Diagnosing Climate Sensitivity
I have been working on devising tests of climate model fidelity to data. The goal here is to discover statistical properties of climate models that are correlated with the models' climate sensitivity, but can be observed using climate data sets of limited duration. Intuitively one expects that such measures should exist: after all the climate feedbacks that will determine climate sensitivity just involve processes (cloud formation, sea ice formation and melting, changes in the humidity of the atmosphere) that happen all the time in the real atmosphere over a huge range of temperatures and other conditions that dwarfs the expected change in temperature due global warming.
As it turns out, the problem is extremely challenging. It turns out that when one examines the output of the climate models submitted to the IPCC, and compares various properties of the models, one does not easily come up with measures whose variation across the models correlates with the variation in climate sensitivty. The situation is even worse than that: different definitions of climate sensitivity are not even well correlated among the models. For instance, consider the following two sets of experiments: one in which a set of climate models is run for thirty years with carbon dioxide concentrations increasing by one percent per year, and another in which the models are run for a century with a rising concentration of carbon dioxide, and then run for another century to come into an approimate equilibrium. It is not even true that the models which warm the most in the first set of experiments will warm the most in the second.
My work in this area has thus far involved understanding the limits of statistics based on the lag-autocorrelation of temperature data to diagnose climate sensitivity. A system with larger climate sensitivity should have longer memory---if you push it away from equilibrium, positive feedbacks tend to reinforce the push, and extend the time the system stays out of equilibrium. Using a very simplified climate model I've been able to show how models with more than a single heat capacity can fool the analyst, introducing substantial biases into estimate of climate sensitivity from observed climate fluctuations.