Charles Curry, Bruno Sanso and Chris Forest

12/31/2005 09:00 AM

Applied Mathematics & Statistics

Modern climate system models make use of numerous parameters to control their behavior. By matching model outputs with observed data we can perform inference over such parameters. Our work extends previous related analyses through the application of model interpolation, Markov-chain Monte Carlo simulation and model selection. The analysis is focused on three input parameters: climate sensitivity, rate of deep ocean heat uptake, and net aerosol forcing. Climate model output was sampled on a non-uniform grid over the 3D parameter space. We interpolate the climate model output over the empty regions of these grids, providing an approximation to the continuous climate model response for use in a likelihood function. The covariance matrix that is required by the likelihood is estimated from general circulation model output. Several model selection criteria are applied to estimate the number of non zero eigenvalues of such matrix. We find significant posterior uncertainty in all three climate model properties. Posterior parameter distributions differ according to the choice of climate model, likelihood and covariance estimate.