Interpretable machine learning predictive model for mechanical properties of AZ31 magnesium alloy rolled sheets

To investigate the complex relationship between rolling process parameters and mechanical properties of AZ31 magnesium alloy rolled sheets, the Leave-One-Out Cross-Validation (LOOCV) and parameter tuning were applied to optimizing hyper-parameters for the four (BPNN, SVR, RF, and KNN) machine learning models. An interpretable prediction model based on machine learning and SHapley Additive exPlanations (SHAP), as well as an analytical method combining the SHAP model and the Pearson Correlation Coefficient (PCC), were proposed. The results showed that among the four models, the SVR model was able to simultaneously and accurately predict the ultimate tensile strength (UTS) and elongation (EL). According to the combination analysis of PCC and the magnesium alloy rolling forming mechanism, it was found that strain rate and reduction displayed a negative and positive correlation with UTS, respectively, while rolling temperature and reduction illustrated a positive and negative correlation with EL, respectively. Through the SHAP method, which could interpret the output results of the SVR machine learning model, it was deduced that reduction and strain rate played an important role in the SVR model of the outputs of the UTS and EL, respectively. Combining SHAP with PCC, it was found that strain rate and reduction had a greater influence on the UTS than rolling temperature, whereas strain rate and rolling temperature had more influence on the EL compared to reduction.