Characterizing Multi‐Source DNAPL Source Zone Architecture in Channelized Aquifers Using a Geophysics‐Informed Hybrid Deep Learning Framework

Abstract Accurate characterization of dense nonaqueous phase liquid (DNAPL) source zone architecture (SZA) is essential for contaminated site remediation. However, highly channelized aquifer heterogeneity and multi‐source contamination pose significant challenges. This study proposes a geophysics‐informed hybrid deep learning framework for SZA characterization under such conditions. First, an improved generative adversarial network (GAN) is employed to generate channelized aquifer structures. Then, a convolutional neural network (CNN)‐based surrogate model is developed for rapid DNAPL multiphase flow simulation. Finally, GAN and CNN are integrated with a geophysical model to construct a data assimilation‐based inversion framework, termed GAN‐CNN‐Geophysics‐Inversion (GCGI). This framework achieves SZA identification by assimilating electrical resistivity tomography observations. A case study demonstrates that GCGI accurately identifies channelized permeability fields, leakage locations and rates, enabling reliable reconstruction of DNAPL saturation distributions. This study highlights the potential of integrating deep learning and geophysical methods for characterizing subsurface multiphase flow systems.