Uncertainty Quantification Working Group
May 15, 11:30 AM, CNLS Conf. Room, TA-3, Bldg. 1690

Estimating the error and modeling uncertainty in simulation prediction over a domain of validation

Francois M. Hemez, ESA-WR

Abstract

This study addresses the assessment of accuracy of simulation predictions. A procedure is developed to validate a simple non-linear model defined to capture the hardening behavior of a foam material subjected to a short-duration transient impact. Validation means that the predictive accuracy of the model must be established, not just in the vicinity of a single testing condition, but for all settings or configurations of the system. The modelís predictive accuracy is established by training a metamodel of prediction error. The prediction error is not assumed to be systematic. Instead, it depends on which configuration of the system is analyzed. The study shows how predictive accuracy can be assessed even in the presence of a calibrated model. Finally, the prediction errorís confidence bounds are estimated by propagating the uncertainty associated with specific modeling assumptions.

Return to UQWG meetings