• : Motivated by a challenging computer model calibration problem from the oil and gas industry, involving the design of a so-called honeycomb seal, we develop a new Bayesian calibration methodology to cope with limitations in the canonical apparatus stemming from several factors. We propose a new strategy of on-site experiment design and surrogate modeling to emulate a computer simulator acting on a high-dimensional input space that, although relatively speedy, is prone to numerical instabilities, missing data, and nonstationary dynamics. Our aim is to strike a balance between data-faithful modeling and computational tractability within an overarching calibration framework--tuning the computer model to the outcome of a limited field experiment.  Situating our on-site surrogates within the canonical calibration apparatus requires updates to that framework.  In particular, we describe a novel yet intuitive Bayesian setup that carefully decomposes otherwise prohibitively large matrices by exploiting the sparse blockwise structure thus obtained.  We illustrate empirically that this approach outperforms the canonical, stationary analog, and we summarize calibration results on a toy problem and on our motivating honeycomb example.
  • : Jiangeng Huang, Robert B. Gramacy, Mickael Binois, Mirko Libraschi
  • : Department of Statistics, Virginia Tech; Argonne National Laboratory; Baker Hughes, a GE Company
  • : Jiangeng Huang
  • : big_data
  • : introductory/practitioner
  • : huangj@vt.edu
  • : 540-394-0243
On-site Surrogates for Large-scale Calibration