Nonlinear Bayesian Inference for High-dimensional Systems

Peter Jan van Leeuwen
Atmosphereic Science, Colorado State University

ABSTRACT: In the geosciences, Bayesian inference is known as data assimilation, and many powerful methodologies have been developed, for instance for weather prediction. At present, weather prediction centers try to find useful solutions to the data-assimilation problem in a state space of size 10 billion or more. These methods are all based on linearizations of the problem, while with ever increasing model resolution and more sophisticated observations the problem has become highly nonlinear. Hence, there is a call for fully nonlinear methods. A new possibility has been developed based on so-called particle flows. Van Leeuwen demonstrates the usefulness of this new approach in simple toy problems and in high-dimensional ocean systems discussing theoretical and practical issues. Finally, van Leeuwen discusses how the present-day weather prediction scheme can be adapted to transform weather prediction from an approximate linearized solution to a fully nonlinear solution of the Bayesian Inference problem.

View vanLeeuwen's presentation

Start date
Friday, Feb. 28, 2020, 10:10 a.m.
End date
Friday, Feb. 28, 2020, 11 a.m.
Location

George J. Schroepfer Conference Theater, 210 Civil Engineering Building

Peter Jan van Leeuwen

Share