Integration of adversarial autoencoders with residual dense convolutional networks for inversion of solute transport in non-Gaussian conductivity fields

摘要: Characterization of a non-Gaussian channelized conductivity field in subsurface flow and transport modeling through inverse usually leads to high-dimensional problem requires repeated evaluations the forward model. In this study, we develop convolutional adversarial autoencoder (CAAE) network parameterize fields using low-dimensional latent representation deep residual dense (DRDCN) efficiently construct surrogate model for The two networks are both based on multilevel learning architecture called residual-in-residual block. strategy connection structure block ease training networks, enabling us build deeper that have an essentially increased capacity approximating mappings very high-complexity. CCAE DRDCN incorporated into iterative local updating ensemble smoother formulate inversion framework. integrated method is demonstrated synthetic solute Results indicate CAAE robust parameterization with Gaussian conductivities within each facies. able obtain accurate highly-complex concentration relatively limited data. paramterization approach together significantly reduce number runs required achieve results.