Deep Learning to Infer Ocean Dynamics

Abstract

Inferring ocean dynamics remains challenging due to the complexity of multiscale nonlinear interactions, sparsity of observations, and limitations of numerical models. Deep learning (DL) has emerged as a powerful data-driven framework that exploits nonlinear dependencies across heterogeneous datasets, complementing traditional approaches. Here, we review recent progress in applying DL to ocean dynamics, including spatiotemporal interpolation of satellite and in situ data, estimation of unobserved ocean variables, neural data assimilation, ocean forecasting, and the development of eddy parameterizations for hybrid models. DL methods have excelled at fusing multimodal observations, reconstructing multiscale fields, learning complex distributions, and providing computationally efficient predictions that often match or exceed those of conventional statistical approaches. Despite the rapid methodological progress, opportunities remain in uncertainty quantification, ensuring physical consistency, generalization to unseen conditions, and leveraging DL beyond obtaining operational gains to advance scientific discovery and human-level understanding of ocean dynamics.