Unveiling similarities and differences in oxidation processes of oxidants and derived reactive oxygen species through machine learning interpretation of oxidation rate constants

Accurately understanding the similarities and differences in the oxidation processes of oxidants and derived reactive oxygen species (ROS) is crucial for machine learning models to achieve chemical accuracy and rationality when quantitatively predicting ROS oxidation rate constants (kAOP). In this study, we utilized thirteen molecular fingerprints to express pollutant structures and combined six machine learning models to accurately predict the kAOP of three typical reactive oxygen species, ozone (O3), hydroxyl radical (center dot OH), and sulfate radical (SO4 center dot-), while exploring their degradation mechanisms. Our established models achieved high determination coefficients of 0.73, 0.84, and 0.85 for O3, center dot OH, and SO4 center dot-, while the external validation coefficients were 0.73, 0.83, and 0.85, respectively, which were higher than those reported in published studies. The excellent generalization ability of the models was demonstrated by their accurate prediction of kAOP outside the data sets with a threshold of less than 0.1. Furthermore, we systematically interpreted the models using the SHAP method and chemical theory. Our results showed that SO4 center dot- prefers to destroy aromatic rings and amines by single electron transfer, center dot OH tends to attack thioether structures by radical adduct formation, and O3 likes to degrade hydroxyl groups and unsaturated carbon atoms by hydrogen atom abstraction.