Knowledge discovery in climate science

Imme Ebert-Uphoff, Colorado State University

In this talk we discuss how machine learning methods, primarily causal discovery methods, can be used to discover scientific knowledge from data.  Causal discovery seeks to identify potential cause effect relationships from data.  We will start with a quick tutorial on causal discovery, specifically of constraint-based structure learning for probabilistic graphical models.  The goal is for everyone in the audience to understand the basic concepts that make causal discovery possible, as well as some of the limitations of the approach.  In the process we review applications of causal discovery in climate science, such as identifying pathways of interactions in the earth' atmosphere and deriving causal signatures for the output of climate models for the purpose of error checking.  We will illustrate the importance of close collaboration between the climate scientist and the machine learning researcher at every step of the research.

Finally, we will briefly touch on more general strategies in which machine learning methods can be used for knowledge discovery in climate science.  Namely, we believe that many machine learning models developed for specific purposes, e.g. prediction or classification, actually have important information embedded in the model itself.  Thus we promote the viewpoint of training a machine learning model from data and then studying the model itself to gain insights into the dynamics of the studied system.

Link to Recording: