Machine Learning Gravity Wave Parameterization Generalizes to Capture the QBO and Response to Increased CO ₂

We present single-column gravity wave parameterizations (GWPs) that use machine learning to emulate non-orographic (GW) drag and demonstrate their ability generalize out-of-sample. A set of artificial neural networks (ANNs) are trained the momentum forcing from a conventional GWP in an idealized climate model, given only one view annual cycle phase Quasi-Biennial Oscillation (QBO). investigate sensitivity offline online performance choice input variables complexity ANN. When coupled with moderately complex ANNs accurately generate full cycles QBO. model is forced enhanced CO2, its response ANN matches generated physics-based GWP. That can existing scheme new regimes limited data suggests potential for developing GWPs observational estimates GW transport.