Description
Date depot: 15 avril 2022
Titre: Physically Constrained Deep Learning for CFD and PDE Modeling. Application to Steady Navier-Stokes Equations
Directeur de thèse:
Patrick GALLINARI (ISIR (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: The objective of this thesis is to build physically-based DL models in the spirit of replacing CFD solvers by an end-to-end and differentiable neural models to improve the accuracy of prediction while being able to understand the decisions of DL models from physics standpoint. The motivation of DL models with physical constraints is to propose a good approximate solution offering quick conception insights. The recent advances in DL, namely automatic differentiation process and the neural PDE/ODE architectures open perspectives to explicitly incorporate the Navier-Stokes equations in the optimization procedure of DL models and/or adding its residual terms to the loss function in order to physically guide the gradient updates.
Doctorant.e: Bonnet Florent