Atmosphere Ocean Science Friday Seminar

Towards ML acceleration of a ray-tracing gravity wave parameterization, with emphasis on the “towards”

Speaker: David Connelly, CAOS

Location: Warren Weaver Hall 512

Date: Friday, September 22, 2023, 4 p.m.


Parameterizations of atmospheric gravity waves often only include effects on the mean flow associated with wave breaking at critical levels. Ray tracing is an attractive framework in which to develop more complete parameterizations that include non-dissipative pseudomomentum fluxes. However, ray tracing can be quite computationally expensive, especially in the many-Lagrangian-particle regime necessary for convergence. I am in the early stages of efforts to use machine learning to accelerate MS-GWaM, a ray tracing parameterization. As this project is still in its infancy (or at least its toddlerhood), I will begin with an overview of the relevant WKB theory and ray tracing equations. I will then propose several approaches to machine learning acceleration of this sort of system, paying particular attention to how each will respect (or fail to respect) the appropriate wave action conservation properties of the system.