From data to currents: generative models for short-term forecasting of oceanic currents

We investigate whether short-term forecasts of ocean surface currents can be learned from two-dimensional observations alone, without access to the full three-dimensional ocean state. Using daily surface currents from the GLORYS12 reanalysis and 10-metre winds from ERA5 as the only inputs, we train and compare three neural architectures -a U-Net, a Swin Transformer, and a flow matching generative model -on a regional domain covering the Agulhas current south of Africa. All three models produce 10-day forecasts with similar accuracy, reaching a root mean square error of approximately 16 cm/s at day 10. This suggests that forecast skill is limited by the intrinsic predictability of mesoscale ocean dynamics at this resolution rather than by the choice of architecture. The flow matching model additionally produces an ensemble of physically plausible forecast trajectories, from which spatially coherent uncertainty estimates can be derived at no extra modelling cost. We further show that inference with a single Euler step and a lookback of only 2 days is sufficient for near-optimal performance, making the approach well suited for operational deployment. These results support the feasibility of lightweight, uncertainty-aware ocean current forecasting based on surface observations only.