Speaker
Description
Global climate change projections, such as those from the Coupled Model Intercomparison Project phase 6 (CMIP6), are still subject to substantial modelling uncertainties. These uncertainties limit our ability to fully assess future climate risks and to effectively inform policy. At the same time, running high-resolution, global climate models remains computationally intensive, posing further challenges.
In this talk, I will highlight three key ways in which machine learning (ML) can help advance climate science:
(a) using observational constraints, i.e. to constrain uncertainty in CMIP6 projections on the basis of high-dimensional relationships in Earth observations,
(b) developing ML parameterizations of complex Earth system processes that can be both more accurate and computationally efficient than traditional parameterizations or explicit process representations, and
(c) building fast AI-driven emulators of climate models to enable rapid exploration of a variety of future scenarios.
I will emphasize the importance of addressing core challenges such as extrapolation in a changing climate system and understanding causal relationships, which are central to consider when working with purely data-driven ML models.
