A physics-based ocean ensemble forecast works by running the same simulation many times with slightly different initial conditions. Each run is expensive — the same fluid dynamics, the same thermodynamics, the same grid, solved again. Twenty ensemble members means twenty times the computation. The ensemble is buying uncertainty estimates at the cost of redundant physics.
FuXi-ONS (arXiv:2603.19591) is the first machine-learning ensemble system for global oceans: 5-day forecasts on a 1° grid, extending to 365 days for sea-surface temperature, sea-surface height, subsurface temperature, salinity, and currents. The key difference from physics-based ensembles: it learns structured perturbations rather than adding random noise to initial conditions.
“Structured” is doing the work. Random perturbations grow in physically arbitrary ways — a small error in temperature might amplify where the flow is unstable, but the perturbation doesn't know that in advance. Learned perturbations are trained to grow in physically meaningful directions — along the unstable manifold of the ocean dynamics, where uncertainty naturally concentrates. The perturbation is informed by the dynamics it perturbs.
An atmospheric encoding module stabilizes long-range forecasts by coupling ocean state to atmospheric forcing — a physical constraint that prevents the ocean model from drifting independently. The result: orders of magnitude faster than conventional ensembles, competitive with seasonal forecast benchmarks for Niño3.4 variability and SST.
The insight: ensemble forecasting is uncertainty sampling. Physics-based methods sample by brute force. Learned methods sample by structure. The structure is the physics, compressed.