Chlorine transport and decay analysis in water distribution systems using hydraulic simulation and physics- and chemistry-informed graph neural networks

Maintaining compliant chlorine residuals across water distribution networks is challenging because concentrations evolve under time-varying hydraulics, advective transport, and bulk and wall decay kinetics, while field monitoring is sparse. Physics-based simulators such as EPANET can represent these processes but require careful calibration and can be difficult to use for rapid, network-wide decision support, while purely data-driven models may violate basic transport and reaction constraints and often need dense measurements. This study develops a physics- and chemistry-informed graph neural network (GNN) that learns network-wide chlorine dynamics from hydraulic simulations while embedding domain features that represent connectivity, flow behavior, water age, and decay parameters. Two GNN architectures—a Graph Convolutional Network (GCN) and a Gated Graph Neural Network (GGNN)—are trained on features including pressure, demand, chlorine concentration, flow velocity, headloss, and decay coefficient that capture the system’s hydraulic and chemical behavior. The approach is demonstrated on a subsection of Ward 4 in Reno, Nevada using a 240-hour EPANET extended-period simulation dataset. Compared with a graph convolutional network baseline, the proposed GGNN improves predictive accuracy, reducing the mean absolute error from about 0.38 to 0.22 mg/L and the root mean square error from about 0.57 to 0.35 mg/L, while increasing coefficient of determination (R 2 ) from about 0.93 to about 0.97. At a representative snapshot, results show that the GGNN significantly outperforms the GCN, achieving lower error and higher R 2 (>0.99), due to its iterative message-passing mechanism that captures long-range dependencies. Both models generalize well across hydraulic conditions, enabling accurate virtual sensing for real-time chlorine monitoring in sparsely instrumented networks. • The study introduces a hybrid modeling framework that integrates hydraulic simulations with graph neural networks (GNNs) to predict chlorine residuals across water distribution networks (WDNs). • Leveraging the natural graph-like structure of WDNs, the GNNs developed in this study naturally mirror the pipe-and-junction topology of distribution systems and offer a unique advantage by enabling message-passing between connected nodes. • The approach captures spatial dependencies and flow-related interactions more effectively than conventional machine learning models. • The manuscript addresses key challenges in chlorine monitoring, offering a robust alternative to physical sensor deployment and traditional hydraulic models and data-driven approaches. By embedding both hydraulic and chemical principles into a machine learning architecture, this approach offers a scalable and promising tool for real-time water quality management and anomaly detection, especially in regions with limited sensor coverage. • The framework developed can be generalized across systems and scales. Beyond water distribution systems, the findings of this study can aid troubleshooting in chemical processes that involve reaction transport phenomena.