A long short-term memory-based constitutive modeling framework for capturing strain path dependence in plastic deformation

Abstract

Macroscopic models struggle to capture the strain path-dependent behavior of metallic materials, particularly under random loading conditions. While crystal plasticity models effectively describe complex strain path dependence due to their physical basis, they suffer from significant computational inefficiencies and limited scalability. To address these challenges, this study introduces an LSTM-based constitutive modeling framework, a novel data-driven approach. The framework starts with fundamental experiments, optimized using a BPNN method to derive constitutive parameters for a crystal plasticity model. An extensive dataset is generated by simulating crystal plasticity along various random strain paths, which is used to train the LSTM network. The resulting model demonstrates exceptional computational efficiency, providing predictions in under 5 s—far faster than the 30-min crystal plasticity simulations. The LSTM-based model accurately predicts responses for strain paths outside the training dataset, exhibiting low RMSE and MAE values. Experimental results from six strain paths confirm the model's accuracy, capturing behaviors such as the Bauschinger effect and orthogonal hardening/softening. This framework offers a promising alternative to traditional constitutive models, extending crystal plasticity to macroscopic processes and enabling precise engineering predictions. The framework is also adaptable to other materials and holds potential for solving time-series related challenges.