|
Abstract:
|
Models of complex systems often produce large quantities of outputs that can be compared against actual system data, in order to learn about the model-parameters ('calibration'). Where the model is expensive to evaluate, this process typically requires the use of an emulator, to augment the limited number of model-evaluations with additional judgments about the model (e.g., about smoothness, limiting behavior, nonlinearity and interactions, and monotonicity). The outer-product emulator (Rougier, 2008, JCGS) is a new emulation framework particularly suited to large problems, based on a conditionally Gaussian process. Here it is used to calibrate three parameters of the TIE-GCM model of global high-atmospheric processes to observational data on diurnal electron drift, collected at a small number sites around the world.
|
