Conditional generative adversarial networks for groundwater contamination characterization and source identification

Abstract

Accurate characterization of sources and spatial distribution of contamination remains crucial in groundwater survey and remediation efforts. Nevertheless, conventional geo-statistical methods, like kriging, usually produce overly smooth concentration fields that violate the transport physics. Furthermore, the identification of contaminant sources based on kriging methods may be misleading. As a deep learning approach, Generative adversarial networks (GANs) is being increasingly employed to extract patterns and insights from intricate geospatial data. The conditional variant of GANs (CGANs), can reconstruct complete spatial data from sparse point observations. In this study, we propose a deep learning framework with CGANs for the purpose of groundwater contamination characterization and source identification. In ensure compliance with the principles of transport physics, we use a numerical transport model as the underlying prior model for training the CGANs. Through the adaptation of network architecture, the generator transforms low-dimensional random variables and observation data into concentration images, while the discriminator is responsible for extracting source information from these images. The developed framework is tested through two numerical cases with channelized random hydraulic conductivity fields, which pose a challenge due to the complexity of the concentration fields. The results demonstrate that the proposed method can give accurate results for both contaminant characterization and source identification in groundwater. CGANs excel in reconstructing concentration fields compared to traditional interpolation methods such as Kriging and inverse distance weighting (IDW). Additionally, CGANs offer the capability to estimate uncertainty and exhibit a certain level of error robustness.