Predicting the elasto-plastic response of short fiber composites using deep neutral networks trained on micro-mechanical simulations

Abstract

The mechanical modeling of short fiber composites has proven to be difficult. This is partly owing to the high degree of anisotropy and fiber discontinuity. Being able to accurately predict the behavior of short fiber composites with varying fiber orientations and fiber volume fractions is highly relevant in the design and production of injection molded parts. It has therefore become popular to abandon classical constitutive models in favor of data driven models. Artificial neural networks is a popular and efficient method of using large amounts of data to teach an algorithm the underlying rules of a phenomenon, enabling it to make predictions for data it has never before encountered. In this work a recurrent deep neural network model utilizing Gated Recurrent Units (GRU) is trained to predict the elasto-plastic stress response of a short fiber composite material given the strain path. The model is designed to have the ability to make predictions for arbitrary fiber orientations and varying fiber volume fractions. The training data is generated by performing micro-mechanical simulations utilizing mean field methods in the commercially available software digimat-mf. The strain data is generated by a random walk scheme in strain space. The finished model performs well, and the mean error for a typical load cycle usually stays below 10% of the matrix yield stress.