Calibration of Computer Models with Informative Failures

• Activity Number: 69
• Type: Topic Contributed
• Date/Time: Sunday, August 9, 2015 : 4:00 PM to 5:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Peter Marcy* and Curtis Storlie
• Companies: Los Alamos National Laboratory and Los Alamos National Laboratory
• Keywords: computer experiment ; Gaussian process ; emulator ; discrepancy ; latent process ; uncertainty quantification

Abstract

Gaussian process emulators are widely used to calibrate deterministic computer codes (simulators) to experimental/field data. The goal of such an analysis is to determine which values of the physical parameters used in the computer model are most consistent with the experimental observations. However, there may be incomplete simulated data as some simulators can fail to produce output for particular combinations of the inputs. Under the assumption that the failed model runs correspond to regions of the parameter space which are not physically feasible, the missing data can be incorporated into a Bayesian calibration routine. In this talk I detail the procedure and then illustrate it using a computational fluid dynamics model for carbon capture.