Jonathan Poterjoy, National Center for Atmospheric Research (NCAR)
Abstract for special seminar on Wednesday, October 26, 2016 at 3:30pm in Math 101
A major challenge for producing reliable probabilistic forecasts of the atmosphere is the specification of initial conditions, boundary conditions, and their errors for weather prediction models. Monte Carlo filtering methods—typically in the form of ensemble Kalman filters—are often used for this purpose. These methods operate under linear and Gaussian approximations for the model and its errors, respectively, to form probabilistic estimates of state variables from partial measurements of the atmosphere. The underlying assumptions of these methods, however, limit their effectiveness for applications where nonlinear physical processes lead to highly non-Gaussian errors, or when observations relate nonlinearly to model state variables. In this talk, I will introduce a new nonparametric filter, based on particle filtering strategies, that has potential for nonlinear/non-Gaussian problems encountered in weather models. Benefits of this method will be discussed using applications where convectively driven atmospheric motions determine the evolution of the dynamical system (i.e., severe thunderstorms and squall lines).