skip to main content
OSTI.GOV title logo U.S. Department of Energy
Office of Scientific and Technical Information

Title: Multi-fidelity Gaussian process regression for prediction of random fields

Abstract

We propose a new multi-fidelity Gaussian process regression (GPR) approach for prediction of random fields based on observations of surrogate models or hierarchies of surrogate models. Our method builds upon recent work on recursive Bayesian techniques, in particular recursive co-kriging, and extends it to vector-valued fields and various types of covariances, including separable and non-separable ones. The framework we propose is general and can be used to perform uncertainty propagation and quantification in model-based simulations, multi-fidelity data fusion, and surrogate-based optimization. We demonstrate the effectiveness of the proposed recursive GPR techniques through various examples. Specifically, we study the stochastic Burgers equation and the stochastic Oberbeck–Boussinesq equations describing natural convection within a square enclosure. In both cases we find that the standard deviation of the Gaussian predictors as well as the absolute errors relative to benchmark stochastic solutions are very small, suggesting that the proposed multi-fidelity GPR approaches can yield highly accurate results.

Authors:
 [1];  [2];  [3];  [4]
  1. Department of Engineering and Architecture, University of Trieste (Italy)
  2. Department of Applied Mathematics and Statistics, University of California Santa Cruz (United States)
  3. Department of Mechanical Engineering, Massachusetts Institute of Technology (United States)
  4. Division of Applied Mathematics, Brown University (United States)
Publication Date:
OSTI Identifier:
22622285
Resource Type:
Journal Article
Resource Relation:
Journal Name: Journal of Computational Physics; Journal Volume: 336; Other Information: Copyright (c) 2017 Elsevier Science B.V., Amsterdam, The Netherlands, All rights reserved.; Country of input: International Atomic Energy Agency (IAEA)
Country of Publication:
United States
Language:
English
Subject:
71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSICS; BENCHMARKS; COMPUTERIZED SIMULATION; ERRORS; FORECASTING; GAUSSIAN PROCESSES; KRIGING; MATHEMATICAL SOLUTIONS; NATURAL CONVECTION; OPTIMIZATION; RANDOMNESS; STOCHASTIC PROCESSES

Citation Formats

Parussini, L., Venturi, D., E-mail: venturi@ucsc.edu, Perdikaris, P., and Karniadakis, G.E.. Multi-fidelity Gaussian process regression for prediction of random fields. United States: N. p., 2017. Web. doi:10.1016/J.JCP.2017.01.047.
Parussini, L., Venturi, D., E-mail: venturi@ucsc.edu, Perdikaris, P., & Karniadakis, G.E.. Multi-fidelity Gaussian process regression for prediction of random fields. United States. doi:10.1016/J.JCP.2017.01.047.
Parussini, L., Venturi, D., E-mail: venturi@ucsc.edu, Perdikaris, P., and Karniadakis, G.E.. Mon . "Multi-fidelity Gaussian process regression for prediction of random fields". United States. doi:10.1016/J.JCP.2017.01.047.
@article{osti_22622285,
title = {Multi-fidelity Gaussian process regression for prediction of random fields},
author = {Parussini, L. and Venturi, D., E-mail: venturi@ucsc.edu and Perdikaris, P. and Karniadakis, G.E.},
abstractNote = {We propose a new multi-fidelity Gaussian process regression (GPR) approach for prediction of random fields based on observations of surrogate models or hierarchies of surrogate models. Our method builds upon recent work on recursive Bayesian techniques, in particular recursive co-kriging, and extends it to vector-valued fields and various types of covariances, including separable and non-separable ones. The framework we propose is general and can be used to perform uncertainty propagation and quantification in model-based simulations, multi-fidelity data fusion, and surrogate-based optimization. We demonstrate the effectiveness of the proposed recursive GPR techniques through various examples. Specifically, we study the stochastic Burgers equation and the stochastic Oberbeck–Boussinesq equations describing natural convection within a square enclosure. In both cases we find that the standard deviation of the Gaussian predictors as well as the absolute errors relative to benchmark stochastic solutions are very small, suggesting that the proposed multi-fidelity GPR approaches can yield highly accurate results.},
doi = {10.1016/J.JCP.2017.01.047},
journal = {Journal of Computational Physics},
number = ,
volume = 336,
place = {United States},
year = {Mon May 01 00:00:00 EDT 2017},
month = {Mon May 01 00:00:00 EDT 2017}
}
  • Expanding upon the work of Way and Srivastava we demonstrate how the use of training sets of comparable size continue to make Gaussian process regression (GPR) a competitive approach to that of neural networks and other least-squares fitting methods. This is possible via new large-size matrix inversion techniques developed for Gaussian processes (GPs) that do not require that the kernel matrix be sparse. This development, combined with a neural-network kernel function appears to give superior results for this problem. Our best-fit results for the Sloan Digital Sky Survey (SDSS) Main Galaxy Sample using u, g, r, i, z filters givesmore » an rms error of 0.0201 while our results for the same filters in the luminous red galaxy sample yield 0.0220. We also demonstrate that there appears to be a minimum number of training-set galaxies needed to obtain the optimal fit when using our GPR rank-reduction methods. We find that morphological information included with many photometric surveys appears, for the most part, to make the photometric redshift evaluation slightly worse rather than better. This would indicate that most morphological information simply adds noise from the GP point of view in the data used herein. In addition, we show that cross-match catalog results involving combinations of the Two Micron All Sky Survey, SDSS, and Galaxy Evolution Explorer have to be evaluated in the context of the resulting cross-match magnitude and redshift distribution. Otherwise one may be misled into overly optimistic conclusions.« less
  • Excursion sets of Gaussian random fields (GRFs) have been frequently used in the literature to model two-phase random media with measurable phase autocorrelation functions. The goal of successful modeling is finding the optimal field autocorrelation function that best approximates the prescribed phase autocorrelation function. In this paper, we present a technique which uses convex quadratic programming to find the best admissible field autocorrelation function under a prescribed discretization. Unlike previous methods, this technique efficiently optimizes over all admissible field autocorrelation functions, instead of optimizing only over a predetermined parametrized family. The results from using this technique indicate that the GRFmore » model is significantly more versatile than observed in previous studies. An application to modeling a base-catalyzed tetraethoxysilane aerogel system given small-angle neutron scattering data is also presented.« less
  • We present a novel class of models for Type Ia supernova time-evolving spectral energy distributions (SEDs) and absolute magnitudes: they are each modeled as stochastic functions described by Gaussian processes. The values of the SED and absolute magnitudes are defined through well-defined regression prescriptions, so that data directly inform the models. As a proof of concept, we implement a model for synthetic photometry built from the spectrophotometric time series from the Nearby Supernova Factory. Absolute magnitudes at peak B brightness are calibrated to 0.13 mag in the g band and to as low as 0.09 mag in the z =more » 0.25 blueshifted i band, where the dispersion includes contributions from measurement uncertainties and peculiar velocities. The methodology can be applied to spectrophotometric time series of supernovae that span a range of redshifts to simultaneously standardize supernovae together with fitting cosmological parameters.« less
  • This article considers the challenging task of estimating geologic properties of interest using a suite of proxy measurements. The current work recast this task as a manifold learning problem. In this process, this article introduces a novel regression procedure for intrinsic variables constrained onto a manifold embedded in an ambient space. The procedure is meant to sharpen high-dimensional interpolation by inferring non-linear correlations from the data being interpolated. The proposed approach augments manifold learning procedures with a Gaussian process regression. It first identifies, using diffusion maps, a low-dimensional manifold embedded in an ambient high-dimensional space associated with the data. Itmore » relies on the diffusion distance associated with this construction to define a distance function with which the data model is equipped. This distance metric function is then used to compute the correlation structure of a Gaussian process that describes the statistical dependence of quantities of interest in the high-dimensional ambient space. The proposed method is applicable to arbitrarily high-dimensional data sets. Here, it is applied to subsurface characterization using a suite of well log measurements. The predictions obtained in original, principal component, and diffusion space are compared using both qualitative and quantitative metrics. Considerable improvement in the prediction of the geological structural properties is observed with the proposed method.« less
  • Shot peening is the primary surface treatment used to create a uniform, consistent, and reliable sub-surface compressive residual stress layer in aero engine components. A by-product of the shot peening process is random surface roughness that can affect the measurements of the resulting residual stresses and therefore impede their NDE assessment. High frequency eddy current conductivity measurements have the potential to assess these residual stresses in Ni-base super alloys. However, the effect of random surface roughness is expected to become significant in the desired measurement frequency range of 10 to 100 MHz. In this paper, a new Multi-Gaussian (MG) auto-correlationmore » function is proposed for modeling the resulting pseudo-random rough profiles. Its use in the calculation of the Apparent Eddy Current Conductivity (AECC) loss due to surface roughness is demonstrated. The numerical results presented need to be validated with experimental measurements.« less