Atmosphere Ocean Science Colloquium

---

Coupling Neural Networks to GCMs

Speaker: Noah Brenowitz, U Washington

Location: Warren Weaver Hall 1302

Date: Wednesday, October 16, 2019, 3:30 p.m.

Synopsis:

One of the most obvious ways to improve climate models is to reduce the errors made by sub-grid-scale parameterizations. While there have been steady improvements to traditional sub-grid-scale parametrizations over the past decades, machine learning presents a much more direct path to reducing these errors based on realistic datasets. However, ML parameterization is not a straightforward prediction problem (e.g image classification) because any scheme must interact well when coupled with the dynamical core of a GCM. For instance, the ML schemes should avoid numerical instability. In this talk, I will discuss our efforts to couple a neural network parametrization of moist physics to a GCM. To help solve the numerical stability problem, we use a linearized framework to study the sensitivity of the neural network to its inputs.