Abstract:

The objective of this talk is to describe a Bayesian hierarchical spatiotemporal statistical model developed in the context of glaciology, which is extensible to other physical systems. A feature of this model is to represent the discrepancy between the output of a computer simulator and the real physical process with a multivariate random walk. In addition to the model, I will present a few ways to make Bayesian inference computationally feasible. In particular, these ways include using bandwidth limited linearalgebraic routines, the use of an approximation to the likelihood, and the use of firstorder emulation of a computer simulator (e.g., a numerical partial differential equation solver). The model assumptions, in addition to the speed and accuracy of computational methods used, are assessed using a test system from glaciology; as such, a discussion of Icelandic glaciers and the physics of glaciers will also be included.
