Calibration and Reliability in Groundwater Modelling: Credibility of Modelling

(Proceedings of ModelCARE 2007 Conference, held in Denmark, September 2007). IAHS Publ. 320, 2008, 299-304.

Can conditioning to transmissivity data worsen model predictions?

JAOUHER KERROU¹, HARRIE-JAN HENDRICKS FRANSSEN²,
PHILIPPE RENARD¹ & IVAN LUNATI³

1 Centre for Hydrogeology, University of Neuchâtel, Emile Argand 11-CP158, CH-2009 Neuchâtel, Switzerland

jaouher.kerrou@unine.ch

2 Institute of Environmental Engineering, ETH Zurich, CH-8093 Zurich, Switzerland

3 Laboratory of Environmental Fluid Mechanics and Hydrology, EPF Lausanne, CH-1015 Lausanne, Switzerland

Abstract It is reasonable to think that spatially variable transmissivity fields often follow non-multi-Gaussian statistics. Nevertheless, in groundwater flow and mass transport studies multi-Gaussian models are very popular. This paper investigates the consequences of adopting a wrong Random Function (RF) model. Previous studies have shown that conditioning to hydraulic head data, adopting a multi-Gaussian approach, only very marginally detects connected structures typical for non-multi-Gaussian fields. In addition, several numerical simulations performed have given us a hint that conditioning on a large number of transmissivity data might prevent head conditioning from being effective. We consider non-multi-Gaussian T fields (with braided structures) and compare the results obtained by using the T data only for computing the variogram with those obtained by additionally conditioning to T data (erroneously, a multi-Gaussian RF model is assumed). The preliminary results presented here do not clearly show an improvement when only part of the data is used for T conditioning (this confirms the primary role played by the RF model in deteriorating field characterization). However, evidence is found that conditioning to T data yields a systematic loss of connectivity behind a distance of the order of the variogram range. This fact prevents the inverse problem from identifying elongated capture zones. Conditioning to h data, instead, generally yields an increase in connectivity, which is more effective at distances larger than the variogram range, and seems to allow a partial recovery of non-Gaussian structures.

Key words aquifer characterization; inverse modelling; variogram; conditioning data; stochastic simulations; uncertainty; connectivity