Assimilation of real observations using the GFS model and the Maximum Likelihood Ensemble Filter


The Maximum Likelihood Ensemble Filter (MLEF) has been developed at the Colorado State University. It is a reduced-rank method based on the minimization of the cost function. This implies that the solution of the minimization is the conditional mode of the Probability Density Function. The ensembles are essentially used to estimate the curvature of the cost function, which is directly related to the inverse Hessian of the control problem. Thus, the analysis error covariance is obtained from the minimization algorithm as an updated inverse Hessian, rather than as a sample covariance. The Hessian preconditioning is in principle similar to the transformation of the Ensemble Transform Kalman Filter.

In THORPEX related research, the MLEF algorithm is installed on the NCEP IBM SP computer, and used with the GFS T62L28 model and the SSI data assimilation system for assimilation of real observations. In the preliminary phase, the MLEF is assimilating surface pressure, temperature and winds. In the near future all operational observations will be assimilated, including satellite and radar. Preliminary results of the MLEF with GFS and SSI will be shown and discussed, as well as some NCEP-specific details of the algorithm.

Milija Zupanski and Arif Albayrak