final_program_abstracts[1]

11 IMSC Session Program
Artificial neural network assisted Bayesian calibration of the
planet simulator general circulation model
Monday - Poster Session 9
T. Hauser, L. Tarasov, A. Keats , E. Demirov
Memorial University of Newfoundland, St. John's, Canada
Earth systems models that attempt to make long-term climate predictions are sensitive
to the approximations they employ. These approximations crucially depend upon
model parameters whose values and uncertainties ought to be defined using objective
and repeatable methods. In this study we approach this problem by using
observational data to generate Bayesian posterior probability distributions for the
model parameters. This allows us to determine high-probability parameter values
along with their credible intervals, and accounts for the observational uncertainties
related to the calibration data. However, for complex climate models, evaluating the
posterior can require a prohibitive degree of computational expense. In the
experiments presented here, artificial neural networks (ANNs) are trained with output
from a general circulation model (GCM) to emulate the model response to different
parameter sets. The ANNs are used as surrogate models to allow a computationally
efficient Markov chain Monte Carlo (MCMC) sampling of the Bayesian posterior of
the GCM calibrated against seasonal climatologies of temperature, pressure, and
humidity. To reduce complexity, for these initial investigations we vary only five
model parameters, which influence temperature and radiation transport. We validate
the methodology with the results of a calibration against a default model run with
added noise. These experiments serve as benchmark tests, and allow us to determine
sensitivity to noise in the constraint data. Using observational climatologies, we also
examine calibration sensitivity to the spatial resolution of the constraint data.
Increasing the number of data points used eventually increases the complexity of the
emulation problem, to the point where the ANNs are no longer accurate enough to
effectively direct the MCMC sampling. To reduce the complexity of the emulation we
experiment with using constraint data produced by empirical orthogonal function
(EOF) analysis of the observations. Finally, we summarize remaining issues to
address in order to create a fully validated operational methodology for objective
model calibration.
Abstracts 47