Exploring decisive operating factors for micropollutants fate in ultraviolet-based advanced oxidation processes using the integrated clustering-classification model

In this study, we present an integrated clustering-classification machine learning (ML) model for optimizing the removal of micropollutants in water using ultraviolet-based advanced oxidation processes (UV-AOPs). We evaluated the degradation of 28 micropollutants using five UV-AOPs: UV/TiO 2 , UV/H 2 O 2 , UV/Fenton, UV/ peroxydisulfate (PDS), and UV/Cl 2 , and thereafter, investigated the optimization of these AOPs using six ML models (Random Forest, Gradient Boosted Decision Tree, Extremely Randomized Trees, Extreme Gradient Boosting (XGBoost), Categorical Boost, and k-nearest neighbors). Among the six models, XGBoost showed superior performance, exhibiting the highest prediction accuracy (>0.9) and shortest computing time. It also provided valuable insights into the classification criteria for ensuring a high micropollutant removal rate. SHAP analysis further identified photoreactivity and controllable features (e.g., oxidant dose and type) as the key factors driving the degradation of micropollutants in different clusters. Second-order reaction rate constants (e.g., that for •OH in the UV/Fenton process and SO 4 •- in the UV/PDS process) also played critical roles in specific clusters. These findings highlight the potential of ML-based optimization in facilitating the selection of appropriate process types and operating conditions for micropollutant degradation. Additionally, the application of these findings may enable a more effective evaluation of micropollutants removal efficiency in water treatment. • ML enables the analysis of micropollutant (MP) degradation in UV-based AOPs. • Graph-based clustering of MPs by physicochemical properties improves model accuracy. • XGBoost shows the highest prediction accuracy (>0.9) among six ML models. • Photoreactivity and controllable parameters contribute differently to MP fate.