On a modular coupling strategy for high-performance computing of multiphysics problems

Modeling and simulating multiphysics problems (involving mechanical, fluid, thermal, chemical and magnetostatic aspects, etc.) is a major challenge when it comes to building digital twins that are as close as possible to the real systems they represent. These high-fidelity twins are used to predict behavior, assess reliability, perform predictive maintenance, etc. Some approaches to solving these problems are now mature, but many challenges remain: constructing numerical twins in the case of strongly coupled problems, using models optimized for each of the physics (discretizations in time and space), enriching the models with data (the latter may complement or even substitute for one of the physics) and finally, numerically solving the models thus constructed within a timeframe compatible with analysis constraints. Artificial intelligence tools are also likely to be an ingredient of choice in this landscape, and deserve to be investigated. This is the challenge that this thesis focuses on in the context of coupling solid mechanics with other physics, with the use of neural networks potentially enriched by physics, on the strength of the methods developed at the LMPS which are reaching maturity and which make it possible to confront such a challenge. In addition, the thesis is in line with the challenges of the ecological transition for sustainable development (ETSD) by proposing to develop multiphysics simulation tools with the aim of reducing the environmental impact of the design and maintenance of systems implementing several strongly coupled physics.

Doctoral Researcher: Elise Foulatier

Scientific Advisors: Pierre-Alain Boucard, David Néron, François Louf, Philipp Junker