An efficient surrogate-based simulation-optimization method for calibrating a regional MODFLOW model

Mingjie Chen, Azizallah Izady, Osman A. Abdalla

Research output: Contribution to journalArticle

3 Citations (Scopus)

Abstract

Simulation-optimization method entails a large number of model simulations, which is computationally intensive or even prohibitive if the model simulation is extremely time-consuming. Statistical models have been examined as a surrogate of the high-fidelity physical model during simulation-optimization process to tackle this problem. Among them, Multivariate Adaptive Regression Splines (MARS), a non-parametric adaptive regression method, is superior in overcoming problems of high-dimensions and discontinuities of the data. Furthermore, the stability and accuracy of MARS model can be improved by bootstrap aggregating methods, namely, bagging. In this paper, Bagging MARS (BMARS) method is integrated to a surrogate-based simulation-optimization framework to calibrate a three-dimensional MODFLOW model, which is developed to simulate the groundwater flow in an arid hardrock-alluvium region in northwestern Oman. The physical MODFLOW model is surrogated by the statistical model developed using BMARS algorithm. The surrogate model, which is fitted and validated using training dataset generated by the physical model, can approximate solutions rapidly. An efficient Sobol’ method is employed to calculate global sensitivities of head outputs to input parameters, which are used to analyze their importance for the model outputs spatiotemporally. Only sensitive parameters are included in the calibration process to further improve the computational efficiency. Normalized root mean square error (NRMSE) between measured and simulated heads at observation wells is used as the objective function to be minimized during optimization. The reasonable history match between the simulated and observed heads demonstrated feasibility of this high-efficient calibration framework.

Original languageEnglish
Pages (from-to)591-603
Number of pages13
JournalJournal of Hydrology
Volume544
DOIs
Publication statusPublished - Jan 1 2017

Fingerprint

simulation
method
calibration
alluvial deposit
groundwater flow
discontinuity
well
history
parameter

Keywords

  • Bagging MARS
  • Calibration
  • Groundwater
  • MODFLOW
  • Surrogate

ASJC Scopus subject areas

  • Water Science and Technology

Cite this

An efficient surrogate-based simulation-optimization method for calibrating a regional MODFLOW model. / Chen, Mingjie; Izady, Azizallah; Abdalla, Osman A.

In: Journal of Hydrology, Vol. 544, 01.01.2017, p. 591-603.

Research output: Contribution to journalArticle

@article{d723916fb29847549892c40b9e1c04ff,
title = "An efficient surrogate-based simulation-optimization method for calibrating a regional MODFLOW model",
abstract = "Simulation-optimization method entails a large number of model simulations, which is computationally intensive or even prohibitive if the model simulation is extremely time-consuming. Statistical models have been examined as a surrogate of the high-fidelity physical model during simulation-optimization process to tackle this problem. Among them, Multivariate Adaptive Regression Splines (MARS), a non-parametric adaptive regression method, is superior in overcoming problems of high-dimensions and discontinuities of the data. Furthermore, the stability and accuracy of MARS model can be improved by bootstrap aggregating methods, namely, bagging. In this paper, Bagging MARS (BMARS) method is integrated to a surrogate-based simulation-optimization framework to calibrate a three-dimensional MODFLOW model, which is developed to simulate the groundwater flow in an arid hardrock-alluvium region in northwestern Oman. The physical MODFLOW model is surrogated by the statistical model developed using BMARS algorithm. The surrogate model, which is fitted and validated using training dataset generated by the physical model, can approximate solutions rapidly. An efficient Sobol’ method is employed to calculate global sensitivities of head outputs to input parameters, which are used to analyze their importance for the model outputs spatiotemporally. Only sensitive parameters are included in the calibration process to further improve the computational efficiency. Normalized root mean square error (NRMSE) between measured and simulated heads at observation wells is used as the objective function to be minimized during optimization. The reasonable history match between the simulated and observed heads demonstrated feasibility of this high-efficient calibration framework.",
keywords = "Bagging MARS, Calibration, Groundwater, MODFLOW, Surrogate",
author = "Mingjie Chen and Azizallah Izady and Abdalla, {Osman A.}",
year = "2017",
month = "1",
day = "1",
doi = "10.1016/j.jhydrol.2016.12.011",
language = "English",
volume = "544",
pages = "591--603",
journal = "Journal of Hydrology",
issn = "0022-1694",
publisher = "Elsevier",

}

TY - JOUR

T1 - An efficient surrogate-based simulation-optimization method for calibrating a regional MODFLOW model

AU - Chen, Mingjie

AU - Izady, Azizallah

AU - Abdalla, Osman A.

PY - 2017/1/1

Y1 - 2017/1/1

N2 - Simulation-optimization method entails a large number of model simulations, which is computationally intensive or even prohibitive if the model simulation is extremely time-consuming. Statistical models have been examined as a surrogate of the high-fidelity physical model during simulation-optimization process to tackle this problem. Among them, Multivariate Adaptive Regression Splines (MARS), a non-parametric adaptive regression method, is superior in overcoming problems of high-dimensions and discontinuities of the data. Furthermore, the stability and accuracy of MARS model can be improved by bootstrap aggregating methods, namely, bagging. In this paper, Bagging MARS (BMARS) method is integrated to a surrogate-based simulation-optimization framework to calibrate a three-dimensional MODFLOW model, which is developed to simulate the groundwater flow in an arid hardrock-alluvium region in northwestern Oman. The physical MODFLOW model is surrogated by the statistical model developed using BMARS algorithm. The surrogate model, which is fitted and validated using training dataset generated by the physical model, can approximate solutions rapidly. An efficient Sobol’ method is employed to calculate global sensitivities of head outputs to input parameters, which are used to analyze their importance for the model outputs spatiotemporally. Only sensitive parameters are included in the calibration process to further improve the computational efficiency. Normalized root mean square error (NRMSE) between measured and simulated heads at observation wells is used as the objective function to be minimized during optimization. The reasonable history match between the simulated and observed heads demonstrated feasibility of this high-efficient calibration framework.

AB - Simulation-optimization method entails a large number of model simulations, which is computationally intensive or even prohibitive if the model simulation is extremely time-consuming. Statistical models have been examined as a surrogate of the high-fidelity physical model during simulation-optimization process to tackle this problem. Among them, Multivariate Adaptive Regression Splines (MARS), a non-parametric adaptive regression method, is superior in overcoming problems of high-dimensions and discontinuities of the data. Furthermore, the stability and accuracy of MARS model can be improved by bootstrap aggregating methods, namely, bagging. In this paper, Bagging MARS (BMARS) method is integrated to a surrogate-based simulation-optimization framework to calibrate a three-dimensional MODFLOW model, which is developed to simulate the groundwater flow in an arid hardrock-alluvium region in northwestern Oman. The physical MODFLOW model is surrogated by the statistical model developed using BMARS algorithm. The surrogate model, which is fitted and validated using training dataset generated by the physical model, can approximate solutions rapidly. An efficient Sobol’ method is employed to calculate global sensitivities of head outputs to input parameters, which are used to analyze their importance for the model outputs spatiotemporally. Only sensitive parameters are included in the calibration process to further improve the computational efficiency. Normalized root mean square error (NRMSE) between measured and simulated heads at observation wells is used as the objective function to be minimized during optimization. The reasonable history match between the simulated and observed heads demonstrated feasibility of this high-efficient calibration framework.

KW - Bagging MARS

KW - Calibration

KW - Groundwater

KW - MODFLOW

KW - Surrogate

UR - http://www.scopus.com/inward/record.url?scp=85006789259&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85006789259&partnerID=8YFLogxK

U2 - 10.1016/j.jhydrol.2016.12.011

DO - 10.1016/j.jhydrol.2016.12.011

M3 - Article

VL - 544

SP - 591

EP - 603

JO - Journal of Hydrology

JF - Journal of Hydrology

SN - 0022-1694

ER -