A New Causal Discovery Framework Accounting for Nonlinear Emergent Physics, Applied to Convective Initiation and Tropical Cyclone Rapid Intensification - NASA/ADS

Causal discovery in the geosciences requires addressing a unique set of problems, such as a plethora of sparse data, variables that are almost always continuously distributed, interacting nonlinear processes, and many hidden processes that are potentially important. We have realized a causal discovery framework which fully addresses the issues of continuous distribution and nonlinearity, whereas most existing attempts address only one of these issues at a time. Uniquely, the framework also determines the size of the contribution from missing processes, so we know how complete our causal inference is. By exploring information theoretic relations, it offers a unique decomposition of causal information into direct causation by one process, as well as joint, or emergent, causation when several processes work together to drive another process. Most existing literature stops at calculating what we call direct causation, thereby failing to capture crucial emergent physics. Furthermore, by dividing the emergent causations among their constituent drivers, we can calculate the total causation by each driver, including its interactions with other drivers, which has not been done before. <P />We will present the application of our framework to the Lorentz 1963 dynamical system and to two real-world systems. For the Lorenz system, the direct contributions tell us about the larger-scale dynamical structure (including the attractor) of the system, while the total causations inform about the underlying differential equations. In our application to soil moisture's role in convective initiation we found, among others, that convective inhibition is one of the main determiners of whether or not there is convection, while convective available potential energy was surprisingly a rather weak determiner. We are now exploring the causes for the rapid intensification of Hurricane Patricia using this new powerful framework.

saved by

• David A. Lee