The research project targets the development of Deep Learning methods in the field of computational fluid dynamics (CFD). Fluid flow modeling underlies the simulation of weather, climate, aerodynamics, and many other fields. Direct numerical simulation (DNS) is intractable outside of limited cases. Large volumes of data from measurements and simulations allow researchers to revisit existing CFD bottlenecks by leveraging learning-based strategies. This topic presents typical interdisciplinary challenges, with the end goal of improving the cost accuracy ratio of physical system description and forecasting.

The project targets the learning of surrogate and reduced order models for building approximate and fast simulation when DNS are too costly. This involves solving the following machine learning (ML) challenges: (i) developing models closely integrating CFD solvers and ML components, (ii) solving generalization problems allowing us to extrapolate to whole ranges of situations (ii) developing mesh-free spatio-temporal predictive approaches allowing us to work with different resolutions and meshes. In particular, the project presents an application to hypersonic reentry, targeting the European initiative on Clean Space and sustainable missions, and possibly hydrogen combustion in a scramjet engine.

PhD student: Ismaël ZIGHED

PhD supervisors: Dr Taraneh SAYADI (Director)

Research laboratory: UMR 7190 Institut Jean Le Rond D'Alembert