Operator-Based Machine Learning Framework for Generalizable Prediction of Unsteady Treatment Dynamics in Stormwater Infrastructure

Stormwater infrastructure experiences uncontrolled and unsteady hydraulic and pollutant loadings driven by episodic rainfall-runoff events. Evaluating its in situ treatment dynamics and performance remains a fundamental challenge for cost-effective system design, planning, and management. Computational fluid dynamics (CFD) offers robust predictive capabilities of these systems by resolving the underlying turbulent transport and pollutant fate physics; however, its high computational cost constrains its application, particularly over extended periods. To address this limitation, this study developed a composite neural network (CPNN) framework to predict the three-dimensional (3D), transient dynamics of stormwater infrastructure under unsteady hydraulic and particulate matter (PM) loading. This framework was evaluated on a common urban stormwater treatment system, a hydrodynamic separator (HS). A total of 496 CFD simulations of the HS under various storm events were used for training, 62 for validation, and 62 for test. The results demonstrate a high predictive capability of the proposed CPNN framework relative to CFD, achieving an R2 score above 0.8 in 95.2% of test cases for hydrodynamic predictions. For PM concentration predictions, the R2 score exceeds 0.8 in 72.6% of test cases and remains between 0.4 and 0.8 in 22.6%. The study highlights challenges in learning spatial-temporal dynamics under complex loading conditions spanning multiple orders of magnitude. Furthermore, CPNN’s automatic differentiation capability is leveraged to assess the sensitivity of PM transport to loading conditions. The potential of CPNN for evaluating stormwater infrastructure performance over long-term periods is discussed. The framework is transferable to other water systems and represents a critical step toward enabling robust and climate-aware stormwater infrastructure planning and management.