Assessing the impact of the new mixed layer eddy parameterization based on machine learning in NEMO

Mixed layer eddies (MLE) are submesoscale structures, characterized by spatial and temporal scales of O(10 km) and O(1 day), generated by mixed layer instability under conditions of strong horizontal buoyancy gradient and weak stratification.  MLE produces mixed layer restratification, which has important implications for global ocean and climate dynamics. Existing parameterizations represent MLE effects with a streamfunction that depends on the horizontal buoyancy gradient, mixed layer depth, and the Coriolis parameter. Machine learning techniques have recently been proposed for improving existing MLE parameterizations. Bodner et al., (2024) proposed an approach for predicting submesoscale vertical buoyancy fluxes using a convolutional neural network (CNN), showing an improvement compared to previous parameterizations.

In this study,  we analyze the impact of a new MLE parameterization - based on Bodner et al. (2024) - in a global ocean model simulation performed with NEMO (eORCA25). The implementation of the CNN parameterization in NEMO is performed through EOPHIS (https://github.com/meom-group/eophis/). The CNN simulation (MLE-CNN) is compared with a simulation with a standard  parametrization and a simulation without MLE parameterization. With the CNN parameterization, maximum winter mixed layer depths are reduced by 10% with respect to the simulation without parameterization, which is comparable to the reduction obtained with the standard parameterization. The CNN parameterization differs from the standard parameterization in terms of  spatial variability.  For example, in the tropical region, the CNN produces a vertical heat flux across the mixed layer that can reach twice the magnitude of the standard parameterization. Mixed layer depth from simulations will be compared with observations.