Uncertainty Quantification in Climate Prediction by xgi59866

VIEWS: 48 PAGES: 26

									Uncertainty Quantification
in Climate Prediction
                      Charles Jackson (1)
                        Mrinal Sen (1)
                       Gabriel Huerta (2)
                         Yi Deng (1)
                       Ken Bowman (3)

(1) Institute for Geophysics, The University of Texas at Austin
(2) Department of Mathematics and Statistics, University of New
    Mexico
(3) Department of Atmospheric Science, Texas A&M University
(IPCC 2001)
Surface air temperature




       clouds




                 (AchutaRao et al., 2004)
Where can clouds go wrong?
Are current approaches to climate
model development convergent?

Address question using:
• Bayesian inference
• Stochastic sampling
  – Simulated annealing to focus sampling
  – Multiple search attempts for uncertainties
Posterior probability density for
        3 parameters:
     MVFSA
      Grid Search


        Metropolis




                      Metropolis


                     MVFSA
Target: Match observed climate
1990-2001

One 11-year climate model
integration takes 11 hours over 64
processors of an Intel-based
compute cluster.
                Results
• Analysis of top six performing model
  configurations
• T42 CAM3.1, forced by observed SST
  March 1990 to February 2001.
• ~400 experiments completed (so far).
Histogram of configurations
     with Improved skill
Convergence in predictions of global
 means does not imply predictions
          are correct.
Much improved simulation of
 rain intensities over tropics.
            climateprediction.net




27,000 experiments completed in past year on 10,000 personal computers
(Stainforth et al., Nature 2005)
              Conclusions
• Stochastic optimization of CAM3.1
  suggests the model may provide
  convergent results of global mean
  predictions.
  – Assumes parameters tested are key sources
    of uncertainty.
  – Hadley Center model supports inference.
  – Unanticipated gains in model skill.

• Important differences at regional scales
  remain.
Each parameter affects many aspects of the model
   There are multiple ways to combine
parameter values to yield better model skill.
Definition of model-observational data mismatch



                                                        
                N
                
                      1
      E (m)            (dobs  g (m))T C1(dobs  g (m)) i
                i 1
                     2N


                                    K
                (dobs  g (m))      a j  EOF j
                                    j 1


                                N   K a2 
                                          j 
                                
                           1
                  E (m) 
                          2 N i 1  j 1 2 
                                          j i
Villagran-Hernandez et al. (in prep)

								
To top