A major challenge in the design of a coupled ocean-atmosphere data assimilation system is the existence of a wide range of growing instabilities. An effective data assimilation for the longer time scales has to be able to incorporate the slow instabilities of the ENSO background flow into, for example, the background error covariance, since the forecast errors will have a strong projection on these instabilities. Without a special effort to isolate the slow modes in a coupled data assimilation system, the faster but less relevant instabilities that dominate linear tangent models will wipe out the slower but important coupled processes from the estimated forecast and analysis errors (but not from the real analysis and forecast errors!).

To study whether it is possible to isolate the slow, coupled instabilities in the background flow, we have done experiments with breeding, a simple process that mimics ensemble data assimilation. We will present results of breeding in the NSIPP coupled ocean-atmosphere data assimilation system and in a perfect model coupled simulation. The crucial condition is that the interval for rescaling the ocean-atmosphere perturbations is large (one month) allowing atmospheric noise to saturate. The results are very encouraging and suggest that coupled data assimilation designed for seasonal and interannual prediction is feasible and could be based on a coupled Ensemble Kalman Filter using similarly long intervals between the coupled assimilation cycles. Simple experiments with a coupled Lorenz model confirm this “finite Lyapunov vector” approach should work.