[Book Title], Edited by [Editor’s Name]

Assessments of uncertainty in climate prediction require a meaningful ensemble of model configurations that represent the limitation of observations to constrain arbitrary choices in model formulation such as the values assigned to its multiple, non-linearly related parameters. We present a data directed importance sampling strategy that makes optimal use of distributed computing resources and improves the sampling efficiency over uniform or random sampling strategies by several orders of magnitude. The algorithm is then further developed to take into account uncertainties in our ability to establish a Wiley STM / Editor: Advanced Computational Infrastructures for Parallel and Distributed Adaptive Applications, Chapter 6 / Authors Jackson, Sen, Stoffa, Huerta / filename: ch6.doc page 2 normalized metric of model-data distance owing to the correlations that exist in space and time among the 15 fields that are used to evaluate model skill. By incorporating estimates of this uncertainty in the prior, further improvements can be attained in the accuracy and efficiency of the stochastic sampling algorithm. The algorithm is then applied to estimating an a posteriori probability density for 6 parameters important to clouds and convection of the Community Atmosphere Model version 3.1. The top six performing parameter sets improved model skill by 10% with nearly identical skill scores, but for different reasons related to the wide range of selected parameter values. Efficiently quantifying these uncertainties will provide important insights into the limitations of climate model predictive skill within a time frame practical for climate model development.