Machine Learning and the Physics of Climate
Coordinators: Annalisa Bracco, Henk A. Dijkstra, Claire Monteleoni, and Laure Zanna
Predicting climate variability and change at regional or local scales remains challenging, while urgently needed, because regional climate closely interacts with and feeds back to global climate. The challenge is compounded by the rich complexity of interactions between different climate system components, including the atmosphere, ocean, land and cryosphere. Rapid advances could potentially be stimulated by the exponentially growing archive of observational climate data and modeling data from simulations of increasing resolution and complexity. But these big data archives present their own challenge for computation, inference, and understanding. At this stage, data-driven and machine-learning approaches to capturing scale and component interactions , paired with modeling experiments and in-depth research in model process representation, have become necessary to move the regional climate prediction problem forward.
The program seeks to identify and contribute to solutions for problems related to the functioning of the climate system on global and regional scales for which machine learning approaches, coupled with first-principles physical understanding, could effectively close existing gaps. The program seeks to convene world experts across the disciplines represented in current climate science practice within the collaborative environment of KITP. The aims of the program are:
1) to summarize present understanding and outlook for climate science and machine learning;
2) to identify the latest ideas, algorithms and techniques relevant to the recently available model results and observations;
3) to foster productive collaborations between the climate-science and machine-learning communities; and
4) to identify crucial objectives and open questions that will drive coherent advancement of climate science in the coming years through joint work by these communities.