Machine Learning-Based Causally Informed Atmospheric Parametrizations for Climate Models
Abstract
Earth system models are fundamental to understanding and projecting climate change. The models have continued to improve over the years, but considerable biases and uncertainties in their projections remain. A large contribution to this uncertainty stems from differences in the representation of phenomena such as clouds and convection that occur at scales smaller than the resolved model grid. These long-standing deficiencies in cloud parametrizations have motivated developments of global high-resolution cloud-resolving models (horizontal grid resolution of a few kilometers). They can explicitly resolve clouds and convection, but due to high computational costs cannot be run at climate timescales for multiple decades or longer. In this talk we show, how short regional and global ICOsahedral Non-hydrostatic (ICON) high-resolution simulations can be used to train machine learning (ML)-based atmospheric parameterizations such as cloud cover and convection. While unconstrained neural networks often learn spurious relationships that can lead to instabilities in climate simulations, we also show that causal discovery can mitigate this problem by identifying direct physical drivers of subgrid-scale processes, using the Super Parameterized Community Atmosphere Model v3.0 (SPCAM) as a testbed.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMNG16A..03E