Yanlin Zhang, Xue Yang, D N Wang, Jiaquan Zhang
Thermally activated persulfate oxidation is an effective advanced oxidation process for the removal of refractory organic contaminants, yet quantitative prediction of degradation kinetics and interactions between process parameters remain challenging due to the complicated reaction mechanisms and heterogeneity of experimental conditions. In this study, a machine learning framework was established to predict apparent degradation rate constants in thermally activated persulfate systems using a curated dataset of 580 data points collected from 53 peer-reviewed studies. Fourteen input variables were considered, with the degradation rate expressed as -log(k) as the target variable. Six supervised regression models were developed and evaluated, including CatBoost, XGBoost, LightGBM, random forest, support vector regression, and artificial neural networks. Among them, CatBoost achieved the most robust predictive performance, exhibiting high accuracy and generalization capability. Model interpretation using permutation importance, SHapley Additive exPlanations, and partial dependence plots identified temperature, initial contaminant concentration, and oxidant dose as the dominant factors governing degradation kinetics, whereas molecular descriptors played a secondary role within the studied domain. Applicability domain analysis further confirmed the reliability of model predictions across most experimental conditions. This work provides an interpretable data driven approach for analyzing and optimizing thermally activated persulfate oxidation processes in water treatment.