Seasonal ensemble prediction with a coupled ocean-atmosphere model

Abstract

nsemble prediction methods, in which the control initial conditions are perturbed by coupled ocean-atmosphere non-linear instabilities obtained by a breeding method, have been applied within a coupled ocean-atmosphere model with prognostic primitive equation atmospheric and oceanic components. The bred vectors have a distinct annual cycle of growth rates with the maximum occurring during boreal winter and a minimum during boreal summer. Bred vector amplitudes peak in boreal spring and have a minimum in boreal autumn. The leading empirical orthogonal functions of the 50 m ocean temperature of bred vectors have maxima in the equatorial Pacific between 120 – 150°W, as well as in the western Pacific, while the associated 500 hPa streamfunction fields have large-scale wave-trains in extratropical regions that are strongly influenced by ENSO. Coupled control and ensemble forecasts for one year have been initiated on the first day of each month during a period in the 1990s to examine the sensitivity of enhanced ensemble mean skill, compared with the control, on the number and type of ensemble perturbations. The focus has been on the skill of predicting the 50 m ocean temperature in the equatorial region. For forecasts longer than about two months the root mean square errors in the ensemble mean forecasts are smaller than for the control based on averages over all the forecasts. There is considerable variability in the skill of both control and ensemble forecasts during the 1990s. In particular, forecasts through the 1997 El Niño and 1998 La Niña regime transitions tend to be less skilful than in more quiescent periods. Forecast errors tend to increase with time as expected and there are peaks in error amplitudes for forecasts verifying in boreal spring. Forecast skill increases with increasing numbers of bred vectors but saturates with little additional gain in using 64 members compared with 32. Ensemble forecasts with cyclic mode perturbations, the non-linear generalisations of finite time normal modes, are found to be more skilful than bred vectors, with eight cyclic modes producing similar error reduction in three to nine-month forecasts as 32 to 64 bred vectors. Our results suggest that a contributing cause of the boreal spring predictability barrier is the fact that large-scale atmospheric teleconnection patterns and instabilities peak in boreal spring and in turn couple to the ocean.