This article is licensed under CC-BY-NC-ND 4.

http://pubs.acs.org/journal/acsodf Article

Machine Learning Assisted Prediction of Porosity and Related

Properties Using Digital Rock Images
Md Irfan Khan and Aaditya Khanal*

Cite This: ACS Omega 2024, 9, 30205−30223 Read Online

ACCESS Metrics & More Article Recommendations

ABSTRACT: Accurately estimating reservoir rock properties is para-

mount for modeling the storage and flow of fluids (hydrocarbon, carbon
dioxide, and groundwater) in porous media. However, existing laboratory
techniques to measure rock properties are usually time-consuming,
expensive, and computationally intensive. This work proposes an efficient
workflow that uses the machine learning algorithm, based on the
convolutional neural network (CNN) framework, to predict rock
properties from microcomputed tomography (micro-CT) X-ray images.
The workflow involves data preprocessing, label extraction, training, and
prediction using the segmented images of the rock to predict porosity,
throat area, and pore surface area, which are essential for pore-scale
modeling. The model was trained and validated on the Bentheimer
sandstone, which was then used to predict properties of other sandstones
(Castlegate and Leopard) with different pore structures and flow
properties. The model yielded a good prediction for the throat and pore surface area but a significant error for porosity.
Subsequently, a new complex model was trained and validated using diverse images from Bentheimer and an additional rock
Castlegate, which was then used to predict the properties of Leopard sandstone. The new model improved the prediction of each
property, resulting in mean absolute percentage error (MAPE) values of 2.19%, 3.04%, and 6.08% for porosity, pore surface area, and
throat area with the binary images, respectively. In addition, we present a novel data-driven method using a simple regression model
to predict the absolute permeability of a digital rock sample using the pore network parameters as predictors. The extreme gradient
boost (XGBoost), which performed the best among several machine algorithms, was trained and validated using digital rock images
from Bentheimer and Castlegate sandstone. The generated model was then used to predict the absolute permeability of the Leopard
sandstone with an R2 of 0.813, which was a significant improvement over the model generated solely by using either the Bentheimer
or the Castlegate sandstone images. Furthermore, our analysis showed that the tortuosity had the most significant effect on the
absolute permeability prediction of the rock sample. This study showed that we can reliably predict the morphological properties of
porous media using computationally efficient models generated from digital rock images, which can be used to build a regression
model to predict the crucial petrophysical properties needed to model the flow of fluids in porous media.

1. INTRODUCTION whereas other complex properties (capillary pressure, relative

The reservoir rocks contain complex networks of interconnected permeability, and wettability) require SCAL. However, the data
and isolated pores, which create an ideal environment for collected from the core analysis may have significant errors due
trapping hydrocarbons in the porous media. Furthermore, the to issues such as the presence of hydratable clays, miscalibration
connected pore networks dictate the flow behavior and of equipment, absorption of atmospheric water in the core, loss
production potential of the trapped hydrocarbons by directly of core samples due to chipping, and other reasons.5 As a result,
affecting the intrinsic permeability.1 Thus, the accurate the core analysis data may not always be reliable for modeling
determination of the reservoir rock properties, such as porosity, the storage and flow behavior of the reservoir rocks.6 For these
throat area, and permeability, is essential for predicting the initial
hydrocarbon in place, estimating the ultimate recovery,
identifying the storage potential for CO2 sequestration and Received: December 22, 2023
hydrogen storage, and modeling groundwater flow in subsurface Revised: April 9, 2024
porous media.2−4 These properties are usually determined from Accepted: May 7, 2024
reservoir rock samples or the core through routine and special Published: June 25, 2024
core analyses (RCA/SCAL) in the lab. The fundamental
properties, such as porosity, are measured during the RCA,
© 2024 The Authors. Published by
American Chemical Society https://doi.org/10.1021/acsomega.3c10131
30205 ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

reasons, there is a significant push toward reliable and prediction, and acquiring high-quality 3D images of rock
economical alternative and complementary methods to estimate samples is expensive. This complexity may explain why, in
these properties. The three-dimensional rock images obtained many previous studies,31−33 artificially generated sphere packs
from advanced imaging techniques, often referred to as the are used to predict the petrophysical properties of rock samples.
“digital rock”, provide an alternative to laboratory-based Second, the 3D CNN algorithm is highly computationally
methods. intensive and requires excessive memory.34 In contrast, simple
Digital rock technology is an effective and accepted tool for regression models like random forest (RF), support vector
extracting various petrophysical and flow properties from machine (SVM), and gradient boost (GB) algorithms are much
reservoir rocks. Numerical simulation can be applied to data easier to interpret and less memory intensive.35,36 Further
obtained from digital rocks to compute mechanical properties,7,8 investigation is required to define a systematic workflow
predict fluid topology,9 estimate flow properties,10,11 and focusing on data preprocessing and predicting various
identify electrical properties12 of the rock by solving the petrophysical properties and microstructural characteristics of
governing equations of different pore-scale phenomena. The a digital rock sample. Furthermore, although numerous
typical digital rock physics workflow can be divided into the previously published works have used an image-based approach
following steps: (a) acquiring raw grayscale digital rock images, using CNN to predict permeability, significantly less focus has
(b) segmentation of digital rock images, and (c) calculating been placed on simple regression models, which are computa-
petrophysical properties by numerical simulation.13 Segmenta- tionally less intensive than CNN.
tion is a process of identifying all the phases (pore vs matrix) in In this work, a step-by-step procedure for predicting the
an image correctly and distinctly. This process is crucial in digital morphological descriptors, namely, porosity (Ø), pore surface
rock analysis and influences pore-scale characterization.14,15 area (PSA), and throat area (TA) from the digital rock, is
Inaccurate segmentation of the digital rock leads to significant presented using a convolutional neural network (CNN).
errors in subsequent quantitative analysis. There are numerous Utilizing the framework, the end user can extract the properties
segmentation methods, among which machine learning (ML) related to the pore network of an unknown rock sample using
algorithms are becoming more popular due to their relative only raw micro-CT images, which can later be used to build a
accuracy and ease over traditional methods.16,17 ML methods data-driven machine-learning model to predict permeability. To
have been found to be useful in solving complex problems in train the machine-learning model using CNN, we used micro-
several areas of science and engineering by identifying CT images of multiple rocks (Bentheimer, Castlegate, and
relationships between inputs and outputs, deciphering patterns, Leopard sandstones) with different pore structures. The trained
and generating solutions to complex problems. Petroleum model is then used to predict the morphological properties of
industry also finds extensive use of ML in different areas, other rocks with different properties. First, three models are
including reservoir engineering,18 production engineering,19 trained with binary and grayscale images of two different rocks.
drilling engineering,20 and other related areas.21,22 Two of these models are only trained with images of either
The application of ML continues to grow for other Bentheimer and Castlegate sandstone. The third model is
engineering domains, including digital rock analysis for property trained with the images of both Bentheimer and Castlegate
and flow behavior estimation. Tembely and AlSumaiti23 sandstone. Next, the trained model was used to predict the
developed a workflow for fast and accurate prediction of the properties of Leopard sandstone, which were unseen by each of
permeability of complex networks using deep learning methods. the three models. Our investigation reveals that predata
Their work suggested that the deep neural network performs processing of the rock images significantly improves prediction
slightly better than gradient boosting and linear regression for accuracy; thus, this is a crucial step. In addition, the results
the permeability calculation. Linden et al.24 investigated the showed that skewed data sets and defective images could
relationship between fluid flow at the macroscale and internal negatively affect the performance of the trained model. The
pore structure in porous media by applying machine learning novel workflow presented here expands on the limited literature
techniques. They reported that the pore network closeness for porosity and related property prediction using ML methods.
centrality, a metric for describing the centrality of the shortest We utilized actual rock images instead of synthetic images.
paths between pores, was the most crucial network feature to While synthetic images have advantages,37−39 we believe that
predict the permeability. Many researchers incorporate physical actual rock images provide a more realistic representation of
equations into ML to reduce the dependency on the data. This porous media structure and properties. We focused on the
method is usually referred to as physics-informed machine quantitative analysis of rock images, specifically measuring pore-
learning.25 Tian et al.26 suggested a combination of artificial related properties and their distribution. While previous studies
neural networks (ANNs) and genetic algorithms to predict have primarily used imaging techniques such as thin section, CT
permeability. Their study uses the genetic algorithm to tune the scan, and SEM images to qualitatively analyze rocks,40−42 our
ANN architecture, whereas the ANN is applied to learn the approach provides a more comprehensive and quantitative
nonlinear relationships. They reported that tortuosity, number understanding of the rock properties. We further developed a
of pores, and pore−throat ratio were inversely proportional to data-driven machine learning model to predict the absolute
permeability. There are other studies that focus on the permeability of a new rock sample by using the properties
prediction of flow properties,27,28 but studies for the prediction extracted from the pore network of known rock samples. This
of pore networks and porosity remain limited. study confirms that a simple regression model can be developed
The convolutional neural network (CNN) is capable of to predict permeability by using accurate morphological
extracting task-related data from the images with high accuracy; descriptors of porous media. In addition, our study offers
hence, in many studies,29,30 a CNN-based surrogated model has practical techniques for image preprocessing and segmentation,
been conducted to predict rock permeability. However, the use which can be useful for other geoscientists working with similar
of CNN is not out of the scope of limitations. First, a CNN types of data. The remainder of the paper is organized as follows:
model requires a large number of training data for better Section 2 provides a brief description of the materials (rock
30206 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 1. Tomogram layer of (A) Bentheimer sandstone, (B) Castlegate sandstone, and (C) Leopard sandstone.

samples) used in this study, and Section 3 presents the methods images, extracting the labels from binary and grayscale images,
for image segmentation, label extraction, and machine learning preprocessing of the label data, creating a CNN model, and
model architectures. Section 4 is dedicated to the results and finally testing, training, and validation of the CNN model.
discussion of this study. Finally, Section 5 provides a summary Figure 3 shows the workflow for developing a regression
and significant conclusions from this work. model to predict the permeability. The workflow involves
subdividing the rock samples, segmenting the images, extracting
2. MATERIALS input and target parameters, and finally training, testing, and
This study uses three digital rock samples: Bentheimer, evaluating the accuracy of the regression model. Additional
Castlegate, and Leopard. The Bentheimer sandstone is selected details on each step and the relevant morphological properties
as it is considered to be the model sandstone for reservoir are provided in Section 3.2.2.
studies,43 and the other two rocks (Castlegate and Leopard) are 3.1. Image Acquisition and Pre-processing of Data.
chosen as they have similar fractions of the primary minerals as Data pre-processing is a crucial machine learning step as it
Bentheimer sandstone: quartz, feldspar, and kaolinite. The significantly affects model accuracy. As a part of data pre-
digital versions of these sandstones contain micro-CT images in processing, we applied data cleaning, denoising, and augmenta-
the netCDF format and are collected from the “Digital Rock tion to the input data set and data transformation to the target
Portal”.44 The Bentheimer sandstone has a lab-measured data set. Data cleaning is a process of detecting incorrect and
average porosity of 23.5% and a permeability of 1.30 noisy data and correcting or removing it from the data set.
Darcy.43,45 The porosity of the Castlegate and Leopard Usually, the micro-CT images have noise and some artifacts,46
sandstone is 26.0% and 22.8%, respectively, and the correspond- which can lead to incorrect label extraction. After extracting the
ing permeability is 1.05 and 0.08 Darcy, respectively.45 Further 2D images, the 3D tomographic data, and subdividing them into
details on the other properties of the selected rock samples are the regions of interest (ROIs) measuring 200 × 200 pixels, it was
given in Shikhov et al.45 Figure 1 displays the Bentheimer, observed that some images have higher noise labels (Figure 4).
Leopard, and Castlegate sandstone subsamples, and Table 1 Hence, we manually checked the input images before any further
provides the specifics of the tomograms. steps and deleted the images which had higher noise
interference. We applied a Gaussian blur filter to reduce the
Table 1. Size and Resolution of the Tomograms of signal-to-noise ratio of the rest of the images, as the Gaussian
Bentheimer, Castlegate, And Leopard Sandstone blur filter effectively smooths out noise while preserving crucial
features of the image.47
Sandstone name Size (voxels) Resolution (μm) As the voxel sizes of the individual rock samples are different
Bentheimer 800 × 800 × 400 2.15 (Table 1), dividing the 3D tomograms into specific ROIs yields a
Leopard 1240 × 1240 × 180 2.15 variable number of images. For instance, the Bentheimer
Castlegate 800 × 800 × 400 2.28 sandstone sample has 6400 ROI with 16 images per 400 layers.
In order to improve the predictive ability of the trained model,
we increased the number of training data by augmentation. The
3. METHODS data augmentation is implemented in this work by rotating the
This section discusses the methods used to prepare the raw images at different angles (90°, 180°, and 270°). After cleaning
grayscale images and extract the morphological properties (Ø, the data set and data augmentation, 24,880 images were
TA, and PSA) and the absolute permeability. In addition, the generated, from which 17,416 images were used to train the
machine learning models with different training and testing model, 3,732 images for testing, and 3,732 for validation. The
functions for predicting the morphological properties and label data was then checked for skewness factor, as it can
permeability are also briefly described. Figure 2 shows the significantly impact the model’s ability to predict outcomes
general workflow for predicting the morphological properties accurately. A skewed data set has a significant deviation between
from the digital rock image, which involves image acquisition the value of the mean and median. So, the tail region of the
from digital rock, preprocessing of raw grayscale images skewed data set may act as an outlier, and outliers negatively
(augmentation, data cleaning, denoising), segmentation of the affect the model’s performance.48 In this work, the values of the
30207 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 2. Workflow for extracting and predicting physical properties of porous media using CNN: (A) Image acquisition, (B) Raw image pre-
processing (augmentation, data cleaning, denoising), (C) Image segmentation and label data extraction, (D) Label data pre-processing, (E) Training
and testing of the CNN model, and (F) and Model evaluation.

Figure 3. Workflow for developing the regression model to predict permeability.

throat area were positively skewed, and log transformation was 3.2. Image Segmentation and Target Data Extraction.
applied to the target data set to create normally distributed 3.2.1. Image Segmentation. The ground truth or the labels for
training data. training the ML model were obtained from the binary images
30208 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

and extracted pore network, which require image segmentation.

First, we manually segmented 50 images of each rock sample to
create the ground truth using the interactive learning and
segmentation toolkit, Ilastik.49 After that, the segmented results
from the traditional thresholding method and a machine
learning model were compared with the manually segmented
images to evaluate each segmentation method.
The Otsu thresholding50 method was used as a traditional
thresholding tool to segment the individual rock ROIs. The
Otsu threshold separates the pixels into the foreground and
background. It involves iterating through all possible threshold
values and calculating the spread of the foreground and
background pixels.50 The algorithm aims to identify the
threshold value at which the sum of foreground and background
pixels is the least. “Scikit-image”,51 an image-processing library
in Python, was used to implement the Otsu threshold to raw
grayscale images to create binary images.
A U-Net model was developed as a machine learning model to
segment the individual ROIs of the rock samples. The model
architecture is illustrated in Figure 5. Due to the encoder−
decoder architecture, the U-Net model can capture both high-
level features and fine details of the image. The skip connection
Figure 4. Defective ROIs of Leopard sandstone. option facilitates U-Net to preserve the spatial information of
the image. In total, 14,000, 500, and 50 images were used for
training, validating, and testing the model, respectively.

Figure 5. U-Net model architecture for image segmentation.

30209 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

The intersection over union (IoU) was calculated by suggested by Gostick.52 Figure 6 summarizes the steps for data
comparing the segmented images from the Otsu thresholding processing involved in pore network extraction.
method and U-Net model with the manually segmented images Next, the morphological descriptors for predicting the
(Table 2). The results show that both Otsu and U-Net’s permeability were extracted from the pore network model.
The pore network model represents the network topology of the
Table 2. Image Segmentation Results from the U-Net Model porous media in terms of their interconnected pores and throats.
and the Otsu Thresholding Method Table 3 lists the morphological descriptors extracted from the
rock sample to create a data-driven model to predict absolute
Sandstone type IoU value for U-Net IoU value for Otsu thresholding
permeability.
Bentheimer 0.970 0.954
Leopard 0.805 0.780
Table 3. Morphological Properties Extracted for the
Castlegate 0.964 0.949
Quantitative Characterization of Porous Microstructures

performance is satisfactory in segmenting the digital rock Index Morphological descriptors Denotation
images. Though U-Net’s deep learning approach offers more 1 Porosity Ø
versatility and capability for complex image segmentation 2 Pore volume vp
challenges, this study used the Otsu thresholding method to 3 Pore surface area Sp
segment the individual ROIs of different rock samples because of 4 Maximum pore size dmp
simplicity, fast processing, and interpretability. However, it 5 Pore region volume vpr
should be noted that for complex images, ML methods trained 6 Pore size distribution pd
on a large number of images perform better than traditional 7 Pore diameter dp
thresholding methods. 8 Maximum throat size dmt
3.2.2. Morphological Property Extraction. To measure the 9 Pore equivalent diameter deqp
porosity, first, the pixels of the binary images were separated 10 Throat diameter dt
according to their intensity. For example, the white pixels have 11 Throat length lt
an intensity closer to 255 and represent the pore area. Then, the 12 Throat volume vt
total number of pore pixels was added up, and finally, the 13 Throat size distribution td
porosity was calculated by dividing it by the area of the image. 14 Pore inscribed diameter dip
The average porosity calculated from this workflow matches the 15 Pore extended diameter dep
porosity obtained from the laboratory measurements reported in 16 Throat inscribed diameter dit
the literature.43 17 Throat total length lTt
The pore network was extracted from the segmented images 18 Throat perimeter lt
using the method proposed by Gostick.52 This technique uses 19 Throat cross-sectional area Sct
the watershed algorithm to create a distance map of the 20 Throat equivalent diameter deqt
segmented image to locate the pores and throats in the 21 Throat spacing mt
tomograms. After the peaks from the watershed were identified 22 Effective diffusivity Deff
using the maximum filter, the spurious peaks on saddles and 23 Tortuosity τ
plateaus of the distance map are eliminated. Following that, the
algorithm automatically merges the nearby peaks and considers 3.2.3. Permeability Extraction. The 3D tomogram of the
them as a single large pore. Finally, in order to achieve the sandstone was first subdivided into small ROI (100 × 100 ×
segmented pore areas and extracted pore networks, a marker- 100) to ease the computational resource requirements prior to
based watershed algorithm was used. The value of sigma in the permeability calculation. An increment of 100 was applied in
Gaussian filter was set to 0.4, and the radius of the spherical each direction to increase the data set. Then, the images were
structuring element was set to 4, which is similar to the settings segmented using the Otsu thresholding to distinguish and label

Figure 6. (A) Tomogram of a raw grayscale image for Bentheimer sandstone, (B) ROI (200 × 200) of the tomogram, and (C) segmented image.

30210 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 7. Workflow for extracting permeability from a digital rock image.

solid grains and pore space. After that, based on the defined detectors trained by the filters can be reused several times
geometry and grid resolution, a computational mesh was throughout the input image, which makes CNN suited for
generated using Palabos, and absolute permeability for each computer vision tasks. In this study, the inputs for the CNN
block was measured. Palabos,53 an open-source software library, model were the ROIs obtained from the 2D images, which were
was used to implement the Lattice Boltzmann Method (LBM) generated from the 3D tomographic data. Figure 9 shows the
for simulating fluid flow. LBM is a popular numerical method for model architecture proposed in this work. In the following
simulating fluid dynamics in porous media. It is a numerical sections, we briefly discuss the rationale for choosing the major
approach that simulates fluid flow in porous media by functions used in this study (ReLU, Dropout, Max-pooling,
representing the fluid as a collection of particles. Each particle Adam optimizer, and Huber loss) and their working principle.
is associated with a particle velocity distribution function located The convolutional layer convolves over the input image using
at each discrete lattice node. At regular time intervals, the filters to extract the key features such as pore and throat spaces,
particles collide with one another, and the properties associated which are important for analyzing the porosity, pore surface area,
with the lattice nodes are updated accordingly. These collisions and throat area of the sandstone. Figure 10 shows the sequential
are governed by rules designed to ensure that the time-averaged feature extraction at each activation layer from a single ROI
motion of the particles aligns with the Navier−Stokes using the proposed CNN model.
equations,54 which describe the motion of fluids in a continuum. The convolution process is just a simple mathematical
We used the D3Q19 scheme, representing three-dimensional operation of two matrices. For example, consider that an input
fluid motion and utilizing 19 associated velocity vectors. Palabos image has a matrix of (7 × 7); the filter has a matrix of (3 × 3);
calculates the permeability of the porous medium using Darcy’s the stride is 2; and the padding is the same. Then, a sum-product
law.55 The workflow for extracting the permeability is illustrated operation between the individual elements of the two matrices
in Figure 7. was done to generate the elements of the feature map with a
3.3. Machine Learning Models. 3.3.1. Convolutional matrix of (3 × 3). Equation 1 controls the output dimension of
Neural Network for the Morphological Property Prediction. the convolutional process.
The convolutional neural network (CNN) architecture uses n + 2p f n + 2p f
multiple layers of nonlinear transformations, which are trained +1 × +1
iteratively to make predictions close to the expected ground S S (1)
truth targets. The CNN consists of convolution, pooling, and where n × n is the dimension of the matrix; p is the padding; f is
fully connected layers (Figure 8). the filter size; and S is the stride.
The function of the first two layers (convolution and pooling) The activation function introduces nonlinearity into the
is to extract the features, and the function of the third layer (fully output of a neural network, allowing it to model more complex
connected layer) is to map the extracted features into the final relationships between inputs and outputs. There are several
output. The convolution layers preserve the spatial relationship activation functions to construct the CNN, such as Sigmoid,
between inputs and feature maps, whereas fully connected layers Tanh, Softplus, and the Rectified Linear Unit (ReLU). In this
do not preserve the spatial structure of their inputs. In CNN, work, ReLU was used as an activation function as it is simple to
calculate, has a higher convergence speed, is cost-effective, and
prevents the emergence of vanishing gradient problems, which is
common for Sigmoid and Tanh functions.56
Pooling and subsampling were used to reduce the spatial size,
hence reducing the number of features and computational
complexity of the network. Three standard functions can be used
in the pooling operation: average pooling, maximum pooling,
and minimum pooling. The max-pooling function with a kernel
size of (2 × 2) and stride of 2 pixels was applied for the images
used in this work. The purpose of the max-pooling function is to
downsample the input representation by picking up the highest
value from the region of the feature map covered by the filter,
which tends to retain the most prominent features and edges
present in the image. In contrast, average pooling and min
Figure 8. Structure of the convolutional neural network. pooling can potentially blur important image details and may
30211 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 9. Proposed CNN model architecture.

Figure 10. Key feature extraction using the proposed CNN model.

not be as effective at capturing the essential characteristics of the

image. Figure 11 illustrates the maximum-pooling operation in
CNNs.
The “dropout” method was applied to the fully connected
layer to avoid overfitting the training data. After implementing
dropout, the network becomes thinner than the standard neural
network, thus enhancing the resistance of overfitting and making
training faster.57 In this work, 20% of the neurons in the fully
connected layers were dropped out.
The model was trained to minimize errors between its Figure 11. Max-pooling operation in CNNs.
predictions and expected ground truth labels. In this work,
“Huber loss”58 was used as a loss function to compute the error best of these loss functions: robustness and stability. Equation 2
(eq 2). Huber loss is the hybrid of Mean Absolute Error (MAE) was used to measure the distance and backpropagate the error
and Mean Squared Error (MSE) loss functions, and it takes the through the network to compute gradients. The hyperparameter
30212 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Table 4. Comparison between Different Evaluation Functions for Model A, Model B, and Model C
Porosity Pore surface area Throat area
Evaluation functions Binary image Grayscale image Binary image Grayscale image Binary image Grayscale image
Model A (Trained with Bentheimer Images)
MAPE (%) 1.01 1.43 3.2 4.21 1.97 2.55
MSPE (%) 0.02 0.04 0.18 0.29 0.07 0.11
RMSPE (%) 1.24 1.95 4.29 5.37 2.7 3.34
Model B (Trained with Castlegate Images)
MAPE (%) 0.84 1.51 3.44 3.93 7.31 2.91
MSPE (%) 0.0136 0.201 0.1589 0.3187 0.867 0.139
RMSPE (%) 1.169 4.49 3.986 5.645 9.311 3.73
Model C (Trained with Bentheimer and Castlegate Images)
MAPE (%) 3.9 4.13 2.26 3.29 1.17 1.28
MSPE (%) 0.25 0.28 0.08 0.19 0.02 0.03
RMSPE (%) 5.03 5.31 2.93 4.41 1.4 1.66

Figure 12. Prediction of porosity against actual labels from (left) binary and (right) grayscale images of Leopard sandstone using Model C.

δ was set to 0.5 in this study. The “Adam”59 optimizer was used 3.3.2. Regression Models for Permeability Prediction. The
in this experiment to update the network parameters (weights morphological descriptors and absolute permeability of each of
and biases). The Adam optimizer updates the learning rate for the subdivided ROIs (200 × 200 × 200) of Bentheimer
each network weight individually, and it takes less time to sandstone were used to construct proxy models using four
compute and requires fewer parameters for tuning. different machine learning algorithms: random forest (RF),
l
o 1 multilayer perceptron (MLP), gradient boosting (GBoost), and
o
o (y yi )2 for |yi yi |
o
o 2 i extreme gradient boosting (XGBoost). The best algorithm was
o
Huber Loss = o
m
o then selected to develop a combined model for predicting an
o
o ij 1 yz
o
o jj|yi yi | zz for |yi yi | > unknown rock’s permeability (Leopard sandstone in this study).
o
o
n k 2 { (2) After that, three models were trained with two different rock
where ŷi is the true value; yi is the predicted value; and δ is the samples: the first with the morphological properties, of
hyperparameter. Bentheimer, the second with only Castlegate properties, and
The data set was too large to be processed in a single batch, so the third with combined Bentheimer and Castlegate properties.
it was divided into smaller batches (240 images each) to train the We used 80% of the data set to train the model while allocating
model. We set optimal values for the learning rate through the remaining 20% to validate the model’s performance. Log
manual adjustment to optimize the individual model accuracy transformation was applied to reduce the skewness of the data
and computational speed. The learning rate was set to 1 × 10−6
while training the model for predicting porosity and pore surface set, and GridSearchCV was used to find the best hyper-
area, and for throat area, it was set to 1 × 10−4. The parameters for the model by optimizing for the minimum mean
“TensorFlow”60 platform was used to train the network in this squared error. Metrics such as mean squared error (MSE), mean
work. Google Colaboratory was used to run the code, and each absolute error (MAE), root mean squared error (RMSE), and R-
step was processed in around 26 s using a Tesla T4 GPU. square score were computed to evaluate the model.
30213 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 13. Prediction of pore surface area (PSA) against actual labels from (left) binary and (right) grayscale images of Leopard sandstone using
Model C.

4. RESULTS AND DISCUSSION similar to those of the Leopard sandstone, which is further
This section presents training, validation, and testing results for discussed in Section 4.4.
predicting the morphological properties and permeability of a 4.2. Pore Surface Area (PSA). Model C was trained for 300
rock sample using CNN and regression models, respectively. epochs to predict the pore surface area (PSA) of Leopard
The CNN estimation accuracy increases with every elapsed sandstone (Figure 22). The R2 score for predicting PSA using
epoch. Various metrics have been used to evaluate the model Model C is 0.977 and 0.944 for the binary and grayscale images,
such as mean absolute percentage error (MAPE), mean squared respectively (Figure 13). The MAPE for predicting the PSA of
percentage error (MSPE), and root mean squared percentage Leopard sandstone utilizing Model C is 3.04% for binary images
error (RMSPE). Table 4 shows the evaluation metrics calculated and 8.35% for grayscale images. The MAPE to predict the PSA
for Model A, Model B, and Model C. Model A was trained, of Leopard sandstone using Model A is much higher than those
validated, and tested using the Bentheimer sandstone images, using Model B and Model C, 6.52% and 9.56% for binary and
Model B with Castlegate images, and Model C with Bentheimer grayscale images, respectively. For Model B, the error (MAPE)
and Castlegate images combined. is 3.6%, and 9.45% is the value for binary and grayscale images,
Next, we used these trained models (Model A, Model B, and respectively.
Model C) to predict the morphological descriptors of an 4.3. Throat Area (TA). The pore throat area computed from
unknown rock (Leopard sandstone) that the models do not see. the Gostick algorithm (Section 3.2.2) was positively skewed,
4.1. Porosity (Ø). Figure 12 shows the porosity predictions with a significant number of calculated values larger than 280
with binary and grayscale images against the test data set μm. The mean of the throat area was 133.9 μm, with a standard
generated from Leopard sandstone using Model C. The model deviation of 56.1 μm. The skewness in training data can
was trained for 200 epochs, after which the average error loss negatively impact the predictive ability of an ML model. Hence,
stays constant (Figure 21). The performance of Model C in we applied the log transformation method to reduce the
predicting porosity from binary and grayscale images demon- skewness, after which the mean of the throat area became 4.82
strates a significant improvement compared to the results μm, with a standard deviation of 0.43. The log transformation is
reported in other literature.61 This improvement can be a mathematical operation that reduces skewness and compresses
attributed to several factors, such as image preprocessing, data large values while expanding smaller ones. It involves taking the
cleaning, and the CNN model architecture used in this study. logarithm of the original values, which results in a new set of
The R-squared (R2) metric in Model C achieved 0.992 for values that are proportional to the original values but with a
predicting porosity with binary images and 0.913 for predicting more manageable range of variations. Mathematically, the log
porosity with grayscale images. Furthermore, when employing transformation can be presented as eq 3.
Model C, the MAPE for porosity prediction of Leopard y = log(x) (3)
sandstone stands at 2.19% with binary images and at 6.25%
for grayscale images, which marks a notable improvement where x is the original value, and y is the transformed value.
compared to the porosity prediction from Model A and Model Figure 14 illustrates the distribution and probability plot
B. Model A resulted in the errors (MAPE) of 16.19% for binary changes of the TA after the log transformation. The red line in
images and 19.91% for grayscale images, whereas Model B the normal probability plot represents the expected values of a
showed errors of 2.32% with binary images and 6.59% with standard normal distribution, which is used to compare the
grayscale images. This result shows that the pore network and distribution of the transformed variable against a normal
other morphological characteristics of the Castlegate rock are distribution. Although the skewness of the training data
30214 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 14. Distribution and probability plot of the throat area (A, B) before log transformation and (C, D) after log transformation.

decreases after transformation, the data becomes slightly 4.4. Permeability Prediction. The permeability prediction
negatively skewed. from digital rock images using CNN, which we previously used
Figure 15 presents the prediction of TA from binary and for the porosity prediction, is significantly complicated and
grayscale images of Leopard sandstone with Model C after 300 resource-intensive. The permeability prediction based on CNN
epochs (Figure 23). The R2 values for TA prediction using requires complex architecture and significant computational
binary and grayscale images of Leopard sandstone are 0.776 and resources, which may not be tenable if a large number of samples
0.744, respectively. Model B results in a MAPE of 6.08% for are to be analyzed. Hence, we used a data-driven method that is
binary images and 6.37% for grayscale images when applied to significantly easier and more efficient to implement than image-
the Leopard sandstone. However, Model A results in higher based approaches.
MAPE values: 7.34% for binary images and 7.4% for grayscale The morphological properties extracted from Bentheimer
images. Also, with Model B, the MAPE is higher than that of sandstone were trained with four different algorithms (XGboost,
Model C, with 6.74% and 10.04% for grayscale and binary GBoost, RF, and MLP). Eighty percent of the data set of
images, respectively. Bentheimer properties was used for training and 20 percent for
Overall, it is observed, in terms of predicting Leopard testing. Among other models, the XGboost demonstrated the
sandstone properties, that Model B outperforms Model A, best accuracy for predicting the permeability of the Bentheimer
indicating the similarity between Castlegate and Leopard sandstone, achieving an R2 score of 0.934. Figure 16 shows the
morphological properties. However, overall, the combined permeability prediction of Bentheimer sandstone blocks against
model (Model C) trained with both Bentheimer and Castlegate the LBM simulation results using the XGBoost model. The
sandstone exhibits higher accuracy in predicting Leopard XGboost provides the best accuracy among the tested models as
sandstone’s morphological properties, indicating the importance it can handle complex relationships between features in the data.
of diverse training data for robust model performance. It also includes L1 and L2 regularization terms, which help
30215 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 15. Prediction of throat area against actual labels from (left) binary and (right) grayscale images of Leopard sandstone using Model C.

Table 5. Comparison between Different Evaluation

Functions for the Models Used for Predicting Permeability
Evaluation functions XGBoost GBoost RF MLP
MAE 0.087 0.117 0.158 0.217
MSE 0.013 0.018 0.051 0.117
RMSE 0.114 0.136 0.227 0.342
R2 score 0.933 0.906 0.738 0.407

However, when the models were tested with Leopard sandstone,

the model trained with the combined properties of Bentheimer
and Castlegate sandstone provided the most accurate results.
The MAE, RMSE, and R2 for permeability prediction of Leopard
sandstone using all the models are summarized in Table 6.

Table 6. R2, MAE, and RMSE Errors for Predicting the

Permeability of Leopard Sandstone Using Models Trained on
Three Distinct Datasets
Sandstone Properties for model training MAE RSME R2
Bentheimer 0.112 0.146 0.466
Castlegate 0.069 0.105 0.710
Figure 16. Comparison between targets and predictions for predicting Bentheimer and Castlegate 0.049 0.084 0.813
Bentheimer sandstone permeability with the XGBoost model.
Across all of the metrics, the combined model delivers
prevent overfitting and make the model more robust.62 significantly improved rock property prediction. The improve-
However, in this study, the MLP model becomes less accurate ment is more apparent when comparing the combined model to
for predicting the permeability. Table 5 shows the values of the the model trained with just the Bentheimer sandstone. However,
evaluation matrices for different models (XGBoost, GBoost, RF, we observe improvement in prediction even for the combined
and MLP). model compared with the model trained with the Castlegate
Next, three different models were trained using Bentheimer sandstone.
and Castlegate rock samples with XGboost to predict the The improvement in accuracy for the combined model was
permeability of an unknown rock sample (Leopard). The first also observed in predicting the morphological properties
two models were trained on the morphological properties of (Sections 4.1−4.3). Figure 17 illustrates the accuracy in the
individual rock samples, and the third model was trained on the prediction of Leopard sandstone permeability with models
properties of both Bentheimer and Castlegate sandstones. trained with different data sets.
Each model resulted in an acceptable accuracy for Feature importance analysis was conducted to understand the
permeability prediction when tested with the same rock sample. critical parameters influencing the permeability prediction of the
30216 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 17. Prediction of permeability of Leopard sandstone with models trained with (A) Bentheimer sandstone properties, (B) Castlegate sandstone
properties, and (C) Bentheimer and Castlegate sandstone properties.

Figure 18. Feature importance plot for permeability prediction, based on the XGBoost model trained with the combined properties of Bentheimer and
Castlegate sandstone.

rock samples. Previous studies reported that morphological importance plot for permeability prediction using the data-
properties like porosity, tortuosity, and pore-to-throat ratio driven model with morphological properties.
correlate with the porous media’s permeability.63−65 Our In order to understand the reason for the improved prediction
findings underscore the paramount role of tortuosity, which of both morphological properties and permeability using the
emerged as the most significant factor with an importance individually trained model trained with Castlegate sandstone, we
percentage of ∼28%, indicating the complexity of fluid paths analyzed the underlying pore-scale properties of each rock,
within the rock as a crucial determinant of permeability. which showed that the statistical distributions of features in the
Additionally, porosity and pore region volume were identified as Castlegate sandstone are closer to those of Leopard sandstone.
important features with importance percentages of ∼9%, The key features such as pore region volume, porosity, pore/
highlighting the significance of void spaces within the rock for throat ratio, effective diffusivity, and tortuosity in Castlegate
fluid storage and transmissibility. The analysis also revealed the sandstone share mean values and distributions that align more
importance of effective diffusivity in influencing permeability closely with those in Leopard sandstone, which are critical in
(∼8%), pointing toward the relevance of fluid movement ease defining permeability. Figure 19 shows the box plot for various
between connecting pores. Figure 18 shows the feature features used for training the machine learning model, and the
30217 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 19. Box plot showing the key features for Castlegate, Bentheimer, and Leopard sandstones. The morphological properties of Castlegate
sandstone are significantly similar to those of Leopard sandstone.

Figure 20. MAPE for predicting properties (Ø, PSA, and TA) of Leopard sandstone by using (A) grayscale images and (B) binary images.

descriptive statistics of these features are summarized in only used the images of Bentheimer sandstone (Model A);
Appendix, Table 7. another one was with only Castlegate sandstone images (Model
B); and the last one was with both Bentheimer and Castlegate
5. SUMMARY AND CONCLUSIONS sandstone images (Model C). The training of Model C required
This study proposed a workflow to predict the morphological an average duration of 2.5 h, utilizing the Tesla T4 GPU offered
properties of an unknown rock sample using micro-CT images, by Google Colaboratory. The number of epochs to train the
which can later be used to predict the permeability of the porous models was selected based on the training and validation loss.
media. Various flow properties, including porosity and Training and validation losses decreased with each additional
permeability, can be extracted by using the proposed workflow epoch. Figure 20 summarizes the MAPE for predicting various
without extensive laboratory experiments. These flow properties morphological properties for an unknown rock sample.
are crucial in modeling the flow in porous media. This work used In the case of Model C while using grayscale images to predict
Leopard sandstone as an unknown rock sample to predict the properties of Leopard sandstone, the MAPE is 6.25% for
morphological properties (Ø, TA, and PSA). Three distinct porosity, 8.35% for pore surface area, and 6.37% for throat area
models were trained using two distinct rock samples: one model (Figure 20). On the other hand, while using binary images, the
30218 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 21. Average error loss during training and validation of Model C to predict porosity.

MAPE is lower: 2.19%, 3.04%, and 6.08% for predicting We also demonstrated a data-driven workflow for predicting
porosity, pore surface area, and throat area, respectively (Figure the permeability of porous media using various morphological
20). This is because the model generated with grayscale images data extracted from the rock sample. The XGboost model
is more sensitive to noise and defects, resulting in a higher resulted in the most accurate prediction of permeability for an
prediction error than the model developed with binary images.66 unknown rock sample. The model trained with Castlegate
Accurate rock image segmentation is essential for extracting the sandstone outperformed the model trained with the Bentheimer
precise properties of the rock sample. This study applied two sandstone data set in terms of predicting both the morphological
different segmentation methods, the U-Net model and the Otsu property and permeability of the Leopard sandstone, which
thresholding method, to extract the properties. Then, the indicates the similarity between Castlegate and Leopard
segmented images were evaluated against the manually sandstone. The feature importance analysis revealed that the
segmented images. The results show that with proper data statistical distributions of key features in the Castlegate
preprocessing Otsu thresholding can provide satisfactory sandstone are closer to those of Leopard sandstone (Figure
segmentation results and improve model accuracy. We found 19). For predicting the permeability of Leopard sandstone, the
that deleting the defective images and reducing skewness models trained using Bentheimer and Castlegate sandstone
properties yielded R2 values of 0.466 and 0.710, respectively
reduced the MAPE of property prediction by around 3%.
(Figure 17). However, the combined model (trained with both
Furthermore, this study also shows that it is possible to create an
Bentheimer and Castlegate sandstone) was more accurate in
enhanced model trained from a large variety of rock samples to predicting the properties of Leopard sandstone, indicating the
predict the accurate morphological descriptors of an unknown importance of diverse training data for robust model perform-
rock sample. The workflow can also be applied to predict other ance. The R2 value for predicting the permeability of Leopard
morphological properties using high-resolution micro-CT sandstone was 0.813 by using the model that is trained with both
images. To validate the workflow presented in this study, the Bentheimer and Castlegate sandstone properties.
average pore size was calculated from the extracted network of The model presented in this work for predicting the
the rock images and trained with the proposed CNN model morphological descriptors of porous media still has room for
(Model C). Previous studies have indicated that model accuracy improvement. Methods like the generative adversarial network
tends to be lower when predicting properties extracted from the (GAN)66 and hybrid stochastic deep-learning (HSDL)67 can be
pore network.61 However, the results show that the proposed applied to increase the image resolution and thus increase the
model (Model C) provides good accuracy in predicting the prediction accuracy. In addition, the hyperparameters of the
average pore size of an unknown rock sample (Leopard CNN model can also be tuned for better model performance.
sandstone), with an R2 value of 0.91 and 0.94 for grayscale For the models in this study, the decrease in the learning rate and
and binary images, respectively. increase in the number of epochs improved the property
30219 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 22. Average error loss during training and validation of Model C to predict pore surface area.

Figure 23. Average error loss during training and validation of Model C to predict throat area.

prediction with grayscale images. The MAPE while training from 24.60% with a learning rate of 1 × 10−3 to 1.43% when the
Model A with grayscale images for porosity prediction decreased learning rate was 1 × 10−6. Future studies will focus on
30220 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

Table 7. Descriptive Statistics of Key Features of Castlegate, Bentheimer, and Leopard Sandstone Subdivided Blocks for Machine
Learning Model Training
Bentheimer Sandstone
Tortuosity Effective Diffusivity Pore/Throat Porosity Pore Region Volume
min 7.51 × 10−03 2.11 × 10−09 2.02 × 1014 0.10 6.32 × 10−12
−08
25% 3.21 × 1000 4.36 × 10 3.24 × 1014 0.19 7.35 × 10−12
−08
Median 1.10 × 1001 8.07 × 10 4.11 × 1014 0.23 7.64 × 10−12
75% 2.91 × 1001 1.31 × 10−07 4.61 × 1014 0.26 8.03 × 10−12
−08
Mean 3.60 × 1001 9.55 × 10 3.99 × 1014 0.23 7.63 × 10−12
−07
max 6.70 × 1001 2.62 × 10 6.66 × 1014 0.36 8.93 × 10−12
Castlegate Sandstone
Tortuosity Effective Diffusivity Pore/Throat Porosity Pore Region Volume
min 8.31 × 10−03 1.11 × 10−09 1.85 × 1014 0.12 6.75 × 10−12
25% 2.50 × 1000 4.29 × 10−08 2.45 × 1014 0.20 7.34 × 10−12
Median 5.97 × 1000 6.44 × 10−08 2.81 × 1014 0.23 7.63 × 10−12
75% 1.84 × 1001 9.76 × 10−08 3.18 × 1014 0.26 7.93 × 10−12
Mean 2.21 × 1001 7.54 × 10−08 2.80 × 1014 0.23 7.60 × 10−12
max 4.01 × 1001 1.80 × 10−07 4.00 × 1014 0.32 8.60 × 10−12
Leopard Sandstone
Tortuosity Effective Diffusivity Pore/Throat Porosity Pore Region Volume
min 1.58 × 10−02 4.02 × 10−09 2.03 × 1014 0.11 5.95 × 10−12
25% 1.13 × 1000 2.95 × 10−08 2.64 × 1014 0.16 7.33 × 10−12
Median 3.77 × 1000 5.17 × 10−08 2.84 × 1014 0.20 7.87 × 10−12
75% 1.35 × 1001 8.44 × 10−08 3.15 × 1014 0.26 8.28 × 10−12
Mean 1.39 × 1001 6.22 × 10−08 2.91 × 1014 0.22 7.70 × 10−12
max 3.10 × 1001 1.64 × 10−07 3.77 × 1014 0.40 8.77 × 10−12

improving the predictive power of the proposed model using Society Petroleum Research Fund (PRF # 62679-UNI9) and
other advanced algorithms and tuning different hyperpara-
meters. partly by the National Science Foundation Award under CBET-
2245484. Any opinions, findings, conclusions, or recommenda-
6. APPENDIX
The average error loss during the training and validation for tions expressed in this material are those of the author(s) and do
predicting porosity, pore surface area, and throat area using not necessarily reflect the views of the funding agencies.
Model C is presented in Figures 21, 22, and 23.

■ AUTHOR INFORMATION
Corresponding Author
■ TABLE OF ABBREVIATIONS
CNN, Convolutional neural network
Aaditya Khanal − The Jasper Department of Chemical 2D, Two dimensional
Engineering, The University of Texas at Tyler, Tyler, Texas 3D, Three dimensional
75799, United States; Russell School of Chemical Engineering Micro-CT, Microcomputed tomography
and McDougall School of Petroleum Engineering, The RCA, Routine core analysis
University of Tulsa, Tulsa, Oklahoma 74104, United States; SCAL, Special core analysis
orcid.org/0000-0003-1333-7457; ML, Machine learning
Email: [email protected], aadityakhanal@ ANN, Artificial neural network
gmail.com Ø, Porosity
PSA, Pore surface area
Author
TA, Throat area
Md Irfan Khan − The Jasper Department of Chemical
ROI, Region of interest
Engineering, The University of Texas at Tyler, Tyler, Texas
ReLU, Rectified linear unit
75799, United States; orcid.org/0000-0001-9790-7386
MAE, Mean absolute error
Complete contact information is available at: MSE, Mean squared error
https://pubs.acs.org/10.1021/acsomega.3c10131 MAPE, Mean absolute percentage error
MSPE, Mean squared percentage error
Notes RMSPE, Root mean squared percentage error
The authors declare no competing financial interest. GAN, Generative adversarial network

■ ACKNOWLEDGMENTS
We thank the anonymous reviewers for their edits and
HSDL, Hybrid stochastic deep learning
XGBoost, Extreme gradient boost
GBoost, Gradient boost
suggestions, which improved our manuscript. This material is RF, Random forest
based upon work supported partly by the American Chemical MLP, Multilayer perceptron
30221 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

■ REFERENCES
(1) Khanal, A.; Weijermars, R. Comparison of Flow Solutions for
unconventional shale reservoirs. J. Nat. Gas Sci. Eng. 2022, 105,
No. 104720.
(22) Ebadi, M.; Armstrong, R. T.; Mostaghimi, P.; Wang, Y. D.;
Naturally Fractured Reservoirs Using Complex Analysis Methods
Alqahtani, N.; Amirian, T.; James, L. A.; Parmar, A.; Zahra, D.; Hamze,
(CAM) and Embedded Discrete Fracture Models (EDFM):
H.; Koroteev, D. Predictive Soft Computing Methods for Building
Fundamental Design Differences and Improved Scaling Method.
Digital Rock Models Verified by Positron Emission Tomography
Geofluids. 2020, 2020, 1−20.
(2) Khanal, A.; Irfan Khan, M.; Fahim Shahriar, M. Comprehensive Experiments. Water Resour. Res. 2022, DOI: 10.1029/2021WR031814.
parametric study of CO2 sequestration in deep saline aquifers. Chem. (23) Tembely, M.; AlSumaiti, A. Deep Learning for a Fast and
Eng. Sci. 2024, 287, No. 119734. Accurate Prediction of Complex Carbonate Rock Permeability From
(3) Khanal, A.; Shahriar, M. F. Physics-Based Proxy Modeling of CO2 3D Micro-CT Images. In SPE-197457-MS, 2019. DOI: 10.2118/
Sequestration in Deep Saline Aquifers. Energies. 2022, 15, 4350. 197457-MS.
(4) Bachu, S.; Bonijoly, D.; Bradshaw, J.; Burruss, R.; Holloway, S.; (24) van der Linden, J. H.; Narsilio, G. A.; Tordesillas, A. Machine
Christensen, N. P.; Mathiassen, O. M. CO2 storage capacity estimation: learning framework for analysis of transport through complex networks
Methodology and gaps. Int. J. Greenh. Gas Control. 2007, 1, 430−443. in porous, granular media: A focus on permeability. Phys. Rev. E 2016,
(5) S. Edition, Recommended practices for core analysis. API Recomm. 94, 022904.
Pract. 1998. (25) Tang, M.; Liu, Y.; Durlofsky, L. J. A deep-learning-based
(6) McPhee, C.; Reed, J.; Zubizarreta, I. Best Practice in Coring and surrogate model for data assimilation in dynamic subsurface flow
Core Analysis. Core Analysis - A Best Practice Guide 2015, 64, 1−15. problems. J. Comput. Phys. 2020, 413, No. 109456.
(7) Berg, C. F.; Lopez, O.; Berland, H. Industrial applications of digital (26) Tian, J.; Qi, C.; Sun, Y.; Yaseen, Z. M.; Pham, B. T. Permeability
rock technology. J. Pet. Sci. Eng. 2017, 157, 131−147. prediction of porous media using a combination of computational fluid
(8) Amabeoku, M. O.; BinNasser, R. H. Quality Control/Quality dynamics and hybrid machine learning methods. Eng. Comput. 2021,
Assurance Assessments of Core Analysis Data from Multiple 37, 3455.
Commercial Laboratories. All Days, SPE 2012, 605−616. (27) Tembely, M.; AlSumaiti, A. M.; Alameri, W. S. Machine and deep
(9) Alpak, F. O.; Berg, S.; Zacharoudiou, I. Prediction of fluid topology learning for estimating the permeability of complex carbonate rock
and relative permeability in imbibition in sandstone rock by direct from X-ray micro-computed tomography. Energy Reports. 2021, 7,
numerical simulation. Adv. Water Resour. 2018, 122, 49−59. 1460−1472.
(10) Iraji, S.; Soltanmohammadi, R.; Munoz, E. R.; Basso, M.; Vidal, (28) Tian, J.; Qi, C.; Sun, Y.; Yaseen, Z. M. Surrogate permeability
A. C. Core scale investigation of fluid flow in the heterogeneous porous modelling of low-permeable rocks using convolutional neural networks.
media based on X-ray computed tomography images: Upscaling and Comput. Methods Appl. Mech. Eng. 2020, 366, No. 113103.
history matching approaches. Geoenergy Sci. Eng. 2023, 225, (29) Kamrava, S.; Tahmasebi, P.; Sahimi, M. Linking Morphology of
No. 211716. Porous Media to Their Macroscopic Permeability by Deep Learning.
(11) Malki, M. L.; Saberi, M. R.; Kolawole, O.; Rasouli, V.; Sennaoui, Transp. Porous Media. 2020, 131, 427−448.
B.; Ozotta, O. Underlying mechanisms and controlling factors of (30) Hong, J.; Liu, J. Rapid estimation of permeability from digital
carbonate reservoir characterization from rock physics perspective: A rock using 3D convolutional neural network. Comput. Geosci. 2020, 24,
comprehensive review. Geoenergy Sci. Eng. 2023, 226, No. 211793. 1523−1539.
(12) Dong, H.; Sun, J.; Lin, Z.; Fang, H.; Li, Y.; Cui, L.; Yan, W. 3D (31) Takbiri, S.; Kazemi, M.; Takbiri-Borujeni, A.; McIlvain, J. A deep
pore-type digital rock modeling of natural gas hydrate for permafrost learning approach to predicting permeability of porous media. J. Pet. Sci.
and numerical simulation of electrical properties. J. Geophys. Eng. 2018, Eng. 2022, 211, No. 110069.
15, 275−285. (32) Wang, H.; Yin, Y.; Hui, X. Y.; Bai, J. Q.; Qu, Z. G. Prediction of
(13) Andrä, H.; Combaret, N.; Dvorkin, J.; Glatt, E.; Han, J.; Kabel, effective diffusivity of porous media using deep learning method based
M.; Keehm, Y.; Krzikalla, F.; Lee, M.; Madonna, C.; Marsh, M.; on sample structure information self-amplification. Energy AI. 2020, 2,
Mukerji, T.; Saenger, E. H.; Sain, R.; Saxena, N.; Ricker, S.; Wiegmann, No. 100035.
A.; Zhan, X. Digital rock physics benchmarks�Part I: Imaging and (33) Zhou, X.-H.; McClure, J. E.; Chen, C.; Xiao, H. Neural network−
segmentation. Comput. Geosci. 2013, 50, 25−32. based pore flow field prediction in porous media using super resolution.
(14) Niu, Y.; Mostaghimi, P.; Shabaninejad, M.; Swietojanski, P.;
Phys. Rev. Fluids. 2022, 7, No. 074302.
Armstrong, R.T. Digital Rock Segmentation for Petrophysical Analysis
(34) Saxena, N.; Hows, A.; Hofmann, R.; Alpak, F. O.; Freeman, J.;
With Reduced User Bias Using Convolutional Neural Networks. Water
Hunter, S.; Appel, M. Imaging and computational considerations for
Resour. Res. 2020, DOI: 10.1029/2019WR026597.
image computed permeability: Operating envelope of Digital Rock
(15) Karimpouli, S.; Tahmasebi, P. Segmentation of digital rock
images using deep convolutional autoencoder networks. Comput. Physics. Adv. Water Resour. 2018, 116, 127−144.
(35) Fu, J.; Wang, M.; Chen, B.; Wang, J.; Xiao, D.; Luo, M.; Evans, B.
Geosci. 2019, 126, 142−150.
(16) Wang, H.; Dalton, L.; Fan, M.; Guo, R.; McClure, J.; Crandall, A data-driven framework for permeability prediction of natural porous
D.; Chen, C. Deep-learning-based workflow for boundary and small rocks via microstructural characterization and pore-scale simulation.
target segmentation in digital rock images using UNet++ and IK-EBM. Eng. Comput. 2023, 39, 3895.
J. Pet. Sci. Eng. 2022, 215, No. 110596. (36) Tian, J.; Qi, C.; Sun, Y.; Yaseen, Z. M.; Pham, B. T. Permeability
(17) Phan, J.; Ruspini, L. C.; Lindseth, F. Automatic segmentation prediction of porous media using a combination of computational fluid
tool for 3D digital rocks by deep learning. Sci. Rep. 2021, 11, 19123. dynamics and hybrid machine learning methods. Eng. Comput. 2021,
(18) Patel, A. K.; Chatterjee, S. Computer vision-based limestone 37, 3455−3471.
rock-type classification using probabilistic neural network. Geosci. Front. (37) Houston, A. N.; Otten, W.; Falconer, R.; Monga, O.; Baveye, P.
2016, 7, 53−60. C.; Hapca, S. M. Quantification of the pore size distribution of soils:
(19) Wang, S.; Chen, Z.; Chen, S. Applicability of deep neural Assessment of existing software using tomographic and synthetic 3D
networks on production forecasting in Bakken shale reservoirs. J. Pet. images. Geoderma. 2017, 299, 73−82.
Sci. Eng. 2019, 179, 112−125. (38) Temizel, C.; Odi, U.; Balaji, K.; Aydin, H.; Santos, J. E.
(20) Unrau, S.; Torrione, P.; Hibbard, M.; Smith, R.; Olesen, L.; Classifying Facies in 3D Digital Rock Images Using Supervised and
Watson, J. Machine Learning Algorithms Applied to Detection of Well Unsupervised Approaches. Energies. 2022, 15, 7660.
Control Events. In SPE-188104-MS, 2017. DOI: 10.2118/188104-MS. (39) Santos, J. E.; Pyrcz, M. J.; Prodanović, M. 3D Dataset of binary
(21) Jha, H. S.; Khanal, A.; Seikh, H. M. D.; Lee, W. J. A comparative images: A collection of synthetically created digital rock images of
study on outlier detection techniques for noisy production data from complex media. Data Br. 2022, 40, No. 107797.

30222 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223
ACS Omega http://pubs.acs.org/journal/acsodf Article

(40) Willson, C. S.; Lu, N.; Likos, W. J. Quantification of Grain, Pore, Kaiser, L.; Kudlur, M.; Levenberg, J.; Mane, D.; Monga, R.; Moore, S.;
and Fluid Microstructure of Unsaturated Sand from X-Ray Computed Murray, D.; Olah, C.; Schuster, M.; Shlens, J.; Steiner, B.; Sutskever, I.;
Tomography Images. Geotech. Test. J. 2012, 35, No. 20120075. Talwar, K.; Tucker, P.; Vanhoucke, V.; Vasudevan, V.; Viegas, F.;
(41) Al-Bazzaz, W. H.; Al-Mehanna, Y. W. Porosity, Permeability, and Vinyals, O.; Warden, P.; Wattenberg, M.; Wicke, M.; Yu, Y.; Zheng, X.
MHR Calculations Using SEM and Thin-section Images for Character- TensorFlow: Large-Scale Machine Learning on Heterogeneous
izing Complex Mauddud-Burgan Carbonate Reservoir. In SPE-110730- Distributed Systems. arXiv 2016, DOI: 10.48550/arXiv.1603.04467.
MS, 2007. DOI: 10.2118/110730-MS. (61) Alqahtani, N.; Alzubaidi, F.; Armstrong, R. T.; Swietojanski, P.;
(42) Hakimi, M. H.; Shalaby, M. R.; Abdullah, W. H. Diagenetic Mostaghimi, P. Machine learning for predicting properties of porous
characteristics and reservoir quality of the Lower Cretaceous Biyadh media from 2d X-ray images. J. Pet. Sci. Eng. 2020, 184, No. 106514.
sandstones at Kharir oilfield in the western central Masila Basin, Yemen. (62) Chen, T.; Guestrin, C. XGBoost: A Scalable Tree Boosting
J. Asian Earth Sci. 2012, 51, 109−120. System. Proceedings of the 22nd ACM SIGKDD International Conference
(43) Peksa, A. E.; Wolf, K.-H.A.A.; Zitha, P. L. J. Bentheimer on Knowledge Discovery and Data Mining; 2016, pp 785−794..
sandstone revisited for experimental purposes. Mar. Pet. Geol. 2015, 67, (63) Cai, J.; Zhang, Z.; Wei, W.; Guo, D.; Li, S.; Zhao, P. The critical
701−719. factors for permeability-formation factor relation in reservoir rocks:
(44) Cui, I. 3D Sandstone Sample Images. Digit. Rocks Portal. 2022, Pore-throat ratio, tortuosity and connectivity. Energy. 2019, 188,
DOI: 10.17612/FRCM-CN23. No. 116051.
(45) Shikhov, I.; D’Eurydice, M. N.; Arns, J.-Y.; Arns, C. H. An (64) Goral, J.; Panja, P.; Deo, M.; Andrew, M.; Linden, S.; Schwarz, J.-
Experimental and Numerical Study of Relative Permeability Estimates O.; Wiegmann, A. Confinement Effect on Porosity and Permeability of
Using Spatially Resolved $$T_1$$-z NMR. Transp. Porous Media. Shales. Sci. Rep. 2020, 10, 49.
2017, 118, 225−250. (65) Fu, J.; Wang, M.; Chen, B.; Wang, J.; Xiao, D.; Luo, M.; Evans, B.
(46) Sidorenko, M.; Orlov, D.; Ebadi, M.; Koroteev, D. Deep learning A data-driven framework for permeability prediction of natural porous
in denoising of micro-computed tomography images of rock samples. rocks via microstructural characterization and pore-scale simulation.
Comput. Geosci. 2021, 151, No. 104716. Eng. Comput. 2023, 39, 3895−3926.
(47) Huang, T. T. Y.; Jones, A. S.; He, L. H.; Darendeliler, M. A.; (66) Hou, Z.; Cao, D.; Ji, S.; Cui, R.; Liu, Q. Enhancing digital rock
Swain, M. V. Characterisation of enamel white spot lesions using X-ray image resolution with a GAN constrained by prior and perceptual
micro-tomography. J. Dent. 2007, 35, 737−743. information. Comput. Geosci. 2021, 157, No. 104939.
(48) Cankaya, S. A comparative study of some estimation methods for (67) Kamrava, S.; Tahmasebi, P.; Sahimi, M. Enhancing images of
parameters and effects of outliers in simple regression model for shale formations by a hybrid stochastic and deep learning algorithm.
research on small ruminants. Trop. Anim. Health Prod. 2009, 41, 35−41. Neural Networks. 2019, 118, 310−320.
(49) Berg, S.; Kutra, D.; Kroeger, T.; Straehle, C. N.; Kausler, B. X.;
Haubold, C.; Schiegg, M.; Ales, J.; Beier, T.; Rudy, M.; Eren, K.;
Cervantes, J. I.; Xu, B.; Beuttenmueller, F.; Wolny, A.; Zhang, C.;
Koethe, U.; Hamprecht, F. A.; Kreshuk, A. ilastik: interactive machine
learning for (bio)image analysis. Nat. Methods. 2019, 16, 1226−1232.
(50) Otsu, N. A Threshold Selection Method from Gray-Level
Histograms. IEEE Trans. Syst. Man. Cybern. 1979, 9, 62−66.
(51) van der Walt, S.; Schönberger, J. L.; Nunez-Iglesias, J.; Boulogne,
F.; Warner, J. D.; Yager, N.; Gouillart, E.; Yu, T. scikit-image: image
processing in Python. PeerJ. 2014, 2, No. e453.
(52) Gostick, J. T. Versatile and efficient pore network extraction
method using marker-based watershed segmentation. Phys. Rev. E
2017, 96, No. 023307.
(53) Latt, J.; Malaspinas, O.; Kontaxakis, D.; Parmigiani, A.; Lagrava,
D.; Brogi, F.; Belgacem, M. B.; Thorimbert, Y.; Leclaire, S.; Li, S.;
Marson, F.; Lemus, J.; Kotsalos, C.; Conradin, R.; Coreixas, C.;
Petkantchin, R.; Raynaud, F.; Beny, J.; Chopard, B. Palabos: Parallel
Lattice Boltzmann Solver. Comput. Math. with Appl. 2021, 81, 334−
350.
(54) Mohammed Azarudeen, J.; Kaja Bantha Navas, R.; Vikas Reddy,
G.; Gowsik Saran, R.S.; Prakash, S.; Anderson, A. Optimizing the
process parameters for Powder metallurgy electrode in Electrical
Discharge Machining. Mater. Today Proc. 2021, 46, 9372.
(55) Bear, J.; Bachmat, Y. Introduction to Modeling of Transport
Phenomena in Porous Media; Springer Netherlands: Dordrecht, 1990.
DOI: 10.1007/978-94-009-1926-6.
(56) Lin, G.; Shen, W. Research on convolutional neural network
based on improved Relu piecewise activation function. Procedia
Comput. Sci. 2018, 131, 977−984.
(57) Wang, S.-H.; Muhammad, K.; Hong, J.; Sangaiah, A. K.; Zhang,
Y.-D. Alcoholism identification via convolutional neural network based
on parametric ReLU, dropout, and batch normalization. Neural
Comput. Appl. 2020, 32, 665−680.
(58) Huber, P. J. Robust Estimation of a Location Parameter. Ann.
Math. Stat. 1964, 35, 73−101.
(59) Kingma, D. P.; Ba, J. Adam: A Method for Stochastic
Optimization. arXiv 2014, DOI: 10.48550/arXiv.1412.6980.
(60) Abadi, M.; Agarwal, A.; Barham, P.; Brevdo, E.; Chen, Z.; Citro,
C.; Corrado, G.S.; Davis, A.; Dean, J.; Devin, M.; Ghemawat, S.;
Goodfellow, I.; Harp, A.; Irving, G.; Isard, M.; Jia, Y.; Jozefowicz, R.;

30223 https://doi.org/10.1021/acsomega.3c10131
ACS Omega 2024, 9, 30205−30223