Employing deep learning in non-parametric inverse visualization of elastic–plastic mechanisms in dual-phase steels

Abstract

Enhancing the interpretability of machine learning methods for predicting material properties is a key, yet complex topic in materials science. This study proposes an interpretable convolutional neural network (CNN) to establish the relationship between the microstructural evolution and mechanical properties of non-uniform and nonlinear multisystem dual-phase steel materials and achieve an inverse analysis of the elastic-plastic mechanism. This study demonstrates that the developed CNN model achieves an accuracy of 94% in predicting the stress-strain curves of dual-phase steel microstructures with different compositions and processes, with the mean absolute error not exceeding 50 MPa, representing merely 5.26% of the average tensile strength of dual-phase steels in the dataset. The reverse visualization results of the CNN model indicate that, during tensile deformation, the grain boundaries maintain deformation coordination within the grains by impeding dislocation slip. This results in a significant stress concentration at the grain boundaries, with stresses at the boundaries being higher than those borne by the martensitic phase and minimal stresses in the ferrite phase. Moreover, compared with traditional crystal plasticity models, the CNN model exhibits a substantial improvement in computational efficiency. This method provides a generic plan for improving the interpretability of machine learning methods for predicting material properties and can be easily applied to other alloy systems.