Interpretable Machine-Learning Prediction of Pollutant Adsorption and Degradation by MnO2 in Aqueous Systems

Abstract Manganese dioxide (MnO2) is a key redox-active mineral governing pollutant migration and transformation, making quantitative prediction of MnO2–pollutant interactions essential for evaluating contaminant fate. Here, we compiled two data sets covering MnO2 adsorption and abiotic degradation across 7 MnO2 structural categories, 48 organic pollutants, and 27 inorganic pollutants. By integrating molecular descriptors, separate machine learning models were developed for predicting log qmax and log kobs. Random forest achieved a mean test R2 of 0.69 for adsorption, whereas CatBoost reached 0.82 and 0.83 for degradation using PaDEL and RDKit descriptors, respectively. Permutation importance and SHAP analysis showed that adsorption was weighted primarily toward environmental conditions, whereas degradation predictions depended jointly on environmental conditions and pollutant properties. SHAP dependence and PDP analyses further clarified feature–prediction relationships, identifying pollutant concentration as the dominant adsorption predictor, likely reflecting concentration-dependent site saturation and isotherm design, while reaction pH and molecular descriptors were the dominant predictors of degradation. The models were integrated into a graphical user interface to enable rapid prediction of MnO2 reactivity. By integrating contaminant retention and transformation into a unified, interpretable framework, this study provides predictions and testable mechanistic hypotheses for contaminant fate in Mn-bearing aquatic systems.