Recurrent neural network model predicting elasto-plasticity and matrix fracture in fiber-reinforced composites

An enhanced recurrent neural network model is proposed to predict the fracture initiation in addition to the elasto-plastic stress–strain response of solids. The stress-state dependency of the fracture initiation in an epoxy is characterized using shear, plane-strain tension and equi-biaxial tension experiments. A Drucker-Prager elasto-plasticity model with Hosford-Coulomb fracture is chosen to describe the experimentally-observed epoxy behavior. After detailing the extended minimal state cell (MSC) formulation, we first demonstrate that the RNN model is able to learn the deformation response of the matrix material based on synthetic data. Subsequently, an RVE model predicting the in-plane response of a carbon-fiber reinforced epoxy is built and used to generate training data for different types of random-walks in strain space. Based on the results of a comprehensive hyperparameter study, it is shown that a compact MSC model with five state variables and 6500 parameters is able to predict the homogenized stress–strain response of a fiber-reinforced composite along with the fracture initiation with the RVE. After demonstrating the model’s self-consistent behavior upon refining strain paths, it is also validated at the structural level by comparing its predictions for a three-point bending problem with those of a detailed heterogeneous model that discretizes all fibers in the beam.