Joint use of ground-penetrating radar and hydraulic-partitioning tracer tomography to improve characterization and long-term prediction of DNAPL source zones in multi-facies aquifers.

Facies-controlled heterogeneity, including fine-grained lenses and abrupt textural transitions, produces irregular and non-Gaussian DNAPL source zone architecture (SZA) that strongly influences mass-transfer and long-term dissolution. Reconstructing both permeability and DNAPL saturation therefore requires estimating high-dimensional, coupled parameter fields that are poorly constrained by sparse hydraulic and tracer observations. Geophysical methods such as electrical resistivity tomography (ERT) and ground-penetrating radar (GPR) can supplement these data, yet ERT often lacks resolution in shallow, thinly layered systems, whereas GPR is sensitive to dielectric contrasts associated with facies and DNAPL distributions. We develop an inversion framework that integrates hydraulic-partitioning tracer tomography with ground-penetrating radar using a deep-learning parameterization and ensemble data assimilation. A convolutional variational autoencoder (CVAE) is trained to provide a non-Gaussian parameterization (prior) of permeability and DNAPL saturation. An ensemble smoother with multiple data assimilation (ESMDA) updates the latent variables z conditioned on hydrogeophysical observations while keeping the CVAE weights fixed. Synthetic experiments show that GPR resolves facies and SZA geometry more effectively than ERT, and that combining GPR with hydraulic-partitioning tracer tomography reduces permeability error by 7.9% and DNAPL saturation error by 25.7% relative to HPTT-only inversion, while lowering false positives and false negatives in SZA delineation. Enhanced SZA reconstruction further yields more reliable long-term predictions of DNAPL mass decay and dissolved-plume evolution. The results demonstrate the value of joint use of GPR and tracer data to characterize complex, non-Gaussian DNAPL source zones in shallow aquifers.