Atmosphere Ocean Science Student Seminar

Regression tree ensemble emulators of gravity wave parameterizations

Speaker: Dave Connelly, CAOS

Location: Warren Weaver Hall 1314

Date: Friday, October 22, 2021, 4 p.m.

Notes:

Orographic and convective gravity waves are important drivers of large-scale circulation patterns in the atmosphere, notably the quasi-biennial oscillation (QBO). However, the intermittency and relatively small spatiotemporal scales associated with individual waves means their effects must be parameterized in general circulation models. The growing availability of rich observational datasets and high-resolution numerical simulations suggests drawing on techniques from machine learning to build data-driven parameterizations. 

I will present experiments in using one family of machine learning methods, regression tree ensembles, to emulate a physics-based gravity wave parameterization. I will discuss design choices and offline performance for two ensemble architectures — random forests and gradient-boosted forests. I will then show the results of coupling regression tree emulators with an idealized circulation model, evaluating online accuracy, stability, and simulation of the QBO.