Insights from large ensembles with perturbed physics

Abstract

State-of-the-art climate models are characterized by substantial deficiencies and their projections for the 21st century differ considerably depending on their specific initial conditions and physical parameterizations. Perturbed physics ensembles represent a promising tool to delimit the range of model uncertainty in a probabilistic sense. Yet large ensembles with complex climate models are still constrained by available computer resources. Here, an extended energy balance model with a stochastic weather term, feedback, and natural and anthropogenic forcings is used to investigate the potential of large perturbed physics ensembles. Up to 100,000 simulations for the 20th and 21st centuries are realized, assembling appropriate random numbers for 9 different model parameters. Bayesian model averaging is applied to filter the simulations in the face of observed temperatures and measurement error. The resulting ensemble mean has learned noticeably from the data and is very close to the observed time series of global-mean temperature. It differs markedly from the CMIP3 and CMIP5 multi-model ensemble means, especially in terms of more realistic decadal temperature variations due to the learning effect by means of the Bayesian approach. The sensitivity of the model and the Bayesian filter is assessed with respect to the model design and the estimated observational error, respectively.