At the National Center for Atmospheric Research, a 48-member ensemble adjustment Kalman filter (EnKf) is being used to assimilate daily subsurface temperature and salinity data into the Parallel Ocean Program (POP) nominal 1° global ocean model. EnKf systems are typically initialized with an ensemble of model states that represent a climatological distribution. Over successive cycles of the assimilation system, the ensemble narrows into a distribution that is a function of the internal variability of the system and the observations that are constraining it. A well- equilibrated ensemble should not be strongly influenced by the choice of the initial ensemble. EnKF methods also employ algorithms to maintain the viability of the ensemble distribution and reduce the spurious impact of observations through mechanisms such as inflation and localization.

The POP/EnKf system was initialized in model-yr 1998 with the goal of having an equilibrated ensemble of ocean states by model-yr 2000. While the dynamic timescales of the upper ocean support this choice of "burn-in" time, the deep ocean remains sensitive to the choice of initial ensemble for at least a decade. We illustrate this with a case-study from the equatorial Atlantic ocean, where the choice of initial ensemble leads to spurious, transient behavior that impacts large-scale climate variables like the vertically integrated northward heat transport and the meridional overturning circulation.

Adaptive inflation is intended to combat the sampling error for each state variable and prevent filter divergence and has been used with success in other geophysical models. However, in the presence of model bias and unresolved small-scale dynamics, the application of adaptive inflation can lead to instabilties in frontal regions such as western boundary currents. These instabilities can grow into unphysical ocean states, resulting in failure of the model numerics. Several amelioration techniques are explored.