The realism of climate models used to estimate the sensitivity of climate to increases in infrared-absorbing gases has improved slowly. From the earliest calculation by Arrhenius of the global average warming produced by a step-function increase in atmospheric carbon dioxide to the fully three dimensional models that include coupled, circulating oceans and a gradual increase of atmospheric carbon dioxide, each step has produced simulations that more nearly replicate today's climate and that are more useful in explaining past climate changes. In the mean time, progress in the assembly of paleoclimate data sets has allowed a variety of comparisons between models and observations, and these comparisons have generally increased the confidence of climate scientists that they are on the right track [ IPCC, 1992]. Even so, models constructed by different groups continue to show different rates of climate warming for similar increases in forcing.
The improvements that have occurred in global models have not appreciably lessened the central problem of those wishing to use models as inputs to impact studies. Impacts take place on a local or regional scale, and modelers have so far been able to give but little help in describing the local manifestations of a global, human-induced climate change. The calculation grid of the models--limited both by computer size and by lack of sufficient understanding of small-scale processes--has been too large to allow reasonable interpolation to local values, and even the values computed on the coarse grid differ from observed averaged values in several regions. Impact studies also frequently require usefully realistic precipitation scenarios, and so far models have not produced them. Finally, studies have shown that some agricultural crops and other climate-sensitive systems are more sensitive to weather extremes than to changes in average conditions. Katz and Brown [1992] have also shown that the probability of an extreme event is relatively much more sensitive to variability than to changes in the mean. So models are now being called upon to simulate variability in addition to average conditions and have so far achieved only limited success. Mearns [1993] has discussed the role of both mean and variability of climate in the production of droughts and the ability of current climate models to simulate these conditions. See also Schimel and Sulzman, Variability in the Earth System: Decadal and Longer Timescales, in this issue.
In response to these difficulties a number of climate scenarios have been developed that are consistent with a global model projection of warming but which supply missing regional detail either from knowledge of the local climate under consideration [ Robock et al., 1993], or from a stochastic daily weather simulation model [ Reed and Desanker, 1992; Wilks, 1992]. There also has been progress in imbedding a fine-mesh regional model within a coarse-grid global model to improved the simulation of the climate in a single region [ Giorgi and Mearns, 1991; Marinucci and Giorgi, 1992; Giorgi et al., 1994], and in the empirical downscaling approach [ Giorgi and Mearns, 1991].
In addition to the limitations described above, there remains a persistent worry that models and other scenario development techniques do not adequately allow for the possibility of a major surprise, such as a large shift of an ocean current or the onset of a positive-feedback situation that would destroy a large forest. Such a surprise would likely have rather different and more severe impacts than a gradual change, but so far no modeling or other approach has been developed to estimate the probability of such an event.