Unified prediction of uniaxial ratcheting deformation at elevated temperatures with physics-informed multimodal network

Abstract

The ratcheting behavior of plastic deformation accumulation under asymmetric loading poses significant risks to the safe service of engineering structures. For accurate prediction of the ratcheting behavior of the material, a physics-informed multimodal network named Dual Stream GRU (DSGRU) model is proposed with training and validation of 316LN stainless steel samples. By incorporating the unrecoverable characteristic of ratcheting behavior into the loss function, there is a significant improvement in the prediction and generalization performance of the DSGRU model. Meanwhile, the multimodal network enables the model to consider material properties at different temperatures. Through sufficient constitutive simulation samples, the DSGRU model with optimal architecture is well-trained and transferred to small sample experimental samples with fine-tuning method. Whether in pre-training or transfer learning processes, the physics-informed loss function ensures the physical consistency of predicted results.