You are on page 1of 38

Journal Pre-proofs

Fast prediction of methane adsorption in shale nanopores using kinetic theory


and machine learning algorithm

MengCheng Huang, HengYu Xu, Hao Yu, HouLin Zhang, Marembo Micheal,
XinHeng Yuan, HengAn Wu

PII: S1385-8947(22)02710-3
DOI: https://doi.org/10.1016/j.cej.2022.137221
Reference: CEJ 137221

To appear in: Chemical Engineering Journal

Received Date: 11 April 2022


Revised Date: 23 May 2022
Accepted Date: 24 May 2022

Please cite this article as: M. Huang, H. Xu, H. Yu, H. Zhang, M. Micheal, X. Yuan, H. Wu, Fast prediction of
methane adsorption in shale nanopores using kinetic theory and machine learning algorithm, Chemical
Engineering Journal (2022), doi: https://doi.org/10.1016/j.cej.2022.137221

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover
page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version
will undergo additional copyediting, typesetting and review before it is published in its final form, but we are
providing this version to give early visibility of the article. Please note that, during the production process, errors
may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

© 2022 Elsevier B.V. All rights reserved.


1 Fast prediction of methane adsorption in shale nanopores using kinetic
2 theory and machine learning algorithm
3 MengCheng Huang#, HengYu Xu#, Hao Yu*, HouLin Zhang, Marembo Micheal, XinHeng Yuan
4 and HengAn Wu*
5 CAS Key Laboratory of Mechanical Behavior and Design of Materials, Department of Modern
6 Mechanics, University of Science and Technology of China, Hefei 230027, China
7 #These authors contributed equally to this work.
8 *Corresponding email: yuhaoo@ustc.edu.cn (Hao Yu); wuha@ustc.edu.cn (HengAn Wu)
9

10 Abstract
11 Understanding the gas adsorption behavior in shale nanopores is essential for the reservoir
12 estimation and performance prediction of shale gas, which is still unclear considering the complexity
13 of geological environment and nanoporous structure of shale. In general, traditional methods based on
14 experiments and molecular dynamics (MD) simulations are always expensive and time consuming. In
15 this work, a machine learning (ML) framework to predict the methane adsorption behavior in shale
16 nanopores is constructed from the microscopic and kinetic theory perspectives, where three novel
17 parameters related to potential energy distribution (PED) are proposed to represent the methane
18 adsorption characteristics of shale slit nanopores. Machine learning algorithm based on the uniformly
19 constructed dataset is introduced to realize fast and accurate prediction of methane adsorption behavior,
20 which is well validated by typical inorganic and organic models. Moreover, the application of the
21 proposed ML model to predict the adsorption behavior of methane in different geological conditions
22 (e.g., pressure and temperature) is performed, indicating its feasibility to predict the gas adsorption
23 behavior in shale nanopores with ultra-fast computation speed, that would be beneficial for the
24 exploitation and development of shale gas reservoirs. The insights gained in this work is also
25 instructive for the prediction of nano-confined fluid behavior under complex environmental factors.
26

27 Keywords: shale nanopores, methane adsorption behavior, kinetic theory, molecular simulations,
28 machine learning

1
1 1. Introduction

2 As a relatively clean energy source, shale gas has lower carbon emissions as compared to other
3 conventional fossil fuels (like oil, coal, etc.) [1–3]. For the same energy output, other fossil fuels
4 produce twice as much carbon emissions as shale gas [4]. Therefore, replacing the highlighted
5 conventional fossil fuels with relatively clean energy sources like shale gas, is meaningful. Global
6 reserves of shale gas are abundant [5, 64], and because of the unique storage state of shale gas, efforts
7 are still needed to exploit the gas without endangering the environment by ensuring low carbon
8 footprints [6,7]. The shale gas storage capacity arises from three parts: free gas in pore and fractures,
9 adsorbed gas on the shale matrix, and dissolved gas in kerogen and formation fluid [8]. The adsorbed
10 gas in shale accounts for 20-85% of total gas-in-place, which gradually releases from the matrix as the
11 pressure drops. The abundant proportion of adsorbed gas in shale reservoirs guarantees the long-term
12 stable production in exploitation [9]. Therefore, quantitative evaluation of methane adsorption capacity
13 is of great necessity for accurate estimation of reservoir reserves and production prediction in shale
14 gas reservoirs.
15 Shale gas is stored in the nanoporous shale matrix [10] composed of inorganic matters (e.g., clay,
16 quartz, and calcite) and organic matter (i.e., kerogen). Many experiments have been applied to reveal
17 the gas adsorption behavior in shale matrix [11]. For example, Zhang et al., [12] carried out methane
18 adsorption experiments on crushed samples from Eocene Green River Formation, Devoniane-
19 Mississippian Woodford Shale, and Upper Cretaceous Cameo coal. This work found that there is a
20 strong correlation between the type of kerogen and thermal maturity on the total adsorbed amount. The
21 methane adsorption isotherms indicate a decrease in the capacity in the following sequence kerogen
22 Type III > Type II > Type I. Chareonsuppanimit et al., [13] carried out high pressure adsorption
23 measurements on New Albany shales from the Illinois basin using nitrogen, methane, and carbon
24 dioxide. The adsorption ratios of carbon dioxide to nitrogen and carbon dioxide to methane are
25 considerably greater than the results measured in coals. However, the cognition from the experimental
26 approach is relatively macroscopic to evaluate the overall adsorption properties of shale matrix, it's
27 still challenging to obtain the underlying adsorption mechanisms in shale nanopores owing to the
28 limitation of resolution and accuracy of experiments [14, 15], which is critical for the construction of
29 theoretical framework for gas adsorption and storage in shale.
2
1 To precisely investigate the methane adsorption behaviors in shale from a microscopic
2 perspective, molecular simulation methods such as Monte Carlo (MC) and molecular dynamics (MD)
3 simulations have been carried out [14,15]. Many efforts have been devoted to constructing different
4 types of realistic shale models in molecular simulations [10,16–18]. For example, inorganic minerals
5 of shale slit nanopores composed by calcite [19,20], illite [21], kaolinite [15], and montmorillonite
6 [23] and organic minerals of shale slit nanopores such as multilayer graphene slit [15,24,25], carbon
7 nanotube [26], graphene oxide (GO) model [21,22], and kerogen model [29–31, 68] are built to explore
8 the adsorption mechanisms of shale gas [66]. Upon these efforts, much progress has been successfully
9 achieved regarding the understanding of shale gas adsorption behavior from the view of atomic-scale,
10 which highlights the significant role of molecular simulations in this topic. For example, Tian et al.,
11 [67] provided important fundamental understandings and insights into accurate estimation of gas-in-
12 place in shale reservoirs based on molecular simulation. Wang et al., [68, 69] used GCMC simulations
13 to study the competitive adsorption between CO2 and typical hydrocarbon components (C𝐻4, 𝐶2𝐻6, and
14 𝐶3𝐻8), and gained a more in-depth understanding of multi-component gas recovery mechanisms
15 within realistic shale kerogen nanopores and sheds light on the CO2 sequestration in shale reservoirs.

16 Huang et al., [70] acquired deep insights into the effect of kerogen maturity and moisture content on
17 the interaction between 𝐶𝐻4/𝐶𝑂2 and kerogen at microscopic scale by using molecular simulation. At

18 the same time, it is undeniable that such accurate atomic-scale simulation brings about the difficulties
19 on the modeling and solving processes, where the high performance computing clusters are needed to
20 implement the corresponding molecular simulations. Even so, several or more days may be required
21 for simulating one case that merely considers a combination of nanopore model and geological
22 parameters (i.e., temperature and pressure). In this context, a new method to quickly and accurately
23 characterize and analyze the methane adsorption behavior in various shale nanopores with different
24 environmental parameters is urgently needed.
25 In recent years, machine learning method has been applied in many research fields and made
26 significant progress that is difficult for classic methods to achieve [32–34], including the prediction on
27 gas adsorption behavior in shale [36-39]. For instance, Meng et al., [35] demonstrated the capability
28 of machine learning for prediction of shale gas adsorption, and they found that a well-trained model
29 can potentially be built into a numerical frame to optimize production curves of shale gas. Fanourgakis
30 et al., [36] proposed a new set of descriptors that are appropriate for machine learning methods to
3
1 accurately predict the gas adsorption capacities of nanoporous materials. Amar et al., [37] utilized two
2 rigorous data-driven techniques, namely gene expression programming (GEP) and group method of
3 data handling (GMDH) to provide accurate and reliable explicit mathematical expression for
4 predicting methane adsorption. To a certain extent, the above applications show that machine learning
5 method can significantly accelerate the shale gas adsorption capacity prediction process in a broad
6 range of storage conditions. These applications always focus on predicting gas adsorption capacity
7 which is a crucial characteristic of gas adsorption behavior. In this work, the methane adsorption
8 profile is investigated using machine learning method in order to precisely and accurately capture
9 methane adsorption behavior. Considering that the methane adsorption behavior is microscopic in
10 nature, the analysis from the microscopic perspective provides more information and cognitions, to
11 reveal the adsorption mechanisms of shale gas. Besides, the kinetic theory is an important basis for
12 methane adsorption behavior and molecular simulation [63, 65]. Therefore, in this work, the
13 parameters used in machine learning method are derived from slit nanopore structures and the
14 interactions between methane molecule and slit nanopore are involved from microscopic perspectives
15 based on kinetic theory.
16 The methane adsorption behavior is microscopic in nature and the microstructure of shale
17 nanopores plays a key role in the methane adsorption behavior. Some studies have made significant
18 progress in discussing the definition of accessible pore width according to the microstructure of the
19 wall of the shale nanopores. Ghasemzadeh et al. [38] and Yu et al. [39] contrasted some different
20 methods used to define the accessible pore width of slit nanopore. Both of them found that the method
21 which uses the He probe to acquire accessible pore width, performs best and can illustrate the
22 accessible pore width more reasonably. However, it is always difficult to characterize wall
23 microstructure of shale nanopores directly. Considering kinetic theory and the inter-molecular
24 interaction, the methane adsorption behavior is mainly determined by temperature, pressure, and the
25 potential energy distribution (PED) of the shale nanopores. In particular, PED can be utilized to
26 represent the role of the micro configuration of shale nanopores during the methane adsorption process
27 [40–42]. Therefore, some novel parameters on the basis of the PED are proposed to represent
28 characteristic of the micro configuration of the shale nanopores during methane adsorption process.
29 Besides, the parameters utilized to analyze methane adsorption behavior are not limited by the
30 complexity of the shale nanopores. A group of uniformly defined carbon slit nanopores are precisely
4
1 constructed and simulated to extract characteristic parameters and methane adsorption behavior results.
2 These parameters and results will be used to train machine learning model and the trained model will
3 be utilized to realize the prediction of methane adsorption behavior of the realistic slit nanopore model.
4 A workflow is sequentially arranged to illustrate the methane adsorption behavior rapid prediction
5 process in a realistic model, as shown in Fig. 1. The main purpose of this work is to rapidly predict the
6 adsorption profiles and analyze methane adsorption behavior in realistic shale slit nanopore models
7 through machine learning. To achieve the rapid and accurate prediction, a series of datasets consisting
8 of simulation results are established to train the machine learning model. Then, the slit nanopore
9 methane adsorption capability parameter (𝛾) acquisition algorithm is utilized to characterize the
10 properties of the realistic shale slit nanopore. The results of the analysis are regarded as the input
11 parameters for the final prediction of machine learning process. Overall, the rapid prediction of
12 methane adsorption behavior in shale nanopores can be realized through the research workflow which
13 is based on the kinetic theory and machine learning method. The context of this article is organized as
14 follows: In section 2, the simulation method and machine learning method are introduced, three novel
15 parameters used to represent the slit nanopore structure characteristic and the parameter 𝛾 acquisition
16 algorithm are proposed. Section 3 presents and analyzes the results acquired by the simulation method
17 and machine learning method, discusses the effectivity of the parameter 𝛾 acquisition and application
18 algorithm, and investigates methane adsorption behavior in shale nanopores. Section 4 briefly
19 concludes the progress made in this work and comes up with some shortcomings along with some
20 solutions.

21 2. Methodology and model

22 2.1 Molecular simulation implementation

23 2.1.1 Construction of the molecular model

24 The generation of training dataset of methane adsorption requires the construction of slit
25 nanopores with various potential energy distribution (PED) characteristics. For the inorganic and
26 organic slit nanopores, their structures are complicated since they contain specific PED features and it
27 is tedious to unify the reference position used to determine the accessible pore width. Therefore, the

5
1 dataset used to train machine learning model should have different PED characteristics and uniformly
2 defined reference position to infer the accessible pore width. Many studies have illustrated the accuracy
3 of methane adsorption simulation within crystal carbon slit nanopores, hence this structure is adopted
4 to construct the carbon plate of slit nanopores for the training dataset. The width of the carbon slit
5 pores are 4 nm, and the roughness of the slit nanopore is introduced by rearranging the coordinates of
6 atoms located on the surface of the plate [43, 67]. Half of these atoms are randomly selected and moved
7 along the direction perpendicular to the plate. The maximum moving distance is set as 0, 0.5, 1.0, 1.5,
8 2.0, and 2.5 Å, respectively, to create slit nanopores with different roughness. The random moving
9 distance satisfies the normal distribution. Every plate is constructed independently to ensure their
10 uniqueness, and the final configurations of the carbon slit nanopores are shown in Fig. 2(a-f). These
11 carbon slit nanopores will be used as the training dataset in machine learning method to investigate
12 the slit nanopore methane adsorption behavior.
13 Considering that the nanopores in shale are composed of inorganic and organic matters, some
14 plates with specific substances are created to construct these nanopores. For the inorganic nanopores,
15 calcite slit nanopore is a kind of typical structure and it will be utilized to represent inorganic slit
16 nanopore to investigate gas adsorption behavior [20]. Here, the standard crystal structure of calcite is
17 collected from Yu et al.'s work [44]. Two pieces of plates are placed in parallel to form the 4 nm
18 inorganic slit nanopore as shown in Fig. 2(g). As for the organic nanopores, kerogen is generally
19 recognized as the main component of organic matters in shale, therefore, it is utilized to create organic
20 plates in this work. Ungerer et al. [45] proposed six kerogen molecule structures, and based on their
21 proposal, type II-C kerogen monomer is selected for this work. Two pieces of the kerogen plates are
22 independently created. Their density is 1.2 g/cm3, which is consistent with the experimental detection
23 result [46]. The 4 nm organic nanopore is constructed by placing the two pieces of kerogen plates in
24 parallel, as shown in Fig. 2(h). The solid walls of both inorganic and organic slit nanopores are
25 recognized as rigid in the simulations. Note that, these slit nanopores representing realistic shale
26 nanopores, will be used as verification data in machine learning method to investigate the methane
27 adsorption behavior. The details about constructing molecular models are put in Supplementary
28 Material.

29 2.1.2 GCMC/MD simulations


6
1 The adsorption behavior of methane in slit nanopores under different conditions are calculated by
2 Grand Canonical Monte Carlo (GCMC) simulations. This method can exchange methane particles
3 with an imaginary ideal gas reservoir at the specified temperature and chemical potential, and attempts
4 Monte Carlo move for every particle [47]. In this work, the chemical potential of the system is
5 converted to pressure according to the equation proposed by Frenkel et al. [36]. Therefore, the
6 temperature and pressure of each condition are respectively set as 300, 325, 350, 375, 400 K, and 1 to
7 60 MPa with the interval of 1 MPa. These density profiles of methane in slit nanopores obtained from
8 simulations, as shown in Fig. 3, will be used as the dataset for machine learning to investigate the slit
9 nanopore methane adsorption behavior.
10 The surface properties of realistic models significantly affect the adsorption behavior of methane
11 molecules, and these properties are introduced in the training models by various roughness of the
12 crystal carbon surface. To quantitatively evaluate the properties, the potential energy distribution (PED)
13 of each solid wall is calculated. PED reflects the interaction energy between adsorbate particles and
14 wall surface at all vicinages, which has been widely used to reveal the dynamic behaviors of methane
15 on solid walls from molecular point of view [40–42]. The calculation method of PED is to employ a
16 probe near the solid wall, and then the potential energy of this probe is calculated and recognized as
17 the potential energy of this position. Through moving this probe to each location in the vicinage of the
18 solid wall, the PED of the surface is captured. In this work, the methane molecule serves as a probe to
19 obtain the PED of each slit nanopore wall surface. The initial distance between the probe and the wall
20 surface for PED scanning is 0.02 nm, and the scanning step of each move is 0.02 nm. The scanning is
21 performed in both x, y, and z directions, and the ending distance between probe and wall surface is 2
22 nm (half of the nanopore width).
23 To maintain consistency with previous studies [48], in this work, the methane molecules are
24 regarded as single particles, and the three boundaries of the simulation box are periodic. The details of
25 GCMC/MD simulations and the validations are put in Supplementary Material (Fig. S1).

26 2.1.3 Force field

27 Force fields are used for describing the basic intra- and inter-molecular forces in the simulation.
28 In this work, we use Lennard-Jones (L-J) 12-6 potential to describe the interactions between slit
29 nanopores and methane molecules. L-J potential is a common force field to investigate the flow [44,49]
7
1 and adsorption [50,51] of hydrocarbon in shale nanopores, and the results show great consistency and
2 reliability. The expression of L-J potential is shown as [19]:

3 𝑉𝑖𝑗 = 𝜀𝑖𝑗
𝑅𝑖𝑗 12
[( )
𝑟𝑖𝑗 ( )]
―2
𝑅𝑖𝑗 6
𝑟𝑖𝑗
(1)

4 where 𝑉𝑖𝑗 is the L-J potential between atom 𝑖 and 𝑗, 𝑟𝑖𝑗 is the distance between the atom 𝑖 and 𝑗, 𝜀𝑖𝑗
5 represents the depth of the L-J potential well, 𝑅𝑖𝑗 represents the distance between two atoms when their

6 potential energy is zero. The profile of L-J potential is shown as the blue profile in Fig. 3. The
7 parameters of L-J potential used in this work are obtained from previous studies [48,52,53].
8 The interaction force of two atoms can be calculated by the gradient of the L-J potential.

9 𝐅(𝑟𝑖𝑗) = ― ∇𝑉(𝑟𝑖𝑗) = 12𝜀𝑖𝑗 [ 𝑅𝑖𝑗12


𝑟𝑖𝑗 13
𝑅𝑖𝑗6
―𝑟
𝑖𝑗
7 ]r
𝑖𝑗 (2)

10 The interaction force profile between two atoms is shown as the orange profile in Fig. 3. One can see
11 that the force between atoms 𝑖 and 𝑗 is repulsive when 𝑟𝑖𝑗 < 𝑅0, and changes to attractive when 𝑟𝑖𝑗 >
12 𝑅0.

13 2.2 Characterization parameters of the slit nanopore

14 The adsorption behavior of methane in the slit nanopore is greatly determined by the surface
15 properties of slit and the interactions between methane molecules and slit walls. As discussed above,
16 the surface properties of a slit can be effectively described by its potential energy distribution (PED).
17 Fig. 4 shows the PEDs of some training models with different roughness. For the specific (𝑥,𝑦) position,
18 its potential energy varies with the changing coordinate of z. Consequently, each (𝑥,𝑦) position has
19 different minimum potential energy in the specific z coordinate. Fig. 4(a)-(c) illustrates the minimum
20 potential energy surface of each PED, and it can be seen that as the roughness of wall increase, the
21 regularity of the energy surface increases. Considering that the minimum potential energy of each (𝑥,𝑦)
22 position denotes the affinity of methane molecules when methane adsorbs here, the average value (𝜔)
23 and regularity (𝜉) of the minimum potential energy surface are used to reflect its properties. These two
24 factors are calculated by:
𝑚 𝑛
∑𝑖 ∑𝑗 |𝑚𝑖𝑛𝑖𝑚𝑢𝑛 𝑝𝑜𝑡𝑒𝑛𝑡𝑖𝑎𝑙 𝑒𝑛𝑒𝑟𝑔𝑦(𝑥𝑖, 𝑦𝑗)|
25 𝜔= 𝑚𝑛
(3)

𝑚 𝑛
∑𝑖 ∑𝑗 (|𝑚𝑖𝑛𝑖𝑚𝑢𝑚 𝑝𝑜𝑡𝑒𝑛𝑡𝑖𝑎𝑙 𝑒𝑛𝑒𝑟𝑔𝑦(𝑥𝑖,𝑦𝑗)| ― 𝜔)2
26 𝜉= 2
(4)
𝑚𝑛
8
1 Besides, according to Fig. 3 and Eq. 2, when the potential reaches the minimum, the interaction
2 force gets the turning points change from repulsion to attraction. Therefore, the distance between
3 methane molecules and the reference surface of slit walls, that makes the potential energy of (𝑥,𝑦)
4 position reach the minimum, reflects the starting position to adsorb methane molecules in the location.
5 Fig. 4(d)-(f) is the potential-minimum distance of each (𝑥,𝑦) position, the regularity of which is also
6 positively correlated to the surface roughness. To evaluate the adsorption position on the silt walls, the
7 average potential-minimum distance (𝑟𝑚𝑖𝑛) is calculated through:
𝑚 𝑛
∑𝑖 ∑𝑗 𝑝𝑜𝑡𝑒𝑛𝑡𝑖𝑎𝑙 ― 𝑚𝑖𝑛𝑖𝑚𝑢𝑚 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒(𝑥𝑖,𝑦𝑗)
8 𝑟𝑚𝑖𝑛 = 𝑚𝑛
(5)

9 Since the L-J potential is the main governing principle that controls the slit nanopore methane
10 adsorption behavior in the GCMC/MD simulation process, the adsorption capability of the slit
11 nanopore is mainly related to the attractive force between methane molecule and slit walls. Specifically,
12 the methane adsorption capability in slit nanopore should be positively correlated to the shaded region
13 shown as the inter-molecular force graph in Fig. 3. The area under the shaded region can be calculated
14 by integrating the profile function as:

15
+∞
𝑆 = ∫𝑅 𝐅(𝑟)𝑑𝑟 = ∫𝑅
+∞
―12𝜀 [ 𝑅12
𝑟13 ]
𝑅6
― 𝑟7 𝑑𝑟 = 𝜀 (6)

16 The integral value 𝜀 is approximately equal to the average minimum potential energy 𝜔. Therefore,
17 the slit nanopore methane adsorption capability is positively correlated to the average minimum
18 potential energy 𝜔. Besides, when the 𝜔 is fixed, the smaller 𝑟𝑚𝑖𝑛 intuitively results into stronger

19 methane adsorption capability. Therefore, it is inferred that the slit nanopore methane adsorption
20 capability is positively correlated to 𝜔, and negatively correlated to 𝑟𝑚𝑖𝑛, which is defined as the slit

21 nanopore methane adsorption capability parameter 𝛾 in this work:


𝜔
22 𝛾 = 𝑟𝑚𝑖𝑛 (7)

23 The dimension of parameter 𝛾 is:


dim 𝜔 𝐿2𝑀𝑇 ―2
24 dim 𝛾 = dim 𝑟𝑚𝑖𝑛 = 𝐿 = 𝐿𝑀𝑇 ―2 (8)

25 According to Eq. (6), the dimension of the adsorption capability parameter, 𝛾, is the same as the
26 dimension of force, that is, the adsorption capability parameter 𝛾 can be regarded as force.
27 Herein, the adsorption characteristics of the slit nanopore are represented by parameters 𝜔, 𝜉, and
28 𝛾. For each slit nanopore, these parameters will be extracted and then used as part of the input
9
1 parameters for the machine learning method, to investigate the slit nanopore methane adsorption
2 behavior.

3 2.3 Machine learning implementation

4 2.3.1 Fully Connected Neural Network (FCNN)

5 Inspired by the structure of human brain neurons, artificial neural network (ANN) is emerged [54].
6 Fully Connected Neural Networks (FCNN) is a type of ANN where the architecture is in such a way
7 that all nodes, or neurons, in one layer are connected to the neurons in the next layer, as shown in the
8 middle graph of Fig. 5. Generally speaking, machine learning method can be divided into three main
9 types: supervised learning, unsupervised learning, and reinforcement learning [55]. The supervised
10 learning is the most widely used method because of its ability to find the implicit relationship between
11 the features and the properties. This work is based on the FCNN and supervised learning. The principle
12 of supervised machine learning [26] is analogous to the standard fitting procedure: it tries to find the
13 implicit function that connects the known inputs to unknown outputs. This desired result within
14 unknown domains is estimated based on the extrapolation of patterns found in the labeled training
15 data.
16 FCNN model contains one input layer, several hidden layers, and one output layer [56]. In each
17 layer, there are multiple nodes that receive values from the predecessor nodes, perform computations
18 using activation functions, and then deliver the outputs to the successor nodes. In this work, the
19 activation function adopts the rectified linear unit (ReLU) equation [57]:
20 𝑅𝑒𝐿𝑈(𝑥) = 𝑚𝑎𝑥 (0,𝑥) (9)
21 According to the complexity of this problem, the network consists of two hidden layers, both of the
22 layers have 400 neurons with the ReLU activation function. Networks are trained over 𝑛𝑒𝑝𝑜𝑐ℎ𝑠 = 2000

23 using Adam optimizer [58].


24 In order to quickly and accurately predict the methane adsorption profiles, 1800 methane
25 adsorption profiles in carbon slit nanopores are prepared as the dataset. These profiles are randomly
26 divided into five parts, in which one part (360 profiles) is taken as the test dataset and the rest of the
27 four parts (1440 profiles) are regarded as training datasets. This data partitioning process is performed
28 five times, and in each time, one of the five parts is set as test dataset in turn. This process is called 5-

10
1 fold cross-validation method, and it can avoid the error caused by the specific data division. As
2 discussed above, the methane adsorption profile in slit nanopore is mainly determined by environment
3 temperature 𝑇, pressure 𝑃, the potential energy distribution (PED) of the slit nanopore, and the
4 interaction between the slit walls and methane molecules. The latter two factors can be generally
5 represented by 𝜔, 𝜉, and 𝛾 of the slit nanopore, as shown in the graph on the left of Fig. 5. The
6 parameter 𝛾 is inferred by the minimum-potential distance (𝑟𝑚𝑖𝑛), relating to the uniformly defined

7 reference position of walls, so the 𝛾 can only be directly acquired in the carbon slit nanopores. As for
8 other slit nanopores, the methane adsorption capability parameter 𝛾 cannot be calculated directly by
9 its formula (as stated in Eq. (7)) because the reference position of the slit nanopore cannot be
10 determined accurately. Therefore, an algorithm to acquire parameter 𝛾 from another perspective, will
11 be introduced in section 2.4. Besides, methane adsorption profiles are discretized to distance-
12 adsorption pair (𝑥,𝑦) to adapt the FCNN model architecture. Therefore, parameters 𝜔, 𝜉, 𝛾, T, P, and
13 𝑥 are used as the input layer parameters and the adsorption amount 𝑦 is used as output layer parameter
14 to train the FCNN model, as shown in Fig. 5. Subsequently, the predicted adsorption profile 𝑦 =
15 𝜑prediction(𝜔, 𝜉, 𝛾, 𝑇, 𝑃, 𝑥) can be obtained quickly through the trained FCNN model. Later in the
16 proceeding subsections, the prediction speed and accuracy of the trained FCNN model will be
17 discussed.

18 2.3.2 Evaluation metrics

19 There are many forms of evaluation metric systems to test the performance of a machine learning
20 model. Since the problem that is being resolved is a regression problem, root mean square error (RMSE)
21 and coefficient of determination (𝑅2) are commonly used as the corresponding evaluation metrics.
22 𝑅𝑀𝑆𝐸 and 𝑅2 can be represented as:
1 𝑛
23 𝑅𝑀𝑆𝐸 = ∑ (
𝑛 𝑖 = 1 𝑦𝑖 ― 𝑦𝑖)2 (10)
∑(𝑦𝑖 ― 𝑦𝑖)2
24 𝑅2 = 1 ― ∑ ( (11)
𝑦𝑖 ― 𝑦𝑖)2

25 where 𝑦𝑖 is the true value and 𝑦𝑖 is the predicted value.

26 2.4 Acquisition and application algorithm of parameter 𝜸


27 Parameter 𝛾 defines the interaction between methane molecules and slit walls, and is an essential
28 parameter for the prediction capability of a machine learning model. To acquire this parameter through
11
1 its formula (as stated in Eq. (7)), the minimum-potential distances (𝑟𝑚𝑖𝑛) should be accurately

2 calculated, which requires a clear definition of the reference surface of slit walls. For the training model
3 of crystal carbon structure, the reference surface is uniformly defined considering its regular structure.
4 However, for other realistic models, it is hard to define the proper reference surface. Therefore, an
5 alternative method to acquire 𝛾 is urgently need sooner than later. Fig. 6 shows an algorithm to fix this
6 problem, where the type II-C kerogen slit nanopore is utilized as an example of a realistic slit nanopore
7 to demonstrate how this algorithm works. After the type II-C kerogen slit nanopore model is
8 constructed, its average minimum potential energy 𝜔0 and the minimum potential energy standard
9 deviation 𝜉0 are calculated through simulation. To obtain the methane adsorption profile in the type
10 II-C kerogen slit nanopore at one arbitrary temperature (𝑇0) and pressure (𝑃0) condition, two different
11 strategies are provided in the algorithm. On the one hand, the methane adsorption profile 𝜑𝑠𝑖𝑚𝑢𝑙𝑎𝑡𝑖𝑜𝑛(
12 𝜔0,𝜉0,𝑇0, 𝑃0) can be acquired by GCMC/MD simulation, shown in Fig. 6 as the simulation path. On
13 the other hand, a group of parameters 𝛾𝑖 (𝑖 = 1, 2,…,𝑛) are set as input parameters of the trained FCNN
14 model, then the methane adsorption profiles 𝜑𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛(𝜔0, 𝜉0, 𝑇0, 𝑃0, 𝛾𝑖), 𝑖 = 1,2,…,𝑛 are acquired by

15 the trained FCNN model, shown in Fig. 6 as the machine learning path. Afterwards, to find out the

16 best matched parameter 𝛾0, the difference between 𝜑𝑠𝑖𝑚𝑢𝑙𝑎𝑡𝑖𝑜𝑛 and 𝜑𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝛾𝑖 is calculated

17 ( 𝑖 = 1, 2,…,𝑛), and the minimum difference can thus be calculated as the discriminant formula:

18 𝑚𝑖𝑛 (𝑒𝑟𝑟) = 𝑚𝑖𝑛 (∑ 𝑥𝑚


|𝜑𝑠𝑖𝑚𝑢𝑙𝑎𝑡𝑖𝑜𝑛(𝑥𝑚) ― 𝜑𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛(𝑥𝑚)𝛾 |)
𝑖
𝑖 = 1, 2,…, 𝑛 (12)

19 The parameter 𝑥𝑚 represents each discrete x-coordinate of the adsorption profile, and the discrete
20 spacing is 0.01 nm. The 𝛾𝑖 (𝑖 = 1, 2,…,𝑛) corresponding to the minimum difference is recognized as
21 the best matched methane adsorption capability parameter 𝛾0 of the type II-C kerogen slit nanopore.
22 Finally, the parameters 𝜔0, 𝜉0, and 𝛾0 are used to predict the methane adsorption profile in type II-C

23 kerogen slit nanopore at the arbitrary temperature 𝑇 and pressure 𝑃 condition through trained FCNN
24 model 𝜑𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛(𝜔0, 𝜉0, 𝑇, 𝑃, 𝛾0), shown in Fig. 6 as the parameter 𝛾0 acquisition and application path.

25 3. Results and discussion

26 3.1 Analysis from GCMC/MD Simulations

27 GCMC/MD simulations are applied to obtain the methane adsorption profiles in carbon slit

12
1 nanopores with different roughness, as well as the potential energy distribution (PED) of slit walls
2 represented by the average minimum potential energy 𝜔, minimum potential energy standard deviation
3 𝜉, and methane adsorption capability parameter 𝛾, as shown in Fig. 7(a). From Fig. 7(a), it can be seen
4 that the 𝜔 has a negative correlation to the solid roughness, implying that the rough wall contains weak
5 affinity to the methane molecule. Meanwhile, the 𝜉 is positively correlated to the solid roughness,
6 which indicates that the minimum-potential surface of rough solid is irregular and thus contains less
7 adsorption sites for methane adsorption. Therefore, the rougher slit has poorer ability to adsorb
8 methane, which is also proved by the declining tendency of the 𝛾. From the view of methane adsorption
9 profile at the condition of 300 K and 60 MPa, as shown in Fig. 7(b), it also demonstrates the
10 relationship between the methane adsorption capacity and the roughness of the slit walls. Methane
11 contains the obvious multilayer adsorption pattern in carbon slit nanopores. Specifically, the
12 adsorption profile shows dramatic first and second adsorption peaks in all slit nanopores, and slight
13 third peak for some smooth slit nanopores. Through analysis, it is found that for those carbon slit
14 nanopores with lower roughness (Roughness = 0.0 and 0.5), the distance between adjacent adsorption
15 peaks is approximately 0.38 nm which is the same as the diameter of the methane molecule. In these
16 slit nanopores, methane molecules adsorb tightly on the wall, which means the corresponding carbon
17 slit nanopore has strong affinity to the methane molecules. In contrast, for those rough carbon slit
18 nanopores (Roughness = 2.0 and 2.5), the distance between adjacent adsorption peaks obviously
19 increases, implying a loose adsorption state of methane in the vicinity of the slit walls.
20 From the simulation results, it is seen that the PED of slit walls, represented by 𝜔, 𝜉, and 𝛾 can
21 effectively sketch the adsorption capacity of methane in carbon slit nanopores. Thus, it is reasonable
22 to regard these parameters as inputs to the machine learning model. In the next subsection, these
23 parameters, as well as the corresponding methane adsorption profiles will be utilized as the dataset to
24 train the machine learning model for the purpose of accurate and fast prediction.

25 3.2 Prediction by Machine learning algorithm

26 The machine learning model is trained based on the existing dataset, which is composed of many
27 sets of carbon slit nanopore characteristic parameter, temperature, pressure, and the corresponding
28 methane adsorption profile. By using the 5-fold cross-validation method, five groups of evaluation
29 results are acquired as shown in Table. 1, The cost curves during the training process of machine
13
1 learning model are put in Supplementary Material (Fig. S2). It is illustrated that in all training processes,
2 the R2 values between the prediction and simulation results are similar and usually higher than 0.99,
3 indicating that the constructed FCNN models can effectively predict the methane adsorption profiles.
4 Among them, the first fold performs best (𝑅𝑀𝑆𝐸 = 3.8987 and 𝑅2 = 0.9990), thus, it will be applied
5 in the proceeding actions to predict methane adsorption profiles and parameter 𝛾 acquisition algorithm.
6 In the latter context, this model which has displayed the best performance is termed as the trained
7 FCNN model. The prediction performance of this trained FCNN model represented by the 𝑅2 graph is
8 shown in Fig. 8(a). One can see that the prediction results of the trained FCNN models highly match
9 the simulation results, resulting that data points are almost located near the diagonal of the graph. Since
10 the trained FCNN model has shown superior performance on the prediction of methane adsorption in
11 carbon slit nanopores, the physical analysis is attempted on the basis of its result.

12 Table 1. The evaluation results of 5-fold cross-validation method.

5-fold cross-validation 1 2 3 4 5
𝑅𝑀𝑆𝐸 3.8978 4.4661 4.0280 4.4493 4.4102
𝑅2 0.9990 0.9987 0.9989 0.9987 0.9987

13 Fig. 8(b)-(c) shows the predicted methane adsorption profiles by the trained FCNN model. They
14 illustrate the multilayer adsorption patterns of methane in carbon slit nanopores. Besides, the density
15 of the first peak and bulk phase decreases with the increasing temperature and declining pressure,
16 which is consistent with the previous conclusions proposed by a host of MD simulations [59,60].
17 Considering that the time spent for simulations is remarkably more than that for trained FCNN model,
18 the machine learning can save much time and computing resource in the field of adsorption research.
19 Meanwhile, the prediction results also prove the rationality to use the parameter 𝛾 to evaluate the
20 adsorption capacity of slit nanopores, as shown in Fig. 8(d), where the 𝛾 positively correlates to the
21 methane adsorption capacity. With the trained FCNN model, the methane adsorption profiles can be
22 quickly predicted by putting parameters 𝜔, 𝜉, 𝛾, 𝑇, 𝑃, and 𝑥 as the input layer data into the trained
23 FCNN model. Among the input parameters, 𝛾 is an important parameter because it directly reflects the
24 adsorption capability of slit nanopores. Considering the difficulty to directly calculate 𝛾 of realistic slit
25 nanopores through its definition, an algorithm is proposed in subsection 2.4, and its rationality and
26 accuracy are discussed in the next subsection.
14
1 3.3 Application for realistic slit nanopore
𝜔
2 Slit nanopore methane adsorption capability parameter 𝛾 is defined as 𝛾 = 𝑟𝑚𝑖𝑛. However, it is hard

3 to precisely calculate 𝛾 for the realistic slit nanopore through this definition because of the ambiguity
4 of its reference surface. Therefore, the parameter 𝛾 acquisition algorithm is proposed. To validate its
5 rationality, the realistic calcite [Inset picture in Fig. 9(a)] and kerogen [Inset picture in Fig. 9(d)] slit
6 nanopore models are constructed. The properties of these slit nanopores, as well as the methane
7 adsorption profiles are acquired by the simulation method. Then, the parameter 𝛾 acquisition and
8 application algorithm is utilized on the basis of the simulation results of the methane adsorption at the
9 350 K and 30 MPa condition to calculate the parameter 𝛾. After getting this critical parameter (𝛾0), the

10 trained FCNN model is applied to predict the methane adsorption profiles in these realistic slit
11 nanopores when the temperature and pressure of the system are set as 320, 350, 380 K, and 10, 30, 50
12 MPa, respectively.
13 Through the algorithm, the 𝛾0 of the calcite slit nanopore is calculated to be 0.2628. Meanwhile,

14 the comparison between the predicted and simulated adsorption profiles at 320 K and 10 MPa
15 condition is shown in Fig. 9(a). The results show great consistency, where even the slight second
16 adsorption peaks of the two profiles are coincident. For other temperature and pressure conditions, the
17 prediction results also illustrate perfect accuracy. As for all compared temperature and pressure
18 conditions, the 𝑅2 value of the comparison reaches 0.9681, indicating the high prediction accuracy.
19 For kerogen slit nanopores, the 𝛾0 is 0.2321, and the comparison between the prediction and simulation

20 results is illustrated in Fig. 9(d). The results also show superior consistency and the R2 value of the
21 comparison is 0.8808. Besides, the methane adsorption capacity of slit nanopore is a critical methane
22 adsorption behavior parameter, the methane adsorption capacity of simulation results and prediction
23 results are calculated. Comparing the simulation and prediction data, as shown in Fig. 9(b) and Fig.
24 9(e), the prediction results match the simulation results, resulting that data points are almost located
25 near the diagonal of the graph, and the compared R2 value based on calcite and kerogen slit nanopore
26 data, are 0.9025 and 0.9757, respectively. Results from the comparison undoubtedly prove the
27 rationality and the accuracy of the prediction method proposed in this work. On the basis of these
28 reliable prediction results, more analysis can be implemented.
29 Considering that the 𝛾0 of the calcite slit nanopore is larger than that of kerogen, the methane
15
1 adsorption ability for the calcite slit nanopore is also stronger than the latter. Comparing Fig. 9(a) and
2 Fig. 9(d), at the 320 K and 10 MPa condition, the density of the first adsorption peak for calcite is
3 nearly 600 kg/cm3, while that of kerogen is only about 300 Kg/cm3. However, the densities of the
4 second adsorption peak and bulk phase for calcite and kerogen are almost the same. This phenomenon
5 implies that for multilayer adsorption, the nanopore walls mainly influence the first adsorption peak,
6 which will finally lead to a different adsorption capacity. Besides, as shown in Fig. 9(c) and Fig. 9(f),
7 methane adsorptions in both calcite and kerogen nanopore slits obey the Langmuir adsorption:
𝑏𝑃
8 𝑁 = 𝑁𝑚𝑎𝑥1 + 𝑏𝑃 (13)

9 where, N and P are the adsorption quantity and the corresponding pressure, 𝑁𝑚𝑎𝑥 and b denote the

10 Langmuir maximum adsorption capacity and Langmuir constant. Through fitting, it is shown that the
11 𝑁𝑚𝑎𝑥 of the calcite is obviously higher than that of the kerogen, indicating the former can attract more
12 methane adsorption, which coincides with the conclusion drawn from the parameter 𝛾0. Besides, the

13 machine learning method significantly shortens the simulation and analysis time compared to when
14 machine learning method was not yet used. These findings indicate that with the aid of the machine
15 learning method, more robust conclusions about methane adsorption phenomenon, can be draw with
16 limited time and computing resource.
17 From the above results, the model proposed herein has been proved to be practical in predicting
18 the methane adsorption profiles and analyzing methane adsorption mechanism. The machine learning
19 model is trained based on the dataset consist of a series of uniformly defined carbon slit nanopores and
20 the simulation results acquired by GCMC/MD simulation methods. With the trained machine learning
21 model and the acquisition and application algorithm of parameter 𝛾, the methane adsorption profiles
22 of different realistic slit nanopores under different temperature and pressure conditions can be
23 predicted quickly and accurately. Besides, the model built from microscopic perspective in our work
24 is fundamental and physics informed. The trained machine learning model can be utilized to acquire
25 methane adsorption profiles quickly under almost all possible conditions. Meanwhile, the predicted
26 results can be post processed to acquire some key parameters in methane adsorption behavior. On this
27 way, the methane adsorption mechanism can be analyzed more comprehensively and effectively.

28 3.4 Prediction for different nanopores under various geological parameters

16
1 In the above sections, it has been proved that, the prediction capabilities of our proposed trained
2 FCNN model for methane adsorption, are both fast and precise, which prompts the research of the
3 adsorption mechanism from the micro-scale perspective. Hence, in this section, other possible
4 applications of this fast prediction method are introduced.
5 In regard to adsorption capability, the adsorption capability parameter, 𝛾, is important in
6 estimating the reservoir reserves, different adsorption capability parameters of shale nanopore lead to
7 different investigation strategy. As for two different slit nanopores with different adsorption capability
8 parameter 𝛾, the methane adsorption profiles are different under the same temperature and pressure
9 condition. The slit nanopore structure is fixed and cannot be changed to make the two slit nanopores
10 have the same adsorption behavior, while the temperature and pressure can be modified to change the
11 adsorption behavior in the slit nanopore. Thanks to the fast prediction ability of the trained FCNN
12 model, the adsorption profiles under different temperature and pressure conditions can be acquired
13 quickly. Then the predicted adsorption profiles can be used to contrast with the target adsorption
14 profile to obtain the best matched temperature and pressure condition. Besides, the information of
15 whether the adsorption behavior difference between two different slit nanopores can be offset by
16 modifying temperature and pressure, can be obtained through the contrast results. Two examples are
17 shown in Fig. 10(b) and Fig. 10(c). There are two different slit nanopores with the same 𝜔 and 𝜉 values,
18 which are 0.9708 and 0.1499, respectively. The two slit nanopores have different 𝛾 values, one is
19 0.2628 which is set as the target (Fig. 10(a)) and the other is 0.2555 which is set to modify its
20 temperature and pressure condition to match the target adsorption profile whose condition is 350 K
21 and 30 MPa. The trained FCNN model is utilized to traverse the temperature and pressure, then the
22 best matched condition is found at 321.8 K and 24.3 MPa, and the contrast graph is shown in Fig.
23 10(b) which shows a better match performance. It means the adsorption behavior difference can be
24 offset by modifying temperature and pressure under this condition. However, the adsorption behavior
25 difference cannot always be offset by modifying temperature and pressure, as shown in Fig. 10(c)
26 which shows a poor match performance. In this condition, the 𝛾 values of the two different slit
27 nanopores are 0.2628 and 0.2455, respectively.
28 With respect to the traverse ability and the rapidity of the trained machine learning model, this
29 study provides a new perspective to investigate the relationship between the methane adsorption
30 behavior and temperature as well as pressure. Adsorption peak value of the methane adsorption profile
17
1 is a key parameter to analyze the methane adsorption behavior, and this can only be analyzed on limited
2 conditions because of the limitations of the traditional methods. With the methods proposed in this
3 work, the methane adsorption behavior of shale slit nanopore can be analyzed in a continuous
4 environment space. For example, as shown in Fig. 10(d), the adsorption peak values of calcite slit
5 nanopore under different temperature and pressure conditions are captured in the graph, herein,
6 researchers can have an intuitive understanding on how temperature and pressure influence the value
7 of the methane adsorption peaks in calcite slit nanopore. This can be extended to many other aspects
8 to help researchers analyze the methane adsorption behavior in shale nanopores quickly and accurately.
9 Noting that, the model in this work is built based on a series of uniformly defined slit nanopore
10 models that are carbon slit nanopores with different roughness for simulation, training, prediction and
11 analysis. Therefore, the prediction of the adsorption behavior performs well for the similar nanopores,
12 and as for other walls with different potential energy distribution, the corresponding methane
13 adsorption profiles could be obtained by expanding the variety of basic slit nanopore models. For
14 example, choosing different atom type, lattice structure and wall roughness for the basic slit nanopore
15 models. In this context, the results predicted by the machine learning model can be adapted to different
16 analysis requirements to some extent, further expanding the practicality of current model.

17 4. Conclusion

18 Methane adsorption behavior is critical for commercial and effective shale gas recovery. In this
19 work, a framework based on machine learning has been constructed to help researchers better
20 understand the methane adsorption behavior in slit nanopores and quickly predict methane adsorption
21 profiles. The main conclusion could be summarized as follows:
22 1. Proposing and utilizing the average minimum potential energy 𝜔, the minimum potential
23 energy standard deviation 𝜉, and the slit nanopore methane adsorption capability parameter 𝛾 to
24 represent the adsorption characteristics of the slit nanopore structure considering kinetic theory;
25 2. Predicting methane adsorption profile at different temperatures and pressures in different slit
26 nanopores quickly and accurately based on the machine learning method;
27 3. Acquiring inorganic and organic slit nanopore methane adsorption capability parameter 𝛾,
28 predicting the adsorption profiles in different temperature and pressure conditions based on the
29 parameter 𝛾 acquisition and application algorithm and analyzing the methane adsorption behavior
18
1 combined with prediction results;
2 4. Providing novel perspectives on investigating the methane adsorption behavior in shale
3 nanopores through the methods proposed in this work.
4 With these unparalleled progresses, the methane adsorption behavior in shale nanopores can be
5 analyzed through new approaches more efficiently and comprehensively. However, just like most
6 machine learning applications, there are a few shortcomings in the analysis process, for a slit nanopore
7 whose methane adsorption characteristic parameter 𝜔 and 𝜉 are out of range of the machine learning
8 dataset, the prediction performance is not satisfactory. Such a shortcoming can be resolved by
9 extending the machine learning dataset or using semi-supervised machine learning, for example, we
10 can build more different carbon slit nanopores with different structures to extend the 𝜔 and 𝜉 ranges
11 of the dataset. Moreover, in actual adsorption process of shale, there exists some other gases, therefore,
12 more complex slit nanopore structures and vast amounts of gases will be considered in our future
13 works to further analyze the gas adsorption behavior.

14 Acknowledgements
15 This work is jointly supported by the National Key Research and Development Program of China
16 (2019YFA0708700), the National Postdoctoral Program for Innovative Talents (BX2021285), and the
17 Fundamental Research Funds for the Central Universities (WK2090000042). The numerical
18 calculations have been done on the supercomputing system in the Supercomputing Center of
19 University of Science and Technology of China.

20 Captions of figures
21 Fig. 1. Workflow of the rapid prediction process of the methane adsorption behavior.

22 Fig. 2. (a)-(f) the carbon slit nanopore with roughness 0, 0.5, 1.0, 1.5, 2.0, and 2.5, respectively; (g)
23 calcite slit nanopore; (h) kerogen slit nanopore. The colored balls denote different elements: gray
24 (carbon), white (hydrogen), red (oxygen), blue (nitrogen), yellow (sulfur) and green (calcium).

25 Fig. 3. The methane and carbon slit nanopore model, gray (carbon) and yellow (methane molecule);
26 the inter-molecular force graph, blue line: L-J potential profile, orange line: interaction force profile;
27 simulation results graph: methane adsorption profiles in different conditions.

28 Fig. 4. (a-c) the minimum potential energy distribution in carbon slit nanopore with roughness 0, 1.0,
19
1 and 2.0, respectively; (d-f) the potential-minimum distance distribution in carbon slit nanopore with
2 roughness 0, 1.0, and 2.0, respectively.

3 Fig. 5. Input parameters, FCNN model architecture and prediction adsorption profiles. The FCNN
4 model was implemented using Keras [61] deep learning library with the TensorFlow [62] backend.

5 Fig. 6. Acquisition and application algorithm of parameter 𝛾.

6 Fig. 7. (a): the average minimum potential energy 𝜀, minimum potential energy standard deviation 𝜎,
7 and methane adsorption capability parameter 𝛾 of six carbon slit nanopores with different roughness;
8 (b): the methane adsorption profiles under the condition of 300 K and 60 MPa in carbon slit nanopore
9 with roughness of 0.0, 0.5, 1.0, 1.5, 2.0, and 2.5, respectively.

10 Fig. 8. (a): The R-squared graph; (b): the predicted profiles under different temperature; (c): the
11 predicted profiles under different pressure; (d): the predicted methane adsorption capacity under
12 different slit nanopore methane adsorption capability parameter 𝛾.

13 Fig. 9. (a, d): the comparison of methane adsorption profile between simulation results and prediction
14 results of calcite and kerogen slit nanopore, respectively; (b, e): the R2 graph of the methane adsorption
15 capacity of the simulation results and prediction results of calcite and kerogen slit nanopore,
16 respectively; (c, f): the fitted Langmuir curves at different temperature based on the prediction data of
17 calcite and kerogen slit nanopore, respectively.
18 Fig. 10. (a): the reference adsorption profile; (b, c): well-matched and ill-matched examples of
19 modified temperature and pressure to offset the difference due to the change of parameter 𝛾; (d): the
20 contour profile of the adsorption peak value under various temperature and pressure conditions

21 References
22 [1] D.W. Keith, Why capture CO2 from the atmosphere?, Science 325 (2009) 1654–1655.
23 [2] J.D. Hughes, A reality check on the shale revolution, Nature 494 (2013) 7–9.
24 [3] G. Guerrier, E. D’Ortenzio, Shale gas is a fraught solution to emissions, Nature 513 (2014) 315.
25 [4] A. Busch, Y. Gensterblum, B.M. Krooss, R. Littke, Methane and carbon dioxide adsorption-
26 diffusion experiments on coal: Upscaling and modeling, Int. J. Coal Geol. 60 (2004) 151–168.
27 [5] R. a. Kerr, Natural Gas From Shale Bursts Onto the Scene, Science 328 (2010) 1624–1626.
28 [6] R.W. Howarth, A. Ingraffea, T. Engelder, Should fracking stop?, Nature 477 (2011) 271–275.
20
1 [7] A. Yethiraj, A. Striolo, Fracking: What can physical chemistry offer?, J. Phys. Chem. Lett. 4
2 (2013) 687–690.
3 [8] S.R. Etminan, F. Javadpour, B.B. Maini, Z. Chen, Measurement of gas storage processes in
4 shale and of the molecular diffusion coefficient in kerogen, Int. J. Coal Geol. 123 (2014) 10–
5 19.
6 [9] C.M. Freeman, G. Moridis, D. Ilk, T.A. Blasingame, A numerical study of performance for tight
7 gas and shale gas reservoir systems, J. Pet. Sci. Eng. 108 (2013) 22–39.
8 [10] T. Wang, S. Tian, G. Li, L. Zhang, M. Sheng, W. Ren, Molecular simulation of gas adsorption
9 in shale nanopores: A critical review, Renew. Sustain. Energy Rev. 149 (2021) 111391.
10 [11] H. Aljamaan, M. Al Ismail, A.R. Kovscek, Experimental investigation and Grand Canonical
11 Monte Carlo simulation of gas shale adsorption from the macro to the nano scale, J. Nat. Gas
12 Sci. Eng. 48 (2017) 119–137.
13 [12] T. Zhang, G.S. Ellis, S.C. Ruppel, K. Milliken, R. Yang, Effect of organic-matter type and
14 thermal maturity on methane adsorption in shale-gas systems, Org. Geochem. 47 (2012) 120–
15 131.
16 [13] P. Chareonsuppanimit, S.A. Mohammad, R.L. Robinson, K.A.M. Gasem, High-pressure
17 adsorption of gases on shales: Measurements and modeling, Int. J. Coal Geol. 95 (2012) 34–46.
18 [14] W. Pang, Y. Wang, Z. Jin, Comprehensive review about methane adsorption in shale
19 nanoporous media, Energy Fuels 35 (2021) 8456–8493.
20 [15] K. Lin, Q. Yuan, Y.P. Zhao, Using graphene to simplify the adsorption of methane on shale in
21 MD simulations, Comput. Mater. Sci. 133 (2017) 99–107.
22 [16] H. Yu, Y. Zhu, X. Jin, H. Liu, H. Wu, Multiscale simulations of shale gas transport in micro /
23 nano-porous shale matrix considering pore structure influence, J. Nat. Gas Sci. Eng. 64 (2019)
24 28–40.
25 [17] H. Yu, J. Fan, J. Xia, H. Liu, H. Wu, Multiscale gas transport behavior in heterogeneous shale
26 matrix consisting of organic and inorganic nanopores, J. Nat. Gas Sci. Eng. 75 (2020) 103139.
27 [18] S. Wang, Q. Feng, M. Zha, F. Javadpour, Q. Hu, Supercritical methane diffusion in shale
28 nanopores: effects of pressure, mineral types, and moisture content, Energy Fuels 32 (2018)
29 169–180.
30 [19] S. Ravipati, M.S. Santos, I.G. Economou, A. Galindo, G. Jackson, A.J. Haslam, Monte Carlo
21
1 molecular simulation study of carbon dioxide sequestration into dry and wet calcite pores
2 containing methane, Energy Fuels. (2021).
3 [20] H. Yu, J. Fan, J. Chen, Y. Zhu, H. Wu, Pressure-dependent transport characteristic of methane
4 gas in slit nanopores, Int. J. Heat Mass Transf. 123 (2018) 657–667.
5 [21] J. Zhang, M.B. Clennell, K. Liu, M. Pervukhina, G. Chen, D.N. Dewhurst, Methane and carbon
6 dioxide adsorption on illite, Energy Fuels 30 (2016) 10643–10652.
7 [22] W. Zhou, H. Wang, Y. Yan, X. Liu, Adsorption mechanism of co2/ch4 in kaolinite clay: insight
8 from molecular simulation, Energy Fuels 33 (2019) 6542–6551.
9 [23] S. Wang, Q. Feng, F. Javadpour, Q. Hu, K. Wu, Competitive adsorption of methane and ethane
10 in montmorillonite nanopores of shale at supercritical conditions: A grand canonical Monte
11 Carlo simulation study, Chem. Eng. J. 355 (2019) 76–90.
12 [24] W. Song, J. Yao, J. Ma, A. Li, Y. Li, H. Sun, L. Zhang, Grand canonical Monte Carlo
13 simulations of pore structure influence on methane adsorption in micro-porous carbons with
14 applications to coal and shale systems, Fuel 215 (2018) 196–203.
15 [25] K. Mosher, J. He, Y. Liu, E. Rupp, J. Wilcox, Molecular simulation of methane adsorption in
16 micro- and mesoporous carbons with applications to coal and gas shale systems, Int. J. Coal
17 Geol. 109–110 (2013) 36–44.
18 [26] T. Zhang, Y. He, Y. Yang, K. Wu, Molecular simulation of shale gas adsorption in organic-
19 matter nanopore, J. Nat. Gas Geosci. 2 (2017) 323–332.
20 [27] P. Psarras, R. Holmes, V. Vishal, J. Wilcox, Methane and co2 adsorption capacities of kerogen
21 in the eagle ford shale from molecular simulation, Acc. Chem. Res. 50 (2017) 1818–1828.
22 [28] Y. Liu, J. Wilcox, Molecular simulation studies of CO2 adsorption by carbon model compounds
23 for carbon capture and sequestration applications, Environ. Sci. Technol. 47 (2013) 95–101.
24 [29] K. Falk, B. Coasne, R. Pellenq, F.J. Ulm, L. Bocquet, Subcontinuum mass transport of
25 condensed hydrocarbons in nanoporous media, Nat. Commun. 6 (2015) 1–7.
26 [30] P. Billemont, B. Coasne, G. De Weireld, Adsorption of carbon dioxide, methane, and their
27 mixtures in porous carbons: Effect of surface chemistry, water content, and pore disorder,
28 Langmuir. 29 (2013) 3328–3338.
29 [31] K. Falk, R. Pellenq, F.J. Ulm, B. Coasne, Effect of chain length and pore accessibility on alkane
30 adsorption in kerogen, Energy Fuels 29 (2015) 7889–7896.
22
1 [32] J. Schmidt, M.R.G. Marques, S. Botti, M.A.L. Marques, Recent advances and applications of
2 machine learning in solid-state materials science, Npj Comput. Mater. 5 (2019).
3 [33] D. Kang, X. Wang, X. Zheng, Y.P. Zhao, Predicting the components and types of kerogen in
4 shale by combining machine learning with NMR spectra, Fuel 290 (2021) 120006.
5 [34] S. Wang, C. Qin, Q. Feng, F. Javadpour, Z. Rui, A framework for predicting the production
6 performance of unconventional resources using deep learning, Appl. Energy 295 (2021) 117016.
7 [35] M. Meng, R. Zhong, Z. Wei, Prediction of methane adsorption in shale: Classical models and
8 machine learning based models, Fuel 278 (2020) 118358.
9 [36] G.S. Fanourgakis, K. Gkagkas, E. Tylianakis, G. Froudakis, A generic machine learning
10 algorithm for the prediction of gas adsorption in nanoporous materials, J. Phys. Chem. C. 124
11 (2020) 7117–7126.
12 [37] H.B. Sharma, S. Panigrahi, A.K. Sarmah, B.K. Dubey, Modeling of methane adsorption
13 capacity in shale gas formations using white-box supervised machine learning techniques, J.
14 Petrol. Sci. Eng. (2021) 135907.
15 [38] H. Ghasemzadeh, S. Babaei, S. Tesson, J. Azamat, M. Ostadhassan, From excess to absolute
16 adsorption isotherm: The effect of the adsorbed density, Chem. Eng. J. 425 (2021).
17 [39] C. Methods, A. Mech, X. Yu, J. Li, Z. Chen, K. Wu, L. Zhang, Effects of an adsorbent accessible
18 volume on methane adsorption on shale, Comput. Methods Appl. Mech. Eng. 370 (2020)
19 113222.
20 [40] W. Zhang, B. Shen, Y. Chen, T. Wang, W. Chen, Molecular dynamics simulations about isotope
21 fractionation of methane in shale nanopores, Fuel 278 (2020) 118378.
22 [41] H. Yu, H. Xu, J. Xia, J. Fan, F. Wang, H. Wu, Nanoconfined transport characteristic of methane
23 in organic shale nanopores: The applicability of the continuous model, Energy Fuels 34 (2020)
24 9552–9562.
25 [42] K. Lin, Y.P. Zhao, Entropy and enthalpy changes during adsorption and displacement of shale
26 gas, Energy 221 (2021) 119854.
27 [43] H. Yu, H. Xu, J. Fan, F. Wang, H. Wu, Roughness factor-dependent transport characteristic of
28 shale gas through amorphous kerogen nanopores, J. Phys. Chem. C. 124 (2020) 12752–12765.
29 [44] H. Yu, H. Xu, J. Fan, Y.B. Zhu, F. Wang, H. Wu, Transport of Shale Gas in
30 Microporous/Nanoporous Media: Molecular to Pore-Scale Simulations, Energy Fuels 35 (2021)
23
1 911–943.
2 [45] P. Ungerer, J. Collell, M. Yiannourakou, Molecular modeling of the volumetric and
3 thermodynamic properties of kerogen: Influence of organic type and maturity, Energy Fuels 29
4 (2015) 91–105.
5 [46] S.R. Kelemen, M. Afeworki, M.L. Gorbaty, M. Sansone, P.J. Kwiatek, C.C. Walters, H. Freund,
6 M. Siskin, A.E. Bence, D.J. Curry, M. Solum, R.J. Pugmire, M. Vandenbroucke, M. Leblond,
7 F. Behar, Direct characterization of kerogen by X-ray and solid-state 13C nuclear magnetic
8 resonance methods, Energy Fuels 21 (2007) 1548–1561.
9 [47] D. Frenkel, B. Smit, Understanding molecular simulation: from algorithms to applications,
10 Elsevier (2002).
11 [48] J. Chen, H. Yu, J. Fan, F. Wang, D. Lu, H. Liu, H. Wu, Channel-width dependent pressure-
12 driven flow characteristics of shale gas in nanopores, AIP Adv. 7 (2017) 045217.
13 [49] H. Yu, J. Chen, Y.B. Zhu, F.C. Wang, H.A. Wu, Multiscale transport mechanism of shale gas
14 in micro/nano-pores, Int. J. Heat Mass Transf. 111 (2017) 1172–1180.
15 [50] L. Huang, Z. Ning, Q. Wang, W. Zhang, Z. Cheng, X. Wu, H. Qin, Effect of organic type and
16 moisture on CO2/CH4 competitive adsorption in kerogen with implications for CO2
17 sequestration and enhanced CH4 recovery, Appl. Energy 210 (2018) 28–43.
18 [51] J. Zhou, Z. Jin, K.H. Luo, Insights into recovery of multi-component shale gas by CO2 injection:
19 A molecular perspective, Fuel 267 (2020) 1–29.
20 [52] R.T. Cygan, J.J. Liang, A.G. Kalinichev, Molecular models of hydroxide, oxyhydroxide, and
21 clay phases and the development of a general force field, J. Phys. Chem. B. 108 (2004) 1255–
22 1266.
23 [53] C.I. Bayly, K.M. Merz, D.M. Ferguson, W.D. Cornell, T. Fox, J.W. Caldwell, P.A. Kollman,
24 P. Cieplak, I.R. Gould, D.C. Spellmeyer, A second generation force field for the simulation of
25 proteins, nucleic acids, and organic molecules, J. Am. Chem. Soc. 117 (1995) 5179–5197.
26 [54] J. Schmidhuber, Deep Learning in neural networks: An overview, Neural Networks 61 (2015)
27 85–117.
28 [55] Y. Lecun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (2015) 436–444.
29 [56] M. Meng, Z. Qiu, R. Zhong, Z. Liu, Y. Liu, P. Chen, Adsorption characteristics of supercritical
30 CO2/CH4 on different types of coal and a machine learning approach, Chem. Eng. J. 368 (2019)
24
1 847–864.
2 [57] M.J. Brown, L.A. Hutchinson, M.J. Rainbow, K.J. Deluzio, A.R. De Asha, Rectified linear units
3 improve restricted boltzmann machines, Int. Conf. Mach. Learn. 33 (2017) 384–387.
4 [58] D.P. Kingma, J. Lei Ba, ADAM: A method for stochastic optimization, ICLR (2015) 1–15.
5 [59] Y. Pang, X. Hu, S. Wang, S. Chen, M.Y. Soliman, H. Deng, Characterization of adsorption
6 isotherm and density profile in cylindrical nanopores: Modeling and measurement, Chem. Eng.
7 J. 396 (2020) 125212.
8 [60] F. Xiong, G. Rother, D. Tomasko, W. Pang, J. Moortgat, On the pressure and temperature
9 dependence of adsorption densities and other thermodynamic properties in gas shales, Chem.
10 Eng. J. 395 (2020) 124989.
11 [61] F. Chollet et al., Keras, http://keras.io (2015).
12 [62] M. M. Abadi et al., TensorFlow: large-scale machine learning on heterogeneous distributed
13 systems, http://www.tensorflow.org/ (2015).
14 [63] Kennard, Earle H, Kinetic theory of gases, Vol. 483. New York: McGraw-hill (1938).
15 [64] M. Micheal, W.L. Xu, H.Y. Xu, J.N. Zhang, H.J. Jin, H. Yu, H.A. Wu, Multi-scale modelling
16 of gas transport and production evaluation in shale reservoir considering crisscrossing fractures,
17 J. Nat. Gas Sci. Eng. 95 (2021) 104156.
18 [65] L. Chen, L. Zuo, Z. Jiang, S. Jiang, K. Liu, J. Tan, L. Zhang, Mechanisms of shale gas
19 adsorption: Evidence from thermodynamics and kinetics study of methane adsorption on shale,
20 Chem. Eng. J. 361 (2019) 559-570.
21 [66] R. Hu, W. Wang, J. Tan, L. Chen, J. Dick, G. He, Mechanisms of shale gas adsorption: Insights
22 from a comparative study on a thermodynamic investigation of microfossil-rich shale and non-
23 microfossil shale, Chem. Eng. J. 411 (2021) 128463.
24 [67] Y. Tian, C. Yan, Z. Jin, Characterization of methane excess and absolute adsorption in various
25 clay nanopores from molecular simulation, Scientific reports. 7 (2017) 12040.
26 [68] S. Wang, Y. Liang, Q. Feng, F. Javadpour, Sticky layers affect oil transport through the
27 nanopores of realistic shale kerogen, Fuel. 310(2022) 122480.
28 [69] S. Wang, X. Yao, Q. Feng, F. Javadpour, Y. Yang, Q. Xue, X. Li, Molecular insights into carbon
29 dioxide enhanced multi-component shale gas recovery and its sequestration in realistic kerogen,
30 Chem. Eng. J. 425 (2022) 130292.
25
1 [70] L. Huang, Z.F. Ning, Q. Wang, R.R. Qi, Molecular simulation of adsorption behaviors of
2 methane, carbon dioxide and their mixtures on kerogen: Effect of kerogen maturity and moisture
3 content, Fuel. 211 (2018) 159-172.
4

5 The authors declare that they have no known competing financial interests or personal relationships
6 that could have appeared to influence the work reported in this paper.
7

8 Highlights
9  A machine learning (ML) framework to predict the methane adsorption is constructed.
10  Three novel parameters related to potential energy distribution (PED) are proposed.
11  Machine learning algorithm based on the uniformly constructed dataset is introduced.
12  The prediction performance is well validated by typical inorganic and organic models.
13  The application in different geological conditions (e.g., pressure and temperature) is performed.
14

15 Supplementary Material for


16 Fast prediction of methane adsorption in shale nanopores using kinetic
17 theory and machine learning algorithm
18 MengCheng Huang#, HengYu Xu#, Hao Yu*, HouLin Zhang, Marembo Micheal, XinHeng Yuan and HengAn
19 Wu*
20 CAS Key Laboratory of Mechanical Behavior and Design of Materials, Department of Modern Mechanics,
21 University of Science and Technology of China, Hefei 230027, China
22 #These authors contributed equally to this work.
23 *Corresponding email: yuhaoo@ustc.edu.cn (Hao Yu); wuha@ustc.edu.cn (HengAn Wu)

24

26
1 Construction of the molecular model
2 For the construction of carbon slit nanopores, each carbon plate is composed of face centered cubic (FCC)
3 carbon crystal with a lattice constant of 0.3 nm, and its dimension is 4.2 nm × 8.1 nm × 1.2 nm. After generating
4 these carbon plates, they are fixed as rigid and placed in the top and bottom of the simulation box to form the 4 nm
5 slit nanopores. Meanwhile, the roughness of the slit nanopore is introduced by rearranging the coordinates of atoms
6 located on the surface of the plate [1, 2]. Half of these atoms are randomly selected and moved along the direction
7 perpendicular to the plate. The maximum moving distance is set as 0, 0.5, 1.0, 1.5, 2.0, and 2.5 Å, respectively, to
8 create slit nanopores with different roughness. The random moving distance satisfies the normal distribution. Every
9 plate is constructed independently to ensure their uniqueness, and the final configurations of the carbon slit nanopores
10 are shown in Fig. 2(a-f).
11 For the construction of calcite slit nanopore, the standard crystal structure of calcite is collected from Yu et al.'s
12 work [3]. The unit cell of calcite is supercelled to create the plate with the size of 4 nm × 8 nm × 2 nm. Two pieces
13 of these plates are placed in parallel to form the 4 nm inorganic slit nanopore as shown in Fig. 2(g). For the
14 construction of kerogen slit nanopores, Ungerer et al. [4] proposed six kerogen molecule structures, and based on
15 their proposal, type II-C kerogen monomer is selected for this work. Fourteen kerogen monomers are randomly
16 placed in the simulation box, and then they are relaxed in isopiestic-isothermal (NPT) ensemble with the pressure of
17 20 MPa. During this process, the simulation box gradually deforms to create an approximate kerogen plate with
18 dimension 4 nm × 8 nm × 2 nm. Two pieces of the kerogen plates are independently created. Their density is 1.2
19 g/cm3, which is consistent with the experimental detection result [5]. The 4 nm organic nanopore is constructed by
20 placing the two pieces of kerogen plates in parallel, as shown in Fig. 2(h).

21

22

23

24 GCMC/MD simulations

25 The GCMC process is performed in a canonical (NVT) ensemble to achieve the grand canonical (μVT) ensemble,
26 in which, the chemical potential, volume, and temperature of the system are constants. This process continues 1 ns
27 to reach the equilibrium. Afterward, the molecular dynamics (MD) simulation is performed. To obtain a stable
28 adsorption state of methane, the system is ran in the NVT ensemble for another 10 ns, and the density distribution of
29 methane in the latter 5 ns of this period is collected. To maintain consistency with previous studies [6], in this work,
30 the methane molecules are regarded as single particles. The time step is 1 fs, and the three boundaries of the
31 simulation box are periodic. The temperature and pressure of the simulation system are controlled by the Nosé-

27
1 Hoover thermostat. The Large Scale Atomic/Molecular Massively Parallel Simulator (LAMMPS) software is utilized
2 to implement the simulations mentioned above.

10

11

12

13

14

15

16

17

28
1 Validation of numerical model

2
3 Fig. S1. The validation of MD simulations. The density (a) and viscosity (b) of methane in slit nanopore through MD
4 simulations are compared with the experiment results from NIST (National Institute of Standards and Technology.
5 Thermophysical Properties of Fluid Systems, 2011, http://webbook.nist.gov/chemistry/fluid/), which indicates the
6 accuracy of the molecular modeling to mimic the gas adsorption behavior in shale nanopores.

10

11

12

13

14

15

16

17

18

19

20
29
1 Cost curves of machine learning model

3 Fig. S2. The cost curves of the 5-fold cross-validation method during the training process. (a)-(e) correspond to fold
4 one to fold five, respectively. As shown in the figure, the loss value drops fast in both training dataset and validation
5 dataset, which means the training process is accurate and effective.

10

11

12

13

14

15

16
30
1

31
1 References
2 [1] H. Yu, H. Xu, J. Fan, F. Wang, H. Wu, Roughness factor-dependent transport characteristic of
3 shale gas through amorphous kerogen nanopores, J. Phys. Chem. C. 124 (2020) 12752–12765.
4 [2] S. Wang, Y. Liang, Q. Feng, F. Javadpour, Sticky layers affect oil transport through the nanopores
5 of realistic shale kerogen, Fuel. 310 (2022) 122480.
6 [3] H. Yu, H. Xu, J. Fan, Y.B. Zhu, F. Wang, H. Wu, Transport of Shale Gas in
7 Microporous/Nanoporous Media: Molecular to Pore-Scale Simulations, Energy Fuels 35 (2021) 911–
8 943.
9 [4] P. Ungerer, J. Collell, M. Yiannourakou, Molecular modeling of the volumetric and
10 thermodynamic properties of kerogen: Influence of organic type and maturity, Energy Fuels 29 (2015)
11 91–105.
12 [5] S.R. Kelemen, M. Afeworki, M.L. Gorbaty, M. Sansone, P.J. Kwiatek, C.C. Walters, H. Freund,
13 M. Siskin, A.E. Bence, D.J. Curry, M. Solum, R.J. Pugmire, M. Vandenbroucke, M. Leblond, F. Behar,
14 Direct characterization of kerogen by X-ray and solid-state 13C nuclear magnetic resonance methods,
15 Energy Fuels 21 (2007) 1548–1561.
16 [6] J. Chen, H. Yu, J. Fan, F. Wang, D. Lu, H. Liu, H. Wu, Channel-width dependent pressure-driven
17 flow characteristics of shale gas in nanopores, AIP Adv. 7 (2017) 045217.
18
19
20

32
1

33
1

34
1

35
1

36
1

37

You might also like