Intellectual Merit: This study will develop new efficient methods to estimate climate model prediction uncertainties stemming from the combined influence of multiple, non-linearly related parameters and explain the large disparity that exists among climate models in their response to projected increases in atmospheric CO2 concentration. Initial attempts to quantify parameter uncertainties will use Bayesian Stochastic Inversion based on Multiple Very Fast Simulated Annealing (Multiple VFSA), a stochastic importance sampling technique developed and successfully applied to the interpretation of geophysical data. In parallel, stochastic sampling strategies will be reviewed from a Monte-Carlo Markov Chain (MCMC) perspective to advance this theory's applicability to problems that are currently limited by the number of samples that may be feasible to consider either because of computational cost (e.g. climate models) or the size of data sets (e.g. interpretations of geophysical data). Specific objectives include 1) estimating a multidimensional probability distribution for a set of parameters related to clouds, convection, and radiation within the latest version of CAM2 (the climate model developed at NCAR) that are significant sources of climate model prediction uncertainty, 2) evaluate sampling biases within Multiple VFSA and prototype new efficient and robust methods of statistical inference from a MCMC perspective, 3) the calculation of an ensemble of doubled CO2 experiments with a coupled atmosphere-mixed layer ocean model that is representative of the combined uncertainty in these parameters, and 4) evaluate the climate processes that explain differences in model response to changes in CO2 forcing.
Broader Impacts: Quantifying the main sources of climate model uncertainty is the first step toward reducing differences between different models forced by the same inputs. There is potentially useful data from both the instrumental record as well as paleoclimate archives that have not been fully exploited. The proposed research provides a methodological foundation for using these data sources to address these objectives. The results from the proposed research will demonstrate for the first time the feasibility of exploring atmospheric GCM parameter space in a comprehensive and systematic manner. With current inexpensive but powerful distributed computing architectures and the ability to automatically navigate parameter space, parameter choices can be more easily optimized for model performance. The proposed research activity will also contribute to the training and support of two graduate students.
