Machine Learning for Parameterization of Moist Processes in the Atmosphere

The representation of sub-grid processes contributes to uncertainty in climate predictions by earth system models. Specifically, the parameterization of convection and clouds is a major contributor to the uncertainty in changes in temperature, rainfall distribution, and severe storm frequency. An increasing number of studies show that machine learning can be used to build data-driven parameterizations directly from high-resolution model output. However, the resulting parameterizations do not always lead to stable and accurate simulations when implemented in a coarse-resolution model. Previous work suggests that a machine learning approach based on an ensemble of decision trees (random forest) can be used to robustly emulate a conventional moist convection scheme. Here we describe the use of random forests to learn a sub-grid parameterization from coarse-grained output of a quasi-global high-resolution simulation of the atmosphere with hypohydostatic rescaling. We discuss the performance of the parameterization when implemented at coarse resolution in the same model with a focus on statistics of mean and extreme precipitation. Furthermore, we discuss the performance of the parameterization at different coarse resolutions and with different sub-grid processes included in the parameterization, and we demonstrate that high accuracy of the parameterization in offline tests does not always lead to accurate simulations.