Fluid Phase Equilibria 558 (2022) 113423

Contents lists available at ScienceDirect

Fluid Phase Equilibria

journal homepage: www.elsevier.com/locate/fluid

A generalized machine learning-assisted phase-equilibrium calculation

model for shale reservoirs
Fangxuan Chen a, Sheng Luo b, Shihao Wang b, Hadi Nasrabadi a,∗
a
Harold Vance Department of Petroleum Engineering, Texas A&M University, United States
b
Chevron Company United States

a r t i c l e i n f o a b s t r a c t

Article history: In compositional reservoir simulation, a significant portion of the CPU time is consumed in phase equilib-
Received 4 November 2021 rium calculations. Previous studies have incorporated the machine learning (ML) technique to accelerate
Revised 14 February 2022
and stabilize the phase equilibrium calculations. However, there are two main limitations: 1) previous
Accepted 20 February 2022
work mainly focuses on conventional reservoirs, which cannot be extended to unconventional reservoirs;
Available online 2 March 2022
2) previous studies are limited to fluid compositions with specific hydrocarbon components that narrows
Keywords: their application. In this paper, we propose a novel ML-assisted framework for phase equilibrium calcu-
Proxy phase-equilibrium calculator lations in shale reservoirs. A general set of pseudo-components is considered to allow users to customize
Shale reservoirs the composition of hydrocarbon mixtures. A pore size-dependent EOS is applied to simulate the fluid
Multi-layer perceptron phase behavior in nano-scale conditions. In the stability test, the multilayer perceptron (MLP) is trained
Physics-informed neural network to predict the fluid phase state: single-phase or two-phase. For the fluid labeled as two-phase condi-
Machine learning
tion, the phase-split computation is performed to obtain the equilibrium ratio. Instead of using the initial
estimate from the stability test, the MLP and the physics-informed neural network (PINN) are applied
to obtain the initial estimates for the minimization program. The results show that, with the assistance
of ML technique, we are able to reduce the computation time needed for the nano-scale phase equi-
librium calculations by more than two orders of magnitude while maintaining 97% accuracy. Compared
with MLP, PINN can accurately predict the equilibrium ratios with a limited range of input variables but
require more training time. The progress of this study present a ML-assisted framework for phase equilib-
rium calculations and the generalized proxy phase-equilibrium calculator can be compiled into reservoir
simulator to accelerate flash calculation.
© 2022 Elsevier B.V. All rights reserved.

1. Introduction 1.1. Phase behavior investigation methods

The widespread application of horizontal wells and multi-stage Methods commonly used to investigate the phase behavior
hydraulic fracturing has facilitated the development of shale reser- of shale reservoirs include experimental studies, molecular sim-
voirs [1,2]. One important difference between conventional and ulation, and equation of state (EOS) modeling [5]. The experi-
shale reservoirs is their pore-size distribution. Distinguished from mental methods for phase-behavior analysis include adsorption-
conventional reservoirs, which are dominated by macro-scale pores desorption experiments [6–8], differential scanning calorimetry
(d > 50nm ), shale reservoirs have a large number of mesopores (DSC) techniques [9–13], diffusion experiments [14], and lab-on-a-
(2nm < d < 50nm ) and micropores (d < 2nm ) [3]. The volume of chip technology [15–17]. Unlike experimental methods that inves-
mesopores and micropores can account for 40% of the total pore tigate phase behavior by measurement and observation, molecular
volume in shale reservoirs [4]. Due to the confinement effect, fluid- simulation provides an atomic-level perspective to reveal the po-
phase behavior in mesopores and micropores deviates significantly sitions of fluid molecules [18–21]. Although experimental methods
from bulk fluidphase behavior. As a result, it is necessary to find a and molecular simulation can provide relatively accurate analyses
fast and accurate way to describe the fluid-phase behavior of shale of phase behavior, the two methods are laborious and time con-
reservoirs. suming. Equation of state (EOS) modeling is efficient and can be
applied to phase-equilibrium calculations in reservoir simulations.
∗
The PengRobinson equation of state [22] (PR EOS) is a commonly
Corresponding author.
used EOS in petroleum engineering. However, it becomes inaccu-
E-mail address: [email protected] (H. Nasrabadi).

https://doi.org/10.1016/j.fluid.2022.113423
0378-3812/© 2022 Elsevier B.V. All rights reserved.
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

scale conditions [30–32]; and 4) using the generalized van der

Nomenclature Waals theory [33,34] to extend the application of PR EOS [35,36].
Although the modified PR EOS shows a higher accuracy, the re-
ANN Artificial neural network quired CPU time is increased due to the complexity of the mod-
BP Backpropagation ified PR EOS, particularly with the extended PR EOS (PR-C EOS)
DSC Differential scanning calorimetry [36–38]. A significant portion of the total CPU time in a compo-
EOS Equation of state sitional reservoir simulator is spent on phase-equilibrium calcula-
LSSVM Least-squares support vector machine tions [39], which need to be calculated repeatedly. As a result, the
LSTM Long short-term memory acceleration of phase-equilibrium calculations can significantly re-
MAE Mean absolute error duce the CPU time needed in reservoir simulations.
ML Machine learning
MLP Multi-layer perceptron 1.2. Machine-learning application in vapor-liquid equilibria studies
MSE Mean square error
ODE Ordinary differential equation In recent years, machine learning (ML) techniques have been
PAH Polycyclic aromatic hydrocarbon utilized in petroleum engineering to facilitate production forecast-
PCNN Physics-constrained neural network ing [40–43], parameter optimization [44–47], compositional sim-
PGNN Physics-guided neural network ulation [39,48–50], and image recognition [51–53]. Specifically,
PINN Physics-informed neural network many researchers have applied ML techniques in the stabiliza-
PR EOS Peng-Robinson equation of state tion and acceleration of vapor-liquid equilibria (VLE) calculations
PR-C EOS Peng Robinson-confined equation of state [54–57]. Gaganis and Varotsis [39,58] used ML to speed up the
ReLU Rectified linear unit compositional reservoir simulation. To determine phase stability,
RVM Relevance vector machine the Support Vector Machines (SVM) [59] algorithm was used.
SVM Support vector machine Rather than collecting data points uniformly, data points around
VLE Vapor-liquid equilibrium the phase boundary were more likely to be selected to train the
A Helmholtz free energy ML model. Data points with temperatures higher than criconden-
ap Confinement-modified energy parameter therm or pressures higher than cricondenbar were ignored, as fluid
BIP Binary interaction parameter under those conditions can be directly determined. For phase-split
bp Confinement-modified volume parameter computation, Gaganis and Varotsis utilized a feed-forward artificial
d Diameter neural network (ANN) to estimate the equilibrium ratio, which is
Econ f Configurational energy an important coefficient in the RachfordRice equation [60].
K Equilibrium ratio Kashinath et al. [61] improved upon Gaganis and Varotsiss work
k Number of folds in cross validation by applying the Relevance Vector Machine (RVM) [62] to pre-
kB Boltzmann constant dict fluid-phase stability. Compared to SVM, RVM set a discrim-
L Loss function inant threshold for the classification results. The predicted result
MW Molecular weight was accepted only when the posterior probability was higher than
N Number of molecules the threshold, therefore reducing the misclassification rate. Groven
Nc Coordination number et al. [49] used ANN to predict the VLE properties of polycyclic
n Number of moles aromatic hydrocarbons (PAHs). The predicted results were consis-
nc Number of components tent with the experimental results. Mesbah et al. [63] applied the
ns Number of samples Least-Squares Support Vector Machine (LSSVM) [64] algorithm to
P Pressure predict the saturation pressure of a (CO2 )-hydrocarbon binary mix-
Pc Critical pressure ture. The predicted results were in good agreement with the re-
q Intramolecular contribution sults calculated by EOS. Wang et al. [65] proposed a proxy phasee-
Q Canonical partition function quilibrium calculator that could predict the phase condition for a
rp Pore radius fluid system and provide the initial guess for phase-split compu-
T Temperature tation if the fluid system was in a two-phase condition. Different
Tc Critical temperature from the previous work, they incorporated the effect of capillary
V Volume pressure into the proxy phaseequilibrium calculator to better pre-
Vf Free volume dict the fluid-phase behavior in shale reservoirs.
w Acentric factor Based on our analysis of the previous work on ML-assisted VLE
V SP Volume shift parameter calculation, there are two main limits. First, most of the proposed
λ Translational contribution ML models are only applicable to conventional reservoirs. Although
δp Square well width Wang et al. [65] incorporated the effect of capillary pressure in
εp Square well depth their VLE calculations, this method is not accurate when the pore
κ Normalized fluid-pore surface affinity factor diameter is less than 10 nm [23]. In their work, the pore diameter
μ Chemical potential is set to a range of 60–200 nm, which does not cover the major-
ρmax,i Packing density ity of pore sizes in shale reservoirs [66]. Second, the previous ML
σ Effective parameter models were trained by a specific set of components, which means
the models are only accurate for a particular type of hydrocarbon
mixture. When given a new hydrocarbon mixture, users must re-
rate when pore diameter is reduced to a nanometer. For applicabil- generate the data sets and then train a new ML model, which is
ity with shale reservoirs, researchers have primarily improved PR time consuming and laborious. As a result, a generalized ML model
EOS in the following three areas: 1) incorporating capillary pres- that is suitable for all possible hydrocarbon mixtures is required
sure to model the confinement effect [23–27]; 2) combining the for broader application.
effect of capillary pressure and adsorption on phase equilibrium In this paper, we propose a generalized proxy-phase equilib-
[28,29]; 3) tuning critical properties of pure components to nano- rium calculator for shale reservoirs, which have resolved the two

2
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

concerns mentioned above. The PR-C EOS [35] is applied to per-

form VLE calculations in nano-scale porous media. This method
has a high accuracy at small pore diameter ranges but requires a
long running time due to its complexity. The introduction of the
ML technique can significantly cut down the running time with-
out losing significant accuracy. The combination of the PR-C EOS
and ML techniques allowed our proxy phase-equilibrium calculator
to conduct VLE calculations in a fast and accurate way. In addi-
tion, a representative set of pseudo-components was considered in
this generalized proxy phase-equilibrium calculator. Unlike previ-
ous ML models in which the number of components and the types
of components were fixed, the general set of pseudo-components
allows users to customize the composition of a hydrocarbon mix-
ture, avoiding concerns of generating different ML models for var-
ious fluid compositions. In the stability test [67], the fluid phase
behavior was predicted by the MLP with an accuracy of about 97%.
In the phase-split computation [68], the PINN was applied to pro-
vide initial guesses for the Helmholtz free energyminimization al- Fig. 2. Schematic figure of a cylindrical pore. The dark green regions (Regions I
gorithm by incorporating the underlying scientific principles into and II) are the surface-adsorbed regions, while the light green region (Region III) is
neural networks, which ensured the rationality of the output and the confined core region. Region II represents the surface-adsorbed region that is
accessible to the mass center of the fluid molecules. Fpr is defined as the volume
reduced the number of iterations. In total, the generalized proxy
ratio of Region II over Regions II and III. (For interpretation of the references to
phaseequilibrium calculator cut down CPU time by 99% compared colour in this figure legend, the reader is referred to the web version of this article.)
to the traditional phase-equilibrium calculation method.
The paper is organized as follows. Section 2 detailly describes
the PR-C EOS and the algorithm of phase-equilibrium calcula- where T denotes the temperature, V represents the total volume
tion. Section 3 shows the general set of pseudo components, re- of the system, N is the total number of molecules, nc denotes the
quired parameters of PR-C EOS and the generation of data set. number of components in the system, q denotes intramolecular
Section 4 elaborates the optimized parameters and structures of contribution, λ is translational contribution, V f denotes the free
the neural networks in the ML models. Section 5 presents the anal- volume, and Econ f represents the configurational energy. Detailed
ysis of the results and Section 6 is the summary and conclusions equations are listed in Appendix A.
of our work. The PR-C EOS divides a pore into two parts: the surface ad-
sorbed region (Region I and II) and the confined core region (Re-
gion III), as shown in Fig. 2. The fluid molecules in the confined
2. Forward modeling core region are only affected by fluid-fluid intermolecular interac-
tions, while the fluid molecules in the surface adsorbed region are
2.1. Extended PR EOS (PR-C EOS) subject to both fluid-fluid and fluidpore wall interactions. Based on
this assumption, the configurational energy (Econ f ) is modeled as
Travalloni et al. [35] proposed an extended PR EOS (PR-C EOS) follows:
to investigate fluid-phase equilibrium in nano-scale conditions. The  

nc 
nc
Nj 
nc
fluidpore wall interaction is described by a square-well potential Econ f = − Nc,i j εi j − (Ni Fp,i ε p,i ) (2)
with the parameters of square-well depth (ε p ) and square-well 2
i=1 j=1 i=1
width (δ p ), as shown in Fig. 1. The PR-C EOS reverts to PR EOS
where Nc,i j is the coordination number, εi j denotes the interaction
when the pore diameter reaches macro-scale conditions.
energy between components i and j, ε p,i denotes the interaction
Based on the generalized van der Waals theory [33,34], the
energy between component i and the pore wall, and Fp,i repre-
canonical partition function (Q ) was calculated as follows:
sents the percentage of molecules of component i in the surface-
  T

nc
qi Ni Econ f adsorbed region. The first part of Eq. (2) represents the fluid-fluid
Q (T , V, N1 , N2 , . . . , Nnc ) = V f exp(
N
dT ) interactions, while the second part represents the fluidpore wall
i=1 λi Ni !
3Ni
∞
kB T 2
interactions. The term Fp,i is modeled by the empirical expression
(1) as shown below:
  ε  xi ρ
θi
p,i
Fp,i = Fpr,i + (1 − Fpr,i ) 1 − exp − 1− (3)
kB T ρmax,i
(r p − σi /2 )2 − (r p − σi /2 − δ p,i )2
Fpr,i = (4)
( r p − σi / 2 ) 2
The term Fpr,i is the volume fraction of the surface-adsorbed re-
gion accessible to the mass center of the fluid molecules in a pore
(calculated as the volume ratio of Region II over Regions II and III),
which is affected by the value of δ p,i . The term δ p,i is the square-
well width of the fluidpore wall interaction potential of component
i. When the distance between a fluid molecule and a pore wall is
less than δ p,i , the fluid molecule is subject to fluidpore wall inter-
action. In this paper, the square-well width (δ p ) of different types
of hydrocarbons was set as 0.5σi , which is widely used for square-
Fig. 1. Square-well potential model for the fluidpore wall interaction. well potential models in the existing literature [69–71].

3
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

The Helmholtz free energy (A) and chemical potential of com-

ponent i (μi ) were obtained using the following thermodynamic
relations [72]:
A(T , V, N1 , N2 , . . . , Nnc ) = −kB T lnQ (T , V, N1 , N2 , . . . , Nnc ) (5)
 
∂A
μi = (6)
∂ Ni T,V,N j=i

Herein, the PR-C EOS model was used in data generation in the
present study. The volume shift parameter (VSP) was incorporated
into the PR-C EOS to calibrate the prediction of liquid density [73],
and the binary interaction parameter (BIP) was considered to ac-
count for the deviation from ideal conditions [74]. In summary, the
input parameters of the forward model were pressure (P), temper-
ature (T), pore radius (r p ), number of moles (n), critical pressure
(Pc ), critical temperature (Tc ), molecular weight (MW), acentric fac-
tor (w), volume shift parameter (VSP), square-well depth (ε p ) and
binary interaction parameter (BIP). The Helmholtz free energy (A)
and chemical potential (μ) were then calculated using Eqs. (5) and
(6) based on the input parameters.

2.2. Stability test

Fig. 3. Flowchart of stability test using PR-C EOS. The fluid is first assumed to be
in single-phase condition, and the chemical potential of each component (μi I ) is
The basic principle of the stability test using the PR-C EOS is calculated using Eq. (6). Then, every possible composition with the same chemical
the equality of the chemical potential of component i in the vapor potential (μi II ) as μi I is searched using the NewtonRaphson iteration. If the new
and liquid phases, as shown in the equation below: phase composition (ni II ) is the same as the initial single-phase composition (ni I ),
the solution is trivial, and the fluid is in single-phase condition. Otherwise, the
μi V = μi L (7) fluid is in two-phase condition. If the fluid is in two-phase condition, the phase-
where μiV and μi L represent the chemical potential of compo- split computation is conducted to obtain the compositions of the vapor and liquid
phases.
nent i in vapor and liquid phases, respectively. We first assumed
that the fluid was in a single-phase condition and calculated the
chemical potential of each component (μi I ) using Eq. (6). We then components with PR-C EOS parameters is considered. Based on
searched for every possible composition with the same chemical the literature review of compositional analysis [75–77], we se-
potential (μi II ) as μi I . The Newton-Raphson iteration was used in lected six pseudo-components to represent the composition of var-
the search task. If the new phase composition (ni II ) was the same ious types of reservoir fluids, including C1 − N2 , C2 − CO2 − C3 , C4−6 ,
as the initial single-phase composition (ni I ), the solution was con- C7−12 , C13−24 , C25+ . When the fluid composition is given, it can be
sidered trivial, and the fluid was considered in single-phase condi- categorized as one of the six pseudo-components, and the PR EOS
tion; otherwise, the fluid was considered in two-phase condition. If parameters of each pseudo-component (ni , Pc,i , Tc,i , MWi , wi , V SPi ,
the fluid was in two-phase condition, the phase-split computation and BIPi, j ) can be obtained. If these PR EOS parameters are con-
was conducted to obtain the compositions of the vapor and liquid sidered to be the input variables for the ML model, the ML model
phases. The flowchart of the stability test is shown in Fig. 3. can be suitable for any fluid composition, which extends the appli-
cation of our approach from fixed-fluid composition to generalized
2.3. Phase-split computation
conditions.
The phase-split computation was performed by directly mini-
mizing the total Helmholtz free energy (Atotal ), which was calcu- 3.2. Data range and acquisition
lated using the equation below:
Atotal = AI + AII (8) The input parameters of the generalized proxy phase-
equilibrium calculator included P, T, r p , n, Pc , Tc , MW, w, VSP, ε p
where AI and AII represent the Helmholtz free energy of phases and BIP. The range of P and T followed the reference [52], and
I and II, which can be calculated by Eq. (5). The variables of the the range of r p was based on the pore-size distribution of shale
Helmholtz free energyminimization algorithm are the equilibrium reservoirs [66], as shown in Table 1. The lower bound of r p was
ratios of all pseudo-components (Ki ), which can be used to calcu- set at 2 nm because the pore diameter should be larger than the
late the number of moles of each component (ni ) in the vapor and effective diameter (σ ) of the fluid molecules, as required by the
liquid phases. The initial estimate of the equilibrium ratio was cal- PR-C EOS.
culated based on the ni I and ni II obtained from the stability test. The values of the PR EOS parameters (n, Pc , Tc , MW, w, VSP, and
Since the fluid is determined to be in two-phase condition, the to- BIP) were determined by applying the mixing rule based on the
tal Helmholtz free energy is minimized by changing the values of mole fraction of each real component in a pseudo-component, as
Ki . The lowest total Helmholtz free energy corresponds to the de-
sired value of Ki .
Table 1
3. Dataset generation Range of parameters of reservoir conditions.

Parameter Unit Minimum Maximum

3.1. General set of pseudo-components
Pressure (P) MPa 1 80
◦
Temperature (T) C 40 100
In the generalized proxy phaseequilibrium calculator, instead Pore radius (r p ) nm 2 50
of using a fixed set of real components, a general set of pseudo-

4
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

Table 2
Range of parameters of the extended PR EOS.

Parameter Unit Minimum Maximum

Critical pressure (Pc ) Pa 736,768 5,455,021

Critical temperature (Tc ) K 175 954
Molecular weight (MW) g/mol 16 533
Acentric factor (w) dimensionless 0.0084 1.3189
Volume shift parameter (VSP) dimensionless -0.1678 0.3535
Binary interaction parameter (BIP) dimensionless 0 0.1303
Square-well depth (ε p ) K 228 2788

shown in the equation below:


S pseudo = xreal,i Sreal,i (9)

where S denotes one of the PR EOS parameters and xreal,i is the

mole fraction of the real component i in a pseudo-component.
Each pseudo component was categorized under several real com-
ponents for the simplicity of data generation, as shown in Table B.1
in Appendix B. The values of the PR EOS parameters of the real
components are listed in Tables B.2 and B.3. To clarify Eq. (9),
an example of calculating the critical pressure of C2 − CO2 − C3 is
shown below:

Pc,C2 −CO2 −C3 = Pc,C2 · xC2 + Pc,CO2 · xCO2 + Pc,C3 · xC3 (10)

xC2 + xCO2 + xC3 = 1 (11)

where Pc,C2 −CO2 −C3 , Pc,C2 , Pc,CO2 , and Pc,C3 denote the critical pressure
of C2 − CO2 − C3 , C2 , CO2 , and C3 respectively and xC2 , xCO2 , and xC3
represent the mole fraction of C2 , CO2 , and C3 in C2 − CO2 − C3 .
The PR-C EOS parameter (ε p ) was calculated based on the equa-
tion [78] shown below
ε pδp
κ= (12) Fig. 4. Structure of the MLP of stability test. The numbers represent the number of
MW nodes in each layer.

where κ is the normalized fluid-pore surface affinity factor, which

can be obtained from the reference [78]. The values of ε p of real 4.1. Stability test
components are listed in Table B.4 The value of ε p of a pseudo-
component was calculated using the same method as shown in The function of the ML model of stability test was to determine
Eq. (9). the fluid phase state: single-phase or two-phase. MLP is applied
In summary, for a sixpseudo-component system, each parame- in the ML model to predict the phase state. The input variables
ter of n, Pc , Tc , MW, w, VSP, and ε p contained six values. The matrix of the MLP were the parameters of the reservoir conditions and
of BIP had a symmetric distribution, and the values of the diagonal EOS, a total number of 60, as mentioned in Section 3.2. The out-
elements in the matrix were zeros. As a result, for a sixpseudo- put variable was the label of the phase-state condition, including
component system, 36 elements in the BIP matrix could be de- single-phase and two-phase conditions. The detailed structure of
scribed by 15 different values. Together with the parameters of the the MLP is shown in Fig. 4, which includes an input layer, five
reservoir conditions (P, T, r p ), the total number of input parameters hidden layers, and an output layer. The activation function of the
for the generalized model was 60. The ranges of different parame- hidden layers was ReLU [80], whereas the activation function of
ters of the extended PR EOS are summarized in Table 2. the output layer was sigmoid. The loss function was set as binary
cross-entropy because the stability test is a binary classification
3.3. Data samples problem. The Adam optimization algorithm [81] was used in the
MLP, which is computationally efficient and requires less memory.
The samples were generated using the Latin hypercube sam- During the training process, the model is optimized by continu-
pling technique [79] to ensure randomness in a multidimensional ously changing the weight and bias values of each node. All input
distribution. A total of 123,050 samples were generated for the variables were scaled based on the normalization equation shown
training and testing of the ML model. below:
yscaled = (y − ymin )/(ymax − ymin ) (13)
4. ML models for phase-equilibrium calculation
where y denotes a variable among the 60 input variables.
The phase-equilibrium calculation consisted of two parts: the A total of 123,050 samples were used for the stability test. The
stability test and the phase split computation. In the stability test, k-fold cross-validation technique [82,83] was applied in the train-
the MLP was applied to predict the fluid-phase state (single-phase ing process. In the present study, k was set as 10, which means
or two-phase) due to its high flexibility and efficiency. In the the samples were divided into 10 parts. Each time, one part was
phase-split computation, the MLP and PINN models were trained selected as the test set and the remaining nine parts were used
to obtain the initial estimates of equilibrium ratios, rather than us- for the training set. The training process was performed 10 times
ing the initial estimate from the stability test. until each part had been selected as the test set. The accuracy ob-

5
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

Table 3
The optimized parameters of MLP.

Parameters of MLP Optimal results

Number of hidden layers 5

Number of nodes 64, 128, 128, 128, 64
Batch size 64
Number of epochs 80
Learning rate of Adam 0.001

tained using k-fold cross-validation was more robust and trustwor-

thy compared to the method of splitting data directly, as each sam-
ple was considered for training and testing.
To increase the accuracy of our ML model, several parameters
of the MLP were optimized separately. First, the number of hidden
layers were optimized to provide a proper MLP structure. The re-
sults showed that the structure with five hidden layers had the
highest accuracy. Less layers may not be able to fully interpret
the complex relationship between the input and output variables,
while too many layers may lead to overfitting. Next, different com-
binations of nodes in each layer tested, and the results showed that
this specific set of nodes (64, 128, 128, 128, and 64) had the best
performance. By balancing accuracy and CPU time, the optimized
batch size and the optimized number of epochs were 64 and 80,
respectively. The optimal learning rate of Adam was 0.001. The op-
timized parameters of the MLP are summarized in Table 3.

4.2. Phase-split computation

The function of the ML model of phase-split computation

was to obtain the initial guess of the equilibrium ratio for the
Helmholtz free energyminimization algorithm. A good initial guess Fig. 5. Structure of the MLP of phase split computation. The numbers represent the
was shown to largely reduce the number of iterations and, there- number of nodes in each layer.
fore, lower the CPU time. In the forward modeling, the initial esti-
mate of the equilibrium ratio was obtained from the stability test. Table 4
In the generalized proxy phase-equilibrium calculator, the MLP and The optimized parameters of MLP.

PINN models of phase-split computation were trained to obtain the Parameters of MLP Optimal results
equilibrium ratio as the initial estimate for the Helmholtz free en-
Number of hidden layers 4
ergyminimization algorithm. Number of nodes 64, 128, 128, 64
Batch size 64
Number of epochs 40
4.2.1. MLP neural network of phase split computation Learning rate of Adam 0.001
The input variables were the parameters of the reservoir con-
ditions and EOS, whereas the output variables were the equilib-
rium ratios of the six pseudo-components. The structure of the 4.2.2. PINN of phase-split computation
MLP neural network of the phase-split computation is shown in Physics-based models are typically limited by the assumptions
Fig. 5, which is similar to the MLP neural network used for the of ideal conditions, which may not fully interpret the complex
stability test. The activation functions of the hidden layers are ReLU relationships in engineering systems. Black-box ML models often
whereas the activation function of the output layer is sigmoid since encounter problems in that the predictions are inconsistent with
the value of the output is in the range of 0 and 1. The optimization basic physical laws. In addition, ML models require a substantial
algorithm used for this model was also Adam. Both the input and amount of data for training and fail to predict accurately when
output variables were normalized using Eq. (13). The loss function the values of the input variables are beyond the range of the
was the mean square error (MSE), which was calculated based on training set. To avoid the limitations and leverage the advantages
the equation shown below: of the physics-based and ML models, many hybrid physics-data
models have been proposed by researchers to incorporate physical
1 
ns nc
2
MSE = Kˆi, j − Ki, j (14) mechanisms into ML models, including the physics-informed neu-
ns · nc ral network (PINN), physics-constrained neural network (PCNN),
j=1 i=1
and physics-guided neural network (PGNN).
where ns represents the number of samples and Kˆi, j is the pre- The PINN was initially designed to solve partial differential
dicted equilibrium ratio of component i in sample j. equations or to determine potential partial differential equations to
The total number of samples for phase-split computation was govern a system [84]. Recently, PINNs have been applied in
52,102. The MLP neural network of phase-split computation was petroleum engineering studies. Klie et al. [85] applied a PINN in
optimized following the same method for the optimization of the the prediction of oil and gas production. Given that oil and gas
MLP neural network for the stability test. The optimized results are rates are functions of time, the ordinary differential equation (ODE)
summarized in Table 4. set with respect to time was incorporated to train the ML model

6
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

Fig. 7. The average CPU time for phase-equilibrium calculation for one sample us-
ing PR EOS and PR-C EOS.

Fig. 6. Flow chart of the PINN of phase-split computation. The structure of the neu-
ral network is shown in the blue dashed box. The input variables are the parameters
shown in Fig. 7. In the previous flash calculation studies [88], the
of reservoir conditions and EOS, whereas the output parameters of the neural net- computation time per flash using cubic EOS is about 10 microsec-
work are the equilibrium ratios. Atotal is calculated based on Eqs. (5) and (8) using Kˆ onds, which is much faster compared with our simulator. However,
predicted by the neural network. During the iterations, L is continuously minimized the aim of this section is not to optimize and accelerate tradi-
and therefore, the neural network is trained to give more accurate predictions of Kˆ .
tional phase-equilibrium calculator, but to show that the phase-
(For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.) equilibrium calculation using PR-C EOS is much more complex
compared with the phase-equilibrium calculation using PR EOS.
The reasons are two folded: 1) the interaction between the fluid
in a specific direction. Sarma et al. [86] applied the PINN in their and pore wall is incorporated in PR-C EOS. As shown in Section 2.1,
steam-injection model to reduce operational costs and increase the more parameters are introduced in PR-C EOS when the effect of
production rate. The equations of mass conservation, energy con- fluid-pore wall interaction is taken into consideration. Compared
servation, and Darcy’s law were incorporated into the PINN to en- with PR EOS, these extra computations require more CPU time;
sure the rationality and accuracy of the output. In this paper, PINN 2) the algorithm of solving the equilibrium of chemical potential
is applied in the phase-split computation to obtain the initial guess (Eq. (7)) is more difficult with PR-C EOS. In the phase-equilibrium
for the Helmholtz free energy-minimization algorithm. calculation using PR EOS, the method of tangent plane distance
In the PINN, the input variables are the parameters of reser- (TPD) is often applied to find the solutions. However, in the phase-
voir conditions and EOS whereas the output parameters are the equilibrium calculation using PR-C EOS, the TPD method is not
equilibrium ratios. The flow chart of the PINN of phase split com- applicable. The fluid is considered to be in single-phase condi-
putation is shown in Fig. 6. The structure of the neural network tion only when all possible roots are checked to be trivial roots,
in the PINN was the same as the neural network of the MLP for which is much more time-consuming. As a result, the application
the phase-split computation, as shown in the blue dashed box in of ML technique provided a more significant speedup in the phase-
Fig. 6. Different from the MLP model that directly minimizes the equilibrium calculations using PR-C EOS.
deviations between the real values and the estimated values of the
equilibrium ratios, the equilibrium ratios (Kˆ ) predicted by the neu- 5.2. Stability test
ral network are used to calculate the total Helmholtz free energy
(Atotal ) based on Eqs. (5) and (8). During the iterations, Atotal is con- Based on the optimized MLP-based stability test, the average
tinuously minimized and therefore, the neural network is trained accuracy of the test set using k-fold cross validation technique
to give more accurate predictions of Kˆ . It should be noted that, (k = 10) was 96.96%. We randomly selected 500 samples from the
during the training process, the real values of the equilibrium ra- data set and performed the stability test by applying the algorithm
tios are not needed, which means that the training of PINN does in forward modeling, as well as the MLP-based stability test. The
not require a data set. What we need to prepare is only the differ- average CPU time for performing the stability test for one sample
ent combination of the values of input variables. The optimization using the two methods is shown in Fig. 8. The results indicated
algorithm used was Adam, with a learning rate of 0.001. The loss
function was calculated based on the equation shown below:

1 
ns
L= Atotal, j /107 (15)
ns
j=1

5. Results and discussion

In this work, all the ML models were trained on Keras [87] with
the CPU of Intel Core i7-1065G7 and an installed memory of 16
GB. A comparison between the forward modeling and the proxy
phaseequilibrium calculator is shown below.

5.1. Phase-equilibrium calculator using PR EOS

Herein, we first compare the average simulation time for the Fig. 8. Average CPU time for performing the stability test for one sample using for-
phase-equilibrium calculation using PR EOS and PR-C EOS, as ward modeling and the MLP.

7
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

The mean absolute error (MAE) of the predicted equilibrium ra-

tios was calculated based on the equation shown below:

1   Kˆi, j − Ki, j
ns nc
MAE = (16)
ns · nc Ki, j
j=1 i=1

The comparisons between the deviations of the predicted equilib-

rium ratios by the initial estimate from the stability test and the
ML models are shown in Fig. 9. The results show that the de-
viations of the equilibrium ratios obtained by MLP with 52,102
samples and PINN with 10 0 0 input combinations are 77.26% and
82.12% lower than the deviations obtained from the stability test,
respectively. The predicted results from the MLP with 10 0 0 sam-
Fig. 9. Mean absolute error (MAE) of equilibrium ratios obtained by the initial es-
timate from the stability test and ML models: (a) the initial estimate obtained from ples are worse than the initial estimate from the stability test, indi-
the stability test, (b) MLP neural network trained with 10 0 0 samples, (c) MLP neu- cating that the training of MLP requires a large data set to achieve
ral network trained with 52,102 samples, (d) PINN trained with 10 0 0 samples. The high accuracy. The results show that, by incorporating the PR-C
MAEs are calculated using Eq. (16).
EOS into the training of neural network, PINN can accurately pre-
dict the equilibrium ratios with a limited range of input variables.
However, for the MLP neural network, a huge data set is required
that the utilization of the ML technique can significantly improve to achieve a high accuracy. As a result, for the condition that a
the computational efficiency by 963 times compared to the tradi- huge data set is not available, PINN can be applied to avoid the
tional method. concern of generating the data set since PINN directly minimizes
the loss function.
Although the deviations of equilibrium ratios obtained by the
ML models are significantly reduced compared with the initial es-
5.3. Phase-split computation timate obtained from the stability test, the minimization program
cannot be completely replaced by the ML models. The reason is
For the phase split computation, the PINN model was trained that the equilibrium ratios will be used to determine the compo-
using 10 0 0 input combinations, while the MLP models were sition of each phase. A small deviation of equilibrium ratios can
trained with 10 0 0 and 52,102 samples, respectively. have significant effect on fluid composition. In addition, the error

Fig. 10. Flow chart of the traditional PR-C EOS phase-equilibrium calculator and the generalized proxy phaseequilibrium calculator. In the traditional PR-C EOS phase-
equilibrium calculator, the stability test is conducted by NewtonRaphson iterations to find possible solutions to the equality of chemical potentials, and the initial guess of
the equilibrium ratio for the Helmholtz free energyminimization algorithm is obtained from the result in stability test. In the generalized proxy phaseequilibrium calculator,
the phase behavior is directly predicted by the ML model, and the initial guess for the Helmholtz free energyminimization algorithm is obtained from the PINN of the
phase-split computation.

8
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

of the ML model to achieve high accuracy. For the stability test, the
MLP model was trained based on the input variables of the reser-
voir conditions and the EOS parameters to determine whether the
fluid was in single-phase or two-phase condition. If the fluid was
in two-phase condition, the PINN model was used to obtain the
initial guess of the equilibrium ratio for the Helmholtz free ener-
gyminimization algorithm instead of the estimate in stability test.
The generalized proxy phaseequilibrium calculator was shown to
be suitable for any fluid composition and could be implemented
into reservoir simulators to perform phase-equilibrium calculations
for shale reservoirs in a fast and accurate manner. In summary, we
have the following findings and contributions:
(1) The extended PR EOS was applied to investigate the fluid-
Fig. 11. Average CPU time of the phase-equilibrium calculation for one sample per- phase behavior in shale reservoirs, which ensures the accuracy of
formed by the traditional PR-C EOS phase-equilibrium calculator and the general- the generalized proxy phase-equilibrium calculator under nano-
ized proxy phaseequilibrium calculator. scale conditions.
(2) Compared with previous works in which the number of
caused by the deviations will accumulate and magnify in the reser- components and the types of components were fixed, the general
voir simulation process, which leads to a large discrepancy. set of pseudo-components allows users to customize the compo-
sition of a hydrocarbon mixture, avoiding concerns of generating
5.4. Phase-equilibrium calculation different ML models for various fluid compositions.
(3) By incorporating the PR-C EOS into the training of neural
Herein, the stability test and phase-split computation were in- network, PINN can accurately predict the equilibrium ratios with a
tegrated as phase-equilibrium calculations. The algorithms of the limited range of input variables, which indicates that, for the con-
traditional PR-C EOS phase-equilibrium calculator and the gener- dition that a huge data set is not available, PINN can be applied to
alized proxy phaseequilibrium calculator are shown in Fig. 10. In avoid the concern of generating the data set.
the traditional PR-C EOS phase-equilibrium calculator, the stabil- (4) For the stability test, the accuracy of the MLP model was
ity test is conducted by NewtonRaphson iterations to find possible 96.96%, and the average CPU time was reduced by 963 times com-
solutions to the equality of chemical potentials. In the gener- pared to the traditional method using Newton-Raphson iterations.
alized proxy phaseequilibrium calculator, the results of the sta- (5) For the phase-split computation, the deviations of the equi-
bility test are directly obtained from the ML model. For the librium ratios obtained by the PINN were 82.12% lower than the
phase-split computation, the initial guess of the equilibrium ra- deviations obtained from the stability test. The MLP model had
tio for the Helmholtz free energyminimization algorithm is ob- similar achievements when given a large data set.
tained from the result in stability test in the traditional PR-C EOS (6) In total, the average CPU time for one phase-equilibrium
phase-equilibrium calculator, whereas the initial guess is obtained calculation was reduced by 627 times using the generalized proxy
from the PINN of the phase-split computation in the generalized phase-equilibrium calculator.
proxy phaseequilibrium calculator. In the generalized proxy phase-
equilibrium calculator, we have included an additional phase- Declaration of Competing Interest
stability check: if the minimized Atotal is higher than Asingle , the
fluid is considered in single-phase condition; otherwise, the sys- The authors declare that they have no known competing finan-
tem is in two-phase condition. Different from the traditional PR-C cial interests or personal relationships that could have appeared to
EOS phase-equilibrium calculation, here, the result of stability test influence the work reported in this paper.
is directly obtained from the ML prediction result, which has an
accuracy of about 97%. The additional phase-stability check further
CRediT authorship contribution statement
inspects the accuracy of the ML-based phase stability analysis, and
therefore increases the accuracy and the reliability of the phase
Fangxuan Chen: Conceptualization, Methodology, Visualization,
equilibrium calculations.
Software, Writing – original draft. Sheng Luo: Conceptualization,
We randomly selected 100 samples from the data set and
Methodology, Software, Writing – review & editing. Shihao Wang:
performed the phase-equilibrium calculation using the traditional
Conceptualization, Methodology, Writing – review & editing. Hadi
PR-C EOS phase-equilibrium calculator and the generalized proxy
Nasrabadi: Conceptualization, Supervision, Writing – review &
phaseequilibrium calculator. The average CPU time of the phase-
editing.
equilibrium calculation for one sample is shown in Fig. 11. Com-
pared to the traditional method, the average CPU time was reduced
by 627 times with the assistance of the ML technique. Acknowledgement

6. Conclusion We appreciate the support from the Texas A&M High Perfor-
mance Research Computing Facility in performing the required
In the present work, we proposed a generalized ML-assisted phase-equilibrium calculations to generate the data set.
phase-equilibrium calculation model to perform stability tests and
phase-split computations for shale reservoirs. Compared to the Appendix A
previous work of ML-assisted phase-equilibrium calculation, our
model has three main innovations as follows: 1) the extended PR The free volume (V f ) is calculated based on the equa-
EOS was applied to investigate the fluid-phase behavior in shale tions shown below:
reservoirs; 2) a general set of pseudo-components was selected to

nc
Ni
allow users to customize the composition of the hydrocarbon mix- Vf = V − (A.1)
ture; 3) the physical equations were incorporated into the training
i=1
ρmax,i

9
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

Table B.1 Table B.2

The lumped real components of 6 pseudo components. The PR EOS parameters of all real components.

Pseudo Components Real Components Real Components Pc (Pa) Tc (K) MW(g/mol) w VSP

C1 − C1 C1 4,600,155 190.6 16 0.008 −0.1595

N2 N2 N2 3,394,388 126.2 28 0.040 −0.1927
C2 − C2 C2 4,883,865 305.4 30 0.098 −0.1134
CO2 − CO2 CO2 7,376,460 304.2 44 0.225 −0.0817
C3 C3 C3 4,245,518 369.8 44 0.152 −0.0863
C4−6 C4 C4 3,799,688 425.2 58 0.193 −0.0675
C5 C5 3,374,123 469.6 72 0.251 −0.0390
C6 C6 2,968,823 507.4 86 0.296 −0.0080
C7−12 C7 C7 2,735,775 540.2 100 0.351 0.0033
C9 C9 2,310,210 594.6 128 0.444 0.0408
C12 C12 2,191,660 663.9 161 0.522 0.0695
C13−24 C13 C13 2,070,070 682.4 175 0.560 0.0775
C19 C19 1,525,955 771.0 163 0.790 0.1225
C24 C24 1,254,404 823.2 324 0.940 0.1835
C25+ C25 C25 1,199,688 832.7 337 0.965 0.1990
C35 C35 902,806 905.9 445 1.179 0.2883
C45 C45 722,346 957.8 539 1.330 0.3583

The parameter θi in Eq. (3) is calculated as follows:

 rp

rp
ρmax,i σi = 1.158 − 0.479exp 0.621(0.5 − )
3
θi = (A.4)
σ δ p,i + σi /2
 rp
 i
+ 0.595exp 4.014(0.5 − ) (A.2) The confinement-modified energy parameter (a p,i j ) between
σi components i and j and the confinement-modified volume param-
eter (b p,i ) of component i were calculated as follows:
1.158bi  
σi = 3
(A.3)
2 σi j
Nav
a p,i j = ai a j 1− (A.5)
5 rp
The parameters ρmax,i and σi represent the packing density of pure
component i and the effective diameter of component i, respec- Nav
b p,i = (A.6)
tively. ρmax,i

Table B.3
The values of BIP of all real components.

Real Components C1 (Pa) N2 C2 CO2 C3 C4 C5 C6

C1 0.000000 0.025000 0.002689 0.105000 0.008537 0.014749 0.020640 0.028330

N2 0.025000 0.000000 0.010000 0.000000 0.090000 0.095000 0.110000 0.110000
C2 0.002689 0.010000 0.000000 0.130000 0.001662 0.004914 0.008578 0.013840
CO2 0.105000 0.000000 0.130000 0.000000 0.125000 0.115000 0.115000 0.115000
C3 0.008537 0.090000 0.001662 0.125000 0.000000 0.000866 0.002712 0.005978
C4 0.014749 0.095000 0.004914 0.115000 0.000866 0.000000 0.000515 0.002305
C5 0.020640 0.110000 0.008578 0.115000 0.002712 0.000515 0.000000 0.000643
C6 0.028330 0.110000 0.013840 0.115000 0.005978 0.002305 0.000643 0.000000
C7 0.035185 0.110000 0.018844 0.115000 0.009436 0.004615 0.002055 0.000400
C9 0.047012 0.110000 0.027939 0.115000 0.016229 0.00968 0.005761 0.002566
C12 0.054287 0.110000 0.033750 0.115000 0.020795 0.013300 0.008640 0.004592
C13 0.058520 0.110000 0.037195 0.115000 0.023560 0.015555 0.010480 0.005967
C19 0.081500 0.110000 0.056500 0.115000 0.039669 0.029200 0.022170 0.015390
C24 0.094117 0.110000 0.067493 0.115000 0.049172 0.037552 0.029600 0.021720
C25 0.096000 0.110000 0.069170 0.115000 0.050638 0.038853 0.030740 0.022730
C35 0.116688 0.110000 0.087486 0.115000 0.066883 0.053464 0.044000 0.034450
C45 0.132875 0.110000 0.102154 0.115000 0.080148 0.065610 0.055240 0.044580
Real Components C7 C9 C12 C13 C19 C24 C25 C35 C45
C1 0.035185 0.047012 0.054287 0.058520 0.081500 0.094117 0.096000 0.116688 0.132875
N2 0.110000 0.110000 0.110000 0.110000 0.110000 0.110000 0.110000 0.110000 0.110000
C2 0.018844 0.027939 0.033750 0.037195 0.056500 0.067493 0.069170 0.087486 0.102154
CO2 0.115000 0.115000 0.115000 0.115000 0.115000 0.115000 0.115000 0.115000 0.115000
C3 0.009436 0.016229 0.020795 0.023560 0.039669 0.049172 0.050638 0.066883 0.080148
C4 0.004615 0.009680 0.013300 0.015555 0.029200 0.037552 0.038853 0.053464 0.065610
C5 0.002055 0.005761 0.008640 0.010480 0.022170 0.029600 0.030740 0.044000 0.055240
C6 0.000400 0.002566 0.004592 0.005967 0.015390 0.021720 0.022730 0.034450 0.044580
C7 0.000000 0.000942 0.002289 0.003289 0.010895 0.016334 0.017220 0.027663 0.036900
C9 0.000942 0.000000 0.000295 0.000713 0.005470 0.009517 0.010200 0.018615 0.026410
C12 0.002289 0.000295 0.000000 0.000091 0.003234 0.006486 0.007055 0.014300 0.021260
C13 0.003289 0.000713 0.000091 0.000000 0.002243 0.005050 0.005554 0.012143 0.018627
C19 0.010895 0.005470 0.003234 0.002243 0.000000 0.000566 0.000744 0.004000 0.008064
C24 0.016334 0.009517 0.006486 0.005050 0.000566 0.000000 0.000012 0.001560 0.004378
C25 0.017220 0.010200 0.007055 0.005554 0.000744 0.000012 0.000000 0.001297 0.003931
C35 0.027663 0.018615 0.014300 0.012143 0.004000 0.001560 0.001297 0.000000 0.000716
C45 0.036900 0.026410 0.021260 0.018627 0.008064 0.004378 0.003931 0.000716 0.000000

10
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

Table B.4 [9] S. Luo, H. Nasrabadi, J.L. Lutkenhaus, Effect of confinement on the bub-
The values of ε p of real components. ble points of hydrocarbons in nanoporous media, AlChE J. 62 (5) (2016)
1772–1780.
Real Components ε p /kB (K) [10] S. Luo, J.L. Lutkenhaus, H. Nasrabadi, Confinement-induced supercriticality
C1 228.57 and phase equilibria of hydrocarbons in nanopores, Langmuir 32 (44) (2016)
N2 444.71 11506–11513.
[11] S. Luo, J.L. Lutkenhaus, H. Nasrabadi, Use of differential scanning calorimetry
C2 490.48
to study phase behavior of hydrocarbon mixtures in nano-scale porous media,
CO2 1202.76
J. Pet. Sci. Eng. 163 (2018) 731–738.
C3 1215.24 [12] X. Qiu, S.P. Tan, M. Dejam, H. Adidharma, Isochoric measurement of the evap-
C4 1842.76 oration point of pure fluids in bulk and nanoporous media using differential
C5 1913.58 scanning calorimetry, PCCP 22 (13) (2020) 7048–7057.
C6 1975.45 [13] X. Qiu, S.P. Tan, M. Dejam, H. Adidharma, Experimental study on the criticality
C7 2030.54 of a methane/ethane mixture confined in nanoporous media, Langmuir 35 (36)
C9 2125.72 (2019) 11635–11642.
C12 2242.81 [14] P. Zeigermann, M. Dvoyashkin, R. Valiullin, J. Kärger, Assessing the pore critical
C13 2276.95 point of the confined fluid by diffusion measurement (2009).
C19 2448.06 [15] M. Alfi, H. Nasrabadi, D. Banerjee, Experimental investigation of confinement
C24 2561.16 effect on phase behavior of hexane, heptane and octane using lab-on-a-chip
C25 2581.55 technology, Fluid Phase Equilib. 423 (2016) 25–33.
[16] M. Alfi, H. Nasrabadi, D. Banerjee, Effect of confinement on bubble point
C35 2756.77
temperature shift of hydrocarbon mixtures: experimental investigation us-
C45 2896.25
ing nanofluidic devices, in: SPE Annual Technical Conference and Exhibition,
OnePetro, 2017.
[17] Q. Yang, B. Jin, D. Banerjee, H. Nasrabadi, Direct visualization and molecular
simulation of dewpoint pressure of a confined fluid in sub-10 nm slit pores,
Applying the mixing rule based on the mole fraction of each com- Fuel 235 (2019) 1216–1223.
ponent in a mixture, the confinement-modified energy parameter [18] Z. Sun, X. Li, W. Liu, T. Zhang, M. He, H. Nasrabadi, Molecular dynamics of
(a p ) and the confinement-modified volume parameter (b p ) of the methane flow behavior through realistic organic nanopores under geologic
shale condition: pore size and kerogen types, Chem. Eng. J. 398 (2020) 124341.
fluid mixture were calculated as follows: [19] B. Jin, R. Bi, H. Nasrabadi, Molecular simulation of the pore size distribution ef-
fect on phase behavior of methane confined in nanopores, Fluid Phase Equilib.

nc 
nc
ap = (xi x j a p,i j ) (A.7) 452 (2017) 94–102.
[20] B. Jin, H. Nasrabadi, Phase behavior of multi-component hydrocarbon systems
i=1 j=1 in nano-pores using gauge-GCMC molecular simulation, Fluid Phase Equilib.
425 (2016) 324–334.

nc [21] R. Bi, H. Nasrabadi, Molecular simulation of the constant composition expan-
bp = xi b p,i (A.8) sion experiment in shale multi-scale systems, Fluid Phase Equilib. 495 (2019)
59–68.
i=1
[22] D.-Y. Peng, D.B. Robinson, A new two-constant equation of state, Ind. Eng.
The parameters of the PR EOS used in this paper were determined Chem. Fundam. 15 (1) (1976) 59–64.
[23] B. Nojabaei, R.T. Johns, L. Chu, Effect of capillary pressure on phase behavior
by the equations shown below: in tight rocks and shales, SPE Reservoir Eval. Eng. 16 (03) (2013) 281–289.
 [24] N. Siripatrachai, T. Ertekin, R. Johns, Compositional simulation of discrete frac-
0.3746 + 1.5423wi − 0.2699wi 2 0 < wi  0.5 tures incorporating the effect of capillary pressure on phase behavior, in: SPE
mi =
0.3796 + 1.4850wi − 0.1644wi 2 + 0.01677wi 3 wi > 0.5 Improved Oil Recovery Conference, OnePetro, 2016.
[25] L. Du, L. Chu, Understanding anomalous phase behavior in unconventional oil
(A.9) reservoirs, in: SPE Canadian Unconventional Resources Conference, OnePetro,
2012.
  2 [26] D.R. Sandoval, M.L. Michelsen, W. Yan, E.H. Stenby, VT-based phase envelope
0.4572R2 Tc,i 2 T
ai = 1 + mi 1 − (A.10) and flash calculations in the presence of capillary pressure, Ind. Eng. Chem.
Pc,i Tc,i Res. 58 (13) (2019) 5291–5300.
[27] D.R. Sandoval, W. Yan, M.L. Michelsen, E.H. Stenby, The phase envelope of mul-
ticomponent mixtures in the presence of a capillary pressure difference, Ind.
0.07780RTc,i
bi = (A.11) Eng. Chem. Res. 55 (22) (2016) 6530–6538.
Pc,i [28] D.R. Sandoval Lemus, W. Yan, E.H. Stenby, Phase equilibrium in shale includ-
ing porous media effects, in: Abu Dhabi International Petroleum Exhibition &
Conference, OnePetro, 2019.
[29] D.R. Sandoval, W. Yan, M.L. Michelsen, E.H. Stenby, Influence of adsorption and
References capillary pressure on phase equilibria inside shale reservoirs, Energy Fuels 32
(3) (2018) 2819–2833.
[1] H. Zhang, J. Sheng, Optimization of horizontal well fracturing in shale gas [30] L. Jin, Y. Ma, A. Jamili, Investigating the effect of pore proximity on phase be-
reservoir based on stimulated reservoir volume, J. Pet. Sci. Eng. 190 (2020) havior and fluid properties in shale formations, in: SPE Annual Technical Con-
107059. ference and Exhibition, OnePetro, 2013.
[2] N. Nagel, M. Sanchez-Nagel, F. Zhang, X. Garcia, B. Lee, Coupled numerical eval- [31] X. Xiong, D. Devegowda, G. Michel, R.F. Sigal, F. Civan, A fully-coupled free and
uations of the geomechanical interactions between a hydraulic fracture stimu- adsorptive phase transport model for shale gas reservoirs including non-darcy
lation and a natural fracture system in shale formations, Rock Mech. Rock Eng. flow effects, in: SPE Annual Technical Conference and Exhibition, OnePetro,
46 (3) (2013) 581–609. 2012.
[3] K.S. Sing, Reporting physisorption data for gas/solid systems with special ref- [32] N.S. Alharthy, T.N. Nguyen, T.W. Teklu, H. Kazemi, R.M. Graves, Multiphase
erence to the determination of surface area and porosity (recommendations compositional modeling in small-scale pores of unconventional shale reser-
1984), Pure Appl. Chem. 57 (4) (1985) 603–619. voirs, in: SPE Annual Technical Conference and Exhibition, OnePetro, 2013.
[4] R.F. Sigal, Pore-size distributions for organic-shale-reservoir rocks from nucle- [33] S.I. Sandler, The generalized van der Waals partition function. I. Basic theory,
ar-magnetic-resonance spectra combined with adsorption measurements, SPE Fluid Phase Equilib. 19 (3) (1985) 238–257.
J. 20 (04) (2015) 824–830. [34] K.-H. Lee, M. Lombardo, S. Sandler, The generalized van der Waals partition
[5] X. Liu, D. Zhang, A review of phase behavior simulation of hydrocarbons in function. II. Application to the square-well fluid, Fluid Phase Equilib. 21 (3)
confined space: implications for shale oil and shale gas, J. Nat. Gas Sci. Eng. 68 (1985) 177–196.
(2019) 102901. [35] L. Travalloni, M. Castier, F.W. Tavares, Phase equilibrium of fluids confined in
[6] K. Morishige, H. Fujii, M. Uga, D. Kinukawa, Capillary critical point of argon, porous media from an extended Peng–Robinson equation of state, Fluid Phase
nitrogen, oxygen, ethylene, and carbon dioxide in MCM-41, Langmuir 13 (13) Equilib. 362 (2014) 335–341.
(1997) 3494–3498. [36] S. Luo, B. Jin, J.L. Lutkenhaus, H. Nasrabadi, A novel pore-size-dependent equa-
[7] P. Russo, M.R. Carrott, P. Carrott, Trends in the condensation/evaporation and tion of state for modeling fluid phase behavior in nanopores, Fluid Phase Equi-
adsorption enthalpies of volatile organic compounds on mesoporous silica ma- lib. 498 (2019) 72–85.
terials, Microporous Mesoporous Mater. 151 (2012) 223–230. [37] L. Travalloni, M. Castier, F.W. Tavares, S.I. Sandler, Thermodynamic modeling
[8] S. Kittaka, M. Morimura, S. Ishimaru, A. Morino, K. Ueda, Effect of confinement of confined fluids using an extension of the generalized van der Waals theory,
on the fluid properties of ammonia in mesopores of MCM-41 and SBA-15, Chem. Eng. Sci. 65 (10) (2010) 3088–3099.
Langmuir 25 (3) (2009) 1718–1724.

11
F. Chen, S. Luo, S. Wang et al. Fluid Phase Equilibria 558 (2022) 113423

[38] S. Luo, F. Chen, D. Zhou, H. Nasrabadi, Multiscale pressure/volume/temperature [62] M.E. Tipping, Sparse Bayesian learning and the relevance vector machine, J.
simulation of decreasing condensate/gas ratio at greater than dewpoint pres- Mach. Learn. Res. 1 (Jun) (2001) 211–244.
sure in shale gas-condensate reservoirs, SPE J. (2021) 1–13. [63] M. Mesbah, E. Soroush, V. Azari, M. Lee, A. Bahadori, S. Habibnia, Vapor liq-
[39] V. Gaganis, N. Varotsis, Machine learning methods to speed up compositional uid equilibrium prediction of carbon dioxide and hydrocarbon systems using
reservoir simulation, in: SPE Europec/EAGE Annual Conference, OnePetro, 2012. LSSVM algorithm, J. Supercrit. Fluids 97 (2015) 256–267.
[40] C. Temizel, C.H. Canbaz, O. Saracoglu, D. Putra, A. Baser, T. Erfando, S. Kr- [64] H.-Z. Huang, H.-K. Wang, Y.-F. Li, L. Zhang, Z. Liu, Support vector machine
ishna, L. Saputelli, et al., Production forecasting in shale reservoirs using LSTM based estimation of remaining useful life: current research status and future
method in deep learning, in: SPE/AAPG/SEG Unconventional Resources Tech- trends, J. Mech. Sci. Technol. 29 (1) (2015) 151–163.
nology Conference, Unconventional Resources Technology Conference, 2020. [65] S. Wang, N. Sobecki, D. Ding, L. Zhu, Y.-S. Wu, Accelerating and stabilizing the
[41] H.-H. Liu, J. Zhang, F. Liang, C. Temizel, M.A. Basri, R. Mesdour, Incorporation of vapor-liquid equilibrium (VLE) calculation in compositional simulation of un-
physics into machine learning for production prediction from unconventional conventional reservoirs using deep learning based flash calculation, Fuel 253
reservoirs: a brief review of the gray-box approach, SPE Reservoir Eval. Eng. (2019) 209–219.
(2021) 1–12. [66] C.R. Clarkson, N. Solano, R.M. Bustin, A. Bustin, G. Chalmers, L. He, Y.B. Mel-
[42] M. Mehana, E. Guiltinan, V. Vesselinov, R. Middleton, J.D. Hyman, Q. Kang, nichenko, A. Radliński, T.P. Blach, Pore structure characterization of North
H. Viswanathan, Machine-learning predictions of the shale wells performance, American shale gas reservoirs using USANS/SANS, gas adsorption, and mercury
J. Nat. Gas Sci. Eng. 88 (2021) 103819. intrusion, Fuel 103 (2013) 606–616.
[43] Y. Wang, H. Liu, Y. Zhou, Development of a deep learning-based model for the [67] R. Bi, A. Firoozabadi, P.C. Myint, Efficient and robust phase-split computations
entire production process of steam-assisted gravity drainage (SAGD), Fuel 287 in the internal energy, volume, and moles (UVN) space, Fluid Phase Equilib.
(2021) 119565. 526 (2020) 112729.
[44] S. Gorell, J. Andrews, J. Browning, Rationalizing the concept of net-to-gross [68] R. Bi, A. Zidane, A. Firoozabadi, Efficient and robust stability analysis in the in-
with respect to reservoir heterogeneity and flow behavior utilizing machine ternal energy, volume, and moles (UVN) space, Fluid Phase Equilib. 512 (2020)
learning analyses, SPE Eastern Regional Meeting, OnePetro, 2019. 112468.
[45] A. Almasov, M. Onur, A.C. Reynolds, Production optimization of the CO2 huf- [69] J.K. Singh, S.K. Kwak, Surface tension and vapor-liquid phase coexistence of
f-n-puff process in an unconventional reservoir using a machine learning confined square-well fluid, J. Chem. Phys. 126 (2) (2007) 024702.
based proxy, in: SPE Improved Oil Recovery Conference, OnePetro, 2020. [70] G.A. Chapela, S.E. Martínez-Casas, C. Varea, Square well orthobaric densities
[46] L. Cadei, G. Rossi, M. Montini, P. Fier, D. Milana, A. Corneo, L. Lancia, D. Lof- via spinodal decomposition, J. Chem. Phys. 86 (10) (1987) 5683–5688.
freno, E. Purlalli, F. Carducci, et al., Machine learning advanced algorithm to [71] D.D. Carley, Thermodynamic properties of a square-well fluid in the liquid and
enhance production optimization: an ann proxy modelling approach, in: Inter- vapor regions, J. Chem. Phys. 78 (9) (1983) 5776–5781.
national Petroleum Technology Conference, OnePetro, 2020. [72] T.L. Hill, An Introduction to Statistical Thermodynamics, Courier Corporation,
[47] I. Al Selaiti, C. Mata, L. Saputelli, D. Badmaev, Y. Alatrach, E. Rubio, R. Mohan, 1986.
D. Quijada, Robust data driven well performance optimization assisted by ma- [73] R. Hosein, R. Dawe, A parametric methodology in tuning the Peng-Robinson
chine learning techniques for natural flowing and gas-lift wells in Abu Dhabi, (PR) equation of state for gas condensate systems, Pet. Sci. Technol. 32 (6)
in: SPE Annual Technical Conference and Exhibition, OnePetro, 2020. (2014) 662–672.
[48] O. Sola-Aremu, An inferable machine learning approach to predicting PVT [74] X. Xu, H. Chen, C. Liu, C. Dang, Prediction of the binary interaction parameter
properties of niger delta crude oil using compositional data, in: SPE Annual of carbon dioxide/alkanes mixtures in the pseudocritical region, ACS Omega 4
Technical Conference and Exhibition, OnePetro, 2019. (8) (2019) 13279–13294.
[49] S.D. Groven, C. Desgranges, J. Delhommelle, Prediction of the boiling and crit- [75] Ø. Fevang, K. Singh, C.H. Whitson, et al., Guidelines for choosing compositional
ical points of polycyclic aromatic hydrocarbons via wang-landau simulations and black-oil models for volatile oil and gas-condensate reservoirs, in: SPE
and machine learning, Fluid Phase Equilib. 484 (2019) 225–231. Annual Technical Conference and Exhibition, Society of Petroleum Engineers,
[50] K. Wang, J. Luo, Y. Wei, K. Wu, J. Li, Z. Chen, Practical application of machine 20 0 0.
learning on fast phase equilibrium calculations in compositional reservoir sim- [76] Y.-K. Li, L.X. Nghiem, A. Siu, Phase behaviour computations for reservoir fluids:
ulations, J. Comput. Phys. 401 (2020) 109013. effect of pseudo-components on phase diagrams and simulation results, J. Can.
[51] Y. Wu, S. Misra, C. Sondergeld, M. Curtis, J. Jernigen, Machine learning for lo- Pet. Technol. 24 (06) (1985).
cating organic matter and pores in scanning electron microscopy images of [77] T. Newley, R. Merrill, Pseudocomponent selection for compositional simulation,
organic-rich shales, Fuel 253 (2019) 662–676. SPE Reservoir Eng. 6 (04) (1991) 490–496.
[52] R. Akkurt, T.T. Conroy, D. Psaila, A. Paxton, J. Low, P. Spaans, Accelerating and [78] S. Luo, J.L. Lutkenhaus, H. Nasrabadi, Multiscale fluid-phase-behavior simula-
enhancing petrophysical analysis with machine learning: a case study of an tion in shale reservoirs using a pore-size-dependent equation of state, SPE
automated system for well log outlier detection and reconstruction, SPWLA Reservoir Eval. Eng. 21 (04) (2018) 806–820.
59th Annual Logging Symposium, OnePetro, 2018. [79] M. Stein, Large sample properties of simulations using latin hypercube sam-
[53] Y. Wang, H. Liu, M. Guo, X. Shen, B. Han, Y. Zhou, Image recognition model pling, Technometrics 29 (2) (1987) 143–151.
based on deep learning for remaining oil recognition from visualization exper- [80] V. Nair, G.E. Hinton, Rectified linear units improve restricted Boltzmann ma-
iment, Fuel 291 (2021) 120216. chines, Icml, 2010.
[54] S. Mohanty, Estimation of vapour liquid equilibria for the system carbon [81] D.P. Kingma, J. Ba, Adam: a method for stochastic optimization,
dioxide–difluoromethane using artificial neural networks, Int. J. Refrig. 29 (2) arXiv preprint arXiv:1412.6980(2014).
(2006) 243–249. [82] K. Abou-Moustafa, C. Szepesvári, An a priori exponential tail bound for k-folds
[55] R. Sharma, D. Singhal, R. Ghosh, A. Dwivedi, Potential applications of artifi- cross-validation, arXiv preprint arXiv:1706.05801(2017).
cial neural networks to thermodynamics: vapor–liquid equilibrium predictions, [83] B.G. Marcot, A.M. Hanea, What is an optimal value of k in k-fold cross-valida-
Comput. Chem. Eng. 23 (3) (1999) 385–390. tion in discrete Bayesian network analysis? Comput. Stat. (2020) 1–23.
[56] R. Petersen, A. Fredenslund, P. Rasmussen, Artificial neural networks as a [84] M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics informed deep learning
predictive tool for vapor-liquid equilibrium, Comput. Chem. Eng. 18 (1994) (part i): Data-driven solutions of nonlinear partial differential equations, arXiv
S63–S67. preprint arXiv:1711.10561 (2017).
[57] V.D. Nguyen, R.R. Tan, Y. Brondial, T. Fuchino, Prediction of vapor–liquid equi- [85] H. Klie, B. Yan, A. Klie, Transfer learning for scalable optimization of unconven-
librium data for ternary systems using artificial neural networks, Fluid Phase tional field operations, in: Unconventional Resources Technology Conference,
Equilib. 254 (1–2) (2007) 188–197. 20–22 July 2020, Unconventional Resources Technology Conference (URTEC),
[58] V. Gaganis, N. Varotsis, An integrated approach for rapid phase behavior cal- 2020, pp. 2315–2333.
culations in compositional modeling, J. Pet. Sci. Eng. 118 (2014) 74–87. [86] P. Sarma, S. Kyriacou, M. Henning, P. Orland, G. Thakur, D. Sloss, Redistribu-
[59] C.J. Burges, A tutorial on support vector machines for pattern recognition, Data tion of steam injection in heavy oil reservoir management to improve EOR
Min. Knowl. Discov. 2 (2) (1998) 121–167. economics, powered by a unique integration of reservoir physics and machine
[60] H.H. Rachford, J. Rice, Procedure for use of electronic digital computers in learning, in: SPE Latin America and Caribbean Petroleum Engineering Confer-
calculating flash vaporization hydrocarbon equilibrium, J. Pet. Technol. 4 (10) ence, OnePetro, 2017.
(1952). 19 [87] F. Chollet, et al., Keras: the python deep learning library, Astrophys. Source
[61] A. Kashinath, M.L. Szulczewski, A.H. Dogru, A fast algorithm for calculating Code Library (2018) ascl–1806.
isothermal phase behavior using machine learning, Fluid Phase Equilib. 465 [88] M.L. Michelsen, W. Yan, E.H. Stenby, A comparative study of reduced-vari-
(2018) 73–82. ables-based flash and conventional flash, SPE J. 18 (05) (2013) 952–959.

12