Data assimilation combines a model forecast of a physical system and observations of the system to produce improved estimates of the system's state. Assimilation can also provide estimates of the errors associated with the forecasts and the observations. In geophysics, state-of-the-art data assimilation has traditionally been developed for numerical weather prediction in the troposphere. Until recently, assimilation systems for weather were model- and observation-specific and required enormous amounts of personnel resources for development.

In the 1990s, a new class of assimilation algorithm, ensemble assimilation, was developed for ocean and tropospheric applications. Ensemble assimilation systems are Monte Carlo approximations to bayesian filters. They are relatively easy to develop and apply and can require fewer model- or observation- specific details. A basic understanding of how ensemble filters work will be provided in the first half of the presentation.

Using ensemble assimilation systems with a small number of ensembles members (model integrations) in conjunction with large models and/or observation sets is not necessarily straightforward. In tropospheric weather prediction, two heuristic algorithmic enhancements, 'localization' and 'inflation', are required for good performance. 'Localization' is predicated on assumptions about the covariance structure of the state estimate that should be produced by the assimilation. 'Inflation' is required to compensate for assimilation system and model errors in forecast estimates of both the mean and variance of the state. A discussion of requirements for successful application of ensemble filters will make up the second half of the talk. Space weather experts will be provided with sufficient information to evaluate if ensemble assimilation is practical for their particular modeling/observing problems.