Related Experiment Video
Updated: Aug 28, 2025
08:02
Reservoir Condition Pore-scale Imaging of Multiple Fluid Phases Using X-ray Microtomography
Published on: February 25, 2015
12.6K
Improved recurrent generative model for reconstructing large-size porous media from two-dimensional images
Fan Zhang1,2, Qizhi Teng1, Xiaohai He1
1College of Electronics and Information Engineering, Sichuan University, Chengdu 610065, China.
Physical Review. E
|September 16, 2022
Summary
This study introduces an improved recurrent generative model for 3D porous media reconstruction, achieving higher resolution and accuracy. The new model enhances detail preservation and continuity in 3D structures, crucial for physical property analysis.
Area of Science:
- Porous media research
- Computational modeling
- Deep learning applications
Background:
- 3D structure modeling of porous media is vital for physical property analysis.
- Existing deep learning methods for 3D reconstruction have limitations in resolution and detail preservation.
- Previous models like 3D-PMRNN expanded reconstruction size but still required down-sampling, losing information.
Purpose of the Study:
- To develop an improved recurrent generative model for enhanced 3D porous media reconstruction.
- To overcome the limitations of existing deep learning models in terms of reconstruction size and detail preservation.
- To enable accurate simulation and analysis of physical properties by improving 3D structure reconstruction.
Main Methods:
- Proposed an improved recurrent generative model with enhanced reconstruction ability up to 512^3.
- Incorporated a hybrid receptive field for convolutional neural network kernels.
- Integrated an attention-based module and introduced a section loss for improved Z-direction continuity.
- Utilized a single 3D training sample and layer-by-layer generation approach.
Main Results:
- Achieved high-resolution 3D reconstruction (512^3) with improved accuracy, diversity, and generalization.
- Demonstrated effective preservation of detailed information from original images.
- Validated the effectiveness of the section loss in enhancing structural continuity.
- Experimental results confirmed the model's superior reconstruction capabilities compared to existing methods.
Conclusions:
- The proposed improved recurrent generative model significantly enhances 3D porous media reconstruction capabilities.
- The model's architecture and novel components (hybrid receptive field, attention, section loss) contribute to superior performance.
- This advancement facilitates more accurate analysis of physical properties in porous media through high-fidelity 3D modeling.
