Neural Operators for Continuous Bias Correction of Water Level Forecast Guidance

Abstract The National Oceanic and Atmospheric Administration's (NOAA) Global Two‐Dimensional Surge and Tide Operational Forecast System (STOFS‐2D‐Global) provides global operational tidal, subtidal, and total water level forecast guidance with a 7.5‐day horizon. The system uses the ADvanced CIRCulation (ADCIRC) finite element model, with a large unstructured grid of approximately 12.7 million nodes and 24.9 million elements for STOFS‐2D‐Global version 2.1. Despite being one of the most accurate nondata assimilated global tide models, a difference has been observed between its simulated total water level and the observed one. To address this issue, we introduce NeurOCAST, a neural operator‐based deep learning model designed to correct biases in water level forecast guidance. The key strength of NeurOCAST is its ability to learn the underlying function of spatiotemporal outputs from a limited number of points and generalize it to gridded outputs at different resolutions. Our finding indicates that NeurOCAST improves water level accuracy by a significant reduction in bias at observational locations and times not included in the training data. NeurOCAST enhances forecast skill over the entire forecast horizon. To the best of our knowledge, this study represents the first application of a neural operator‐based model for improving oceanic forecast guidance using limited observational data. These advancements strengthen and support coastal resilience and safe navigation by providing more accurate prediction of water levels which in turn improve disaster preparedness and infrastructure planning and support informed decision‐making globally.