While about one million ocean observations are available daily, only 300 daily profiles actually inform on the ocean interior. To understand fine-scale processes (1 to 100km, days to weeks) is also a challenge. Both are sparsely observed, predominantly during dedicated campaigns at sea or by autonomous platforms. Since 2023 the SWOT (Surface Water and Ocean Topography) satellite mission has been providing unprecedented high resolution altimetry fields. We focus on a specific area: the Western Mediterranean, where the BioSWOT-Med campaign was conducted in spring 2023 - during the SWOT fast sampling phase. Many in situ measurements were taken during this period and can be exploited by a generative model.

Our objective is twofold. First, we aim to improve our understanding of the physics involved during the campaign, and obtain the most precise state estimation as possible at this period and location. Second, we want to generalize the model to other situations where only satellite data (and some sparse in situ data) are available, and where information on the upper ocean would be useful.

In this perspective, we build a generative artificial intelligence method for surface variables. We aim at reconstructing a 3D state (Temperature, Salinity, horizontal Velocities) mostly from horizontal satellite observations and sparse vertical profiles. In order to learn how to map simulated observations to physically consistent 3D states, the neural network (NN) is trained on the Mediterranean Sea Physics Analysis and Forecast products from CMEMS. Satellite data and Argo profiles are then used to guide the inference process, while other observations are kept for validation purposes. On a longer term perspective, the NN should reconstruct daily states.

The first results will be presented and discussed in comparison with other findings from both in situ and numerical studies of fine-scale dynamics.