Journal of Petroleum Science and Engineering 36 (2002) 149 – 158

www.elsevier.com/locate/jpetscieng

Porosity images from well logs

Anna Ilcéa Fischetti, André Andrade *
Centro de Geociências, Universidade Federal do Pará, Rua dos Pariquis 2902/604, Cramacao, Belem, Para, Brazil
Received 22 January 2002; accepted 5 August 2002

Abstract

Porosity images are graphical representations of the lateral distribution of rock porosity estimated from well log data. We
present a methodology to produce this geological image entirely independent of interpreter intervention, with an interpretative
algorithm approach, which is based on two types of artificial neural networks. The first is based on neural competitive layer and is
constructed to perform an automatic interpretation of the classical qB – /N cross-plot, which produces the log zonation and
porosity estimation. The second is a feed-forward neural network with radial basis function designed to perform a spatial data
integration, which can be divided in two steps. The first refers to well log correlation and the second produces the estimation of
lateral porosity distribution.
This methodology should aid the interpreter in defining the reservoir geological model, and, perhaps more importantly, it
should help him to efficiently develop strategies for oil or gas field development. The results or porosity images are very similar
to conventional geological cross-sections, especially in a depositional setting dominates by clastics, where a color map scaled in
porosity units illustrates the porosity distribution and the geometric disposition of geological layers along the section.
D 2002 Elsevier Science B.V. All rights reserved.

Keywords: Well log interpretation; Neural networks; Porosity prediction; Reservoir characterization

1. Introduction We introduce the concept of interpretative algo-

In recent years, many studies have shown the
applicability and efficiency of artificial neural net-
works to solve a great number of geophysical problems
(Ali and Chawathe, 2000; Banchs and Michelena,
2000; Chang et al., 2000; Jamialahmadi and Javadpour,
2000; Crocker et al., 1999; FitzGerald et al., 1999). A
common point in the majority of these studies is the
need of interpreter intervention to present the right
answer in the training phase of neural networks (Vander
Baan and Jutten, 2000).
* Corresponding author.
E-mail addresses: ilcea@ufpa.br (A.I. Fischetti),
aandrade@ufpa.br (A. Andrade).
rithm, which should integrate all available information
from geology, geophysics and well logging to produce
an autonomous and geologically plausible interpreta-
tion of this data on a field-wide basis.
Here, we apply this concept to a geological well log
interpretation problem. Our purpose is to obtain the
porosity image or cross-section representing the lateral
distribution of subsurface rock porosity. This interpre-
tative algorithm achieves an efficient integration of
well log data and classical well log interpretation
procedures with artificial neural networks.
Conventionally, geological well log interpretation
tasks require a large degree of interpreter expertise
(Luthi, 2001). However, we show that the interpreta-
tive algorithm can simulate the decision processes

0920-4105/02/$ - see front matter D 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 0 - 4 1 0 5 ( 0 2 ) 0 0 2 9 2 - 9
150 A.I. Fischetti, A. Andrade / Journal of Petroleum Science and Engineering 36 (2002) 149–158
involved in this well log interpretation problem, par-
ticularly in a depositional setting dominated by clastics
(sandstones and shales) and where the diagenetic
process has not produced large differences between
various parts of the field.
In the next section, we present an application to
actual well log data from the Lagunillas formation in
the Lake Maracaibo basin, located in western Vene-
zuela. The geologic evolution of the Maracaibo basin
relevant to development of the petroleum systems
includes sedimentation and tectonics from Cretaceous
to Holocene time. The Lagunillas formation is deriving
from the lower-middle Miocene, which consists of
sandstone, shale, and coal deposited in fluvial to
deltaic environments (Talukdar and Marcano, 1994).
2. Methodology
We start with comments about the foundations of
artificial neural networks. We do not intend to provide
a complete discussion about the theory of artificial
neural networks, but rather to focus on their theoretical
basis on why they are indispensable in understanding
of interpretative algorithm here presented. We finish
this section by discussing the steps involved in the
interpretative algorithm design.
2.1. Competitive neural networks
The simplest architecture of an artificial neural
network with competitive layer corresponds to one
input layer composed only by sensorial units full
connected with one processing layer composed by
competitive neurons. Each competitive neuron is con-
nected with all sensorial units by synaptic weights
(weight connections), which will be changed in the
network training process.
In competitive learning, the competitive neurons
compete with each other for each input data to produce
a winner which then carries the output signal, and at a
given time, only one competitive neuron is active
(Haykin, 2001). The output signal corresponding to
the winner neuron is assumed to be 1 (one) and the
output signal of all other competitive neurons is null.
The most common competitive learning uses the
Kohonen rules (Kohonen, 1989), where the compet-
itive neuron that most resembles the input vector wins
the competition and has its synaptic weights moved
closer to the input vector. Thus, each competitive
neuron has its synaptic weights progressively migrated
close to a group of input vectors, and after some
iterations the network stabilizes, having each neuron
at the center of the cluster it represents. If a neuron
does not answer to a particular input vector, then no
learning will happen in that neuron or its synaptic
weights are unchanged.
The data analysis realized by this type of neural
network is commonly called feature-discover scheme,
which performs clustering or quantization of the input
space (Saggaf and Nebrija, 2000) and can be used to
detect statistically significant patterns in the input
space and to classify them (Haykin, 2001).
2.2. Radial basis function neural networks
A basic radial basis function neural network, in-
volves three layers with totally different functions.
Sensorial units that connect the neural network to its
environment constitute the input layer. The second
layer, or hidden layer, applies a non-linear transforma-
tion of the input space to the hidden space. The transfer
functions, or radial basis functions, in the hidden
neurons are similar to the multivariate Gaussian den-
sity function
!
2
Nx  cj N
uj ðxÞ ¼ exp ð1Þ
2r2j
where j refers to jth hidden neuron, Nx  cjN2 is the
Euclidean distance between input vector (x) and the
center (cj) of radial function and rj is its width.
The output layer is linear, supplying the answer of
the neural network to activation patterns applied to the
input layer. The operation performed in the output
layer is given by
X
N
yk ðxÞ ¼ wkj uj ðxÞ; ð2Þ
j¼1
where yk is the kth output neuron for the input vector x,
wkj is the weight connection between the kth output
neuron and the jth hidden neuron.
In this paper, we limit our considerations to the
Gaussian functions with a constant width, which is the
same for all neurons in the hidden layer. The adjust-
 A.I. Fischetti, A. Andrade / Journal of Petroleum Science and Engineering 36 (2002) 149–158
ment of weight connection in Eq. (2) is performed
using a least square solution after the selection of
centers and the width of radial basis functions from
the training set samples.
The radial basis neural networks exhibit the same
properties as error-backpropagation neural networks
such as generalization and robustness; they also have
the advantage of fast learning and ability to detect
outliers during estimation (Devaraj et al., 2002).
2.3. Porosity images
To obtain the porosity image, we construct an
interpretative algorithm comprised by three independ-
ent and sequential procedures involving artificial
neural networks. The first takes as its input the density
and neutron porosity logs to produce the so-called
porosity zonation log. This processed log, obtained in
a reference borehole, will dictate the construction of
the training set necessary in the second process. The
next process receives a set of conventional well logs
(gamma ray, density, neutron porosity, clay volume)
and the porosity zonation log as input to obtain the
correlation lines or the lines that show the reservoir
stratigraphy. Finally, the last procedure uses all infor-
mation generated in the early processes to produce the
cross-section image of porosity distribution.
2.4. Porosity zonation log
The porosity zonation log is composed by one curve
scaled in porosity units for each lithology crossed by
the borehole. We construct an interpretative algorithm
based on competitive layer neural network architecture
to obtain the porosity zonation log. This interpretative
algorithm will try to simulate the interpretation of the
classical qB – /N cross-plot, which receive as input the
neutron porosity log(/N) in the x-axis and the density
log(qB) in y-axis.
We assume that the neutron porosity log is sand-
stone calibrated. Thus, the clean sand line in qB – /N
cross-plot is represented by a straight line, with a 45j
inclination, where each point characterized by an
ordinate pair (/N,qB), represents one porosity value
for clean sand samples.
Each depth point in each input log represents one
ordinate pair (/N,qB), or one point in the qB – /N
plane. If we plot these points from a selected depth
151
interval, we will see that they are arranged in clusters.
From the interpretative criteria of the qB – /N cross-
plot and considering the sand – shale depositional
model, we can identify two important clusters by
inspection of qB –/N plane. The clean sand samples
are placed around the clean sand line—the sand
points. The sand points may show a wide spread
around the clean sand line because of the presence
of clay or hydrocarbon in the pore space, or the matrix
effect—the presence of some amount of feldspar, for
example. The other important cluster for our propose
is located far from the clean sand line, in the lower
right—the shale points, which represent the input
points, whose matrix is dominated by shale.
Our first interest here is to construct a competitive
layer neural network to produce the log zonation, or
separate the points with matrix dominated by sand from
those dominated by shale. In other words, we want to
identify the cluster in the qB –/N plane corresponding
to shale.
From the knowledge of the clean-sand line position
and the relative location of shale cluster, we con-
structed a competitive neural network with two input
neurons; each one receives, as input, the density log
and the neutron porosity log for each depth point. In
our case, we use four neurons in the competitive layer;
representing, the clean sand (neuron 1), sand –clay
(neuron 2), sand – hydrocarbon (neuron 3), and shale
(neuron 4). The training process starts with the syn-
aptic weights of all neurons close to the center of the
qB – /N plane and ends when they stabilize close the
centroid of each cluster. Of course, by virtue of the
wide spread of sand points, and to avoid the misinter-
pretation of sand – clay cluster as shale cluster, we use
more neurons to characterize the sand cluster.
After the competitive training, we retain the syn-
aptic weights of the neuron that stabilizes far from the
clean sand line in the shale direction—the shale
neuron. The synaptic weights of this neuron represent
one characteristic point in qB – /N plane—the shale
point, used to correct the sand porosity for clay content
(Ellis, 1987) and plays an important function in our
approach of log zonation, where we assume that this
particular neuron identifies the shale cluster correctly.
Now, we are able to produce the log zonation and
estimate the porosities of a well log depth interval. To
do it, we introduce the concept of shale line and give a
geological interpretation of the competitive neural
152 A.I. Fischetti, A. Andrade / Journal of Petroleum Science and Engineering 36 (2002) 149–158

network processing. In the first step, we define the

shale line in the qB – /N cross-plot as a line parallel to
clean-sand line and passing by the shale point (iden-
tified above). The next step is the association of one
porosity value to each neuron in the competitive layer.
It means that we assume a convenient porosity scale
and associate each ordinate pair (/N,qB) of each point
of clean-sand line and the shale line as the synaptic
weight connecting each neuron in the competitive
layer to the input neurons. Thus, the number of
neurons in the competitive layer reflects the precision
in the estimated porosity. In our approach, we estimate
the porosity with precision of 0.01 porosity units (PU)
and assume for sands and shales a porosity interval
ranging from 0 to 40 PU.
To show the construction and behavior of this
competitive layer, consider, for example, a competitive
layer composed by 82 neurons. In this case, the
porosity estimation has the precision of 1 PU. The
neurons are labeled in crescent order. The first 41 Fig. 1. Block diagram representing the interpretative algorithm with
neurons represent the points in the clean sand line competitive neural network to obtain the porosity zonation log.
and the last 41 neurons represent the points in the shale
line. If the neuron labeled by the number 20 wins the
competition for a particular input data, it means that this where /D is the porosity calculated from density log,
data comes from sand and has a total porosity equal to qbM is the matrix density (2.65, for sand), qbF is the
19 PU. For shale, we take the same mechanism. For fluid density (1.1, for salt water) and qb is the density
example, if the neuron 65 wins the competition the log. Then, we correct the porosity values in the sand
porosity is equal to 23 PU. Thus, the neural network porosity zonation log by (Crain, 1986)
operation is quite simple. Each pair of input logs forces
the competition between all neurons. The winner /D /NSH  /N /DSH
/¼ ð4Þ
neuron has synaptic weights closest to the input pair. /NSH  /DSH
The position of the winner neuron in the competitive
layer shows the lithology and the porosity value of the where / is the effective porosity and /NSH, /DSH are
depth point correspondent to this input pair. This pro- the synaptic weights of shale neuron.
cess is resumed in the block diagram shows in Fig. 1.
As for the results of all operations involved in this 2.4.1. Correlation lines
interpretative algorithm, we plot the porosity values The correlation lines connect the depth points
associated to sand in the first track and this one associated to the top and bottom of the same layer
associated is to shale in the second track. Thus, we crossed by different boreholes. We assume that a pair
have constructed the porosity zonation log scaled in of correlation lines identifies the top and bottom parts
total porosity. The discontinuities in both tracks mark of a layer.
the vertical limits of the layers. We present an interpretative algorithm to do the
To refine the porosity calculation, we have to cor- well log correlation. It has a radial basis function
rect the sand porosities for clay content (Ellis, 1987). neural network as its kernel and takes four conven-
First, we calculate the density porosity by tional well logs (gamma ray, density, neutron porosity,
clay volume) and the porosity zonation log as input.
 b M  qb
/D ¼ ð3Þ Generally, to make the well correlation, we need the
bM  qbF aid of a correlation datum that should represent a
 A.I. Fischetti, A. Andrade / Journal of Petroleum Science and Engineering 36 (2002) 149–158
regional event that can be identified in the logs of all
boreholes. We use as correlation data the shale layers
that are identified in the porosity zonation logs.
The neural network architecture is composed of five
input neurons that receive one depth point of each log
at a time and the output layer has one neuron that
answer 1 (one) for true correlation and null for false
correlation. We use a routine that add one radial basis
neuron at a time to the hidden layer until the sum-
square error of network output and the desired output
falls beneath the error goal.
Five columns compose the input matrix, where
each column represents one input log with N samples
take from a depth window in the neighborhood of
each shale layer in the shale porosity zonation log.
The desired answer matrix is built with five equal
columns. Each column is formed in such a way that
where the shale porosity zonation log is different from
zero, its value is equal to 1 (one) and equal to zero for
others in the depth window, with N samples. The work
performed in the training phase corresponds to map-
ping each interval of non-zero value of shale porosity
zonation log window to unit in all input logs. It means
that we need to promote an isolated training for each
shale layer.
We work with two boreholes in the operation
phase—the reference well and the test well. This
process can be interpreted as finding the shale layers
in the test well, as they were trained in the reference
well. In practice, it proves a rather delicate issue to be
realized by the neural network itself.
The generalization of the neural network may
produce more than one answer, but for very simple
geological settings, we may avoid it using two rules.
The first is the relative depth position of the shale layer
in each borehole, where it is assumed as the absence of
faults and/or discontinuity planes. The second takes a
statistical approach. The actual network output is a
vector that contains values close to zero for non-
correlation and values close to one when a correlation
is established. We take the median value. If the median
value is closer to the unit, then there is lateral equiv-
alence between these shale layers. These criteria are
good enough to avoid ambiguity in the shale layer
correlation or correlation lines crossover and assures
that layers of different thickness will be correctly
correlated. The process continues for the other shale
layers in the test well.
153
The definition of the correlation lines for shale may
establish the ancient horizontal depositional plane,
modified by later tectonic events. Thus, if we work
in simple geological settings, the correlation lines may
represent very well the lithostratigraphic equivalen-
cies between the same layers crossed by different
boreholes in an oil or gas field.
2.4.2. Porosity image
The porosity image is a graphical representation of
the rock porosity lateral distribution estimated from
well log data. It is closely related to common geo-
logical cross-sections, but here, colors represent a
porosity scale, which may also permit us to infer the
lithology and the layer’s geometry.
We construct an interpretative algorithm based on
radial basis function neural network to perform the
lateral porosity mapping. Now, we explore the inter-
polation capacity of the radial basis function neural
network. This mapping is guided by the direction
dictated by the correlation lines that link all boreholes
involved in the process. It works as if we put a number
of hypothetical wells between each two actual wells,
so we can estimate the porosity distribution and pre-
sent the result as a cross-section, showing the geo-
metric behavior of layers and the lateral porosity dis-
tribution as the third dimension.
When the neural network process is finished, we
have one matrix—the porosity matrix, where the
columns represent the borehole positions and the lines
represent the depths. Each element of this matrix
represents one porosity value. Zeros are added where
no porosity values exist. The porosity matrix is con-
verted to RGB or gray level matrices by conventional
image processing methods to create the porosity
image.
3. Results
We apply the methodology detailed here in three
wells drilled in the Lagunillas formation of Lake
Maracaibo basin, Venezuela. The wells are referred
to as Well A, Well B and Well C. The absolute depths
and relative horizontal distances are not indicated in
the figures, but the relative horizontal displacement
among these wells and the depth position of each
logged interval are kept unchanged.
154 A.I. Fischetti, A. Andrade / Journal of Petroleum Science and Engineering 36 (2002) 149–158

Fig. 2. qB – /N crossplot showing an example of automatic interpretation. The black dot shows the shale point. The shale line and the clean sand
line are indicated.

To exemplify the presented methodology to obtain crosses represent the input ordinate pairs (/N,qB)
the porosity zonation log, we show, in Fig. 2, the qB – from neutron porosity and density log for each
/N cross-plot obtained from Well A, where the depth sample. The black point is the shale point,

Fig. 3. Porosity zonation log for Well A. First track is the gamma ray log. The second is the sand porosity zonation log and the third is the shale
porosity zonation log.
 A.I. Fischetti, A. Andrade / Journal of Petroleum Science and Engineering 36 (2002) 149–158 155

Fig. 4. Porosity zonation log for Well B. First track is the gamma ray log. The second is the sand porosity zonation log and the third is the shale
porosity zonation log.

close the shale line and the other line is the clean sand tracks, two curves are in fact generated: one being the
line. porosity of sands with zero values where there are
The resulting porosity zonation log of Well A is shales, and the other showing the porosity values for
shown in Fig. 3. As can be seen in the second and third shales, with zero values where sand is present. The

Fig. 5. Porosity zonation log for Well C. First track is the gamma ray log. The second is the sand porosity zonation log and the third is the shale
porosity zonation log.
156 A.I. Fischetti, A. Andrade / Journal of Petroleum Science and Engineering 36 (2002) 149–158

Fig. 6. Example of training set construction for a radial basis neural network. The horizontal lines highlight a depth interval used in the
correlation process.

porosity zonation logs for the Well B and Well C are agreement of lithological separation, considering a
shown in the Figs. 4 and 5, respectively. The gamma prior geological knowledge of occurrence of only
ray logs are also shown in order to consider the good clastics lithologies, in order to avoid gamma ray log

Fig. 7. Correlation lines linking the shale layer in the gamma ray logs of three wells in Lake Maracaibo, Venezuela.
 A.I. Fischetti, A. Andrade / Journal of Petroleum Science and Engineering 36 (2002) 149–158 157

Fig. 8. A view of lateral porosity estimation. Each black log represents a porosity log for three wells in Lake Maracaibo, Venezuela.

misinterpretation. As can be seen in the gamma ray according to the correspondence defined by correlation
logs, many sand – shale horizons are classified as sand lines. We show in Fig. 8 the results of the interpolation
in the porosity zonation logs. process performed by the neural network. The logs
The well log correlation is accomplished through a representing the actual porosity log for each borehole
random choice of a reference borehole. Its well logs are marked. The other lines represent the interpolated
will be used in the radial basis function neural network
training. We show in Fig. 6 an example of the training
set construction, with data taken from Well A. In our
tests, the choice of reference borehole did not influence
the correlation result or the depth of correlation lines.
In Fig. 7, we present the depths of the correlation lines
for shale layers in these three wells, obtained from the
network operation with the training set showed in Fig.
6. These lines correspond to the top and bottom of the
shales. These results are in close agreement with
previous studies (Luthi et al., 1995; Bryant et al.,
1998). There are, in fact, many more correlation lines
for the clean sands but these are not shown for the sake
of clarity.
Once the depth correspondence of all zones is
established, we carry out a simple non-linear interpo-
lation of the porosities along the correlation lines using Fig. 9. Porosity image, where the gray scale shows the porosity
the radial basis function neural network, trained with variations. Dark gray represents low porosity values and light gray
porosity values of the three wells by adjusting the depth represents high porosity values.
158 A.I. Fischetti, A. Andrade / Journal of Petroleum Science and Engineering 36 (2002) 149–158
logs. Well A is on the left-hand side; Well C is on the
right-hand side and Well B in the middle of the section.
The correspondent porosity image is shown in Fig.
9. We choose a gay level scale to represent the porosity
values distribution along the section. Black indicates
zero porosity values or the vertical limits of each log
interval. The light color indicates high porosity values
that are associated with shales. The gray color close to
the middle of the section shows dirty sandstone, which
loses its clay content toward the sides of the section. At
the bottom of the section is a sand body that shows a
decrease in porosity values from the left to the right-
hand side of the section.
4. Conclusions
We introduce the concept of interpretative algo-
rithms, which may be a new concept in well log
interpretation. We do not present a complete treatment,
but rather an example of the kind of work that can be
done by this class of computational algorithms.
The interpretative algorithm uses two types of
neural networks as a kernel of a decision-making
procedure, in order to provide a powerful data reduc-
tion tool and an interactive memory of the reservoir
interpretation. The interpretative algorithm can keep
the early porosity interpretation and integrate a new
data set in a very efficient manner. It allows the
interpreter to establish porosity cross-sections and
ultimately build a reservoir model in a very easy way
and with a large degree of automation.
The methodology presented here can easily be
expanded further to include more sophisticated and
geologically more relevant input logs such as borehole
images, nuclear magnetic resonance logs or geochem-
ical logs. It may also be extended to include other
geological output parameters such as mineralogy, bed-
ding or chronostratigraphy of the layers.
Acknowledgements
The authors would like to acknowledge the support
extended by CNPq and ANP/FINEP-UFPa for this
work.
References
Ali, M., Chawathe, A., 2000. Using artificial intelligence to predict
permeability from petrographic data. Computers and Geoscien-
ces 26, 915 – 925.
Banchs, R., Michelena, R., 2000. Well log estimates and confidence
intervals by using artificial neural networks. CD-rom of the
Society of Exploration Geophysicists Annual Meeting.
Bryant, I.D., Baygün, B., Herron, M.M., Matteson, A., Ramamoor-
thy, R., Stewart, L., Tariq, S.M., Coll, C., Cordova, P., Villar-
roel, H.G., Hernández, L., Jimenez, Z., León, K., Rondón, L.,
1998. Integration of old and new measurements to optimize
redevelopment of the lower Lagunillas Reservoir of Bloque
IV, Lake Maracaibo, Venezuela. SPE Reservoir Evaluation and
Engineering, June.
Chang, H.C., Chen, H.C., Durrans, S.R., Kopaska-Merkel, D.C.,
2000. Lithofacies identification using multiple adaptive reso-
nance theory neural networks and group decision expert system.
Computer and Geosciences 26 (5), 591 – 601.
Crain, E.R., 1986. The Log Analysis Handbook. PennWell Books,
Tulsa, USA.
Crocker, H., Fung, C.C., Wong, K.W., 1999. The stag oil field
formation evaluation—a neural network approach. Australian
Petroleum Production and Exploration Association Journal 38,
451 – 459.
Devaraj, D., Yegnanarayana, B., Ramar, K., 2002. Radial basis
function networks for fast contingency ranking. Electrical Power
and Energy Systems 24, 387 – 395.
Ellis, D.V., 1987. Well Logging for Earth Scientists. Elsevier, USA.
FitzGerald, E.M., Bean, C.J., Reilly, R., 1999. Fracture – frequency
prediction from borehole wireline logs using artificial neural
networks. Geophysical Prospecting, 1031 – 1044.
Haykin, S., 2001. Redes Neurais: Princı́pios e Práticas Bookman,
SP, Brasil.
Jamialahmadi, M., Javadpour, F.G., 2000. Relationship of perme-
ability, porosity and depth using an artificial neural network.
Journal of Petroleum Science and Engineering 26, 235 – 239.
Kohonen, T., 1989. Self-Organization and Associative Memory.
Springer-Verlag, Berlin. 312 pp.
Luthi, S., 2001. Geological Well Logs—Their Use in Reservoir
Modeling. Springer-Verlag, Berlin, 367 pp.
Luthi, S., Bryant, I., Villaroel, H., 1995. Well log correlation using
backpropagation neural networks. IV International Congress of
Brazilian Geophysical Society 2, 733 – 736.
Saggaf, M.M., Nebrija, E.M., 2000. Estimation of lithologies and
depositional facies from wire-line logs. AAPG Bulletin 84 (10).
Talukdar, S.C., Marcano, F., 1994. Petroleum Systems of the Mar-
acaibo Basin, Venezuela. The petroleum system—from source
to trap. AAPG Memoir 60.
Vander Baan, M., Jutten, C., 2000. Neural networks in geophysical
applications. Geophysics 65 (4), 1032 – 1047.