A strategy combining interpretable machine learning, NSGA-III optimization model and strengthening and toughing mechanism to predict microstructure for rolled AZ31 magnesium alloy sheets

This study constructed an interpretable prediction model for establishing the relationship between rolling parameters and microstructure parameters in AZ31 magnesium alloy rolled sheets using machine learning and Shapley Additive exPlanations (SHAP), and a NSGA-III optimization model combined with strengthening and toughing mechanisms was used to find better process parameters. By coupling the SHAP model with Pearson correlation coefficient (PCC), the relationships between rolling process parameters (temperature, average strain rate, and reduction) and microstructure parameters (average grain size (AGS) and twin density (TD)) were revealed. The NSGA-III algorithm was employed to identify the optimal range of process parameters, establishing a reliable method for rapidly optimizing the rolling process. By comparing evaluation metrics across BP-IPSO, SVR, RF, and KNN machine learning models, it is found that the SVR model demonstrated superior performance in predicting AGS, while the KNN model with an augmented dataset exhibits higher prediction accuracy for TD. Integrating the PCC and SHAP model, it is inferred that AGS and TD are mainly affected by average strain rate. Based on strengthening and toughing mechanisms and the multi-objective genetic algorithm NSGA-III, the optimal process parameter range is determined to be temperature of 398∼409 °C, average strain rate of 3.2–7.2 s⁻¹, and reduction of 72∼76 %. Finally, the validation experiments confirm that the predictions obtained from the proposed method are consistent with the experimental results, thereby verifying the accuracy and practical effectiveness of integrating interpretable machine learning, the NSGA-III multi-objective optimization model, and strengthening-toughening mechanisms.