Machine learning decoding of electronic thresholds in prediction of oriented generation of solar-induced oxidative reactive species and antibiotics degradation

The oriented generation of solar-induced reactive species from catalysts is crucial for the degradation of antibiotics. This study establishes a multi-scale framework combining theoretical calculation and machine learning (ML) for designing solar-driven heterojunction catalysts to degrade antibiotics. A database of 453 catalysts (including metal oxides, metal sulfides, g-C₃N₄, and BiOX) with 36 key parameters was constructed. The CatBoost-based dual-task model demonstrates excellent performance in predicting degradation kinetics for antibiotics (R² = 0.8073) and identifying reactive species generation (0.9360 test accuracy and 0.9480 precision). Notably, Shapley Additive exPlanation (SHAP) analysis coupled with recursive feature reveals d-electron amount of metal elements and the electronegativity of non-metal thresholds govern reactive species orientation. A nonlinear relationship exists between reaction time/specific surface area and degradation efficiency, with a plateau observed in the degradation kinetics beyond a critical time and specific surface area threshold. Experimental validation and theoretical calculation on five customized heterojunctions (e.g., SrTiO₃/Bi₂O₃) confirm model reliability, with degradation efficiency errors of 10 % and reactive species distributions aligning with 100 % predictions. This work provides an interpretable ML paradigm overcoming traditional trial-and-error limitations in multi-component environmental catalytic optimization, enabling targeted design of antibiotic degradation systems through electronic structure-property correlation analysis.