WP2: Model Order Reduction, Surrogate, Scientific Machine Learning methods
WP2 develops non-intrusive approaches for ultra-fast surrogate models of complex physical problems, combining data-driven and model-driven techniques with a focus on physics-based neural network models.
Recent Highlights (2024-2025)
PINNs Innovation
Domain-decomposed PINNs with perfectly matched layers
Newton Solver Acceleration
Neural operators (FNO) accelerating Newton solvers: 80-200% CPU savings
Navier-Stokes ROMs
Reduced-order models for parametric Navier-Stokes
Exascale Optimization
Multi-fidelity surrogate-based optimization on 8000 GPUs (LUMI)
1. Objectives
WP2 uses non-intrusive approaches for designing ultra-fast surrogate models and strategies for leveraging these surrogates:
Data-Driven Techniques
-
Comparisons between reduced basis methods and neural network-based methods
-
Deep learning operators (FNO, DeepONet) for PDE solutions
-
Data-driven model order reduction without intrusive access to simulators
Model-Driven Techniques
-
Physics-informed neural networks (PINNs) respecting conservation laws
-
Neural operators learning PDE operators from data
-
Hybrid approaches combining physical constraints with data
3. Key Tasks
T2.1: Physics-Driven Deep Learning
-
PINNs for PDEs with physical constraints
-
Domain decomposition strategies for large-scale problems
-
Perfectly matched layers for boundary conditions
T2.2: Neural Operators
-
Fourier Neural Operators (FNO) for parametric PDEs
-
DeepONet and operator learning frameworks
-
Accelerating Newton iterations with learned operators
T2.3: Data-Driven Model Order Reduction
-
Proper Orthogonal Decomposition (POD)
-
Dynamic Mode Decomposition (DMD)
-
Autoencoders for nonlinear dimensionality reduction
T2.4: Non-Intrusive Reduced Basis
-
Regression-based reduced basis methods
-
Greedy sampling for snapshot selection
-
A posteriori error estimation
T2.5: Multi-Fidelity Modeling
-
Coupling low-fidelity and high-fidelity models
-
Adaptive fidelity selection
-
Cost-accuracy trade-offs for large ensembles
T2.6: Super-Resolution Methods
-
Real-time models with spatial/temporal upsampling
-
Neural network-based mesh refinement
-
Generative models for fine-scale details
4. Leads & Partners
Lead Institution |
Inria |
Co-Leaders |
UNISTRA, Sorbonne Université, CEA |
Duration |
Months 1-60 |