QSAR-enhanced machine learning for mechanistic insights and real-time prediction of DBPs in drinking water distribution systems.

Disinfection by-products (DBPs) remain a major health concern in drinking water, but unified prediction across diverse species with mechanistic clarity is still difficult. Current approaches also struggle with limited dataset size, lack of interpretability, and overreliance on laboratory data restricting their practical use in drinking water distribution systems (DWDSs). In this study, we develop a unified, explainable machine learning framework that integrates online water quality parameters with quantitative structure-activity relationship (QSAR) descriptors to enable real-time DBP prediction in DWDSs. By applying vertical and horizontal data augmentation, six ensemble tree-based algorithms were systematically evaluated, among which CatBoost consistently achieved the best performance (R2 = 0.84; RMSE = 2.288 μg/L; MAE = 1.125 μg/L). An optimized 14-feature hybrid set (six online parameters, four molecular descriptors, four fingerprint bits) improved online prediction by +10.4% in R2 and -7.6%/-9.4% in RMSE/MAE versus models without structural features, and generalized in external validation (R2 = 0.65). Model interpretation with SHAP, PDP/ICE, and LiNGAM revealed that carbonaceous DBPs are mainly driven by NOM and chlorine, nitrogenous DBPs by nitrogen, and that cross-domain interactions uncovered hidden mechanistic pathways absent from bulk-only models. These results advance real-time DBP prediction and mechanistic insights by coupling chemical structure information with online monitoring data, offering a pathway toward real-time, risk-based management of DWDSs.