Impact of geostatistical reconstruction approaches on model calibration for flow in highly heterogeneous aquifers

Our study is aimed at assessing the extent at which relying on differing geostatistical approaches may affect characterization of the connectivity of geomaterials (or facies) and, in turn, model calibration outputs in highly heterogeneous aquifers. We set our study within a probabilistic framework, by relying on a numerical Monte Carlo (MC) approach. The reconstruction of the spatial distribution of geomaterials and flow simulations are patterned after a field scenario corresponding to the aquifer system serving the city of Bologna (Northern Italy). Two collections of MC realizations of facies distributions, conditional on available lithological data, are generated through two alternative geostatistically-based techniques, i.e., Sequential Indicator and Transition-Probability simulation. Hydraulic conductivity values of the least- and most-conductive facies are estimated within each MC simulation in the context of a Maximum Likelihood (ML) approach by considering available piezometric data. We provide evidence that the choice of the facies reconstruction technique (1) impacts the degree of connectivity of facies whose proportions are close to the percolation threshold while (2) is not sensibly affecting the connectivity associated with facies whose proportions are much larger than the percolation threshold. By relying on the unique (lithological and hydrological) data-set at our disposal, we also explore the performance of ML-based model identification criteria to (1) discriminate amongst competitive facies reconstruction geostatistical models and (2) quantify the (posterior probabilistic) weight associated with each model. We then show that ML-based model averaging provides estimates of hydraulic heads which are slightly more in agreement with available data when compared to the best-performing realization in the T-PROGS set than considering its counterpart associated with the SISIM-based collection