example on how to (intelligently) augment the nuclear-data ......nuclear-data pipeline with machine...

UNCLASSIFIED

Managed by Triad National Security, LLC for the U.S. Department of Energy’s NNSA LA-UR-YY-XXXXX

Example on how to (intelligently) augment the nuclear-data pipeline with machine learning

D. Neudecker Feb. 1, 2021Human pipeline for nuclear data, WANDA 2021

LA-UR-21-20366

UNCLASSIFIED Slide 2

The “human pipeline” vs machine learning and automation - a dichotomy??

• Machine learning helps us where human brain is overwhelmed with wealth of data. Conversely, we can integrate expert knowledge in these methods.

• Where can we use automation to free the humans in the nuclear-data pipeline from repetitive work?

• Machine learning is an exciting subject area that draws students/ Postdocs into nuclear data. (Would not say “how to train young students for this new paradigm”, but rather how can we learn along with them-with the help of a datascientist.) Students/ Postdocs publications in red


Where in the nuclear-data pipeline can automation and ML help us, a few examples:

Differential Experiment

Nuclear Theory

Evaluation ProcessingSensitivity Validation

TuningCorrections

General-purpose Library

AdjustedLibraries Application

Integral Exp. Designed for V&V/Appl.

Whewell et al., NIMA 978, 164305 (2020).

• Grechanuk et al., J. Comput. & Theor. Transport 47, 552 (2019).

• Neudecker et al., NDS 167, 36 (2020).

Lovell et al., J. Phys. G 47, 114001 (2020).

Schnabel et al., arxiv:/2009.00521

Caliva, De Sousa Ribeiro, et al., IEEE, 2018.

• Michaud, Kleedtke et al., ANS Transactions 121, 1035 (2019).

• Siefman et al., ANE 151 (2021).

Arthur et al., ANE 133, 853 (2019).


2 examples of LANL work to show how we transfer knowledge to ML and that ML augments the pipeline

Differential Experiment

Nuclear Theory

Evaluation ProcessingSensitivity Validation

TuningCorrections

General-purpose Library

AdjustedLibraries Application

Integral Exp. Designed for V&V/Appl.

Whewell et al., NIMA 978, 164305 (2020).

• Grechanuk et al., J. Comput. & Theor. Transport 47, 552 (2019).

• Neudecker et al., NDS 167, 36 (2020).

Lovell et al., J. Phys. G 47, 114001 (2020).

• Michaud, Kleedtke et al., ANS Transactions 121, 1035 (2019).

• Siefman et al., ANE 151 (2021).

Schnabel et al., arxiv:/2009.00521

Caliva, De Sousa Ribeiro, et al., IEEE, 2018.


Question: What features of differential exp. lead to systematic discrepancies/ outliers in a database?Why traditional techniques fail: the problem has 37 feature categories (100 values) for 24 measurements. ML can help us find trends in data, where experts are overwhelmed with data.

Benefit: We are investigating the physics reasons related to discrepancies between experiments or outliers. This can help us:• Add missing unc/ reject data based on physics reasons → more reliable

nuclear data and uncertainties,• Design experiments with features known to produce reliable exp. data.

Needle in the haystack

…

FluxIOCH

PPAC

UNCLASSIFIED

ML find physics expected features and unexpected features related to outliers, brings value to the field.


FluxIOCH

PPAC

Expert knowledge fed to ML:• Exp, data.• Uncertainties• Features




FluxIOCH

PPAC

Find outliers

Flux

PPACFlux

IOCH

Accepted Data Outliers





FluxIOCH

PPAC

Find outliers

Flux

PPACFlux

IOCH

Accepted Data Outliers

Find features common to outliers

Flux IOCH Various

ML answer: Features related to outliers.



Knowledge gained by ML can influence our evaluated data.

It is always up to the physicist to decide if the results are helpful. ML augments but does not replace expert judgment.

CAVEAT: for this particular problem, we lack the infrastructure to apply ML on a large scale.Natural language processing and SG-50 might help.


Question: What nuclear data leads to bias between simulated and experimental criticality?Why traditional techniques fail: we simulate 1 criticality value with 1000s of nuclear data. ML can help us find trends in data, where experts are overwhelmed with data.

Benefit: This information provides input on what nuclear data might need to be revisited and corrected: • Resolve timely issues in nuclear data → better data for applications,• Identify need for integral/ differential experiments or new evaluations.


…

Neutron Data Standards . . . NUCLEAR DATA SHEETS A.D. Carlson et al.

V. TABULAR DATA FOR THE NEUTRONSTANDARDS

Tabular data for each of the cross section standards andthe additional cross sections obtained in the cross sectionstandards evaluation process are given in the Tables XII–XX. For all the evaluations other than those for the lightelement standards, the tabular output is directly fromGMAP. For the 6Li(n,t), 10B(n,α) and 10B(n,α1γ) crosssections the GMAP output was fitted with EDA code asdescribed in Sec. A. The tables for those cross sectionswere provided as point-wise values from EDA. The H(n,n)and C(n,n) cross sections had been evaluated using EDAand the tables are direct output from EDA as point-wisevalues.

The evaluation of the 252Cf PFNS obtained fromthis work led to only very small changes in the spec-trum obtained by Mannhart. It is recommended thatthe Mannhart evaluation be used for any applications. Itis available at https://www-nds.iaea.org/standards/ref-spectra/ together with the evaluated 235U ther-mal prompt fission neutron spectrum. The reference fis-sion cross sections for 209Bi(n,f), natPb(n,f), 235U(n,f),238U(n,f) and 239Pu(n,f); and the prompt γ-ray pro-duction reference Cross Sections for 7Li(n,n’γ) and48Ti(n,n’γ) will be listed and updated on the site https://www-nds.iaea.org/standards/. As noted previously,the 3He(n,p) cross section was not re-evaluated. The pub-lication on the 2006 standards [1] contains the 3He(n,p)evaluation.

The GMAP evaluation estimates a point-wise cross sec-tion and its uncertainty at energy E using experimentaldata in the energy range from E1 to E2. However, forthe 235U(n,f) cross section an integral from 7.8–11 eVis produced with a node average energy 9.4 eV. The in-terval corresponding to the node at 0.15 keV starts at0.1 keV both for 235U(n,f) and 239Pu(n,f) cross sections.From there on, all intervals are located half-way betweengiven GMA nodes. The results from 1 keV up to 150 keVcorrespond to the average of low resolution experiments.For the 238U(n,f) cross section below 2 MeV (below theregion where it is a standard) results with a denser gridare marked by “x” and one corrected point is labelledby “xx”. Smoothing has been applied for regions wherescatter of data needs to be removed since the standardsshould be smooth. For all the tabular data, the values inthe standards energy region are recommended to be usedas standards for measurements. The fitted unsmoothedvalues were included into the evaluated ENDF-B/VIII.0general-purpose files in the standard region.

0.01 0.1 1 10 100

Cros

s Section (b)

1.4

1.6

1.8

2.0

2.2

2.4

2.6

2.8

3.0DS644, Li Jingwen, 1982 DS620, I. Szabo, 1970 DS621, I. Szabo, 1971 DS622, I. Szabo, 1973 DS611, 615, 616, 617TUD/KRI, 1983-1985 DS640, M. Davis, 1978 DS521, K. Kari, 1978 DS619, J. Perkin, 1965 DS623, I. Szabo, 1976 DS612, M. Cance, 1978 DS672, W. Allen, 1957 DS628, C. Uttley, 1956 DS657, A. Moat, 1958 DS589, D. Gayther, 1975, shape DS671, W. Allen, 1957, shape DS1038, Zhou Xian-Jian, 1982 2006 Standard2017 Standard

239Pu(n,f)

Incident Neutron Energy (MeV)

(a) 239Pu(n,f) experimental data from 4 keV up to 200 MeV.

0.01 0.1 1 10 100

Cros

s Section

Ratio

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0DS633, I. Garlea, 1983 DS637, M. Mahdavi, 1972 DS653, B. Fursov-1, 1977 DS654, B. Fursov-2, 1977 DS600, G. Carlson, 1978 DS602, J. Meadows, 1983 DS685, J. Meadows, 1986 DS605, E. Pfletschinger, 1970 DS626, W. Poenitz, 1970, shape DS549, C. Wagemans, 1980, shape DS608, P. White, 1965 DS609, P. White, 1967 DS631, K. Zhuravlev, 1977 DS666, M. Varnagy, 1982 DS1012, O. Scherbakov, 2001 DS1014, P. Staples, 1998 DS635, W. Lehto, 1970, shapeDS536, L. Weston, 1983, shape2006 StandardF. Tovesson, X=14402009, 2014 2017 Standard

Incident Neutron energy (MeV)

239Pu(n,f)/235U(n,f)

(b) 239Pu(n,f) to 235U(n,f) cross section ratio from 1 keV up to200 MeV.


1 10 100

Cro

ss S

ect

ion

Ratio

1.0

1.2

1.4

1.6

DS1012, O. Scherbakov, 2001 DS1014, P. Staples, 1998 2017 StandardF.Tovesson, X4=14271003, 2017 2006 Standard DS1029, P.W. Lisowski, 1991

239Pu(n,f)/235U(n,f)

(c) 239Pu(n,f) to 235U(n,f) cross section ratio from 200 keV upto 200 MeV.

FIG. 37. (Color online) Comparison of the 2017 and 2006standards evaluations, together with experimental data for the239Pu(n,f) cross section (a) and for the 239Pu(n,f) to 235U(n,f)cross section ratio (b,c).

172

0 0.05

0.1 0.15

0.2 0.25

0.3 0.35

0.4 0.45

0.01 0.1 1 10

PFN

S (1

/MeV

)

Outgoing Neutron Energy (MeV)

239Pu(n500keV,f) PFNS

1 EvaluationEvaluation

Staples, 500keVKnitter, 215keV

Starostov, thermalLajtai, thermal

Lestone, 500keVENDF/B-VII.1

stmentdisk

wireguides

Air I rod

Center

r supportlugs

Glory

fvtass adsafetyblock

collet

●

inlet

Fig. 5, l?re active portion of ori~”nal Jezebel, the bare plutonium assembly. Cooling air blows wtt of the locating armsthat nudeon taut wires.

thickness was assumed in apportioning the nickelbetween external and internal surfaces. Lack ofplaneness, however, was assumed to introduce an average0.001 -in. gap between each of the three principal pairsof internal surfaces.

Average densities were established by adjustingmeasured materiaI densities to allow for the nominalvolume of internal nickel coating and voids. Voidsremaining after correction for internal nickel wereredistributed uniformly (with compensating surface-massadjustment, Ref. 3) so that values of average densitywere retained.*

*A restate m e nt of the inverse-square relationship betweendensity and critical mass is that a given mass increment is threetimes as effective when distributed uniformly as it is when addedto the surface.

As shown in Figs. 6, 7, and 8, the three Jezebelsystems differed somewhat in shape, which led todifferent corrections for asphericity. Further, aluminumadapters re uired to fit the thin steel clamps (Fig. 5) tothe small !2 s u p arts a d d e d to the incidental reflection

for that assembly. Otherwise, corrections were similar.Captions of Figs. 6, 7, and 8 give the critical or

slightly subcritical Jezebel configurations from whichcritical masses are derived. Also shown arecorresponding masses corrected for the fiiling of majorvoids left by missing mass-adjustment plugs or glory-holeinserts, and by retracted control rod. These correctionsrely upon calibrations of the control rod and plugs.

The further corrections for asphericit y, nickelcoating, incidental reflection by clamps andsurroundings, homogenization, etc., are listed in Table I.The resulting critical masses apply to isolated barespheres of uniform plutonium or uranium.

5

.

Simulating the Jezebel critical assembly

239Pu233U

19F 2.5

3

3.5

4

4.5

5

5.5

6

6.5

7

10-1110-10 10-9 10-8 10-7 10-6 10-5 10-4 10-3 10-2 10-1 100 101

Aver

age

Prom

pt N

eutro

n M

ultip

licity


ENDF/B-VII.1ENDF/B-VIII.0

Experimental Data


ML points towards potential issue in 19F ENDF/B-VIII.0 nuclear data relevant for validation exp.

Incident Energy (MeV)

Cro

ss S

ectio

n (b

arns

)

10.5 2

1

0.5

2

5

1

0.5

2

5

10.5 2

ENDF/B-VIII.0

JENDL-4.0

Broder 1969

19F(n,inl)

Several 19F nuclear data observables, over a broad energy range, were highlighted as important to predict bias. → Correlation effects known from traditional validation studies hamper ML because it is inherent in the data!!!!

Random Forest Results


Conclusions:

Ø Can ML help our nuclear-data pipeline: Absolutely!!!

Ø ML’s strength: find trends in large amounts of data where human brains are overwhelmed. This information may be crucial to improve our nuclear data.

Ø HOWEVER: ML is no silver bullet. It is critical to feed it expert knowledge and use physics intuition to interpret results. We need to:Ø Develop infrastructure and tools to provide data in an easily readable

and unambiguously interpretable format (e.g., EXFOR format),Ø Develop experimental data and theory to solve physics questions,Ø Bring statisticians and nuclear-data experts together to correctly

interpret the results.

Bottom line: ML is a great tool. We need to use these algorithms along with developing physics data, tools and infrastructure.

UNCLASSIFIED

Research reported in this publication was supported by the U.S. Department of Energy LDRD program at Los Alamos National Laboratory.

We gratefully acknowledge the support of the Advanced Simulation and Computing (ASC) program at Los Alamos National Laboratory.

This work was supported in part by the DOE Nuclear Criticality Safety Program, funded and managed by the NNSA for the Department of Energy.

Acknowledgements

example on how to (intelligently) augment the nuclear-data ......nuclear-data pipeline with machine...

Documents