Harnessing data to explore the dynamics of extreme events in the atmosphere

A second opportunity to present research closer to home. On Halloween. In Greenwich Village.

Abstract

In a presentation geared towards newly arrived members of our community, I will discuss a research project led by a former PhD student. The goal was to use data-driven methods (statistical and machine learning models) to explore the dynamics of extreme events in our atmosphere. The talk will highlight work by Huan Zhang to understand atmospheric blocking events, a (somewhat) rare synoptic pattern of variability associated with extreme weather. Here, a deep learning predictor (a convolutional neural network, CNN) was trained to predict the persistence of blocking events. While the CNN had impressive skill, our true goal was to use it to back out dynamical insights into persistent blocking patterns and explore how similar predictions could be made with scarce reanalysis data with the use of transfer learning. This project was the fruit of a collaboration with Jonathan Weare and Dorian Abbot (Chicago), and I hope to encourage our PhD students and postdocs to collaborate with faculty across the math department and beyond.