Improved monitoring of dense non-aqueous phase liquid (DNAPL) source zone remediation through hydrogeophysical inversion using variational autoencoder and Ensemble Kalman filter

Abstract:

Monitoring the depletion of dense non-aqueous phase liquid (DNAPL) source zones at contaminated sites is important for evaluating remediation efficiency. However, it is difficult to capture the evolution of the highly irregular and localized DNAPL source zone architecture (SZA) at a fine resolution because traditional drilling investigations are costly and only provide limited information at a few locations. With those limited data, the estimation accuracy of traditional geostatistical inversion methods is strongly affected by the parameterization of the prior description of the unknown DNAPL SZA. To improve monitoring performance, we parameterize the DNAPL saturation field using a machine learning approach with a physics-based SZA data set. Firstly, we train a convolutional variational autoencoder (CVAE) using data from multiphase modeling that captures the physics of DNAPL infiltration and depletion. The trained CVAE network, which leverages an efficient nonlinear low-dimensional representation of SZA, is then used in a data assimilation framework using the ensemble Kalman filter (EnKF) to estimate the evolution of the SZA. To overcome difficulties from the limited information at the sparse monitoring locations, we incorporate both hydrogeological and geophysical datasets in the proposed framework. To evaluate the performance of our method, we conducted two dimensional numerical experiments in a hypothetical heterogeneous aquifer with realistic DNAPL SZAs. The results show that the CVAE is an effective parameterization method which can satisfactorily capture the spatial patterns of the evolving DNAPL SZAs and can provide accurate real-time monitoring of DNAPL source zone evolution.