Fast oceanic flow prediction using lattice- and morphology-informed approaches

This study investigates the potential of two advanced modeling techniques to enhance the predictability of circulations in the regional ocean. One technique is the Boundary Small Eddy Addition technique II (BSEA II), a random field generation method to simulate three-dimensional small-scale fluctuations of currents informed by the lattice grid. The other is a technique that employs a Deep Learning approach informed by the morphologic surroundings to model intermediate-scale circulations. The morphology-informed approach utilizes convolutional neural networks (CNNs) to predict intermediate-scale coherent features based on large-scale structures provided by the Ocean Global Circulation Models. Experimental results demonstrate that the BSEA II technique effectively reproduces small-scale features; however, it struggles to accurately resolve intermediate-scale structures (10–20 km) in complex coastal regions. A specific part of the implemented algorithm for morphology-informed modeling is the architecture of using a CNN-based UNet, showing promise in effectively capturing intermediate-scale energy spectra. By integrating two proposed approaches, this study provides a framework for bridging the resolution gap in ocean modeling enabling faster and more efficient predictions for ocean dynamics. Nevertheless, while the model performs well during the first half of a simulation year (January to June), somewhat noticeable errors emerge in the second half as submesoscale activity intensifies, highlighting its limitations in addressing seasonal variability and complex dynamics. This indicates that more information should be implemented into the training of the morphology-informed model, specifically.