Accurately simulating and predicting the marine environment requires a thorough understanding of oceanic processes occurring across a range of spatial scales. While oceanic global circulation models (OGCMs) are designed to represent large-scale ocean ...
Accurately simulating and predicting the marine environment requires a thorough understanding of oceanic processes occurring across a range of spatial scales. While oceanic global circulation models (OGCMs) are designed to represent large-scale ocean phenomena, they often lack the spatial resolution necessary to capture smaller-scale ocean dynamics. To address this limitation, dynamical downscaling techniques have been introduced to improve the representation of regional-scale oceanographic processes. By constructing regional ocean models, these approaches enable more detailed simulations of fine-scale dynamics. However, significant errors often arise during the downscaling process because small-scale oceanic structures are insufficiently included in the input data derived from OGCMs.
This study aims to overcome these limitations by developing a novel dynamical downscaling framework capable of effectively reproducing small- and intermediate-scale ocean dynamics in coastal regions. First, to assess the capabilities of traditional downscaling methods, a regional ocean model was implemented for the East Sea, and its performance was evaluated. The results revealed that conventional methods alone could not adequately reconstruct small-scale oceanic structures within a short simulation period, especially in regions with weak dynamic forcing. Even over longer simulation periods, the representation of submesoscale structures remained limited, and the kinetic energy at small and intermediate scales was significantly underestimated.
To address these challenges, a spectra-informed modeling framework, named BSEA Ⅱ, was developed. This method enables the three-dimensional reconstruction of small-scale features inherently absent from low-resolution input data. The BSEA Ⅱ model showed considerable improvements in restoring small-scale kinetic energy, particularly in cases with large spatial resolution gaps between the input and target grids. However, it remained insufficient for fully capturing submesoscale dynamics, which are essential in coastal environments.
To further enhance model performance, a machine learning-based approach termed submesoscale-informed modeling was developed. This technique uses low-resolution OGCM outputs as input and employs convolutional neural networks (CNNs) to reconstruct submesoscale ocean structures. It also predicts vertical velocity fields, which are not provided by OGCMs but are critical outputs of regional models. Incorporating mixed layer depth (MLD) as an additional input improved the accuracy of vertical velocity predictions, highlighting the value of including physically relevant variables. Moreover, this ML-based approach significantly reduced computational costs compared to traditional regional ocean simulations.
In conclusion, the integration of spectra-informed and submesoscale-informed modeling approaches demonstrated both high accuracy and computational efficiency in reproducing coastal ocean dynamics. These findings highlight the potential of combining machine learning with physics-based modeling to substantially improve regional ocean simulations. Furthermore, the reduced computational demand of the proposed framework suggests strong applicability for real-time or short-term ocean prediction tasks, such as marine pollutant tracking and vessel navigation.