Abstract
Computer models provide useful tools in understanding and predicting quantities of interest for structural dynamics. Although computer models (simulators) are useful for a specific context, each will contain some level of model-form error. These model-form errors arise for several reasons e.g., numerical approximations to a solution, simplifications of known physics, an inability to model all relevant physics etc. These errors form part of model discrepancy; the difference between observational data and simulator outputs, given the ‘true’ parameters are known. If model discrepancy is not considered during calibration, any inferred parameters will be biased and predictive performance may be poor. Bayesian history matching (BHM) is a technique for calibrating simulators under the assumption that additive model discrepancy exists. This ‘likelihood-free’ approach iteratively assesses the input space using emulators of the simulator and identifies parameters that could have ‘plausibly’ produced target outputs given prior uncertainties. This paper presents, for the first time, the application of BHM in a structural dynamics context. Furthermore, a novel method is provided that utilises Gaussian Process (GP) regression in order to infer the missing model discrepancy functionally from the outputs of BHM. Finally, a demonstration of the effectiveness of the approach is provided for an experimental representative five storey building structure.
Highlights
Calibration of computer models is often an important aspect of creating predictions that adequately match observational data
One of the primary novel contributions of this paper is in providing a methodology for inferring the functional form of the model discrepancy after calibration via Bayesian History Matching (BHM). This is performed by using the maximum a posteriori (MAP) estimate of the inferred parameter posterior distribution such that Gaussian Process (GP) regression models can be inferred to map between the simulator output and a set of training observational data
Large discrepancies between the experimental observations and simulator outputs occur, especially for the first and fifth natural frequencies. This illustrates that the simulator has model-form errors, that would lead to incorrect parameter inference if model discrepancy was not considered in the calibration process
Summary
Calibration of computer models ( defined as simulators) is often an important aspect of creating predictions that adequately match observational data. This is performed by using the maximum a posteriori (MAP) estimate of the inferred parameter posterior distribution such that Gaussian Process (GP) regression models can be inferred to map between the simulator output and a set of training observational data. This novel combined approach provides an alternative method to conventional Bayesian calibration methods that consider model discrepancy within the parameter inference process [1,2,14,15,16].
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
