HClimRep-Ocean delivers lowest RMSE on OceanBench by emulating FESOM2 directly on unstructured mesh

The model ingests the native FESOM2 mesh rather than remapping to latitude-longitude grids, preserving fine-scale eddies and complex coastlines that dominate ocean kinetic energy. It receives atmospheric state only at initialisation and then evolves freely, isolating oceanic predictability. This design yields field-specific skill: geostrophically balanced currents outperform persistence and prior emulators at 30 days, while temperature and salinity remain better captured by damped-anomaly persistence because they respond rapidly to atmospheric forcing.

Prior ML ocean models were restricted to regular grids, limiting fidelity near boundaries and in eddy-rich regions. HClimRep-Ocean’s native-mesh approach closes that gap and matches the mesh used in operational AWI-CM3 simulations, enabling direct coupling to existing climate workflows. The 209-year training set supplies sufficient internal variability to learn mesoscale dynamics without atmospheric nudging after day zero.

On the independent OceanBench benchmark the reanalysis-trained variant undercuts all competing systems in root-mean-square error, confirming that unstructured-mesh emulation is now competitive. The main remaining constraint is the single-model training corpus; multi-model ensembles and longer reanalysis integrations would strengthen claims of generalisability.

Next steps include coupling the emulator to atmospheric ML models for fully data-driven seasonal forecasts and testing real-time initialisation from satellite and Argo observations within the next 12 months.