From CP-FFT to CP-RNN: Recurrent neural network surrogate model of crystal plasticity

Recurrent Neural Network (RNN) based surrogate models constitute an emerging class of reduced order models of history-dependent material behavior. Recently, the authors have proposed an alternative RNN formulation that provides stress-responses independent of the time-discretization of the input-path, making it appropriate for integration into explicit finite element (FE) frame-works. Herein, we apply the same methodology to 2D and 3D datasets corresponding to the effective mechanical behavior of an aluminum alloy as obtained through Crystal Plasticity sim-ulations. In both cases, we obtain reasonable approximations of the behavior using RNN models of size ranging from 5'000 to 100'000 parameters. We also develop a methodology to reduce observed numerical instabilities of the finite element implementations.

Automated summary

In this study, RNN models were used to investigate the mechanical behavior of aluminum alloys and reasonable approximations were obtained. A methodology to reduce numerical instabilities in the finite element implementation was also proposed.