Title: Balanced Ensemble Localization with Normal Mode Initialization

Authors: Tom Rosmond, Science Application International Corp, Forks, WA
Craig Bishop (Naval Research Lab, Monterey, CA)
Dave Kuhl (Naval Research Lab Post Doc, Washington, DC)
Liz Satterfield (Naval Research Lab, Monterey, CA)

Ensemble based covariances provide flow dependence that makes them are an attractive feature for 4-D VAR data assimilation. In most applications, however, computational costs limits the size of ensembles, and the resulting covariances have non-local features that cause problems for data assimilation. A pragmatic solution to this problem is applying localization functions to the covariances that filter the non-local structures, yielding covariances that are a reasonable approximation to those possible with very large ensembles.

A troubling by-product of localization is that the wind-mass balance present in the original covariance structures is essentially lost, and this imbalance propagates to the increment fields produced in the data assimilation system. This imbalance can compromise the effectiveness of that system.

At NRL-Monterey this problem has been solved by adding the TLM and adjoint of the NOGAPS non-linear normal mode initialization to the accelerated representer cost function upon which our ensemble DA system is based. Each conjugate gradient iteration the interim solution is 'balanced', and at convergence the increment fields are as well. There are several variations possible with the procedure, including special treatment in the stratosphere to prevent spurious inflation of increments. A mathematical discussion of the method will be presented, along with preliminary results of system performance for some of the variations.

GMAO Head: Michele Rienecker Global Modeling and Assimilation Office NASA Goddard Space Flight Center

Curator: Nikki Privé Last Updated: May 27 2011