Ocean models for climate applications : progress expected from Machine Learning

As an ocean and climate modeller, I propose to expose a few venues of ocean modelling where Machine Learning (ML) is expected to break through persistent challenges. My prime target is the numerical representation of the global ocean, with distinguishable coarse spatial scale (25 to 100 km) and long duration (at least 100 years). Observations are not sufficient (too sparse in space, particularly at depth, and too short in time, spanning only the last few decades) to be used directly as the sole ground truth. Hence it is compulsory to consider perfect model set-ups, besides training on observed database. Current challenges in ocean modelling that ML could contribute to solving, are the following : equilibration of simulations, quantification of sensitivity to parameters, parameterizations of unresolved processes (due to reduced spatial resolution and/or complexity) and quantification of structural uncertainties. I will introduce a few ML-based solutions to these challenges based on recent bibliography and my own activities. Overall, we need to build capacity in bridging the gaps between these centennial global ocean simulations, useful for climate applications, process models at regional scale, global ocean hindcasts (simulations with data assimilation), large eddy simulations and models of the past, present and future climate. To reach this goal, I advocate combining various ML architectures, factoring in uncertainties of every pieces of this hierarchy.