Explainable machine learning reveals how molecular descriptors govern micropollutant degradation in UV/H2O2 oxidation

Abstract

Micropollutants (MPs) pose significant risks to aquatic ecosystems and human health because of their persistence and potential for bioaccumulation. UV/H2O2 oxidation effectively degrades a wide range of MPs through the generation of hydroxyl radicals (center dot OH). However, predicting MP degradation under varying operational conditions remains challenging due to the complex interactions among process parameters and molecular properties. This study advances the understanding of MP degradation in UV/H2O2 systems by integrating experimental data from published studies with explainable machine learning (ML). Nine ML algorithms were developed and evaluated for predicting the degradation rate constant (ln(k)) using process conditions, classical molecular descriptors, and quantum chemical reactivity indices. Model interpretability was assessed through Shapley Additive Explanations (SHAP), and virtual simulations were performed to explore the effects of water matrix characteristics, process conditions, and molecular structure on MPs reactivity. Results show that the best-performing model, a gradient boosting decision tree (GBDT), achieved high predictive accuracy and provided explanations consistent with established center dot OH reaction theory. Process conditions-particularly UV fluence rate, H2O2 concentration, and pH-exerted a stronger influence on ln(k) prediction than molecular properties. Nevertheless, quantum chemical descriptors such as chemical hardness, maximum absorption wavelength, and Fukui indices offered valuable insights into structure-reactivity relationships. By integrating explainable ML with molecular and process descriptors, this study bridges mechanistic understanding and predictive modeling, supporting more informed design and optimization of advanced water treatment systems.