Deep learning for global three-dimensional ocean modeling with physical response consistency

With the advent of the AI era, deep learning has been actively applied to global weather prediction, achieving remarkable progress. Furthermore, these deep learning-based global prediction models are being utilized for seasonal forecasting, with efforts underway to extend the forecast lead time. Leveraging the memory effect of the ocean is essential for such advancements. In this study, we developed a deep learning-based global three-dimensional ocean model, incorporating three key innovations: (1) expanding the receptive field and reducing the number of parameters using a visual attention network, (2) eliminating ocean/land boundary effects through the application of partial convolution, and (3) aligning prediction value distributions with observations using adversarial loss. Compared to persistence forecasts and NMME models, our model demonstrated global three-dimensional ocean simulation capabilities comparable to state-of-the-art coupled general circulation models, achieving significant improvements, particularly in predicting horizontal ocean currents. Furthermore, the model realistically simulated the ocean’s response to surface boundary forcing. These results highlight the potential for developing a deep learning-based ocean-atmosphere coupled general circulation model.