Integrating ozone-reactive moieties with molecular representations of trace organic contaminants improves the prediction of ozone reaction kinetics by machine learning.

Ozone (O3) is widely applied in water treatment to abate trace organic contaminants (TrOCs), yet accurate prediction of its reaction kinetics remains challenging. Existing machine learning (ML) models generally operate as "black boxes", treating TrOC molecules as whole entities with limited mechanistic insight into O3 reactivity. As a strong electrophile, O3 reacts with TrOCs in a manner governed by the types and electronic properties of O3-reactive moieties. Hence, we developed an interpretable ML framework integrating O3-reactive moieties with molecular representations to predict the absolute second-order reaction rate constants (kabs) of O3 with TrOCs. A dataset comprising 477 kabs values for structurally diverse TrOC species spanning broad O3 reactivities was compiled. Compared with models using molecular representations alone, the integrated models consistently exhibited improved prediction accuracy across molecular representations and ML algorithms. For all three molecular representations, optimally integrated models achieved test-set coefficients of determination above 0.85, with root mean square errors and mean absolute errors below 1.15 and 0.89, respectively. The support vector machine with Mordred descriptors exhibited the best performance by capturing structure-reactivity relationships and enabling identification of dominant reactive moieties, as demonstrated by SHapley Additive exPlanations (SHAP) analysis. The absolute SHAP values of key reactive moieties (typically ≥ 1, e.g., olefins) were clearly distinguished from others (≤ 0.5). External validation with six dissociating TrOCs supported the prediction accuracy and identification of dominant reactive moieties. These findings highlight the value of explicitly incorporating mechanism-based reactive moiety information into ML models for accurate and interpretable prediction of O3 reaction kinetics.