From Model Development to Mitigation: Machine Learning for Predicting and Minimizing Iodinated Trihalomethanes in Water Treatment

Disinfection processes in water treatment produce disinfection byproducts (DBPs), such as iodinated trihalomethanes (I-THMs), which pose significant health risks. Mitigating I-THMs remains challenging due to the complex interactions among water quality parameters, disinfectants, and iodine sources, compounded by the difficulty of predicting their formation under varying treatment conditions. This study leverages a data set of 1534 samples from published studies to predict I-THM formation using machine learning (ML). Among five evaluated ensemble models, CatBoost Regression achieved the best performance. Incorporating domain-specific features (iodine/DOC and oxidant/DOC ratios) improved model accuracy and interpretability. Recursive feature elimination revealed that nearly half of the features could be excluded without compromising performance, simplifying model development and reducing experimental effort, an advantage often overlooked in prior research. Feature analysis identified key predictors and mitigation strategies, including minimizing iodine and bromide concentrations, reducing iodine/DOC, UV₂₅₄ and SUVA levels, and optimizing chlorine dose. The model further enabled rapid identification of the optimal chlorine dose to minimize I-THMs using incremental and Bayesian optimization. Achieving an <i>R</i>² of 0.67 on an external validation data set, the model demonstrated strong generalizability. This study establishes ML as a powerful tool for predicting and mitigating I-THMs, offering actionable strategies for safer drinking water treatment.