HClimRep-Ocean: an ocean simulator that operates on the irregular grid of physical models and outperforms the benchmarks for 30-day currents

Trained on 209 years of simulation data, it predicts ocean currents more accurately than any of the benchmark models, but not temperature or salinity, which depend on the atmosphere.

Machine learning emulators have transformed weather forecasting, but the ocean is still lagging behind: its eddies are about ten times smaller than their atmospheric equivalents, and coastlines, straits, and sea ice impose irregular meshes that AI models, built on latitude-longitude grids, were unable to exploit.

HClimRep-Ocean operates directly on the unstructured mesh of the FESOM2 ocean model. It has been trained on an AWI-CM3 control simulation covering 209 years and only receives the state of the atmosphere at the start, which makes it possible to isolate the predictability carried by the ocean itself.

A result that makes sense physically

For currents, which are largely geostrophic and generated internally, it surpasses all 30-day references. For surface temperature and salinity, which are controlled by the weather, a simple damped persistence of anomalies is more accurate. A variant trained on reanalysis also obtains the lowest error against GLORYS on the OceanBench test bench.