Inferring earthquake ground motion fields with Bayesian Networks

Pierre Gehl, John Douglas, Dina D'Ayala

Research output: Contribution to journalArticle

  • 1 Citations

Abstract

Bayesian Networks (BNs) have the ability to perform inference on uncertain variables given evidence on observed quantities, which makes them relevant mathematical tools for the updating of ground-motion fields based on strong-motion records or macroseismic observations. Therefore the present article investigates the use of BN models of spatially correlated Gaussian random fields as an accurate and scalable method for the generation of ground-motion maps. The proposed BN model is based on continuous Gaussian variables, as opposed to discrete variables as in previous formulations, and it is built to account for cross-correlated ground-motion parameters as well as macroseismic observations. This approach is validated with respect to the analytical solution (i.e., conditional multivariate normal distributions) and it is also compared to the USGS ShakeMap method, thus demonstrating a better ability to model jointly the inter- and intra-event error terms of ground-motion models. The scalability of the approach, i.e. its capacity to be applied to large grids, is ensured by a grid sub-division strategy, which appears to be computationally efficient and accurate within an error rate of a fraction of percent. Finally, the BN implementation is demonstrated on a real-world example (the Mw 6.2 Kumamoto, Japan, 2016 foreshock), where vector-valued shake-maps of cross-correlated intensity measures are generated, along with the integration of macroseismic observations.
LanguageEnglish
Pages2792-2808
JournalBulletin of the Seismological Society of America
Volume107
Issue number6
DOIs
StatePublished - 1 Dec 2017

Fingerprint

ground motion
earthquakes
earthquake
foreshock
grids
strong motion
inference
normal density functions
division
Japan
formulations
method

Keywords

  • Bayesian network
  • earthquake
  • strong ground motion
  • ShakeMap
  • ground-motion prediction
  • engineering seismology

Cite this

@article{242e88157a5a4c0fb7e2b1c65b2774c9,
title = "Inferring earthquake ground motion fields with Bayesian Networks",
abstract = "Bayesian Networks (BNs) have the ability to perform inference on uncertain variables given evidence on observed quantities, which makes them relevant mathematical tools for the updating of ground-motion fields based on strong-motion records or macroseismic observations. Therefore the present article investigates the use of BN models of spatially correlated Gaussian random fields as an accurate and scalable method for the generation of ground-motion maps. The proposed BN model is based on continuous Gaussian variables, as opposed to discrete variables as in previous formulations, and it is built to account for cross-correlated ground-motion parameters as well as macroseismic observations. This approach is validated with respect to the analytical solution (i.e., conditional multivariate normal distributions) and it is also compared to the USGS ShakeMap method, thus demonstrating a better ability to model jointly the inter- and intra-event error terms of ground-motion models. The scalability of the approach, i.e. its capacity to be applied to large grids, is ensured by a grid sub-division strategy, which appears to be computationally efficient and accurate within an error rate of a fraction of percent. Finally, the BN implementation is demonstrated on a real-world example (the Mw 6.2 Kumamoto, Japan, 2016 foreshock), where vector-valued shake-maps of cross-correlated intensity measures are generated, along with the integration of macroseismic observations.",
keywords = "Bayesian network, earthquake, strong ground motion, ShakeMap, ground-motion prediction, engineering seismology",
author = "Pierre Gehl and John Douglas and Dina D'Ayala",
year = "2017",
month = "12",
day = "1",
doi = "10.1785/0120170073",
language = "English",
volume = "107",
pages = "2792--2808",
journal = "Bulletin of the Seismological Society of America",
issn = "0037-1106",
number = "6",

}

Inferring earthquake ground motion fields with Bayesian Networks. / Gehl, Pierre; Douglas, John; D'Ayala, Dina.

In: Bulletin of the Seismological Society of America, Vol. 107, No. 6, 01.12.2017, p. 2792-2808.

Research output: Contribution to journalArticle

TY - JOUR

T1 - Inferring earthquake ground motion fields with Bayesian Networks

AU - Gehl,Pierre

AU - Douglas,John

AU - D'Ayala,Dina

PY - 2017/12/1

Y1 - 2017/12/1

N2 - Bayesian Networks (BNs) have the ability to perform inference on uncertain variables given evidence on observed quantities, which makes them relevant mathematical tools for the updating of ground-motion fields based on strong-motion records or macroseismic observations. Therefore the present article investigates the use of BN models of spatially correlated Gaussian random fields as an accurate and scalable method for the generation of ground-motion maps. The proposed BN model is based on continuous Gaussian variables, as opposed to discrete variables as in previous formulations, and it is built to account for cross-correlated ground-motion parameters as well as macroseismic observations. This approach is validated with respect to the analytical solution (i.e., conditional multivariate normal distributions) and it is also compared to the USGS ShakeMap method, thus demonstrating a better ability to model jointly the inter- and intra-event error terms of ground-motion models. The scalability of the approach, i.e. its capacity to be applied to large grids, is ensured by a grid sub-division strategy, which appears to be computationally efficient and accurate within an error rate of a fraction of percent. Finally, the BN implementation is demonstrated on a real-world example (the Mw 6.2 Kumamoto, Japan, 2016 foreshock), where vector-valued shake-maps of cross-correlated intensity measures are generated, along with the integration of macroseismic observations.

AB - Bayesian Networks (BNs) have the ability to perform inference on uncertain variables given evidence on observed quantities, which makes them relevant mathematical tools for the updating of ground-motion fields based on strong-motion records or macroseismic observations. Therefore the present article investigates the use of BN models of spatially correlated Gaussian random fields as an accurate and scalable method for the generation of ground-motion maps. The proposed BN model is based on continuous Gaussian variables, as opposed to discrete variables as in previous formulations, and it is built to account for cross-correlated ground-motion parameters as well as macroseismic observations. This approach is validated with respect to the analytical solution (i.e., conditional multivariate normal distributions) and it is also compared to the USGS ShakeMap method, thus demonstrating a better ability to model jointly the inter- and intra-event error terms of ground-motion models. The scalability of the approach, i.e. its capacity to be applied to large grids, is ensured by a grid sub-division strategy, which appears to be computationally efficient and accurate within an error rate of a fraction of percent. Finally, the BN implementation is demonstrated on a real-world example (the Mw 6.2 Kumamoto, Japan, 2016 foreshock), where vector-valued shake-maps of cross-correlated intensity measures are generated, along with the integration of macroseismic observations.

KW - Bayesian network

KW - earthquake

KW - strong ground motion

KW - ShakeMap

KW - ground-motion prediction

KW - engineering seismology

U2 - 10.1785/0120170073

DO - 10.1785/0120170073

M3 - Article

VL - 107

SP - 2792

EP - 2808

JO - Bulletin of the Seismological Society of America

T2 - Bulletin of the Seismological Society of America

JF - Bulletin of the Seismological Society of America

SN - 0037-1106

IS - 6

ER -