Deep Learning of Unresolved Turbulent Ocean Processes in Climate Models


Climate models are an approximate representation of the laws of physics describing the evolution of the ocean and atmosphere dynamics. Due to limited computational resources, many ocean processes, which are crucial for the transport of heat and carbon, occur at scales smaller than the grid resolution of climate models. Therefore, we rely on approximations, called parameterizations, to represent these unresolved processes in climate models. Parameterizations, traditionally derived from semi-empirical or idealized theories, are often imperfect and can lead to biases in climate models. Machine learning algorithms, and deep learning (DL) algorithms in particular, could provide an avenue to improve the representation of unresolved processes in ocean models by efficiently extracting information from high-resolution ocean simulations and/or observational data, potentially enhancing the skill of parameterizations in climate models.

In Deep learning for the Earth Sciences (eds G. Camps-Valls, D. Tuia, X.X. Zhu and M. Reichstein)
Laure Zanna
Laure Zanna
Professor of Mathematics & Atmosphere/Ocean Science [She/Her]

My research interests include Climate Dynamics, Physical Oceanography and Data Science.