The glaciological, climate, and earth system modelling communities
have been slow to incorporate, in any statistically self-consistent
way, the objective determination of model and data uncertainties into
their results. Though ensemble calculations offer a first step,
statistically self-consistency requires the propagation of model
parameter and constraint data uncertainties into the ensemble results.
Bayesian model calibration addresses this key issue. I'll describe a
Bayesian framework for model calibration based on a combination of
artificial neural networks and Markov Chain Monte Carlo methods. The
calibration provides a posterior distribution for model parameters
(and thereby in my case modelled glacial histories) given
observational constraint data sets. This methodology therefore also
takes into account constraint data uncertainty. This approach is
highly applicable to cluster computing environments in which one can
generate order 100 model runs per month of real time with dozens of
ensemble parameters. It also allows the incorporation of large and
diverse sets of constraint data into the calibration procedure and is
suitable for complex non-linear models. The presentation will focus
on the on-going issues and challenges encountered in the real-world
application of this methodology to the calibration of the 3D MUN/UofT
Glacial Systems Model. |