Discovering Novel Eddy Parameterizations with Machine Learning

Current climate models do not fully resolve ocean eddies due to computational constraints, and consequently their effects on the large-scale flow must be parametrized. Recent studies have successfully used neural networks to construct eddy parametrizations. However, this approach can sometimes sacrifice physical interpretability. Our study aims to use machine learning to accurately capture the effects of unresolved eddies, while retaining a clear physical interpretation of the end result. Specifically, we discover a closed-form equation for a parametrization of eddy momentum fluxes - as opposed to training a black-box function - using an iterative algorithm based on relevance vector machines. We show that this method can reveal simple and interpretable eddy parametrizations, which capture a significant proportion of the true eddy forcing variance in idealised models. The expressions can also be constrained to respect physical conservation principles. When the parametrization is implemented in idealised models, the turbulent energy of the flow is increased, and the model bias compared to higher resolution simulations is decreased.