Unraveling pollutant-catalyst-oxidant matching in persulfate-based advanced oxidation processes through interpretable deep learning

Persulfate-based advanced oxidation using peroxymonosulfate (PMS) or peroxydisulfate (PDS) has been widely investigated for organic contaminant degradation, yet pollutant-catalyst-oxidant matching remains largely empirical. Here, we compiled 1,487 kinetic entries from 95 publications and developed a dual-branch deep learning framework that separately represents catalyst/condition information and pollutant molecular descriptors, preserving pollutant-specific structural signals in heterogeneous datasets. The model achieved a best-run test R² of 0.89. A triangulated attribution strategy integrating data variability, statistical association, and model contribution was combined with pollutant clustering to identify robust and class-resolved structure-reactivity relationships. Subsequent joint analysis of SHapley Additive exPlanations (SHAP) response curves and partial dependence plots (PDPs) resolved the response directions and nonlinear effects of key descriptors across the two oxidant systems. Molecular aspect ratio (L/D) and molecular density (ρ_mol) emerged as important geometric differentiators of PMS and PDS kinetics, complementing conventional electronic descriptors. The resulting dual-regime framework links PMS more strongly to geometric accessibility and localized interfacial activation, while PDS is more closely associated with electronic characteristics related to interfacial electron transfer. Cross-catalyst experiments using materials absent from the training dataset further reproduced the identified material- and pollutant-dependent trends. These results establish a structure-aware framework for interpreting PMS/PDS selectivity and for generating experimentally testable pollutant-catalyst-oxidant matching hypotheses.