Decadal Predictions

The GEOS-5 AOGCM has been used for a suite of decadal predictions. The forecast suite consists of a 11-member ensemble initialized each 1 December from 1960 to 2009. The ensemble perturbations are generated using a breeding method with the first member being unperturbed and the other perturbations calculated based on a difference between each out of 2-11 ensembles and their mean. The rescaling norm was selected to be the root mean square difference of the monthly-mean top 500 m heat content averaged over the Atlantic region [90oW-20oE, 20o-70oN]. The rescaling interval is 5 years.

Model Configuration:

iODAS: EnOI with 50 EOFs every 5 days


The decadal re-analysis uses the same in-situ observations as the analysis used for seasonal forecast initialization, but the earlier period (prior to 1982) required some special care due to a) sparsity of available observations, b) inconsistency between observations of the same quantity from different instruments/processing. To deal with the former issue, a special so-called running climatology was constructed by binning all available in-situ observations within 3° bins and smoothing the resulting time series by applying a 10-yr sliding-window average. By assimilating this data the model's state was kept from drifting away from the observed climatological state. The latter issue manifested itself during the previous decadal re-analysis experiment, when the SST data stream switched from CMIP5 to Reynolds in 1982: the respective bias between the two data sets climatologies resulted in a sudden unrealistic jump in the re-analysis SST. To correct this problem, a special blended SST data set was computed: climatologies for both CMIP5 and Reynolds were calculated over an overlapping period (1971-2000) and then CMIP5 anomalies of the earlier (prior to 1982) period were superimposed onto the Reynolds climatology.