PANEL DISCUSSION: INTEGRATING DATA FROM MULTIPLE SIMULATION MODELS OF DIFFERENT FIDELITY

Computer models are used to simulate physical processes in almost all areas of science and engineering. A single evaluation of these computation models (or computer codes) can take as little as a few seconds or as long as weeks or months. In either case, experimenters use the model outputs to learn something about the physical system. In some settings, outputs from several computational models, with varying levels of fidelity, are available to researchers. In addition, observations from the physical system may also be in hand. In this panel discussion we address issues relating to model formulation, estimation, prediction and extrapolation using multi-fidelity computer models are addressed. In the first presentation, Bayesian methods are used to build a predictive model using low and high fidelity computational models with different inputs and also field observations. The second presentation deals with the difficult computational issues facing computer model calibration and prediction using a Bayesian framework that are typically remedied through the use of Markov Chain Monte Carlo techniques. While the computational burden is substantial, we review faster alternatives to standard MCMC techniques that are particularly useful in the multi-fidelity simulator problem. In the final presentation, calibration of computational models is discussed in the context of validation and extrapolation, with introduction to developments in stochastic model calibration.