Watershed-scale hydrodynamic modeling of Hurricane Helene flooding in Southern Appalachians

Study region: Southern Appalachian headwaters of western North Carolina and eastern Tennessee, USA Study focus: This study develops and applies the XLA-accelerated Water Model (XWM), a differentiable, well-balanced shallow-water solver implemented in JAX (a Python library), to hindcast Helene’s flooding at 10 m resolution over 20,208 km 2 (202 million cells). XWM couples rainfall–runoff and open-channel hydraulics on a single grid and uses a mixed-precision strategy (64-bit accumulators, 32-bit kernels) to run efficiently on GPUs. The (uncalibrated) model performance is evaluated against 30 United States Geological Survey (USGS) gauges, 2375 U.S. Army Corps of Engineers (USACE) high-water marks, and post-event National Oceanic and Atmospheric Administration (NOAA) and Civil Air Patrol (CAP) aerial imagery. The event inundation is compared with the Federal Emergency Management Agency (FEMA) National Flood Hazard Layer (NFHL). New hydrologic Insights: XWM (uncalibrated) reproduces hydrograph timing and shape across the domain (Pearson correlation coefficient r > 0 . 8 ) and matches surveyed high-water marks within ± 1 m (MAE of 0.71 m). Compared to FEMA’s NFHL, XWM predicts 40% more overbank inundation on low-order tributaries, a finding corroborated by post-event NOAA and CAP aerial imagery. These findings highlight hidden headwater flood exposure in steep Appalachian catchments and support integrating watershed-scale, meter-resolution hydraulics into regional risk assessments and emergency planning. • XLA-accelerated water model (XWM) for watershed-scale flood simulations • Mixed-precision FP32/FP64 yields 12x speed-up without numerical instability • Simulated 202M cells (20,208 km 2 ) at 10 m res in 23 h on H100 • Modeled USGS hydrographs (r > 0.8) and 2375 HWMs (MAE 0.71 m) match observations • XWM predicts 40% more overbank flooding on low-order tributaries vs FEMA NFHL