Mapping the UV response of natural organic matter to chlorination: From fraction-resolved fingerprints to operational monitoring of DBP formation

• Chlorination reveals chromophore-specific reactivity in the humic substance fraction • Four non-redundant wavelengths identified from HS differential absorbance clustering • UV fingerprint maps aromatic activation, electron reactivity and bleaching • Multiwavelength model outperforms 254 nm in independent validation for DBP prediction • Transfer from HPSEC fraction-resolved to bulk absorbance for operational monitoring Predicting disinfection by-product (DBP) formation in drinking water requires optical surrogates that are both mechanistically grounded and operationally feasible. While UV absorbance at 254 nm (A 254 ) is widely used, its performance is inconsistent across water sources and it provides limited insight into DBPs speciation. This study develops a data-driven framework to identify multiple and non-redundant UV wavelengths based on chlorine-induced spectral changes in humic substances (HS), the dominant and most reactive fraction of natural organic matter in surface waters. Using HPSEC-DAD-OCD analysis of samples from three drinking water treatment plants with contrasting drinking water sources differing in NOM composition and bromide levels (bromine-rich river vs low-bromide reservoir), differential absorbance spectra (DAS) of HS were subjected to hierarchical clustering, consistently identifying four spectral domains represented by wavelengths at 220, 252, 290, and 362 nm. These wavelengths capture complementary information on aromatic activation, electron-transfer chromophores and conjugated-system bleaching during chlorination. To assess practical applicability, the four wavelengths were transferred to bulk UV absorbance measurements and used to build predictive models. Validation with a fourth independent plant (n=80, 10-fold cross-validation) demonstrated that the multiwavelength model (MWL4) significantly outperformed conventional surrogates, achieving R 2 CV = 0.90 for trichloromethane and 0.85 for total trihalomethanes, compared to 0.77 and 0.75 with A 254 . For brominated species and haloacetic acids, essentially unpredictable with A 254 (R 2 CV < 0.10), MWL4 achieved moderate but meaningful accuracy (R 2 CV = 0.47–0.58). This framework provides a chemically grounded approach for operational DBP risk assessment, with initial validation at an independent site.