Physics-Informed Neural Network Stormwater Modeling for Surface-Subsurface Flow in Stormwater Infiltration Systems

Physics-Informed Neural Networks (PINNs) represent a transformative paradigm for hydrological modeling. However, their application to stormwater control measures remains unexplored. Industry-standard models often rely on lumped approximations and static parameterization of within-SCM hydrological processes, causing spatial blindness and failing to capture complex real-world variabilities. We developed SWE-GA-PINN, a framework coupling the shallow water equations with the Green-Ampt equation, representing a pioneering attempt to integrate surface and subsurface water dynamics within a PINN for improved SCMs simulation. We first constructed a baseline model that replicates industry-standard accuracy with low relative errors (2.30% for , 6.74% for , and 4.21% for overflow). We then deployed a high-fidelity model instance of the framework to overcome spatial blindness and static parameter rigidity through distributed hydrodynamic reconstruction and dynamic mechanism discovery. Specifically, the model reconstructs and fields from sparse sensors with high accuracy (). Furthermore, the model demonstrates a unique ability to discover latent, unknown physical patterns, the spatiotemporal decay of hydraulic conductivity , without explicit physical constraints. This enables unsupervised subsurface infiltration () inference solely from surface observations , yielding a 7.59% relative error () without training data beyond the inlet. To evaluate environmental robustness against sensor interference, a severe 50% Gaussian noise stress-test demonstrates that relative errors for remain restricted to 9.21%. By enhancing within-SCM hydrological processes simulation accuracy, this framework improves prediction reliability under climatic extremes and reduces catchment-scale cumulative errors, serving as a high-fidelity foundational unit for hydrological digital twins to optimize stormwater management.