Primary Submission Category: Causal Discovery
Climate Dynamics via Spatially Informed Causal Discovery
Authors: J. Jake Nichol, Michael Weylandt, Laura P. Swiler,
Presenting Author: J. Jake Nichol*
Understanding the Earth’s climate is perhaps today’s largest, most complex, and most important scientific challenge, and poses many difficulties for causal inference. Accurate characterization of the underlying causal mechanisms is essential for the design and analysis of potential climate change mitigation strategies. Climate data science is challenging due to both the scale of underlying data and the difficulty in obtaining meaningful replicates and counterfactuals, even in simulation. Earth systems data typically consists of hundreds of quantities of interest, each of which is observed at over 4.5 million distinct spatial locations; by contrast, less than a thousand observations are typically available. In this ultra-high dimensional regime, popular methods for causal discovery, such as the PCMCI algorithm (Runge, et al. Science Advances, 2019), exhibit high false positive rates and are unable to separate the causal wheat from the correlative chaff. To address these challenges, we propose new approaches for causal discovery that leverage regionally homogeneous spatial dynamics to create informative pseudo-replicates, improving statistical performance and interpretability. We demonstrate the effectiveness of our approach in simulation and in an application to an important open question in atmospheric dynamics. Our approach enables causal discovery in massive spatiotemporal data and provides an important toolkit for understanding climate dynamics.