Advancing Building-Level Flood Damage Assessment Beyond Conventional Depth-Damage Functions with Physics-Constrained Deep Learning

Conventional depth-damage functions estimate residential flood losses by treating homes within broad occupancy classes as interchangeable, ignoring how size and internal layout influence damage outcomes. We develop a physics-constrained neural network, trained on component-level Monte Carlo simulations of 471 residential layouts, that predicts damage for individual homes using flood depth and building attributes (floor area, stories, bedrooms, bathrooms, and garage spaces). The outputs include the mean and standard deviation of percentage damage while preventing implausible decreases in damage as water rises. On unseen homes, the model achieves $R^2=0.945$ for mean damage and $R^2=0.854$ for damage variability, consistent with trends in established depth-damage functions. Crucially, at 10 ft of inundation, predicted damage varies by roughly 26.5 percentage points across layouts, demonstrating that flood depth alone does not determine losses. Our framework successfully recovers this building-specific variation that traditional curves collapse, providing a physically reliable foundation to guide disaster response, insurance, climate adaptation, and resilience planning.