Interpretable Prediction of Mechanical Properties in Hot Strip Rolling by Combining Machine Learning with Shapley Additive Explanations

Accurate prediction of mechanical properties is essential for quality control in hot strip rolling (HSR), where the relationships among chemical composition, process parameters, and mechanical properties are highly nonlinear under industrial conditions. In this work, a data-driven framework was established for the prediction and interpretation of yield strength (YS), tensile strength (TS), and elongation (EL) of hot-rolled strips based on industrial production data. A high-quality dataset was constructed through data collection, outlier removal, and feature selection. Six machine learning (ML) models were developed and compared, and particle swarm optimization (PSO) was employed for hyperparameter tuning. The results showed that random forest (RF) achieved the best overall predictive performance, with R2 values of 0.979, 0.986, and 0.959 for YS, TS, and EL, respectively. In addition, faster convergence and better optimization performance were obtained by PSO than by genetic algorithm (GA) and artificial bee colony (ABC). Shapley additive explanations (SHAP) were further introduced to reveal both global feature importance and local feature contributions. The proposed framework provides an effective approach for mechanical property prediction and alloy design in HSR.