A new preprint by Kacper Nowak and colleagues, HClimRep-Ocean: A Global Ocean Emulator on an Unstructured Mesh, presents a machine-learning emulator developed to predict ocean dynamics directly on the native unstructured mesh of the FESOM2 ocean model. This approach avoids the need to interpolate ocean fields onto a regular grid, offering a promising direction for machine-learning-based ocean modelling.
The study demonstrates encouraging results for predicting horizontal ocean currents at a 30-day forecast horizon, outperforming the reference methods considered. The OceanBench platform provides a visual comparison of HClimRep-Ocean’s performance against other ocean forecasting systems across different ocean variables and forecast lead times (Figure).
Figure. Comparison of ocean forecasting systems at 50 m depth on the OceanBench platform. The table shows root-mean-square deviation (RMSD) scores across different forecast lead times; lower values indicate smaller prediction errors.
Source: Modified from OceanBench.
However, predicting ocean temperature and salinity remains challenging, highlighting the need for further research to improve the representation of different ocean variables.
The work contributes to the HClimRep project’s broader efforts to advance machine learning for climate modelling and explore the potential of data-driven approaches for Earth system simulations.
