Petroleum 4 (2018) 178e186

Contents lists available at ScienceDirect

Petroleum
journal homepage: www.keaipublishing.com/en/journals/petlm

Development of new correlations for the oil formation volume factor

in oil reservoirs using artificial intelligent white box technique
Salaheldin Elkatatny, Mohamed Mahmoud*
Petroleum Engineering Department, College of Petroleum and Geosciences, King Fahd University of Petroleum & Minerals, 31261 Dhahran, Saudi Arabia

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

Article history: Oil formation volume factor (OFVF) is considered one of the main parameters required to characterize the
Received 1 January 2017 crude oil. OFVF is needed in reservoir simulation and prediction of the oil reservoir performance. Existing
Received in revised form correlations apply for specific oils and cannot be extended to other oil types. In addition, big errors were
27 August 2017
obtained when we applied existing correlations to predict the OFVF. There is a massive need to have a
Accepted 27 September 2017
global OFVF correlation that can be used for different oils with less error.
The objective of this paper is to develop a new empirical correlation for oil formation volume factor
Keywords:
(OFVF) prediction using artificial intelligent techniques (AI) such as; artificial neural network (ANN),
Oil formation volume factor
Artificial intelligent
adaptive neuro-fuzzy inference system (ANFIS), and support vector machine (SVM). For the first time we
Reservoir management changed the ANN model to a white box by extracting the weights and the biases from AI models and
Artificial neural network form a new empirical equation for OFVF prediction. In this paper we present a new empirical correlation
extracted from ANN based on 760 experimental data points for different oils with different compositions.
The results obtained showed that the ANN model yielded the highest correlation coefficient (0.997)
and lowest average absolute error (less than 1%) for OFVF prediction as a function of the specific gravity
of gas, the dissolved gas to oil ratio, the oil specific gravity, and the temperature of the reservoir
compared with ANFIS and SVM. The developed empirical equation from the ANN model outperformed
the previous empirical correlations and AI models for OFVF prediction. It can be used to predict the OFVF
with a high accuracy.
© 2018 Southwest Petroleum University. Production and hosting by Elsevier B.V. on behalf of KeAi
Communications Co., Ltd. This is an open access article under the CC BY-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction volume factor is a strong function of the reservoir pressure and in

The properties of the reservoir oil are crucial in reservoir engi-
neering computational applications. These properties include; oil
formation volume factor, bubble point pressure, gas solubility in oil,
etc. The oil formation volume factor (Bo) determines the ratio be-
tween the oil volume in the reservoir with the dissolved gas to the
volume of the oil at the surface (stock tank). All material balance
calculations need the oil formation volume factor to determine the
reservoir volume after the depletion process. The oil formation
* Corresponding author.
E-mail address: mmahmoud@kfupm.edu.sa (M. Mahmoud).
Peer review under responsibility of Southwest Petroleum University.
Production and Hosting by Elsevier on behalf of KeAi
certain cases the data for the oil formation volume factor (OFVF) is
not available. Correlations and models were developed to predict
the FVF that can be used in reservoir computational models.
Reservoir engineering applications such as material balance
equation, reservoir simulation, and well testing need the PVT
properties of the petroleum reservoir fluids and one of these
properties is the oil formation volume factor.
Several empirical correlations were developed to predict the
OFVF for specific regions and for specific oils. Labedi [1] came up
with an empirical correlation to determine the OFVF at the bubble
point pressure. He developed the correlation using 129 data sets.
Al-Marhoun [2] developed a correlation for the FVF at the bubble
point pressure. He used 11,728 measured values to develop the
correlation for 700 reservoirs from North America and Middle East.
He used the least square linear regression method to develop the
correlation and he obtained the following correlation:

https://doi.org/10.1016/j.petlm.2017.09.009
2405-6561/© 2018 Southwest Petroleum University. Production and hosting by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under
the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 S. Elkatatny, M. Mahmoud / Petroleum 4 (2018) 178e186 179


 develop the new empirical equation for OFVF prediction.

gg T  60
Bob ¼ 1 þ a1 Rs þ a2 Rs þ a3 Rs þ a4 ðT  60Þ (1)
go 1  go
1.1. Artificial intelligent techniques
where Rs is the solution gas oil ratio, go is the oil gravity, gg is the
Artificial Neural network is the most powerful statistical tool to
gas gravity, T is the reservoir temperature,  F, a1 ¼ 0.177342  103,
recognize and classify complex patterns and system which human
a2 ¼ 0.220163  103, a3 ¼ 4.292580  106, and
brain cannot do [17]; in fact the artificial neural network technique
a4 ¼ 0.0.528707  103. The average absolute percentage error of
is inspired from biological neurons that are found in human brain
the developed correlation to predict the FVF was as maximum as
[18].
0.5.
An ANN model consists of fundamental processing unit, termed
Vazquez and Beggs [3,4] developed empirical correlation for the
as neurons. The neural network models are structured on three
oil formation volume factor. They used 6004 data points from
components, learning algorithm, transfer function and network
different oil reservoirs worldwide. Glasø [5] used 41 data points for
architecture [19] (Lippmann 1987). The network model comprises
oil collected from the North Sea to develop a correlation for the oil
of at-least three layers, input layer, hidden layer and output layer.
formation volume factor.
Each layer connects with other layers with the help of weights. The
Macary and El-Batanoney [6] developed empirical correlations
network performance is solely based on the adjustment of weights
for the oil formation volume factor for 30 different oil reservoirs in
between these layers [20,21]. Hidden layers assigned with transfer
the Gulf of Suez area, Egypt. They used 90 data sets in their cor-
function usually ‘log-sigmoidal’ or ‘tan-sigmoidal’. Output layer is
relation. Dokla and Osman [7] generated differently correlation for
assigned with ‘pure linear’ activation function.
the UAE oils to determine the oil formation volume factor and
Adaptive neuro-fuzzy inference system (ANFIS) is also gained
bubble point pressure. They used 51 data sets to develop these
dominant importance in petroleum industry. Many researchers
correlations. Omar and Todd [8] proposed correlations for the for-
used ANFIS to delineate complex concepts in the petroleum in-
mation volume factor and bubble point pressure based on Standing
dustry [22,23]. ANFIS is the combination of neural network and
[9] correlation for oil reservoirs in Malaysia. They developed the
fuzzy logic and its very robust supervised learning technique. It is
correlations using 93 data sets. Kartoatmodjo and Schmidt [10,11]
the kind of neural network that uses Sugeno fuzzy inference system
developed several correlations for different PVT properties
[24]. ANFIS has the capability to extract the benefits of both
(including the oil formation volume factor) using 5392 data sets
mentioned AI techniques in single platform. In order to get best out
collected from different oil reservoirs worldwide. Al-Mehaideb [12]
of this technique one should use any evolutionary algorithm to
developed set of correlations for PVT properties of the UAE oils
optimize the parameters of ANFIS [25].
using data from UAE oil reservoirs, he used 62 data sets. Petrosky
Fuzzy logic maps input parameters to input membership func-
and Farshad [13] developed different PVT correlations for Gulf of
tions, converting input membership functions to set of fuzzy rules,
Mexico crude oils and they used 90 data sets to develop their
converting set of fuzzy rules to output characteristics, then convert
correlations.
output characteristics to output membership functions and finally
Gharbi and Elsharkawy [14] introduced a neural network model
this membership function to one valued output or any classification
to predict the oil formation volume factor. They collected 498
based on output [26]. In ANFIS instead of just fixing the shape of
experimental data points for different oil sample from the Middle
membership function, it automatically assigned the type and shape
East. They introduced a black box model that can accurately predict
of membership function by analyzing the data [27].
the oil formation volume factor.
Support Vector machine is the type of supervised learning that
Al-Marhoun and Osman [15] developed new models to predict
is mostly used for regression and pattern recognition purposes
the oil formation volumed factor at the bubble point pressure using
[28,29]. Based on soft margin hyper-plane support vector machine
artificial neural networks (ANN). They developed the models for
have been introduced as new artificial intelligence tool framework
Saudi oil reservoirs based on 283 data sets measured for oil samples
for both classification and function approximation [30,31]. Instead
from the field. The models predicted the oil formation volume
of sigmoidal type transfer function like in artificial neural network,
factor with high accuracy and the absolute relative error was 5.9%.
support vector machine stands on the kernel neuron function
They used 142 data sets to train the ANN model, 71 data sets for
which definitely allows projection to higher planes and able to
validation, and 70 data sets for testing the developed models. Their
solve more complicated and complex highly nonlinear problems
model was a black box and they did not generate empirical corre-
[32]. SVM applications can be found in many fields like medical,
lations out of the ANN models. Osman and Al-Marhoun [16] used
business, civil and electrical [33].
the Multi-Layer-Preceptor (MLP) and Radial Basis Function (RBF)
neurula networks to develop models for PVT for oil and brine
samples. Their models predicted the PVT properties accurately 1.2. Application of artificial intelligence in petroleum engineering
compared to the published correlations but they did not introduce
mathematical equations, they only presented black box models. Artificial intelligence models were applied in many areas in
The objective of this research is to develop a new empirical petroleum engineering. Semi-clathrate hydrate pressure of carbon
correlation to determine the oil formation volume factor (OFVF) dioxide, methane, nitrogen, and hydrogen sulfide in the presence of
based on the specific gravity of gas (gg), the dissolved gas to oil ratio TBAB ionic was obtained using SVM and coupling of SVM with
(Rs), the oil specific gravity (API), and the temperature of the genetic algorithm (GA-SVM). The obtained correlation coefficient
reservoir (Tf). Three artificial intelligent techniques were used to was 0.97759 and 0.99944 for SVM and GA-SVM, respectively be-
develop the OFVF model such as; artificial neural network (ANN), tween the experimental and predicted values [34]. The clathrate
adaptive neuro-fuzzy inference system (ANFIS), and support vector hydrate formation pressure of carbon dioxide in the presence of
machine (SVM). The obtained results of the three models will be 1,4-dioxine was predicted using ANFIS with a correlation coefficient
compared and the one who will gave the highest correlation co- of 0.9969 and a mean square error of 0.0034, [35].
efficient and lowest average absolute error percent will be used to The moisture content of natural gas dried by calcium chloride
change the AI to a white box by extracting the weight and biases to dehydrator units was predicted using least squares support vector
180 S. Elkatatny, M. Mahmoud / Petroleum 4 (2018) 178e186

outliers and applying normalization technique if required.

Fig. 1. Flowchart for the methodology.
machines (LSSVM) coupling with genetic algorithm (GA-LSSVM)
with a coefficient of determination of 0.9986 and a mean square
error of 2.7898 [36]. The same technique GA-LSSVM was used to
determine the water dew point of a natural gas stream in equilib-
rium with a triethylene glycol (TEG) solution at different condition
[37].
Feed-forward artificial neural network (ANN) was optimized
using different techniques and was used to predict asphaltene
precipitation with high accuracy [38e41].
Least square support vector machine (LS-SVM) was developed
to predict minimum miscible pressure (MMP) [42] and to predict
CO2ebrine solubility [43].
Hydrogen Sulphide solubility in various ionic liquids was pre-
dicted with a high accuracy using LS-SVM [44] and also using
artificial neural networks (ANNs) trained with back-propagation
(BP) and particle swarm optimization (PSO) [45].
The dew point pressure of retrograde condensate gas reservoir
was estimated using different AI tools [46e48].
Reservoir porosity and permeability were determined with a
high degree of accuracy using artificial neural network [49,50] and
s least square support vector machine [51].
The combustion front velocity in heavy oil recovery process was
specified using least-squares support vector machine [52,53].
Monitoring different gases (CO2/O2/CO) was performed using a
robust intelligence based approaches such as artificial neural
network (ANN) and hybrid methods [54].
The second step is to determine the relative importance be-
tween the input parameters and the target to obtain initial sug-
gestion about the main input parameters that can be used to
predict the target.
The third step is to apply artificial intelligence technique. In this
step, several sensitivity analyses are performed to select the pa-
rameters for AI techniques, in addition select the number of input
parameters.
The fourth step is to optimize the AI models by optimizing the
number of neurons based on the selected algorithm and the main
input parameters.
The fifth step is to apply the AI model for a case study and select
the best AI technique for this application, step six. Followed by field
validation of the best AI technique, step seven and finally, we
generalize the model and recommend it for future work. Fig. 1
shows the flow chart of the methodology.
2.1. Data description and analysis
Table 1 lists a sample of the collected data for different crudes
(760 data points). This data was collected from published
papers [4,5,8,55e57]. This data included different oil sources with
different concentrations. Data from Middle East, Al-Marhoun
(1988) [56], data from Malaysian Crudes, Omar and Todd (1993)
[8], data from North Sea Glasø (1980) [5], data from fields all over
the world. Vazquez and Beggs (1980) [4], and data from Mediter-
ranean Basin, Africa, the Persian Gulf and the North Sea, Ghetto
(1994) [57].
The data includes oil formation volume factor (OFVF), the spe-
cific gravity of gas (gg), the dissolved gas to oil ratio (Rs), the oil
specific gravity (API), and the temperature of the reservoir (Tf). The
dissolved gas to oil ratio had a range of 2207 with a minimum value

2. Methodology

The first step in our methodology is the data acquisition in

which the data is collected from different sources. Data processing
step includes the quality check of the available data, removing the Fig. 2. Relative importance of input parameters to oil formation volume factor.

Table 1
Collected data from published papers.

Sample ID Input Parameters Output Parameter

Solution Gas Oil Ratio, scf/bbl Gas Gravity Oil Gravity,  API Reservoir Temperature,  F Oil Formation Volume Factor, OFVF, bbl/STB

1 494 0.677 44.5 230 1.315

2 267 0.884 31.4 174 1.173
3 956 0.811 43.2 226 1.538
4 242 0.824 31.4 180 1.17
5 494 0.677 44.5 230 1.315
6 1355 0.877 48.8 228 1.843
7 1440 0.788 50.3 250 1.954
. . . . . .
. . . . . .
759 0.767 37.4 193 1.577 0.767
760 0.946 39.2 210 1.596 0.946
 S. Elkatatny, M. Mahmoud / Petroleum 4 (2018) 178e186 181

Table 2
Statistical analysis of the training data.

Statistical Parameter Input Parameters Output Parameter

Solution Gas Oil Ratio, scf/bbl Gas Gravity Oil Gravity,  API Reservoir Temperature,  F Oil Formation Volume Factor, bbl/STB

Minimum 9.00 0.59 15.30 74.00 1.03

Maximum 2216.00 1.37 59.50 290.00 2.21
Range 2207.00 0.78 44.20 216.00 1.18
Mean 536.99 0.92 35.00 167.71 1.32
Mode 585.00 0.77 31.20 100.00 1.39
Standard Error 18.20 0.01 0.31 2.11 0.01
Standard Deviation 419.80 0.17 7.12 48.57 0.23
Sample Variance 176233.48 0.03 50.63 2359.38 0.05
Kurtosis 1.24 0.04 0.20 0.80 1.98
Skewness 1.08 0.74 0.14 0.10 1.33

of 9 and a maximum value of 2216. The specific gravity of gas had a
range of 0.778 with a minimum value of 0.589 and a maximum
value of 1.367. The oil specific gravity had a range of 44.2 with a
minimum value of 15.3 and a maximum value of 59.5. The reservoir
temperature had a range of 220 with a minimum value of 74  F and
a maximum value of 294  F. The OFVF had a range of 1.182 with a
minimum value of 1.028 rbbl/STbbl and a maximum value of 2.21
rbbl/STbbl.
The first step in data analysis is to assess the relative importance
between the input parameters (the specific gravity of gas (gg), the
dissolved gas to oil ratio (Rs), the oil specific gravity (API), and the
temperature of the reservoir (Tf)) and the output parameter (OFVF).
Fig. 2 shows that the OFVF was a strong function of the dissolved
gas to oil ratio, where the correlation coefficient was found to be
0.97. OFVF was a moderate function of the oil specific gravity (API)
as the correlation coefficient was 0.59. The correlation coefficient of
the OFVF was 0.38 and 0.28 for the temperature of the reservoir
and the specific gravity of gas, respectively, indicating the poor
relation between these parameters and OFVF.
2.2. Artificial intelligent models
2.2.1. Seating parameters for AI techniques
A back propagation neural network algorithm was implemented
to model OFVF. The proposed ANN model was based on the specific
gravity of gas (gg), the dissolved gas to oil ratio (Rs), the oil specific
gravity (API), and the temperature of the reservoir (Tf), with one
hidden layer and one output parameter. The number of neurons in
the hidden layer were varied between 5 and 50. The optimum
number of neurons was found to be 20. Tan-sigmoidal type acti-
vation function was used as a transfer function between input and
hidden layer and linear type activation function was used between
hidden and output layer. Levenberg-Marquardt back propagation
algorithm was selected as the training algorithm to obtain the
weights and biases. In order to avoid the model to stuck on local
minima 10,000 realizations were performed with the initialization
of different weights and biases during training and cross-validation
phases of the modeling. After training, the weights and biases from
the optimum model were extracted.
ANFIS model was based on Genfis2 (Subtractive clustering) type
with cluster radius 0.5. Optimum cluster radius size was found by
running the sensitivity of cluster radii between 0.1 and 1. The op-
timum epoch size was found to be 500. In order to avoid the model
to stuck on local minima 10,000 realizations were performed.
SVM model was based upon Gaussian type kernel function with
regularization parameter C of value 5000. ‘Gaussian type’ kernel
function was selected by comparing the prediction performance
with ‘Polynomial type’ kernel function. Since ‘Gaussian type’ kernel

Fig. 3. Prediction of oil formation volume factor (OFVF) by three AI techniques for the training data.
182 S. Elkatatny, M. Mahmoud / Petroleum 4 (2018) 178e186

250 and 5000.

2.2.2. Building the artificial intelligent models

The first step in building the AI model is to normalize all the
parameters that will be used to build the AI model. The value of the
parameters (input and output) is normalized between 1 and 1 by
using two points slope, Eqs. (2) and (3).

Y  Ymin X  Xmin
¼ (2)
Ymax  Ymin Xmax  Xmin

X  Xmin
Y¼ ð2Þ  1 (3)
Xmax  Xmin

where; Y is the input parameter in the normalized form, Ymax ¼ 1,

Fig. 4. Coefficient of determination for OFVF prediction using AI techniques for the
training data.
function predicted better results than polynomial type kernel
function. Optimum value of regularization parameter C was found
out to be 5000 by performing the sensitivity of C values between
Ymin ¼ 1, Xmax is the maximum value of input data, Xmin is the
minimum value of input data, X is the input parameter to be
normalized. For example, the minimum value of the OFVF (Xmin) is
1.028 rbbl/STbbl, and the maximum value of OFVF (Xmax) is 2.21
rbbl/STbbl, so for the value of OFVF equal to 1.85 rbbl/STbbl, the
normalized value will equal to 0.39.
The second step is to train the model. Seventy percent of the
data (532 data points) was used to train the AI models. Table 2 lists

Table 3
Statistical analysis of the Testing data.

Statistical Parameter Input Parameters Output Parameter

Solution Gas Oil Ratio, scf/bbl Gas Gravity Oil Gravity,  API Reservoir Temperature,  F Oil Formation Volume Factor, bbl/STB

Minimum 12.00 0.59 17.30 80.00 1.03

Maximum 1660.00 1.37 53.40 278.00 2.00
Range 1648.00 0.78 36.10 198.00 0.97
Mean 555.85 0.91 35.64 166.23 1.33
Mode 61.00 0.80 41.40 100.00 1.16
Standard Error 27.27 0.01 0.49 3.19 0.01
Standard Deviation 409.94 0.15 7.34 48.01 0.22
Sample Variance 168049.93 0.02 53.92 2304.76 0.05
Kurtosis 0.17 0.58 0.57 0.73 0.09
Skewness 0.76 0.86 0.00 0.22 0.91

Fig. 5. Prediction of oil formation volume factor using three AI techniques for the testing data.
 S. Elkatatny, M. Mahmoud / Petroleum 4 (2018) 178e186 183

Fig. 6. Coefficient of determination for OFVF prediction using AI techniques for the
testing data.
the statistical analysis of the trained data. Fig. 3 shows that the ANN
model was able to predict the OFVF with a correlation coefficient
(R) of 0.997 and an average absolute error of 0.85% when comparing
the actual and predicted value of the OFVF. ANFIS was able to
predict the OFVF with a correlation coefficient (R) of 0.996 and an
average absolute error of 0.97 when comparing the actual and the
predicted values of the OFVF. While SVM was able to predict the
BPP with a correlation coefficient (R) of 0.964 and an average ab-
solute error of 3.35% when comparing the actual and predicted
values of the OFVF, Fig. 3.
Fig. 4 shows that for the training data (532 data points), the ANN
yielded higher coefficient of determination (R2) of 1.0 for predicting
the OFVF than ANFIS and SVM where the R2 was 0.99 and 0.93,
respectively.
The third step is to assess the developed AI models by testing the
three models using unseen data (228 data points), which is 30% of
the collected data. Table 3 lists the statistical analysis of the testing
data. It is clear that the data range of the input parameters of the
testing data is within the range of the training data to be sure of the
accuracy of the developed model.
Fig. 5 shows that, ANN model was able to predict the OFVF with
a correlation coefficient of 0.997 and an average absolute error of
0.93%. ANFIS was able to predict the OFVF with a correlation co-
efficient of 0.993 and an average absolute error of 1.16%. While SVM
was able to predict the OFVF with a correlation coefficient of 0.960
and an average absolute error of 3.67%, Fig. 5. Fig. 6 shows that for
the testing data (228 data points), the ANN yielded a coefficient of
determination (R2) of 0.99 for predicting the OFVF and ANFIS and
SVM yielded R2 of 0.99 and 0.92, respectively.
It can be concluded that ANN model was able to predict the
OFVF based on the specific gravity of gas (gg), the dissolved gas to
oil ratio (Rs), the oil specific gravity (API), and the temperature of
the reservoir (Tf) with higher accuracy than ANFIS and SVM. The
correlation coefficient was around 0.997 and the average absolute
error was less than 1.0%. Based on these results, ANN model was
selected to develop a new empirical correlation for predicting the
OFVF and change the ANN model to a white box by extracting the
weights and the biases from the model to develop the mathemat-
ical equation.
2.2.3. Development of empirical equation using artificial neural
network
The mathematical model was driven from the artificial neural
network model by extracting the weights associated with input
layer/hidden layers and hidden layer/outer layer and the biases of
the hidden layer and the output layer. The weights between input
layer and hidden layer termed as w1, the weights between hidden
layer and outer layer termed as w2, the bias of the hidden layer
termed as b1, and the bias of the output layer terms as b2, are given
in Table 4.
Eq. (4) can be used to predict the BPP in normalized form as a
function of the specific gravity of gas (gg), the dissolved gas to oil
ratio (Rs), the oil specific gravity (API), and the temperature of the
reservoir (Tf). To obtain the de-normalized form of the OFVF, Eq. (5)
can be used.

Table 4
Weights and biases for OFVF prediction, Eq. (4).

Input Layer Weight Matrix Input Layer Bias Vector Hidden Layer Weight Vector Output Layer Bias Vector

W1 b1 W2 b2

1 2 3 4

2.1593 1.5149 0.0852 0.1906 0.9474 3.2883 0.2684

0.7850 0.1607 2.2100 0.1100 0.2113 1.2696
1.7225 0.2184 3.0365 0.9670 1.5727 2.8570
2.5881 0.6721 1.6742 0.7588 0.7756 2.4502
2.0772 1.0964 0.7072 0.6274 0.0502 2.3060
1.3760 0.2392 1.9517 2.4830 1.3455 2.7692
1.7120 1.5807 0.8155 1.8059 0.3180 1.4524
0.0613 0.1051 2.0990 1.7291 0.0445 0.2137
1.7336 1.0141 0.9776 0.8812 0.2818 0.6828
1.1465 1.5530 1.0866 1.9053 0.0254 0.3363
1.3604 0.5688 0.5973 1.4252 0.4832 0.7771
1.5185 2.0828 0.1229 2.0578 0.1216 0.8256
0.0906 1.4082 2.5519 1.2860 0.0511 0.1820
1.6835 1.6969 0.0089 1.6219 0.1811 0.8942
1.5940 0.4728 0.6624 0.1171 0.3483 1.1129
0.9380 1.4633 2.0629 1.4521 0.1847 1.6013
1.5683 0.9960 0.5413 0.5240 0.1861 1.4197
0.3586 0.2499 1.2845 2.2493 0.1026 2.4283
2.2021 0.1181 2.0524 0.3511 0.0615 2.3963
1.4285 1.0958 1.8721 3.1844 0.7279 3.8435
184 S. Elkatatny, M. Mahmoud / Petroleum 4 (2018) 178e186

2 0 13
6X
N
B 
2
C 7
OFVFn ¼4 w2i @ A 5 þ b2 (4)
i¼1 2 w1i;1 Rsn þ w1i;2 GGn þ w1i;3 APIn þ w1i;4 Tf n þ b1i
1þe

was optimized to be 20 to give the highest accuracy between the

OFVF ¼ 0:78*OFVF þ 1:808 (5) estimated and actual values of the bubble point pressure.

where; Rsn is the normalized value of the solution gas oil ratio, GGn
is the normalized value of the gas specific gravity, APIn is the
normalized value of the oil gravity, Tfn is the normalized value of
the reservoir temperature, N is the number of neurons (the number
of neurons should be optimized to have good match with less er-
ror); W1 is weight of hidden layer; W2 is weight of the output layer;
b1 is bias of the hidden layer, and b2 is bias of the output layer.
Table 4 lists the input parameters for Eq. (4). The number of neuron
2.2.4. Validation of the developed mathematical equation
To assess the developed equation, 30% of the data (228 data
points) which was unseen by ANN model was used to calculate the
OFVF using Eq. (5). Fig. 7 shows that the developed equation (Eq.
(5)) was able to predict the OFVF with a coefficient of determination
(R2) of 0.995 when comparing the actual and calculated values of
the OFVF and the average absolute error was about 0.86%.
Table 5 lists the input parameters for 20 data point that were
used for further validation [14]. The developed empirical equation
was compare with previous ANN model [14] and other empirical
equations [4,5,9,and56]. Table 6 list the results of the comparison.
Fig. 8 confirms that the developed empirical equation for OFVF
prediction yielded the lowest average absolute error (0.99%) when
compared with the previous models.
Fig. 9 shows that Eq. (5) was able to predict the OFVF with a
coefficient of determination (R2) of 0.99 for unseen data published
by Gharbi and Elsharkawy [14], while Gharbi (1999) [14] yielded a
coefficient of determination (R2) of 0.97, Al-Marhoon (1988) [56]
yielded a coefficient of determination (R2) of 0.94, Al-Marhoon
(1992) [4] yielded a coefficient of determination (R2) of 0.96,
Standing (1947) [9] yielded a coefficient of determination (R2) of
0.89, and Glasø (1980) [5] yielded a coefficient of determination
(R2) of 0.89.

3. Conclusions

Fig. 7. Determination of oil formation volume factor using the developed imperial Three AI models were developed to estimate the OFVF as a
correlation. function of the specific gravity of gas (gg), the dissolved gas to oil

Table 5
Input data for comparison with previous models [14].

Sample ID Input Parameters Output Parameter

Solution Gas Oil Ratio, scf/bbl Gas Gravity Oil Gravity,  API Reservoir Temperature,  F Oil Formation Volume Factor, OFVF, bbl/STB

1 104 1.126 27.49 160 1.10

2 433 1.188 31.14 100 1.22
3 305 0.918 36.95 153 1.22
4 631 1.047 33.42 244 1.49
5 580 0.802 38.16 100 1.26
6 408 0.856 37.76 240 1.39
7 466 0.972 38.98 239 1.36
8 464 0.836 35.56 220 1.29
9 730 0.68 45.60 232 1.67
10 604 0.669 36.55 180 1.33
11 811 0.812 36.55 185 1.45
12 1039 0.91 33.03 193 1.61
13 1172 0.758 52.03 291 2.01
14 1393 0.9 43.84 306 2.03
15 1052 0.767 37.35 193 1.58
16 891 0.975 28.03 200 1.38
17 1452 1.034 47.61 249 1.90
18 1784 0.827 36.15 231 2.06
19 2216 0.909 33.61 250 2.16
20 1745 0.708 41.69 288 2.10
 S. Elkatatny, M. Mahmoud / Petroleum 4 (2018) 178e186 185

Table 6
Comparison of the developed empirical correlation for OFVF prediction with previous models.

# Actual OFVF, bbl/STB New Eq. Gharbi (1999) Al-Marhoon (1988) Al-Marhoon (1992) Standing (1947) Glasø (1980)

1 1.10 1.10 1.088 1.105 1.09 1.1 1.074

2 1.22 1.23 1.213 1.228 1.18 1.259 1.239
3 1.22 1.20 1.24 1.201 1.18 1.195 1.166
4 1.49 1.48 1.483 1.439 1.39 1.466 1.424
5 1.26 1.26 1.311 1.272 1.27 1.292 1.271
6 1.39 1.39 1.355 1.321 1.31 1.303 1.26
7 1.36 1.36 1.395 1.361 1.33 1.353 1.311
8 1.29 1.33 1.348 1.323 1.32 1.315 1.276
9 1.67 1.67 1.53 1.452 1.52 1.443 1.4
10 1.33 1.33 1.364 1.328 1.39 1.331 1.295
11 1.45 1.45 1.464 1.444 1.46 1.48 1.446
12 1.61 1.61 1.555 1.558 1.55 1.648 1.6
13 2.01 2.00 2.032 1.757 1.86 1.781 1.717
14 2.03 2.04 1.973 1.882 1.94 1.989 1.906
15 1.58 1.57 1.598 1.549 1.61 1.611 1.571
16 1.38 1.42 1.484 1.49 1.47 1.573 1.536
17 1.90 1.92 2.017 1.938 1.85 2.066 1.987
18 2.06 1.95 2.064 1.923 2.01 2.12 2.029
19 2.16 2.16 2.275 2.161 2.21 2.482 2.327
20 2.10 2.06 2.113 1.943 2.19 2.074 1.978

(1) Artificial neural network is the best AI technique that can

predict the OFVF as a function the specific gravity of gas (gg),
the dissolved gas to oil ratio (Rs), the oil specific gravity (API),
and the temperature of the reservoir (Tf).
(2) ANN model has a correlation coefficient of 0.997 and an
average absolute error less than 1.0%.
(3) The developed empirical equation from the ANN model
outperformed the previous models for OFVF prediction.
(4) The developed correlation of the OFVF from the ANN model
can be used to predict the OFVF with high accuracy
(R2 ¼ 0.994 and the average absolute error less than 1%).

The ANN model was change to a white box by developing the

empirical correlation for the OFVF. This will eliminate the need for
special software or equipment to run the model. Excel spread sheet
can be used to solve the model for bubble point pressure prediction.
This development will help the reservoir engineer better manage
Fig. 8. The developed equation yields the lowest average absolute error for OFVF
the reservoir and predict the OFVF.
prediction as compared with the previous models.

Nomenclature

OFVF The oil formation volume factor, bbl/STB

GG The Gas gravity
RS The dissolved gas to oil ratio, scf/bbl
API The oil gravity
Tf The temperature of the reservoir,  F

References

Fig. 9. Prediction of OFVF for the published data (data taken form Gharbi and
Elsharkawy [14]).
ratio (Rs), the oil specific gravity (API), and the temperature of the
reservoir (Tf). Based on the results obtained, the following con-
clusions can be drawn:
[1] R. Labedi, Use of production data to estimate volume factor density and
compressibility of reservoir fluids, J. Pet. Sci. Eng. 4 (1990) (1990) 375e390.
[2] M.A. Al-Marhoun, New correlation for formation volume factors of oil and gas
mixtures, J. Can. Pet. Technol. 31 (3) (1992) 22e26.
[3] M.E. Vazquez, Correlations for Fluid Physical Property Predictions, MS Thesis,
The University of Tulsa, Tulsa, Oklahoma, 1976.
[4] M.E. Vazquez, H.D. Beggs, Correlations for fluid physical property prediction,
J. Pet. Technol. 32 (6) (1980) 968e970. SPE-6719-PA.
[5] O. Glasø, Generalized pressure-volume-temperature correlations, J. Pet.
Technol. 32 (5) (1980) 785e795. SPE-8016-PA.
[6] S.M. Macary, M.H. El-Batanoney, Derivation of PVT correlations for the Gulf of
Suez crude oils, in: Paper Presented at the EGPC 11th Petroleum Exploration
and Production Conference, 1992 (Cairo, Egypt).
[7] M. Dokla, M. Osman, Correlation of PVT properties for UAE crudes, SPE Form.
Eval. J. 7 (1) (1992) 41e46. SPE-20989-PA.
[8] M.I. Omar, A.C. Todd, Development of new modified black oil correlation for
malaysian crudes, in: Paper SPE 25338 Presented at the 1993 SPE Asia Pacific
Oil and Gas Conference and Exhibition, 1993, pp. 8e10. Singapore, Feb.
186 S. Elkatatny, M. Mahmoud / Petroleum 4 (2018) 178e186
[9] M.B. Standing, A pressure-volume-temperature correlation for mixtures of
California oils and gases, Drill Prod. Pract., API (1947) 275e287, 1947.
[10] T. Kartoatmodjo, Z. Schmidt, New correlations for crude oil physical proper-
ties, 1991. Paper SPE 23556.
[11] T. Kartoatmodjo, Z. Schmidt, Large data bank improves crude physical prop-
erty correlations, Oil Gas J. (1994) 51. July 1994.
[12] M.R.A. Almehaideb, Correlations for UAE crude oils, in: Paper SPE 37691
Presented at the SPE Middle East Oil Show and Conference, 1997, pp. 15e18.
Bahrain, March.
[13] J. Petrosky, F. Farshad, Pressure volume temperature correlation for the Gulf of
Mexico, in: Paper SPE 26644 Presented at the 1993 SPE Annual Technical
Conference and Exhibition, 1993, pp. 3e6. Houston, TX, Oct.
[14] R.B. Gharbi, A.M. Elsharkawy, Neural network model for estimating the PVT
properties of Middle East crude oils, SPE Reserv. Eval. Eng. 2 (3) (1999)
255e263.
[15] M.A. Al-Marhoun, E.A. Osman, Using Artificial Neural Networks to Develop
New PVT Correlations for Saudi Crude Oils, in: Paper SPE 78592 Presented at
the 10th Abu Dhabi International Petroleum Exhibition and Conference
(ADIPEC), October 8-11, 2002. Abu Dhabi, UAE.
[16] E.-S.A. Osman, M.A. Al-Marhoun, Artificial neural networks models for pre-
dicting PVT properties of oil field brines, in: Paper SPE 93765-MS Presented at
the SPE Middle East Oil and Gas Show and Conference, 2005, pp. 12e15.
March, Kingdom of Bahrain.

[17] A.A.D. Castillo, E. Santoyo, O. Garcia-Valladare, A new void fraction correlation
inferred from artificial neural networks for modeling two-phase flows in
geothermal wells, Comput. Geosci. 41 (2012) 25e39.
[18] M. Graves, S. Liwicki, R. Fernandez, H. Bertolami, J. Bunke, Schmidhuber,
A novel connectionist system for improved unconstrained handwriting
recognition, IEEE Trans. Pattern Anal. Mach. Intell. 31 (5) (2009) 855e868.
[19] R.P. Lippmann, An introduction to computing with neural nets, IEEE ASSP
Mag. 4 (2) (1987) 4e22.
[20] G.E. Hinton, S. Osindero, Y. Teh, A fast learning algorithm for deep belief nets,
Neural Comput. 18 (7) (2006) 1527e1554.
[21] S. Mohaghegh, S. Ameri, Artificial neural network as a valuable tool for pe-
troleum engineers, 1995. SPE Pap. 29220-prepared A. T. unsolidated Pap. SPE.
[22] M.H. Rammay, A. Abdulraheem, Automated history matching using combi-
nation of adaptive neuro fuzzy system (ANFIS) and differential evolution al-
gorithm, in: Paper SPE 172992 Presented at the SPE Large Scale Computing
and Big Data Challenges in Reservoir Simulation Conference and Exhibition
Held in Istanbul, 2014. Turkey.
[23] H.A. Nooruddin, F. Anifowose, A. Abdulraheem, Applying artificial intelligence
techniques to develop permeability predictive models using mercury injec-
tion capillary-pressure data, in: Paper SPE 168109 Presented at SPE Saudi
Arabia Section Technical Symposium and Exhibition, 2013. Al-Khobar, Saudi
Arabia.
[24] J. Jang, S R. ANFIS: adaptive-network-based fuzzy inference system, IEEE
Trans. Syst. Man Cybern. (3) (1993) 23, https://doi.org/10.1109/21.256541.
[25] P.A. Tahmasebi, Hybrid neural networks-fuzzy logic-genetic algorithm for
grade estimation, Comput. Geosci. 42 (2012) 18e27.
[26] G. Klir, B. Yuan, Fuzzy Sets and Fuzzy Logic: Theory and Applications, Prentice-
Hall, Englewood Cliffs, 1995.
[27] H. Kaydani, A. Mohebbi, A. Baghaie, Neural fuzzy system development for the
prediction of permeability from wireline data based on fuzzy clustering, J. Pet.
Sci. Eng. 30 (19) (2012) 2036e2045.
[28] R. Burbidge, M. Trotter, B. Buxto, Drug design by machine learning: support
vector machines for pharmaceutical data analysis, Comput. Chem. 26 (1)
(2001) 5e14.
[29] A.H. Ben-Hur, S.H. David, V. Vapnik, Support vector clustering, J. Mach. Learn.
Res. 2 (2001) 125e137.
[30] J.T. Jeng, C.C. Chuang, S.F. Su, Support vector interval regression networks for
interval regression analysis, Fuzzy Sets Syst. 138 (2) (2003) 283e300.
[31] H. William, S.A. Teukolsky, W.T. Vetterling, B.P. Flannery, Support Vector
Machines. Numerical Recipes: the Art of Scientific Computing, third ed.,
Cambridge University Press, New York, 2007. ISBN 978-0-521-88068-8.
[32] K. Trontl, T. Smuc, D. Pevec, Support vector regression model for the esti-
mation of g-ray buildup factors for multi-layer shields, Ann. Nucl. Energy 34
(12) (2007) 939e952.
[33] R. Gholami, A.R. Shahraki, P.M. Jamali, Prediction of hydrocarbon reservoirs
permeability using support vector machine, Math. Problems Eng. (2012),
https://doi.org/10.1155/2012/670723.
[34] A. Baghbana, M.A. Ahmadi, B. Pouladi, B. Amanna, Phase equilibrium modeling
of semi-clathrate hydrates of seven commonly gases in the presence of TBAB
ionic liquid promoter based on a low parameter connectionist technique,
J. Supercrit. Fluids 101 (2015) 184e192, 2015.
[35] M.A. Ahmadi, M. Lee, A. Bahadori, Prediction of a solid desiccant dehydrator
performance using least squares support vector machines algorithm, J. Taiwan
Inst. Chem. Eng. 50 (2015) 115e122, 2015.
[36] M.A. Ahmadi, M. Ebadi, A. Samadi, M.Z. Siuki, Phase equilibrium modeling of
clathrate hydrates of carbon dioxide þ 1,4-dioxine using intelligent ap-
proaches, J. Dispersion Sci. Technol. 36 (2015) 236e244, 2015.
[37] M.A. Ahmadi, A. Bahadori, Prediction performance of natural gas dehydration
units for water removal efficiency using a least square support vector ma-
chine, Int. J. Ambient Energy 37 (5) (2016) 486e494.
[38] M.A. Ahmadi, S.R. Shadizadeh, New approach for prediction of asphaltene
precipitation due to natural depletion by using evolutionary algorithm
concept, J. Fuel 102 (2012) 716e723, 2012.
[39] M.A. Ahmadi, Neural network based unified particle swarm optimization for
prediction of asphaltene precipitation, Fluid Phase Equilib. 314 (2012) 46e51,
2012.
[40] M.A. Ahmadi, M. Golshadi, Neural network based swarm concept for predic-
tion asphaltene precipitation due natural depletion, J. Pet. Sci. Eng. 98e99
(2012) 40e49, 2012.
[41] Prediction of asphaltene precipitation using artificial neural network opti-
mized by imperialist competitive algorithm, J. Pet. Explor. Prod. Technol. 1
(2011) 99e106.
[42] M.A. Ahmadi, M. Mohammad Zahedzadeh, S.R. Seyed Reza Shadizadeh,
R. Abbassi, Connectionist model for predicting minimum gas miscibility
pressure: application to gas injection process, Fuel 148 (2015) 202e211, 2015.
[43] M.A. Ahmadi, A. Ahmadi, Applying a sophisticated approach to predict CO2
solubility in brines: application to CO2 sequestration, Int. J. Low-Carbon
Technol. 11 (2016) 325e332.
[44] M.A. Ahmadi, B. Pouladi, Y. Javvi, S. Alfkhani, R. Soleimani, Connectionist
technique estimates H2S solubility in ionic liquids through a low parameter
approach, J. Supercrit. Fluids 97 (2015) 81e87, 2015.
[45] A. Shafiei, M.A. Ahmadi, S.H. Zaheri, A. Baghbana, A. Amirfakhrian,
R. Soleimani, Estimating hydrogen sulfide solubility in ionic liquids using a
machine learning approach, J. Supercrit. Fluids (2014) 525e534, 2014.
[46] A.M. Ahmadi, A. Ebadi, Robust intelligent tool for estimation dew point
pressure in retrograded condensate gas reservoirs: application of particle
swarm optimization, J. Pet. Sci. Eng. 123C (2014) 5e17, 2014.
[47] A.M. Ahmadi, A. Elsharkawy, Robust correlation to predict dew point pressure
of gas condensate reservoirs, Petroleum (2016), https://doi.org/10.1016/
j.petlm.2016.05.001, 2016.
[48] A.M. Ahmadi, A. Ebadi, Evolving smart approach for determination dew point
pressure through condensate gas reservoirs, Fuel 117 (2014) 1074e1084,
2014.
[49] S.M. Elkatatny, M.A. Mahmoud, T. Zeeshan, A. Abdulraheem, New insights into
the prediction of heterogeneous carbonate reservoir permeability from well
logs using artificial intelligent network, Neural Comput. Appl. (2017), https://
doi.org/10.1007/s00521-017-2850-x, 2017.
[50] M.A. Ahmadi, B. Pouladi, Y. Javvi, S. Alfkhani, R. Soleimani, Connectionist
approach estimates gas-oil relative permeability in petroleum reservoirs:
application to reservoir simulation, Fuel 140 (2015) 429e439, 2015.
[51] M.A. Ahmadi, M.R. Ahmadi, S.M. Hosseini, M. Ebadi, Connectionist model
predicts porosity and permeability of petroleum reservoirs by means of petro-
physical logs: application of artificial intelligence, J. Pet. Sci. Eng. 123 (2014)
183e200, 2014.
[52] M.A. Ahmadi, M. Masoumi, R. Askarinezhad, Evolving smart model to predict
combustion front velocity throughout in-situ combustion process employ-
ment, Energy Technol. (3) (2015) 128e135, 2015.
[53] M.A. Ahmadi, M. Masoumi, R. Askarinezhad, Evolving connectionist model to
monitor efficiency of the in-situ combustion process: application to heavy oil
recovery, J. Energy Technol. 2 (9e10) (2014) 811e818.
[54] M.A. Ahmadi, M. Masoumi, R. Kharrat, H. Amir, A.H. Mohammad, Gas analysis
by in situ combustion in heavy oil recovery process: experimental and
modeling studies, J. Chem. Eng. Technol. 37 (3) (2014) 409e418.
[55] D.L. Katz, Prediction of shrinkage of crude oils, Drill Prod. Pract., API (1942)
137e147.
[56] M.A. Al-Marhoun, PVT correlations for Middle East crude oils, J. Pet. Technol.
40 (5) (1988) 650e666.
[57] G. Ghetto, D.F. Aone, M. Villa, Reliability Analysis on PVT Correlation, Paper
SPE 28904 presented at The SPE European Petroleum Conference, London, UK,
1994. October 25-27.