OceanFlow: a flux-matching model produces better-calibrated sets of ocean forecasts

Rather than perturbing a single forecast, OceanFlow directly learns the distribution of future ocean states around a stable mean trajectory.

AI models can now predict the global state of the ocean in a deterministic way with good accuracy, but probabilistic forecasting, which is essential for maritime operations or climate analysis, remains in the background. Ensembles obtained by disturbing the initial conditions create dispersion without the joint distribution of members being optimized, which often results in poorly calibrated uncertainties and physically inconsistent states.

An average anchor, then a learned uncertainty

Yuan Gao, Fan Xu, Hao Wu and Xiaomeng Huang propose OceanFlow, which proceeds in two steps: the model first establishes a net and stable average trajectory, then a flow matching mechanism shapes the distribution of deviations around this anchor at each time step, as a function of the current state of the ocean. A factored dynamics operator parameterizes both the anchor and the velocity of the residual flow.

According to the authors, the approach improves ensemble mean accuracy, probabilistic reliability, and physical consistency diagnostics compared to reference methods. The code is published in open access.