WHAT IS THE SKILL OF CLIMATE PARAMETER ESTIMATION METHODS? A CASE STUDY WITH GLOBAL AVERAGE OBSERVATIONAL CONSTRAINTS.
Open Access
- Graduate Program:
- Geosciences
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- June 01, 2010
- Committee Members:
- Klaus Keller, Thesis Advisor
- Keywords:
- Climate sensitivity
- Bayesian methods
- climate models
- ocean diffusivity
- climate change
- Abstract:
- Future climate model projections are deeply uncertain. One key driver of this uncertainty if the uncertainty in model parameters, such as climate sensitivity. Recent studies have used suite of model runs together with observations to estimate these key parameters. In these studies, Markov Chain Monte Carlo (MCMC) methods are often employed to obtain posterior probability distributions for the parameters. Despite ubiquitous use of such methods, their skill at recovering known parameter values has not been thoroughly evaluated. This study quantifies the skill of an MCMC parameter estimation method to recover true parameter values using pseudo-observations generated from an University of Victoria Earth System Climate Model (UVic ESCM). Specifically, the work addresses several key questions. First, what is the effect of reducing the combined model and observational error on the skill of the method? Second, what is the skill of different pseudo-observations to constrain model parameters? Third, what is the effect of random realizations of the combined model and observational error on results of the parameter estimation? I first run an ensemble of UVic ESCM model runs spanning the last two centuries. I vary the parameters of climate sensitivity, background vertical ocean diffusivity and strength of effects of anthropogenic aerosols between the ensemble members. I then implement a simple MCMC method to estimate the parameters using global observations of temperature and upper ocean heat content. The inversion accounts for uncertainty in the statistical properties of the potentially correlated model-data residuals, and reduces biases due to sparse sampling of the parameter space using emulation. I perform a set of perfect model experiments using pseudo-observations of globally average temperature and ocean heat content, derived from the UVic model using various assumptions about the observational and model error. I show that at current estimates of combined model and observational error the results of parameter estimation hinge critically on random realizations of the combined error process, but that the skill of the method increases rapidly if the combined error decreases. Using both temperature and ocean heat uptake observations improves the skill of the method compared to cases where only individual observations are used, except for the case of background vertical ocean diffusivity at current combined error estimates. Implications of the results for parameter estimation work are discussed and strategies for future research are outlined. In addition, I compare the the probabilistic UVic ESCM hindcasts of global average near-surface temperature with those from more complex General Circulation Models (GCMs). I show that a well calibrated intermediate complexity model such as UVic ESCM can perform comparably to GCMs in terms of skill at reproducing global mean historical temperature observations.