Assessing Data-Driven Eddy-Parameterizations in an Atlantic Sector Model

Mesoscale eddies are the dominant reservoir of kinetic energy in the ocean and a key component of the Earth’s climate system. Current state-of-the-art climate models can only partially resolve mesoscale processes due to the limited spatial resolution imposed by computational resources, making the accurate parameterization of unresolved kinetic energy transfers crucial. In recent years, a growing number of studies use machine learning tools to build data-driven eddy parameterizations. In this work, we implement two such parameterizations, Zanna and Bolton (2020) (ZB20) and Guillaumin and Zanna (2021) (GZ21), into the NEMO ocean model. We tune and evaluate them against a higher-resolution ground truth, and compare them to a baseline backscatter parameterization in an idealized Atlantic basin configuration. GZ21 is unaware of the model grid, resulting in a systematic bias in the predicted fluxes tied to the local grid spacing. It does not improve the coarse simulation concerning any metrics presented here. ZB20 accurately predicts subgrid fluxes and improves both local kinetic energy spectra and large-scale circulation. It yields an improved mean state comparable to that of the baseline parameterization, consistent with previous studies. We emphasize the importance of carefully designed training data, which takes into account all resolution-dependent model components, such as viscosity and model geometry. Our findings provide guidance for the future development of robust and generalizable data-driven eddy parameterizations.