Physics-informed neural network for constitutive modeling of cyclic crystal plasticity considering deformation mechanism

Neural networks (NNs) have become increasingly crucial tools in the multiscale constitutive modeling of elastoplastic materials. In the present paper, a physics-informed neural network-based constitutive model within the crystal plasticity framework is proposed. The architecture of the neural network is designed and decoupled according to the crystal slip mechanism. The neural network training combines slip internal variables with hardening equations, enabling learning with or without labeled data. As a result, the monotonic and cyclic mechanical behaviors of single-crystal materials are accurately predicted. The extrapolation ability of the neural network, which is enhanced by the guidance of physical information and the designed network architecture, is evaluated by both FE simulations and a simple experimental case under complex loading conditions. Moreover, an online self-training scheme on top of conventional offline learning strategies is developed. The online learning serves as a corrective measure when the accuracy of the offline model is insufficient, which achieves the same results as FE simulations. The current work provides a novel and efficient method for calculating single-crystal elastoplastic behaviors. Moreover, it offers a new perspective for constitutive modeling with neural networks by embedding deformation mechanisms.