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

Area of Science:

• Environmental Chemistry
• Water Quality Management
• Computational Science

Background:

• Disinfection byproducts (DBPs) form from reactions between disinfectants and dissolved organic matter (DOM) in drinking water.
• Harmful concentrations of DBPs pose risks to public health.
• Predicting DBP formation is crucial for effective water treatment and risk management.

Purpose of the Study:

• To explore the application of machine learning (ML) models for predicting the formation of regulated and emerging DBPs.
• To assess the utility of DOM spectroscopic variables and hydrochemical parameters in ML model training.
• To develop a scalable workflow for data-driven DBP management in drinking water.

Main Methods:

• Laboratory chlorination experiments were conducted on filtered raw water samples.
• DOM spectroscopic variables and hydrochemical parameters were used to train ML models (neural networks, bagging trees, boosted regression, SVM).
• Model performance was evaluated for quantitative prediction and binary presence-absence classification of ten DBP species.

Main Results:

• ML models accurately predicted ten DBP concentrations (average R² = 0.86) and classified five additional species (95.6% accuracy).
• DOM spectroscopic variables, particularly humic-like fluorophores and UV-Vis absorbance, were most influential.
• Inclusion of hydrochemical parameters like dissolved organic carbon (DOC) showed only marginal improvement.

Conclusions:

• Machine learning models trained on DOM spectroscopic data offer a powerful tool for predicting DBP formation.
• This approach provides a proof-of-concept for scalable, data-driven management of DBPs in drinking water.
• The findings support risk-based management strategies for ensuring global drinking water safety.