MCMC for (too?) large climate models

Document Type

Presentation Abstract

Presentation Date

12-7-2009

Abstract

New tools for computational statistics have emerged that enable proper reliability analysis of nonlinear models, e.g., the MCMC (Markov chain Monte Carlo) approach is widely used. Typically, the sampling-based methods require a large number of model evaluations. The standard use of the approach is thus limited to models of low computational cost.

Here, we discuss the seemingly intractable task of applying MCMC methods for large models with really high computational cost. Climate models provide an important case. The equations are expressed using a computational grid. However, the grid is insufficient for accurate computation of many important physical processes, such as formation of clouds and their interaction with solar radiation. These processes operate in scales much smaller than the grid interval. Parametrizations of these sub-grid scale processes leads to a closure problem where some free parameters necessarily appear. The climate simulation results thus depend on the specified values of the closure parameters, sometimes called the 'tuning parameters' of the models. We discuss the possibility of employing MCMC to analyze the sensitivity of the climate predictions with respect to these parameters. The challenges include high CPU times, short sampling chains, grid size effects, proper formulation of the likelihood. Solutions may be provided by parallel chains, use of databanks, various approximation methods, surrogate models, and early rejection diagnostics in MCMC sampling.

Additional Details

Monday, 7 December 2009
3:10 p.m. in Math 103
4:00 p.m. Refreshments in Math Lounge 109

This document is currently not available here.

Share

COinS