Mathematical Analysis of Large Data Sets

Monday, February 6, 2006, 4 p.m. PSA 113

Eric Kostelich

Dept Math & Stats

Data Assimilation: Finding the Initial Conditions in Large Dynamical Systems

Abstract One of the most significant problems in modeling a spatio-temporal dynamical system, such as a weather forecast model, is finding the initial conditions with which to start the simulation. Present-day global forecast models incorporate on the order of 100 million dynamical variables, and operational forecast centers collect on the order of 1 million measurements every six hours. Data assimilation refers to the problem of incorporating the measurements into the dynamical models and updating the initial conditions. This is the most expensive part of numerical weather prediction. By the end of the decade, new satellite observing systems will produce an order of magnitude more data than current assimilation systems can handle.

In my talk, I will describe a new, model-independent approach to the problem that promises to be much more accurate than current schemes, can incorporate huge amounts of data, and is amenable to efficient implementation on parallel computers. Applications to models other than weather forecast models will be discussed, and implementation strategies on parallel computers ranging from modest Beowulf clusters to IBM's Blue Gene will be described.

For further information please contact: Anne Gelb