Development of a submesoscale-informed modeling for the fast ocean prediction

Among the multiple spatial and temporal scales of ocean circulation, submesoscale structures play a critical role in energy cascades and regional oceanic processes. Ocean General Circulation Models (OGCMs) effectively capture large-scale ocean dynamics but lack the resolution needed to resolve submesoscale processes. Although dynamic downscaling with Ocean Regional Circulation Models (ORCMs) improves spatial resolution, it often fails to accurately reconstruct intermediate-scale structures. To address these limitations, this study proposes a submesoscale-informed modeling framework based on a Convolutional Neural Network (CNN)-based UNet architecture to improve the representation of these features in downscaling applications.

The CNN-based UNet effectively reconstructs intermediate-scale features, showing strong correlation with high-resolution reference data from the ORCM between January and June 2021, and successfully capturing structures absent from OGCM outputs. It also improves spectral energy representation at scales of 10–20 km. However, prediction accuracy decreases from July to December 2021, likely due to increased submesoscale activity and seasonal variability. Model performance is highest in winter and lowest in summer, suggesting the potential benefits of incorporating additional physical constraints or seasonal tuning.

This study demonstrates that integrating deep learning with dynamic downscaling enhances the representation of submesoscale features in coastal ocean models. The proposed framework effectively addresses resolution gaps in OGCMs, and future work will focus on refining the model architecture and incorporating additional oceanographic variables to improve seasonal generalization. This approach offers a computationally efficient pathway toward high-resolution ocean forecasting.