Mardi 2 juillet 2024 à 11h00 / Amphithéâtre François Canac, LMA
Anders THORIN, Research Engineer at CEA List and Assistant Prof. at École Polytechnique, will expose the research carried out at the Interactive Simulation Laboratory (CEA List) on the use of Neural Networks (NNs) to learn and accelerate finite elements computations for interactive simulations. Computation cost is indeed essential in such simulations: for the user to be able to interact in “real-time” with the computations, the governing equations must be solved fast enough to take into account his/her actions at every time step (typically: a few milliseconds). Taking advantage of the fact that NNs are (usually) cheaper to evaluate than finite elements (or similar), costly computations can be replaced by NNs. NNs and surrogate models will be seen here as a type of Reduced-Order Models, capable of approximating/interpolating a database composed of finite elements computations, considered as the reference solutions.
The presentation will be dedicated to:
Quickly covering the different possibilities to inject physical knowledge into NNs;
Presenting a NN architecture capable of learning finite strain plasticity;
Presenting a NN architecture capable of learning hyperelasticity;
NNs for solid dynamics in the context of interactivity (loading is not known in advance);
Preliminary results regarding active learning strategies to minimize the number of finite element simulation.