Morays-community: a framework to share reproducible hybrid Machine Learning and Ocean modeling experiments.

The combination of Machine Learning (ML) with geoscientific models has become an active area of research, but many technical challenges still remain because of the heterogeneous nature of programming languages, library environments and hardwares. Much efforts have been made over the recent years to propose different frameworks to perform online deployment of ML components within geoscientific models. One common drawback to all these solutions is the complexity of the required software environment. The latter often relies on versioned libraries and codes, both for the geoscientific and the ML models. Thus, ensuring the reproducibility of hybrid geoscientific model experiments is challenging, as it requires describing several tools and how to deploy them. This becomes even more problematic as the number of coupling solutions for hybrid modeling increases and may be unfamiliar to the members of the different modeling communities.

Here, we introduce Morays as an example of a community-based workflow for sharing reproducible hybrid ocean model experiments. Morays uses a GitHub organization to host hybrid experiments material that leverage the OASIS coupler (https://oasis.cerfacs.fr/en), which is widely used in European climate models. Our framework is based on a Python library (https://github.com/meom-group/eophis) that facilitates the use of OASIS for deploying hybrid modeling pipelines bridging FORTRAN solvers and ML models implemented in Python. The geoscientific model and ML scripts are executed separately and exchange data through the coupling API. 

In this presentation, we will showcase several successful deployments of hybrid ocean model experiments with the NEMO ocean/sea-ice modeling framework. These experiments implement ML-based parameterizations and model correction schemes for improving different aspects of model solution (vertical physics, eddy parameterization, surface fluxes). All the experiments are shared openly in a dedicated GitHub organization (https://github.com/morays-community), as individual repositories following a standard template. We will present the material available to the community (tutorials, test cases), explain how to contribute, and discuss the broader perspective of reproducible workflow for future hybrid geoscientific models.