Prediction of SAG mill liner wear and geometry optimization using interpretable machine learning

Abstract

Geometric parameters are considered a critical factor influencing the wear of semi-autogenous grinding (SAG) mill liners. However, traditional liner design optimization studies only focus on single variables and incur substantial time and financial costs. Therefore, this paper proposes a framework that integrates the discrete element method (DEM) with interpretable machine learning and that offers an effective approach to address this issue. In this study, a dataset was constructed based on the numerical simulation results using the DEM. Subsequently, four machine learning models were constructed to predict the wear volume (W). The SHapley Additive exPlanations (SHAP) method was employed to interpret the best prediction model for identifying the response relationship between liner geometric parameters and W. The results showed that variations in the geometric parameters of lifters exhibit complex and nonlinear effects on liners wear. The best W prediction model is the multilayer perceptron (MLP) model with superior accuracy, stability and generalizability. The face angle emerged as the most significant geometric factor influencing liner wear with a relative importance of 48.1 %. Optimal lifter designs should prioritize the following geometric parameter intervals: Face angle < 50.57°, Height < 186.09 mm, Width < 91.27 mm, and Number < 33. The findings of this study indicate that interpretable machine learning combined with DEM provides a promising technique for liner design.