DELUGE Predicts Continental-Scale Daily Pluvial Flood Damage.

Pluvial flooding, caused by heavy rainfall, is a major source of flood damage and is notoriously difficult to predict at high resolution across large geographical areas. Existing models are often limited by scale, resolution, or computational intensity, making them unsuitable for daily, continental-scale use. The DELUGE framework addresses these limitations with a multimodal deep learning approach. It predicts daily pluvial flood damage at approximately 1 km resolution across the continental United States, focusing on the highest-claim areas. The model is trained on historical flood insurance claims and incorporates hazard, exposure, and vulnerability factors. A key innovation is its interpretable conditioning scheme, using parametric modules conditioned on terrain descriptors and AlphaEarth foundation-model embeddings. This design provides architecture-level interpretability, allowing direct inspection of hydrological response parameters. DELUGE significantly outperforms traditional machine learning baselines, especially for rare, high-cost claims.