Learning model discrepancy: A Gaussian process and sampling-based approach. (1st May 2021)

• Record Type: Journal Article
• Authors: Gardner, P.; Rogers, T.J.; Lord, C.; Barthorpe, R.J.
• Is Part Of: Mechanical systems and signal processing. Volume 152(2021)
• Publication Date: 2021-05-01
• Subjects: Model discrepancy -- Gaussian process regression -- Importance sampling -- Bayesian history matching; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2020.107381
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: A new GP regression and importance sampling method for inferring model discrepancy. A new decoupled two stage process for calibration and model discrepancy inference. Method benchmarked against a joint hierarchal Bayesian approach. Approach demonstrated to be effective on numerical and experimental case studies. Parametric (and hyperparametric) uncertainty quantified in the model discrepancy. Abstract: Predicting events in the real world with a computer model ( simulator ) is challenging. Every simulator, to varying extents, has model discrepancy, a mismatch between real world observations and the simulator (given the 'true' parameters are known). Model discrepancy occurs for various reasons, including simplified or missing physics in the simulator, numerical approximations that are required to compute the simulator outputs, and the fact that assumptions in the simulator are not generally applicable to all real world contexts. The existence of model discrepancy is problematic for the engineer as performing calibration of the simulator will lead to biased parameter estimates, and the resulting simulator is unlikely to accurately predict (or even be valid for) various contexts of interest. This paper proposes an approach for inferring model discrepancy that overcomes non-identifiability problems associated with jointly inferring the simulator parameters along with the model discrepancy. Instead, the proposed procedure seeks to identify model discrepancy given some parameter distribution, which could come from a 'likelihood-free' approach that considers the presence of model discrepancy during calibration, such as Bayesian history matching. In this case, model discrepancy is inferred whilst marginalising out the uncertain simulator outputs via a sampling-based approach, therefore better reflecting the 'true' uncertainty associated with the model discrepancy. Verification of the approach is performed before a demonstration on an experiential case study, comprising a representative five storey building structure.