Machine learning-based constitutive model for metal plasticity

Abstract

In this work, it is proposed a machine learning-based constitutive model to predict elasto-plastic behavior for metal plasticity. An artificial neural network (ANN) model is used for machine learning, and the conventional theoretical constitutive model composed of J2 plasticity, isotropic hardening model, and associated flow rule was replaced. Furthermore, the proposed ANN-based constitutive model is a problem-independent model, which is applicable to arbitrary sheet metal forming simulation. For the problem-independent generality of the proposed ANN-based model, sufficient amount of training data are required where the dataset should represent the constitutive relation of the targeted material. In this work, a training dataset suitable for the aforementioned requirement was investigated. In addition, in order to reduce the size of the training dataset, the return mapping scheme in principal stress space was employed. Before training an ANN-based model, a case study on neural architectures was performed, and then the ANN-based model was implemented on User MATerial interface (UMAT) for a commercial FE code (ABAQUS2016/STANDARD). The accuracy of the ANN-based model was verified with two types of finite element simulations: single element simulation under three typical loading paths and dog-bone tensile test simulation. The simulation results predicted from the ANN-based constitutive model show a good agreement with those from the conventional theoretical constitutive model. Finally, the proposed ANN-based constitutive model was applied for circular cup drawing simulation. The cup profiles predicted from the conventional theoretical model and the proposed ANN-based model show a good agreement. In terms of computational cost, the ANN-based model shows better computational efficiency than the conventional theoretical constitutive model.