Study of LSTM-based constitutive modeling for coupled multimodal microstructure time-series damage in DZ125 superalloy

Abstract

A novel modeling method for DZ125 superalloy has been proposed, integrating a long short-term memory (LSTM) network into the Chaboche unified viscoplasticity constitutive model. Initially, the modified Chaboche constitutive model, incorporating the multimodal microstructure coupled with time-series damage, was developed and implemented using the UMAT subroutine in ABAQUS. Subsequently, damage parameters were determined based on the extraction of three microstructural features, enabling the establishment of an LSTM network for predicting the damage variable, which was then embedded into the UMAT subroutine. Finally, the comparative analysis indicated that the LSTM model achieved nearly the highest prediction accuracy and shortest calculation time, while the UMAT-LSTM model uniquely enabled the prediction of mechanical behavior responses at any given service time. The UMAT-LSTM model developed in this study achieved cross-platform integration, effectively combining the embedded LSTM network’s data-driven learning capability with the constitutive model’s physical mechanism. This approach provides a cost-effective and time-efficient nondestructive solution for predicting the mechanical properties of hot section components.