Hybrid Data‐Driven and Physics‐Informed Learning of Cyclic Plasticity For Pipe Steel

ABSTRACT An efficient and explainable machine learning (ML) approach is presented, replacing conventional material models based on the radial return mapping (RRM) algorithm for the constitutive modeling of cyclic plasticity in 3D. The application of transfer learning, based on an existing model for a separate class of steel, leads to a significant reduction of computational effort and training time. High accuracy and stability are achieved by physics‐informed regularization and including back stress information. The loss function is designed to stipulate several qualitative restrictions: deviatoric character of internal variables, compliance with the flow rule, and the differentiation between elastic and plastic steps. The approach is validated by means of a case study for the carbon steel API 5L X65 pipe. The base model used for transfer learning represents the plastic behavior of steel alloy 4130. The validation shows cyclic stability and deviations in normal directions of less than 2% at peak values, which is comparable to the order of measurement inaccuracies. The accuracy in a pure shear test is limited and inspires further improvements in the sampling strategy.