Toward Neural Network Models to Model Multi-phase Solids

Abstract

In this study, a neural network model is developed to describe the large deformation response of a multi-phase material, i.e., a two-dimensional perforated plate. Using the finite element, virtual experiments are performed to generate stress–strain data for monotonic biaxial loading paths. Subsequently, a combination of fully connected and recurrent neural network models are trained and validated using the results from the virtual experiments. The predictions of a network show a remarkable good agreement with all the experimental data. The suggested neural network-based constitutive model does provide a robust solution to the problem at hand, providing a fully anisotropic, three-dimensional material model capable of covering all physical material properties. The suggested procedure promises to be generally applicable to any material class and can be paired with any numerical method.