A cyclic elasto-plastic constitutive model based on physics informed neural network of a pure polycrystalline copper under uniaxial loading

Abstract

A cyclic elasto-plastic constitutive model based on physics informed neural network is constructed to describe the cyclic hardening as well as non-Massing behaviors of oxygen free highly conductive copper under fully reversed strain-controlled loading. By translating and rotating the data points, the tension branch and compression branch in stress–strain hysteresis loop are unified. Internal variables of de-hardening coefficient, unified back stress, and saturation back stress are summarized from the unified stress–strain data. In order to construct the relationship between the internal variables and plastic deformation with limited experimental data, the physics informed neural network is established by combining the physics information constraints and the artificial neural networks. And four multilayer perceptrons are trained to identify the internal variables. A plastic prediction-elastic correction algorithm is proposed in conjunction with multi-layer perceptron to solve the constitutive. To evaluate the model, an Armstrong-Frederick type model with strain memory surface is conducted for comparison. The results show the model based on physics informed neural network can obtain accurate results while avoiding a large number of material parameters and partial differential equations.