A machine learning-driven framework for optimizing disinfection in drinking water treatment

Abstract

Drinking water disinfection effectively prevents waterborne disease outbreaks but inevitably results in the formation of carcinogenic disinfection by-products (DBPs). Mitigating DBP formation without compromising microbial safety remains a critical challenge. Here we developed DISoptimizer, a water-quality-adaptive optimization framework designed to determine the optimal chlorine dose across heterogeneous water matrices, thereby maintaining a user-defined residual chlorine target as an operational surrogate for disinfection reliability while minimizing four trihalomethane (THM4) formation. Using machine learning models trained on experimental datasets covering diverse water quality conditions, DISoptimizer predicts residual chlorine and THM4 formation after 24 h, representing the point of delivery. Crucially, DISoptimizer incorporates a weighting factor to encode user-defined preferences, allowing utilities to adjust the operational emphasis on maintaining residual chlorine margins versus mitigating THM4 formation in response to fluctuating water quality and different management priorities. Computational simulations coupled with external validation demonstrated that DISoptimizer achieved a 5–35% reduction in THM4 formation compared to the empirical fixed-dosing strategies, while consistently maintaining residual chlorine margins. This work advances disinfection management from DBP concentration prediction towards proactive chlorine-dosing decisions, supporting the management of health risks associated with microbial contamination and DBP exposure in drinking water. A machine learning-based framework is designed to adaptively optimize chlorine dosing across diverse water conditions to maintain disinfection reliability while reducing the formation of carcinogenic disinfection by-products.