Predictability of a Coupled Ocean-Atmosphere Model

The predictability and variability of a coupled ocean-atmosphere model has been investigated by examining the growth of small initial perturbations during the evolution of the coupled system. The ocean model is first integrated in a forced mode for a duration of over 24 years beginning with January 1964 in which wind stress forcing for each month is prescribed from the observations. This provides surrogate analysis or control run with which predictions from the coupled model can be initiated and compared. Starting from January 1970 with each of the next 181 initial states from the control run, a prediction experiment was carried out for a duration of 36 months each using the fully coupled model. With this large ensemble of prediction experiments, a detailed analysis of growth of initial error and forecast errors was carried out. The SST forecasts are compared with observations as well as the control run. The root-mean-square difference between control and forecasts becomes larger than the standard deviation of the control as well as persistence error in about three months. As a result of differences between the simulated SST in the control run and observations, the forecasts are forced to have initial errors that are comparable to the standard deviation of the observations. Some significant systematic errors in the model are also noted. There is an indication that the forecasts may be improved to some extent by averaging a few of the most recent available forecasts and removing the known systematic error. Also carried out is a large ensemble of identical twin experiments, each for a duration of 15 years. In one of each pair of experiments a small random perturbation is introduced at the initial time in the surface winds. These experiments have shown that the growth of small initial errors in the coupled model is governed by processes with two well-separated time scales. The fast time scale process introduces errors that have a doubling time of about 5 months, while the slow time scale process introduces errors that have a typical doubling time of about 15 months. The existence of a slow time scale gives us optimism about long-range forecasts of ENSO-type events. However, the fast growth rate tends to saturate at a level that is comparable to the climatological standard deviation. Thus, a key to long-range forecasting of ENSO-type events may lie in the ability to identify those initial states that are not too sensitive to the processes associated with fast growth rate. The diagnostic analysis shows that the first three empirical orthogonal functions (EOF) of the observed wind stress together explain only about 36% of the total variance. Although the observed wind stress has considerable amplitude in the higher EOFs, it is shown that only the first three components are important for forcing the observed interannual variations using this model. The atmospheric component of the coupled model is not able to simulate these large-scale components of the observed wind stress accurately. This is partly because the atmospheric model is mainly driven by the underlying sea surface temperature anomalies (SSTA) and partly due to the structural differences between the SSTA simulated by the model and the observed SSTA. Thus, a combination of the atmospheric component's tight coupling to the ocean and the ocean model's inability to simulate the SST anomalies correctly seems to be responsible for the rather rapid growth of prediction errors.