Journal of Applied Geophysics 183 (2020) 104207

Contents lists available at ScienceDirect

Journal of Applied Geophysics

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

A new approach for porosity and permeability prediction from well logs
using artificial neural network and curve fitting techniques: A case study
of Niger Delta, Nigeria
Job G. Urang ⁎, Ebong D. Ebong, Anthony E. Akpan, Emmanuel I. Akaerue
Applied Geophysics Programme, Physics Department, University of Calabar, PMB 1115, Calabar, Cross River State, Nigeria

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

Article history: A new approach based on artificial neural network and standardized non-linear regression analysis is proposed
Received 27 February 2020 for the prediction of reservoir properties from limited well log data. Two separate multi-layer-perceptron neural
Received in revised form 28 October 2020 networks using the Bayesian framework were employed and mathematical models were developed from the
Accepted 29 October 2020
networks. These neural networks were used to predict reservoir properties from well log data. The porosity neu-
Available online 04 November 2020
ral network (NN) utilized density log (RHOB) as input and core porosity as target, while the permeability NN
Keywords:
employed RHOB and water saturation as input and core derived permeability as target. Mean square errors
ANN (MSEs) for training, validation and test for the porosity network were 1.20889e−6, 2.22574e−6, and
Curve fitting 1.37204e−6 respectively while 5.62849e−1, 3.13709e−1, and 4.83728e−1 were obtained for the permeability net-
Porosity work. These networks were later applied to four wells to predict porosity and permeability. Very good correlation
Permeability coefficients (R-values) between the NN predicted data and the target data were obtained in each of these wells.
Prediction For the porosity network R-values of 0.99987, 0.99099, 0.99474 and 0.83749 were obtained, whereas 0.97584,
Niger Delta-Nigeria 0.83594, 0.97002 and 0.83512 respectively were observed from the permeability network. For the model devel-
opment, standardized nonlinear regression curve fitting tool was utilized. The polynomial model that determines
the coefficients of the fitted model through the Gauss-Newton optimization method was used. Several criteria
such as robust fitting, normalization and scaling were employed to further reduce and randomize the errors so
that it follows a Gaussian distribution with zero mean and constant variance. Goodness of the fit was tested
with the observable statistical parameters. For the porosity model, the sum of squared error (SSE) of 2.2715
e−04, R-square of 0.9997, Adjusted R-square of 0.9997 and root mean square error (RSME) of 4.3562 e−04 was
observed while the permeability model, SSE of 7.8315e01, R-square of 0.9753, Adjusted R-square of 0.9752 and
RSME of 8.092 was observed. This approach provides a simple and cost effective way of estimating reservoir
properties.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction (Saemi et al., 2007). Even though this method has proven to give accu-
rate results, it is not totally reliable as information derived from such
Porosity and permeability are important reservoir properties that methods are not continious. More so, only selected depths in wells are
plays key role in hydrocarbon and groundwater exploration. These cored in a particular field and only few wells are sampled, since the pro-
properties are directly related to the amount of fluid that can be stored cess is time consuming and very expensive (Martley, 2000). Conse-
in the pore volume and those that have the capacity to flow. Thus, accu- quently, the need for a method that can provide a continious and
rate evaluation of these properties can enhance production perfor- wider range of information is inevitable. Data from wireline logs tend
mance of the reservoir (Aminian and Ameri, 2005). Additionally, these to proffer solution to the problem of lack of continuity in core informa-
properties are utilized during reservoir modelling and simulation in tion and are almost available for all wells in a field. This provides an in-
the development of static and dynamic models and are often used as in- direct method of estimating reservoir properties. However, the pitfall in
puts during reserve estimation and design of hydrocarbon recovery pro- this method is its dependence on empirical relationships that may not
cess (Madj and Hezarkhani, 2011). Conventionally, these reservoir represent the true reservoir conditions (Helle et al., 2001; Lim, 2005).
properties can be determined directly from core sample analysis Other problems associated with the wireline logs include, missing
data during preservation period and missing information during log-
⁎ Corresponding author. ging, due to factors such as excessive temperature, pressure and corro-
E-mail address: joburang@unical.edu.ng (J.G. Urang). siveness of subsurface media, tool malfunctioning and operator errors

https://doi.org/10.1016/j.jappgeo.2020.104207
0926-9851/© 2020 Elsevier B.V. All rights reserved.
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

Field C
Well 4
Field A
Well 3

Well 1

Field B
Well 2

4oE 8oE Sokoto

12oN Basin Borno Basin
LEGEND OPUAMA
CHANNEL Basement
Connental plaorm
(upper deltaic plain) Complex
6oN 6oN Upper
Plaorm turbides Bida Benue
(lower deltaic plain) Trough
Basin Abuja
Turbides environment 8oN Middle
Benue
Afam & Opuama Ibadan Lower Trough
Channels Benue
Trough
Kwa-Iboe collapse
High gas risk zone BELT Lagos Niger Republic of
(GOR>1000) (Deep play) Delta Cameroun
Mobile shale
4oN a Basin
-BELT 0 200 km
Paleo-front Delta
Onshore and Offshore
main structural limits 4oE 8oE 12oE
Compressive front/belt OFFSHORE Terary volcanics
EQUATORIAL
GUINEA Quaternary and Terary sediments

Cretaceous sediments
BIOCO
ISLAND
DEEP - OFFSHORE Basement Complex

3oN 3oN
Thrust front belt

b 0 100 km
LTA

4oE 8oE

Fig. 1. Geologic map of Nigeria showing the Niger Delta Basin (Adapted from Ebong et al., 2017) (a) and sectional map of the Niger Delta depo-belts, structural limits and study location
(b) (Redrawn from Doust and Omatsola, 1990).

that limits logging capabilities, thereby creating gaps in sensor record- et al. (2001), motivated by the results of ANN modelling of permeability
ing (Ayoub and Mohamed, 2015; Lopes and Jorge, 2018). These gaps in offshore Eastern Canada, developed neural nets for converting well
can hinder the process of building representative stratigraphic models logs in the North Sea to porosity and permeability. The porosity and per-
and creating uncertainties regarding what to expect during the next meability were estimated using two separate back-propagation algo-
drilling program (Churikov and Grafeeva, 2018). Hence, the need to rithms. Verma et al. (2012) applied the ANN technique to estimate
seek alternative ways of addressing these problems is highly recom-
mended. Our study combines core information and well log data to gen-
erate suitable neural network models that can help to overcome these
problems, since the model can mimic the true subsurface conditions
down the hole.
Recently, artificial neural network (ANN) have been proven to be a
strong tool for predicting reservoir properties from well log data as it
can identify complex relationships among permeability, porosity, fluid
saturations, depositional environments, lithology and well log data
(Rezaee et al., 2007; Wang et al., 2013; Anifowose et al., 2017). Over
the years, tremendous success has been recorded in the application of
ANN to solve problems such as completion analysis, formation damage
prediction, field development and reservoir characterization in the oil
and gas industries (Madj and Hezarkhani, 2011; Mohaghegh et al.,
1996). Neural network presents an alternative and easier method of es-
timating these reservoir properties (Hamada and Elshafei, 2010). Helle Fig. 2. A schematic diagram of a multilayered perception network.

2
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

Fig. 3. Neural network architecture for permeability prediction.

permeability and porosity of the formation using few well log data in Al- in equal proportions. However, the upper portion is mostly sand with
berta Deep basin in Canada. Aminian and Ameri (2005) utilized the ANN only minor shale interbeds (Kogbe, 1976). At the base of the Delta is
technique to predict the flow units and flow unit permeability of a het- the marine Akata Formation, dominated by thick shale sequences (po-
erogeneous reservoir from well log and limited core data. Wong et al. tential source rock), turbidite sand (potential reservoirs in deep
(2002) used back-propagation network method to compare the predic- water), and minor amounts of clay and silt. Beginning in the Paleocene
tion of porosity from genetic and non genetic approach of reservoir and through Recent, the Akata Formation was formed during lowstands
characterization in oil field NSW of Australia. These authors have uti- when terrestrial organic matter and clays were transported to deep
lized several inputs (i.e., three and above) in arriving at suitable models water areas characterized by low energy conditions and oxygen defi-
for their predictions. However, we are proposing a novel approach that ciency. The formation underlies the entire delta, and is usually
utilizes a maximum of two inputs for permeability and one for porosity overpressured. The petroleum system recognized in the Niger Delta is
predictions that can provide reliable results in areas where data are lim- known as the Tertiary Niger Delta (Agbada-Akata) Petroleum System.
ited. This study is aimed at developing a simple neural network and This is because nearly all the oil produced from the area comes from
mathematical models that can predict porosity and permeability. the Agbada Formation and partly from the Akata Formation in deep off-
shore fields.
2. Geology of the study area

3. Working principle of ANN

The Niger Delta covers an area of 300,000km2 with average sedi-
ment thickness of over 10 km in the basin depo-centre (Kaplan et al.,
The ANN is a functional imitation of the natural biological neurons
1994). It has prograded south-westward from Eocene to present, lead-
which is its fundamental data processing element. Like the biological
ing to the development of depo-centres (Doust and Omatsola, 1990).
neurons in the brain, it is made up of simple but highly interconnected
It is among the most prolific and largest hydrocarbon basins in the
processors also called neurons. Fundamentally, a biological neuron take
world. The Niger Delta is subdivided into five onshore and shelf
sub-basins, termed depo-belts that range in age from Eocene to Plio-
Pleistocene (Fig. 1). The depo-belts include the Northern Delta (Late Eo-
Table 1
cene – Middle Miocene), Greater Ughelli (Oligocene – Early Miocene),
Summary statistics of the trained porosity network using neutron and density log as input
Central Swamp I & II (Middle – Late Miocene), Coastal Swamp I & II and core porosity as target.
(Late Miocene - Plioceene), and Offshore (Middle – Late Pliocene)
Sample division No of sample Percentage (%) of sample R-value MSE
(Steele et al., 2009). The present day Niger Delta is characterized by
three major lithostratigraphic units that portray three main cycles of de- Training 839 70 0.99818 1.51025e−9
positional sequence in the region (Kogbe, 1976; Koledoye et al., 2003). Validation 180 15 0.99609 6.93251e−9
Testing 180 15 0.99245 2.03184e−9
The uppermost unit is the Benin Formation -a continental Latet Eocene
to Recent deposits of alluvial and upper coastal plain sands that are up to
2000 m thick in some locations. The sands and sandstones are coarse to
fine, grannular in texture, partly unconsolidated and represent braided
Table 2
stream point bars and channel fills and/or crevasse splay deposit. The
Summary statistics of the trained porosity network using only density log as input and
shales are few, thin and may represent back swamp deposits. Underly- core porosity as target.
ing the Benin Formation is the Eocene to Recent Agbada Formation, the
Sample division No of sample Percentage (%) of sample R-value MSE
major reservoir rock in the Niger Delta. The formation consists of paralic
siliciclastic sequences and is over 3700 m thick. The clastics accumu- Training 839 70 0.99999 1.20889e−6
lated in delta-front, delta-topset, and fluvio-deltaic environments. In Validation 180 15 0.99906 2.22574e−6
Testing 180 15 0.99997 1.37204e−6
the lower Agbada Formation, shale and sandstone beds were deposited

3
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

Error Histogram with 20 Bins

1200
Training
Validation
1000 Test
Zero Error

800
Instances
600

400

200

0
-0.01199
-0.01128
-0.01056
-0.00985
-0.00914
-0.00842
-0.00771

0.000127
-0.007
-0.00629
-0.00557
-0.00486
-0.00415
-0.00344
-0.00272
-0.00201
-0.0013
-0.00059

0.00084
0.001553
Errors = Targets - Outputs

Fig. 4. Error histogram for the porosity network.

inputs from different sources, puts them together, performs a nonlinear between the target (i.e., desired output) and the output is computed
operation on the input data and brings out the result (Lim, 2005). and back propagated through the network by the adjustment of weights
Weighted links are the source of neurons connection and signals are and biases (Akpan et al., 2013). This process is repeated by means of it-
passed from one neuron to another by these connections. A number of erations until a minimum error is obtained. A single iteration according
input signals are received by each neuron through its connections and to Hagan et al. (1996) is given by Eq (1).
transmitted through outgoing neuron connections (i.e., axon in biolog-
ical neuron) as output. The strength of each neuron input is demon- Z kþ1 ¼ Z k −α k g k ð1Þ
strated by the weights and its adjustment recurrently, leads to the
where gk is the current gradient, Zk is a vector of current weights and
neural network learning process. Basically, there are two types of neural
biases and αk is the learning rate. In other to adjust the connection
networks, the unsupervised and supervised. In the unsupervised net-
weights of a particular neuron i at a certain iteration p, Negnevitsky
work the training algorithm only requires the input that is classified
(2005) reported the process as shown in Eq. (2).
into groups and clusters to be fed into the network. The learning process
goes on without the true output being shown to the network wi ðp þ 1Þ ¼ wi ðpÞ þ Δwi ðpÞ ð2Þ
(Ramgulam, 2006). Aminian and Ameri (2005) recognized Kohonen
network as the most common unsupervised network that has a simple where wi(p) is the weight of the ith neuron at p iteration, and Δwi(p) is
architecture of one layer input and output. As with all unsupervised net- the weight correction factor that can be calculated using Eq. (3).
works, the Kohonen network is provided with the input data and learns
Δwi ðpÞ ¼ αxi ðpÞeðpÞ ð3Þ
without being shown the true output. The supervised networks recog-
nize patterns and make decisions based on the patterns of inputs and
outputs fed to it (e.g., back-propagation network). Providing the desired
result is an integral part of the learning/training process. During the
training process, the input data is exclusively divided into test, valida- Table 3
Summary statistics of the trained permeability network using density log and water satu-
tion and training sets. The development of the network is done using
ration as input and core derived permeability as target.
the training set, the validation set is used to measure generalization of
the network (Saemi et al., 2007), while the ability of the trained net- Sample division No of sample Percentage (%) of sample R-value MSE
work to predict events, is confirmed using the test set (Aminian and Training 839 70 0.99589 5.62849e−1
Ameri, 2005). In the oil and gas industry the supervised network is Validation 180 15 0.96306 3.13709e−1
mostly used due to its ability to solve complex problems (Mohaghegh, Testing 180 15 0.99164 4.83728e−1

2000; Ramgulam, 2006). Its advantage over the unsupervised network

lies in its capacity to predict more reliable and repeatable results from
new set of data.
Table 4
Summary statistics of the trained permeability network using only water saturation as in-
put and core derived permeability as target.
3.1. Backpropagation algorithm
Sample No of Percentage (%) of R-value MSE
division sample sample
Backpropagation refers to the way a multilayered network com-
putes its error gradient from performance function (Hagan et al., Training 839 70 0.60915 3393.6768
1996). As a supervised network, a pair of input and target presented Validation 180 15 0.49852 16,318.90239
Testing 180 15 0.53346 4010.7591
to it, is propagated through it and an output is generated. The error

4
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

Error Histogram with 20 Bins

1000 Zero Error

900

800

700
Instances 600

500

400

300

200

100

0
-0.00964
-0.03644
-0.02304

0.003759
0.01716
0.03056
0.04396
0.05736
0.07076
0.08416
0.09755
0.111
0.1244
0.1378
0.1512
0.1646
0.178
0.1913
0.2047
0.2181
Errors

Fig. 5. Error histogram for the permeability network.

Table 5 wi ðpÞ ¼ α γi ðpÞδk ðpÞ ð4Þ

Training parameters of the selected porosity and permeability network.

Parameter Porosity network Permeability network where δk(p) is the error gradient at neuron k in the output layer at
Time 10s 125 s
iteration p. Eq. (4) is called the delta rule. Backpropagation algorithm
Epoch (No. of iterations) 16 1000 are of different types; Quasi-Newton, Conjugate gradient, Levenberg-
Best validation performance 1.1354 e−06 3.0271e−04 Marquardt algorithms, Gradient descent with momentum, and Gradi-
Mu 1 e−10 0.01 ent descent (Hagan et al., 1996). The last two algorithms are usually
Gradient 1.4154e−05 0.19508
slow, while the first three are comparatively faster in solving prob-
lems. In this work, the Levenberg-Marquardt algorithm was used
Alternatively, the expression for Δwi(p) for the jth neuron in a hid- for both networks because it is reputed for generating results faster
den layer γi according to Akpan et al. (2013) and Negnevitsky (2005) when used to train moderate-sized feed forward neural network
can be expressed as (Hagan et al., 1996).

0
Gradient = 1.4154e-05, at epoch 16
10
gradient

-5
10

0
Mu = 1e-10, at epoch 16
10

-5
mu

10

-10
10

Validation Checks = 6, at epoch 16

10
val fail

0
0 2 4 6 8 10 12 14 16
16 Epochs

Fig. 6. Neural network training state for the porosity network.

5
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

Best Validation Performance is 1.1354e-06 at epoch 10

Train
-2 Validation
10
Test

Mean Squared Error (mse)

Best

-4
10

-6
10

-8
10

0 2 4 6 8 10 12 14 16
16 Epochs

Fig. 7. Performance plot for the porosity network.

4. Materials and methods Development Company (SPDC) of Nigeria was used (Fig. 1b). A base
map that shows Field A with two wells, Field B and C with one well
4.1. Materials each is provided. Fields A and B are located within the Coastal Swamp
Depo-belt while Field C is located within the Central Swamp Depo-
For this work, datasets from four wells (Wells 1, 2, 3 and 4) from belt (Fig. 1b). A dataset consisting of core porosity (ϕ), derived core per-
three different fields within the Niger Delta operated by Shell Petroleum meability (K), neutron log, density log (RHOB) log and water saturation

Training: R=0.99999 Validation: R=0.99906

Output ~= 0.99*Target + 0.0038

Data
Output ~= 1*Target + 0.0011

Data
0.4 Fit 0.4 Fit
Y=T Y=T
0.35 0.35

0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15

0.15 0.2 0.25 0.3 0.35 0.4 0.15 0.2 0.25 0.3 0.35 0.4
Target Target

Test: R=0.99997 All: R=0.99987

Output ~= 0.99*Target + 0.0038

Output ~= 0.99*Target + 0.0017

Data Data
0.4 Fit 0.4 Fit
Y=T Y=T
0.35 0.35

0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15

0.15 0.2 0.25 0.3 0.35 0.4 0.15 0.2 0.25 0.3 0.35 0.4
Target Target

Fig. 8. Regression plot for the training data set of the porosity network.

6
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

10
Gradient = 0.19506, at epoch 1000
10

gradient
0
10

-10
10

-2
Mu = 0.01, at epoch 1000
10

-3

mu
10

-4
10

Validation Checks = 0, at epoch 1000

1
val fail

-1
0 100 200 300 400 500 600 700 800 900 1000
1000 Epochs

Fig. 9. Neural network training state for the permeability network.

(Sw) from Well 1 belonging to Field A, was used in training the net- consists of a hidden layer with 10 neurons linking the input layer and
works. The permeability is described as derived permeability (K) be- 1 neuron linking the output layer for both networks (Fig. 2). Two trans-
cause it was calculated using core porosity and water saturation based fer functions (i.e., tan-sigmoid and linear) were used. The tan-sigmoid
on Wyllie and Rose (1950) empirical relationship (Eq. (5)). transfer function was used at the input layer to constrain the output
values to fall within the range of −1 to +1 since the given input values
aϕ4:4 ranged from negative to positive infinity (Fig. 3). At the output layer, the
K¼ ð5Þ
Sbwirr linear transfer function was used so that the network's output can take
on any value. In designing the porosity network both density (RHOB)
where the tortuosity factor (a) is 250, saturation exponent (b) is 2, ϕ is and neutron log data were used as input in the first instance (Fig. 3),
porosity and Swirr is irreducible water saturation. while only density (RHOB) was used in the second scenario as input.
However, core porosity was used as target for both networks. Three
4.2. Network design models were designed for the permeability network, the first having
neutron log and water saturation as input, the second utilized density
Every neural network model have basically three types of layers; the (RHOB) and water saturation as input (Fig. 3) and the third had only
input, hidden and output layers. However, the network design can vary water saturation as input with core derived permeability serving as
based on input conditions, layer architecture (i.e., number of hidden the target for the three networks. For each of the networks, a total of
layers and number of neurons on each layer) and the desired output 1199 vectors of the input and target datasets were randomly divided
(Zhang et al., 2018). In this work, both the permeability and porosity into 839 vectors representing 70% for training, 180 vectors representing
network have similar design, except that each varies essentially with 15% for testing and 180 vectors representing 15% for validation of the
input parameters and the final output. The network architecture networks. After training the networks, one network each was selected

Best Validation Performance is 0.00030271 at epoch 1000

6 Train
10
Validation
Test
Best
Mean Squared Error (mse)

4
10

2
10

0
10

-2
10

-4
10

0 100 200 300 400 500 600 700 800 900 1000
1000 Epochs

Fig. 10. Performance plot for the permeability network.

7
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

Training: R=0.99589 Validation: R=0.96306

3000
Data Data

Output ~= 1.3*Target + -68

4000

Output ~= 1*Target + -5.2

Fit Fit
2500
Y = T Y = T

2000 3000

1500
2000

1000
1000
500

0 0
0 1000 2000 3000 0 1000 2000 3000 4000
Target Target

Test: R=0.99164 All: R=0.97584

1000
Data Data

Output ~= 1.1*Target + -22

4000
Output ~= 1*Target + -8.4

Fit Fit
800
Y = T Y = T
3000
600

400 2000

200 1000

0
0
0 200 400 600 800 1000 0 1000 2000 3000 4000
Target Target

Fig. 11. Regression plot for the permeability network.

Table 6 and Wang, 2008). The general problem with this technique is that of
Correlation coefficient (R-value) for both porosity and permeability in all the wells. finding a member of the family of curves or surfaces that can “best” fit
Well no. R-value for porosity R-value for permeability a given number of data points. This criterion finds application in orthog-
onal distance regression, where the sum of squares of the orthogonal
1 0.99987 0.97584
2 0.99099 0.83598 distances from the data points to the surface is minimized (Atieg and
3 0.99474 0.97002 Watson, 2003). It is practically a classical optimization problem that
4 0.83749 0.83512 can be solved using optimization techniques such as decent methods
and Gauss-Newton methods. For this kind of fitting problem, the idea
is to minimize the sum of squares of the orthogonal distances from
the data points to the surface using the Gauss–Newton iteration tech-
for the porosity and permeability prediction in the four wells. The selec-
niques that forces orthogonality to hold at every iteration point (Atieg
tion was done based on the following criteria; (i) where there are lim-
and Watson, 2003). In the field of Meteorology and Pattern recognition,
ited data, the network can still be used to predict the reservoir
the least squares orthogonal distance technique is preferred in solving
properties and (ii) if the chosen model has a very strong correlation
fitting problems because it is iterative and fast in bringing out the rela-
with the desired output. Based on these conditions, the trained network
tionship between shape parameters and location parameters especially
with water saturation and density as input and core derived permeabil-
when the data points are many (Liu and Wang, 2008). Given a family of
ity as target was selected for the permeability model. The porosity
parametric curves or surfaces represented by a function F(L; tj) ⊂ Rs, for
model has only density as input and core porosity as the target. In
a set of points {Xj}nj=1 ⊂ Rs to be approximated by F(L; t), where t = (t1,
order to ascertain the efficacy of the ANN models, the ANN porosity
…,tm) ∈ Rm is the location parameter and L = (L1, …, Lr) is the shape pa-
values were compared with empirically calculated porosity values
rameter, the standard nonlinear least square representation of this
from density log and their differences with regards to the core porosity
problem is given by Eq. (6).
were plotted.
Min n

L, t1...tm
∑j¼1 ‖ F L; tj −Xj ‖2 ð6Þ
4.3. Model generation by curve fitting
where, tj is associated with the data points Xj. The traditional way of
The procedure described above in 4.1 was to develop a neural net- solving Eq. (6) involves two basic steps. The first step is parameteriza-
work model that can be used in predicting permeability and porosity tion that involves assigning the location parameter tj to each data point
that is similar to core derived porosity and permeability within the Xj. The second step involves solving a linear least squares problem to de-
Niger Delta with limited data. In other to generalise these models, math- termine the shape and location parameters (i.e., coefficients) if the
ematical models were generated from the robust neural network shape parameters occur in linear form. Where the problems are non-
models that can be used to calculate these reservoir properties else- linear, there can be linearized (Liu and Wang, 2008). By executing these
where. To achieve this, the curve fitting analysis technique that is two steps iteratively through computer algorithms, improved location
widely applied in computational sciences and engineering to solve and shape parameters are obtained easily (Liu and Wang, 2008). This
such problems (Liu and Wang, 2008) was utilized. It is a regression anal- is analogous to solving a system of first-order partial derivative of F
ysis tool used to identify relationship between variables whose para- (L; t) generated from Eq. (6) to determine the coefficient of the polyno-
metric form is known but parametric values are to be determined (Liu mial iteratively using the Gauss-Newton method which immediately

8
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

a b

d c

Fig. 12. Comparative analysis of fractional porosity derived from cores, ANN and empirical calculation for Well 1 (a), Well 2 (b), Well 3 (c) and Well 4 (d).

a b

d c

Fig. 13. Fraction porosity estimation error derived from the difference between core porosity and ANN predicted porosity and the difference between core porosity and empirically
calculated porosity from density log for Well 1 (a), Well 2 (b), Well 3 (c) and Well 4 (d).

9
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

a b

d c

Fig. 14. Comparison of core derived permeability and ANN predicted permeability for Well 1 (a), Well 2 (b), Well 3 (c) and Well 4 (d).

gives the correct values of the coefficients. Computer codes that are used additional scale factor (i.e., weight) is required in the fitting process.
in solving such problems are embedded in Matlab. These coefficients The choice of weight depends on the justification that either it follows
when determined are fitted into the desired polynomial function, gen- a particular form or by computation using a known variance as given
erally of the form in Eq. (8),

Gðx, yÞ ¼ a0 þ a1 x þ a2 y . . . . . . þ e ð7Þ 1
Wi ¼ ð8Þ
σ2
where x and y are two independent variable otherwise referred to as
predictor 1 and 2 and G(x,y) the dependent variable also known as out- where Wi is the weight attached to individual data points. If the variance
come (Shanock et al., 2010). Eq. (7) is a third order polynomial with two is not known, the weights can be computed using,
independent variables. In this work, we fitted a third order degree two  n −1
and a second order degree one polynomial from the NN model already 1
Wi ¼ ∑ ðyi −yÞ2 ð9Þ
developed for permeability and porosity prediction in the Niger Delta. n i¼0
The idea is to develop mathematical models that can easily be used to
calculate these reservoir properties, if inputs are fed into the models. where n is the number of replicate data, yi is the individual data points
This was achieved using the curve fitting tool. Since the permeability and y the mean of the data. Though, it is better to compute the weights
model involves two inputs, i.e., the independent variables and one out- using variance, it was observed that in fitting the permeability model,
put, i.e., the dependent variable, density was selected as the X-data and the weight at each data point followed a particular trend similar to the
water saturation as the Y-data. The NN predicted permeability was se- value of water saturation at that point. Thus, water saturation value
lected as the Z-data. We needed an empirical model that is simple and was used as the weight and this further reduced the error estimate.
can maximize the flexibility of the fitted data; hence, the polynomial For the porosity model, no weight was attached, since the model was
model was selected. Generally, polynomials of any degree can provide linear with a near perfect fit with minimal error. The main disadvantage
good fits within a data range but can diverge widely especially if higher of least squares fitting is that outliers have large influence on the fit be-
degree polynomials are involved in the fit. During extrapolation, higher cause squaring the residuals magnifies the effects of these extreme data
degree polynomials tend to be unstable and results in scaling problem points (Hadi and Sabh, 2014). To minimize the influence of outliers, the
(Hadi and Sabh, 2014). To handle this problem the data were normal- data was fitted using robust least squares regression that are basically of
ised by centring it at zero mean and scaling it to unit standard deviation. two types; Least absolute residuals (LAR) and Bisquare weights (BW).
When fitting data with random variations, it is assumed that error exists The LAR method finds a curvature that minimizes the absolute differ-
in the response data, not in the predictor data. Thus, the errors can be ence of the residuals, rather than the squared differences. Therefore, ex-
random and follow a normal (Gaussian) distribution with zero mean treme values have a lesser influence on the fit. Whereas the BW scheme
and constant variance (Chai and Draxler, 2014). This implies that, the minimizes a weighted sum of squares, where the weight given to each
response data should be of equal quality to give better fit. However, this data point depends on how far the point is from the fitted line. Points
is not usually the case as fits are most times unduly influenced due to near the line gets full weight, while points farther from the line are
poor data quality. To improve the fit and minimize error estimates, an given reduced weight. In most cases, the BW scheme is preferred over

10
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

LAR because it simultaneously seeks to find a curve that fits the bulk of in permeability different from Well 1 in the same field due to heteroge-
the data using the least squares approach, and minimizes the effect of neity in structural styles which controls permeability (i.e., dynamic
outliers. However, the LAR was used in fitting the permeability model property) and the location of the well within the boundary of two differ-
because it gave a better result in terms of minimizing the error. Having ent Depo-belts (i.e., Coastal and Central Swamp Depo-belts).
found a solution that satisfied all conditions of constraints and optimal- Having being satisfied with the performance of both networks,
ity, certain statistical indices were calculated. mathematical models were generated from the NN models used for
the predictions afterwards. Taking advantage of the linear inverse
5. Results and discussion relationship between porosity and bulk density (Figs. 15 and 16), the
porosity network was fitted with the polynomial model as shown in
The first NN porosity model has neutron and density (RHOB) as the the surface diagram (Fig. 17) and a linear relationship of the form
input and core porosity as the target output (Table 1). For the second NN
FðxÞ ¼ P1 x þ P2 ð10Þ
porosity model, only density log was used as the input with core poros-
ity as the target output (Table 2). Fig. 4 show an error histogram that il-
was generated. F(x) is the porosity and x the density log value, P1 and P2
lustrates the distribution of errors among the three data sets for the
are coefficients of the polynomial calculated to be P1 = −0.6026
selected porosity network. Similarly, the permeability network had
(−0.6032, −0.6021) and P2 =1.599 (1.597, 1.6) with 95% confidence
two models tested for the purpose of selecting a better network. The
bounds. The values in the brackets represent the lower and upper
NN permeability model with density (RHOB) and water saturation as
limits of the coefficients while the one outside the brackets is their
input provided a better result than the one with only water saturation
average. Hence, inputting the values of the coefficient, the model
as presented in Tables 3 and 4 respectively. The error histogram show-
for porosity becomes.
ing the distribution of the errors among the three data sets used for the
permeability network is displayed in Fig 5. The training of the selected ϕ ¼ −0:6026  þ1:599 ð12Þ
NN porosity model revealed gradient of 1.4154e−05, model uncertainty
(Mu) of 1e−10, validation checks of 6 after 16 epochs (Table 5) and best Eq. (12) is an empirical model generated from the network for estimat-
validation performance of 1.1354e−06 after 10 iterations (Figs. 6 and 7). ing porosity. The degrees of accuracy of this model have been presented
Although the MSE was relatively high, the second NN porosity model statistically in Table 7. The permeability network was also fitted using
was selected because the R-values for the training, validation and test the polynomial model (Fig. 17). A functional relationship with three co-
sets, i.e., 0.99999, 0.99906 and 0.99997 respectively (Table 2 and efficients was observed (Eq. (13)).
Fig. 8) were very close to unity and meets our desired objective of cre-
ating a NN that can utilize very limited data to predict porosity. This im-
plies good in-sample fit of the model. The MSE for the training and
Sw RHOB Por (Frac.)
testing sets are almost of the same magnitude, hence no evidence of
underfitting and/or overfitting of the networks was observed 0.00 1.00 0.00 g/cm3 2.50 0.00 ft3/ft3 0.50
(Koerhsen, 2018). Also, the selected NN permeability model for the
6200

training state show gradient of 0.19508, Mu of 0.01, validation checks

of 0 at 1000 epochs (Table 5) and best validation performance of
3.0271e−04 at 1000 epochs (Figs. 9 and 10). Based on the MSE and the
R-value (Table 3 and Fig. 11), the first NN permeability model was cho-
6300

sen. Chai and Draxler (2014) reported acceptable limits of MSEs for
values that ranged between 1 and 10,000 to be between 0.1 and 10.
Hence, the permeability network was able to predict permeability
values with minimal amount of uncertainty. The selected NN for both
porosity and permeability, were applied to new field data (inputs)
6400

from the remaining three wells in order to predict these reservoir prop-
Depth (ft)

erties. For the porosity network, R-values of 0.99099 in Well 2, 0.99474

in Well 3 and 0.83749 in Well 4 were observed, while the permeability
network have R-values of 0.97584, 0.83598, 0.97002, and 0.83512 re-
6500

spectively (Table 6). Very strong correlations between the ANN and
core porosity were observed in Wells 1, 2 and 3 (i.e., 0.99987, 0.99099,
and 0.99474). However, a relatively low value of R (i.e., 0.83749) was
observed in well 4. This may be due to slight changes in geology
resulting from depositional facies pattern, grain sizes or textural varia-
6600

tions (Ebong et al., 2019a). Similarly, the permeability network show,

R-values of 0.97584, 0.83598, 0.97002 and 0.83512 for wells 1, 2, 3
and 4 respectively. Since our aim is to develop a NN that can predict
these properties when presented with limited input data, we compared 6700
6700

the core derived fractional porosity with the ANN predicted and the
conventional empirical calculated porosity from density log (Fig. 12),
estimated the error between the fraction porosity predicted from ANN
and empirically calculated porosity with regards to the core fractional
porosity. The ANN predicted porosity correlated well with the core po-
6800

rosity and showed minimal error than the conventional empirically cal-
culated porosity (Fig. 13). Core derived permeability and ANN predicted
permeability was also compared for all four wells (Fig. 14). Fields A
(i.e., location of Wells 1 and 3) and B (i.e., location of Well 2) share
the same depositional facies attributes since they are within the Coastal Fig. 15. Composite log from Well 1 showing neural network input parameters. Water
Swamp Depo-belt. However, Well 3 showed some degree of variability saturation Sw, bulk density RHOB and core porosity Por (Frac.)

11
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

Porosity vs Density

0.45 Porosity vs. Density

Porosity model fit 1
0.4

0.35
Porosity

0.3

0.25

0.2

0.15

1.9 2 2.1 2.2 2.3 2.4

Density

Fig. 16. Surface diagram showing the functional relationship between the dependent variable (porosity) and the independent variables (density log).

(−2774, −2408) with 95% confidence bounds. Hence, the permeability

Fðx, yÞ ¼ P00 þ P10 x þ P01 y ð13Þ
values can be predicted using the model.
F(x,y) represents the permeability, x is the density log value normalised K ¼ 8056–3116  –2591y ð14Þ
by mean 2.126 and standard deviation (std) of 0.04552, and y is the wa-
ter saturation normalised by mean 0.1135 and std. of 0.07266. P00, P10 The accuracy of this model is evident from the statistical parameters
and P01 are coefficients of the polynomial calculated to be P00 = 8056 presented in Table 7. Curve fitting result for the permeability network
(7928, 8185), P10 = −3116 (−3299, −2933) and P01= −2591 along the line of perfect agreement (i.e., the diagonal between water

Fig. 17. Surface diagram showing the functional relationship between the dependent variable (permeability) and the two independent variables (water saturation and density log).

12
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

Table 7
Statistical parameters showing the goodness of fitted model for the porosity and permeability networks.

Network model Fit type SSE R- square DFE Adjusted R-square RMSE

Porosity Poly 1 2.2715 e−04 0.9997 1197 0.9997 4.3562 e−04

Permeability Poly 11 7.8315 e−04 0.9753 1196 0.9752 8.092

saturation and density) show positive relationship and a convex surface porosity values of interest observed across the wells were between
along the line of incongruency (Fig. 12). The implication of this as sug- 0.20 and 0.35 (~ 27% on average) and are described as ranging from
gested by Shanock et al. (2010) is that the two support variables (den- very good to excellent (Table 8). These ranges of porosity dominates
sity and water saturation), can affect the outcome (permeability) over 80% of the reservoir across all the wells (Fig. 12) and agrees with
either negatively or positively depending on the direction of the dis- reports of Okwoli et al. (2015); Nwankwo et al. (2015); Akaerue et al.
crepancy. As density and water saturation decreases towards the line (2018); and Ebong et al. (2019a) within the Niger Delta Basin. The per-
of perfect agreement, permeability increases. However, moving away meability values that dominated over 60% of the reservoir in each of the
from the line of perfect agreement, as the water saturation increases wells, were between 50 and 500 mD which are fair to very good
and density decreases, K increases. This implies that the discrepancy be- (Table 8). However, permeability values ≥500 mD were also observed
tween the predictor variables becomes lower but on the other hand, as in almost all the wells implying excellent. The primary influence on
density increases and water saturation decreases, K decreases up to a the porosity and permeability values observed are direct consequences
point it begins to increase again (Fig. 17). This accounts for the convex of depositional sequences and the differential responses resulting from
curvature along the line of incongruency. Although, permeability may the various diagenetic regimes within the basin.
be related to porosity as empirical models suggest (Ebong et al.,
2019b; Rafik and Kamel, 2016), however it does not depend totally on
porosity except for interconnected pores. Thus, reduction in density of 6. Conclusion
the subsurface material due to unconsolidated materials can imply in-
crease in porosity (Figs. 15 and 16). This agrees with the reports of The essential ingredient in reservoir development and production
Akaerue et al. (2018). Rahmouni et al. (2014) opined that there is a di- optimization is a good knowledge of the reservoir properties. Porosity
rect relationship between permeability and porosity. This may probably and permeability are two of the reservoir properties needed for efficient
explain the reason why K increases as density and water saturation de- and effective reservoir characterization. Core analysis tends to provide
creases towards the line of perfect agreement (Eq. (5)). Again it throws an accurate and more reliable measurements of these properties but
more light on the second scenario where density decreases, water satu- its inherent challenge include high cost, timing and selective sampling
ration increases and K increases as this may actually be a proportional of depths and wells to mention a few. Artificial neural network tends
increase with slightly higher porosity value thus causing an increase to provide a quick and alternative solution to estimating these reservoir
in K. Also, a very good explanation to increase in density, reduction in properties at all depths. In view of this, different neural network models
water saturation and decrease in K in the third scenario mentioned with limited data were developed with the aim of selecting the best for
above, can be attributed to compaction as clayey or shaly materials fill predicting porosity and permeability within the Niger Delta region. Two
the pore spaces (Scherer, 1987; Ekine and Iyabe, 2009). The porosity separate networks were selected; one for predicting porosity and the
surface diagram displayed an inverse linear relationship (Fig. 16). This other for permeability. The porosity NN model utilized only density log
indicates an additive model though with a negative slope (Shanock data (RHOB) as the input and core porosity as the target while the per-
et al., 2010). Thus, as the predictor variable (i.e., density) increases, meability NN used both water saturation and RHOB as the input and
the outcome (i.e., porosity) decreases and vice versa. The goodness of core derived permeability as the target. The porosity network with
the fit were examined by the following statistical parameters; SSE of MSE of 1.20889e−6, 2.22574e−6, and 1.37204e−6 for training, validation
2.2715 e−04, R-square of 0.9997, Adjusted R-square of 0.9997 and and test respectively and the permeability network with MSE of
RSME of 4.3562 e−04 for the porosity fitted model (Table 7). For a 5.62849e−1, 3.13709e−1, and 4.83728e−1 later applied to four wells to
small range of values of porosity between 0 and 0.5, statistical parame- predict porosity. Very good correlation coefficients (R) of the NN pre-
ters presented above indicate a very good model. For the permeability dicted and real core data were obtained in each of the wells. For the po-
model, SSE of 7.8315e01, R-square of 0.9753, Adjusted R-square of rosity network R-values of 0.99987, 0.99099, 0.99474 and 0.83749 were
0.9752 and RSME of 8.092 were observed. Previous work done by obtained, whereas 0.97584, 0.83594, 0.97002 and 0.83512 respectively
Chiamogu and Ehinola (2010); Aigbedion (2004) showed that regres- were observed also from the permeability network. The NN was
sion models such as linear, power-law, logarithmic, and exponential modeled mathematically using a nonlinear regression tool. The
can predict permeability and porosity reasonably well at intermediate polynomial model was assumed and the location and surface parame-
to deeper depths for composite sandstone sequence. However, in this ters otherwise called coefficients were iteratively determined using
work both linear and nonlinear regressions were applied for developing the Gauss-Newton optimization method. The goodness of fits were
the porosity and permeability models respectively. The ranges of tested based on observable statistical parameters such as RMSE, ad-
justed R-squared, R-squared and standard error. The result showed
good model within the permissible error limits. The models generated
can be used to predict reservoir properties with minimal error else-
where with similar geologic characteristics like the Niger Delta province.
Table 8
Permeability and porosity ranges and their description (After Etu-Efeotor, 1997).

Porosity ϕ (Frac.) Remark Permeability K (mD) Remark Credit Author Statement

0.01 < ϕ < 0.05 Very low 0 < K < 10 Very low
0.05 < ϕ < 0.10 Good 10 < K < 50 Fair Job G. Urang: Conceptualization, Methodology, Software, Writing-
0.10 < ϕ < 0.30 Very good 50 < K < 500 Good original draft, Project Administration, Validation, Formal analysis, Inves-
ϕ > 0.30 Excellent K > 500 Excellent tigation Ebong D. Ebong: Data curation, Writing-Review and Editing, Vi-
sualization, Investigation, Resources, Formal analysis Anthony E. Akpan:

13
J.G. Urang, E.D. Ebong, A.E. Akpan et al. Journal of Applied Geophysics 183 (2020) 104207

Supervision, Methodology, Writing-Review and Editing Emmanuel I. Hamada, G.A., Elshafei, M.A., 2010. Neural network prediction of porosity and permeabil-
ity of heterogeneous gas sand reservoirs using NMR and conventional logs. NAFTA 61
Akaerue: Resources, Investigation, Writing-Review and Editing. (10), 451–460.
Helle, H.B., Bhatt, A., Ursin, B., 2001. Porosity and permeability prediction from wireline
logs using artificial neural networks: a North Sea case study. Geophys. Prospect. 49,
431–444. https://doi.org/10.1046/j.1365-2478.2001.00271.
Declaration of Competing Interest Kaplan, A., Lusser, C.U., Norton, I.O., 1994. Tectonic Map of the World, Panel 10: American
Association of Petroleum Geologists Tulsa, Scale 1:10,000,000.
The authors declare that they have no known competing financial Koerhsen, W., 2018. Overfitting vs. Underfitting: A Complete Example. http://www.
towardsdatascience.com/overfitting-vs-underfitting-a-complete-example-
interests or personal relationships that could have appeared to influ- d05dd7e19765s.html (Accessed: 15 January 2020).
ence the work reported in this paper. Kogbe, C.A., 1976. Geology of Nigeria. Elizabeth Publishing Company, Lagos.
Koledoye, B.A., Aydin, A., May, E., 2003. A new process-based methodology for analysis of
shale smear along normal faults in the Niger Delta. Am. Assoc. Pet. Geol. Bull. 87,
Acknowledgements 445–463.
Lim, J., 2005. Reservoir permeability determination using artificial neural network.
The authors specially appreciate Shell Petroleum Development J. Korean Soc. Geosyst. Eng. 40, 232–238.
Liu, Y., Wang, W., Chen, F., Juttler, B., 2008. A revisit to least squares orthogonal distance
Company (SPDC), Nigeria for providing the data used in this work. We
fitting of parametric curves and surfaces. Advances in Geometric Modeling and Pro-
acknowledge the Department of Petroleum Resources (DPR) for their cessing, Lecture Notes in Computer Science. 4975, pp. 384–397.
faith in this project and directed SPDC to release the data. Also, the au- Lopes, R.L., Jorge, A.M., 2018. Assessment of predictive learning methods for the comple-
tion of gaps in well log data. J. Pet. Sci. Eng. 162, 873–886.
thors are grateful to the University of Calabar for providing the enabling
Madj, T.M., Hezarkhani, A., 2011. Estimation of spatial distribution of porosity by using
environment in terms of relevant software, funding and good working neural networks method in one of oil fields in south of Iran. Aust. J. Basic Appl. Sci.
condition that made this research work possible. 5 (8), 182–189.
Martley, A.O., 2000. Lecture Notes on Petrophysics. Shell intensive training program, Port
Harcourt.
References Mohaghegh, S., 2000. Virtual-intelligence applications in petroleum engineering: part I.
Artificial neural networks. J. Pet. Technol. 52, 64–73.
Aigbedion, I., 2001. Petrophysical analysis of some onshore wells in the Niger Delta. A Ph. Mohaghegh, S., Ameri, S., Aminian, K., 1996. A methodological approach for reservoir het-
D thesis submitted to University of Benin. pp. 70–73. erogeneity characterization using artificial neural networks. J. Pet. Sci. Eng. 16,
Akaerue, E.I., Ita, O., Okonkwo, A.C., 2018. Petrophysical evaluation of Etu field coastal 263–274.
swamp I depobelt Niger Delta, Nigeria. Elixir Earth Sci. 119, 50989–50994. Negnevitsky, M., 2005. Artificial Intelligence: A Guide to Intelligent System. Adison-wes-
Akpan, A.E., Narayanan, M., Harinarayana, T., 2013. Estimation of subsurface temperatures ley, Australia.
in the Tattapani geothermal field, Central India from limited volume of Nwankwo, C.N., Ohakwere-Eze, M., Ebeniro, J.O., 2015. Hydrocarbon reservoir volume es-
magnetotelluric data and borehole thermograms using a constructive back propaga- timation using 3-D seismic and well log data over an X-field, Niger Delta Nigeria.
tion neural network. Earth Interact. 18, 1–24. J. Pet. Explor. Prod. Technol. 5, 453–462.
Aminian, K., Ameri, S., 2005. Application of artificial neural networks for reservoir charac- Okwoli, E., Obiora, D.N., Adewoye, O., Chukudebelu, J.U., Ezema, P.O., 2015. Reservoir char-
terization with limited data. J. Pet. Sci. Eng. 49 (3–4), 212–222. acterization and volumetric analysis of “lona” Field, Niger Delta, using 3-d seismic
Anifowose, F.A., Labadin, J., Abdulraheem, A., 2017. Ensemble machine learning: an un- and well log data. J. Pet. Coal. 57 (2), 108–119.
tapped modeling paradigm for petroleum reservoir characterization. J. Pet. Sci. Eng. Rafik, B., Kamel, B., 2016. Prediction of permeability and porosity from well log data using
151, 480–487. the nonparametric regression with multivariate analysis and neural network, Hassi
Atieg, A., Watson, G.A., 2003. A class of methods for fitting a curve or surface to data by R’Mel Field, Algeria. Egypt. J. Pet. https://doi.org/10.1016/j.ejpe.2016.10.013.
minimizing the sum of the squares of orthogonal distances. J. Comp. Appl. Maths Rahmouni, A., Boulanouar, A., Boukalouch, M., Géraud, Y., Samaouali, A., Harnafi, M.,
158 (3), 277–296. Sebbani, J., 2014. Relationships between porosity and permeability of calcarenite
Ayoub, M.A., Mohamed, A.A., 2015. Estimating the lengthy missing log interval using rocks based on laboratory measurements. J. Mater Environ. Sci. 5 (3), 931–936.
group method of data handling (GMDH) technique. Applied Mechanics and Mate- Ramgulam, A., 2006. Utilization of Artificial Neural Network in the Optimization of His-
rials. 695. Trans Tech Publications Ltd, pp. 850–853. tory Matching. An unpublished M.Sc thesis. Department of Energy and
Chai, T., Draxler, R.R., 2014. Root mean square error (RMSE) or mean absolute error GeoEngineering, Pennsylvania State University, USA.
(MAE)? – Arguments against avoiding RMSE in the literature. Geosci. Model Dev. Rezaee, M.R., Ilkhchi, A.K., Barabadi, A., 2007. Prediction of shear wave velocity from
(7), 1247–1250. petrophysical data utilizing intelligent systems: an example from a sandstone reser-
Chiamogu, A.G., Ehinola, A.O., 2010. Distribution patterns of porosity and permability in voir of Carnarvon Basin, Australia. J. Pet. Sci. Eng. 55 (3–4), 201–212.
the Hydrocarbon bearing sands of Agbada formation, Niger Delta Continental shelf, Saemi, M., Ahmadi, M., Yazdian, V.A., 2007. Design of neural networks using genetic algo-
Nigeria. AAPG International Convention and Exhibition, Calgary, Alberta, Canada Sep- rithm for the permeability estimation of the reservoir. J. Pet. Sci. Eng. 0920–4105.
tember 12–15, 2010. Scherer, M., 1987. Parameters influencing porosity of sandstones: a model for sandstone
Churikov, N.S., Grafeeva, N.G., 2018. Recovering gaps in the gamma-ray logging method. porosity prediction. AAP G Bull. 71, 485–491.
18th International Multidisciplinary Scientific GeoConference SGEM 2018. p. 7. Shanock, L.R., Baran, B.E., Gentry, W.A., Pattison, S.C., Heggestad, E.D., 2010. Polynomial re-
Doust, H., Omatsola, E., 1990. Niger Delta. In: Edwards, J.D., Santogrossi, P.A. (Eds.), Diver- gression with response surface analysis: a powerful approach for examining moder-
gent/passive Margin Basins, AAPG Memoir 48. American Association of Petroleum ation and overcoming limitations of difference scores. J. Bus. Psychol. 25 (4),
Geologists, Tulsa, pp. 239–248. 543–554.
Ebong, E.D., Akpan, A.E., Emeka, C.N., Urang, J.G., 2017. Groundwater quality assessment Steele, D., Ejedawe, J., Adeogba, T., Grant, C., Filbrandt, J., Ganz, H., 2009. Geological Frame-
using geoelectrical and geochemical approaches: case study of Abi Area, southeastern work of Nigeria linked Shelf Extension and DeepwaterThrust Belts. AAPG Hedberg
Nigeria. J. Appl. Water Sci. 7 (5), 2463–2478. Conference, Tirrnia, Italy October 4–9, 2009.
Ebong, E.D., Akpan, A.E., Ekwok, S.E., 2019a. Stochastic modeling of spatial variability of Verma, K.A., Cheadle, A.B., Routray, A., Mohanty, K.W., Mansinha, L., 2012. Porosity and
petrophysical properties in parts of the Niger Delta Basin, southern Nigeria. J. Pet. Permeability Estimation using Neural Network Approach from Well Log.
Explor. Prod. Technol. https://doi.org/10.1007/s13202-019-00787-2. Geoconvention 2012:Vision, Canada.
Ebong, E.D., Akpan, A.E., Urang, J.G., 2019b. 3D structural modelling and fluid identifica- Wang, B., Wang, X., Chen, Z., 2013. A hybrid framework for reservoir characterization
tion in parts of Niger delta basin, southern Nigeria. J. Afr. Earth Sci. https://doi.org/ using fuzzy ranking and an artificial neural network. Comput. Geosci. 57, 1–10.
10.1016/j.jafrearsci.2019.103565. Wong, P.M., Gideon, T.D., Taggart, I.J., 2002. An improved technique in porosity predic-
Ekine, A.S., Iyabe, P., 2009. Petrophysical characterisation of the Kwale Field reservoir tion: a neural network approach. J. Geosci. Remote Sens. IEEE Trans. 33 (4), 971–980.
sands (OML 60) from Wireline logs, Niger Delta, Nigeria. J. Appl. Sci. Environ. Wyllie, M.R.J., Rose, W.D., 1950. Some theoretical considerations related to the quantita-
Manag. 13 (4), 81–85. tive evaluation of the physical characteristics of reservoir rock from electrical log
Etu-Efeotor, J.O., 1997. Fundamentals of Petroleum Geology; Paragraphics. Port Harcourt, data. J. Pet. Technol. 2 (04), 105–118.
Nigeria, pp. 51–63. Zhang, X., Zhou, X., Lin, M., Sun, J., 2018. Shufflenet: an extremely efficient convolutional
Hadi, I., Sabh, M., 2014. Behavior formula extraction for object trajectory. Int. J. Comput. neural network for mobile devices. Proceedings of the IEEE Conference on Computer
Appl. 104 (2), 0975–0987. Vision and Pattern Recognition, pp. 6848–6856.
Hagan, M.T., Demuth, H.B., Beale, M., 1996. Neural Network Design. PWS Publishing Com-
pany, USA.

14