Real-time wave model error correction via coupled neural networks and WAM under extreme weather

Accurate forecasts of wave parameters, especially significant wave height, are essential for maritime operations, yet predicting wave heights during extreme weather remains difficult due to rapid error growth in numerical models. This study presents a real-time error correction framework that couples a spatiotemporal attention-based neural network with the WAM wave model. The correction network is trained using CFOSAT satellite observations and dynamically coupled with WAM via a Fortran–Python interface. Applied to 114 typhoon events in the Northwest Pacific, the system reduces significant wave height (SWH) root mean square error (RMSE) by 24.6 % and increases the structural similarity index (SSIM) by 26.3 %, compared to WAM predictions made with default tuning parameters. Validation across 32 tropical cyclones with diverse intensities in the Gulf of Mexico shows strong generalization, achieving up to a 47 % reduction in RMSE and enhancing wave spectral accuracy by >30 %. These results highlight the robustness and scalability of this hybrid AI-physics framework, demonstrating its practical value for real-time wave forecasting during extreme weather events.