The overarching objective of this research was to develop and demonstrate a comprehensive approach for field characterization of dense non-aqueous phase liquid (DNAPL) source zones, which quantifies the key features that control plume response. The intent was to integrate targeted (local-scale) in situ tests with transect-based observations of downstream contaminant flux or concentration and information on subsurface geologic variability. Specific objectives included: (1) identification of the most information rich metrics for linking NAPL architecture to plume response; (2) development and refinement of in situ test methods and modeling tools that can be used to quantify identified metrics in targeted regions of the source zone; (3) integration of these metrics and tools with current machine learning characterization methods for an overall source zone assessment protocol; and (4) development of simplified models for prediction of plume response.

The research approach integrated batch, column, and aquifer cell experiments with mathematical modeling and data processing tools to identify and quantify features of the DNAPL architecture controlling downgradient plume response.

A comprehensive set of experiments and numerical simulations was undertaken to quantify important source zone metrics that control downgradient plume response and to create observation data for subsequent characterization modeling tool development. A large library of three-dimensional (3D) field-scale DNAPL source zone scenarios was created to encompass a realistic range of release conditions and permeability distributions. Stochastic permeability fields were generated using two distinct models: sequential Gaussian simulation and Markov-chain transitional probability. Simulations confirmed the strong influence of source zone architecture (SZA) (e.g., pool fraction) on long-term dissolution behavior. Modeling results indicated that the DNAPL release rate, capillary pressure parameters, and the distribution of persistent low permeability layers were the most significant factors influencing DNAPL source zone metrics and downgradient plume response. For highly heterogeneous permeability fields, results suggest that analyses must be conducted in 3D to provide reliable predictions. Heterogeneous source zone aquifer cell experiments confirmed the mathematical model results, demonstrating the strong influence of SZA on dissolution behavior. Experiments with discrete lenses, as well as with more realistic stochastic permeability fields modeled using Markov-Chain transitional probability theory, were observed to give rise to two-stage mass flux behavior, attributed to persistent pools. Comparisons between mathematical model predictions and laboratory measurements revealed the importance of considering measurement scale in defining local saturations and average source zone metrics. A consistent methodology was recommended and demonstrated to determine saturation distributions from aquifer cell light transmission data.

The second task of the project dealt with the estimation of local saturation and SZA metrics using partitioning tracer tests. This task was completed through the coupling of mathematical modeling with batch, column, and aquifer cell experiments. Analytical models for push-pull tracer tests analysis typically assume linear equilibrium partitioning, a uniform NAPL saturation, and a homogenous medium. The reasonableness of each of these assumptions was investigated. Application of equilibrium partitioning behavior failed to reproduce column experimental observations. A linear driving force model, however, was shown to provide excellent predictions. Diffusion limitations within the NAPL and surface accumulation at the NAPL interface were found negligible for realistic source zone conditions. Tracer test interpretation was also shown to be highly sensitive to the tracer water/NAPL partitioning coefficient, with application of common group contribution methods for partitioning coefficient estimation potentially leading to underestimation of NAPL saturations. Batch measurements for representative tracers demonstrated that partitioning behavior is nonlinear but may be approximated with a linear function below specified concentrations. Application of kinetic partitioning with the assumption of a uniform NAPL saturation was shown to provide a reasonable match to recovery curves obtained in heterogeneous 2D aquifer cell experiments. An empirical approach was developed to interpret recovery curves, offering potential for identifying the distance to and the fraction of the vertical domain occupied by NAPL. Development and application of a coupled adjoint sensitivity method to the transport equations suggests the promise of this method for estimation of total DNAPL mass, average saturation, and distance of DNAPL from the push-pull well from concentration breakthrough observations.

In the third project task, machine learning methods were developed and used to successfully estimate SZA metrics (mass of pools, mass of ganglia, and pool fraction) from observations of plume concentrations in a downgradient transect. The library of source zone realizations created in the first task was used to train and test the models. A fundamentally new machine learning processing method was developed employing ideas from manifold learning and embedding for estimating each metric individually. This approach was also extended using ideas from multi-task learning to determine the three metrics jointly based on enforcing a physical relationship among the three. Strong performance can be obtained both with densely sampled data as well as when a more limited sampling of transect concentration data are available, as is the case with typical laboratory experimental field applications.

In task four, an improved upscaled model was developed to predict flux averaged concentration evolution downgradient of DNAPL source zones. Inputs to the model include the initial pool fraction, the initial flux-averaged concentration, and the initial fraction of that flux eluting from pool zones. This enhanced model improved on the upscaled model developed under previous SERDP support by enabling the prediction of two-stage mass recovery behavior that may be observed in the field, especially at “aged” sites. An upscaled model was also developed and validated for application for the interpretation of two-well and push-pull partitioning tracer tests. The model incorporates rate-limited partitioning though an effective mass transfer coefficient correlation that depends on four SZA parameters: the vertical spread of the DNAPL; the injection/extraction rate; pool fraction; and average DNAPL saturation. The application of the upscaled model was demonstrated for simulations of interwell and push-pull tracer tests, producing accurate estimates of average NAPL saturation (DNAPL mass).

This research provides an improved understanding of the coupling of downgradient plume response to DNAPL architecture and highlights the importance of measurement scale and mass transfer limitations to source zone characterization. The suite of data processing techniques and upscaled models developed in this research offers site managers specific tools that can be employed for source zone characterization and remedy screening. These tools include methods for the design and evaluation of localized push-pull partitioning tracer tests; machine learning techniques for the estimation of DNAPL source zone metrics from plume transect concentration observations; and simplified screening models that incorporate these metrics to predict plume evolution and persistence.