PRISM at Discovery Park

"A Framework for UQ: Calibration, Validation, and Computational Resource Allocation"

February 1 @ 12:30 PM - 1:30 PM - BRK 1001 - Guest speakers: You Ling and Joshua Mullins, Vanderbilt University

Guest speakers: You Ling and Joshua Mullins

Abstract: A comprehensive view of uncertainty quantification (UQ) includes several activities in order to account for various sources of errors and uncertainty in engineering systems. In this talk, we will first focus on two of these UQ activities: model calibration and model validation, and then discuss an efficient uncertainty propagation algorithm based on model selection.

In our approach, model calibration is performed using a Bayesian network built upon the Kennedy-O’Hagan framework, which includes a discrepancy term to account for inadequacy in the form of the physics model. The model discrepancy may include both bias and variance corrections, each of which may be either constant or varying with respect to the model inputs. 

In model validation, we assess the predictive capability of the model under uncertainty. Two model validation metrics will be introduced: Bayes factor and a reliability-based metric, each of which provides a clear probabilistic interpretation.

Propagating uncertainties to a system-level quantity of interest (QoI) can be very expensive when system models require a large amount of computation. To address this issue, we often utilize cheaper models in the form of mathematical surrogates or reduced-order and reduced-physics models. We discuss a procedure for efficiently selecting among the available options with the goal of accurately predicting the distribution of the QoI.

Bookmark and Share

Contact Details

More Information

«September 2014»
SuMTuWThFSa
1 2 3 4 5 6
7 8 9 10 11 12 13
14 15 16 17 18 19 20
21 22 23 24 25 26 27
28 29 30