Fitting-free mechanical response prediction in dual-phase steels by crystal plasticity theory guided deep learning

Abstract

Generic prediction of the work hardening behavior and plasticity of metallic structural materials is a long-standing problem. Accurate reflection of the microstructure–property relationship requires complex modification of the crystal plasticity (CP) constitutive equations and the re-measurement of intrinsic parameters in different compositions or processing conditions, significantly limiting the modeling efficiency and extension of the models. In this study, a CP model with concise constitutive equations and non-fitted fixed parameter values is introduced into a convolutional neural network (CNN) using local stress nephograms. The proposed model uses the fundamental CP theory to constrain the training direction of the deep learning strategy. It accurately predicts the stress–strain curves, work hardening behavior, and necking initiation point of dual-phase (DP) steels with different processing or/and composition conditions. The effect of CP constitutive equations on the prediction accuracy of the CP-CNN model is also investigated and has been proven to exhibit strong robustness. The CNN architecture is optimized using the dot product method to better adapt the model to multimodal data in the field of metallic structural alloys to confirm the proposed model's ability to overcome the adverse effects of parameter sensitivity in different composition/processing conditions for the CP theory. Finally, extensibility validation and efficiency analysis are performed to demonstrate the advantages of the proposed model in terms of time consumption and generality. The method provides a new solution for the design and efficient prediction of mechanical properties of steel materials.