Source-linked AI summary
Stochastic reconstruction of an oolitic limestone by generative adversarial networks
Lukas Mosser, Olivier Dubrule, Martin J. Blunt
TL;DR
The paper addresses the need for representative three-dimensional porous-media samples when micro-CT data are limited and do not reveal structural variability. It trains a GAN on an oolitic Ketton limestone micro-CT dataset and evaluates generated images using morphology, effective properties, and velocity statistics. The generated samples reproduce the evaluated statistical and effective properties, while the authors identify diversity and mode-collapse as open challenges.
Problem
Limited micro-CT sample size provides insufficient insight into the variance of key micro-structural properties needed for representative porous-media modeling.
Method
The paper trains a GAN on gray-scale volumetric micro-CT data from oolitic Ketton limestone to generate stochastic three-dimensional reconstructions.
Results
Generated images capture textural features, Minkowski functionals, statistical properties, effective properties, permeability, and pore-scale velocity distributions of the Ketton limestone data.
Takeaways & Limitations
GANs can represent the evaluated statistical and effective properties of porous-media microstructures while enabling stochastic reconstruction of gray-scale volumetric images.
Takeaways & Limitations
Generated samples show less variation than training samples, and mode collapse or low diversity may remain undetected by morphology and effective-property evaluations.
Abstract
from arXiv · showhide
Stochastic image reconstruction is a key part of modern digital rock physics and materials analysis that aims to create numerous representative samples of material micro-structures for upscaling, numerical computation of effective properties and uncertainty quantification. We present a method of three-dimensional stochastic image reconstruction based on generative adversarial neural networks (GANs). GANs represent a framework of unsupervised learning methods that require no a priori inference of the probability distribution associated with the training data. Using a fully convolutional neural network allows fast sampling of large volumetric images.We apply a GAN based workflow of network training and image generation to an oolitic Ketton limestone micro-CT dataset. Minkowski functionals, effective permeability as well as velocity distributions of simulated flow within the acquired images are compared with the synthetic reconstructions generated by the deep neural network. While our results show that GANs allow a fast and accurate reconstruction of the evaluated image dataset, we address a number of open questions and challenges involved in the evaluation of generative network-based methods.
1 Introduction
The paper motivates stochastic reconstruction as a way to generate representative porous-media volumes beyond limited micro-CT samples. It introduces GANs for reconstructing unsegmented Ketton limestone and evaluating both morphology and flow-related properties.
- Motivation: Micro-CT imaging resolves pore-scale structure but increased resolution limits acquisition speed and sample size.Individual samples also provide little information about variability in micro-structural properties.
- Prior approaches: Existing reconstruction approaches include statistical, object-based, process-based, and training-image-based methods.Examples include simulated annealing, covariance-based simulation, Boolean models, process models, multiple-point statistics, direct sampling, and image quilting.
- Proposed approach: GANs provide a training-image-based approach for fast generation of large volumetric images capturing statistical and morphological features.The method extends prior GAN reconstruction work on segmented porous-media images.
- Study scope: The study applies GANs to an unsegmented gray-scale micro-CT scan of a larger oolitic Ketton limestone sample.Using gray-scale data enables evaluation of Minkowski functionals across gray-level thresholds.
- Evaluation: The reconstructions are compared using Minkowski functionals, permeability, and velocity distributions.These evaluations address both image morphology and simulated flow behavior.
2 Generative Adversarial Networks
GAN reconstruction learns an implicit image distribution by training a generator against a discriminator, then samples new images from noise. Training is computationally demanding and sensitive to architecture choices, whereas generation is comparatively fast.
- GAN framework: GANs generate samples from arbitrary probability distributions without imposing an a priori density model.This implicit formulation supports sampling from high-dimensional, intractable image distributions.
- Training data: The training data comprise 643 non-overlapping voxel sub-domains extracted from the original micro-CT image.Each sub-domain represents an independent part of the acquired dataset.
- Network roles: The generator maps a noise prior z to image samples, while the discriminator estimates whether an image belongs to the true data distribution.The generator and discriminator are represented by differentiable neural networks.
- Optimization: Training alternates between improving the discriminator on real and generated samples and updating the generator to fool the discriminator.The updates use backpropagation and stochastic gradient descent with the other network held fixed in each step.
- Training challenges: GAN training is expensive for three-dimensional data and may require substantial trial and error to select architecture and hyperparameters.Large GPU memory is needed because three-dimensional filter maps and images increase computational demands.
- Generation: Individual sample generation is extremely fast because generator operations are matrix-vector computations executed efficiently on modern hardware.Generation takes on the order of seconds for modern CPUs.
3 Dataset
The study trains a fully convolutional GAN on Ketton limestone micro-CT data and evaluates synthetic reconstructions against original sub-domains using structural, morphological, permeability, and velocity measures.
- Dataset: The dataset comprises oolitic Jurassic limestone imaged at 27.8 µm voxel size and subdivided into non-overlapping 64^3 training images.The original image domain contains 900^3 voxels, and evaluation uses larger 200^3 domains to test generalization.
- Neural Network Architecture: The generator uses a fully convolutional DCGAN architecture with transposed convolutions for upsampling and an added convolutional layer to reduce artifacts.The generator can produce images of arbitrary spatial size by using latent vectors with larger spatial dimensions.
- Neural Network Architecture: The discriminator is a three-dimensional convolutional classifier that distinguishes real 64^3 training images from synthetic realizations.Its layers use convolution, batch normalization, and LeakyReLU activation, with a final sigmoid output representing the probability of belonging to the training set.
- Evaluation: After training, the generator produced 64 synthetic 200^3 reconstructions for comparison with 64 non-overlapping Ketton validation sub-domains.The samples were randomly generated rather than selected for visual or statistical quality.
- Evaluation: The reconstructions reproduced the hole-effect in S2(r), showed good radial-average agreement, and matched porosity, specific surface area, and integral mean curvature across thresholds.The specific Euler characteristic had a 20% error within the global-thresholding range, requiring care during segmentation.
- Evaluation: Synthetic samples captured the average permeability and produced velocity distributions visually similar to the real images, although low velocities were more abundant and high velocities less abundant.For all three flow directions, the velocity-distribution null hypothesis was accepted at the 5% significance level.
4 Discussion
The GAN reconstructions reproduce key morphological and effective-property behavior, while revealing limited sample variation, possible mode collapse, and architectural artifacts. Layer-wise inspection clarifies how global and detailed features emerge and highlights unresolved challenges in GAN evaluation and training.
- Reconstruction quality: Generated images capture the training image’s textural features, while two-point statistics and effective properties show excellent agreement with Ketton subsets.These evaluations were performed on segmented representations of individual sub-domains.
- Reconstruction quality: All effective, statistical, and morphological properties cluster tightly around the training-image average, indicating less variation in generated samples.This reduced variation may reflect mode collapse or other limitations of GAN training.
- Latent-space behavior: Smooth latent-space interpolation indicates that the generator did not memorize the training set and learned meaningful lower-dimensional pore-grain representations.The authors identify high-diversity, mode-collapse-free, stable GAN objectives as an open problem.
- Network interpretation: Early generator layers represent global image features, whereas later layers construct highly detailed structural features through multi-scale upsampling.Layer-wise analysis also exposes repeated noise and architectural deficiencies in intermediate layers.
- Network limitations: Transposed convolutions can produce grid-like artifacts, partly reduced by an added convolution before the final upsampling operation.Sub-pixel or interpolation-based upsampling is suggested as a possible alternative.
- Network interpretation: The discriminator learns pore-space and edge features in shallow layers, while deeper layers encode more abstract representations; its scores may support selecting higher-quality reconstructions.The discriminator is also proposed as a possible future classifier or feature representation.
5 Conclusion
The study demonstrates that GANs can reconstruct porous-media microstructures from grayscale volumetric images while recovering structural and flow-related properties. Training remains difficult because of extensive hyperparameter searches and learning instabilities, but the training cost is useful when many reconstructions are needed.
- GANs reconstruct oolitic Ketton limestone microstructures from grayscale volumetric images.
- The synthetic models recover statistical and effective properties of segmented pore spaces, Minkowski functionals across gray-level thresholds, and pore-scale velocity distributions.
- Effective permeability captures a global average of the velocity field, while velocity distributions provide an additional recovered flow-property assessment.
- The GAN learns a multi-scale pore-space representation through inference from a latent noise prior.
- Large hyperparameter searches and learning instabilities make GAN training difficult, but its computational cost is useful for generating very many stochastic reconstructions.