Climate and Earth system models are important tools to understand and
project climate change. Due to their complexity, they are limited in
their horizontal resolutions, and some processes remain
uncertain. Machine Learning (ML) together with short km-scale
simulations and Earth observations provide new opportunities to reduce
long-standing systematic errors and to improve projection
capability. In this presentation, we argue that ML should be fully
exploited to (a) develop hybrid ML/physics Earth system models with
greater fidelity, (b) to improve detection, attribution, and
forecasting of extremes, and (c) to advance climate model analysis and
understanding of the Earth system. We further discuss how tackling key
ML challenges such as generalization, physical constraints,
uncertainty quantification, explainable artificial intelligence,
causal inference, and benchmarks can help achieve these goals. This
effort will require bringing together ML and climate scientists, while
also leveraging the private sector, to accelerate progress towards
desperately-needed actionable climate science.