Jeffrey L. Anderson (National Center for Atmospheric Research)
Data Assimilation for Weather and Climate Models with Ensemble Filters
Friday 2:00-2:30, Fountain II
Abstract:
Data assimilation combines information from observations and model predictions to produce estimates of the state of the atmosphere. These estimates can be used as initial conditions for subsequent forecasts and studied to learn more about physical processes in the atmosphere. There is also growing interest in using assimilation to assist in model development
by confronting models with observations and studying systematic
errors. For many applications like numerical weather prediction, both the model state vectors and the number of observations to be assimilated are very large. Being able to assimilate millions of observations in models with millions of state variables is essential.
Simple methods for data assimilation can be developed by using Monte Carlo methods, Bayes rule, and a prediction model. These methods can be decomposed into two parts: how does a direct observation of a single variable impact a prior ensemble estimate of that variable?; how should these changes to a single observed variable impact prior estimates of a single unobserved variable? Ensemble algorithms to solve these problems are straightforward but must be augmented to deal with sampling error arising from the use of small ensembles. An overview of the ensemble filter assimilation testbed at NCAR is presented. Results from an assimilation of observations used for operational numerical weather prediction in an atmospheric general circulation model are used to
demonstrate the performance of the ensemble assimilation algorithms in
large problems.