Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
Uncertainty Quantification Area
FASTMath UQ Team
1Sandia National Laboratories, Livermore, CA
2Sandia National Laboratories, Albuquerque, NM
3University of Southern California, Los Angeles, CA
4Massachusetts Institute of Technology, Cambridge, MA
FASTMath All-Hands MeetingArgonne National Lab.
June 10-11, 2019
Overview
UQ work under FASTMath includes
Enhancing deployed UQ capabilities in Dakota & UQTkCoupling with MIT UQ library (MUQ)
Low rank tensor methods for dealing with high-D functionsManifold learningAdaptive basisBayesian inferenceUQ with model errorInference in dynamical systemsMultilevel Multifidelity methodsOptimization under uncertainty
ASCR base funding on networks, stochastic dynamical systems,probabilistic computing, etc has contributed to the development andhardening of our UQ software over a number of years
SNL Najm Comp 2 / 14
FASTMath UQ Team
Sandia – CA
Habib NajmBert DebusschereCosmin SaftaKhachik SargsyanTiernan Casey
MIT
Youssef Marzouk
Sandia – NM
Michael EldredJohn JakemanGianluca Geraci
USC
Roger Ghanem
SNL Najm Comp 3 / 14
Partnerships – 1
Plasma Surface Interactions: Predicting the Performance and Impact ofDynamic PFC Surfaces (PSI2) [SC-FES]
Modeling transport of He gas bubbles in ITER divertor materialBayesian estimation, global sensitivity analysis (GSA), forward UQ
Simulation of Fission Gas in Uranium Oxide Nuclear Fuel [NE]
Modeling noble gas bubble transport in nuclear fuel rod materialBayesian estimation, GSA, forward UQ
Optimization of Sensor Networks for Improving Climate ModelPredictions (OSCM) [SC-BER]
Energy Exascale Earth System Model (E3SM) Land ComponentForward UQ and GSA: high-dimensionality challengesBayesian estimation of model parameters with model structural errorOptimal experimental design for measurement sensor placement
SNL Najm Comp 4 / 14
Partnerships – 2
Probabilistic Sea Level Projections from Ice Sheet and Earth SystemModel (ProSPect) [SC-BER]
Goal: Estimate uncertainty in predictions of sea-level rise due toice-sheet evolutionHigh-dimensionality (Hi-D) challenges of forward UQHi-D challenges of Bayesian inference of model parametersMoving toward using multi-fidelity algorithms for above tasks
SNL Najm Comp 5 / 14
High Dimensional Function Approximation-Dependence
0.0 0.2 0.4 0.6 0.8 1.00.0
0.2
0.4
0.6
0.8
1.0
−2 0 2
−3
−2
−1
0
1
2
3
−1.0 −0.5 0.0 0.5 1.0−1.00
−0.75
−0.50
−0.25
0.00
0.25
0.50
0.75
1.00
Efficiently building surrogates of models parameterized by dependentrandom variables is challenging:
Probabilistic transformations introduce non-linearitiesApproaches not accounting for PDF of variables are inefficient
Solution:
Generate orthonormal polynomial basis and weighted Leja sequenceTarget accuracy in high-probability regions while maintaining stability
John D. Jakeman, Fabian Franzelin, Akil Narayan, Michael Eldred, Dirk Pflüger, “Polynomial chaos expansions for dependent random variables”,Computer Methods in Applied Mechanics and Engineering, Volume 351,2019, Pages 643-666, https://doi.org/10.1016/j.cma.2019.03.049.
SNL Najm Comp 6 / 14
High Dimensional Function Approximation - Low Rank
E3SM Simulation
50°N
40°N
30°N
50°N
40°N
30°N
110°W 80°W
2
3
4
5
6
7
Tensor-train Schematic
L22/ 7Q` +QMiBMmQmb hh 2ti2MbBQMb
Ç .Bb+`2i2f�``�v #�b2/ `2T`2b2Mi�iBQMb Q7 7mM+iBQMb Bb HBKBiBM;Ç *QKTmi�iBQM rBi? 7mM+iBQMb Q7 /Bz2`BM; /Bb+`2iBx�iBQMbÇ :`B/ �/�Ti�iBQM Bb MQM@i`BpB�H- i2MbQ` T`Q/m+i ;`B/b �`2 Q7i2M r�bi27mHÇ GQ+�H BM7Q`K�iBQM �#Qmi bKQQi?M2bb- /Bb+QMiBMmBiB2b- 2i+X- MQi mb2/
Ç � +QMiBMmQmb 2ti2MbBQM, i2MbQ`@i`�BM → 7mM+iBQM@i`�BMÇ AK�;BM2 /Bb+`2iBxBM; � 7mM+iBQM rBi? BM}MBi2 TQBMib BM 2�+? /BK2MbBQM
f(x1, x2, :, x4) =
Re
Total Order Sensitivities
AprMay
JunJul
AugSep
Oct
rg_fracndays_offbr_mrfpgndays_onfroot_leafslatop_decidstem_leafleafcnnue_treecrit_daylgdd_crit
Total Effect Sobol Index
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Seek low-rank functional tensor-train representation to reveal couplings inhigh-dimensional models.
Compressed representation, can result in a sample complexity which isonly weakly dependent on dimensionCan combine local/global approximations in a compact model
Next Steps:
Embed parametric dependencies into a generalized low-rankfunctional tensor-train decompositionAdaptive sampling of mixed stochastic/physical spaces to targetregions of high-probability/non-linear behavior in the joint space
SNL Najm Comp 7 / 14
UQTk Developments
mod
el_A
mod
el_B
mod
el_C
mod
el_D
mod
el_E
0
5
10
15
20
25
log
evid
ence
Importance Evidence Values for polynomial
20.0 20.5 21.0 21.5 22.0 22.5 23.0
0.0
0.5
1.0
1.5
2.0
2.5
3.0
QoI−1
Den
sity
Convergence study of adapted PC for 600−dimensionalcrash analysis problem with hardening plasticity
RF order 3 dim=2RF order 3 dim=2RF order 3 dim=3RF order 3 dim=4RF order 3 dim=5RF order 3 dim=6
In-progress integration of new capabilitiesCoupling between UQTk and MUQ – C++/PythonBasis adaptation with error correctionManifold learning and sampling
Improved User experienceBug fixes, more documentation, better user interfacesNew tutorials – e.g. Bayesian model selection
Version 3.1.0 this summer — push to github for easier distribution
SNL Najm Comp 8 / 14
Dakota highlights
Dakota version 6.10 released May 15, 2019Hardening of multilevel-mutifidelity surrogate approachesAdaptivity through greedy multilevel refinement
Prototyped integration with Legion AMT for ensemble UQ workflowswith Stanford
Exploited special structure w/functional tensor train approximationIntegrated Compressed Continuous Computation (C3) package– from Alex Gorodetsky, U. MichiganMultidimensional functions represented in low-rank format– sample requirements scale linearly w/dimension 𝑂(𝑑𝑟2)
In progressIntegration of MUQ library (w/ MIT and CRREL)Integration of Adaptive Basis for dimension reduction (w/ USC)
SNL Najm Comp 9 / 14
MLMF/OUU Algorithmic highlights
Advancements on multifidelity/multilevel approaches for UQ
Surrogate-based approachesMultilevel expansions
Polynomial chaos, projection, regression, stoch. collocationFunctional tensor train (in progress)
Adaptive refinementIsotropic, anisotropic, index set, expanding front
Sampling-based approaches (more in the poster)A Generalized Framework for Approximate Control Variates
(arXiv: 1811.04988v3)
Optimization Under Uncertainty (OUU)
Verification cases to test SNOWPAC + MC-based UQ with error est.Fall 2019: multifidelity global opt. with GPs (MF EGO)2020: multifidelity OUU for Tokamaks (TDS)
SNL Najm Comp 10 / 14
Advances in Bayesian Methods
Systematic framework for embedded model error quantificationAllows predictive uncertainty attribution into components due to modelsurrogate construction, data noise and model errorWorkflow is implemented in UQTk and is employed in DOE and DODprojects, including turbulence modeling, fusion science and land model
(K. Sargsyan, X. Huan, HNN; Int. J UQ 2019)
MCMC acceleration via local surrogatesNew rate-optimal refinement strategies balance surrogate and sampling error
(Davis, Smith, Pillai, Marzouk 2019)
Explicit control of Lyapunov function guarantees convergence forheavier-tailed distributions, broadens robustness and applicability
Distance metrics for likelihood construction for dynamical systems
Comparison of chaotic dynamicaltrajectories using distances betweendensities defined in manifold coordinates
Global diffusion-map distance forconstructing Bayesian likelihoods inparameter estimation
SNL Najm Comp 11 / 14
Non-SciDAC Engagement – 1
SECURE: An Evidence-Based Approach for CybersecurityC. Safta, J. Jakeman, G. Geraci, B. Debusschere – [SNL Grand Chall. LDRD]
Adaptive sampling for quadrature on discrete random variables– developed and applied to SECURE relevant models
Analysis of redox potentials in H2 advanced water splitting materialsB. Debusschere – [DOE EERE]
Application of UQTk for model comparison and model errorcomputation to water splitting reactions for Hydrogen production.Work led to tutorials that have in turn been incorporated in UQTk
Bayesian Optimal Experimental DesignH. Najm – [DOE BES]
Application of UQTk surrogate construction and Bayesian capabilitiesOptimal design of a mass spectrometry expt at the ALS
SNL Najm Comp 12 / 14
Non-SciDAC Engagement – 2
Partnering w/U. Michigan - flood modeling – K. Sargsyan – [NSF]
Surrogate construction, sensitivity analysis and parameter estimationfor the ecohydrologic model using data observed in Manaus, Brazil
M. C. Dwelle, J. Kim,K. Sargsyan, V. Ivanov, “Streamflow, stomata, and soil pits: sources of inference for complex models with fast, robustuncertainty quantification”, Advances in Water Resources, Volume 125, pp. 13-31, March, 2019
Energy Exascale Earth System Model (E3SM) – K. Sargsyan – [SC-BER]
UQTk is employed for sensitivity analysis and model calibrationD. Ricciuto, K. Sargsyan, P. Thornton, The Impact of Parametric Uncertainties on Biogeochemistry in the E3SM Land Model, Journal of Advances inModeling Earth Systems, Vol. 10, No. 2, p.297–319, 2018
Atmosphere to electrons (A2e) – M. Eldred, G. Geraci – [EERE]
Dakota and multilevel / multifidelity algorithms deployed forforward / inverse UQ and OUU in wind plant applications
SNL Najm Comp 13 / 14
Non-SciDAC Engagement – 3 – R. Ghanem
Risk assessment for spent fuel containment structures – [NNSA]
UQ analysis of a high-fidelity model of a fuel cask systemUQ models to characterize damage following transport
Hazard maps for tsunami storm surge – [NSF]
Assessment of topography effect on run-up during a tsunami event
Modeling and design of a lightweight composite car – [DOE]
Adaptive basis in very high-dimensional coupled physics problems.
Probabilistic machine learning (PML) for physics discovery – [DARPA]
Integration of manifold sampling with genetic programming to carrysymbolic regression for candidate mathematical models.Adaptation of PML to large scale OUU; scramjet design optimization
SNL Najm Comp 14 / 14