Generalization and Calibration: A 1-D QBO model testbed for data-driven gravity wave parameterization
Date:
I attended this meeting virtually, actually tuning in from three different locations. In a classic, left hand doesn’t know what the right is doing, I managed to schedule our moving date to Berlin to coincide with my talk! It only worked out in the end because our sessions were in the late afternoon in Pittsburgh, so I could join our sessions at 10 pm. It started on the last day of our vacation in Tuscany, then I was in Bayreuth for the second day, as we paused on our drive across Europe, reaching Berlin for the final day of the symposium when I gave my talk!
Abstract
Gravity waves, or buoyancy waves, so named because their restoring force is the action of gravity on a stratified fluid, present a challenge to atmospheric modeling. They play an important role in the circulation by transporting momentum, but cannot be properly resolved in global models. Further, many of their sources, e.g., moist convection, are themselves not directly represented. Gravity wave impacts must therefore be approximated, or parameterized, based on the resolved flow. New observations have raised hope for a data driven approach to gravity wave parameterization.
We first demonstrate the potential for machine learning to capture gravity wave momentum transport. We focus on a macroscopic effect of gravity waves on the circulation, the Quasi-Biennial Oscillation (QBO), a 28 month oscillation of jets in the tropical stratosphere. Neural network and regression tree approaches can successfully emulate an existing, physics based parameterization. Most critically, schemes trained on limited data successfully emulate out-of-sample conditions when coupled online with the model. We then turn to the question of calibrating data driven parameterizations to work with an imperfect atmospheric model (emphasizing that all models are imperfect), formulating a 1-D model of the QBO as a testbed. The simple 1-D model allows us to explore techniques to calibrate data drives schemes to compensate for biases in the resolve flow and gravity wave sources.