Optimising SAG mill throughput and circulating load using machine learning models: A multi-objective approach for identifying optimal process parameters

Grinding is a fundamental operation in mineral processing plants. Semi-autogenous grinding (SAG) mills are widely utilised in these circuits. Enhancing SAG mill throughput is crucial due to its significant financial impact. However, controlling the mill discharge particle size is also essential. If the discharge particles are oversized, a substantial portion of the SAG mill output is recirculated to pebble crushers. Therefore, the challenge addressed in this research is how to set controllable process parameters to achieve both maximum mill throughput and minimum circulating load. This research addresses the problem by formulating it as a constrained multi-objective optimisation problem. An industrial dataset, consisting of 66,776 records is utilised. Initially, three gradient boosting methods including CatBoost, XGBoost, and HGBM are compared for modelling both objectives. The most precise prediction models are subsequently employed within four multi-objective optimisation algorithms consisting of NSGA-II, MOEA/D, RVEA, and SPEA2. The comparative results indicated that CatBoost is the best-performing prediction model with an average R 2 of 0.9304 and 0.9031 for mill throughput and circulating load, respectively. NSGA-II is also identified as the best-performing optimiser, achieving the highest average hypervolume of 291,489. A sensitivity analysis demonstrated how changes in input features affect objectives and confirmed the robustness of the optimal solutions. By analysing the top-performing solutions, recommendations for process parameter setting are provided. Additionally, the advantage of the developed framework in this research is providing a set of Pareto-optimal solutions, allowing experts to select the most suitable settings based on real-time process conditions.