Multi-domain data integration for a predictive geometallurgy framework applied to MgO control in phosphate beneficiation

The control of magnesium oxide (MgO) content in apatite concentrates is a critical challenge in phosphate processing plants hosted in carbonatite deposits, as elevated MgO levels degrade concentrate quality and impair downstream fertilizer production. This study presents a data-driven geometallurgical framework to forecast MgO content in the final concentrate by integrating geological, chemical, metallurgical, and operational variables through machine learning. A comprehensive dataset from 258 phosphate ore stockpiles was compiled, encompassing lithological proportions, head-sample chemistry, grinding parameters, and plant performance data. Among five regression algorithms evaluated, Support Vector Regression (SVR) demonstrated the most robust performance under Leave-One-Out Cross-Validation (LOOCV), achieving a satisfactory R2 of 0.69 and a mean absolute percentage error (MAPE) of approximately 13%. Model stability and uncertainty were rigorously assessed using bootstrap confidence intervals and k-fold validation. SHapley Additive exPlanations (SHAP) were employed to interpret the underlying logic of the model, identifying plant feed MgO, micaceous lithologies, and Al-/Fe-bearing phases as the primary controls on MgO deportment. To bridge the gap between complex modeling and industrial routine, a simplified linear surrogate equation was derived and validated against an independent external dataset. The proposed framework provides a transparent, operationally relevant tool for stockpile management and blending strategies, enabling proactive quality control in phosphate beneficiation.