Toward a new parameterization of ocean-atmosphere interactions based on a machine learning approach

In ocean models, coupling with the atmosphere leads to two major phenomena at the mesoscale (1/4° ~ 25 km) : current feedback and thermal feedback. At present, numerical oceanic simulations that include these phenomena require either coupling with an atmospheric model or an atmospheric parameterisation that already includes this coupling. However, a different approach is emerging in the modelling field, using machine learning. We will present a neural network that aims to find the surface wind stress that reproduces the effects of the ocean-atmosphere coupling at the mesoscale. The idea here is to avoid atmospheric simulations and to use an uncoupled atmospheric forcing as one part of the neural network inputs, the other part being data from the ocean simulation. The data used are from the realistic coupled ocean-atmosphere simulation PULSATION, in the tropical channel, 45°N to 45°S. This simulation is based on the oceanic model NEMO, the atmospheric model WRF and the interface OASIS. The training is carried out in turbulent zones where current feedback plays an essential role in reducing eddies. Our experiments show that the surface wind stress are well retrieved, resulting in a current feedback representation as accurate as the one obtained from a standard physical-based parametrisation, while allowing the representation of an accurate thermal feedback.