Causality linking ENSO with North American rainfall: inference by data-driven causal discovery methods and physical evaluation

Abstract:

The core of scientific research is to understand the causality of phenomena. Traditional machine learning methods (e.g., classification and regression) widely used in climate science are essentially black boxes from which physical linkages are difficult to extract. In the climate science community, lagged cross-correlation (also known as Granger causality) is often used to study the interaction among physical processes and derive causal relationships. However, true causality is not necessarily inferred because this approach lacks sensitive in time series data with high auto-correlation and cannot discover the latent causal contributors. Previous studies have shown that North American rainfall exhibits a strong teleconnection with the El Niño Southern Oscillation (ENSO), and different types of ENSO events can result in differing climatic responses over North America. Two major teleconnection mechanisms, via wave trains or eddy-jet streams, have been proposed to explain the impacts of ENSO on middle- and high-latitude climate phenomena. In this study, we employ a new score-based causal inference method to discover the spatial distribution and intensity of North American rainfall affected by ENSO. Further, a causal mechanism chain can be established after incorporating other intermediate contributors. The credibility of this method is evaluated by the consistency with the two teleconnection mechanisms, which paves the way for extending to other earth science applications, including quantifying factors dominating or modulating ENSO. Preliminary results indicate that this method is able to capture the causality of ENSO effects on North America rainfall. The method proves to be a trustworthy and powerful tool to reveal critical causal mechanisms behind large-scale intricate spatiotemporal observational and simulated data for the strongly non-linear climate system.