The Graft-versus-host problem: Calibrating data-driven parameterizations to work in imperfect models
Date:
A virtual journal club run by the Institute of Computing for Climate Science at the University of Cambridge.
Abstract
In the event we successfully develop a well performing data-driven parameterization of un- and under-resolved gravity wave momentum transport, we do not expect it to produce desirable results when first ‘grafted’ into a ‘host’ atmospheric model. First, atmospheric models are capable of producing part of the gravity wave spectrum, and worse, likely simulating it poorly; a parameterization must be calibrated to account for what the host model does on its own. (We can turn this question around, asking how to properly compute the gravity wave momentum transport from a high-resolution integration. There are several potential strategies to isolate gravity waves, all which return slightly different answers, and it is unclear what scales count as unresolved with respect to a global atmospheric model.) Second, many of the sources of gravity waves (convection, frontogenesis) are themselves not well represented in atmospheric models – and errors in surface winds can bias topographically induced wave production – requiring a calibration of the gravity wave source spectrum. Finally, success is measured from on macroscopic behavior, e.g., a reduction in climatological bias, or a good representation of natural variability such as the Quasi-Biennial Oscillation. Current parameterizations serve a dual role of representing a missing process and correcting other biases in the host. We present a toy 1-dimensional model of the Quasi-Biennial Oscillation that allows us to explore calibration strategies in an idealized context. We show that offline-online learning may provide a route for calibration, though its computational feasibility in a full 3D model remains an open question. Even here, however, we worry whether too much calibration defeats the initial purpose of using a data-driven parameterization to accurately represent gravity waves: does a calibrated scheme get the right answer for the right reasons?