Physics‐Guided Deep Learning for Modeling Single‐Point Wave Spectra Using Wind Inputs of Two Resolutions

Abstract Directional Wave Spectra (DWSs) are essential for various applications such as seafaring and ocean engineering. Traditionally, DWSs are modeled with numerical wave models, which, despite their solid physical basis, are often computationally expensive. For any given point in the ocean, once the recent prior wind fields that could affect this point are known, the current DWS at this location can be largely determined. This is because local wave energy is generated by either local winds or recent prior remote winds. This fundamental physical principle can be leveraged for the statistical modeling of single‐point DWSs. This study presents a deep learning approach to directly model single‐point DWS using wind inputs. We utilize wind fields with two different spatial resolutions: specifically, a high‐resolution, small‐area, short‐time local wind field for local wind‐sea spectra, and a low‐resolution, large‐area, long‐time wind field for remotely generated swell spectra. A tailored network architecture and loss function are developed to capture the relevant information from both wind scales. The performance of the proposed model was evaluated using both open ocean spectra from ERA5 and coastal spectra from numerical wave hindcasts. Results indicate that the deep learning model can effectively and efficiently simulate DWSs in both open oceans and coastal regions without relying on predefined spectral shapes, demonstrating its potential for applications in wave forecasting and wave climate studies.