Segmented Bayesian Calibration of Multidisciplinary Models

This paper investigates Bayesian model calibration for multidisciplinary problems that involve several disciplinary models and multiple sources of data regarding individual and combined physics. Asegmented approach is explored as an alternative to simultaneous calibration of the parameters and discrepancy terms of all the component models. Simultaneous Bayesian calibration requires conducting inference on all uncertain parameters using all models and data concurrently. This can lead to significant computational burden and ambiguity regarding each individual model's contribution to the overall prediction uncertainty. Segmented Bayesian model calibration is first investigated with two illustrative mathematical examples and the performance of this strategy is examined for different characteristics of the problem (i.e., model dependence and data availability). The Kullback-Leibler divergence and the Bayes factor metric are used to compare the computational effort and accuracy of the segmented and simultaneous calibration strategies. The segmented approach is observed to yield comparable prediction uncertainty with fewer samples than simultaneous calibration for the multidisciplinary scenarios considered. The strategies are then applied to the estimation of model discrepancy in aerodynamic pressure and heat flux models using high-speed wind-tunnel data.