Statistical Skill in the Emulation of Climate Models
Masters Thesis (2014)
Department of Meteorology, The Pennsylvania State University
A climate model of low complexity can be used to emulate the performance of one with higher complexity by identifying the parameters for each model that yield similar model responses. In this thesis, an energy balance model, the Diffusive Ocean Energy balance CLIMate model (DOECLIM) was calibrated to match the output from 639 simulations of the MIT Integrated Global System Model (IGSM), where each IGSM simulation has a different set of values for three key climate parameters: climate sensitivity, vertical ocean diffusivity and aerosol forcing. The energy balance model estimates the globally averaged climate state based on simplified model physics. The IGSM estimates the zonal mean state of the atmosphere and ocean based on physics in higher complexity climate models, and estimates climate changes that include significant internal variability. The DOECLIM parameters were estimated for each IGSM run to find the parameter settings for the simpler model that would best match results from the more complex model. This allows for the simpler model to be used as an emulator over a range of parameter settings. Two model calibration techniques were used and compared. These techniques are Differential Evolution (DE), a genetic algorithm that produces a single set of parameters providing best fit values, and Markov Chain Monte Carlo (MCMC), which produces a joint probability distribution for the parameters. The study analyzed the statistical skill, including potential biases, that exist when calibrating the energy balance model to IGSM output. In particular, the estimated DOECLIM climate sensitivity values tended to be lower than their corresponding IGSM values, particularly for runs with low ocean diffusivity. The parameter estimates also vary depending on the choice of noise model, AR(1) or AR(0) for the atmosphere and ocean temperatures.