Seminar Series - Learning the weather to predict the climate. Dr. Noah Brenowitz, Vulcan Inc.

Tuesday, November 24 at 2 p.m. via Zoom

Speaker:  Dr. Noah Brenowitz, Senior Machine Learning Scienitist for Climate Modeling (, Vulcan, Inc.

Title: Learning the weather to predict climate.

Abstract:  Modern numerical climate and weather forecasts are a crowning achievement of classical mechanics. With them, we can accurately predict the weather a week in advance and estimate how the climate will respond to emissions of greenhouse gases, but there is still room for improvement. While climate models largely agree about future temperature increases, changes in precipitation are much more uncertain. Because of computational constraints, climate models can only explicitly resolve motions larger than a few 100 km, so anything smaller than that, like a rain cloud, must be parameterized. For simplicity, traditional parameterizations depict small-scale processes in a manner that is easy to write on paper or code in Fortran, but can suffer from poor accuracy. On the other hand, machine learning can improve upon these traditional schemes by learning the coarse-grained effect of unresolved motions directly from much higher resolution (3km) weather simulations. In this talk, I summarize recent efforts in this direction and give examples of how ML can improve upon the accuracy of traditional methods. I will also discuss how we can use ML interpretability techniques to gain new physical insights or confirm existing intuitions.