Understanding, modeling, forecasting and reconstructing fine-scale and large-scale processes and their interactions are among the key scientific challenges in ocean-atmosphere science. Artificial Intelligence (AI) technologies and models open new paradigms to address poorly-resolved or poorly-observed processes in ocean-atmosphere science from the in-depth exploration of available observation and simulation big data. In this context, this proposal aims to bridge the physical model-driven paradigm underlying ocean & atmosphere science and AI paradigms with a view to developing geophysically-sound learning-based and data-driven representations of geophysical flows accounting for their key features (e.g., chaos, extremes, high-dimensionality). We specifically address three key methodological questions: (i) How to learn physically-sound representations of geophysical flows? (ii) Which learning paradigms for the representation of geophysical extremes? (iii) how to learn computationally-efficient representations and algorithms for data assimilation?. Upper ocean dynamics will provide the scientifically-sound sandbox for evaluating and demonstrating the relevance of these learning-based paradigms to address model-to-observation and/or sampling gaps for the modeling, forecasting and reconstruction of imperfectly or unobserved geophysical random flows. To implement these objectives, we gather a transdisciplinary expertise in Numerical Methods (INRIA GRA & Rennes), Applied Statistics (IMT, LSCE), Artificial Intelligence (IMT, LIP6) and Ocean and Atmosphere Science (IGE, INRIA GRA, LOPS), complemented by the participation of two SMEs (Ocean Data Lab and Ocean Next) to anticipate the added value of AI technologies in future earth observation missions and coupled observation-simulation systems.