. When using this matrix to generate a
sequence, the resulting lithologic states are considered as independent random events
(i.e., the current state has no dependency on the previous states). From the limiting
distribution, long sequences generated using these matrices should have their net-to-
gross ratios within the sequence equal to 0.25; however, in our simulation the values
spread around 0.25. Therefore, to separate the net-to-gross effect from the saturation
effect, we select only sequences with net-to-gross ratios ranging from 0.2 – 0.3. Water
saturation (Sw) is set to be the same in all simulations: Sw=0.1 for sand layers and
Sw=1 for the other lithology states). In this case, we show only simulations for
velocity distributions with standard deviation of 0.1 km/s.
Figure 4.40: Sample sequences from three selected transition matrices with the same
limiting distribution: [.25 .25 .25 .25]. From left to right, columns represent retrogradational, progradational, and aggradational stacking patterns, respectively. Red arrows schematically indicate transitions from coarse to fine grains, and vice versa. Note that other interpretations of transitional patterns are possible. The lithologic states are sand (s), shaly sand (sh-s) sandy shale (s-sh) and shale (sh).
(A) .183 .45 .183 .183
.183 .183 .45 .183
.183 .183 .183 .45
.45 .183 .183 .183
(B) .183 .183 .183 .45
.45 .183 .183 .183
.183 .45 .183 .183
.183 .183 .45 .183
(C) .25 .25 .25 .25
.25 .25 .25 .25
.25 .25 .25 .25
.25 .25 .25 .25
Sh S Sh S Sh S
(A) (B) (C)
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 151
To see how well the seismic attributes can differentiate the stacking patterns,
results from wavelet-transform, MDS, and KPCA are used in the cross-validation
scheme described in Section 4.4.2. Three classes considered are retrogradational,
progradational, and aggradational stacking patterns. The classification success rates
are summarized in Table 4.5. When using wavelet-transform attributes (i.e., slope and
intercept), the success rate is only 36%, which is only slightly better than the chance of
guessing the correct class out of three classes (33.33%). The classical, metric, non-
metric MDS, and KPCA with all tested kernels (Gaussian, dynamic similarity, inverse
multi-quadric, and polynomial kernels) yield similar success rates. Even when ten
coordinates or components are included, the success rates are still under 40%. A
parallel coordinate plot of the first ten components of the Gaussian KPCA is shown in
Figure 4.41 as an example to illustrate how each component from the three classes are
largely overlapped each other, causing the low success rate.
Figure 4.41: Parallel coordinate plot of the first ten components of the Gaussian
KPCA result for investigating the effect of stacking pattern on seismic signatures. The solid lines are the median (i.e., the 0.5-quantile) of the component values. The dash lines surrounding the median are the 0.45- and 0.55- quantiles. All lines are color-coded by the types of stacking patterns.
1 2 3 4 5 6 7 8 9 10-0.4
-0.2
0
0.2
0.4
0.6
Coordinate
Coord
inate
Valu
e
(A) Retrograding
(B) Prograding
(C) Aggrading
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 152
Table 4.5: Summary of the methods used to compute seismic attributes for stacking-pattern effects using sequences generated from the fixed-sampling transition matrices. Choices of parameters for each method are also included. Performance of each method is shown as a success rate in classifying a data point into three stacking patterns: retrogradational, progradational, and aggradational patterns.
Method Compared
objects
Distance/ or
Kernel
Classification
success rate
(using 1st, 2nd,
both coordinates)
Notes
Wavelet-
transform
– – 0.34, 0.34, 0.36
Slope/intercept/both
MDS
(classical)
Seismograms Euclidean 0.37, 0.35, 0.36
0.37,0.36,
0.39,0.37
1,2,1&2
(using 1,2,5,10
coordinates)
equivalent to KPCA with
a linear kernel or PCA
MDS
(metric)
Seismograms Euclidean 0.34,0.36,0.39,
0.37
(using 1,2,5,10
coordinates)
MDS
(non-
metric)
Seismograms Euclidean 0.34,0.36,0.39,
0.37
(using 1,2,5,10
coordinates)
KPCA Seismograms Gaussian 0.35,0.35,0.39,
0.39
(using 1,2,5,10
components)
σ2 = 0.5
KPCA Seismograms Dynamic
similarity
0.34, 0.33, 0.36,
0.39
(using 1,2,5,10
components)
σ = 0.1265, m = 80% of
total seismogram length,
r = 2
KPCA Seismograms Inverse
multi-
quadric
0.37,0.36,0.39,
0.39
(using 1,2,5,10
components)
c2 = 1
KPCA Seismograms Polynomial 0.37, 0.37, 0.39,
0.39
c = 1, d = 2
In the above study of stacking-pattern effect on seismic signatures, sequences are
generated following the method of fixed-sampling transition matrices, with a step size
of 0.5 m (Figure 4.40). However, these sequences do not exhibit any thinning- or
thickening-upward trends of the sand layers that are observed in fining- and
coarsening-upward parasequences and parasequence sets (Section 4.3.1). Adjusting
the transition probabilities in the fixed-sampling transition matrices such that the sand
layers become thicker or thinner is possible. However, using only one transition
matrix to generate the entire sequence (i.e., stationary Markov chain model) makes the
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 153
thicknesses of sand layers approximately the same throughout the sequences.
Therefore, one way to incorporate thinning- or thickening-upward trends into the
synthetic sequences is to vary the transition matrices along the sequence. If the
embedded-form transition matrix is used, an alternative way to generate these trends is
to vary parameters of thickness distributions along the sequence (e.g., average values
of lithologic thicknesses that are exponentially distributed).
We create a new set of sequences with the three stacking patterns (e.g.,
retrogradational, progradational and aggradational) by using the embedded-form
transition matrices with varying the average thicknesses along the sequences to
capture the thickening- and thinning-upward trends. For example, we create the
retrogradational pattern by using the transition matrix shown in Figure 4.42 (top left)
and varying the average thicknesses for sand from 0.05 m at the top to 4 m at the
bottom of sequences. Examples of sequences for all three stacking patterns are shown
in Figure 4.42. The total thicknesses of all sequences are between 120 and 130 m. To
separate out the net-to-gross and saturation effects, we select only sequences with net-
to-gross ratios ranging from 0.37 to 0.42 and set water saturations (Sw) to be the same
in all simulations: Sw=0.1 for sand layers and Sw=1 for the other lithology states.
Velocity distributions for lithologic states are assumed to be normally distributed with
standard deviations of 0.1 km/s.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 154
Figure 4.42: Sample sequences using the embedded-form transition matrices shown
on the left with varying averages of exponentially-distributed lithologic thicknesses along the sequences. From (A) to (C), the three stacking patterns are retrogradational (overall thinning- and fining-upward), progradational (overall thickening- and coarsening-upward), and aggradational stacking patterns in parasequence sets, respectively. Red arrows schematically show a series of progradational parasequences within retrogradational, progradational, and aggradational parasequence sets. Note that other interpretations of such parasequence patterns are possible The lithologic states are sand (s), shaly sand (sh-s) sandy shale (s-sh) and shale (sh).
Results from wavelet-transform, MDS, and KPCA are used in the cross-validation
scheme with the three stacking-pattern classes: retrogradational, progradational, and
aggradational. The classification success rates are summarized in Table 4.6. Using
wavelet-transform attributes (i.e., slope and intercept) gives the success rate only 34%,
which is approximately the chance of guessing the correct class out of three classes
(33.33%). However, classical, metric, non-metric MDS, and KPCA with both
Gaussian and inverse multi-quadric kernels yield higher success rates, with the
numbers increasing up to 70% when ten coordinates or components are included.
Since the distance used for classical MDS is not Euclidean but city-block distance,
classical MDS is not equivalent to the linear KPCA. Similar to the other scenarios,
Figure 4.43 illustrates that when the numbers of components included as attributes
Sh S
0
121
Sh S
0
124.2Sh S
0
123
Depth
fro
m a
refe
rence p
oin
t (m
)
(B) (C)0 .2 .3 .5
.5 0 .2 .3
.3 .5 0 .2
.2 .3 .5 0
Transition matrix
Stacking patterns of
parasequence sets:
( A) Retrogradational
(B) Progradational
( C) Aggradational
S: sand
Sh: shale
(A)
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 155
increase, classification success rates also increase. Among the selected kernels for
KPCA, the classification success rate of the dynamic similarity KPCA reaches a
higher success rate with just only a few principal components included; whereas, the
success rates of the linear KPCA (or PCA) show a gradual increase over a larger
number of components. The linear KPCA yields about 95% success rate when using
20 or more principal components. Note that the components are included as attributes
in order of decreasing eigenvalues, not their contributions to the success rate.
Compared to the previous results of fixed layer thicknesses, when thickening- or
thinning-upward trends are incorporated into sequences, stacking patterns can affect
seismic signatures. Seismograms with the retrogradational parasequence sets are the
most distinguishable among the three stacking patterns. This observation can be seen,
for example, from the parallel coordinate plot of the first ten coordinates of the
classical MDS result. From Figure 4.44, some coordinate values corresponding to the
retrogradational pattern are quite separated out from the other patterns, especially the
1st and 4
th coordinates, while most coordinate values of the progradational and
aggradational patterns are very similar. Note that patterns here refer to the patterns of
parasequence sets, which contain a series of parasequences. The parasequences
themselves are modeled in this study as being progradational (i.e., coarsening- and
thickening-upward). Therefore, it is more difficult to distinguish the progradational
pattern (i.e., both parasequences and their stacks are progradational) and the
aggradational pattern (i.e., parasequences are progradational and their stacks have no
change in the overall trend) because both share similar characteristics throughout the
sequences.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 156
Table 4.6: for stacking-pattern effects using sequences generated from the fixed-sampling transition matrices. Choices of parameters for each method are also included. Performance of each method is shown as a success rate in classifying a data point into three stacking patterns: retrogradational, progradational, and aggradational patterns.
Method Compared
objects
Distance/ or
Kernel
Classification
success rate
(using 1st, 2nd,
both coordinates)
Notes
Wavelet-
transform
– – 0.36, 0.36, 0.34
Slope/intercept/both
MDS
(classical)
Seismograms Cityblock 0.49, 0.45, 0.55
0.49,0.55,
0.61,0.71
1,2,1&2
(using 1,2,5,10
coordinates)
MDS
(metric)
Seismograms Cityblock 0.50,0.58,0.63,
0.71
(using 1,2,5,10
coordinates)
MDS
(non-
metric)
Seismograms Cityblock 0.48,0.57,0.63,
0.71
(using 1,2,5,10
coordinates)
KPCA Seismograms Gaussian
0.44,0.47,0.62,
0.68
(using 1,2,5,10
components)
σ2 = 5
KPCA Seismograms Dynamic
similarity
0.67, 0.66, 0.66,
0.72
(using 1,2,5,10
components)
σ = 0.0716, m = 80% of
total seismogram length,
r = 4
KPCA Seismograms Inverse
multi-
quadric
0.44,0.44,0.59,
0.68
(using 1,2,5,10
components)
c2 = 100
KPCA Seismograms Polynomial 0.43, 0.44, 0.59,
0.67
(using 1,2,5,10
components)
c = 10, d = 2
KPCA Seismograms Linear 0.44,0.44, 0.58,
0.67
(using 1,2,5,10
components)
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 157
Figure 4.43: Change in KPCA classification success rate when the number of
principal components included as attributes increases.
Figure 4.44: Parallel coordinate plot of the first ten coordinates of the classical MDS
result. The solid lines are the median (i.e., the 0.5-quantile) of the component
4.4.5 Discussions
The layer thickness in our simulations has a wavelength-to-mean thickness ratio of
about 100. Seismic waves cannot distinguish the boundaries of these thin layers
0 10 20 30 40 50 60 70 80 900.4
0.5
0.6
0.7
0.8
0.9
1
# of principal components
Cla
ssific
ation s
uccess r
ate
Gaussian KPCA
dynamic similarity KPCA
inverse multi-quadric KPCA
polynomial KPCA
linear KPCA
1 2 3 4 5 6 7 8 9 10-20
-15
-10
-5
0
5
10
15
20
Coordinate
Coord
inate
Valu
e
(A) Retrograding
(B) Prograding
(C) Aggrading
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 158
because of their band-limited characteristics. Hence, interpretations of these sub-
resolution layers can only be based on statistical attributes of the seismograms. In this
study, we generate synthetic seismograms of thin sand-shale sequences and extract
seismic attributes using various techniques such as wavelet transform,
multidimensional scaling (MDS), and kernel principal component analysis (KPCA).
These attributes are then related to properties of interest, which are net-to-gross ratio,
water saturation in the sand layers, and stacking patterns. Our results demonstrate that
these attributes can differentiate sequences with different reservoir properties. For
example, KPCA, especially the dynamic similarity KPCA, captures variations in net-
to-gross ratios and saturations well. However, there is no clear cut answer to which
attributes are the best choice because an attribute that performs really well in
differentiating one reservoir property may not work so well for other properties. For
example, the wavelet-transform attributes are able to capture variations in net-to-gross
ratios within the reservoirs, but their performance deteriorates when dealing with
changes in water saturation and stacking patterns.
Increasing the number of coordinates of MDS or components of KPCA results can
improve the classification of various reservoir properties. Our results show that with
only a small number of components used, KPCA with the dynamic similarity kernel
yields high success rates which are often higher than the other kernels with the same
number of components; however, when the number of components increase, the
success rates from the dynamic similarity KPCA do not vary much, while the success
rates from the linear KPCA gradually increase and eventually reach higher success
rates than those from the dynamic similarity KPCA. Thus, both the dynamic similarity
and the linear KPCA are good candidates for being used as attributes. Note that all
parameters used in the kernel formulas for KPCA (e.g., dynamic similarity,
polynomial KPCA) are estimated to yield good classification success rates; however,
it is not our goal in this work to find the exact parameters that optimize the success
rates.
Using a large number of components as attributes can make visualization more
difficult. Moreover, if these attributes are used in other classification methods or
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 159
models such as neural network, a large number of attributes can lead to more complex
models and more computation time. Therefore, instead of just increasing the number, a
subset of coordinates or components which are most relevant to the property of
interest can also be chosen. For example, we heuristically select a few components
that effectively differentiate different property classes using the parallel coordinate
plots. Further analysis is needed for the best possible combination of components.
All seismograms generated in this section (Section 4.4) are noise-free. To
investigate the effect of noise on the seismic attributes (e.g., KPCA), we add noise to
the synthetic seismograms which are generated from transition matrix A with varying
water saturations (Section 4.4.3). Then we run KPCA with a dynamic similarity kernel
and present the results in Figure 4.45. From the first two principal components of the
dynamic similarity KPCA, adding noise does not affect the attributes much. The
success rates of classifying a seismogram into its corresponding saturation class stay
almost the same. Note that the effect of noise on the KPCA attributes may vary
depending on types of kernels and the characteristics of seismograms (e.g., length and
shape).
Figure 4.45: Projections of seismograms with added noise onto the first two principal
components from the dynamic similarity KPCA. Each projected point is color-coded by the water-saturation value of the sand layers within the corresponding sequence. Percentages of noise added are specified in each panel. The success rate of classifying a projected seismogram into its corresponding saturation class is shown in the lower left corner of each panel.
-1 0 1-1
-0.5
0
0.5
1
Second p
rincip
al com
ponent
First principal component
-1 0 1-1
-0.5
0
0.5
1
Second p
rincip
al com
ponent
First principal component
-1 0 1-1
-0.5
0
0.5
1
Second p
rincip
al com
ponent
First principal component
-1 0 1-1
-0.5
0
0.5
1
Second p
rincip
al com
ponent
First principal component
sw =0.1
sw =0.5
sw =1
0% noise
success rate = 0.87
5% noise
success rate = 0.87
10% noise
success rate = 0.87
20% noise
success rate = 0.85
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 160
Results from our study have shown how the seismic signatures can vary with
changes in the properties and spatial arrangements of sub-resolution layers. However,
any patterns in the attribute space associated with changes in net-to-gross ratios, water
saturation, or stacking patterns shown in this study are not universal. In practice, these
patterns can vary depending on rock properties, numbers of lithologic states,
transition-matrix configurations, presence of noise, and other specific parameters.
Even though the study does not investigate all possible scenarios (e.g., all possible
transition-matrix configurations), it shows a promising applicability in characterizing
thin sand-shale reservoirs. In real application, all parameters should be calibrated at
the well locations. By assuming that the stratigraphy in the explored area demonstrates
a lateral continuation within conformable sequences, the inferred transition matrix
from a calibration well could be used to explore statistically how the seismic attributes
(e.g., wavelet-transform attributes, MDS coordinates) would change with varying net-
to-gross ratios and saturations. These statistics of the attributes can then be applied to
observations away from the well to characterize the area and quantify the uncertainties.
We show a numerical example of this application in the next section.
4.5 Net-to-gross estimation from 2-D sections
In the previous section, we investigated seismic signatures of 1-D, thin sand-shale
reservoirs. Various attributes from different methods show patterns relating to changes
in net-to-gross ratios, water saturation, and stacking patterns within the reservoirs.
These patterns can be applied to a real situation by following four main steps: (1)
extract a transition matrix at the well location using well log data, (2) create models
(e.g., 2-D sections) with varying property of interest, (3) extract seismic attributes
corresponding to those models and generate probability distributions of the attributes,
(4) and use the distributions to estimate properties of locations away from the well. In
this section, we demonstrate this four-step procedure through a numerical example.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 161
4.5.1 Model setup
First, we assume a known transition matrix at the well location, which is selected
to be an embedded-form transition matrix with two lithologic states representing sand
and shale. We assign average clay fractions of 0.1 and 0.9 to the sand and shale states,
respectively. Then, we introduce uncertainties by assuming each state has normally-
distributed clay fractions with previously specified average values. Sw of sand and
shale are 0.3 and 1, respectively. The thickness distributions of both states are
exponentially distributed with average thicknesses of 0.3 and 0.5 m, respectively. The
transition matrix and thickness distributions are used to generate 1-D vertical
sequences at the well location only. Then, we create multiple realizations of 2-D
spatial models describing geology away from the well location (i.e., sand layers are
thinning linearly starting from the well location into an area away from the well)
(Figure 4.46). Seismic responses of these realizations are generated at 25 discrete
locations which are equally spaced along the 2-D sections. 5% noise is added to the
seismograms. Then, seismic attributes are extracted and related to the property of
interest, in this case net-to-gross values. For illustration here, we select slope and
intercept attributes from the wavelet-transform analysis. In section 4.5.2 below, we
will explore statistically how the attributes vary with changes in net-to-gross ratios.
These statistics will then be applied to an unknown synthetic seismic section for
estimating net-to-gross ratios of the area away from the well (Section 4.5.3).
Figure 4.46: A realization of one 2-D geologic section used in the numerical example.
The area at the left end marked with the red box corresponds to the well location. Sand and shale are colored in yellow and blue respectively. The thicknesses of sand layers decrease linearly away from the well. The total thickness of reservoir is 150 m. An example of the thickness distribution used to simulate the sequence at the well location is also shown.
0 1 2 3 4 5 6 7 80
20
40
60
80
100
Shale Cap Rock
Shale Cap Rock
decreasing
N/G
Total Thickness
~ 150 m
Thickness (m)
0 1 2 3 4 5 6 7 80
20
40
60
80
100
Shale Cap Rock
Shale Cap Rock
decreasing
N/G
Total Thickness
~ 150 m
Thickness (m)
Shale Cap Rock
Shale Cap Rock
Shale Cap Rock
Shale Cap Rock
decreasing
N/G
Total Thickness
~ 150 m
Thickness (m)
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 162
4.5.2 Results
Slope and intercept results from realizations of 1,000 2-D sections is shown in
Figure 4.47. Each point in the plot represents one seismogram. There is a systematic
change in both slope and intercept values as net-to-gross values increase. Contour
plots show distributions of slope and intercept values for the high (i.e., ≥ 0.6) and low
(i.e., ≤ 0.35) net-to-gross groups (Figure 4.48).
Figure 4.47: Results from realizations of 1,000 2-D sections show how slope and
intercept vary with varying net-to-gross ratios.
Figure 4.48: Contour plots and marginal distributions for the high (i.e., equal to or
greater than 0.6) and low (i.e., equal to or less than 0.35) net-to-gross values.
2.95 2.96 2.97 2.98 2.99 3 3.01-29
-28
-27
-26
-25
-24
-23
Slope
Inte
rce
pt
0
0.2
0.4
0 0.2 0.4
net-to-gross<=0.35
net-to-gross>=0.6
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 163
4.5.3 Net-to-gross estimation using a Bayesian framework
We estimate posterior distributions of net-to-gross ratios at three selected locations
from the unknown seismic section assumed to be acquired from the area. Posterior
distributions of net-to-gross ratios given a pair of attributes (i.e., slope and intercept)
can be obtained using Bayes’ formula:
𝑃 𝑁𝑇𝐺 𝐴𝑡𝑡𝑟 𝛼𝑃 𝐴𝑡𝑡𝑟 𝑁𝑇𝐺 ∙ 𝑃 𝑁𝑇𝐺 ,
4.1
where NTG is the net-to-gross ratio, Attr represents the seismic attributes (e.g., slope
and intercept from wavelet-transform-based analyses), P(NTG) is the prior probability
of the net-to-gross ratios, P(Attr | NTG) is the likelihood function, and P(NTG | Attr)
is the posterior probability of the net-to-gross ratios.
In this example, we use a simple geological model (i.e., thinning of sand layers) to
create realizations of many 2-D rock sections. From these realizations, we can then
estimate P(NTG) or the prior probability of net-to-gross ratios at a specified location
“X” on the 2-D section and use it together with the likelihood function at that location
“X” to obtain the posterior distribution of the net-to-gross ratios at “X”. Results for
three selected locations are shown in Figure 4.49.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 164
Figure 4.49: (Lower left corner) Posterior distributions of net-to-gross ratios for three
selected locations on the unknown seismic section labeled as (1), (2) and (3). The true values for each location are shown in the table at the lower right corner.
The posterior distributions shown in Figure 4.49 are quite narrow and also capture
the true net-to-gross ratios. This is partly because we assume that the geology is
simple and known in the area, and this information helps reduce uncertainty in the
prediction. However, in reality there may see larger uncertainty due to the more
complicated natural system.
4.6 Local net-to-gross estimation in non-stationary sequences
In the previous numerical examples, we focused mostly on sequences in which an
entire sequence is generated using one transition matrix. However, we often see spatial
statistics of layered media change with depth. Examples of non-stationary sequences
are shown in Section 4.4.4, where thickening- and thinning-upward trends are
incorporated into sequences. In the previous sections, we used seismic attributes
extracted from seismograms to represent the entire reservoir interval, and so we could
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 165
only estimate a posterior distribution of net-to-gross ratios for the whole reservoir.
Unlike the previous sections, this section focuses on seismic signatures of thin-bedded
sequences that are generated from two transition matrices and aims to estimate local
distributions of net-to-gross across the reservoir.
4.6.1 Model setup
To generate sequences, we use transition matrices of the fixed sampling type. Only
two lithologic states are considered: sand and shale. The thickness of each individual
layer is 0.5 m, and the total thickness of the reservoir interval is 200 m. To simulate
sequences of lithologic states, we select two fixed-sampling transition matrices (A and
B), as shown in Figure 4.50. At each simulation step, the choice of which matrix to
use for generating the next lithologic state depends on the following rules. At each
step, we draw a random number from a uniform distribution on the interval [0,1]. For
a simulation step within the top half of the reservoir, if the random number is less than
0.8, matrix A is selected (i.e., the probability of using matrix A at each step in the top
half of the reservoir is 0.8). A similar rule is used for every simulation step in the
bottom half of the reservoir, except that matrix A is replaced by matrix B. The
resulting sequence becomes non-stationary. Examples of these sequences are shown in
Figure 4.50.
We assign average clay fractions of 0.1 and 0.9 to the sand and shale states,
respectively. Then we introduce uncertainties by assuming that each state has
normally-distributed clay fractions with the previously specified average values and
standard deviations of 0.01 and 0.03, respectively. Water saturations in sand and shale
layers are equal to 0.1 and 1, respectively. Porosity, velocity and density are then
computed following the methods described in Section 4.3.2. Forward modeling of
these sequence realizations for normal-incident reflected seismograms is performed
using the Kennett algorithm (Kennett, 1983) with a zero-phase Ricker wavelet. The
center frequency is 30 Hz, and the simulations are performed for all reverberations.
The seismograms are noise-free. We run a moving window of 80 ms through each
seismogram, resulting in seismogram segments of equal length. Then the
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 166
corresponding sequence segments and their net-to-gross values are obtained (Figure
4.51). These seismogram segments are then used to extract statistical attributes.
Figure 4.50: Two transition matrices used to generate non-stationary sequences and
examples of sequence realizations. For simulation steps in the top half of the reservoir interval, the probability of using matrix A is 0.8. The same is true for the bottom half of the reservoir with matrix B. The lithologic states considered are sand (S) and shale (Sh).
Figure 4.51: Segmentation of seismograms and sequences. (Left) application of an
80-ms moving window to a seismogram (Right) seismograms segments and their corresponding sequence segments.
P(A) = 0.8
P(B) = 0.8
Sh S Sh S
(A) 0.15 0.85
0.10 0.90
(B) 0.90 0.10
0.85 0.15
Transition matrices
Sh SSh S
Sh S
Sh S
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 167
4.6.2 Results and discussion
Results from the Gaussian KPCA with σ2 = 0.5 are shown in Figure 4.52. Each
point corresponds to a seismogram segment and is color-coded by its net-to-gross
value. Four net-to-gross classes are defined as follows: [0.16, 0.31), [0.31, 0.43), [0.43,
0.59), and [0.59, 0.77). When the first ten components of the Gaussian KPCA is used,
10-fold cross validation repeated ten times (Section 4.4.2) produces the classification
success rate of 82%.
Figure 4.52: The first two components of KPCA with Gaussian kernel. A total of
1200 seismogram segments are represented as points which are color-coded by the net-to-gross values of the corresponding sequence segments.
In real applications, one of the goals of any seismic signature study is to apply the
results to characterize reservoirs in areas away from well locations. To apply our local
net-to-gross estimation in real situations, first non-stationary transition matrices at the
well location are estimated using well log data. By assuming some prior knowledge of
geology in areas around the well, sequence models with varying net-to-gross values
are generated. Seismic attributes corresponding to these models are then used in local
net-to-gross estimation of locations away from the well.
Here we present a numerical example assuming that the non-stationary transition
matrices at the well location are already obtained and that the geology of areas
spatially around the well is not fully known but it can still be described by the same
-0.4 -0.2 0 0.2 0.4-0.4
-0.2
0
0.2
0.4
First principal component
Second p
rincip
al com
ponent
Net-to-gross ratios
0.2
0.3
0.4
0.5
0.6
0.7
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 168
form of the transition matrices inferred at the well. Sequences are modeled using the
same setup as described in Section 4.6.1 and Figure 4.50, except that the probabilities
of using transition matrix A and B in the top and bottom halves of the reservoir
interval respectively are 1, and that the transition matrices A and B inferred at the well
are as follows: 𝐴𝑤𝑒𝑙𝑙 = 0.12 0.880.37 0.63
and 𝐵𝑤𝑒𝑙𝑙 = 0.67 0.330.78 0.22
.
From the matrices at the well, the same forms of matrices which described areas
around the well are 𝐴 = 1 − 𝑎 𝑎1 − 𝑏 𝑏
and 𝐵 = 𝑐 1− 𝑐𝑑 1 − 𝑑
, where 0.51 ≤ a, b, c, and d
≤ 0.99. These matrices are used to generate sequence realizations with varying net-to-
gross ratios. Then, we generate corresponding synthetic seismograms and run a
moving window of 80 ms through each seismogram, resulting in seismogram
segments of equal length. These seismogram segments are then used to extract the
KPCA attributes. The first two components of the Gaussian KPCA with σ2 = 0.5 are
shown in Figure 4.53. These KPCA attributes are then used to estimate local net-to-
gross distributions for three “unknown” seismograms, each of which corresponds to a
sequence with net-to-gross ratios varying from a low value at the top to a high value at
the bottom of the sequence. Each of the “unknown” seismogram is divided into three
80-ms segments, and the nine “unknown” seismogram segments are represented as
black markers in Figure 4.53. By following the Bayesian framework described in
Section 4.5.3, probability distributions of local net-to-gross for the nine samples are
estimated and compared with the true net-to-gross values (shown in red dash lines)
(Figure 4.54). There is a reasonable match between the most probable and the true net-
to-gross values. Moreover, we are able to extract the trend of net-to-gross ratios (e.g.,
low to high) from the top to the bottom of the reservoir and see where most sand is
located. In this example, our seismic attribute analysis of non-stationary sequences is
used to estimate local net-to-gross distributions along the reservoirs, which give us
more insight into how sand is distributed within the reservoir.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 169
Figure 4.53: The first two components of KPCA with Gaussian kernel. A total of
2070 seismogram segments generated from the same forms of transition matrices as described in the text are represented as points which are color-coded by the net-to-gross values of the corresponding sequence segments. Nine seismogram segments are plotted in black and treated as unknowns.
Figure 4.54: Probability distributions of local net-to-gross estimation for the three
(unknown) seismograms. Each row represents the results for each unknown seismogram. Each column represents local net-to-gross estimations at a specified position of the sequences which correspond to the unknown seismograms. The red dash lines represent the true net-to-gross values.
-0.4 -0.2 0 0.2 0.4-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
First principal component
Second p
rincip
al com
ponent
0
0.2
0.4
0.6
0.8
Net-to-gross ratios
0 .2 .4 .6 .8 10
.2
.4
.6
.8
Net-to-gross ratio
Pro
babili
ty
0 .2 .4 .6 .8 10
.2
.4
.6
.8
Net-to-gross ratio
0 .2 .4 .6 .8 10
.2
.4
.6
.8
Net-to-gross ratio
0 .2 .4 .6 .8 10
.2
.4
.6
.8
Net-to-gross ratio
Pro
babili
ty
0 .2 .4 .6 .8 10
.2
.4
.6
.8
Net-to-gross ratio
0 .2 .4 .6 .8 10
.2
.4
.6
.8
Net-to-gross ratio
0 .2 .4 .6 .8 10
.2
.4
.6
.8
Net-to-gross ratio
Pro
babili
ty
0 .2 .4 .6 .8 10
.2
.4
.6
.8
Net-to-gross ratio
0 .2 .4 .6 .8 10
.2
.4
.6
.8
Net-to-gross ratio
Position along the sequence Top Middle Bottom
Unknown 1
Unknown 2
Unknown 3
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 170
4.7 Discussion
In Sections 4.4 – 4.6, we showed three different examples where seismic attributes
were extracted from synthetic seismograms and used to infer reservoir properties of
thin sand-shale sequences. In this section, we discuss broader aspects of the attributes
used in this work.
4.7.1 Comparisons with amplitude attributes
Using the synthetic seismograms in Section 4.4.2, we extract two additional
attributes, the root-mean-square (RMS) amplitude and the absolute amplitude at the
top of the reservoir. The cross-plot between the two amplitudes is illustrated in Figure
4.55. Each point is color-coded by the net-to-gross ratios of the corresponding sand-
shale sequences. Then, we use the stratified 10-fold cross validation scheme described
in Section 4.4.2 to compute the success rate of classifying the seismograms
represented by the two amplitude attributes into their corresponding net-to-gross
classes. The success rate is 80%. Figure 4.56 shows the success rates of three selected
sets of KPCA attributes and the rates of these KPCA attributes combined with the
amplitude attributes. By adding the amplitude attributes to the first two principal
components of Gaussian KPCA, the success rate changes from 81% to 82%. Likewise,
for the case of dynamic similarity KPCA, the success rate changes from 90% to 91%.
However, when the amplitude attributes are added to the first 20 principal components
of linear KPCA, the success rate stays almost the same. These observations show that
the amplitude attributes can be combined with the KPCA attributes to improve the
success rate; however, in this example even though the amplitude attributes alone
yield a success rate as high as 80%, the improvement in the success rate is very small.
This implies that the amplitude attribute do not contribute much “new” information to
the KPCA attributes. In other words, these variations and patterns in the amplitudes
are already captured and included in the KPCA attributes.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 171
Figure 4.55: Amplitude attributes extracted from seismograms which correspond to
sequences with different net-to-gross ratios.
Figure 4.56: Success rates in classifying a data point into three net-to-gross classes
(<0.38, 0.38-0.465, and >0.465) by using the attributes specified on the horizontal axis. The rates are shown on top of each bar. Att1: RMS and absolute amplitude at the top of the reservoir, Att 2: the first two principal components of Gaussain KPCA, Att3: the first two principal components of dynamic similarity KPCA, and Att4: the first 20 principal components of linear KPCA.
4.7.2 Notes on the feature-extraction based attributes
In this work, when performing the feature-extraction based techniques, we only
use amplitude data as the input features for the seismograms. However, it is easy to
combine other features (e.g., instantaneous phase, instantaneous frequency) into the
0 0.02 0.04 0.06 0.08-0.15
-0.1
-0.05
0
0.05
0.1
RMS amplitude
Am
plit
ude a
t th
e t
op o
f re
serv
oir
Net-to-gross ratios
0.25
0.3
0.35
0.4
0.45
0.5
0
0.2
0.4
0.6
0.8
1
Cla
ssific
ation s
uccess r
ate
Att4 Att1+Att4Att1 Att2 Att1+Att2 Att3 Att1+Att3
0.797 0.8090.823
0.8960.905
0.9830.985
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 172
analysis by normalizing different features so that each of the features has zero mean
and unit variance before running the usual steps.
Principal component analysis (PCA), kernel principal component analysis (KPCA),
and multidimensional scaling (MDS) used in this work are unsupervised feature-
extraction techniques, which do not account for class labels. As a result, there is no
guarantee that the selected new features (i.e., principal components) optimize class
separation. Instead, by selecting only a number of these new features as new
representations of the input data, the features with more discriminating power may be
discarded (Avseth et al., 2005). An alternative approach is to use supervised feature-
extraction techniques. For example, linear discriminant analysis (LDA) aims to find
new representations of data, while minimizing within class variance and maximizing
class separation. However, LDA has several limitations. For example, when the
dimensionality of data is high, a large number of training samples is required so that
within-class scatter matrix is nonsingular. To avoid this problem, it is common to first
reduce the dimensionality of the data using techniques such as PCA before applying
LDA (also known as PCA plus LDA) (Zhang, et al., 2006; Liew and Wang, 2009;
Thomas and Wilsey, 2011). We also use this workflow by applying PCA, KPCA, or
MDS and followed by LDA.
PCA involves solving an eigenproblem of the data covariance matrix whose size
grows with the dimensionality of the training data, while the computations in both
KPCA and MDS algorithms involve the matrices whose sizes grow with the number
of the training data. As a result, depending on both the dimensionality and the size of
the training dataset, the feasibility of these techniques may be limited (e.g., van der
Maaten, 2009; Shi and Zhang, 2011). The number of training data can be controlled
since that is the number of simulated seismograms.
With the exception of classical MDS, MDS algorithms solve iterative optimization
problems, so they can be very slow. Furthermore, before running the algorithms, the
desired number of dimension needs to be specified.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 173
4.8 Conclusions
This chapter illustrates a workflow which can be applied to thin sand-shale
sequences for quick interpretation of reservoir properties before performing full
stochastic inversion. The workflow consists of four steps: (1) estimate transition
matrices at the well location from log data, (2) create what-if scenarios with varying
reservoir properties, (3) extract seismic attributes from the synthetic models and these
attributes become a training set, and (4) finally use the training set to classify seismic
signatures or estimate properties of locations away from the well. Most of the seismic
attributes shown in this chapter are derived from feature-extraction techniques, which
compare amplitudes (i.e., features) of entire seismogram segments, find new
representations of these seismograms with a new, smaller set of features, and use them
as attributes for estimating reservoir properties. The resulting attributes do not
necessarily relate to specific rock properties (e.g., Vp/Vs, acoustic impedance), nor do
they have physical meaning. However, it should be kept in mind that physical and
geological models are used as inputs to generate the seismograms from which the
attributes are extracted. Our numerical examples demonstrate that the feature-
extraction based attributes are able to capture variations and patterns in the input
amplitudes corresponding to variations in reservoir properties. The decision of how
many features should be selected depends on both the performance of each algorithm
and the goal of the study. By including more informative features as attributes, the
accuracy of predicting reservoir properties may increase; however, if these attributes
are further used in other algorithms, a larger number of attributes increases complexity
and requires more computation time.
4.9 Acknowledgements
This work was supported by the Stanford Rock Physics and Borehole Geophysics
project and the Stanford Center for Reservoir Forecasting.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 174
4.10 References
Álvarez, G., Sansó, B., Michelena, R.J. and Jiménez, J.R., 2003, Lithologic
Characterization of a Reservoir Using Continuous-Wavelet Transforms: IEEE
Trans. on Geoscience and Remote Sensing, 41(1).
Avseth, P., Mukerji, T. and Mavko, G., 2005, Quantitative Seismic Interpretation,
Cambridge.
Backus, G. E., 1962, Long-wave elastic anisotropy produced by horizontal layering:
Journal of Geophysical Research, 67, 4427–4440.
Bengio, Y., Delalleau, O., Le Roux, N., Paiement, J-F., Vincent, P., and Ouimet, M.,
2006, Spectral Dimensionality Reduction: Studies in Fuzziness and Soft
Computing, 207/2006, 519-550.
Boggs, S. Jr., 2001, Principles of Sedimentology and Stratigraphy, Prentice Hall: New
Jersey.
Buja, A., Swayne, D.F., Littman, M., Hofmann, H., and Chen, L., 2008, Data
Visualization with Multidimensional Scaling: Journal of Computational and
Graphical Statistics, 17(2), 1-29.
Burges, C.J.C., 2010, Geometric Methods for Feature Extraction and Dimensional
Reduction: The Data Mining and Knowledge Discovery Handbook, Eds. L.
Rokach and O. Maimon, 59-92. Kluwer Academic.
Burges, C.J.C., 2010, Dimension Reduction: A Guided Tour: Foundations and Trends
in Machine Learning, 2(4), 275-365.
Chopra, S., and Marfurt, K.J., 2005, Seismic Attributes – A Historical Perspective:
Geophysics, 70(5), 3SO – 28SO.
Cox, T.F. and Cox, M.A.A., 2001, Multidimensional Scaling, Chapman & Hall/CRC.
Chung, H. and Lawton, D.C., 1995, Frequency Characteristics of Seismic Reflections
from Thin Beds: Canadian J. of Explr. Geophysics, 31, 32-37.
Cunningham, P., 2008, Dimension reduction: Machine learning techniques for
multimedia: case studies on organization and retrieval, Springer, Eds. M. Cord and
P. Cunningham.
Dvorkin, J. and Gutierrez, M.A., 2002, Grain Sorting, Porosity, and Elasticity:
Petrophysics, 43(3), a3.
Flórez, J. and Mavko, G., 2003, Porosity-Velocity Distribution in Stratigraphic
Sequences: The Marion-Yin Model, Search and Discovery, no. 40105.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 175
Gassmann, F., 1951, Uber die elastizitat poroser medien: Vier Natur Gesellschaft, 96,
1-23.
Harbaugh, J.W. and Bonham-Carter, G., 1970, Computer Simulation in Geology,
Wiley.
Harms, J.C. and Tackenberg, P., 1972, Seismic signatures of sedimentation models:
Geophysics, 37(1), 45-58.
Hart, B. and M-A., Chen, 2004, Understanding seismic attributes through forward
modeling: The Leading Edge, 23(9), 834–841.
Hart, B., 2008, Stratigraphically Significant Attributes: The Leading Edge, March
2008, 320-324.
Kamola, D.L. and Van Wagoner, J.C., 1995, Parasequence Stratigraphy and Facies
Architecture of the Spring Canyon Member, Upper Cretaceous Blackhawk
Formation: Van Wagoner, J.C., and Bertran G. (eds.), Sequence Stratigraphy of
Foreland Basin Deposits – Outcrop and Subsurface Examples from the Cretaceous
of North America: American Association of Petroleum Geologists, Memoir 64, 27
– 54.
Kennett, B., 1983, Seismic wave propagation in stratified media: Cambridge
University Press, Cambridge.
Khattri, K. and Gir, R., 1976, A Study of the Seismic Signatures of Sedimentation
Models Using Synthetic Seismograms: Geophysics, 34, 454-477.
Krumbein, W.C. and Dacey, M.F., 1969, Markov Chains and Embedded Markov
Chains in Geology: Mathematical Geology, 1(1), 79-96.
Lange, J. and Almoghrabi, H., 1988, Lithology Discrimination for Thin Layers Using
Wavelet Signal Parameters: Geophysics, 53(12), 1512-1519.
Li, B., Change, E., and Wu, Y., 2003, Discovery of a perceptual distance function for
measuring image similarity: Multimedia Systems, 8, 512 – 522.
Liew, A. W-C and Wang, S., 2009, Visual speech recognition: Lip segmentation and
mapping, Idea Group Publishing.
López, M. and Aldana, M., 2007, Facies Recognition Using Wavelet Based Fractal
Analysis and Waveform Classifier at the Oritupano-A Field, Venezuela: Nonlinear
Processes in Geophysics, 14, 325-335.
Marion, D., 1990, Acoustical, Mechanical and Transport Properties of Sediments and
Granular Materials, Ph.D. Thesis, Stanford University.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 176
Marion, D., Mukerji, T., and Mavko, G., 1994, Scale effects on velocity dispersion:
From ray to effective medium theories in stratified media: Geophysics, 59(10),
1613-1619.
Marion, D., Nur, A., Yin, H. and Han, D., 1992, Compressional Velocity and Porosity
in Sand-Clay Mixtures: Geophysics, 60(1), 241-251.
Mitchum, R.M. and Van Wagoner, J.C., 1991, High-Frequency Sequences and Their
Stacking Patterns: Sequence-stratigraphic Evidence of High-frequency Eustatic
Cycles: Sedimetary Geology, 70, 131 – 160.
Mukerji, T., 1995, Waves and scales in heterogeneous rocks, Ph.D. Thesis, Stanford
University.
Mukerji, T., and Mavko, G., 2008, The flaw of averages and the pitfalls f ignoring
variability in attribute interpretations: The leading edge, 27(3), 382-384.
Mukerji, T., Mavko, G., Mujica, D., and Lucet, N., 1995, Scale-dependent seismic
velocity in heterogeneous media: Geophysics, 60(4), 1222-1233.
Mulholland, J.W., 1998, The Parasequence: The Leading Edge, 17(10), 1374 – 1376.
Okaya, D.A., 1995, Spectral Properties of the Earth’s Contribution to Seismic
Resolution: Geophysics, 60(1), 241-251.
Parks, K.P., Bentley, L.R. and Crowe, A.S., 2000, Capturing Geological Realism in
Stochastic Simulations of Rock Systems with Markov Statistics and Simulated
Annealing: Journal of Sedimentary Research, 70(4), 803-813.
Rathi, Y., Dambreville, S., and Tannenbaum, A., 2006, Comparative analysis of kernel
methods for statistics shape learning: Proceeding of the 2nd
ECCV international
conference on computer vision approaches to medical image analysis, 96-107.
Saul, L.K., Weinberger, K.Q., Ham, J.H., Sha, F., and Lee, D.D., 2006, Spectral
methods for dimensionality reduction: Semisupervised Learning, Eds. O. Chapelle,
B. Schoelkopf, and A. Zien, 293-308. MIT Press: Cambridge.
Scheidt, C. and Caers, J., 2009, Representing Spatial Uncertainty Using Distances and
Kernels: Mathematical Geosciences, 41, 397-419.
Schölkopf, B., Smola, A., and Müller, K.-R., 1997, Kernel Principal Component
Analysis: Lecture Notes in Computer Science, 1327/1997, 583-588.
Schwarzacher, W., 1975, Sedimentation Models and Quantitative Stratigraphy,
Elsevier.
Shi, W. and Zhang, W., 2011, The accelerated power method for kernel principal
component analysis: Applied Informatics and Communication, Part II:
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 177
International Conference, ICAIC 2011, Xi’an China, August 20-21, 2011,
Proceedings.
Sinvhal, A. and Khattri, K., 1983, Application of Seismic Reflection Data to
Discriminate Subsurface Lithostratigraphy: Geophysics, 48(11), 1498-1513.
Sinvhal, A. and Sinvhal, H., 1992, Seismic Modeling and Pattern Recognition in Oil
Exploration, Kluwer Academic Publishers.
Stovas, A., Landrø, M., and Avseth, P., 2006, AVO attribute inversion for finely
layered reservoirs: Geophysics, 71(3), C25-C36.
Takahashi, I., 2000, Quantifying Information and Uncertainty of Rock Property
Estimation from Seismic Data, Ph.D. Thesis, Stanford University.
Taner, M.T., 2001, Seismic Attributes: CSEG Recorder, Sep 2001, 48-56.
Tenenbaum, J.B., de Silva, V., and Langford, J.C., 2000, A global geometric
framework for nonlinear dimensionality reduction: Science, 290, 2319-2323.
Thomas, A.A. and Wilsey, M., 2011, A comparative study of feedforward neural
network and simplified fuzzy ARTMAP in the context of face recognition:
Advances in networks and communications: First international conference on
computer sciences and information technology, CCSIT 2011, Bangalore, India,
January 2-4, 2011. Proceedings. N. Meghanathan et al. (Eds.), 277-289.
Van Wagoner, J.C., Posamentier, H.W., and Mitchum, R.M., Vail, P.R., Sarg, J.F.,
Loutit, T.S., and Hardenbol, J., 1988, An Overview of the Fundamentals of
Sequence Stratigraphy and Key Definitions: Sea-Level Change: An Integrated
Approach. SEPM Special Publication, 42.
Van Wagoner, J.C., Mitchum, R.M., Campion, K.M., and Rahmanian, V.D., 1990,
Siliciclastic Sequence Stratigraphy in Well Logs, Cores and Outcrops: Concepts
for High-Resolution Correlation of Time and Facies: AAPG Methods in
Exploration, no. 7, 55 p.
Van der Maaten, L., Postma, E., and Van der Herik, J., 2009, Dimensionality
reduction: A comparative review: Technical report TiCC TR 2009-005.
Velzeboer, C.J., 1981, The theoretical Seismic Reflection Response of Sedimentary
Sequences: Geophysics, 46(6), 843-853.
Webb, A.R., 2002, Statistical Pattern Recognition, 2nd
ed., Wiley.
Williams, C.K.I., 2002, On a Connection between Kernel PCA and Metric
Multidimensional Scaling: Machine Learning, 46, 11-19.
CHAPTER 4: Seismic signatures of thin sand-shale reservoirs 178
Witten, I.H., Frank, E., and Hall, M.A., 2011, Data Mining: Practical Machine
Learning Tools and Techniques. The Morgan Kaufmann Series in Data
Management Systems, 3rd
ed., Morgan Kaufmann.
Xu, H. and MacCarthy, A.J., 1996, Markov Chain Analysis of Vertical Facies
Sequences Using a Computer Software Package (SAVFS): Courtmacsherry
Formation (Tournaisian), Southern Ireland: Computers & Geoscience, 24, 131-140.
Yan, W., Liu, q., Lu, H., and Ma, S., 2006, Dynamic similarity kernel for visual
recognition: Lecture Notes in Computer Science, 4252, 47 – 54.
Yin, H., 1992, Acoustic Velocity and Attenuation of Rocks: Isotropy, Intrinsic
Anisotropy, and Stress Induced Anisotropy, Ph.D. Thesis, Stanford University.
Zhang, D., Jing, X., and Yang, J., 2006, Biometric image discrimination technologies,
Idea Group Publishing.
Zhang, J., 2008, Multidimensional Scaling: Visualization for Information Retrieval.
The Kluwer International Series on Information Retrieval, 23, 143-163.
179
Chapter 5
Seismic signature and uncertainty in
petrophysical property estimation of
thin sand-shale reservoirs: Case
studies
5.1 Abstract
In Chapter 4, we presented a workflow and provided seismic attributes for
interpreting seismic signatures of thin sand-shale reservoirs. In this chapter, we apply
the workflow and these seismic attributes to real data from channelized turbidite
deposits offshore Equatorial Guinea, West Africa. We focus on a short fining-upward
interval from a distal well. Within this interval, both the well log analyses and the core
images show fine-scale sand-shale lamination. Their bed thicknesses are below the log
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 180
resolution. Here we present two case studies. In the first case study, using rock
properties from the well, we generate synthetic seismograms under various scenarios
and treat these seismograms as our “real” traces in order to remove the effects of noise
and wavelet estimation from seismic signatures. The main seismic attribute used here
is obtained from kernel principal component analysis (KPCA), which projects
seismograms onto principal components. KPCA results of the training set in the first
case study show distinct seismic signatures of thin sand-shale reservoirs for varying
net-to-gross ratio, stacking pattern, water saturation, and reservoir thickness. These
results can be used to characterize the “real” traces successfully. We also present the
link between the KPCA attributes and amplitude attributes and investigate other
factors that can affect seismic signatures. In the second case study, we estimate
reservoir properties of sub-resolution sand-shale sequences from a real 2-D seismic
section. We analyze two existing wells that are located on each side of the seismic
section and find that average impedances of both the overburden and underburden
change spatially. Therefore, we incorporate such non-stationarity by adding two
additional parameters which are referred to as overburden and underburden multipliers.
We generate a training set by varying all parameters (e.g., net-to-gross ratios,
thickness) within specified ranges. The KPCA attributes of the training data on the
first two principal components illustrate distinct trends associated with changes in
some seismic-signature parameters, in particular the reservoir thickness while for other
parameters no clear trend can be observed. Since parameters that have greater impact
on seismic signatures can overshadow subtle effects from the other parameters, we
perform sensitivity analyses to rank all the seismic-signature parameters according to
their influence on seismic signatures. We first estimate probability density functions of
the important parameters for all seismic test traces. Then to reduce the effect of these
parameters and to enhance the effect of the others, for each test trace we create a new
training set (or select a subset from the training set) where these newly estimated
parameters are varied within smaller ranges near their mean values, while keeping the
ranges of the other parameters the same. Then this new training set can be used to
estimate the subsequent parameters. After these iterative steps, we obtain posterior
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 181
distributions of reservoir properties from all seismic traces in the 2-D section. To
evaluate estimation results, we add three known traces into the test set. Results show
that for both the well trace and the three additional known traces, their posterior
distributions for all the seismic-signature parameters generally capture the true values.
5.2 Introduction
In chapter 4, we used synthetic examples to illustrate a workflow for petrophysical
property estimation of thin sand-shale sequences. Our workflow has four main steps:
(1) extract transition matrices and calibrate rock properties at well locations, (2) use
the inferred transition matrices and rock-physics models to generate thin sand-shale
sequences with various scenarios (e.g., varying saturations), (3) simulate the
corresponding synthetic seismograms and extract seismic attributes from these
seismograms, and (4) apply the attributes to real seismic data for estimating reservoir
properties. The attributes for our seismic signature study are obtained using wavelet-
transform analysis, multi-dimensional scaling (MDS), and kernel principal component
analysis (KPCA). The results from the synthetic examples show that these attributes
can potentially differentiate thin sand-shale sequences with different net-to-gross
ratios or saturations.
In this chapter, we present two case studies in which we apply our workflow to
real well log and seismic data from deep-water turbidite deposits, offshore Equatorial
Guinea, West Africa. In Section 5.3, we provide geological background for the study
area including depositional environment and preliminary analyses of well log data. In
Sections 5.4, for our first case study we interpret seismic signatures at the well
location by focusing on a short fining-upward interval. In Section 5.5, for our second
case study we interpret seismic signatures within a reservoir zone across the selected
2-D section which extends from proximal to distal directions along a turbidite channel.
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 182
5.3 Geological background
The study area is an oil field located in the Rio Muni basin offshore Equatorial
Guinea, West Africa. The reservoirs are parts of a late-Cretaceous submarine canyon
which is characterized as being erosive and sand-rich, as evidenced by seismic, well
log, and core data (Dutta, 2009; Jobe et al., 2011). Using the available cores from five
wells in the study area, Lowe (2004) identified six main lithofacies: (1) thick-bedded
to massive sandstone, (2) interbedded thin-bedded sandstone and mudstone with beds
2-20 cm thick and >20% sandstone, (3) interbedded thin-bedded sandstone and
mudstone with beds <2 cm thick, (4) carbonate-cemented sediments, (5) conglomerate
and breccia, and (6) mudstone with <10% interbedded sandstone (Figure 5.1). Only
lithofacies (4) is a diagenetic lithofacies, while the others are depositional lithofacies.
One of the basic features observed in the cores is the fining-upward trend in which the
coarse-grained, thick-bedded lithofacies grades upward into successively finer-grained,
thinner-bedded lithofacies and/or mudstone (Lowe, 2004). The fining-upward pattern,
core analyses, and interpretation of seismic data suggest that these sediments are
channelized deposits of high-density turbidity currents in the submarine canyon
system (Dutta, 2009).
Figure 5.1: Six identified lithofacies from the study area offshore Equatorial Guinea, West Africa (Lowe, 2004; Dutta, 2009).
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 183
5.4 Case study 1: Effect of reservoir properties and stacking pattern
on seismic signatures at well location
In this case study, we select a short interval representing a fining-upward sequence
in a reservoir zone from well A, which locates in the distal direction of the
channelized turbidite deposits (Dutta et al., 2007; Dutta, 2009). Using the well log data,
we generate synthetic seismograms to study the effect of net-to-gross ratio, stacking
pattern, saturating fluid, and reservoir thickness on seismic signatures. In Section 5.4.1,
we apply Thomas-Stieber model to porosity and gamma ray values for estimating net-
to-gross ratio of the fining-upward sequence (Chapter 2). Sections 5.4.2 – 5.4.4 show
synthetic seismogram generation for sequences with varying net-to-gross ratio,
stacking pattern, saturating fluid, and reservoir thickness. In Section 5.4.5, we analyze
signatures of those synthetic seismograms using kernel principal component analysis
(KPCA; Chapter 4). These signatures are then used for interpreting the synthetic
seismograms generated from the log data at the well. Section 5.4.6 discusses other
factors that could affect our interpretations.
5.4.1 Net-to-gross estimation from well log data
Log measurements from well A are shown in Figure 5.2. The gamma-ray log in
the selected interval exhibits a fining-upward pattern of the uncemented lithofacies, in
which the thick-bedded or massive sandstone facies (lithofacies (1)) change into the
thinly interbedded sandstone-mudstone facies (lithofacies (2) and (3)) in the upward
direction (Dutta, 2009). As previously noted, detailed core analyses by Lowe (2004)
showed that the sandstone beds in the thin-bedded lithofacies have thicknesses ranging
from <2 cm to 20 cm. Since the individual thin beds of sandstones and mudstones are
too thin to be resolved by the logging tools, well log measurements represent average
properties of multiple beds. To correct for this averaging effect in order to estimate the
amount and the properties of the sub-resolution sand beds, we use the Thomas-Stieber
model (Thomas and Stieber, 1975) as discussed in Chapter 2.
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 184
Figure 5.2: Well A from deep-water turbidite deposits, offshore Equatorial Guinea,
West Africa. The zone of interest is highlighted. From left to right, the curves are gamma ray, bulk density, density-derived porosity, P-wave velocity, and water saturation (SW), respectively.
The Thomas-Stieber model describes how total porosity varies with shale volume
depending on the configuration of shale in the sand-shale mixtures. We focused only
on two shale configurations which are laminar and dispersed. By inputting properties
of the clean sand and shale end-members, we showed how this model can be used
graphically or algebraically to estimate the net-to-gross ratios and volumes of
dispersed shale in the sandy layers. We also demonstrated how to propagate
uncertainties of the input parameters through the model using Monte Carlo simulations
in a Bayesian framework and produced posterior distributions of the estimated
properties. We showed an example of applying the Thomas-Stieber model to the log
data from well A. In that example, the selected well-log interval is exactly the same
fining-upward sequence as shown in Figure 5.2.
In the example involving the log data from well A shown in Chapter 2, the
Thomas-Stieber model was applied to density-derived porosity and gamma ray values
by assuming that the properties of the sand and shale end-members are uncertain and
that they are normally-distributed. The resulting posterior distributions of net-to-gross
ratios along the fining-upward interval are shown in Figure 5.3. Note that The
Thomas-Stieber model for estimating net-to-gross ratios is only applied to the data
from lithofacies 1, 2, and 3. For the rest of the data, their net-to-gross ratios are simply
estimated by 1 – 𝑉𝑠ℎ , where 𝑉𝑠ℎ is the shale volume fraction. The net-to-gross ratio of
0 50 100
1050
1060
1070
1080
1090
1100
1110
Gamma ray
Depth
(m
)
2 2.5
Bulk density
0 0.2
Total porosity
2 3 4
P-wave velocity
0 0.5 1
SW
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 185
the entire fining-upward sequence is then equivalent to an arithmetic average of net-
to-gross ratios of all points along the sequence. We estimated this average value to be
0.48. This estimate can vary depending on the properties of the end-members used in
the Thomas-Stieber model.
Figure 5.3: Petrophysical analysis of the selected well-log interval using the Thomas-
Stieber model. (Left) Total porosity and gamma ray values for three lithofacies in the selected interval. The median total porosity and gamma ray values for each lithofacies are shown in solid circles. From these median points, the up-down or left-right bars indicate the interquartile ranges (i.e., from 1st to 3rd quartiles) of each property. A Thomas-Stieber diagram is also superimposed on the data. (Right) Variation of gamma ray values with depth. Data points are color-coded by lithofacies similar to the left panel. The gamma ray log shows an upward-fining trend.
Figure 5.4: Estimated net-to-gross ratios for the selected well-log interval. (Left)
gamma ray log of the selected interval and (right) posterior distributions of estimated net-to-gross ratios.
20 30 40 50 60 70 80 900.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Gamma ray
Tota
l poro
sity
Lithofacies 1
Lithofacies 2
Lithofacies 3
40 60 80
1070
1075
1080
1085
1090
1095
1100
Gamma ray
Depth
(m
)
40 60 80
1075
1080
1085
1090
1095
Gamma ray
Depth
(m
)
Lithof acies 1
Lithof acies 2
Lithof acies 3
0 0.5 1
1075
1080
1085
1090
1095
Sand fraction
Depth
(m
)
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 186
5.4.2 Seismograms at well location
When using real seismic data, both the noise and the uncertainty from the wavelet
estimation process can affect seismic signatures of thin sand-shale sequences. To
focus on the signatures, in this case study we remove these two sources of
uncertainties by generating a synthetic seismogram at the well location and using it as
our “true” seismic trace. We extract velocity and density values of the target zone
from the logs, together with the upper and the lower 200 points of the target zone
(Figure 5.5, left). Each layer is assumed to be approximately 15-cm thick (i.e., the
spacing in the log measurements). Layer properties (i.e., velocity, density and
thickness) are then input into the Kennett algorithm (Kennett, 1983) to simulate full-
waveform, normally-incident, reflected seismograms using a zero-phase Ricker
wavelet with a central frequency of 30 Hz. The simulations are performed for all
reverberations. Figure 5.5 (right) illustrates the synthetic seismogram at the well
whose gamma ray log exhibits a fining-upward stacking pattern. To expand our
seismogram test set, we generate four additional synthetic seismograms by modifying
the velocity and density logs within the target zone, while both the upper and the
lower parts of the target zone are kept the same.
For the first modification, we turn the velocity and density curves of the target
zone upside down. This new property arrangement corresponds to a “coarsening-
upward” stacking pattern. For the second modification, we randomly shuffle the orders
of the velocity-density pairs within the target zone. This property arrangement
corresponds to a more “serrated” stacking pattern. For the third modification, we apply
fluid substitution to both velocity and density curves to change the saturating fluid
from oil to brine. For the final modification, we decrease the thickness of each layer
within the target zone from 15 cm to 10 cm. Consequently, the total thickness of the
target zone also decreases. The logs produced by the four modifications are then used
to generate corresponding synthetic seismograms. Hereafter, we refer to the
seismograms for the original and the four additional logs as the fining-upward,
coarsening-upward, serrated, brine-saturated, and thinner-reservoir models,
respectively (Figure 5.6).
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 187
Figure 5.5: Acoustic impedance and the corresponding synthetic seismogram of the
target zone. (Left) Acoustic impedance of the target zone shown in magenta, together with 200 data points above and 200 data points below the target zone. (Right) Synthetic seismogram generated using a 30-Hz zero-phase Ricker wavelet.
Figure 5.6: Synthetic-seismogram test set. Note that each panel shows four repetitions
of one seismogram. From left to right, the panels show seismograms that are generated from the original well log (i.e., fining-upward), the inverted log (i.e., coarsening-upward), the shuffled log (i.e., serrated), the brine-saturated log, and the log with smaller layer thicknesses within the target zone.
5.4.3 Transition matrices and rock property calibration
This section shows how we create transition matrices and use them to generate
synthetic earth models within the target zone for varying net-to-gross ratio, stacking
5000 10000
1050
1060
1070
1080
1090
1100
1110
1120
1130
Acoustic impedance
(g/cm3 x m/s)
Depth
(m
)
tim
e (
ms)
tim
e (
ms)
Fining-upward
model
tim
e (
ms)
Coarsening-upward
model
tim
e (
ms)
Serrated model
tim
e (
ms)
Brine-saturated
model
tim
e (
ms)
Thinner-reservoir
model
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 188
pattern, saturating fluid, and layer thickness. We use Markov chain models to create
synthetic earth models within the target zone by assigning two lithologic states – sand
and shale. Using the fixed-sampling transition matrices (Chapter 4), we consider three
main categories of matrices, each of which generates sequences with distinct spatial
patterns in net-to-gross ratios.
5.4.3.1 Interbedded sand-shale sequences with no clear trend in net-to-gross ratios
along the sequences
To generate simple interbedded sand-shale sequences that span a wide range of
net-to-gross ratios, we set the form of transition matrices as 𝑇1 = 𝑎 1 − 𝑎𝑏 1 − 𝑏
, where a
and b are randomly drawn from a uniform distribution on [0,1]. Since we draw new
values of a and b for each realization, the transition matrix always changes from
realization to realization. When using these 𝑇1 matrices, net-to-gross ratios vary along
the sequences, but there is no one particular trend.
5.4.3.2 Interbedded sand-shale sequences with an upward-decreasing trend in net-to-
gross ratios along the sequences
To incorporate an upward-decreasing trend in net-to-gross ratios into the sand-
shale sequences, we first divide the sequences into two equal subsequences. Then we
set the transition matrices for the top and bottom subsequences as follows:
𝑇2(𝑡𝑜𝑝 ) = 𝑎 1 − 𝑎𝑏 1 − 𝑏
, where a and b are drawn from a uniform distribution on [0,0.5],
𝑇2(𝑏𝑜𝑡 ) = 𝑐 1 − 𝑐𝑑 1 − 𝑑
, where e and f are drawn from a uniform distribution on [0.5,1].
By combining both sets of 𝑇2 transition matrices, we produce the sequences which
exhibit an upward decrease in sand to shale ratios.
5.4.3.3 Interbedded sand-shale sequences with an upward-increasing trend in net-to-
gross ratios along the sequences
To generate an upward-increasing trend in net-to-gross ratios along the sequences,
we first divide the sequences into two equal subsequences and then use both sets of
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 189
transition matrices, as shown in Section 5.4.3.2. However, the order of both transition-
matrix sets is reversed:
𝑇3(𝑡𝑜𝑝 ) = 𝑎 1 − 𝑎𝑏 1 − 𝑏
, where a and b are drawn from a uniform distribution on [0.5,1],
𝑇3(𝑏𝑜𝑡 ) = 𝑐 1 − 𝑐𝑑 1 − 𝑑
, where e and f are drawn from a uniform distribution on [0,0.5].
By combining both sets of 𝑇3 transition matrices, we produce the sequences which
exhibit an upward increase in sand to shale ratios.
Figure 5.7 shows sample sequences from the above three categories of transition
matrices which are used to generate arrangements of sand and shale layer within the
target zone. As noted in Section 5.4.2, log measurements are spaced at approximately
15 cm apart, and we assign this spacing as the layer thicknesses when we use well log
data to generate the “true” seismogram at the well location. However, as shown by the
Thomas-Stieber analysis (Section 5.4.1), the target zone contains thin layers below the
vertical resolution of logging tools. Furthermore, core analyses also demonstrate that
sand thicknesses in this zone can be even less than 2 cm (Lowe, 2004; Dutta, 2009).
Therefore, for our synthetic earth models within the target zone, we use the fixed-
sampling transition matrices and select the thickness of the sand or shale layer to be
approximately 1.5 cm.
After generating layer arrangements, we estimate acoustic impedances of both the
sand and the shale lithologic states from the acoustic impedances at the well location.
This estimated sand property represents the oil-saturated sand. Since we also test an
interbedded brine-sand and shale model, we estimate the property of the brine-
saturated sand by applying Gassmann’s fluid substitution to the oil-saturated sand. The
estimated impedances for all three states are used as mean values of normal
distributions with assigned standard deviations (Figure 5.8). We draw values from
these distributions and assign them to layers according to their states.
In summary, we simulate five sets of sand-shale sequences with the properties
shown in Table 5.1. Each set contains 400 realizations.
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 190
Figure 5.7: Sample realizations of synthetic earth models which cover a range of net-
to-gross ratios and various stacking patterns. The sequence realizations are generated from the three categories of transition matrices: with an upward-increasing trend (left), with an upward-decreasing trend (middle), and with no trend in net-to-gross ratios (right).
Figure 5.8: Probability density functions of acoustic impedance for all lithologic states: oil sand (blue), shale (red), and wet sand (black).
No t
rend in n
et-
to-g
ross r
atios
Upw
ard
-decre
asin
g t
rend in n
et-
to-g
ross r
atios
Upw
ard
-incre
asin
g t
rend in n
et-
to-g
ross r
atios
white: sand and black: shale
4.5 5 5.5 60
2
4
6
8
10
Acoustic impedance (g/cm3 x km/s)
Pro
babili
ty d
ensity
oil-saturated
sand
shale
brine-saturated
sand
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 191
Table 5.1: Specifications for synthetic earth models with various net-to-gross ratios, stacking patterns, saturating fluids, and layer thicknesses.
Model Net-to-gross trend Sand property Reservoir
thickness
I upward-decreasing oil-saturated 17 m
II upward-increasing oil-saturated 17 m
III no trend oil-saturated 17 m
IV upward-decreasing brine-saturated 17 m
V upward-decreasing oil-saturated 11 m
5.4.4 Synthetic seismograms
In the previous section, we have generated five sets of sand-shale sequences for
various net-to-gross ratios, stacking patterns, saturating fluids, and layer thicknesses.
By inputting these sequences into the Kennett algorithm (Kennett, 1983), in this
section we simulate the corresponding synthetic seismograms (Figure 5.9). These
seismograms become our training set for characterizing the seismogram test set from
Section 5.4.2. Note that the Kennett algorithm is used here with a 30-Hz Ricker
wavelet and that all seismograms are noise-free. The seismograms from both the
training set and the test set are then input into kernel principal component analysis
algorithms (KPCA; Chapter 4) to generate seismic attributes.
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 192
Figure 5.9: Synthetic seismograms corresponding to five sets of sand-shale sequences.
Model I to V correspond to sequence realizations generated from various transition matrices, sand properties, and layer thicknesses. Refer to Table 5.1 for detailed specifications.
5.4.5 Effect of reservoir properties and stacking pattern on seismic signatures
The seismic attributes generated by KPCA are the projections of the input
seismograms onto the principal components (Figure 5.10). Note that we also run MDS
algorithms (Chapter 4) and KPCA with different kernels (e.g., linear kernel, Gaussian
kernel, and dynamic similarity kernel) and observe similar results. Therefore, here we
show only the results from KPCA with a linear kernel, and we refer to these linear
KPCA attributes as the KPCA attributes, unless otherwise specified. The results from
the training set correspond to points color-coded by net-to-gross ratios from their
corresponding sequences, and the results from the test set (shown in black symbols)
are plotted on top of the results from the training set. Figure 5.10 shows that the
training groups with different stacking pattern, saturating fluid, and reservoir thickness
are nicely separated as distinct clusters. Within these clusters, there are also apparent
trends of net-to-gross ratios. Figure 5.10 also illustrates that all the test points are
placed correctly in their corresponding training groups (e.g., the point corresponds to
the fining-upward log is located on top of the upward-decreasing net-to-gross group).
From the locations of all the test points, we can roughly estimate their net-to-gross
ratios to be between 0.45 and 0.55, which is consistent with what we previously
estimated using the Thomas-Stieber model in Section 5.4.1. Using the training set we
successfully infer the net-to-gross ratio, stacking pattern, saturating fluid, and reservoir
thickness of the test set.
tim
e (
ms)
Model I Model II Model III Model IV Model V
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 193
Figure 5.10: Projections of the noise-free seismograms from the training set and the
test set onto the first two principal components after the application of the linear KPCA. Each point from the training set is color-coded by the net-to-gross ratio from its corresponding sand-shale sequence. The test set, shown in black symbols, is plotted on top of the projections of the training set.
5.4.6 Discussion
In the previous section, we demonstrated that KPCA results of the training set
show distinct seismic signatures which correspond to changes in net-to-gross ratio,
stacking pattern, saturating fluid, and reservoir thickness, and that these results can be
used to characterize the seismogram test set successfully. This section discusses
several factors that could also influence seismic signatures.
5.4.6.1 Link between distance-based attributes and amplitude attributes
Using trace amplitudes as inputs, KPCA algorithms output coordinate values
which are projections of the traces onto the principal component axes. These
coordinate values and the principal components may not directly relate to one
particular physical property of the reservoirs (e.g., porosity). Instead, they might be
more connected to the seismic traces themselves (e.g., shape). In order to illustrate a
possible link between KPCA attributes and characteristics of seismic traces, we extract
-0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
First principal component
Second p
rincip
al com
ponent
color-coded by net-to-gross ratios
0.35
0.4
0.45
0.5
0.55
0.6
Model I
Model II
Model III
Model IV
Model V
fining-upward log
coarsening-upward log
serrated log
brine-saturated,
finning-upward log
thinner-layered,
finning-upward log
thickness
effect
fluid effect
increasing
net-to-gross
ratios
changing
stacking
patterns
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 194
the following amplitude attributes. We pick the maximum amplitude and compute the
root-mean-square amplitude (RMS) from all the seismograms. A cross-plot between
the two amplitude attributes is shown in Figure 5.11. This plot illustrates trends
similar to those in the KPCA results (Figure 5.10). The first and second principal
components of the KPCA results correlate very well with the maximum amplitude and
the RMS amplitude, respectively. The amplitude attributes can also be used
successfully to infer net-to-gross ratio, stacking pattern, saturating fluid, and reservoir
thickness of the sequences in the test set. However, even though the three stacking-
pattern clusters of the amplitude attributes are separated, the separation is relatively
slight. Therefore, small noise added to seismograms can easily alter the amplitude
attributes so that these clusters may become more overlapped. Consequently, when
using the amplitude attributes, presence of noise can reduce the odds of correctly
inferring a stacking pattern of a sequence from an unknown seismogram.
Figure 5.11: Cross-plot between the RMS and the maximum amplitude of noise-free seismograms from both the training set and the test set. The points from the training set are color-coded by the net-to-gross ratios from their corresponding sand-shale sequences. The points from the test set are shown in black symbols. Two distinct trends corresponding to changes in stacking pattern and net-to-gross ratio are marked by the blue and the red arrows, respectively.
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.160.028
0.03
0.032
0.034
0.036
0.038
0.04
Maximum amplitude
RM
S a
mplit
ude
color-coded by net-to-gross ratios
Model I
Model II
Model III
Model IV
Model V
fining-upward log
coarsening-upward log
serrated log
brine-saturated,
fining-upward log
thinner-layered,
finning-upward log
0.35
0.4
0.45
0.5
0.55
0.6
thickness
effect
increasing
net-to-gross
ratios
changing
stacking
patterns
fluid
effect
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 195
5.4.6.2 Effect of noise
To investigate the effect of noise on both the feature-extraction based (e.g., KPCA)
and the amplitude attributes, we add random noise to all seismograms from both the
training set and the test set before extracting their attributes. The linear KPCA and the
amplitude attributes are shown in Figure 5.12 and Figure 5.13, respectively. For a
better look at the noise effect, we use only sequences with varying net-to-gross ratio
and different stacking pattern. KPCA attributes for the noise-free case separate the net-
to-gross and the stacking-pattern trends nicely; however, these trends are smeared out
when level of noise increases. Any interpretation using the smeared trends is subject to
more uncertainty. Note that even though the clusters of points in the noisy case
sometimes change their directions from those of the noise-free case (e.g., from
northeast-southwest to northwest-southeast alignments), these changes do not affect
our interpretation. Any rotation or translation of the KPCA results does not matter.
Both the net-to-gross and stacking-pattern trends in amplitude attributes are also
affected and smeared out by noise. To quantify and compare the effect of noise on
KPCA and amplitude attributes, we count the number of times the linear discriminant
analysis algorithm (LDA) successfully sorts the attribute vectors (i.e., the seismograms)
into their corresponding net-to-gross/stacking-pattern classes. Net-to-gross interval is
divided into three subintervals: (1) NTG ≤0.46, (2) NTG >0.46 and NTG ≤0.54, and (3)
NTG >0.54, and the stacking pattern contains three subsets: fining-upward (F),
coarsening-upward (C), and serrated (S). The combination of both the net-to-gross
subintervals and the stacking-pattern subsets results in a total of nine classes: NTG1 +
F, NTG2 + F, NTG3 + F, NTG1 + C, NTG2 + C, NTG3 + C, NTG1 + S, NTG2 + S, and
NTG3 + S. An example of KPCA attributes with class labels is shown in Figure 5.14.
The classification success rate is estimated using a stratified 10-fold cross validation
(Chapter 4). Figure 5.15 illustrates the success rates for both the KPCA and the
amplitude attributes at different noise levels. Both types of attributes are affected by
noise; however, the amplitude attributes are more severely affected, as shown by their
rapidly decreasing classification success rates. Note that in this example, using more
components of KPCA results as attributes does not increase classification success rate.
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 196
This is possibly because the well-log interval considered here is relatively short. As a
result, the corresponding seismograms with different properties show difference in
amplitudes only on a small portion of the traces (i.e., about two wavelength).
Therefore, with just a few principal components we are able to capture all the
important features. As previously noted, in this case study we show only results from
the linear KPCA because KPCA with other kernels, including Gaussian kernel and
dynamic similarity kernel, yields similar results. When we investigate the effect of
noise on the dynamic similarity KPCA, we observe that its results do not differ much
from those of the linear KPCA (Figure 5.15).
In summary, even though the amplitude attributes can be useful for property
estimation of thin sand-shale sequences, they can be very sensitive to noise. While the
feature-extraction-based algorithms (e.g., KPCA) also extract important information
from seismograms by comparing their amplitudes, the resulting attributes are less
sensitive to noise.
Figure 5.12: Projections of the noisy seismograms from the test set onto the first two principal components from the linear KPCA. The test set shown in black symbols is plotted on top of the projections of the training set, which are color-coded by the net-to-gross ratios from their corresponding sand-shale sequences. The percentages of noise added to the seismograms are shown in the lower left corner of each plot.
-0.05 0 0.05-0.04
-0.02
0
0.02
0.04
First principal component
Second p
rincip
al com
ponent
-0.05 0 0.05-0.04
-0.02
0
0.02
0.04
First principal component
Second p
rincip
al com
ponent
-0.05 0 0.05-0.04
-0.02
0
0.02
0.04
First principal component
Second p
rincip
al com
ponent
-0.05 0 0.05-0.04
-0.02
0
0.02
0.04
First principal component
Second p
rincip
al com
ponent
0.35
0.4
0.45
0.5
0.55
0.6
Model I
Model II
Model III
fining-
upward log
coarsening-
upward log
serrated log
net-to-gross ratios
0% noise 5% noise
10% noise 20% noise
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 197
Figure 5.13: Cross-plot between the RMS and the maximum amplitude of noisy seismograms from both the training set and the test set. The points from the training set are color-coded by the net-to-gross ratios from their corresponding sand-shale sequences, and the points from the test set are shown in black symbols. The percentages of noise added to the seismograms are shown in the lower left corner of each plot.
Figure 5.14: Linear KPCA results (0% noise) with net-to-gross (NTG)/stacking-pattern class labels. NTG1, NTG2, and NTG3 represent NTG≤0.46, NTG>0.46 and NTG≤0.54, and NTG>0.54, respectively. The three stacking patterns are fining-upward (F), coarsening-upward (C), and serrated (S).
0.12 0.14 0.16
0.036
0.038
0.04
0.042
Maximum amplitude
RM
S a
mplit
ude
0.12 0.14 0.16
0.036
0.038
0.04
0.042
Maximum amplitude
RM
S a
mplit
ude
0.12 0.14 0.16
0.036
0.038
0.04
0.042
Maximum amplitude
RM
S a
mplit
ude
0.12 0.14 0.16
0.036
0.038
0.04
0.042
Maximum amplitude
RM
S a
mplit
ude
0.35
0.4
0.45
0.5
0.55
0.6
Model I
Model II
Model III
fining-
upward log
coarsening-
upward log
serrated log5% noise0% noise
10% noise 20% noise
net-to-gross ratios
-0.04 -0.02 0 0.02 0.04 0.06-0.03
-0.02
-0.01
0
0.01
0.02
0.03
Linear KPCA results (0% noise)
color-coded by nine net-to-gross/stacking-pattern classes
First principal component
Second p
rincip
al com
ponent
NTG1 + S
NTG2 + S
NTG3 + S
NTG1 + F
NTG2 + F
NTG3 + F
NTG1 + C
NTG2 + C
NTG3 + C
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 198
Figure 5.15: Classification success rate when using linear KPCA attributes, amplitude attributes, and dynamic similarity KPCA attributes at various noise levels.
5.4.6.3 Effect of layer thickness used to model the training set
In Section 5.4.3, when modeling the sequences for the training set, using
information from the core analysis we assigned each layer a thickness of 1.5 cm,
equivalent to wavelength to thickness ratio of about 50. In this section, we investigate
the effect on the attribute results of varying layer thickness. We increase the layer
thickness to 2.5 cm, equivalent to wavelength to thickness ratio of about 30 and
decrease the number of layers so that the total thicknesses of the sequences are kept
the same. The linear KPCA results and the amplitude results for the two thickness
models are shown in Figure 5.16 and Figure 5.17, respectively. We show only results
from sequences that are all fining-upward, but with varying net-to-gross ratios. To
quantify and compare the effect of different modeling thicknesses on KPCA and
amplitude attributes, we use each set of attributes and compute their success rates for
classifying the seismograms into three net-to-gross classes: (1) NTG ≤0.46, (2) NTG
>0.46 and NTG ≤0.54, and (3) NTG >0.54. The classification success rates are
displayed in the lower corner of each plot in both Figure 5.16 and Figure 5.17. When
the layer thickness in the sequences changes from 1.5 to 2.5 cm, both the KPCA and
the amplitude attributes show little to no change, and their classification success rates
also stay almost exactly the same.
0 5 15 200.2
0.4
0.6
0.8
1
Cla
ssific
ation s
uccess r
ate
% of noise added
Linear KPCA
(the first tw o components)
Amplitude attributes
(maximum and RMS)
Dynamic similarity KPCA
(the first tw o components)
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 199
Figure 5.16: Effect of layer thickness on linear KPCA results. Each plot represents the projections of seismograms onto the first two principal components of linear KPCA when layer thickness is 1.5 cm (left) or 2.5 cm (right). Points are color-coded by the net-to-gross ratios from their corresponding sand-shale sequences.
Figure 5.17: Effect of layer thickness on amplitude results (RMS versus maximum amplitude). Layer thickness used for modeling sequences is 1.5 cm for the left plot and 2.5 cm for the right plot. Points are color-coded by the net-to-gross ratios from their corresponding sand-shale sequences.
5.4.6.4 Effect of impedance contrast
Figure 5.8 illustrates sand and shale impedances used for modeling the sequences
in the training set. In this section, we investigate the effect on the attribute results of
varying impedance contrast between the sand and the shale. We simulate sequences
with two different sets of impedances. The original set represents a larger impedance
contrast between the sand and the shale (i.e., a smaller overlap in impedances), while
the other set represents a smaller contrast (i.e., a larger overlap) (Figure 5.18). . The
-0.04 -0.02 0 0.02 0.04-0.015
-0.01
-0.005
0
0.005
0.01
0.015
First principal component
Second p
rincip
al com
ponent
1.5-cm thick layers
-0.04 -0.02 0 0.02 0.04-0.015
-0.01
-0.005
0
0.005
0.01
0.015
First principal component
Second p
rincip
al com
poent
2.5-cm thick layers
0.4
0.45
0.5
0.55
0.6
net-to-gross ratios
classification
success rate = 0.88
classification
success rate = 0.89
0.13 0.135 0.14 0.145 0.150.0365
0.037
0.0375
0.038
0.0385
0.039
0.0395
0.04
0.0405
Maximum amplitude
RM
S a
mplit
ude
1.5-cm thick layers
0.13 0.135 0.14 0.145 0.150.0365
0.037
0.0375
0.038
0.0385
0.039
0.0395
0.04
0.0405
Maximum amplitude
RM
S a
mplit
ude
2.5-cm thick layers
0.4
0.45
0.5
0.55
0.6
net-to-gross ratios
classification
success rate = 0.87classification
success rate = 0.86
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 200
linear KPCA results and the amplitude results for the two impedance sets are shown in
Figure 5.19 and Figure 5.20, respectively. We show only results from sequences that
are all fining-upward, but with varying net-to-gross ratios. To quantify and compare
the effect of impedance contrast on KPCA and amplitude attributes, we compute the
success rates of classifying the seismograms into three net-to-gross classes, similar to
Section 5.4.6.3. The classification success rates are displayed in the lower corner of
each plot in both Figure 5.19 and Figure 5.20. When the impedances of the sand and
the shale become more overlapped (i.e., small contrast), the linear KPCA attributes
show little to no change, and their classification success rate stays almost exactly the
same. However, the smaller impedance contrast shows a significant effect on the
amplitude attributes by decreasing their success rate from 86% down to 52%.
Figure 5.18: Probability density functions of two sets of acoustic impedance showing small and large contrast (shown in red and green, respectively) between the sand and the shale (shown in solid and dash lines, respectively).
4 4.5 5 5.5 60
2
4
6
8
10
Acoustic impedance (g/cm3 x km/s)
Pro
babili
ty d
ensity
sand
(large contrast)
shale
sand
(small contrast)
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 201
Figure 5.19: Effect of impedance contrast on linear KPCA results. Each plot represents the projections of seismograms onto the first two principal components of linear KPCA when the overlap between the sand and the shale impedances is small (left) or large (right). Points are color-coded by the net-to-gross ratios from their corresponding sand-shale sequences.
Figure 5.20: Effect of impedance contrast on amplitude results (RMS versus maximum amplitude). The overlap between the sand and the shale impedances is small in the left panel and large in the right panel. Points are color-coded by the net-to-gross ratios from their corresponding sand-shale sequences.
5.4.6.5 Effect of wavelet phase rotation
In this section, we investigate how KPCA attributes are affected by wavelet phase.
First, we use the synthetic seismogram generated at the well with a zero-phase Ricker
wavelet. Then we generate two additional seismograms using -20° phase and +40°
phase Ricker wavelets. These three seismograms at the well are our test traces. Linear
KPCA attributes of the three seismograms at the well and the Model I training set
-0.04 -0.02 0 0.02 0.04-0.015
-0.01
-0.005
0
0.005
0.01
0.015
First principal component
Second p
rincip
al com
ponent
Impedances with a small overlap
-0.04 -0.02 0 0.02 0.04-0.015
-0.01
-0.005
0
0.005
0.01
0.015
First principal component
Second p
rincip
al com
ponent
Impedances with a large overlap
0.4
0.45
0.5
0.55
0.6
0.65
classification
success rate = 0.88
net-to-gross ratios
classification
success rate = 0.89
0.12 0.125 0.13 0.135 0.14 0.145
0.035
0.036
0.037
0.038
0.039
Maximum amplitude
RM
S a
mplit
ude
Impedances with a small overlap
0.12 0.125 0.13 0.135 0.14 0.145
0.035
0.036
0.037
0.038
0.039
Maximum amplitude
RM
S a
mplit
ude
Impedances with a large overlap
0.4
0.45
0.5
0.55
0.6
0.65
classification
success rate = 0.86
classification
success rate = 0.52
net-to-gross ratios
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 202
(Section 5.4.5) are shown in Figure 5.21. Since the training set is generated using a
zero-phase Ricker wavelet, the only test trace that overlaps with this training set is the
one with zero phase. In addition to the training and the test traces, we also add the
KPCA representation of the zero trace (i.e., trace that has all elements equal to zero) as
a reference point. The angle between the vector from the reference point to the -20°
phase trace and the vector from the reference point to the zero-phase trace is equal to
20°. Also, the angle between the vector from the reference point to the +40° phase
trace and the vector from the reference point to the zero-phase trace is equal to 40°.
Note that this is also true when the attribute vectors are replaced by the actual seismic
traces. In this example, phase of the wavelet affect the linear KPCA attributes by
rotating each attribute point around the reference point by an angle of that absolute
phase value. The rotation is clockwise if the phase is positive and counterclockwise if
the phase is negative.
Figure 5.21: Effect of phase rotation on linear KPCA results. Points represent projections of the seismograms from the Model I training set and the well location onto the first two principal components from the linear KPCA. Each point is color-coded by the net-to-gross ratio from its corresponding sand-shale sequence. The three seismograms at the well with three different phases shown in black symbols are plotted on top of the projections of the training set. The projection of the zero trace is used as a reference point.
-0.1 0 0.1 0.2 0.3 0.4-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
First principal component
Second p
rincip
al com
ponent
color-coded by net-to-gross ratios
0.4
0.45
0.5
0.55
0.6
Model I
zero trace
0 degree
-20 degree
+40 degree
20o
40o
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 203
5.5 Case study 2: Estimating sub-resolution reservoir properties
from a 2-D seismic section
In the previous case study (Section 5.4), we investigated the effect of net-to-gross
ratio, stacking pattern, saturating fluid, and reservoir thickness on the seismic
signatures by using real well log data, but with synthetic seismograms. In this case
study, we use real seismograms from a 2-D section extracted along a turbidite channel
from proximal to distal directions.
5.5.1 2-D seismic section
The 2-D seismic section is shown in Figure 5.22. The location of well A is marked,
and the proximal direction of the channel is on the left side of the plot. In this case
study, we still focus on the same fining-upward interval as discussed in the previous
case study (Section 5.4). This interval at well A corresponds to a seismic segment
shown in the black rectangle (Figure 5.22). By following the same top and bottom
marks of this interval, we extract seismic segments across the 2-D section.
Figure 5.22: A 2-D seismic section extracted along the turbidite channel from proximal (left) to distal (right) directions. The location of well A is marked.
5.5.2 Rock property calibration at well location
In order to generate realizations of sand-shale sequences and their corresponding
synthetic seismograms which will be used as a training set, we first need to estimate
transition matrices and rock properties for each lithologic state from the information at
Tim
e (
ms)
Trace #
0 10 20 30 40 50
Well A
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 204
the well. However, since we focus on the same fining-upward interval as discussed in
the previous case study (Section 5.4), we simply use the same matrices and rock
properties which are already estimated in Section 5.4.3 and follow the same steps
described in that section to simulate multiple realizations of sand-shale sequences. The
steps are summarized as follows: (1) select a transition matrix that represents the
desired stacking pattern, (2) use the matrix to generate arrangement of sand and shale
layers, and (3) assign rock properties to each layer by drawing from the distribution
corresponding to its lithologic state. In this case study, all the sand layers are assumed
to be oil-saturated. All three stacking patterns are also included in generating the
training set by following the same step shown in Section 5.4.3.2. However, the
transition-matrix sets for the top and bottom sub-sequences are defined as follows:
𝑇(𝑡𝑜𝑝 ) = 𝑎 1 − 𝑎𝑏 1 − 𝑏
, where a and b are drawn from a uniform distribution on [0,1],
𝑇(𝑏𝑜𝑡 ) = 𝑐 1 − 𝑐𝑑 1 − 𝑑
, where e and f are drawn from a uniform distribution on [0,1].
When we generate the training set, we obtain rock properties of the sediments that are
underlain and overlain the reservoir interval directly from well A.
In order to expand our training set, we also vary reservoir thickness which is
assumed to be uniformly distributed between 8.5m and 25.6 m (Figure 5.23, left
column). Net-to-gross ratios for all sequence models are also recorded, and their
overall distribution is shown in Figure 5.23 (right column).
Figure 5.23: Prior distributions for thickness (left) and net-to-gross ratio (right), assigned to sequence models when generating a training set.
10 15 20 250
1000
2000
3000
Thickness (m)
Counts
0 0.5 10
2000
4000
6000
Net-to-gross ratio
Counts
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 205
5.5.3 Non-stationarity when moving away from the well
For the two existing wells, each of which is located on each side of the 2-D
seismic section (Figure 5.24), we observe that the average impedances for both the
overburden and the underburden are different and that these impedances in the
proximal direction are lower than those in the distal direction (Figure 5.25). This
observation shows spatial non-stationarity of rock properties. Another example of the
non-stationarity for this same dataset was provided by Dutta (2009), who showed that
the sand-rich facies becomes better sorted, and its quartz/clay ratio increases spatially
from distal to proximal directions. We observe that the impedance of shale also
increases from the distal to proximal wells, but the increase is very small.
Figure 5.24: Relative locations of two existing wells (well A and well B). Well A (black symbol) is located in the distal direction along the channel, and well B (red symbol) is located in the proximal direction along the channel. Distance is measured from a reference point. The selected 2-D seismic line is shown in blue.
0 0.25 0.5 0.75
0
0.25
0.5
0.75
1
1.25
Distance (km)
Dis
tance (
km
)
well B
(proximal)
well A
(distal)
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 206
Figure 5.25: Average properties of the overburden and underburden of the interval of interest from proximal and distal wells. Three properties are shown: P-wave velocity (Vp), bulk density (RHOB), and acoustic impedance (AI).
We incorporate the non-stationarity of rock properties into our modeling steps as
follows. First, to account for spatial variations in the overburden and underburden, we
vary two additional parameters, which are referred to as the overburden and the
underburden multipliers. These multipliers are simply numbers drawn from uniform
distributions over ranges that are estimated from the wells, and these numbers are used
to multiply the overburden and underburden rock properties extracted from well A.
Second, to account for spatial variations of the sand and shale lithologic states within
the target zone, we assign standard deviations to the distributions of impedances for
both the sand state and the shale state so that these distributions cover such variations.
The distributions are shown in Figure 5.26. Note that since the lateral extent of our 2-
D section is not significantly large, the spatial variations in rock properties can be
dealt with by adjusting the distributions. However, when exploring areas significantly
away from the well such that rock properties change vastly, we recommend creating
additional training sets where the rock properties are chosen appropriately (e.g., using
a linear interpolation between two wells to estimate suitable rock properties). Then a
training set can be selected and applied to a group of test points that are located in the
same neighborhood. In the case where only one well is available, spatial trends in
0
1000
2000
3000
4000
5000
6000
distal well
proximal well
Vp RHOB AI Vp RHOB AI
(m/s) (kg/m3) (kg/m2s) (m/s) (kg/m3) (kg/m2s)
Overburden Underburden
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 207
sedimentological parameters and rock physics models are key tools for extrapolating
the rock properties of the lithologic states in the area away from the well (e.g., Dutta,
2009).
Figure 5.26: Probability density functions of acoustic impedance for the sand (blue) and the shale (red) lithologic states used in generating a training set.
5.5.4 Synthetic seismogram generation and seismic attribute extraction
After obtaining sequence models with assigned rock properties, we generate the
corresponding synthetic seismograms using the Kennett algorithm (Kennett, 1983) to
simulate full-waveform, normally-incident, reflected seismograms using a zero-phase
Ricker wavelet with a central frequency of 50 Hz. This frequency is selected to match
with the seismic data. The simulations are performed for all reverberations. These
seismograms serve as our training set. Since the wavelet of the real seismograms is not
a Ricker wavelet, we transform the real seismograms with their original wavelet to
match the training set by using
𝑆𝑟𝑒𝑎𝑙∗ 𝜔 = 𝑆𝑟𝑒𝑎𝑙 𝜔 ∙ 𝐻 𝜔 ,
where 𝐻 𝜔 is the Fourier transform of a filter function, and
𝐻 𝜔 =𝑆syn
well A 𝜔
𝑆realwell A 𝜔
.
In the above relations, 𝑆𝑟𝑒𝑎𝑙 and 𝑆𝑟𝑒𝑎𝑙∗ are the Fourier transforms of the seismogram
before and after applying the transfer function, respectively. 𝑆synwell A and 𝑆real
well A are the
4.6 4.8 5 5.2 5.4 5.60
2
4
6
8
10
Acoustic impedance (g/cm3 x km/s)
Pro
babili
ty d
ensity
oil-saturated
sand
shale
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 208
Fourier transform of the synthetic seismogram at well A generated using a Ricker
wavelet and the Fourier transform of the real seismogram at well A with its original
wavelet, respectively. After applying the above relations, the Fourier transforms of the
filtered seismograms are transformed back to the time domain.
The filtered seismograms and all the seismograms from the training set are then
input into KPCA algorithms. Here we show only the results from KPCA with a linear
kernel. The projections of all the input seismograms onto the first two principal
components are shown in Figure 5.27. The projections of the training set are color-
coded by reservoir thickness, and the projections of the real seismograms are labeled
by their trace numbers. Each panel in Figure 5.28 shows the results where each
training point is color-coded by the parameter specified at the top of that panel. An
example of results from KPCA with a Gaussian kernel is shown in Figure 5.29, which
shows trends similar to those from the linear KPCA. In this case study, we do not
focus on an extensive comparison of performance of various kernels. Hereafter, we
refer to the linear KPCA attributes as the KPCA attributes, unless otherwise specified.
Figure 5.27: Projections of the training set and the test seismograms onto the first two principal components of KPCA with a linear kernel. The training points are color-coded by the reservoir thicknesses of their corresponding sand-shale sequences. The test points are labeled by their trace numbers.
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 209
Figure 5.28: Projections of the training set onto the first two principal components of KPCA with a linear kernel. In each panel, the training points are color-coded by the reservoir parameter specified at the top of each panel.
Figure 5.29: Projections of the training set onto the first two principal components of KPCA with a Gaussian kernel. In each panel, the training points are color-coded by the reservoir parameter specified at the top of each panel.
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 210
While there are noticeable trends in KPCA attributes associated with changes in
reservoir thickness and the underburden multiplier, there is almost no trend associated
with varying net-to-gross ratio (Figure 5.28). From this observation, net-to-gross ratio
does not seem to affect seismic signatures (i.e., KPCA attributes). However, when all
other parameters are fixed and only net-to-gross ratio is varying, a net-to-gross trend
becomes visible on the KPCA attribute space (Figure 5.30). The visibility of the net-
to-gross trend implies that each parameter does not affect seismic signatures equally
and that effects of some parameters can be so large that they overshadow effects of
other parameters.
Figure 5.30: Projections of the training set onto the first two principal components of KPCA with a linear kernel, when the values of underburden multiplier, overburden multiplier, and reservoir thickness are fixed. Points are color-coded by net-to-gross ratio.
Even though we show the KPCA results on the first two principal components, we
use the first ten components as our attributes in all later analyses. The number of
components included as attributes is chosen based on the eigenvalues of the KPCA
results (Figure 5.31). The first ten principal components account for 99% of the total
variance.
0 2000 4000 6000 8000 10000 12000-1
-0.5
0
0.5
1x 10
4
First principal component
Second p
rincip
al com
ponent
Color-coded by net-to-gross ratio
0.2
0.4
0.6
0.8
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 211
Figure 5.31: The first 20 eigenvalues of the linear KPCA results.
5.5.5 Sensitivity analysis
In the previous case study (Section 5.4), we generate sequence models where we
vary one reservoir parameter at a time. In this case study, we vary all parameters at the
same time. Since different parameters can have different impact on seismic signatures,
a parameter with a greater impact may overshadow subtle effects from other
parameters (e.g., Figure 5.28). In this section, in order to rank the input parameters
(i.e., reservoir properties) according to their impacts, we perform a sensitivity analysis
by following the method of Fenwick et al (2012), in which the estimated sensitivity for
the ith
parameter 𝑠(𝑝𝑖) using
𝑠 𝑝𝑖 =1
𝐾
𝑑 𝐹,𝑘
𝑑 𝐹,𝑘(95)
𝐾
𝑘=1.
In the above equation, the integer i ranges from 1 to 4, and these numbers correspond
to the following parameters: overburden multiplier, underburden multiplier, reservoir
thickness, and net-to-gross ratio. The training set is divided into 3 classes (i.e., 𝐾 = 3)
based on their Euclidean distances in the KPCA space by using K-mean clustering. 𝑐𝑘
is the kth
class, and the integer k ranges from 1 to 3. For each parameter, 𝑑 𝐹,𝑘 is the L-1
norm measure of difference (or distance) between its class-conditional empirical
cumulative distribution 𝐹 (𝑝𝑖|𝑐𝑘) and its prior empirical distribution 𝐹 (𝑝𝑖), and this
difference can be estimated by integrating the area between the two curves. 𝑑 𝐹,𝑘(95)
is the
95th
-percentile of bootstrapped L-1 norm distances. Note that for a given parameter 𝑝𝑖 ,
5 10 15 200
2
4
6
8
10
12x 10
8
Principal component
Eig
envalu
e
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 212
if there exists a class 𝑘 where 𝑑 𝐹,𝑘 ≥ 𝑑 𝐹,𝑘(95)
, then the parameter 𝑝𝑖 is defined as being
sensitive to the classification based on the distances in the KPCA space.
The strategy behind this method is that if a parameter has an impact on seismic
signatures which, in our study, are represented by projections of seismograms into the
principal-component space, then this parameter will separate the projected
seismograms into different classes. Consequently, the statistical distributions of the
sample values of this parameter among all classes will be dissimilar. 𝑑 𝐹,𝑘 is used as a
measure of such dissimilarity. Note that this sensitivity analysis does not account for
any interaction (or co-existence) between two or more parameters that can jointly
affect seismic signatures (Fenwick et al., 2012).
Figure 5.32 illustrates a comparison between a class-conditional empirical
cumulative distribution and a prior empirical distribution of the overburden multiplier,
and the shaded area is the difference measure between the two distributions. Figure
5.33 shows results of our sensitivity analysis. For each parameter, each bar represents
the difference 𝑑 𝐹,𝑘 normalized by 𝑑 𝐹,𝑘(95)
for class 𝑘. The red line represents a threshold.
For a parameter, if there is any class where the normalized difference exceeds this line,
then that parameter is considered to be an impacting parameter. Out of the four
parameters, only the net-to-gross ratio is considered a non-impacting parameter
because none of its classes have the normalized difference exceed the threshold. This
result is consistent with our previous observation that net-to-gross ratio does not seem
to affect seismic signatures (Figure 5.28). However, as previously shown, when all
other parameters are fixed and only net-to-gross ratio is varying, a net-to-gross trend
becomes visible on the KPCA attribute space (Figure 5.30). The pareto plot in Figure
5.34 shows the ranking of the four reservoir parameters according to their estimated
sensitivity values (𝑠). Reservoir parameters in decreasing of the sensitivity values are
reservoir thickness, underburden multiplier, overburden multiplier, and net-to-gross
ratio.
As previously noted, some parameters can overshadow the subtle effects on
seismic signatures from the other parameters. Therefore, for each seismic test trace we
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 213
first estimate probability density functions of the three parameters: reservoir thickness,
underburden multiplier, and overburden multiplier. Then for each test trace we create
a new training set by varying the net-to-gross parameter within the same range, but the
thickness and the two multipliers are varied within smaller intervals around their mean
values. Note that a “new” training set may not be needed if the original training set
contains a sufficiently large numbers of points. In this case, we can simply select a
subset of training points that have the values of the thickness and the multipliers fall
within the desired ranges. By limiting the values of an influential parameter within a
smaller range, we can have a better look at the effect of other parameters. Even though
we are not interested in the values of both the underburden multiplier and the
overburden multiplier, the effects of both parameters overshadow the effect from net-
to-gross ratio. Therefore, we also need to estimate posterior distributions for both
multipliers.
In summary, for each seismic test trace we will estimate probability density
functions of the first three parameters which are reservoir thickness, underburden
multiplier, and overburden multiplier, and after doing so we create a new training set
in which these previously estimated parameters will vary uniformly over small ranges
around their mean values. The new training set is then used to estimate net-to-gross
ratio.
Figure 5.32: Comparison of two empirical cumulative distributions for the overburden multiplier values in class 3. The prior distribution and the class-conditional distribution are shown in blue and red, respectively. The shaded area represents area between the two curves.
0.8 0.85 0.9 0.95 1 1.05 1.10
0.2
0.4
0.6
0.8
1
Overburden multiplier
Cum
ula
tive p
robabili
ty
prior empirical distribution
class-conditional
empirical distribution
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 214
Figure 5.33: Normalized difference measure between a class-conditional empirical cumulative distribution and a prior empirical distribution for a parameter. The red line represents a threshold used in determining whether a parameter has a significant impact on the output response (i.e., seismic signature).
Figure 5.34: Pareto plot of estimated sensitivity values. Parameters are ranked according to their sensitivity values.
5.5.6 Property estimation of the 2-D seismic section
In the previous sections, we showed the KPCA results of both the training set and
the test set (Figure 5.27 – Figure 5.28), then performed sensitivity analyses, and
presented the order of the reservoir parameters to be estimated which are reservoir
thickness, underburden multiplier, overburden multiplier, and net-to-gross ratio. In this
section, we use these results to estimate reservoir properties of the 2-D seismic section
(Figure 5.22).
To estimate a probability density of a reservoir parameter 𝑝 for a test trace, we use
a k-nearest neighbor (kNN) method which finds k closet points (in a training set) to a
0 4 8 12
Underburden multiplier
Overburden multiplier
Reservoir thickness
Net-to-gross ratio
Normalized difference between distributions
class 1
class 2
class 3
sensitivity threshold
0 2 4 6 8 10 12
Net-to-gross ratio
Overburden multiplier
Underburden multiplier
Reservoir thickness
Estimated sensitivity value
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 215
query (or test) point. Assume that the value of 𝑝 is subdivided into equal 𝑛𝑝classes,
each of which is denoted by 𝑝𝑙 for 𝑙 = 1, 2, …, 𝑛𝑝 . We want to find the posterior
probability of parameter 𝑝 given a test point 𝑥 or 𝑃(𝑝𝑙|𝑥). Using the kNN method, this
probability according to Bayes’ theorem is written as
𝑃 𝑝𝑙 𝑥 =𝑃(𝑥, 𝑝𝑙)
𝑃(𝑥, 𝑝𝑗 )𝑛𝑝
𝑗=1
= 𝑘𝑙
𝑘 ,
where 𝑃 𝑥, 𝑝𝑙 is the joint probability estimated as (𝑘𝑙 𝑛 /𝑉. Within a volume 𝑉
surrounding the test point 𝑥, 𝑘𝑙 is the number of the training points that belong to class
𝑝𝑙 , and 𝑛 is the total number of the training points (Tobin et al., 2009).
Using the above equation, we first estimate probability densities for reservoir
thickness, underburden multiplier, and overburden multiplier from the linear KPCA
results (Figure 5.27). Their estimated densities and their mean values are shown in
Figure 5.35 – Figure 5.40. The subsequent estimated probability density and the mean
values for net-to-gross ratio are shown in Figure 5.41 – Figure 5.42. Note that the
estimated underburden multipliers generally increase from left (i.e., proximal) to right
(i.e., distal). This trend is consistent with the observation that the average impedance
of the underburden in the proximal direction is smaller than the impedance in the
distal direction (Figure 5.25). In order to evaluate the performance of our property
estimation, we add three additional seismic traces to the test set, each of which
corresponds to a thin sand-shale sequence with reservoir parameters listed in Table 5.1.
Table 5.1: Summary of reservoir parameters for three seismic traces which are added to the test set for evaluating the performance of property estimation.
Trace # Underburden
multiplier
Overburden
multiplier
Reservoir
Thickness
Water
saturation
Net-to-
gross
ratio
Stacking
pattern
55 (real) 1 1 17 0 ~0.48 Fining-
upward
56
(synthetic)
Well B Well B 17 0 ~0.48 Fining-
upward
57
(synthetic)
0.95 0.87 17 0 ~0.48 Coarsening-
upward
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 216
Figure 5.35: Estimated probability density functions for reservoir thickness for all seismic traces in the test set. The last three traces (#55-57) correspond to the additional three traces which are added for evaluating the performance of our property estimation. Each density function is shown as a column-wise color scale.
Figure 5.36: The mean values of reservoir thickness for all seismic traces in the test set. The last three points are added for evaluating the performance of our property estimation.
Figure 5.37: Estimated probability density functions for the underburden multiplier for all seismic traces in the test set. The last three traces (#55-57) correspond to the additional three traces which are added for evaluating the performance of our property estimation. Each density function is shown as a column-wise color scale.
Figure 5.38: The mean values of underburden multiplier for all seismic traces in the test set. The last three points are added for evaluating the performance of our property estimation.
Trace #
Reserv
oir
thic
kness (
m)
Probability
density
5 10 15 20 25 30 35 40 45 50 55
25.6
22.2
18.8
15.4
11.9
8.53 0
0.1
0.2
5 10 15 20 25 30 35 40 45 50 558.5
14.5
20.5
25.6
trace #
Reserv
oir
thic
kness (
m)
Trace #
Underb
urd
en
multip
lier
Probability
density
5 10 15 20 25 30 35 40 45 50 55
1.1
1.04
0.98
0.92
0.86
0.8
2
46
8
10
5 10 15 20 25 30 35 40 45 50 550.85
0.9
0.95
1
1.05
trace #
Underb
urd
en
multip
lier
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 217
Figure 5.39: Estimated probability density functions for the overburden multiplier for
all seismic traces in the test set. The last three traces (#55-57) correspond to the additional three traces which are added for evaluating the performance of our property estimation. Each density function is shown as a column-wise color scale.
Figure 5.40: The mean values of overburden multiplier for all seismic traces in the
test set. The last three points are added for evaluating the performance of our property estimation.
Figure 5.41: Estimated probability density functions for net-to-gross ratio for all
seismic traces in the test set. The last three traces (#55-57) correspond to the additional three traces which are added for evaluating the performance of our property estimation. Each density function is shown as a column-wise color scale.
Figure 5.42: The mean values of net-to-gross ratio for all seismic traces in the test set. The last three traces (#55-57) correspond to the additional three traces which are added for evaluating the performance of our property estimation.
Trace #
Overb
urd
en
multip
lier
Probability
density
5 10 15 20 25 30 35 40 45 50 55
1.1
1.06
1.02
0.98
0.94
0.924681012
5 10 15 20 25 30 35 40 45 50 55
0.95
1
1.05
trace #
Overb
urd
en
multip
lier
Trace #
Net-
to-g
ross r
atio
Probability
density
5 10 15 20 25 30 35 40 45 50 55
1
0.8
0.6
0.4
0.2
0
1
2
3
4
5
5 10 15 20 25 30 35 40 45 50 550
0.2
0.4
0.6
0.8
1
trace #
Net-
to-g
ross r
atio
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 218
To evaluate the performance of property estimation, we summarize and compare
the result at the well and the results of the three additional points with the true
parameter values (Figure 5.43 – Figure 5.45). Even though the underburden and the
overburden multipliers are not the parameters of interest, we also include their results
here in order to show the performance of the estimation. In general, the true values are
captured within the estimated densities for all the four test traces. Trace #56 shows a
bimodal distribution of the estimated thickness, which is different from the other three
traces. This may be because we generate trace #56 by using rock properties of the
underburden and the overburden sediments from well B, while the rest of the traces
(including the training set) are generated using properties from well A (either with or
without multipliers). Note that since trace #56 is associated with well B, we do not
include the “true” multipliers in Figure 5.44 and Figure 5.45.
Figure 5.43: The estimated probability densities of reservoir thickness for seismic
traces at the well and trace #55-57, which are added for evaluating the performance of our property estimation. The true thicknesses for all four test traces are marked by the magenta line.
Figure 5.44: The estimated probability densities of underburden multiplier for seismic
traces at the well and trace #55-57, which are added for evaluating the performance of our property estimation. The true multipliers for three test traces are marked by the lines shown in the same colors as their densities. Note that since trace #56 is associated with well B, the “true” multiplier of well A is irrelevant and thus not shown here.
10 15 20 250
0.05
0.1
0.15
0.2
Reservoir thickness (m)
Pro
babili
ty d
ensity
well A
Trace 55
Trace 56
Trace 57
0.85 0.9 0.95 1 1.050
2
4
6
8
10
12
Underburden multiplier
Pro
babili
ty d
ensity
well A
Trace 55
Trace 56
Trace 57
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 219
Figure 5.45: The estimated probability densities of overburden multiplier for seismic
traces at the well and trace #55-57, which are added for evaluating the performance of our property estimation. The true multipliers for three test traces are marked by the lines shown in the same colors as their densities. Note that since trace #56 is associated with well B, the “true” multiplier of well A is irrelevant and thus not shown here.
Figure 5.46: The estimated probability densities of net-to-gross ratio for seismic
traces at the well and trace #55-57, which are added for evaluating the performance of our property estimation. The true net-to-gross ratios for all four test traces are marked by the magenta line.
When generating the training set, we account for the variation in rock properties of
the underburden and the overburden (Section 5.5.3). If this variation (non-stationarity)
is not accounted for, the estimated net-to-gross ratios can be erroneous. For example,
Figure 5.47 shows the estimated probability densities of net-to-gross ratio for trace #
57 for both when the non-stationarity is taken into account and when the non-
stationarity is not taken into account. For trace #57, when the non-stationarity is
ignored, the estimated probability density of net-to-gross ratio shifts and the density
curve (almost) does not cover the true net-to-gross ratio.
0.95 1 1.050
2
4
6
8
10
12
14
Overburden multiplierP
robabili
ty d
ensity
well A
Trace 55
Trace 56
Trace 57
0.2 0.4 0.6 0.80
1
2
3
4
Net-to-gross ratio
Pro
babili
ty d
ensity
well A
Trace 55
Trace 56
Trace 57
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 220
Figure 5.47: The estimated probability densities of net-to-gross ratio for seismic trace
# 57 for both when the non-stationarity is taken into account (blue curve) and when the non-stationarity is not taken into account (red curve). The true net-to-gross ratio for trace #57 is marked by the blue line.
5.5.7 Discussion
In the previous sections, we presented a workflow for reservoir property
estimation in thinly-bedded sand-shale reservoirs and showed results of applying this
workflow to real seismic data. In this section, we provide some additional guidelines
for real application.
In the second case study, in order to match real seismic traces with the training set,
we transform those traces with their original wavelet into the traces with the Ricker
wavelet. The advantage of this transformation is that we can use the same training set
for other sets of real seismic traces that have frequency ranges similar to the range in
the training set. Even though variation in wavelets is not accounted for in this study, it
can be incorporated into the analysis by creating multiple filters for seismograms
(Section 5.5.4). Alternatively, instead of applying a transformation to real seismic data,
a wavelet can be estimated from the data and then used to create a training set.
To reduce the effects of reservoir thickness, overburden multiplier, and
underburden multiplier, we decrease the ranges of these parameters and re-create new
training sets. As a result, seismic signatures become more sensitive to parameters such
as net-to-gross ratio. However, instead of decreasing the ranges of the seismic-
signature parameters around their mean values, an alternative way to appropriately
incorporate their posterior distributions into estimation of the next parameters is to
0.2 0.4 0.6 0.80
1
2
3
4
5
Net-to-gross ratio
Pro
babili
ty d
ensity
not accounting for
non-stationarity
accounting for
non-stationarity
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 221
perform Monte Carlo simulations. By doing so, the errors of the previously estimated
parameters are propagated to the estimation of the next parameters.
Prior knowledge of the geology of the area and trends in the sedimentological
parameters (e.g., sorting) can be used to create training sets. For example, when there
is only one well available, if we know that sorting deteriorates in one particular
direction away from the well, then we can use rock physics relations to predict rock
properties and assign these properties to the appropriate lithologic states.
5.6 Conclusions
We apply our workflow for property estimation of thin sand-shale sequences to
real data from channelized turbidite deposits offshore Equatorial Guinea, West Africa.
We focus our analysis on an interbedded sand-shale interval which has bed thickness
below seismic resolution. Within this interval, the average wavelength to thickness
ratio is approximately 35 for massive sandstone, but can be greater than 100 for
interbedded thin-bedded sandstone and mudstone. In the first case study, where real
well log data and synthetic seismograms are used, results show that the effects of net-
to-gross ratio, saturation, thickness, and stacking pattern on seismic signatures can be
distinguished on the KPCA and the amplitude attribute space. However, of the two
types of attributes, the amplitude attribute is more sensitive to noise. In the second
case study, where real well log data and a real 2-D seismic section are used, seismic
signatures are also affected by the properties of the overburden and the underburden
which are non-stationary across the 2-D section. Even though we are not interested in
these two parameters, the sensitivity analysis shows that their impact on seismic
signature is greater than net-to-gross ratio. Therefore, the effect of these two
parameters needs to be reduced before net-to-gross estimation.
The workflow presented here can be extended by combining knowledge of
sedimentological trends For example, if only well B (proximal) is available, we can
incorporate trends such as sorting, quartz/clay ratio (Dutta, 2009) and use rock physics
relations to predict rock properties of lithologic states in the well A area (distal). Then
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 222
these properties can be used to generate training sets for estimating reservoir
properties of sub-resolution sequences in the area around well A.
Additional topics for future research include incorporating saturation effect,
finding the optimal values of parameter k (i.e., number of neighbors) in the kNN
method, comparing performance of other kernels for KPCA algorithms, and
minimizing computational time.
5.7 Acknowledgements
We would like to thank Hess Corporation for providing the data. This work was
supported by the Stanford Rock Physics and Borehole Geophysics project and the
Stanford Center for Reservoir Forecasting.
5.8 References
Dutta, T., Mukerji, T., and Mavko, G., 2007, Quantifying spatial trends of sediment
parameters in channelized turbidite, West Africa: SEG Expanded Abstracts, 26,
1674-1678.
Dutta, T., 2009, Integrating sequence Stratigraphy and rock-pshycis to interpret
seismic amplitudes and predict reservoir quality, Ph.D. Thesis, Stanford University.
Fenwick, D., Scheidt, C., and Caers, J., 2012, A distance-based generalized sensitivity
analysis for reservoir modeling: Computational Geosciences.
Jobe, Z.R., Lowe, D.R., and Uchytil, S.J., 2011, Two fundamentally different types of
submarine canyons along the continental margin of Equatorial Guinea: Marine and
Petroleum Geology, 28, 843-860.
Kennett, B., 1983, Seismic wave propagation in stratified media: Cambridge
University Press, Cambridge.
Lowe, D.R., 2004, Report on core logging, lithofacies, and basic sedimentology of
Equitoral Guinea: Hess internal report.
Thomas, E. C., and Stieber, S. J., 1975, The distribution of shale in sandstones and its
effect upon porosity: 16th Annual Logging Symposium, SPWLA, paper T.
CHAPTER 5: Seismic signatures of thin sand-shale reservoirs: Case studies 223
Tobin, K.W., Chaum, E., Gregor, J., Karnowski, T.P., Price, J.R., and Wall, J., 2009,
Image informatics for clinical and preclinical biomedical analysis: Computational
Intelligence in Medical Imaging, 1st ed. CRC Press.