Dependence of Flow Regime Properties on SST: Validation of large ensemble AGCM simulations and Reanalysis pdf

The dependence of the seasonal-mean atmospheric circulation on sea-surface temperature (SST) has been the primary concern of many studies on seasonal predictability. However, the usefulness of actual seasonal predictions also depends on the ability to simulate the dependence of intra-seasonal variability statistics on the SST forcing using coupled and uncoupled models.

Building on previously published work, we here study the properties of circulation regimes in the Pacific/North American region using results from a very large set of seasonal simulations with the COLA AGCM, encompassing a 55-member ensemble for each winter season from 1981/82 to 1998/99. Non-hierarchical cluster analysis is performed on low-pass-filtered fields of geopotential height from: (a) the full 18*55 experiment dataset, (b) 55 subsets including one member from each winter, and (c) NCEP/NCAR reanalysis data covering both the same 18-year period and the complete 54-year record. A quasi-stationary filter (first proposed by Toth) is important for increasing the statistical significance of the clusters, particularly for the reanalysis data.

The stability of the regime centroid patterns, and of the relationship between regime frequencies and SST patterns, is assessed by comparing results from the full dataset with those obtained from the 55 one-member-per-year sub-samples. Distributions of similarity indices from the subsamples are compared with indices derived from the comparison of full-sample and reanalysis results, in order to assess to what extent discrepancies can be ascribed to internal atmospheric variability. At least for the main circulation regimes, a good correspondence is found between simulations and re-analysis, and a dependence of regime frequency on SST anomalies in the tropical Pacific can be established.

Having validated the simulations, they are used to estimate whether the difference between the 18-year and 54-year reanalysis clusters is due to internal variability, or to the difference in frequency of SST patterns between the two periods. This is accomplished by computing the pdf of similarity indices of random GCM subsamples of length 54 (three GCM simulations for each winter) with equivalent subsamples of length 18 (one simulation per winter). The significance of the 18-year vs. 54-year cluster similarity index can be assessed using this pdf.