Collaborative Research: Improved Minimization Techniques in Meteorological Data Assimilation

点击翻译按钮获取中文摘要

The goal of this project is to develop improved optimization techniques for use in three- and four-dimensional data assimilation in meteorology that will both reduce the computer time required in the assimilation process and increase the reliability of the results. The research will consist of experimental work with existing assimilation codes, in conjunction with a theoretical study of both the optimization algorithms and the physical models. The project represents a collaborative effort among Dr. Jorge Nocedal (Northwestern University), Dr. Stephen Wright (University of Chicago), and researchers at the National Centers for Environmental Prediction (NCEP).The new optimization techniques include enriched limited memory quasi-Newton methods for minimizing the nonlinear cost functions, automatic preconditioners to accelerate the conjugate gradient method, and structured quasi-Newton methods of nonlinear least squares problems. The new techniques will be designed for the case in which the background covariance matrix does not have a simple sparsity structure. The techniques must also be robust when varying levels of resolution are used in the model during computation, a common situation with meteorological models. Thus, a theoretical framework that quantifies the effects of these inaccuracies on the performance of the algorithms will have to be developed.The Divisions of Atmospheric Sciences and Mathematical Sciences jointly support this project.