Seeing macro-dispersivity from hydraulic conductivity field with convolutional neural network

Abstract

Building a quantitative relation between the spatial heterogeneity of the hydraulic conductivity fields and the macroscale behavior of solute transport is fundamental for groundwater environment problem. In this work, the deep learning technique is explored to build the functional mapping between the hydraulic conductivity field and the longitudinal macro-dispersivity. We examine the capability of the deep neural network in estimating macro-dispersivities of conductivity fields with different variances. The universality of the trained deep neural network is investigated. Comparisons of the neural network results and the reference values (macro-dispersivities from transport simulation) suggest the promising potential of deep learning technique in porous media with moderate heterogeneity. For a given size of training datasets, the deep neural network produces better macro-dispersivity estimation for the conductivity field with smaller variance. The trained neural network by conductivity fields with larger variance has stronger universality for macro-dispersivity estimation. This study demonstrates that deep neural network can be an effective alternative for estimating macroscale behavior of solute transport by directly interpreting hydraulic conductivity fields.