Principal Investigators: Charles Jackson and Mrinal Sen.

Funding Agency: NSF OCE-0415738
Start Date 9/1/2004
Expires 8/31/2008

project web site (password protected)

Abstract

A large disparity exists among various climate models in their prediction of global mean surface air temperature when atmospheric CO2 has doubled from present concentrations (figure 1). There are an overwhelming number of reasons why these differences could exist.  Although each climate model has been optimized to reproduce observational means, each model contains slightly different choices of model parameter values as well as different parameterizations of under-resolved physics. The need to understand the sources and impacts of these differences was recently emphasized within the 2001 Third Assessment Report of the Intergovernmental Panel on Climate Change that called for more quantitative evaluations of modeling uncertainty. [read section from IPCC report]

    There are good reasons why the climate modeling community is behind in this important endeavor. In order to quantify the uncertainty resulting from a realistic range of model configurations, one needs to estimate a multi-dimensional probability distribution that quantifies how likely different model parameter combinations are given knowledge of the uncertainties in our observations. The computational cost of mapping a multi-dimensional probability distribution for a climate model using traditional means (e.g. Monte-Carlo or Importance Sampling) is impractical requiring 10⁴ to 10⁶ model evaluations for problems involving less than ten parameters. 

Research Plans

    Over the past decade, there has been significant progress within the mathematical geophysics community for solving non-linear problems in geophysical inversion using statistical methods to account for the possibility of multiple solutions (interpretations) of geophysical data (for a review, see Barhen et al., 2000).  The Institute for Geophysics has been at the forefront of these efforts. Like climate models, geophysical models are complex, computationally expensive, and involve many potential degrees of freedom. Within the geophysics community, particular emphasis has been placed on efficiency, although with some measured compromises.  Similarly dramatic advances have taken place within the statistics community over the past decade on a class of methods of statistical inference known as Monte-Carlo Markov Chain (MCMC). In particular, greater awareness now exists for how different challenges in robust statistical inference can be addressed with a variety of sampling rules that obey the properties of a Markov chain. Our present goal is to advance the feasibility of quantifying climate model uncertainties that stem from multiple, non-linearly related parameters. We have submitted a proposal to NSF to continue the work on testing a variety of statistical sampling strategies for idealized and state-of-the-art climate models, including the most efficient sampling strategies that are found to work well for geophysics problems, while simultaneously reformulating the statistical foundation of these methods from a MCMC perspective to ensure that the estimated uncertainties are robust and are well suited for the class of problems that climate models represent.

We hope to address these questions:

1. How does uncertainty in parameters affecting clouds, convection, and radiation affect a climate model’s sensitivity to CO2 forcing? 

   • How well do observational data constrain parameter values? 

   • What physical processes create differences in sensitivity?

2. What is the potential for reducing systematic biases between model predictions and observations through careful selection of parameter values? 

3. What sampling strategy provides the most efficient and robust strategy for estimating climate model parameter uncertainties?