Bayesian Inversion of Climate Models
Emmet Cleary, California Institute of Technology
Climate predictions contain large uncertainties, in part due to uncertainties in model parameters that arise in closure approximations. Generally, these parameters are tuned manually. However, with simpler computer models, Bayesian inversion is a popular approach to parameter estimation. Bayesian inversion yields a posterior distribution that links mismatches between model output and data to uncertainties in model parameters. But common Bayesian inversion tools, such as Markov chain Monte Carlo (MCMC), are not easily applicable to climate models because they require more forward model evaluations (about 10<sup>5</sup>-10<sup>6</sup>) than are feasible with computationally expensive models. This presentation demonstrates the use of ensemble Kalman inversion (EKI) techniques to estimate parameters in a moist convection scheme of a General Circulation Model. We train a Gaussian Process emulator during the iterations of the EKI and then use standard MCMC algorithms on the emulator to estimate posterior uncertainties. What results is a computationally efficient approximate Bayesian inversion approach that yields accurate results in practice.
Abstract Author(s): Emmet Cleary, Alfredo Garbuno Inigo, Tapio Schneider, Andrew Stuart