The Physical-space/Finite-volume Data Assimilation System (fvDAS) is the next generation global atmospheric data assimilation system in development at the Data Assimilation Office at NASA's Goddard Space Flight Center. It is based on a new finite-volume general circulation model jointly developed by NASA and NCAR, and the Physical-Space Statistical Analysis System (PSAS) developed at the DAO. The data assimilation method implemented in fvDAS incorporates a simplified version of the model bias estimation and correction algorithm, as described by Dee and da Silva (1998).

Traditional data assimilation methods are based on assumptions that both models and observations are unbiased. However, it is generally accepted that neither is true in practice. A great portion of the work in data assimilation at operational weather centers is concerned with observational quality control and bias removal. Recent efforts at the DAO have addressed the issue of model bias correction by means of data assimilation, under the assumption that the remaining bias in the assimilated observations is small compared to the model bias. Dee and Todling (2000) described the implementation of on-line model bias correction applied to the moisture analysis in a global atmospheric data assimilation system, which is currently operational at the DAO.

This presentation describes the implementation of on-line model bias correction in the new fvDAS, applied to the complete multivariate atmospheric analysis. Straightforward application of the algorithm presented in Dee and da Silva (1998) requires an additional solution of the analysis equations for each set of observations used for model bias estimation. The algorithm is flexible, in terms of the choice of observations and with regard to the analysis weights used for model bias estimation. However, doubling the cost of each global analysis may be prohibitively expensive in operation, depending on the available computational resources and on the cost of other components of the DAS. Alternatively, the algorithm can be simplified by (1) using the same set of observations for analysis and bias estimation and (2) modifying the analysis weights for model bias estimation to take advantage of computations already performed in the traditional analysis step.

We will present early results obtained with model bias correction in the fvDAS, using both the original and the simplified versions of the algorithm.