Urban flooding causes billions in damage annually and threatens lives; city planners urgently need fast
“what-if” tools to evaluate flooding under hypothetical rainfall scenarios. Physics-based hydraulic
simulators (e.g., MIKE FLOOD, SWMM) can take hours per event, making real-time scenario analysis
impractical. This project develops a neural surrogate that learns to approximate coupled 1D (underground
drainage) and 2D (surface flow) hydraulic simulations and can generate flood predictions in seconds.
This work intends to find answers for three interconnected research questions:
RQ1 – Generalization across rainfall scenarios: Can a graph neural network trained on historical
rainfall events accurately simulate water levels for unseen events – both within the training intensity range
and for intensities bracketed by training values (e.g., train on 2, 4, 6-inch events; test on 3 and 5-inch)?
RQ2 – Local vs. global mass conservation: Does enforcing local (per-node) mass conservation as a
physics-informed training constraint improve prediction accuracy and reduce autoregressive error
accumulation, compared to the global domain-wide conservation used in the baseline?
RQ3 – 1D drainage network integration: Does explicitly incorporating 1D underground drainage nodes
(manholes, junctions) alongside 2D surface mesh cells improve flood predictions, and does this interact
with whether local conservation is enforced at the 1D–2D coupling interfaces?