A Deep Learning Framework for Extreme Storm Surge Modeling Under Future Climate Scenarios

Abstract

Abstract Coastal regions are increasingly exposed to sea-level rise and intensifying storm surges, underscoring the urgent need for accurate long-term predictions of extreme water levels to support robust adaptation planning. Physics-based hydrodynamic storm surge models remain the gold standard for such projections, but are computationally demanding, limiting their feasibility for producing the large scenario ensembles needed under deep uncertainty. Artificial intelligence surrogate models have emerged as a promising alternative. Yet, current approaches often underrepresent rare extremes and lack validation under future climate conditions, constraining their application for long-term planning. Here, we develop a deep learning surrogate model trained on hydrodynamic simulations from the Global Tide and Surge Model (GTSM), with both historical reanalysis and high-resolution climate projections (CMIP6 HighResMIP). Using New York City, a highly vulnerable urban coastline with extensive surge records, as a testbed, we demonstrate the model's ability to represent extreme storm surges under both historical and mid-21st-century scenarios. To enhance performance on extremes, we propose a novel asymmetric loss function, combining quantile and expectile losses, which substantially improves predictions of rare storm surge events, while maintaining high overall performance. Fine-tuning with climate model outputs further aligns the surrogate's estimates with those of the hydrodynamic model across spatial and temporal scales. Under future climate forcing, projections obtained with the surrogate model closely reproduce the response of GTSM, capturing projected trends in extreme events. This open-data-based framework provides a computationally efficient and globally transferable approach for storm surge projection, enabling the large-scale scenario analyses required for climate-resilient coastal planning.