Using neural networks and high-resolution simulations to improve mesoscale eddy representation in ocean models

Climate models are one of the most useful tools for predicting future climate states, which has become more important than ever in the ongoing climate crisis. However, due to their spatial and temporal resolutions, which are constrained by computing power and resources, climate models are not able to represent all processes in the ocean and atmosphere. Therefore, modelers need to estimate the effects unresolved processes have on the resolved processes.

One such structure is turbulent mesoscale eddies in the ocean. It is known from observations that eddies carry a large amount of kinetic energy and play a significant role in transport of tracers such as temperature and salinity as well as in heat uptake from the atmosphere. Therefore, it is crucial that eddies and their effects on the processes mentioned above are represented accurately in climate models.

To better estimate these effects in low-resolution simulations, high-resolution simulations can be used to constrain the parameters necessary for the estimates. However, tuning these parameters can be subjective and time-consuming. In this project, Machine Learning (ML) methods will be used to facilitate and speed up this process.

In my PhD project, high-resolution data from the FESOM2 ocean model will be used. At low resolution, which is insufficient to represent eddies, FESOM2 estimates the effects of the missing eddies using the Gent-McWilliams (GM) parameterization containing a GM coefficient. With the help of Bayesian Neural Networks, a framework will be developed to calculate a predictor for this parameterization as well as its variability. Using this framework, maps of the GM coefficient for multiple setups with increasing complexity and data volume will be created. The presentation includes a project outline as well as preliminary results.