Adaptive calibration of a computer code with time-series output. (April 2020)

• Record Type: Journal Article
• Authors: Perrin, G.
• Is Part Of: Reliability engineering & system safety. Volume 196(2020)
• Publication Date: 2020-04
• Subjects: Bayesian framework -- Computer experiment -- Dynamic simulator -- Gaussian process -- Multi-output; Reliability (Engineering) -- Periodicals; System safety -- Periodicals; Industrial safety -- Periodicals; Fiabilité -- Périodiques; Sécurité des systèmes -- Périodiques; Sécurité du travail -- Périodiques; 620.00452
• DOI: 10.1016/j.ress.2019.106728
• Languages: English
• ISSNs: 0951-8320

Abstract

Highlights: The paper considers the calibration of a code from high-dimensional observations. Extensions of the Gaussian process regression to the multivariate case are presented. An adaptive sampling framework is introduced to solve the inference problem. Two real-life applications are presented to illustrate the proposed methodology. Abstract: Simulation plays a major role in the conception, the optimization and the certification of complex systems. Of particular interest here is the calibration of the parameters of computer models from high-dimensional physical observations. When the run times of these computer codes is high, this work focuses on the numerical challenges associated with the statistical inference. In particular, several adaptations of the Gaussian Process Regression (GPR) to the high-dimensional or functional output case are presented for the emulation of computer codes from limited data. Then, an adaptive procedure is detailed to minimize the calibration parameters uncertainty at the minimal computational cost. The proposed method is eventually applied to two applications that are based on dynamic simulators.