Emulating Hydrodynamic Models with Neural Operators for Rapid Storm Assessment

Abstract

Timely, accurate, and trustworthy flood models for rapid decision making during and after extreme storms do not exist in an operational capacity, as flood models often compromise between physical accuracy and speed. Physics-based flood models that explicitly simulate mass and momentum balances can be accurate, but are prohibitively computationally expensive. Stochastic models built from synthesizing sparse amounts of existing flood data are timely, but do not model the underlying hydrodynamics of a flood, limiting simulations to water levels alone. Capturing the full impact of a flood disaster requires models that include strong currents that can harm people, flood momentum that damages infrastructure, and debris flows that can block water outflows and evacuation routes. To address this operational gap, we are developing a near-real-time "nowcast" flood disaster simulation fusing physics-based models with machine learning. As a part of this effort, we are creating training datasets using physics-based hydrodynamic models verified against recent historical flood events in the United States. These data will inform a timely flood model that simulates more than water levels: flow properties such as velocity can improve damage and risk calculations, as well as situational awareness of regions that could be hazardous during a flood. With this nowcast, we aim to build the timely, trustworthy foundation of tools that decision-makers can use to prepare for, respond to, and recover from flood disasters.