Adaptive estimation of climate model closure parameters

Abstract

International climate policy is critically dependent on the future climate simulations. While the climate models are becoming increasingly accurate, inh erent uncertainties remain. One of these is the fact that all models contain some free (closure) parameters related to the modelling of sub-grid scale proc esses (such as clouds or turbulence). Optimal closure parameter values are approximately known but their uncertainties are not well understood. Generally, it is very hard to determine the parameter values based on observations. In the ADAEST-project, a computational parameter estimation method, Markov chain Monte Carlo (MCMC), is applied to estimate the closure parameter values and their uncertainties for the ECHAM5 atmospheric general circulation model. We wi ll apply modern adaptive MCMC sampling techniques, developed by the proposing team. The cost function to be minimized consists of the squared differences b etween the dominant modes of the observed and simulated climate variability. The MCMC analysis interprets this cost function as a Bayesian likelihood funct ion. In the Markov chain, a large number of climate simulations are performed, making the approach of the ADAEST-project computationally very demanding. Th e adaptive sampling technique implies however large computational savings. Thanks to the DEISA resources, significant scientific and computational advances can be obtained in the research of the reliability of climate predictions.