Chenchong Wang

The State Key Lab of Rolling and Automation, Northeastern University, Shenyang, 110819, China 

ABSTRACT: The demanding operational enviromnents necessitate increasingly complex mechanical perfo皿ance requirements for advanced metal structural materials in engineering applications. Fundamental mechanical properties such as high strength, high ductility, and robust work hardening behavior remain critical prerequisites to meet engineering demands. Accurately and efficiently predicting the mechanical behavior of metal structural materials has been a long-standing challenge. To accurately reflect the microstructure-property relationship of crystals, complex modifications to crystal plasticity constitutive equations are required. Furthermore, re-measuring constitutive parameters under different compositions or processing conditions significantly limits modeling efficiency and applicability. Establishing the relationship between composition, microstructure, and performance is the foundation for the reverse design and process optimization of materials, particularly complexly structured metal alloys. This study introduces a novel approach that combines traditional crystal plasticity (CP) theory with machine learning. Firstly, a crystal plasticity finite element model considering the real texture and the influence of carbon content on mechanical performance in dual-phase steel is established. This model simulates the stress-strain behavior of dual-phase steel under uniaxial tension for various compositions and processing conditions, yielding local stress and strain distribution maps under different stress loads. These stress distribution maps, obtained through crystal plasticity simulations, along with their corresponding strain values, are utilized as datasets to train a Convolutional Neural Network (CNN) model. This trained CNN model predicts the stress­strain responses of dual-phase steel in different systems under loading conditions. The results show that using only the crystal plasticity model to predict true stress results in errors mostly distributed between 0-80 MPa, with only a small amount of data exhibiting larger deviations between 80-100 MPa. This indicates that specific crystal models are suitable for the corresponding compositions and processing conditions of dual-phase steel organizations. However, when the system changes, the actual material properties change, and the crystal model cannot accurately predict the material's mechanical behavior. In contrast, the CP-CNN model predicts true stress with a more concentrated error distribution, with the majority falling between 0-40 MPa. When predicting the mechanical performance of dual-phase steel in different systems, the well-trained CP-CNN model demonstrates smaller overall prediction errors compared to the crystal plasticity model. Therefore, the established CP-CNN model is more effective in predicting mechanical responses than the crystal plasticity model and addresses the issue of the latter's strong system sensitivity.
KEYWORDS: convolutional neural network; stress-strain curve; crystal plasticity model; dual-phase steels