Projecting Atlantic Storm Surge Risk with DeepSurge, a New Deep-Learning Storm Surge Model

We present DeepSurge, a flexible and efficient neural network approach to modeling tropical cyclone (TC) induced storm surge in the Atlantic basin. The model is trained on output from high-resolution ADCIRC simulations, and validated using historical tide gauge observations collected from the National Oceanic and Atmospheric Administration (NOAA). We find that DeepSurge achieves excellent skill in identifying peak surge levels as measured by the observations, with its predictions achieving a mean absolute error of 0.45 meters and a Pearson's correlation coefficient of 0.67. We use the model to assess the future evolution of North Atlantic storm surge risk in a changing climate by applying the model to TCs generated by the HighResMIP project in modeled historical and future climate periods under the SSP5-8.5 scenario. Our model suggests a substantial increase in future surge risk for the Mississippi River Delta region in the vicinity of New Orleans, with up to 100% increases in the frequency of some strong events. Our model also predicts increasing risk in southern Florida, western Cuba, the Chesapeake Bay, and southern Massachusetts. Conversely, somewhat lower surge risk may be expected in Texas, much of the Carolinas, and the majority of the Caribbean islands, among other regions, though the magnitude of these decreases is often comparatively smaller than the increases seen in other regions. We put these findings into context, and contrast DeepSurge's skill and projections with related methods.