Machine Learning Emulation of Subgrid-Scale Orographic Gravity Wave Drag in a General Circulation Model With Middle Atmosphere Extension

Gravity wave parameterizations contribute to uncertainties in middle atmosphere modeling. To investigate the potential for using machine learning to represent atmospheric gravity waves and the impact of implementing such schemes in a general circulation model (GCM), we train a random forest (RF) emulator on outputs from an existing complex parameterization scheme for orographic gravity wave drag (GWD). The performance of the RF emulator is then evaluated with a focus on stratospheric climatology and variability in climate simulations from the middle atmosphere resolving Beijing Climate Center Atmospheric Circulation Model. In offline tests, the predicted orographic GWD by the RF agrees remarkably well with the target GWD throughout the troposphere and the middle atmosphere. The RF emulator can reproduce the observed climatology of zonal-mean wind and air temperature in the GCM simulation, as well as its target scheme. Compared to the target orographic GWD parameterization scheme, the RF emulator can reproduce the breakdown of the polar vortex in the Southern Hemisphere stratosphere. This study demonstrates the feasibility of using machine learning to emulate parameterized orographic GWD for modeling the stratosphere with a GCM. Machine learning has been utilized to learn atmospheric physical parameterizations such as moist convection, cloud microphysics, radiation, nonorographic gravity waves, etc. However, it remains unknown whether it is feasible to apply machine learning algorithms to parameterize orographic gravity waves and how such schemes perform when interactively coupled in atmospheric general circulation models. Here we employ the widely used random forest algorithm to learn from an existing complex parameterization scheme for orographic gravity wave drag. Then, we implement the random forest scheme in a high-top atmospheric general circulation model to test its performance. We find that the random forest scheme acts as well as the target parameterization scheme. This study demonstrates the potential of using machine learning to emulate parameterized orographic gravity waves for modeling the middle atmosphere. In future work, it is recommended that machine learning may be used to develop new parameterization schemes directly from observations or high-resolution model outputs to overcome critical deficiencies in current orographic gravity wave parameterizations, helping to reduce the common biases in simulating the stratosphere with general circulation models. Parameterized gravity wave drag from unresolved orography can be emulated with a random forest The random forest emulator simulates vertical distributions of the zonal wind and air temperature quite well in a general circulation model (GCM) The GCM configured with the random forest scheme reproduces a reasonable breakdown of the austral polar vortex in the stratosphere