A Data-Driven, Single Column Gravity Wave Parameterization in an Idealized Model

Abstract:

Current state-of-the-art physics-based parameterizations of gravity waves are severely limited by our inability to validate them with observations. Recent advances in machine learning have made developing a validated, data-driven gravity wave scheme a possibility. Here we present a single-column approach to gravity wave parameterization that uses machine learning to emulate an existing physics-based parameterization. A deep artificial neural network (ANN) is trained with a model of an idealized moist atmosphere (MiMA) as a proof of concept that an ANN can learn the salient features of gravity waves directly from resolved flow variables. We demonstrate that when trained on only the eastward phase of the Quasi-Biennial Oscillation (QBO), the ANN can skillfully predict the westward phase. Additionally, the meridional and zonal wind components are the only flow variables necessary to predict horizontal gravity wave tendencies with an R^2 value over .8 when averaged across all pressure levels. In addition to presenting the ability of our trained ANN to predict horizontal gravity wave tendencies, we provide an assessment of the ANN’s numerical stability when coupled with an idealized model and offer an analysis of its performance sensitivity to data availability.