Search for a command to run...
Operational coastal flood forecasting confronts a fundamental “scale gap’: physics-based numerical models provide process fidelity but require computational resources incompatible with real-time decision-making, while machine learning approaches offer speed but lack the spatial physics necessary for inundation mapping. I present a probabilistic ML-physics hybrid framework that bridges this gap by coupling machine learning-derived temporal boundary conditions with a diffusive wave connectivity solver for spatial flood propagation. Validated against 11 official NOAA tide gauge records from Hurricane Isabel (2003) across three Chesapeake Bay sites—Annapolis, Baltimore, and Norfolk-Virginia Beach—my framework achieves a root-mean-square error (RMSE) of 0.015 m, representing a 96\% improvement over the NOAA Storm Surge Total Water Level and Coastal Flooding (STOFS-2D) operational baseline (RMSE = 0.35 m). The physics-based hydraulic connectivity algorithm successfully resolves flood pathways across 1,210 km$^2$ of vulnerable land in Norfolk and 36.8 km$^2$ in Annapolis, eliminating the unrealistic “bathtub’ artifacts that plague simplified approaches. Benchmark comparisons against LISFLOOD-FP demonstrate equivalent flood extent accuracy (Jaccard index 0.85) with 10$\times$ faster runtime, enabling real-time operational deployment. A Monte Carlo uncertainty quantification engine ($n=30$ ensemble members) propagates boundary condition uncertainty through the spatial domain, enabling probabilistic risk assessment. Under climate change scenarios, I demonstrate pronounced nonlinear risk acceleration: a +1.0 m sea level rise amplifies the Norfolk flood risk zone by a factor of 4.8$\times$ (from 1,210 km$^2$ to 5,854 km$^2$). The Python-native implementation achieves 16-second runtime for 42 km$^2$ domains and 139-second runtime for 251 km$^2$ domains, enabling deployment for real-time emergency management and long-term resilience planning.