Intelligently Identifying Patterns of Predictability in the Oceans and Atmosphere Using Interpretable Neural Networks

Jamin Rader, Colorado State University

Photo of Jamin Rader

Extreme events in the climate system (e.g. heat waves, extreme precipitation, droughts) are projected to increase in frequency, intensity, and duration through the 21st century. With these changes, there is a need for enhanced predictability of the atmosphere and oceans on seasonal-to-decadal timescales. However, the climate system is chaotic, complex, and changing. These three c’s encapsulate the barriers to climate prediction. It can be difficult to disentangle which features of our climate are chaotic, and thus cannot be predicted, and which are just complex, requiring novel data science techniques in order to be predictable. Furthermore, as our climate changes, patterns of predictability may change as well. To approach this problem, we use neural networks and other machine learning techniques to identify patterns of climate variability and change. In this presentation, I will introduce three neural network architectures used for Earth System predictability, lessons learned from each, and how we are using this knowledge to intelligently identify patterns of predictability for extreme events.

Abstract Author(s): Jamin K Rader, Elizabeth A Barnes, Imme Ebert-Uphoff, Chuck Anderson, et al.