Understanding how pollutants move underground is difficult because aquifers are far from uniform. Water can travel quickly through some zones and slowly through others, while most of these hidden structures cannot be directly observed. Researchers have now developed an artificial intelligence framework that can represent this three-dimensional complexity more efficiently and use monitoring data to improve predictions of PFOA movement in groundwater.
The study introduces a variational autoencoder-least squares generative adversarial network, or VA-LSGAN , that compresses complex three-dimensional hydraulic conductivity fields into a much smaller set of variables. The approach also preserves important spatial patterns that conventional variational autoencoders can smooth out.
“Groundwater contamination models must account for subsurface heterogeneity, but directly estimating every property at every location quickly becomes computationally difficult,” said corresponding author Zhilin Guo of Southern University of Science and Technology . “Our framework provides a compact way to represent these complex structures while retaining the spatial information that matters for contaminant transport.”
Aquifer properties such as hydraulic conductivity can vary strongly across space. In the study, each simulated conductivity field contained 131,072 grid-scale parameters , which the neural network compressed into a 1,024-dimensional latent representation . This reduction makes it more practical to update subsurface properties during inverse modeling.
The researchers combined a three-dimensional variational autoencoder with adversarial learning. The adversarial component encouraged the reconstructed aquifer fields to retain sharper local contrasts and more realistic spatial continuity. Although the conventional VAE achieved slightly better point-by-point reconstruction metrics, VA-LSGAN reproduced the spatial autocorrelation structure of the reference fields more closely, particularly in the horizontal directions.
The team also introduced a principal component analysis based latent covariance sampling strategy . Instead of independently sampling each latent variable, the method preserves statistical relationships learned from the training data. This substantially improved the similarity between generated and reference hydraulic conductivity distributions. The Wasserstein distance for hydraulic conductivity fell from 1.9219 with standard Gaussian sampling to 0.3817 with PCA based sampling .
To test the framework in a practical setting, the researchers applied it to a PFOA contaminated groundwater site in Pingshan, Shenzhen, China . The AI generated conductivity fields were coupled with MODFLOW and MT3D-USGS groundwater flow and contaminant transport models. Monitoring data were then incorporated using an ensemble smoother with multiple data assimilation.
The results showed a substantial reduction in predictive uncertainty. The width of the simulated PFOA concentration envelope at the monitoring well decreased from 420.9 ng/L to 137.0 ng/L, a 67.5% reduction. The ensemble mean root mean square error also fell from 34.61 ng/L to 1.70 ng/L after data assimilation. Importantly, the method retained multiple plausible subsurface configurations instead of producing a single apparently certain solution.
“The remaining uncertainty is important rather than undesirable,” Guo said. “Different underground structures can sometimes produce similar monitoring responses. Representing those possibilities helps avoid overstating how much can be learned from limited field observations.”
The researchers note that future work should incorporate additional monitoring constraints and evaluate the framework across a broader range of field-scale contamination scenarios.
===
Journal reference: Pan Z; Guo Z; Zhao D; et al. Variational autoencoder-least squares generative adversarial Network (VA-LSGAN): a latent-space parameterization framework linking three-dimensional aquifer heterogeneity characterization and perfluorooctanoic acid (PFOA) inverse modeling. AI Environ. 2026, 1(3): 189-202. DOI: 10.66178/aie-0026-0022
https://www.the-newpress.com/aie/article/doi/10.66178/aie-0026-0022
About the Journal:
Artificial Intelligence & Environment is an international multidisciplinary platform for communicating advances in fundamental and applied research on the intersection of environmental science and artificial intelligence (AI). It is dedicated to serving as an innovative, efficient and professional platform for researchers in the cross-discipline fields of earth and environmental sciences, big data science and AI around the world to deliver findings from this rapidly expanding field of science. It is a peer-reviewed, open-access journal that publishes critical review, original research, rapid communication, view-point, commentary and perspective papers.
Follow us on Facebook , X , and Bluesky .
Experimental study
Variational autoencoder-least squares generative adversarial Network (VA-LSGAN): a latent-space parameterization framework linking three-dimensional aquifer heterogeneity characterization and perfluorooctanoic acid (PFOA) inverse modeling
16-Sep-2026