Forecasting Hurricane-Induced Storm Surge by Forcing STOFS-2D-Global with AI-Generated Nudged Winds

Two large sources of epistemic uncertainty when forecasting hurricane-induced storm surge are the hurricane’s track and intensity. Provided that other sources of uncertainty like mesh refinement, specified bathymetry / topography, and defined bottom and internal tide friction are managed, representing hurricane winds accurately for hydrodynamic model forcing is an important obstacle preventing consistently skillful storm surge forecasts. Assuming that operational forecast models have exceptional five-day storm tracking, their forecasted winds will likely be muted by virtue of their relatively coarse spatial resolution. In fact, most gridded operational forecast models tend to underpredict wind speeds near hurricane cores, and this leads to significant underpredictions of storm surge. These modeled winds are typically nudged to bring them in closer correspondence to reality. In this study, we employ UT-MeteoGAN, a generative adversarial network (GAN)- based model (a modified SRGAN - Super-Resolution GAN), to nudge modeled wind fields towards more accurate representations. Specifically, UT-MeteoGAN adjusts hindcasted wind fields from GraphCast to align more closely with National Hurricane Center (NHC) guidance by emulating the corresponding ERA5 wind structure and enforcing peak intensity. These adjusted wind fields are then used to force STOFS-2D-Global, NOAA’s operational storm surge and tide forecast system, with the goal of improving storm surge predictions. The approach is validated using high water marks provided by the USGS for several hurricanes in the Gulf of Mexico. Additionally, surge predictions using un-nudged ERA5 winds are compared to quantify the improvements achieved through our nudging strategy.