Neural network constitutive model for uniaxial cyclic plasticity based on return mapping algorithm

Structures are subjected to cyclic loading during earthquakes. Therefore, to develop accurate constitutive models for simulating structural response to earthquakes, the cyclic plasticity of structural steels should be considered. Neural network constitutive models reproduce the plastic behavior of materials by applying the learning capability of neural networks. Although numerous monotonic loading models have been developed, there have been relatively few applications of neural network constitutive models to cyclic plasticity. In this study, a neural network constitutive model with good generalization ability and high computation efficiency was developed to model the stress and strain history of materials under cyclic loading. The model employs a feed forward neural network as part of an implicit stress integration scheme implemented by the return mapping algorithm for the hardening model proposed by Chaboche. An appropriate network architecture was constructed using numerical results to set the number of neurons in the hidden layer and the batch size. The results obtained by the trained neural network under random loading agreed well with those obtained by a return mapping algorithm, along with a 70% reduction in calculation time.