Résumé :

Recent years have seen the development of various Machine-Learning based approaches to material modeling. Among them, Recurrent Neural Networks (RNNs) hold a particular place due to their ability to treat sequential data. Notably, they can construct their own internal state-variables in order to reproduce history-dependent material behavior based purely on stress-strain sequences.

Most of the literature on RNN-based mechanical models relies on established RNN architectures called LSTMs and GRUs. In this presentation, we will show that modifying the RNN architecture can provide several advantages. The proposed architectures provide compact models, can reproduce the state-space of phenomenological models, and are usable in complex explicit finite element simulations.

Due to the combined large data requirements of RNNs and technical difficulties in leveraging experimental results, we will focus on stress-strain sequences drawn from numerical examples. We will detail applications to phenomenological models, crystal plasticity and structural simulations.