A global ocean emulator was built to operate directly on the unstructured mesh of the FESOM2 model, trained on a 209-year control integration from AWI-CM3. It runs without atmospheric forcing, only receiving the atmospheric state at initialization, which isolates the predictability coming from the ocean state itself.
The emulator's forecast skill depends on the variable: it outperforms all reference methods for currents at 30-day lead time, but for temperature and salinity a damped-anomaly persistence forecast is still more accurate. When trained on reanalysis data, a variant of the emulator achieves the lowest RMSE against the GLORYS reanalysis in the OceanBench benchmark, confirming that the native-mesh approach is competitive.