An efficient calibration-constrained Monte Carlo technique for evaluating model predictive error

Matthew Tonkin, John Doherty

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

We describe a new Monte Carlo (MC) technique that reduces the computational burden of calibration-constrained MC using the concept of the calibration null space. In the new MC approach, the model is calibrated using a subspace regularization method such as Truncated Singular Value Decomposition (TSVD) or the hybrid Tikhonov-TSVD approach described by Tonkin & Doherty (2005). Next, a stochastic parameter field generator is used to produce many realizations of the parameter field. For each realization, a difference is formed between the stochastic field and the calibration field. This difference is projected onto the calibration null space determined through the calibration process, and added to the calibration field. If the model is no longer calibrated, the underlying field is re-estimated with the null-space-difference field "riding on its back". If this can be undertaken using pre-calculated sensitivities, conditioning may require only a very small number of model runs. The new MC approach can rapidly produce a large number of conditioned stochastic fields, for use in assessing the potential error in a wide range of predictions.

Original languageEnglish
Title of host publicationProceedings of an International Conference on Calibration and Reliability in Groundwater Modelling
Subtitle of host publicationCredibility of Modelling, ModelCARE2007
Pages76-80
Number of pages5
Edition320
Publication statusPublished - 2008
Externally publishedYes
EventInternational Conference on Calibration and Reliability in Groundwater Modelling: Credibility of Modelling, ModelCARE2007 - Copenhagen, Denmark
Duration: 9 Sep 200713 Sep 2007

Publication series

NameIAHS-AISH Publication
Number320
ISSN (Print)0144-7815

Conference

ConferenceInternational Conference on Calibration and Reliability in Groundwater Modelling: Credibility of Modelling, ModelCARE2007
Country/TerritoryDenmark
CityCopenhagen
Period9/09/0713/09/07

Keywords

  • Calibration
  • Monte Carlo
  • Null space
  • Predictive error
  • Stochastic
  • Uncertainty

Fingerprint

Dive into the research topics of 'An efficient calibration-constrained Monte Carlo technique for evaluating model predictive error'. Together they form a unique fingerprint.

Cite this