Multi-Modal Eddy-Resolving Surface Ocean State Estimation Using Generative Data Assimilation

Ocean eddy dynamics are observed using satellite altimetry-derived surface geostrophic currents. However, the sparsity of altimetry observations limits the resolution of gridded sea surface height (SSH) products, causing significant biases in eddy dynamics. Our new AI method, NeurOST, improved global SSH resolution by up to 30% by synthesizing SSH and sea surface temperature (SST) observations, proving the value of synergizing satellite observables using AI. While NeurOST enhances eddy dynamics observation, fine-scale features are smoothed due to the sparsity of real-world training data and NeurOST reconstructs only SSH, omitting key components of the dynamical state such as SST, salinity, and ageostrophic currents.

Building on NeurOST, we here propose an AI method which uses training data from high-resolution simulations to estimate the multi-modal surface ocean state from satellite observables, jointly reconstructing SSH, SST, salinity, and ageostrophic currents from sparse SSH/SST observations and low-resolution SSH, SST, and salinity gridded products. Prior studies of this problem used supervised learning, predicting target variables using pseudo-observations from the simulation during training. However, these methods face a domain gap when applied to real observations, requiring real-world fine-tuning which is challenging for variables not observed by satellites, like ageostrophic currents.

To address this gap, we explore a generative AI approach, namely score-based data assimilation (SDA). We train a score-based diffusion model to generate realistic ocean states resembling the simulation, before guiding its generation using real observations with no re-training of the score network. SDA ensures the state estimates ‘look like’ the simulation, enabling physically realistic joint estimation of multiple ocean variables, while matching observations better than state-of-the-art dynamical data assimilation. We compare the generative and supervised learning strategies through regional experiments in the Gulf Stream using simulated and real-world satellite observations.