Impact of initial conditions on seasonal simulations with an atmospheric general circulation model

Many previous studies have examined the use of very long integrations of atmospheric general circulation models (AGCMs) forced by observed sea surface temperatures (SSTs) as proxies for seasonal atmospheric predictions. These long simulations explore a boundary-value problem in which significant deviations from the model's long-term climatology must be a result of the SST forcing. Seasonal lead simulations starting with observed initial conditions (ICs) for the atmosphere and land surface while retaining observed SST forcing are an intermediate step between the pure boundary-value problem and the pure initial-value forecast problem in which SSTs are also predicted. As part of the Dynamical Seasonal Prediction (DSP) experiment, an ensemble of AGCM integrations with observed atmospheric ICs and model climatology land surface ICs was integrated from mid-December through March for 16 years. These DSP simulation ensembles are compared to ensembles of long boundary-value simulations from the same AGCM in a perfect-model setting (no comparisons of simulations to observations are attempted). Significant differences must be due to the impact of the DSP ICs. Surprisingly large and long-lived differences are found in both the mean and the variance of the ensembles. Many appear to occur because the ICs of the DSP runs are inconsistent with the AGCM climatology; an extended period of model 'spin-up' is the result. Some differences are related to local impacts of the land surface ICs while others, like shifts in the distribution of tropical precipitation and a cooling of the northern hemisphere, are less obviously related to the ICs. The results suggest that care will be needed when inserting observed ICs into seasonal predictions in order to avoid the long-term effects of model spin-up.