Two limits of decadal predictability in two generations of CCSM

A new component of the next IPCC assessment report will be decadal time-scale predictions in which models are initialized with states based on the observed state of the climate system. Since the climate system is chaotic, there is an inherent limit on the range at which the information in these initial states can have an impact on the skill of the forecasts. Actually in the AR5 projections there are two limits on the duration of initial condition influence. One is the classical limit corresponding to when the forecast distribution of states is indistinguishable from the climatological distribution. The other is the forecast range at which information from the initial conditions becomes small compared to the growing forced response to greenhouse gases.

In this presentation we will quantify these two limits for two versions of NCAR's Community Climate System Model, namely CCSM3 and CCSM4. We will concentrate on the predictability of upper ocean heat content in the two northern basins, and we will use two methodologies for estimating the two limits of predictability. First we will use a conventional perfect model integration method in which ensembles of integrations from (40) very similar initial conditions are employed. To make the comparison with the climatological distribution that is needed to measure predictability, we will employ relative entropy, which has several worthwhile properties. By comparing these ensemble forecasts with forecasts in which greenhouse gases are increased, we will find the range at which the forced response becomes large compared to the influence of the initial conditions. To date our results suggest that the limit of initial value influence is roughly a decade for CCSM3, but this influence becomes unimportant compared to the forced response a couple of years earlier than this.

The ensemble method is computationally expensive, so initial ensembles centered on only a few distinct climate states can be considered. To determine whether the limits given by the ensemble experiments are more generally applicable, we employ a second suite of approaches for estimating predictability limits from the statistics of model intrinsic variability. These include analog methods and the fitting of stochastic models to control integrations. These calculations together with the ensemble experiments indicate that when these models are used for decadal forecasts, predictions transition from being initial value problems to forced problems after no more than a decade. Also, the contrasting limits we find for CCSM3 and CCSM4 suggest the need for estimating the predictability of each individual model used in decadal prediction research.