Item

Abstract

New observations and high-resolution simulations provide opportunities for developing data-driven Gravity Wave (GW) parameterizations. GWs play an important role in the large-scale atmospheric circulation, e.g., in driving the Quasi-Biennial Oscillation (QBO), a 28-month oscillation of jets in the tropical stratosphere. Yet, their horizontal scales of 10^2 to 10^5 meters cannot be fully resolved even by the highest-resolution climate models, and hence their impacts on the large-scale flow must be parameterized. We train a variety of machine learning algorithms to emulate the GWs in a 1D QBO model relevant to the physical scenario and the learning problem of GW parameterizations. The model is driven by a stochastic source term that mimics convective GWs with randomly distributed fluxes (related to precipitation) and spectral widths (related to the depth of convection). A key challenge of many machine learning algorithms, which is of practical concern for data-driven GW parameterization, is their ability to generalize. In the context of GW, their sources depend on the resolved flow, making them susceptible to model biases. As a result, the source distribution fed to the parameterization by any model is expected to be somewhat different from the observed distribution on which it was trained. While the different methods successfully learn the “observed” GW source distribution, they struggle to generalize to nearby distributions. The implication of these results is that a calibration step is likely to be required upon implementing a data-driven GW parameterization in operational models.