Interpreting relationships between present-day fidelity and climate change projections (Invited)

Comprehensive models of the atmosphere have been used to estimate the equilibrium climate sensitivity for more than 30 years. These models have certainly improved over time: improved spatial resolution and more sophisticated representation of a wider range of physical processes have led to a greater ability to simulate the Earth’s present-day climate. The range of estimates of climate sensitivity hasn’t changed, though, which implies that simply making models better, or as much better as is currently possible, does not, in itself, lead to narrower estimates of long-term future change. In fact, it has so far proved impossible to identify a robust link between model fidelity, as measured by the agreement between observations and simulations of the present-day climate, and model response (with climate sensitivity being one example). In the absence of such a connection the range of climate sensitivity estimates cannot be further constrained with present-day models. But neither would the presence of a link between model skill and climate sensitivity in a given ensemble of models, in itself, be a guarantee that that link offers insight in nature’s behavior. This talk will discuss relationships between model fidelity and climate change response in two ensembles of climate models: one simple “perturbed-parameter” ensemble, in which a single model is used with varying values of closure parameters, and the more complicated multi-model CMIP3 ensemble on which the IPCC’s Fourth Assessment was based. The two ensembles have similar distributions of fidelity and climate sensitivity and the diversity of climate change responses within each ensemble is driven by the same phenomena. Carefully-crafted comparisons with present-day observations can be used to distinguish between high- and low-sensitivity models in the simpler ensemble. This signal is small relative to other sources of variation in global measures of skill, explaining why links between present-day performance and climate sensitivity are so hard to find. But even these tightly focused metrics fail to predict the behavior of the multi-model ensemble. This suggests that extrapolating relationships from multi-model ensembles to the natural world is not well-founded.