Merging parallel tempering with sequential geostatistical resampling for improved posterior exploration of high-dimensional subsurface categorical fields

Research output: Contribution to journalArticlepeer-review

Standard

Merging parallel tempering with sequential geostatistical resampling for improved posterior exploration of high-dimensional subsurface categorical fields. / Laloy, Eric; Jacques, Diederik; Linde, Niklas; Mariethoz, Grégoire.

In: Advances in Water Resources, Vol. 90, 01.04.2016, p. 57-69.

Research output: Contribution to journalArticlepeer-review

Bibtex - Download

@article{9a2b2baf3058428abe1c5dd12808bdea,
title = "Merging parallel tempering with sequential geostatistical resampling for improved posterior exploration of high-dimensional subsurface categorical fields",
abstract = "The sequential geostatistical resampling (SGR) algorithm is a Markov chain Monte Carlo (MCMC) scheme for sampling from possibly non-Gaussian, complex spatially-distributed prior models such as geologic fa- cies or categorical fields. In this work, we highlight the limits of standard SGR for posterior inference of high-dimensional categorical fields with realistically complex likelihood landscapes and benchmark a parallel tempering implementation (PT-SGR). Our proposed PT-SGR approach is demonstrated using syn- thetic (error corrupted) data from steady-state flow and transport experiments in categorical 7575- and 10,0 0 0-dimensional 2D conductivity fields. In both case studies, every SGR trial gets trapped in a local optima while PT-SGR maintains an higher diversity in the sampled model states. The advantage of PT-SGR is most apparent in an inverse transport problem where the posterior distribution is made bimodal by construction. PT-SGR then converges towards the appropriate data misfit much faster than SGR and partly recovers the two modes. In contrast, for the same computational resources SGR does not fit the data to the appropriate error level and hardly produces a locally optimal solution that looks visually similar to one of the two reference modes. Although PT-SGR clearly surpasses SGR in performance, our results also indicate that using a small number (16–24) of temperatures (and thus parallel cores) may not permit complete sampling of the posterior distribution by PT-SGR within a reasonable computational time (less than 1–2 weeks).",
keywords = "parallel tempering, sequential geostatistical resampling, training image, MCMC, multiple-point statistics",
author = "Eric Laloy and Diederik Jacques and Niklas Linde and Gr{\'e}goire Mariethoz",
note = "Score=10",
year = "2016",
month = apr,
day = "1",
doi = "10.1016/j.advwatres.2016.02.008",
language = "English",
volume = "90",
pages = "57--69",
journal = "Advances in Water Resources",
issn = "0309-1708",
publisher = "Elsevier",

}

RIS - Download

TY - JOUR

T1 - Merging parallel tempering with sequential geostatistical resampling for improved posterior exploration of high-dimensional subsurface categorical fields

AU - Laloy, Eric

AU - Jacques, Diederik

AU - Linde, Niklas

AU - Mariethoz, Grégoire

N1 - Score=10

PY - 2016/4/1

Y1 - 2016/4/1

N2 - The sequential geostatistical resampling (SGR) algorithm is a Markov chain Monte Carlo (MCMC) scheme for sampling from possibly non-Gaussian, complex spatially-distributed prior models such as geologic fa- cies or categorical fields. In this work, we highlight the limits of standard SGR for posterior inference of high-dimensional categorical fields with realistically complex likelihood landscapes and benchmark a parallel tempering implementation (PT-SGR). Our proposed PT-SGR approach is demonstrated using syn- thetic (error corrupted) data from steady-state flow and transport experiments in categorical 7575- and 10,0 0 0-dimensional 2D conductivity fields. In both case studies, every SGR trial gets trapped in a local optima while PT-SGR maintains an higher diversity in the sampled model states. The advantage of PT-SGR is most apparent in an inverse transport problem where the posterior distribution is made bimodal by construction. PT-SGR then converges towards the appropriate data misfit much faster than SGR and partly recovers the two modes. In contrast, for the same computational resources SGR does not fit the data to the appropriate error level and hardly produces a locally optimal solution that looks visually similar to one of the two reference modes. Although PT-SGR clearly surpasses SGR in performance, our results also indicate that using a small number (16–24) of temperatures (and thus parallel cores) may not permit complete sampling of the posterior distribution by PT-SGR within a reasonable computational time (less than 1–2 weeks).

AB - The sequential geostatistical resampling (SGR) algorithm is a Markov chain Monte Carlo (MCMC) scheme for sampling from possibly non-Gaussian, complex spatially-distributed prior models such as geologic fa- cies or categorical fields. In this work, we highlight the limits of standard SGR for posterior inference of high-dimensional categorical fields with realistically complex likelihood landscapes and benchmark a parallel tempering implementation (PT-SGR). Our proposed PT-SGR approach is demonstrated using syn- thetic (error corrupted) data from steady-state flow and transport experiments in categorical 7575- and 10,0 0 0-dimensional 2D conductivity fields. In both case studies, every SGR trial gets trapped in a local optima while PT-SGR maintains an higher diversity in the sampled model states. The advantage of PT-SGR is most apparent in an inverse transport problem where the posterior distribution is made bimodal by construction. PT-SGR then converges towards the appropriate data misfit much faster than SGR and partly recovers the two modes. In contrast, for the same computational resources SGR does not fit the data to the appropriate error level and hardly produces a locally optimal solution that looks visually similar to one of the two reference modes. Although PT-SGR clearly surpasses SGR in performance, our results also indicate that using a small number (16–24) of temperatures (and thus parallel cores) may not permit complete sampling of the posterior distribution by PT-SGR within a reasonable computational time (less than 1–2 weeks).

KW - parallel tempering

KW - sequential geostatistical resampling

KW - training image

KW - MCMC

KW - multiple-point statistics

UR - http://ecm.sckcen.be/OTCS/llisapi.dll/open/11928131

U2 - 10.1016/j.advwatres.2016.02.008

DO - 10.1016/j.advwatres.2016.02.008

M3 - Article

VL - 90

SP - 57

EP - 69

JO - Advances in Water Resources

JF - Advances in Water Resources

SN - 0309-1708

ER -

ID: 854905