Generative Reinforcement Learning for Characterization of Heterogeneous Aquifer Structures in a Latent‐Space Search Framework
Water Resources Research · 2026-11 · 62 (11)
DOI: 10.1029/2025wr042799 ↗Kurz gefasst
Fachartikel in der Zeitschrift Water Resources Research, online erschienen am 14.11.2026. Verfasst von Zidong Pan und 10 weiteren Autor:innen. Die Arbeit nennt 79 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: National Key Research and Development Program of China.
Abstract (Originalsprache)
Accurate characterization of aquifer structure is essential for groundwater management, contamination risk assessment, and remediation design. However, high‐resolution characterization involves inversion problems with more than computational nodes. This study proposes a generative reinforcement learning (GRL) framework for aquifer structure characterization using borehole and tracer‐test data. A β ‐variational autoencoder compresses three‐dimensional facies fields into latent representations, while a Wasserstein generative adversarial network maps 128‐dimensional standard Gaussian variables to the learned latent space, enabling inversion in the 128‐dimensional input space. A Fourier neural operator maps the latent representations to transient three‐dimensional tracer concentration fields. During inversion, concentrations at discrete monitoring locations are extracted from these fields and compared with observations. Candidate solution ensembles generated by Make Variance Reduction Shine (MARS) and ensemble smoother with multiple data assimilation (ES‐MDA) form an inversion experience pool, from which an actor network refines the 128‐dimensional variables using concentration mismatch and training‐derived facies‐proportion constraints. The framework was evaluated using a synthetic heterogeneous‐aquifer tracer test with 5% relative Gaussian observational noise. Actor refinement of the MARS‐derived solutions produced the maximum facies reconstruction accuracy of 84.62% (defined as the proportion of correctly reconstructed facies grid cells), compared with 83.62% before refinement and 82.02% for the unrefined ES‐MDA solutions, while retaining variability among the realizations. In a field tracer‐test application, the best‐performing GRL solution reproduced the observed breakthrough curve with an R 2 of 0.95. These results indicate that the framework can integrate geological conditioning and sparse tracer observations to generate multiple data‐consistent representations of heterogeneous aquifer structure.
Förderung
- National Key Research and Development Program of China
Zitieren
APA
Pan, Z., Guo, Z., Zhao, D., Zhu, Y., Chen, K., Liang, X., Yin, M., Li, H., Zeng, Z., Shen, H., & Zheng, C. (2026). Generative Reinforcement Learning for Characterization of Heterogeneous Aquifer Structures in a Latent‐Space Search Framework. Water Resources Research, 62(11). https://doi.org/10.1029/2025wr042799
Vancouver
Pan Z, Guo Z, Zhao D, Zhu Y, Chen K, Liang X, et al. Generative Reinforcement Learning for Characterization of Heterogeneous Aquifer Structures in a Latent‐Space Search Framework. Water Resources Research. 2026;62(11). doi:10.1029/2025wr042799
BibTeX
@article{Pan2026,
author = {Pan, Zidong and Guo, Zhilin and Zhao, Dongwei and Zhu, Yuguang and Chen, Kewei and Liang, Xiuyu and Yin, Maosheng and Li, Hongkai and Zeng, Zhenzhong and Shen, Huizhong and Zheng, Chunmiao},
title = {Generative Reinforcement Learning for Characterization of Heterogeneous Aquifer Structures in a Latent‐Space Search Framework},
journal = {Water Resources Research},
year = {2026},
volume = {62},
number = {11},
doi = {10.1029/2025wr042799}
}
EUVIMED