Date of Award

Spring 2026

Abstract

Flood damage to residential buildings is commonly estimated using depth-damage functions that treat buildings as interchangeable, ignoring differences in size and layout that influence damage outcomes. In this work, we present a physics-informed neural network that incorporates building characteristics (floor area, bedrooms, bathrooms, stories and garage spaces) and reduces physically implausible predictions in which damage decreases as flood depth rises. The model predicts both the mean and uncertainty of percent damage and is tuned to balance physical consistency and predictive accuracy. Model interpretation shows that building characteristics collectively account for nearly half of predicted damage variability, highlighting the importance of building-level information beyond flood depth alone. We demonstrate its utility through a case study of Hurricane Helene’s impact on Asheville, NC, the most severely affected city in the US, where no prior building level estimates are available. Using these estimates, we examine socioeconomic disparities in flood impacts linked with housing cost burden, evaluate resilience under varying flood severities, and assess the cost-effectiveness of home elevation as a mitigation strategy. More broadly, this framework offers a scalable, physically reliable approach to flood damage assessment to better inform disaster response, insurance, and resilience planning.

Document Type

Master's Thesis

First Advisor

Fei Han

Second Advisor

Cuihong Song

Degree Name

Master of Science

Download

Share

COinS