The Earth system is a highly nonlinear multi-scale system. Since it is impossible to explicitly resolve all relevant processes, especially at small temporal and spatial scales, modelling climate and more generally Earth system dynamics is challenging. Nevertheless, accurate models are crucial for tasks ranging from numerical weather prediction to climate projections, which are needed to identify viable pathways for sustainable Earth system stewardship in the context of anthropogenic climate change.

I will present recent methodological advances toward hybrid Earth system models, combining process-based and machine learning models. I will focus on geometrical approaches to enforce physical conservations laws and other constraints in neural differential equation models derived from data, as well as on generative machine learning methods for bias corrections and downscaling of Earth system model simulations, with emphasis on the representation and prediction of extremes.

Moreover, I will show how combinations of process-based and data-driven models can be used for improved modelling and predictability of potential large-scale critical transitions in the Earth system.

Niklas Boers is leading ClimTip (Climate Tipping Points) one of the two European Horizon projects dedicated to tipping points. He will present us briefly the ClimTip project and give a colloquium on the following subject.

Lieu
ENS – salle Claude Froidevaux – E314
24, rue Lhomond 75005 PARIS