02/24/2020 - 3:30pm to 4:30pm
101 Nierenberg Hall
Prof. Peter Jan van Leeuwen
Colorado State University
Abstract. Determining the drivers of interesting processes in complex systems such as highly nonlinear ocean circulations or biogeochemistry interactions
is crucial for our understanding of these systems. Present-day methods for causal discovery are either linear (e.g. exploring partial correlations) or if they are nonlinear they are incomplete. Furthermore, many processes cannot directly be observed, or the observations are incomplete in space and/or time, and numerical models are needed. But using numerical models can easily lead to causal inference based on model world, and not the real ocean.
We have developed a new causal discovery framework that can handle highly nonlinear relations between different processes and is complete in the sense that it not only looks at direct relations between processes, but also includes how processes influence a target process together. A simple example would be the case in which a driver process influences the target process if it reaches a certain threshold, and the value of that threshold is set by another process. Existing causal discovery frameworks cannot handle this; we can.
To avoid obtaining answers from model world while still using the ability to dynamically spread information we combine the information in the observations and in the model using data assimilation. Since the underlying system is nonlinear we have developed a fully nonlinear data-assimilation method based on so-called particle flows.
I will outline both the causal framework and the nonlinear data assimilation method, each with examples of their use, followed by present-day efforts and future plans for combining these two powerful new tools.
For more information on this event, contact: