Contaminant plume modeling in aquifers through prediction-focused approach and geoelectrical data

Although hydrogeophysics application in studying the groundwater systems has been significantly increased over the two recent decades, the solute concentration quantities obtained from geophysical modeling are of high uncertainty. This is mostly attributed to (1) the regularization procedure in geoelectrical inverse models, particularly in complex geological settings such as heterogeneous aquifers, and (2) the use of petrophysical relationships.The primary goal of this study is to model the spatio-temporal evolution of the injected salt contaminant in a heterogeneous loamy sand aquifer through the prediction-focused approach (PFA) and resistivity data, circumventing the need for classical geoelectric inversion. The primary advantage of the PFA method is that it does not need any regularization step used in the deterministic geoelectric inversion. This methodology only needs to generate the prior dataset without suffering from any spatial bias, spatially and temporally varying resolution or uncertainty in the post inversion petrophysical transformation.In this research, a synthetic heterogeneous two-dimensional aquifer with 30m´30m is generated through a sequential Gaussian simulation. Then, 500 heterogeneous hydraulic conductivity (K) fields with mean of logK = -4.6 are generated. Accordingly, 500 models of flow and solute transport are carried out for each of six time steps of 0.05, 0.1, 0.2, 0.5, 1, and 5 years. Subsequently, 500 corresponding apparent resistivity datasets are generated through forward geoelectrical modeling (dipole-dipole array) for each of six steps using a MATLAB code. After preparing the large 3D matrices of resistivity and concentration variables as inputs for the PFA, canonical correlation analysis is used to explore the relationship between the apparent resistivity (data) and the solute concentrations (forecast variables) in their reduced dimension space. We selected only 12 and 8 first components for the resistivity and saline concentration variables which they both explain more than 99.5 percent of the variance. The principal component analysis and canonical correlation analysis are used on the reduced datasets to maximize the correlation between the components of the resistivity and solute concentration data. Since a linear relationship is established between the data and forecast, the posterior distribution of the solute concentration is directly sampled using a Gaussian process regression. Finally, the reduced dimension space is back-transformed to the original space. Results demonstrate that the modeled contaminant plumes, in addition to their spatio-temporal distributions, are highly consistent with the maximum and minimum concentration values of the reference images. This signifies the robustness of the PFA for hydrogeophyscical investigation.