Statistical approaches for parameter estimation in climate models
To quantify the uncertainties arising in climate prediction it is necessary to estimate a multidimensional probablility distribution. This is know as the calibration problem. The computational cost of evaluating such a probability distribution for a climate model is impractical using traditional methods such as Gibbs/Metropolis algorithms. This talk will describe an optimization based method that has been applied for non-linear problems in geophysics and that is currently in use to calibrate parameters of an atmospheric general circulation model (ACGM).
Furthermore, we will also consider adaptive Monte Carlo based methods in the context of a climate model that is able to approximate the noise and response behavior of the AGCM. Comparisons and efficiency evaluations between approaches will be made. Another aspect of this talk is to overview the current role of spatical methods in providing emulators to climate model output and reducing computational burden. In particular we will discuss the use of Gaussian process (GP) in this context andon potential limitations and challenges for these methods.