Mechanical submodels driven by machine learning : application to structural dynamics

The primary goal of this thesis is to develop efficient and reliable numerical methods and deep learning methods for the reduction of parametric and/or non-parametric contact models in structural dynamics, including impact zone scenarios that can evolve over time on cabin aeronautical equipment. The approach is to determine a zone of interest in the physical model and construct models capable of generating boundary conditions to the physical model around the zone of interest. This modelisation will allow to explore the parametric space using the generative model while keeping the high-fidelity caracteristics of the physical solutions by solving the physical problem in the area of interest, and then use it to test out a variety of impact scenarios. Thus reducing the computational cost of the physical model. Our source code for Europlexus will be used to create the program. There will be more Python development for deep learning methods.