ARS | Publication request: Augmenting an observation network to facilitate flow and transport model discrimination.

Submitted to: Meeting Abstract
Publication Type: Abstract Only
Publication Acceptance Date: October 31, 2011
Publication Date: December 5, 2011
Citation: Pachepsky, Y.A., Kuznetsov, M., Guber, A., Yakirevich, A., Martinez, G., Gish, T.J., Cady, R., Nicholson, T. 2011. Augmenting an observation network to facilitate flow and transport model discrimination. [abstract]. American Geophysical Union. H23H-1372.

Technical Abstract: Improving understanding of subsurface conditions includes performance comparison for competing models, independently developed or obtained via model abstraction. The model comparison and discrimination can be improved if additional observations will be included. The objective of this work was to implement and to test a Bayesian method for the sequential design of the network augmentation. The method is based on (1) generalization of Kullback’s discriminant function and “weights of evidence” for the case of available prior probabilities, (2) ensemble modeling to estimate variance of the predicted values. The method was tested with the data from the tracer experiment at the USDA-ARS OPE3 integrated research site. A pulse of KCL solution was applied to an irrigation plot, and chloride concentrations were measured in the groundwater at three sampling depths in 12 observations wells. The spatial distribution of soil materials was obtained from cores taken from depths of 0-200 cm with 20 cm increment during installation of observation wells. A three-dimension flow and transport model was developed to simulate the flow and chloride transport for the tracer experiment at the OPE3 site. The manual calibration of hydraulic conductivities and dispersivities was performed, and pedotransfer functions were conditioned to calibration results to build ensemble of models. The search of the optimal location of the augmentation wells was done on a 2D grid. Models of different complexity were compared. Both single and multiple responses were used to discriminate models. The outcome of this study can provide the information for the future data collection and monitoring efforts to further reduce the uncertainty