Title: Calibration under Uncertainty of Expensive Computer Simulations with Multivariate Output

Speaker: John McFarland, Vanderbilt University

Date/Time: Thursday, August 9, 2007, 1:00-2:00 pm

Location: Building 880, Room B30A (Sandia/NM)

Brief Abstract:  The use of complex simulation models is continuing to increase in prevalence within the scientific community. These simulations are often used for prediction, parameter studies, and high-consequence decision making, and may be characterized by a large number of input parameters, long run times, and high-dimensional output. In many cases, experiments may be conducted under a set of "moderate" conditions, and such results may be compared against the corresponding model predictions. When these experimental data are used to improve the predictive capability of the model (perhaps by making inference about internal model parameters), we call this model calibration. The work presented here explores model calibration for realistic engineering applications, with an emphasis on taking a comprehensive and understandable account of various uncertainties present using the methods of Bayesian analysis. I explore how Gaussian process surrogate models can be used effectively as representations of expensive simulations, particularly when the response may be a function of time and/or space (multivariate output). I also discuss how characterized experimental uncertainty can be included along with other uncharacterized experimental and modeling uncertainties. In addition, I illustrate a method by which prescribed uncertainties at the modeling level can be accounted for in the calibration under uncertainty process. The complete methodology is thoroughly illustrated for a Calore thermal simulation of a foam component, using a comprehensive database of experimentally observed response values over time and space.

CSRI POC: Laura Painton Swiler , (505) 844-8093