Related Experiment Video
Updated: Oct 22, 2025
06:54
Photorealistic Learned Landscapes for Augmented Reality
Published on: June 27, 2025
311
Fast and scalable earth texture synthesis using spatially assembled generative adversarial neural networks
Sung Eun Kim1, Hongkyu Yoon2, Jonghyun Lee3
1Department of Safety and Environmental Research, The Seoul Institute, Seoul, South Korea; Civil and Environmental Engineering, University of Hawaii at Manoa, Honolulu, HI 96822, USA; Water Resources Research Center, University of Hawaii at Manoa, Hawaii, HI 96822, USA.
Journal of Contaminant Hydrology
|August 30, 2021
Summary
Spatially Assembled Generative Adversarial Networks (SAGANs) efficiently generate large geological textures from limited samples. This method enables realistic geomaterial reconstruction and analysis for subsurface applications like CO2 storage.
Area of Science:
- Geosciences
- Computational Science
- Materials Science
Background:
- Characterizing complex earth textures like shale and carbonate rocks is challenging due to sparse field samples and high costs.
- Generating large-scale geological textures with similar topological structures is crucial for realistic geomaterial reconstruction and hydro-mechanical evaluation.
- Generative Adversarial Networks (GANs) show promise for synthesizing geomaterial images for stochastic analysis, but face limitations in computational cost and output size scalability.
Purpose of the Study:
- To propose a novel Spatially Assembled Generative Adversarial Networks (SAGANs) method for generating arbitrary large geological textures efficiently.
- To evaluate the performance of SAGANs in generating 2D and 3D rock image samples for geostatistical reconstruction.
- To compare pore-scale flow patterns and upscaled permeabilities between training and SAGANs-generated geomaterial images using Lattice-Boltzmann simulations.
Main Methods:
- Development and implementation of the Spatially Assembled Generative Adversarial Networks (SAGANs) framework.
- Application of SAGANs to generate 2D and 3D rock image samples.
- Conducting Lattice-Boltzmann (LB) simulations to analyze fluid flow and permeability.
Main Results:
- SAGANs successfully generate arbitrary large-sized statistical realizations of earth textures with preserved connectivity and structural properties.
- Generated images exhibit flow characteristics similar to the original training images.
- SAGANs demonstrate the ability to produce diverse realizations from a single training image.
- Significant improvement in computational time compared to standard GANs frameworks was observed.
Conclusions:
- SAGANs offer an efficient and scalable solution for generating large, realistic geological textures.
- The method preserves essential geostatistical and hydro-mechanical properties, making it suitable for geomaterial reconstruction and analysis.
- SAGANs provide a versatile tool for creating multiple, statistically similar realizations for subsurface applications.
