Objectives of the chair
Recently, scientific machine learning (SciML) has expanded the capabilities of traditional numerical approaches, by simplifying computational modeling and providing cost-effective surrogates. However, SciML surrogates suffer from absence of the explicit error control, computationally intensive training phase, and the lack of reliability in practice.
HAILSED chair aims to tackle these challenges by
(1) Developing novel types of SciML surrogates, the architecture of which incorporates physical and geometric constraints explicitly, and whose validation error can be controlled a posteriori.
(2) Developing novel training algorithms, which leverage multilevel and domain-decomposition-based approaches and utilize data and model parallelism.
(3) Hybridizing SciML surrogates with state-of-the-art numerical solution methods, which will be achieved by developing AI-equipped nonlinear field-split and domain-decomposition-based preconditioning strategies.
By successfully carrying out the proposed methodologies, HAISLED chair aims to unlock efficient and error-controlled solutions for large-scale multiphysics and multiscale problems through the synergy of SciML surrogate models and classical numerical approaches.
- Approximation bounds for SciML surrogates
- Incorporation of physical and geometric constraints into SciML architectures
- Multilevel and domain-decomposition based training algorithms that are robust to subsampling noise and can efficiently exploit the underlying structure of SciML surrogates
- Globally convergent hybridization and error control of hybrid (AI-augmented) iterative methods with machine learning approaches
- IRT SAINT EXUPÉRY Error bounds for SciML surrogates using the Lipschitz constant, collaboration led by P. Novello
- AIRBUS, VITESCO, et LIEBHERR : Surrogate modeling for industrial design
- EPFL (Switzerland, group of P. Fua), Fully differentiable AI-based pipeline for fast optimization, collaboration led by M. Bauerheim
- SANDIA NATIONAL LABORATORIES (USA, group of E. Cyr), Parallel training algorithms for large-scale machine learning
- UNIVERSITÀ DELLA SVIZZERA ITALIANA
(Saudi Arabia, group of R. Krause), Parallel training algorithms
- Error control for SciML surrogates
- Efficient and parallel training of deep neural networks
- Efficient solution of large-scale, high-fidelity, multiscale and multiphysics problems by means of hybrid AI-numerical simulations
English 
French