High-resolution Characterization Of DNAPL Contaminant Source Zone Architecture By Assimilating Multi-source Data

Toxic organic contaminants,such as chlorinated solvents,often exist in the form of oily phases that do not mix with water and are denser than water.Such dense nonaqueous phase liquids(DNAPLs)represent a challenging environmental problem worldwide.To develop efficient remediation strategies,the quantity and morphology of the DNAPL in the subsurface needs to be accurately identified.However,characterization of the DNAPL source zone architecture(SZA)has proven difficult mainly due to the following reasons:(1)In real-world conditions,the number of available borehole data is usually limited,which may result in an inaccurate DNAPL mapping.(2)High-resolution characterization of the DNAPL SZA fall in the category of high-dimensional inversion,with a large number of unknowns to be estimated,which may result in a significant computational cost.(3)The typical choice for spatially correlated field estimation in the traditional geostatistical approach is a simple Gaussian covariance model,which may not be suitable for the highly non-Gaussian SZA as a result of complex multiphase physics.Therefore,the deviation of the prior assumption from physical reality may lead to biased estimation for SZA,missing the fine structure of the true DNAPL saturation(S_N)field.Therefore,we integrated the hydrogeological and hydrogeophysical data to overcome the difficulty from sparse borehole data.We solved the high-dimensional inversion by the principal component geostatistical approach,in which the dense covariance matrix is replaced by a low-rank approximation,thus alleviating the computational barrier.To handle the non-Gaussianity of the SZA,we parameterized the non-Gaussian SN field using a deep learning-based approach(convolutional variational autoencoder,CVAE).In addition,we assessed the ability of the proposed methods in real-world sandbox experiments and numerical experiments.The research content and the main conclusions are as follow:(1)For the problem of sparse data,we introduced the cost-effective hydrogeophysical methods(i.e.,ERT-electrical resistivity tomography,SP- self potential)to characterize the SZA,and then assimilated these multi-source data in a coupled groundwater-hydrogeophysical inversion framework.By integrating multi-source complementary data,we can achieve a high-resolution characterization with a low cost.First,we demonstrated the ability of ERT to characterize the DNAPL distribution in a real-world three-dimensional sandbox experiment.Then,we conducted numerical experiments to show the advantages of integrating multi-source data.Results show that,by adding hydrogeophysical data(i.e.,self-potential data)to hydrogeological data(i.e.,hydraulic heads and partitioning tracer data),the error is reduced by 68%in DNAPL saturation characterization.(2)To overcome the computational barrier from the SZA inversion problem,we developed a sequential inversion framework in an efficient way,with computational costs reduced by(i)using PCGA to address the computation bottleneck caused by large-dimensional unknown parameters,(ii)harnessing the complementarity of data and using them in sequence so that the information content can be maximized without relying on the expensive multiphase simulations.To evaluate the performance of the proposed framework,we conducted numerical experiments in a 3D aquifer with complex DNAPL sources zones.The results demonstrate that the proposed method can provide an accurate DNAPL imaging in a high computational efficiency.(3)To handle the non-Gaussianity of the SZA,we proposed a physics-based parameterization method to accurately describe the prior distribution of the non-Gaussian SN field.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 Gaussian prior covariances.We then integrated the CVAE network into an iterative ensemble smoother(ESMDA),to formulate a joint inversion framework.We evaluated the performance of the proposed CVAE-ESMDA method in numerical experiments.The results show that the CVAE-based prior can capture the DNAPL infiltration patterns better than a Gaussian prior.By utilizing the CVAE parameterization with the ESMDA,the estimation error of total NAPL mass was reduced by 89%,compared to the standard ESMDA.(4)In addition to effective characterization of the DNAPL source zone architecture(SZA)before remediation,temporal monitoring of the SZ during remediation plays a vital role in all aspects of site stewardship.Therefore,we extended the CVAE parameterization method to real-time monitoring of DNAPL SZA with time-dependent state variables(i.e.,S_N).We combined the CVAE network with the En KF method.The En KF is used to estimate S_N and K_eff fields'temporal evolution by inverting multi-source data.To assess the proposed framework's performance,we conducted numerical experiments with evolving DNAPL SZAs in a 2D aquifer.Results show that the proposed framework significantly improved DNAPL saturation estimates throughout the source zone over time.CVAE-En KF reduced the estimation error of DNAPL mass remediated by 51%from the standard En KF.