Climate as a Chaotic System
The ability of atmosphere-ocean GCMs to predict the climatic effects of human alterations of greenhouse gases and other factors cannot be tested directly with respect to a point in time a hundred years in the future. However, it is still possible to ask—and determine—whether those models can in principle make such predictions to a reasonable degree of accuracy. One way to evaluate this ability is to consider the effects of errors in system initial values. If a system is well-behaved, small initial errors will lead to small future errors, or even damped responses. In a chaotic system, on the other hand, small initial errors will cause trajectories to diverge over time; and for such a system (or model), true predictability is low to nonexistent. In a study addressing initial value errors, Collins (2002) used the HadCM3 model, the output of which at a given date was used as the initial condition for multiple runs in which slight perturbations of the initial data were used to assess the effect of a lack of perfect starting information, as can often occur in the real world. The results of the various experimental runs were then compared to those of the initial control run, assuming the degree of correlation of the results of each perturbed run with those of the initial run is a measure of predictability. As a result of these operations, Collins found “annual mean global temperatures are potentially predictable one year in advance” and “longer time averages are also marginally predictable five to ten years in advance.” In the case of ocean basin sea surface temperatures, it was additionally found that coarse-scale predictability ranges from one year to several years. But for land surface air temperature and precipitation, and for the highly populated northern land regions, Collin concludes, “there is very little sign of any average potential predictability beyond seasonal lead times.”
Collins, M. 2002. Climate predictability on interannual to decadal time scales: the initial value problem. Climate Dynamics 19: 671–692.