An Evaluation of Coupled Machine Learning Emulators for Physical Parameterizations in the Community Atmosphere Model

Abstract:

Atmospheric General Circulation Models (GCMs) are composed of a dynamical core and various physical parameterization schemes, which estimate the physical processes occurring below the horizontal resolution of the grid. This paper extends our previous work that emulates these physics tendencies with machine learning (ML) models in an offline mode. The trained ML models are now coupled back into the GCM and are explored in an online configuration. We test this approach with simplified physics packages in NCAR’s Community Atmosphere Model version 6 (CAM6), which is part of NCAR’s Community Earth System Model (CESM). The idealized parameterizations employed include a simplified dry atmosphere and its corresponding moist counterpart. The moist parameterizations include simplified forcing mechanisms for radiation, boundary layer mixing, surface fluxes, and precipitation. We develop, train, and test several machine learning-based techniques to approximate the full physics tendencies, as well as the large-scale precipitation field in the moist case. The implemented machine learning techniques include random and boosted forests, as well as a variety of deep learning architectures. We thereby shed light on the strengths and weaknesses of the selected ML models on climate time scales.