Predicting regulated and emerging disinfection byproducts in small drinking water catchments using machine learning

• Machine learning used to predict DBP formation using DOM spectra and hydrochemistry. • Ten DBP parameters were accurately predicted with an R 2 of 0.86 and RMSPE of 28% • A further five DBPs were classified for presence/absence with an accuracy of 95.6% • Predictive models were most sensitive to DOC or A 254 and humic-like fluorophores. • DOM spectral features alone could accurately predict regulated DBP formation. Potentially harmful concentrations of disinfection byproducts (DBPs) arise from unintended reactions between chemical disinfectants and dissolved organic matter (DOM) present in raw drinking water sources. We explore the application of machine learning tools trained on DOM spectroscopic variables and hydrochemical parameters to predict the formation of regulated and emerging DBPs from laboratory chlorination experiments. Raw source water samples from two small drinking water catchments and groundwater were subject to filtration (<0.7 µm) followed by chlorination at pH 7 and 25 °C under dissolved organic carbon (DOC) concentrations of 0.6–15 mg C L –1 . Machine learning models included neural networks, bagging tree techniques and a generalised boosted regression model with binary presence-absence classification using a support vector machine. Ten DBP parameter concentrations, namely total trihalomethanes, total haloacetic acids, trichloromethane, bromodichloromethane, dibromochloromethane, dichloroacetic acid, trichloroacetic acid, dichloroacetonitrile, trichloronitromethane (chloropicrin) and trichloropropanone could be quantitatively predicted (average R 2 = 0.86, root mean squared percent error = 27.9 %) with a further five species classified for binary presence-absence only (95.6 % average accuracy). Models were most sensitive to two widely reported humic-like fluorophores together with UV–Vis absorbance at 254 nm. Inclusion of hydrochemical parameters (e.g., DOC) only marginally improved model performance. Our findings demonstrate a proof-of-concept for the utility of machine learning models trained on DOM spectroscopic variables using a relatively small sample size (n = 198) with a scalable workflow freely available to further data-driven, risk-based management of DBP formation in drinking water supplies globally.