J Petrol Explor Prod Technol (2015) 5:157–166

DOI 10.1007/s13202-014-0137-7

ORIGINAL PAPER - EXPLORATION GEOPHYSICS

Generalized regression and feed-forward back propagation neural

networks in modelling porosity from geophysical well logs
Ahmed Amara Konaté • Heping Pan •

Nasir Khan • Jie Huai Yang

Received: 27 December 2013 / Accepted: 21 August 2014 / Published online: 11 September 2014
Ó The Author(s) 2014. This article is published with open access at Springerlink.com

Abstract Geophysical formation evaluation plays a fun- Keywords Porosity prediction  Generalized regression
damental role in hydrocarbon exploration and production neural network Feed-forward back propagation 
processes. It is a process which describes different reser- Computational geophysics  Reservoir evaluation 
voir parameters using well field data. Porosity is one of the Reservoir properties  Geophysical well logs
parameters that determines the amount of oil present in a
rock formation and research in this area is mainly carried
out by engineers and geoscientists in the petroleum Introduction
industry. Accurate prediction of porosity is a difficult
problem. This is mostly due to the failure in the under- A major activity in evaluating reservoir is examining the
standing of spatial porosity parameter distribution. Artifi- impact of reservoir heterogeneities on reservoir behavior.
cial neural networks have proved to be a powerful tool for Heterogeneity in evaluating reservoir is referred to as non-
mapping complicated and non-linear relationships in linear and non-uniform spatial distribution of rock prop-
petroleum studies. In this study, we analyze and compare erties such as porosity, permeability and fluids (oil, gas,
generalized regression neural network (GRNN) and feed- water) saturation (Mohaghegh et al. 1996). However, it is
forward back propagation neural network (FFBP) in difficult to predict rock properties due to the form and
modeling porosity in Zhenjing oilfield data. This study is spatial distribution of these heterogeneities, also the
calibrated on four wells of Zhenjing oilfield data. One well applicability of traditional analytical techniques such as
was used to find an empirical relationship between the well multivariate regression are limited in this context. Several
logs and porosity, while the other three wells were used to authors such as Mohaghegh et al. (1996) and Handhel
test the model’s predictive ability in the field, respectively. (2009) in their related researches buttress these complexi-
The findings proved that the GRN network can make more ties for predicting in heterogeneity reservoir in oil and
accurate and credible porosity parameter estimation than natural gas field studies.
the commonly used FFBP network. Artificial intelligence Understanding the form and spatial distribution of rock
can be exploited as a powerful instrument for predicting properties is fundamental to a successful characterization
reservoir properties in geophysical formation evaluation of petroleum reservoirs (Haldorsen and Damsleth 1993;
and reservoir engineering in petroleum industry. Wong et al. 1995). In this prevalent situation, it is useful to
construct a model that understands rock properties and has
the capabilities to make a good prediction. To build a
model for predicting requires a set of mathematical equa-
tions which describe the dynamic behavior of the process,
A. A. Konaté  H. Pan (&)  N. Khan  J. H. Yang
in other words link a number of input variables with a set of
Institute of Geophysics and Geomatic, China University
of Geosciences, Wuhan 430074, Hubei Province, China results.
e-mail: [email protected] This is a typical problem that can be solved by artificial
A. A. Konaté neural network (ANN) if the phenomenon to be modeled is
e-mail: [email protected] non-linear; such as the one used in this research. ANNs

123
158 J Petrol Explor Prod Technol (2015) 5:157–166

(Hecht-Nielsen 1989) offer an alternative that has the common testing set. It also as well as examines the model’s
potential to establish a model from non-linear, complex ability to predict porosity in the oilfield.
and multidimensional data. They usually take little time to The findings showed that, GRNN compared to FFBP,
predict output response for any input value that falls in the GRNN model for modeling porosity using geophysical
range of the training data. well logs is significant for the geophysical exploration
Artificial neural network offers real benefit over tradi- undertaken in the petroleum industry.
tional modeling, including the ability to handle large
amounts of noisy data from dynamic and nonlinear systems
without a priori information of the processes involved, Database
ANN provides an adequate solution even when the data are
incomplete or ambiguous (Handhel 2009). One of the Four wells named Well#A, Well#B, Well#C and Well#D
interesting properties of ANN is that it makes accurate from Zhenjing oilfield China were used to provide phys-
predictions. This predictive ability is that it has a degree of icals log and core porosity data. The physicals logs con-
liberty that allows it to better capture the non-linearity of a sisted of bulk density (DEN), compensated neutron
system compared to other modeling techniques. porosity (CNL), acoustic (AC) and deep induction resis-
Artificial neural network has become progressively tivity (ILD). The DEN, CNL and AC respond to the
popular in the petroleum industry and has been widely characteristics of rock directly adjacent to the borehole. A
applied in many petroleum industry applications (see combination of these logs provides more accurate esti-
review in van der Baan and Jutten 1992; Poulton 2002). In mations of porosity. These geophysical logs are also
most of these publications, the feed-forward back propa- known as porosity logs. The difference existing between
gation neural network (FFBP) configuration is proposed. these porosity logs is that, the DEN and CNL are nuclear
However, despite the popular applications of FFBP as measurements while AC uses acoustic measurements.
shown by these publications, the FFBP suffers mostly from However, ILD is an electric log that measures the resis-
the weight initialization randomly given, and the local tivity of the un-invaded zone of the formation. A crucial
minima problems, which can often lead the model to use of ILD is the determination of hydrocarbon contained
evolve in an inaccurate direction. More precisely there are within the pore space of the formations traversed by the
chances that the model never converges to the optimal well. Figure 1 shows the geophysical well logs used in
solution. Furthermore, FFBP training time is often slow this study.
and in network architecture optimization, FFBP still have Logging tool responses are badly affected by breakout
too much human interferences, i.e. the determination of the of wall-rock during drilling, as well as stick-and-pull as
hidden layer and the number of hidden layer neurons still logging tools are winched up the well (Yan 2002). Keeping
depend on the user. An alternative way to avoid the above this in mind, during this study, the data set from the three
problem is to employ generalized regression neural net- wells were carefully examined. All geophysical well logs
work (GRNN). As mentioned by Specht (1991), in GRN which exhibited strange, and possibly inappropriate data
network optimization process only one parameter were ignored. In addition, correction of the offset between
(smoothness parameter) has to be adjusted in one pass core depth and logging depth was done, so that the geo-
through the data; no iterative procedure is required; the physical well logs and experimental data may be matched
estimate is confined by the minimum and maximum of the and integrated effectively.
data; convergence guaranteed; fast and stable. The data from Well#A (1046 core and log data) were
Based on the aforementioned assertions, the researchers chosen to provide the training patterns. This well was
were motivated to propose GRNN to predict porosity using chosen because, it had the most complete set of core and
four geophysical well logs. The prediction performances log data. It was randomly divided into training data (70 %)
were quantified, and compared to FFBP. and testing data (30 %). The data from Well#B (152 core
Porosity is one of the key petrophysical parameters in and log data), Well#C (91 core and log data) and Well#D
evaluating reservoirs to optimize the production of oil and (40 core and log data) were used to test the model’s ability
natural gas fields. It is one of the factors that determines the to predict porosity in the oilfield.
amount of oil present in a rock formation and research in
this area is mostly carried out by engineers and geoscien-
tists in the petroleum industry. Methods
This study focuses on developing a model based on
ANN that is applicable to predict porosity at Zhenjing Artificial neural network is a set of computing systems that
oilfield China. It highlights the comparison between GRNN imitates biological processes through the use of intercon-
over FFBP in porosity modeling and estimation on nections between simple artificial neurons. While the

123
J Petrol Explor Prod Technol (2015) 5:157–166 159

Fig. 1 Geophysical well logs used in this study. Well#A

concept may seem to belong to recent technological An artificial neuron is a calculating unit that receives a
developments, it has been discussed long before the current certain number of inputs directly from the environment or
trend in computers with the objective of trying to duplicate from upstream neurons. When the information comes from
the learning abilities of biological neurons that constitute a neuron, it is associated with a weight (w), which repre-
the basic element of the brain. From a technical point of sents the ability of the upstream neuron to excite or inhibit
view, each neuron is connected to others by direct links. downstream neurons. Each neuron is provided with a
Each link is associated with a weight which represents the unique output, which then branches out to supply a variable
information used by the network to solve the problem. number of downstream neurons.

123
160 J Petrol Explor Prod Technol (2015) 5:157–166

layer is enough to approximate any continuous function.

Therefore, one hidden layer was employed in the current
research. Besides, transfer functions for the hidden nodes
are needed to introduce non-linearity into the network. In
this study, the sigmoid was selected as activation function
of the hidden neurons while a linear activation function
was used in the output neurons.
Next, the choice of the optimal number of hidden layer
neurons is an essential decision in the modeling phase. If an
insufficient number of neurons are used, the network will be
unable to model complicated data, and the resulting fit will
Fig. 2 Schematic representation of FFBP
be poor. Many hidden neurons will ensure correct training,
and the network will be able to appropriately predict the
Artificial neural network presumes that the true under- data it has been trained on, but its performance on new data
lying function that governs the relationship between inputs and its ability to generalize will be compromised (Abraham
and outputs is not known a priori. It determines a mathe- 2005). Whereas, with very few hidden neurons, the network
matical function which can properly approach the repre- may be inept to learn the associations between the input and
sentation of inputs and outputs. output variables. In this sense, the error will fail to fall
One of the major aspects of ANN is the training process, below an adequate level (Abraham 2005). Thus, a com-
which can be either supervised or unsupervised. In this promise has to be reached between too many and too few
study, the former was used for prediction approach. It is the neurons in the hidden layer. In this study, the optimal
most widely applied in geophysical fields (van der Baan number in hidden layer was selected by experimental trial
and Jutten 1992; Poulton 2002). Supervised learning, i.e. based on the smallest mean square error (MSE).
guided learning by ‘‘teacher’’; requires a training set which The objective of training the FFBP is to find optimal
consists of input vectors and a target vector associated with connection weights (w*) in such a manner that the value of
each input vector. The advantage of supervised training is calculated outputs for each example matches the value of
that the output can be interpreted based on the training desired outputs. This is typically a non-linear optimization
values. The disadvantage is that a large number of inputs problem, where w* is given by Eq. (1)
and outputs are required to guarantee adequate training. In w ¼ arg minEðwÞ ð1Þ
this study, the given training dataset (1046 core and log
data) is sufficient and requires a supervised learning model. where w is weight matrix and E(w) is an objective function
on w, which is to be minimized.
Feed-forward back propagation neural network (FFBP) The E(w) is evaluated at any point of w given by Eq. (2)
X
EðwÞ ¼ Ep ðwÞ ð2Þ
Feed-forward back propagation neural network is one of p
the most popular ANN models for engineering applications
(Haykin 2007). The FFBP represented in Fig. 2 comprises p is the number of examples in the training set and Ep(w) is
of three layers; the input layer receiving the information on the output error for each example p. Ep(w) is expressed by
the neurons represented by circles and an output layer Eq. (3)
having a single neuron and giving the internal calculation 1X 2
Ep ðwÞ ¼ dpj  ypj ðwÞ ð3Þ
result. Between these two layers, there is another layer not 2 j
visible from the outside called the hidden layer responsible
for performing intermediate computations. where ypj(w) and dpj are the calculated and desired network
Determination of the number of hidden layers, hidden outputs of the jth output neuron for pth example,
neuron and type of transfer function plays an important role respectively. The objective function to be minimized is
in FFBP model constructions (White 1992). The number of represented by Eq. (4):
hidden layers required depends on the complexity of the 1XX 2
EðwÞ ¼ dpj  ypj ðwÞ : ð4Þ
relationship between the input and the target parameters. It 2 p j
has an impact on the quality of the learning, FFBP com-
prising more hidden layers are very rare, given that each For each learning (training) process, the network
new layer increases the quantity of calculations. In calculated output value is compared to the desired output
majority problems only one hidden layer is sufficient. value. If there is a difference between the calculated and
Hornik et al. (1989) proved that FFBP with one hidden desired output network, the synaptic weights which

123
J Petrol Explor Prod Technol (2015) 5:157–166 161

contribute to generate a significant error will be changed more ‘‘pattern layer’’. In the pattern layer, each neuron is a
significantly than the weight that led to a marginal error. The training pattern and its output represents a measure of the
adaptation of the weights begins at the output neurons and distance between the input and the stored patterns. The
then continues toward the input data. There are many hidden layer is fully linked to the third layer, called
algorithms available to perform this weight selection and ‘‘summation layer’’. This later has two different types of
adjustment (see Bishop 1995). One of the most popular is the summation: S-summation neuron (summation units) and
gradient descent, which suffers from slow convergence times D-summation neuron (a single division unit). S-summation
and can easily get trapped in local minima within the vector neuron determines the sum of the weighted outputs of the
space of w during the learning process; this leads the model to hidden layer, whereas the D-summation neuron determines
evolve in an accurate direction. In this research, this algorithm the unweighted outputs of the pattern neurons. As also
was applied with no guarantee in obtaining the optimal mentioned in Fig. 3, the synaptic weight between a neuron
trained network for given data. Therefore, Levenberg– in the hidden layer and an S-summation layer neuron is the
Marquardt algorithm (LMA) was chosen to train the neural target output value corresponding (yi). The summation layer
network. LMA is considered one of the most efficient training and the last layer of the network, ‘‘output layer’’ together
algorithms; the study of Hagan and Menhaj (1994) proved execute a normalization of output set. In the training of the
that LMA is faster and has more stable convergence as network, radial basis function (Gaussian) and linear transfer
compared to gradient descent algorithm. functions are used in hidden and output layers, respectively.
In reference to Specht (1991), let us suppose that f(x, y)
Generalized regression neural network (GRNN) represents the known joint continuous probability density
function of a vector random variable, x, and a scalar random
Generalized regression neural network is related to the variable, y. The regression of y on x is expressed by Eq. (5):
radial basis neural networks, which are found on kernel R1
regression. It can be treated as a normalized radial basis yf ðx; yÞdy
E½y=x ¼ R1 1 : ð5Þ
neural networks in which there is a hidden neuron centered 1 f ðx; yÞdy
at every training case. These radial basis function units are If the density f(x, y) is unknown, it must generally be
generally probability density function such as the Gaussian
predicted (estimated) from a sample of observations of x and
(Celikoglu 2006). The use of a probability density function
y. The probability estimator f^ðx; yÞ given in Eq. (6), is based
is particularly gainful due to its ability to converge to the
upon sample values of the variables x and y represented by xi
underlying function of the data with only limited training
and yi, respectively. n and p represent the number of sample
data available. In GRNN optimization process only one
observations and the dimension of the vector variable x,
parameter (smoothing) has to be adjusted in one pass
respectively:
through the data; no iterative procedure is required; the
estimate is confined by the minimum and maximum of the 1
f^ðx; yÞ ¼
data. Furthermore, GRNN approximates any arbitrary ð2pÞðpþ1Þ=2 rðpþ1Þ
" # " #
function between input and target vectors; fast training and
1X n
ðx  xi ÞT ðx  xi Þ ðy  yi Þ2
convergence to the optimal regression surface as the train-  exp  exp  :
n i¼1 2r2 2r2
ing data becomes very large (Specht 1991). This makes
GRNN a very advantageous tool to perform predictions. ð6Þ
Figure 3 is a representation of the GRNN architecture
A meaningful explanation of the probability estimate
with four layers: an input layer, a hidden layer, a summation
layer, and an output layer. As it can be seen in Fig. 3, the f^ðx; yÞ is that, it allocates sample probability of smoothness
input layer is completely linked to the hidden layer called parameter (r) for each sample xi and yi, and the probability
estimate is the sum of those sample probabilities.
When defining the scalar function given by Eq. (7)

D2i ¼ ðx  xi ÞT ðx  xi Þ: ð7Þ
Therefore, a prediction performed by GRNN, y^ðxÞ to an
unknown input vector x is expressed in Eq. (8)

P
n  
D2
yi exp  2ri2
y^ðxÞ ¼ i¼1
Pn   ð8Þ
D2
exp  2ri2
Fig. 3 Schematic diagram of a GRNN architecture i¼1

123
162 J Petrol Explor Prod Technol (2015) 5:157–166

Table 1 Pearson correlation (r) of observed porosity versus geo- mean observed value; p is mean predicted value and N is
physical well logs data (Well#A) total number of data
Pearson r value r probability Level of
1X N
correlation (r) (p) significance MSE ¼ ðoi  p^i Þ2 ð9Þ
N i¼1
Porosity vs ILD -0.42* 0.000 0.01
Porosity vs AC 0.83* 0.000 0.01 MSE, measures the average of the squares of the errors, i.e.
Porosity vs DEN -0.60* 0.000 0.01 the residual errors, which help scientists to understand and
Porosity vs CNL 0.51* 0.000 0.01 interpret the difference between the observed value and
estimated values. We should keep in mind that, this
* Correlation statistically significant at the 0.01 (p B 0.01)
indicator measures how near a fit line is to data points. The
smaller the MSE, the nearer the fit is to the data points.
0 12
where each sample, xi of x is used as the mean of a normal
distribution. B PN C
B ðoi  oÞ  ð^ pi  pÞ C
As mentioned by Specht (1991), the smoothness B C
R2 ¼ B
Buv i¼1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiC
C ð10Þ
parameter r, is a very important parameter of GRN net- Bu P C
work. We should keep in mind that, if r is bigger, the @t N PN
A
ðo  oÞ2  i ð^p  pÞ2i
predicted (estimated) density is forced to be smooth and in i¼1 i¼1
the limit becomes a multivariate Gaussian with covariance
r2 I (I = unity matrix). In contrary, when r is smaller, the This coefficient is a statistical index that expresses the
predicted density assumes non Gaussian shapes, but with quality of fit estimates of the regression equation and also
the hazard that wild points may have a great effect on the the intensity of the linear relationship. It helps to have a
estimate. The smoothness parameter (r), is still subject to a general idea of the model fit. Its value varies between 0 and
search. 1, and if the R2 value is close to 1 it is sufficient to say that
the fit is good.
Performance criteria PN PN PN
N oi p^i  oi p^i
i¼1 i¼1 i¼1
r ¼ vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!
In diagnostic statistics, there are many ways to quantify the u   2  N 2 !
u
t N P o2  P oi P P
N N N
difference between observed values and predicted (esti- N p^i  2 p^i
i
mated) values. In this study, to evaluate FFBP scheme and i¼1 i¼1 i¼1 i¼1
GRNN scheme, the statistics mean squared error (MSE),
ð11Þ
coefficient of determination (R2) and coefficient of corre-
lation (r) were used to quantify performance. They are r measures the strength of a linear relationship amongst the
given by Eqs. (9), (10) and (11), respectively. Where oi is observed value and estimated variables. In other words, it
observed porosity value, p^i is predicted porosity value, o is is an indicator of the scatter around the fit line. If r is close

Fig. 4 Cross-plots of predicted porosity against observed porosity. Prediction from FFBP (a) and GRNN (b). Training data Well#A

123
J Petrol Explor Prod Technol (2015) 5:157–166 163

Table 2 Statistical performance of GRNN and FFBP scheme of

porosity and observed porosity (Well#A)
Methods Testing data (314 data points) Training data (734 data
non trained point)
R2 r MSE R2 r MSE

GRNN 0.958 0.978 0.278 0.970 0.984 0.383

FFBP 0.940 0.969 0.381 0.960 0.979 0.449

to 1, it means that the relationship between the observed

and estimated variables is positive and thereby indicating
that the data points (dots) fall nearly along a fit line with
positive slope. Whereas, when r is close to -1, the rela-
tionship between the observed and estimated variables is
negative and the dots fall nearly along a fit line with neg-
ative slope. When r is close to zero, it implies a weak
relationship between the observed and estimated variables
and that the data points are scattered around the fit line and
most of data points are not in good agreement with the fit
line.

Results and interpretations

There are four parameters (geophysical well logs) consid-

ered as the inputs for the modeling process. They are bulk
density (DEN), compensated neutron porosity (CNL), Fig. 6 Prediction results a FFBP, b GRNN. Well#B
acoustic (AC) and deep induction resistivity (ILD). Pear-
son’ correlation (r) was used to evaluate the statistical well log and core measured porosity; with a significant
relationship that may exist between each well log and core (p B 0.01) difference from zero. In other words, the rela-
measured porosity. Table 1 shows the relationship between tionship existing between each well log and core measured
each well log and core measured porosity. As can be seen porosity, respectively, is statistically significant. Statisti-
in Table 1, there is significant correlation between each cally significant means that the observed sample dataset

Fig. 5 Cross-plots of predicted porosity against observed porosity. Performance from FFBP (a) and GRNN (b). Testing data Well#A

123
164 J Petrol Explor Prod Technol (2015) 5:157–166

Fig. 7 Prediction results a FFBP, b GRNN. Well#C

Fig. 8 Prediction results a FFBP, b GRNN. Well#D

provides ample evidence to reject the null hypothesis that While GRNN structure was 4 inputs, smoothness parame-
‘‘the population correlation coefficient is zero’’ (H0: q = 0) ter (r) = 0.02 and 1 output.
thereby concluding that q = 0. Thus, each well log Figure 4a, b illustrates cross-plots of predicted porosity
appears linearly correlated to measured porosity, respec- against observed porosity for training dataset. From a
tively. Therefore, in this study, Pearson’ correlation (r) has visual observation of Fig. 4a, b, there is a very positive
confirmed the accuracy of the four geophysical well logs as correlation between predicted porosity from the two neural
inputs parameters to the ANNs. networks scheme and observed porosity, respectively, as
To use the study datasets for ANN training some nor- shown in the alignment results (dots) obtained by the two
malization was performed. The output and all inputs were neural networks around the lines (fit line and ideal line).
normalized between 0 and 1. The data for this study This indicates satisfactory training by these networks.
involved different parameters that have dissimilar physical However, the neural network in generalized regression
meaning and units. To make sure that each variable is structure fit line approaches the ideal line closer than neural
treated similarly in the model, the data were normalized. network in back propagation structure fit line (Fig. 4a, b).
The next step was to develop porosity model, integrate Additionally, the results in Table 2 support the superiority
core porosity data (target) with well log data (inputs) using of GRNN training performance, since GRNN structure
ANN algorithms to establish a satisfactory model for the shows higher R2 and r values and lower MSE value, while
relationship between well log data and rock porosity. The FFBP structure indicates lower R2 and r values and higher
1043 data points from Well#A were randomly divided into MSE value. We can therefore conclude that in this study
training data (70 %) and testing data (30 %). The training GRNN model trains (learns) porosity better than FFBP
data were used in FFBP and GRNN training. While the model.
testing data (not trained) was used to estimate the pre- After training step was done, the two neural networks
diction ability of the models. were tested using testing data (not trained) from Well#A.
After several trials, the optimal architecture of FFBP Figure 5a, b shows cross-plots predicted values versus
was, 4 inputs, one hidden layer of 10 neurons and 1 output. observed values porosity for testing data (Well#A).

123
J Petrol Explor Prod Technol (2015) 5:157–166 165

Table 3 Statistical performance of GRNN and FFBP scheme of porosity and observed porosity (Well#B, Well#C and Well#D)
Methods Well#B Well#C Well#D
2 2
R r MSE R r MSE R2 r MSE

GRNN 0.966 0.982 0.370 0.961 0.980 0.309 0.986 0.992 0.128
FFBP 0.936 0.967 0.544 0.922 0.960 0.519 0.973 0.984 0.188

Analysis of the cross-plot in Fig. 5a, b depicts that the wells production. Porosity is essential in understanding the
two neural networks fit line, respectively, closely coin- crustal heterogeneity of a reservoir. In this light, geo-
cides with the ideal line. Additionally, the dots appear physicists, geologists and engineers are always trying to
alongside the lines (fit line and ideal line). This means find a cost effective, fast and robust method for accurate
that the networks have satisfactorily predicted porosity. A estimation.
visual check in Table 2, clearly depicts that neural net- In this study innovative effort was made to analyze and
work generalized regression scheme surpasses neural compare generalized regression neural network and feed-
network in back propagation scheme since, GRNN forward back propagation neural network in modeling
scheme produces higher R2 and r values and lower MSE porosity. The findings indicate that artificial neural network
value, while FFBP scheme suffers from lower R2 and is an appropriate tool for modeling porosity, despite the
r values and higher MSE value. We can therefore con- high degree of heterogeneity of reservoir in Zhenjing oil-
clude that in this study GRNN scheme fits porosity better field. Additionally, the geophysical well logs (DEN, CNL,
than FFBP scheme. AC and ILD) are significant parameters to be considered
After successful testing from Well#A, the two neural for developing a porosity model.
networks were also tested using data from Well#B, Well#C From all the results in this study we see that, neural
and Well#D. Figures 6, 7 and 8 visually illustrate the network generalized regression scheme is better than neu-
variation of observed porosity values and predicted ral network in back propagation scheme. This obviously
porosity values from the two models, thus providing a indicates GRNN model outperforms FFBP model. This
visual exhibition of accuracy. As it can be seen in Figs. 6, 7 assertion has also been echoed by several authors such as
and 8, neural network in generalized regression scheme is Specht (1991); Cigizoglua and Alp (2006); and Sun et al.
in better concordance with observed porosity values as (2008) in their respective research areas. These researchers,
compared to the neural network in back propagation have mentioned promising advantages of GRNN model
scheme. Table 3 presents the R2, r and MSE statistics for over FFBP model.
testing data from Well#B, Well#C and Well#D. From In conclusion GRNN gives better prediction accuracy in
Table 3, the results confirm the FFBP scheme weakness as predicting porosity than FFBP. The GRNN exhibited better
they exhibit lower R2 and r values and higher MSE value. precision with the core porosity data. Due to it great flex-
This means that the neural network in generalized regres- ibility and capability in dealing with non-linear problem in
sion scheme estimate porosity better than the neural net- actual situation, GRN network scheme can thus serve as a
work in back propagation scheme. cost effective approach for the petroleum industry by way
of reducing the necessity of coring because, it may allow
Summary and conclusions improved prediction in uncored intervals. Furthermore, this
method may be a very useful tool in aiding prediction of
The increasing success of ANN application mostly in many future wells.
techniques can be attributed to its power, adaptability and However, artificial neural network still has some limi-
simplicity. It can be useful to elucidate any complex, non- tations. For example, in network construction and adjusting
linear and dynamic reservoir parameter problems. This is of learning parameters, there are too many human inter-
because it does not require a priori information about the ferences. Nevertheless, all these problems are now being
functional shape to be estimated. investigated which can be expected to provide satisfactory
In the petroleum industry, rock porosity (as well as answers for future use.
permeability, lithology) is one of the major concern in
Open Access This article is distributed under the terms of the
reservoir characterization. It is identified during the geo- Creative Commons Attribution License which permits any use, dis-
physical exploration phase to estimate the capability of an tribution, and reproduction in any medium, provided the original
oilfield and to look for the optimal locations for drilling author(s) and the source are credited.

123
166 J Petrol Explor Prod Technol (2015) 5:157–166

References Hornik K, Stinchcombe MB, White H (1989) Multilayer feed forward

networks are universal approximators. Neural Netw
Abraham A (2005) Artificial neural networks, handbook of measuring 2(5):359–366
system design. In: Sydenham PH, Thorn R (eds) ISBN: 0-470- Mohaghegh S, Arefi R, Ameri S, Aminian K, Nutter R (1996)
02143-8, Wiley, Hoboken Petroleum reservoir characterization with the aid of artificial
Bishop C (1995) Neural networks for pattern recognition. Oxford neural networks. J Petrol Sci Eng 16(4):263–274
Press, Oxford Poulton MM (2002) Neural networks as an intelligence amplification
Celikoglu HB (2006) Application of radial basis function and tool: a review of applications. Geophysics 67(3):979–993
generalized regression neural networks in non-linear utility Specht DF (1991) A general regression neural network. IEEE Trans
function specification for travel mode choice modelling. Math Neural Netw 2(6):568–576
Comput Model 44:640–658 Sun G, Hoff SJ, Zelle BC, Nelson MA (2008) Nelson Development
Cigizoglua HK, Alp M (2006) Generalized regression neural network and comparaison of backpropagation and generalized neural
in modelling river sediment yield. Adv Eng Softw 37:63–68 network models to predict diurnal and seasonal gas and PM10
Hagan MT, Menhaj MB (1994) Training feed forward techniques concentrations and emissions from buildings. Trans ASABE
with the Marquardt algorithm. IEEE Trans Neural Netw 51(2):685–694
5(6):989–993 Van der Baan M, Jutten C (1992) Neural networks in geophysical
Haldorsen HH, Damsleth E (1993) Challenges in reservoir charac- applications. Geophysics 65(4):1032–1047
terization. AAPG Bull 77:541–551 White H (1992) Artificial neural networks. Approximation and
Handhel AM (2009) Prediction of reservoir permeability from wire learning theory. Blackwell, Cambridge
logs data using artificial neural networks. Iraqi J Sci 50(1):67–74 Wong PM, Gedeon TD, Taggart J (1995) An improved technique in
Haykin S (2007) Neural networks: a comprehensive foundation, Third porosity prediction: a neural network approach. IEEE-TGRS
edn. Prentice Hall, Inc., Upper Saddle River 33(4):971–980
Hecht-Nielsen R (1989) Theory of backpropagation neural networks. Yan J (2002) Reservoir parameters estimation from well log and core
In: Presented at IEEE Proceedings, international conference on data: a case study from the North Sea. Pet Geosci 8:63–69
neural network, Washington DC

123