Short-term neural forecasts of ocean dynamics from sparseobservations

Machine Learning solutions for earth system modeling, monitoring and forecasting are growing rapidly. AI-based weather forecasts relying on end-to-end neural schemes [Bi et al., 2023, Lam et al., 2022] reach state-of-the-art performance and are among striking examples of this trend. Recent studies [Garcia et al., 2025, Botvynko et al., 2025, Beauchamp et al., 2025, Martin et al., 2025] support the potential of end-to-end Deep Learning schemes to improve the monitoring and forecasting of the ocean from satellite/in situ observations. In this study we focus on the stochastic extension of the previously developed framework for deterministic short-term neural ocean forecasting workflow [Botvynko et al., 2025]. We define the forecasting task as the training of the 4DVarNet variational neural assimilation scheme adapted to the forecasting of ensemble of ocean states from sparse observations. We present an evaluation framework, and benchmark ensemble 4DVarNet against state-of-the-art assimilation-based and neural forecasts. The results highlight the added value of ensemble formulation of the proposed end-to-end forecasting workflow when compared to its deterministic formulation.