Hydrogeophysical Characterization of Nonstationary DNAPL Source Zones by Integrating a Convolutional Variational Autoencoder and Ensemble Smoother

Abstract

Detailed characterization of dense nonaqueous phase liquid (DNAPL) source zone architecture (SZA) is essential for designing efficient remediation strategies. However, it is difficult to characterize a highly irregular and localized SZA, because traditional drilling investigations provide limited information. With limited data, the estimation accuracy of traditional geostatistical methods is strongly affected by the parameterization of the prior description of the SZA. To improve characterization performance, we parameterized the DNAPL saturation field using a physics‐based approach. We trained a convolutional variational autoencoder (CVAE) using data from multiphase modeling that captures the physics of DNAPL infiltration. The trained CVAE network was used in SZA inversion to obtain an improved prior DNAPL saturation field, instead of the typical stationary prior covariances. We then integrated the CVAE network into an iterative ensemble smoother (ES), to formulate a joint inversion framework. To overcome difficulties from limited/sparse data, we incorporated hydrogeological and geophysical datasets in the proposed inversion framework. To evaluate the performance of our method, we conducted numerical experiments in a hypothetical heterogeneous aquifer with an intricate SZA. The results show that the CVAE was an effective and efficient parameterization method which can capture the DNAPL infiltration patterns better than a Gaussian prior. The improved prior, combined with multisource datasets, can result in better resolution, and overall improved SZA characterization. In contrast to the standard ES method, the proposed framework reconstructed the SZA more accurately. We also demonstrated that DNAPL depletion behavior and dissolved concentration profiles can be predicted accurately using the estimated SZA.

Thermal Experiments for Fractured Rock Characterization: Theoretical Analysis and Inverse Modeling

“…Two complementary strategies for making the inversion feasible for large, complex problems are (a) to reduce the number of forward solves that are necessary for the inversion algorithm to converge and (b) to reduce the computational cost of an individual forward solve. The former strategy includes the development of accelerated Markov chain samplers, Hamiltonian Monte Carlo sampling, iterative local updating ensemble smoother, ensemble Kalman filters, and learning on statistical manifolds (Barajas‐Solano et al., 2019; Boso & Tartakovsky, 2020a, 2020b; Kang et al., 2021; Zhou & Tartakovsky, 2021). The latter strategy aims to replace an expensive forward model with its cheap surrogate/emulator/reduced‐order model (Ciriello et al., 2019; Lu & Tartakovsky, 2020a, 2020b).…”

Deep Learning for Simultaneous Inference of Hydraulic and Transport Properties

“…In terms of inversion complexity, we subdivide recent groundwater-related studies into three categories: estimation of hydraulic conductivity from measurements of hydraulic head and, optionally, of solute concentration (Mo et al, 2019b;Ju et al, 2018); estimation of contaminant release history from concentration measurements, for known flow and transport parameters (Z. Zhou & Tartakovsky, 2021;Zhang et al, 2015); and estimation of both contaminant release history and hydraulic conductivity from hydraulic head and solute concentration data in two- (Mo et al, 2019a;Xu & Gómez-Hernández, 2018;Kang et al, 2021) and three-dimensional (Kang et al, 2020) aquifers. We briefly discuss the latter category to highlight the novelty of our approach.…”

“…A low-dimensional representation of the random log-normal conductivity obtained via the Karhunen-Loève expansion (KLE) in (Mo et al, 2019a) loses its attractiveness if the subsurface environment is highly heterogeneous, exhibiting short correlation lengths and multimodal statistics; additionally, this study relies on a linear transport model. The deep learning-based strategies of ensemble inversion were adopted in (Xu & Gómez-Hernández, 2018;Kang et al, 2021) to estimate both a non-Gaussian conductivity field and the source of contamination, yet their reported accuracy is relatively low. In the adjacent field of petroleum engineering, CNN postprocessing of PCA (CNN-PCA) parameterization and ESMDA were used to estimate both a channelized permeability and the oil/water rate (Tang et al, 2021).…”

Thermal experiments for fractured rock characterization: theoretical analysis and inverse modeling

“…Two complementary strategies for making the inversion feasible for large, complex problems are i) to reduce the number of forward solves that are necessary for the inversion algorithm to converge, and ii) to reduce the computational cost of an individual forward solve. The former strategy includes the development of accelerated Markov chain samplers, Hamiltonian Monte Carlo sampling, iterative local updating ensemble smoother, ensemble Kalman filters, and learning on statistical manifolds (Barajas-Solano et al, 2019;Boso & Tartakovsky, 2020bKang et al, 2021;Zhou & Tartakovsky, 2021). The latter strategy aims to replace an expensive forward model with its cheap surrogate/emulator/reduced-order model (Ciriello et al, 2019;Lu & Tartakovsky, 2020a.…”

“…The latter strategy aims to replace an expensive forward model with its cheap surrogate/emulator/reduced-order model (Ciriello et al, 2019;Lu & Tartakovsky, 2020a. Among these techniques, various flavors of deep neural networks (DNNs) have attracted attention, in part, because they remain robust for large numbers of inputs and outputs (Zhou & Tartakovsky, 2021;Mo et al, 2020;Kang et al, 2021). Another benefit of DNNs is that their implementation in open-source software is portable to advanced computer architectures, such as graphics processing units and tensor processing units, without significant coding effort from the user.…”