A high-fidelity framework for dynamical downscaling with a submesoscale-informed neural network

Global-to-regional dynamical downscaling provides a practical means of increasing resolution from global ocean models to regional ocean circulation models (ORCMs); however, submesoscale dynamics are often not fully resolved, particularly in simulating coastal seas shaped by stratification and circulation. To overcome this gap, we introduce a physics-guided, submesoscale-informed modeling framework that combines deep learning with dynamical downscaling. Using CNN-UNet architectures, the framework consists of a first neural network that reconstructs intermediate-scale horizontal velocity from coarse-scale inputs and a second neural network that predicts surface vertical velocity from high-resolution inputs. The first neural network successfully recovers intermediate-scale dynamics with high fidelity, reproducing spectral energy distributions comparable to the original ones and Rossby number patterns, while capturing their seasonal modulation. The second neural network demonstrates the feasibility of surface vertical velocity prediction, yielding predictions that closely agree with high-resolution reference fields in both spatial structures and statistical metrics. Importantly, while conventional ORCM simulations on 16 CPU cores require nearly 40 hours for a single year of integration, the trained neural network generates comparable high-resolution fields in under 10 minutes on a single GPU, without the need for long dynamical integrations. These results establish submesoscale-informed modeling as a fast, accurate, and computationally efficient complement to conventional dynamical downscaling, offering strong potential for improving regional ocean prediction and coupled forecasting applications.