Predictive Modeling and SHAP-Based Interpretability of Manganese and Iron Dissolution in Multi-Acid Leaching Systems Using Hybrid Machine Learning

Abstract

Hydrometallurgical leaching processes contain complex and nonlinear parameter interactions that are difficult to capture with conventional empirical models. In this study, a multiple hybrid machine learning approach was developed to predict manganese (Mn) and iron (Fe) dissolution efficiency in leaching systems and performed using sulfuric acid (H2SO4), hydrochloric acid (HCl), and nitric acid (HNO3). A large-format dataset consisting of 204 independent leaching experiments was generated in which acid type, acid concentration (0.5-5 M), temperature (25-90 degrees C), solid/liquid ratio (100-200 g/L), leaching time (1-4 h), and eight different reducing agent types were systematically varied. XGBoost, LightGBM, CatBoost, and Random Forest algorithms were individually trained and subsequently combined with a Soft Voting Ensemble architecture. Hyperparameters were optimized using the RandomizedSearchCV method with 3-fold cross-validation. The XGBoost model achieved the highest prediction accuracy for Mn dissolution (R-2 = 0.8993, RMSE = 8.06%), while CatBoost demonstrated the best performance in Fe dissolution (R-2 = 0.8415, RMSE = 4.43%). SHAP analysis suggested that the dosage and type of reducing agents are the most influential predictive features for Mn dissolution, while acid molarity and temperature were identified as the dominant predictors for Fe leaching. Friedman test confirmed that performance differences among both Mn and Fe models were statistically significant (Mn: chi(2) = 32.76, p < 0.001; Fe: chi(2) = 25.96, p < 0.001). The developed models contribute significantly to hydrometallurgical process optimization by predicting the nonlinear effects of leaching parameters on metal dissolution with high accuracy. This study presents a comprehensive and interpretable machine learning framework supported by an extensive experimental dataset, a substantial portion of which has not been previously utilized or comparatively analyzed within a unified multi-acid framework, enabling systematic modeling of selective Mn-Fe dissolution across multiple acid systems and reducing agents.