Research paperReconstructing porous media using generative flow networks
Introduction
Understanding fluid flow in porous media at the micro-scale is relevant to many fields, such as oil and gas recovery, geothermal energy, geological CO2 storage, and battery energy storage. Porous medium properties, such as porosity and permeability, are often calculated from laboratory measurements or direct imaging of the microstructure (Walsh and Frangos, 1968, Ketcham and Carlson, 2001, Josh et al., 2012, Callow et al., 2020). Due to acquisition times and experimental costs, however, it is difficult to evaluate the variability of these properties across multiple macroscopic samples especially as sample permeability decreases. Instead, researchers often use statistical methods to reconstruct porous media based on either two-point or multi-point statistics (Okabe and Blunt, 2007, Wang et al., 2018). Reconstruction using these methods often requires prior knowledge about the pore- and throat-size distributions and can be time-consuming to generate multiple realizations of the same rock sample.
Various methods are available to reconstruct porous media, such as simulated annealing and multi-point statistical (MPS) methods. Simulated annealing can incorporate multiple correlation functions and statistical properties during reconstruction (Yeong and Torquato, 1998, Manwart et al., 2000, Pant and Berkeley, 2016). The method has been used to generate 2D and 3D Fontainebleau and Berea sandstone structures but can take hours to days to create large volumes (Manwart et al., 2000, Čapek et al., 2009, Pant and Berkeley, 2016). Extensions of the simulated annealing technique have also been used to generate multi-scale heterogeneous materials from multiple data sources, as well as generating 3D binary volumes from limited, multimodal data (Chen et al., 2016, Li et al., 2018). MPS techniques have been used to reconstruct large-scale reservoir systems, as well as micro-scale porous media, but also are slow to scale to generate multiple, large volumes (Caers, 2001, Okabe and Blunt, 2004, Okabe and Blunt, 2007, Bai and Tahmasebi, 2020). Pattern-based MPS methods have allowed for faster reconstruction of 3D networks using a cross-correlation based simulation method (CCSIM), but MPS techniques still struggle to capture long-range connectivity (Tahmasebi et al., 2012, Tahmasebi et al., 2014, Tahmasebi et al., 2017). Other recent methods for reconstructing three-dimensional porous volumes include texture synthesis and gradient-based methods as well (Liu and Shapiro, 2015, Fullwood et al., 2008).
Given the multi-scale nature of many rock systems, an ideal reconstruction method is one that recreates different length scales accurately and rapidly from a few datasets and is equally applicable to conventional and unconventional porous media. Recently, advances in machine learning, specifically in generative modeling, have focused on learning from limited datasets and generalizing these learned features for various applications (Goodfellow et al., 2014, Dinh et al., 2015, Zhu et al., 2017).
The major contribution of this work is the application of generative flow models to 3D volumes of porous media exhibiting grains, pore throats, and pore bodies. The advantage of this method is that the training is done on 2D grayscale images. The training time and model size are reduced, thereby improving scalability of the algorithm for training on larger image sizes. We generate 3D grayscale images via a latent space interpolation using a modifiable Gaussian distribution. The volumes generated are shown to be as useful as the original datasets for computing porosity, permeability, mercury injection capillary pressure, and other rock properties. We verify the accuracy of the generated images by calculating morphological parameters and comparing them against the real rock dataset. Accordingly, this paper proceeds by proposing implementation of generative flow models, metrics for evaluating the accuracy of model results, computation of permeability from 3D images, discussing results, and then conclusions.
Section snippets
Generative models
Deep-learning generative modeling is classified into two main approaches: generative adversarial networks (GANs) and likelihood-based methods. GANs are composed of two networks, a generator and discriminator, that work in competition against one another (Goodfellow et al., 2014). The generator attempts to recreate the training image as realistically as possible in order to fool the discriminator, that attempts to discern between a real or fake image. Deep convolutional GANs (DCGANs) are a class
Methods
The main contribution of this work is to present and evaluate a generative flow model to synthesize 3D porous media images and evaluate a series of domain-specific metrics to verify model accuracy. We demonstrate the applicability of the generative flow method using a sandstone dataset because data is widely available and petrophysical properties are fairly well-understood. We first present the generative flow model and interpolation technique, described in further detail in Anderson et al.
Dataset
The sandstone sample is 3 mm in diameter and 1 cm long. Although small, the sample provides significant information about the micro-structure of complex porous media at an optimal image resolution for sandstones. The dataset was obtained from micro-X-ray tomography scans of a Bentheimer sandstone core. This acquisition method produces a 3D volume depicting the rock interior in grayscale values. The full dataset dimensions are voxels with a voxel size of 1.53 µm. In order for the 2D
Results
The progress of training is presented first. This is followed by results for pore morphology and petrophysical results, primarily permeability and porosity.
Discussion
The average chord length of the pore space, a statistical property, is useful to inform accurate 3D generation of a model trained on a 2D dataset. The agreement of the two-point correlation function and Minkowski functionals support this method. As shown in Fig. 4, step size has an important effect on the 3D statistical properties of the generated volume. The step size during interpolation refers to the number of intermediate images generated in between a pair of anchor slices. Deviations
Conclusions
We developed and implemented a deep-learning method to generate rapidly 3D realizations of rock pore structure from 2D grayscale image slices of pore networks used as inputs into generative flow models. We apply the model to create volumes of a sandstone sample to demonstrate feasibility. The agreement of morphological and flow parameters confirm that the volumes created are realistic. The calculated pore properties, such as porosity, pore-size distribution, and permeability agree well with the
CRediT authorship contribution statement
Kelly M. Guan: Performed the NS flow calculations, Developed the image evaluation pipeline, assisted with development of the image synthesis algorithm by testing the algorithm across multiple datasets. Timothy I. Anderson: Developed and implemented the presented image volume synthesis algorithm. Patrice Creux: Provided the LBM, MICP, and PSD calculations and advised on the project. Anthony R. Kovscek: Acquired funding, conceptualized project, and supervised work.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported as part of the Center for Mechanistic Control of Water-Hydrocarbon-Rock Interactions in Unconventional and Tight Oil Formations (CMC-UF), an Energy Frontier Research Center funded by the U.S. Department of Energy, Office of Science, USA under DOE (BES) Award DE-SC0019165.
We thank Dassault Systèmes for providing computational resources for the LBM calculations. We thank the Stanford Center for Computational Earth & Environmental Sciences (CEES) for providing computational
References (61)
- et al.
Multimodal imaging and machine learning to enhance microscope images of shale
Comput. Geosci.
(2020) - et al.
Hybrid geological modeling: Combining machine learning and multiple-point statistics
Comput. Geosci.
(2020) Geostatistical reservoir modelling using statistical pattern recognition
J. Pet. Sci. Eng.
(2001)- et al.
Bonej: Free and extensible bone image analysis in imagej
Bone
(2010) - et al.
Gradient-based microstructure reconstructions from distributions using fast fourier transforms
Mater. Sci. Eng. A
(2008) - et al.
Laboratory characterisation of shale properties
J. Pet. Sci. Eng.
(2012) - et al.
Acquisition, optimization and interpretation of x-ray computed tomographic imagery: Applications to the geosciences
Comput. Geosci.
(2001) - et al.
Random heterogeneous materials via texture synthesis
Comput. Mater. Sci.
(2015) - et al.
MS-CCSIM : Accelerating pattern-based geostatistical simulation of categorical variables using a multi-scale search in fourier space
Comput. Geosci.
(2014) - et al.
Journal of natural gas science and engineering pore-scale simulation of gas flow in microscopic permeable media with complex geometries
J. Natural Gas Sci. Eng.
(2020)
User-controllable multi-texture synthesis with generative adversarial networks
Glow-PyTorch
GitHub Repository
Rockflow: Fast generation of synthetic source rock images using generative flow models
Energies
3D structural analysis: Sensitivity of Minkowski functionals
J. Microsc.
Novel morphometric analysis of higher order structure of human radial peri-papillary capillaries: relevance to retinal perfusion efficiency and age
Sci. Rep.
Optimal X-ray micro-CT image based methods for porosity and permeability quantification in heterogeneous sandstones
Geophys. J. Int.
Stochastic reconstruction of particulate media using simulated annealing: Improving pore connectivity
Transp. Porous Media
Parametric generation of conditional geological realizations using generative neural networks
Comput. Geosci.
Stochastic multi-scale reconstruction of 3D microstructure consisting of polycrystalline grains and second-phase particles from 2D micrographs
Metall. Mater. Trans. A
Reconstruction of porous media from extremely limited information using conditional generative adversarial networks
Phys. Rev. E
Microstructure synthesis using style-based generative adversarial networks
Phys. Rev. E
Generative-Adversarial-Nets
Nips
Generating sequences with recurrent neural networks
Effects of image resolution on sandstone porosity and permeability as obtained from X-Ray microscopy
Transp. Porous Media
Long short-term memory
Neural Comput.
Modeling heterogeneous materials via two-point correlation functions. II. Algorithmic details and applications
Phys. Rev. E
Cited by (14)
A multi-condition denoising diffusion probabilistic model controls the reconstruction of 3D digital rocks
2024, Computers and Geosciences3D reconstruction of digital cores based on multi-discriminator GAN and improved residual blocks
2023, Geoenergy Science and EngineeringMultiscale reconstruction of porous media based on multiple dictionaries learning
2023, Computers and GeosciencesContinuous conditional generative adversarial networks for data-driven solutions of poroelasticity with heterogeneous material properties
2022, Computers and GeosciencesCitation Excerpt :Kadeethum et al. (2021d) presented a data-driven generative adversarial networks (GAN) framework that can parameterize heterogeneous PDEs coefficients (i.e., heterogeneous fields), which has been demonstrated with steady-state cases for both forward and inverse problems. This GAN-based work could be considered as an extension of regression in subsurface physics through GAN model such as Zhong et al. (2019), Laloy et al. (2019) and Lopez-Alvis et al. (2021), in which heterogeneous fields are also parameterized through GAN model (Chan and Elsheikh, 2020; Hao et al., 2022; Guan et al., 2021) and subsequently used to predict the state variables (pressure and displacement responses). In Kadeethum et al. (2021d), the conditional GAN (cGAN) approach (Mirza and Osindero, 2014; Isola et al., 2017) was extended to the heterogeneous field for both generator and critic (i.e., discriminator) where usage of Earth mover’s distance through Wasserstein loss (W loss) and gradient penalty (Arjovsky et al., 2017; Gulrajani et al., 2017) improved the model accuracy and training stability compared to the traditional GAN approach.
Study of mixed convective–radiative fluid flow in a channel with temperature-dependent thermal conductivity
2022, Partial Differential Equations in Applied MathematicsStriving to translate shale physics across ten orders of magnitude: What have we learned?
2021, Earth-Science ReviewsCitation Excerpt :The synthesis is comparable in quality to 3D GANs. When tested on a Bentheimer sandstone, the reconstructions matched the permeability and Minkowski functionals of the training data (Guan et al., 2021). The approach was further generalized to reconstruct multimodal (TXM and SEM) images of shales (Anderson et al., 2020b), making it applicable to other image types.