Two applications of causal discovery in climate science: Imme Ebert-Uphoff

Abstract:This talk presents two applications of causal discovery in climate science. First, we briefly present a small example that studies the potential causal relationships between four prominent modes of atmospheric variability and show what we can learn using causal discovery. The remainder of the talk focuses on the second application, which discusses a new concept in climate science, graphs of information flow. The key idea is to interpret large-scale atmospheric dynamical processes as information flow around the globe and to identify the pathways of this information flow from observed data using causal discovery. We will discuss obstacles encountered and overcome, some interesting results we obtained, as well as the computational challenges we face for future work. Both applications employ algorithms for constraint-based structure learning for the causal discovery process, namely a version of the Peter and Clark (PC) algorithm that utilizes spatio-temporal data.