How well can historical temperature observations constrain climate sensitivity

Future climate projections are strongly influenced by climate sensitivity (CS). Many re- cent studies estimated CS by combining runs of Earth Models of Intermediate Complexity (EMICs) with global mean instrumental observations. Yet, CS estimates remain consistently uncertain. This dissertation addresses four questions: (1) What is the probability distribution function (pdf) of CS implied by global mean surface temperatures and upper ocean warming when a model with full three-dimensional ocean dynamics is used? (2) How sensitive is this pdf to priors? (3) How does the CS estimation uncertainty depend on the true CS of the climate system? and (4) How strongly is the CS uncertainty affected by internal climate variability that is not resolved by the model? These questions are addressed with a 250-member ensemble of UVic ESCM climate model runs varying CS, background vertical mixing in the ocean, and anthropogenic sulfate cooling effects. A Gaussian Process emulator is developed to interpolate UVic ESCM output between the ensemble parameter settings. The emulator is constrained with historical observations of temperature and upper ocean warming using a Bayesian Markov chain Monte Carlo method. The results are combined with prior independent evidence. The method results in the 95% posterior credible interval for CS from 1.8°C to 4.9°C, with the mode of 2.8°C. The CS pdf depends strongly on the CS prior. When the prior evidence is discarded, the 95% interval upper bound increases to 10.2°C. In addition, multiple observation system simulation experiments are performed. First, pseudo-observations of temperature and upper ocean warming are simulated from the emulator at a specified "true" CS. Then, CS is re-estimated using the pseudo-observations and the emulator, and compared to the "true" value. It is found that CS is more difficult to estimate when it is high. This suggests that a high CS might get undetected when relying on the recent instrumental record alone. Unresolved internal climate variability represents a key uncertainty in CS estimates from global observations and EMICs. Using a single observational record, contaminated by internal variability, can result in a sizeable discrepancy (up to several .C) between a CS estimate and the "true" value.