Comprehensive Comparison of Machine Learning Approaches—Deterministic and Stochastic—In Modeling the Production and Power of an SAG Mill: A Case Study of the Chilean Copper Mining Industry

SAG grinding mills represent critical energy-intensive operations in copper concentrators, accounting for 30%–50% of total plant energy consumption. The accurate prediction of mill power draw and production rate under varying operational conditions is essential for real-time control, production planning, and energy management. This study presents a comprehensive comparison of ML algorithms for modeling Production and Power in a Chilean copper mining industry. Deterministic and stochastic models were fitted and validated using industrial data from a Chilean copper operation. More representative models were re-estimated and subsequently evaluated under different operating regimes to examine their predictive performance under aggregated conditions of the feeding variables. This procedure allowed for the identification of the modeling approaches that provide the most robust performance across varying operational regimes. The results show that XGB achieved the best predictive performance, with test RMSE and R2 values of 87.98 and 97.35% for SAG Production, and 431.11 and 95.11% for SAG Power, respectively. Stochastic approaches provided complementary uncertainty quantification, supporting risk-informed decision making under variable operating conditions. The analysis by operational regime indicates that XGB presents better fit in the Thick hydraulic regime, for both responses’ variables, which could be explained why a dense pulp operation provides more predictable grinding dynamics. The comparative analysis reveals trade-offs between model complexity, interpretability, computational requirements, and predictive performance, offering practical guidance for selecting appropriate modeling frameworks based on specific operational objectives and data availability in mineral processing applications.