modeling of physical systems underspeci ed by data
TRANSCRIPT
![Page 1: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/1.jpg)
Modeling of physical systemsunderspecified by data
Daniel M. Tartakovsky
Department of Mechanical & Aerospace Engineering
University of California, San Diego
![Page 2: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/2.jpg)
Outline
1. Physical systems & Stochastic PDEs
2. Data-Driven Domain Decompositions (D4) for SPDEs
3. Implementation of D4
(a) Data analysis & image segmentation
(b) Closure approximations for SPDEs
(c) PDEs on random domains
4. Effective parameters for heterogeneous composites
5. Conclusions
![Page 3: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/3.jpg)
1.1 Background
• Wisdom begins with the acknowledgment of uncertainty—of the
limits of what we know. David Hume (1748), An Inquiry Concerning
Human Understanding
• Most physical systems are fundamentally stochastic (Wiener, 1938;
Frish, 1968; Papanicolaou, 1973; Van Kampen, 1976):
• Model & Parameter uncertainty
– Heterogeneity– Lack of sufficient data– Measurement noise∗ Experimental errors∗ Interpretive errors
00.2
0.40.6
0.81
0
0.2
0.4
0.6
0.8
1
• Randomness as a measure of ignorance
![Page 4: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/4.jpg)
1.2 Probabilistic Framework
• Parameter,
p(x) : D ∈ Rd→ R
• Random field,
p(x, ω) : D × Ω→ R
• Ergodicity,
〈p〉 ≈ 1||D||
∫D pdx
00.2
0.40.6
0.81
0
0.2
0.4
0.6
0.8
1
parameter p(x1, x2)
• Governing equations become stochasticL(x;u) = f(x), x ∈ DB(x;u) = g(x), x ∈ ∂D
=⇒
L(x, ω;u) = f(x, ω), x ∈ DB(x, ω;u) = g(x, ω), x ∈ ∂D
![Page 5: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/5.jpg)
1.3 Stochastic PDEs & UQ
• Consider a physical processL(x;u) = f(x), x ∈ DB(x;u) = g(x), x ∈ ∂D
– System parameters pi = p(xi),i = 1, . . . , N
– System states uj = u(xj),j = 1, . . . ,M
• Modeling steps
1. Use data to construct a probabilistic description of p(x, ω)2. Solve SPDEs to obtain a probabilistic description of u(x, ω)
3. Assimilate uj = u(xj) to refine prior distributions
![Page 6: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/6.jpg)
1.4a Statistical Methods for UQ
• (Brute-force) Monte Carlo methods
– Convergence rate (CR): 1/√N
– CR is independent of the number of random variables
• Monte Carlo based methods
– Quasi MC (QMC)– Markov chain MC (MCMC)– Importance sampling (Fishman, 1996)
• Variance reduction techniques
– Problematic when the number of RVs is large
• Response Surface Methods (RSM)
– Interpolation in the state space reduces the number of realizations– Problematic when the number of RVs is large
![Page 7: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/7.jpg)
1.4b Stochastic Methods for UQ
• “Indirect” methods
– Fokker-Planck equations
– Moments equations
• “Direct” methods
– Interval analysis: Maximum output bounds
– Operator-based methods
∗ Weighted integral method (Takada, 1990; Geodatis, 1991)
∗ Neumann expansion (Shinozuka, 1988)
• Polynomial chaos expansions
– Grounded in rigorous mathematical theory of Wiener (1938)
– Arbitrary inputs: Generalized polynomial chaos
![Page 8: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/8.jpg)
1.5 Modeling Dichotomy
• Real systems are characterized by
– Non-stationary (statistically inhomogeneous)– Multi-modal– Large variances– Complex correlation structures
• Standard SPDE techniques require
– Stationarity (statistically homogeneity)– Small variances– Simple correlation structures (Gaussian, exponential)– Uni-modality
• Our goal is
– to incorporate realistic statistical parameterizations– to enhance predictive power– to improve computational efficiency
![Page 9: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/9.jpg)
1.5 Reasons for Modeling Dichotomy
• Perturbation expansions
– Small variance σ2p
• Spectral methods / polynomial chaos expansions
– large correlation length lp– uni-modality
• Mapping closures
– “Nice” parameter distributions, e.g., Gaussian
• Stochastic upscaling (homogenezation)
– Regularity requirements, e.g., spatial periodicity– Small σ2
p
• Monte Carlo simulations
– N increases with variance σ2p
– Resolution depends on σ2p and lp
![Page 10: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/10.jpg)
Outline
1. Physical systems & Stochastic PDEs
2. Data-Driven Domain Decompositions (D4) for SPDEs
3. Implementation of D4
4. Effective parameters for heterogeneous composites
5. Conclusions
![Page 11: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/11.jpg)
2. Data-Driven Domain Decomposition
γR. D. D.
ΩΩ
f(Π) f(Π,Γ)
Multi-modal distributions Uni-modal distributions
High variances Low variancesComplex correlation structures Simple correlation structures
σ2p = Q1σ
2p1
+Q2σ2p2
+Q1Q2 [〈p1〉 − 〈p2〉]2
![Page 12: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/12.jpg)
2.1 Strategy for Domain Decompositions
• Step 1: Decomposition of the parameter space (image processing
techniques; probability maps)
• Step 2: Conditional statistics (noise propagation; closures)∫LΠu f(Π|γ) dΠ → 〈u|γ〉
• Step 3: Averaging over random geometries
〈u〉 =
∫〈u|Γ〉 f(Γ) dΓ
![Page 13: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/13.jpg)
2.2 Implementation of Domain Decompositions
1. Probabilistic decomposition of the parameter space
• Geostatistical reconstruction of internal geometries• (nonstationary) Bayesian spatial statistics• Statistical Learning Theory (Support Vector Machines)• Risk-based parameterization
2. Conditional statistics from SPDEs
• Perturbation expansions• Polynomial chaos expansions• Collocation methods• Nonlinear Gaussian mappings
3. PDEs on random domains
• Monte Carlo simulations• Stochastic mappings• Perturbation expansions
![Page 14: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/14.jpg)
Outline
1. Physical systems & Stochastic PDEs
2. Data-Driven Domain Decompositions (D4) for SPDEs
3. Implementation of D4
3.1 Decomposition of the parameter space• Spatial statistics (geostatistics)• MCMC with Metropolis sampling• Support Vector Machines
3.2 Conditional Statistics
3.3 Averaging over random geometries
4. Effective parameters for heterogeneous composites
5. Conclusions
![Page 15: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/15.jpg)
Probabilistic Reconstruction of Facies
Mean boundary (a) and the boundary with 95% probability (b)
![Page 16: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/16.jpg)
Outline
1. Physical systems & Stochastic PDEs
2. Data-Driven Domain Decompositions (D4) for SPDEs
3. Implementation of D4
3.1 Decomposition of the parameter space
3.2 Conditional Statistics
• Perturbation solutions
• Polynomial chaos expansions
• Stochastic collocation on sparse grids
3.3 Averaging over random geometries
4. Effective parameters for heterogeneous composites
5. Conclusions
![Page 17: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/17.jpg)
Conditional Statistics
L(x, ω;u) = f, x ∈ DB(x, ω;u) = g, x ∈ ∂D
, p = p(x, ω), ω ∈ Ω
γR. D. D.
ΩΩ
L(x, ω;u) = f, x ∈ DB(x, ω;u) = g, x ∈ ∂Dcontinuity conditions, x ∈ γ
, p =
p1(x, ω), x ∈ D1, ω ∈ Ω1
p2(x, ω), x ∈ D2, ω ∈ Ω2
![Page 18: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/18.jpg)
3.2a Perturbation Solutions
• ∇ · k(x)∇u = Qδ(x− x0)
Yi = ln ki – Gaussian, Y 1 = 3Y 2
σ2Y1
= σ2Y2
= 1, σ2Y ≈ 30
exponential correlation functions
b is log-normal
• Perturbation in σ2Yi
• Monte Carlo simulations for b
– 〈u(x)〉: Eu ≈ 2%
– σ2u(x): Eσ ≈ 5%
1H H 2
b
K
K
2
1
tran
sver
se c
oord
inat
e, x
2
longitudinal coordinate, x1
0 2 4 6 8 10 120
2
4
6
8
10
12
![Page 19: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/19.jpg)
3.2a Perturbation Solutions (cntd)
![Page 20: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/20.jpg)
3.2a Perturbation Solutions (cntd)
Assign weights Wk to each boundary Γk, so that
〈u〉 =
∫〈u|Γ〉 f(Γ) dΓ ≈
∑k
Wk〈u|Γk〉
Mean (a) and variance (b) of u
![Page 21: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/21.jpg)
3.2b Polynomial Chaos Expansions
• Stochastic PDEL(x, ω;u) = f(x, ω), x ∈ DB(x, ω;u) = g(x, ω), x ∈ ∂D
• Generalized polynomialchaos expansions
p(x, t, ω) =
∞∑i=1
ai(x, t)Ψi(ω)
– An approximation
p(x, t, ω) ≈N∑i=1
ai(x, t)Ψi(ω)
Correspondence between the type of the Wiener-Askey
polynomial chaos and their underlying random variables.
Distribution Polynomials
Gaussian HermiteGamma Laguerre
Beta JacobiUniform LegendrePoisson Charlier
Binomial KrawtchoukNegative binomial MeixnerHypergeometric Hahn
G. Em. Karniadakis. etc.
![Page 22: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/22.jpg)
3.2b Polynomial Chaos Expansions (cntd)
• Parameter representation
p(x, t, ω) =
N∑i=1
ai(x, t)Ψi(ω)
– Choose an orthogonal polynomial basis– Reduce N
• Advantages:
– Nonperturbative
– Large fluctuations
• Drawbacks:
– Finite correlations
– Unimodal distributions
−15 −10 −5 0 5 10 15 20 25 30 350
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
F X(x
)
exact CDF5th−order Hermite expansion
a highly non-Gaussian unimodal field
0 0.5 1 1.5 2 2.5 30
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
xF X
(x)
exact CDF3th−order Hermite expansion5th−order Hermite expansion10th−order Hermite expansion
a bimodal field
![Page 23: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/23.jpg)
3.2b Polynomial Chaos Expansions (cntd)
Problem: ∇ · k∇u = 0
Parameter measurements ki = k(xi):
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10.5
1
1.5
2
2.5
x
Mea
sure
men
t of K
(x)
0.5 1 1.5 2 2.5 30
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
K
p(K
)
Solution statistics: ensemble mean (a) and standard deviation (b)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.2
0.4
0.6
x
mea
n he
ad
RDDPCMCS
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
x
head
STD
RDDPCMCS
![Page 24: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/24.jpg)
3.2c Stochastic Collocation Methods
• Choice of sampling points
– Tensor products of 1-D collocation point sets– Sparse grids (nested or non-nested)
• Choice of weight functions
ϑ(ξ) = δ(ξ − ξk)
• Advantage:
– Nonintrusive
• Disadvantage
– Can be less accurate than PCE
x3
h
x1
x2
meanσh
x1
x2
x3
variance
![Page 25: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/25.jpg)
Outline
1. Physical systems & Stochastic PDEs
2. Data-Driven Domain Decompositions (D4) for SPDEs
3. Implementation of D4
3.1 Decomposition of the parameter space
3.2 Conditional Statistics
3.3 Averaging over random geometries
4. Effective parameters for heterogeneous composites
5. Conclusions
![Page 26: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/26.jpg)
Computational Approach
• A deterministic equation in a random domainL(x;u) = f(x), x ∈ D(ω)
B(x;u) = g(x), x ∈ ∂D(ω)
• Step 1: Stochastic mapping
ξ = ξ(x;ω), x = x(ξ;ω) x ∈ D(ω), ξ ∈ E
• Step 2: Solving a stochastic equation in a deterministic domainL(ξ, ω;u) = f(ξ, ω), ξ ∈ EB(ξ, ω;u) = g(ξ, ω), ξ ∈ ∂E
![Page 27: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/27.jpg)
Random Mapping
• Surface parameterization
– Karhunen-Loeve
representations
– Fourier-type expansions
– Etc.
• Numerical mappings, e.g.,
∇2ξxi = 0 xi|∂E = xi|∂D(ω)
• Analytical mappings
x1
x2
ξ1
ξ2
D(ω)E
Q1 Q2
Q3
Q4
P1 P2
P3P4
ξ=ξ(x)
x=x(ξ)
Random mapping, D(ω)→ E
![Page 28: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/28.jpg)
Computational examples
• Steady-state diffusion
∇2u(x;ω) = 0, x ∈ D(ω)
– Random bottom of a channel
– Random exclusion
– Numerical mapping
• Transport by Stocks flow in a pipe
∂u
∂t+ v · ∇u = a∇2u, x ∈ D(ω)
– Lubrication approximation
– Analytical mapping
u = 1
u = 0 u = 0
u = 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
x1
x 2
u=1
u=0
u=0
u=0
u=0
00.5
11.5
22.5
33.5
44.5
5
−0.20
0.2
−0.20
0.2
![Page 29: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/29.jpg)
Steady Diffusion
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
x1
x 2
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
ξ1
ξ 2
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
0.4
0.5
x1
x 2
• Random surface: exponential correlation function, λ = 1
• 10-term Karhunen-Loeve representation
• 10-dimensional (N = 10) random space
![Page 30: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/30.jpg)
Steady Diffusion: Results
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
ξ1
ξ 2
0.166670.33333
0.50.66667
0.83333
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
ξ1
ξ 2
0.00300430.0060087
0.0090130.012017
0.015022
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
0.02
x
STD
SG: 1st−orderSG: 2nd−orderSG: 3rd−order
Mean & standard deviation PCE vs. MCS
![Page 31: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/31.jpg)
4. Effective Parameters
• Step 1: Conditional effective conductivity
– Keffi(x|γ) = 〈Ki〉I + ki(x) ki = [I−Bi]−1Ai
– Closure by perturbation
• Step 2: Averaging over random geometries γ
0.0 0.2 0.4 0.6 0.8 1.0coordinate, x
0
100
200
300
400
500
600
700
appa
rent
con
duct
ivity
, Kap
p
σβ = 0.1
σβ = 0.25
σβ = 1.0
homogeneous model
![Page 32: Modeling of physical systems underspeci ed by data](https://reader034.vdocuments.us/reader034/viewer/2022051319/627b69ed8fdb4060fd70d893/html5/thumbnails/32.jpg)
Conclusions
• Stochastic PDEs provide a natural framework for uncertainty
quantification
• In real applications, SPDEs require a complex statistical
parameterization
• D4 allows for incorporation of expert knowledge and diverse data
• New concept of effective properties
• Ability to make predictions and quantify uncertainty in realistic
setting