Predicting Reaction Rate Constant of Organic Compounds with Singlet Oxygen and Revealing Its Contributors Using Machine Learning

Singlet oxygen (¹O₂), as an electrophilic reagent, is widely used for oxidative degradation of organic compounds (OCs). The ¹O₂ reaction rate constant (k) is an important index for evaluating the OCs degradation rate. Under limited conditions, owing to the difficulty in experimentally determining k of all OCs, researchers have gradually developed predictive models. However, research on how to evaluate the best modeling method from multiple models is still lacking. In addition, this study also introduces the environmental variable (pH) in log k model development for the first time. On this basis, eight models were developed using linear and nonlinear methods, and a comprehensive evaluation system based on the TOPSIS method was proposed. The evaluation results indicated that the CatB model (R_tra² = 0.980, Q_Loo² = 0.775, Q_test² = 0.868) exhibited outstanding comprehensive performance and the best prediction effect. The interpretation based on the SHAP values revealed that the key influences of E_(HOMO), pH, electronegativity, van der Waals surface area and molecular structure had major effects on the degradation of OCs by ¹O₂. Interpretable machine learning methods provide a potential shortcut for predicting k, offering significant guidance for optimizing experimental design and improving environmental management efficiency.