You are here:
SELECTION AND CALIBRATION OF SUBSURFACE REACTIVE TRANSPORT MODELS USING A SURROGATE-MODEL APPROACH
Citation:
MATOTT, L. S. AND A. J. RABIDEAU. SELECTION AND CALIBRATION OF SUBSURFACE REACTIVE TRANSPORT MODELS USING A SURROGATE-MODEL APPROACH. Presented at American Geophysical Union Fall 2006 Meeting, San Francisco, CA, December 11 - 15, 2006.
Impact/Purpose:
The primary goals are to: (1) Construct a 400-node PC-based supercomputing cluster supporting Windows and Linux computer operating systems (i.e. SuperMUSE: Supercomputer for Model Uncertainty and Sensitivity Evaluation); (2) Develop platform-independent system software for the management of SuperMUSE and parallelization of EPA models and modeling systems for implementation on SuperMUSE (and other PC-based clusters); (3) Conduct uncertainty and sensitivity analyses of the 3MRA modeling system; (4) Develop advanced algorithmic software for advanced statistical sampling methods, and screening, localized, and global sensitivity analyses; and (5) Provide customer-oriented model applications for probabilistic risk assessment supporting quality assurance in multimedia decision-making.
Description:
While standard techniques for uncertainty analysis have been successfully applied to groundwater flow models, extension to reactive transport is frustrated by numerous difficulties, including excessive computational burden and parameter non-uniqueness. This research introduces a novel surrogate-based method that overcomes such difficulties by utilizing a global search technique that samples from a hierarchy of candidate reactive transport models. During calibration, the search algorithm is driven toward least-squares minimization of the most parsimonious model that best matches the available calibration data. This surrogate-model approach is demonstrated via application to several nitrate contamination problems. Results indicate that, due to the utilization of simpler models in regions of parameter insensitivity, the method is able to identify quality model fits at reduced computational expense, relative to traditional techniques. Furthermore, comparisons with a formal multi-model ranking procedure suggest the new approach is a promising tool for multi-model ranking and selection.