Estimation of hydraulic conductivity fields through generative adversarial networks and Bayesian inference

Date

2025-6-20

Author

Çelebi, Emrehan Berkay

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

31
views

0
downloads

Cite This

Accurate characterization of the subsurface is critical for many engineering applications, ranging from groundwater supply to carbon sequestration. Mostly, inverse modeling is ill-posed due to numerous parameters to estimate and heterogeneity of the subsurface. Traditional solutions rely on simplification of conditions, however, emerging data heavy methods like deep learning algorithms can make more realistic assumptions about subsurface structure. This two-phase work employs generative adversarial networks (GAN) for accurate modelling of simulated hydraulic conductivity fields. First phase focuses on methodology development for 1-D aquifers and utilizes groundwater head data for inversion. Conditioning of GAN results is subsequently achieved through the Bayesian Inference based Markov Chain Monte Carlo (MCMC) algorithm. This phase also utilizes the ensemble Kalman filter (EnKF) method as a basis of comparison. Second phase expands the developed methodology to 3-D and aims to map the distribution of hydraulic conductivity in high resolution. Phase two also tests two additional deep learning models, convolutional neural networks and conditional GAN (cGAN). Furthermore, it compares the conditioning performances of MCMC and cGAN. For both phases, results show that MCMC conditioned GAN and cGAN can create highly accurate reconstructions of target fields. Results from the second phase additionally indicate that cGAN can achieve better conditioning than MCMC within the scope of this study. This work demonstrates the impressive performance of generative deep learning models in aquifer modelling. With increasing computing power and availability of data, such models are expected to become the established approach in handling ill-posed inversions.

Subject Keywords

Deep learning, Ensemble Kalman filter, Generative adversarial networks (GAN), Hydraulic conductivity, Markov Chain Monte Carlo (MCMC)

URI

https://hdl.handle.net/11511/115106

Collections

Graduate School of Natural and Applied Sciences, Thesis