History matching with subset simulation

authored by

Z. T. Gong, F. A. DiazDelaO, P. O. Hristov, M. Beer

Abstract

Computational cost often hinders the calibration of complex computer models. In this context, history matching (HM) is becoming a widespread calibration strategy, with applications in many disciplines. HM uses a statistical approxi-mation, also known as an emulator, to the model output, in order to mitigate computational cost. The process starts with an observation of a physical system. It then produces progressively more accurate emulators to determine a non-implausible domain: a subset of the input space that provides a good agreement between the model output and the data, conditional on the model structure, the sources of uncertainty, and an implausibility measure. In HM, it is essential to generate samples from the nonimplausible domain, in order to run the model and train the emulator until a stopping condition is met. However, this sampling can be very challenging, since the nonimplausible domain can become orders of magnitude smaller than the original input space very quickly. This paper proposes a solution to this problem using subset simulation, a rare event sampling technique that works efficiently in high dimensions. The proposed approach is demonstrated via calibration and robust design examples from the field of aerospace engineering.

Organisation(s)

Institute for Risk and Reliability

External Organisation(s)

CRRC Sifang Co. Ltd.
University College London (UCL)
University of Liverpool

Type

Article

Journal

International Journal for Uncertainty Quantification

Volume

11

Pages

19-38

No. of pages

20

ISSN

2152-5080

Publication date

2021

Publication status

Published

Peer reviewed

Yes

ASJC Scopus subject areas

Statistics and Probability, Modelling and Simulation, Discrete Mathematics and Combinatorics, Control and Optimization

Electronic version(s)

https://doi.org/10.1615/Int.J.UncertaintyQuantification.2021033543 (Access: Closed)