.pdf
WP6: Uncertainty Quantification
WP6 quantifies uncertainties in complex multiscale simulations through sensitivity analysis, surrogate modeling, and high-dimensional integration, leveraging exascale computing for tractable UQ workflows.
Recent Highlights (2024-2025)
Kernel-Based Sensitivity
Global sensitivity analysis methods with kernel-based dependence measures
UQ Platforms
Uranie and OpenTURNS platforms for uncertainty quantification
Integrated Analysis
Sensitivity analyses linked with WP3 solvers and WP2 surrogates
EuroHPC UQ
Initial UQ pipelines exercised on EuroHPC resources
1. Objectives
Uncertainty Quantification includes several critical steps:
-
Uncertainty propagation through complex multiscale models
-
Sensitivity analysis to understand important often correlated factors
-
Reduce uncertainty through modeling improvements in WP1
-
Establish robust inversion or optimization under uncertainties
3. Key Tasks
T6.1: Kernel-Based Sensitivity Analysis
-
Handle very high-dimensional and multivariate data in exascale applications
-
Develop tractable extensions of sensitivity analysis built upon kernel-based dependence measures
-
Investigate optimized computing schemes of high-dimensional integrals
-
Support uncertainty propagation step
T6.2: UQ in a PDE Solving Framework
-
Propagate uncertainties (initial conditions, coefficients) in complex PDE solutions
-
Use machine learning and stochastic spectral methods for suitable approximations
-
Address calibration challenges that exceed underlying problem size by orders of magnitude
-
Make HPC strategies mandatory for tractable solutions
T6.3: Surrogate Modeling for UQ
-
Address complex multi-physics and/or multi-scale problems with coupled, nested and chained numerical codes
-
Build and calibrate global metamodel assembling all prediction uncertainties
-
Requires HPC for this formidable task
T6.4: Acceleration via Exascale Calculations
-
Integrate methodological developments of Tasks 6.1 to 6.3
-
Use opensource platforms Uranie and OpenTURNS dedicated to uncertainty quantification
-
Take advantage of exascale computational properties
-
Conduct benchmarking on exascale applications
4. Leads & Partners
Lead Institution |
École Polytechnique |
Co-Leaders |
CEA, UNISTRA |
Duration |
Months 1-60 |
5. Addressed Exascale Bottlenecks
WP6 targets bottlenecks B8 (Discovery, design, and decision algorithms) and B13 (Opportunity to integrate uncertainties):
-
Enabling ensembles of many small runs for UQ
-
Integrating uncertainties directly into core calculations
-
Handling high-dimensional integrals at scale
-
Efficient parameter sweeps for sensitivity analysis