Egyptian Journal of Petroleum (2017) 26, 763–778

H O S T E D BY
Egyptian Petroleum Research Institute

Egyptian Journal of Petroleum

www.elsevier.com/locate/egyjp
www.sciencedirect.com

FULL LENGTH ARTICLE

Prediction of permeability and porosity from well

log data using the nonparametric regression with
multivariate analysis and neural network, Hassi
R’Mel Field, Algeria
Baouche Rafik a,*, Baddari Kamel b

a
University of Boumerdes, Algeria
b
Department of Geophysics, University of Bouira, Algeria

Received 21 February 2016; revised 18 August 2016; accepted 17 October 2016

Available online 16 November 2016

KEYWORDS Abstract Most commonly, to estimate permeability, we can use values of porosity, pore size dis-
Well logs; tribution, and water saturation from logging data and established correlations. One benefit of using
Permeability; wireline log data to estimate permeability is that it can provide a continuous permeability profile
Multivariate statistics; throughout a particular interval.
Neural network; This study will focus on the evaluation of formation permeability for a sandstone reservoir in the
Hassi R’Mel; reservoir formations of Hassi R’Mel Field Southern from well log data using the multivariate meth-
Algeria ods. In order to improve the permeability estimation in these reservoirs, several statistical regression
techniques have already been tested in previous work to correlate permeability with different well
logs. It has been shown that statistical regression for data correlation is quite promising. We pro-
pose a two-step approach to permeability prediction that utilizes non-parametric regression in con-
junction with multivariate statistical analysis. First we classify the well log data into electrofacies
types. A combination of principal component analysis, model-based cluster analysis and discrimi-
nant analysis is used to characterize and identify electrofacies types. Second, we apply non-
parametric regression techniques to predict permeability using well logs within each electrofacies.
Three non-parametric approaches are examined via alternating conditional expectations (ACE),
generalized additive model (GAM) and neural networks (NNET) and the relative advantages
and disadvantages are explored. The results are compared with three other approaches to perme-
ability predictions that utilize data partitioning based on reservoir layering, lithofacies information
and hydraulic flow units. An examination of the error rates associated with discriminant analysis
for uncored wells indicates that data classification based on electrofacies characterization is more
robust compared to other approaches.

* Corresponding author.
E-mail address: [email protected] (B. Rafik).
Peer review under responsibility of Egyptian Petroleum Research Institute.
http://dx.doi.org/10.1016/j.ejpe.2016.10.013
1110-0621 Ó 2016 Egyptian Petroleum Research Institute. Production and hosting by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
764 B. Rafik, B. Kamel

These methods are tested and compared at the heterogeneous reservoirs in Triassic formations of
Hassi R’Mel. The results show that permeability prediction is improved by applying variable selec-
tion to non-parametric regression ACE while tree regression is unable to predict permeability.
In comparing the relative predictive performance of the three regression methods, the alternating
conditional expectations with ACE method appears to outperform the other two methods.
Ó 2016 Egyptian Petroleum Research Institute. Production and hosting by Elsevier B.V. This is an open
access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction Southern directed anticline structure is located at the Western

end of the Triassic province of Algeria (Fig. 1). The principal
The knowledge of permeability is critical to developing an studies carried out in this formation are mainly focalized
effective reservoir description. Permeability data can be in electrofacies characterization and non-parametric regres-
obtained from well tests, cores or logs. Normally, using well sion ACE techniques [7] and are proven to be successful in
log data to derive estimates of permeability is the lowest cost predicting permeability in highly heterogeneous sandy
method. In the Hassi R’Mel area (Fig. 1), the principal reser- reservoirs.
voir is located in the deposits of Triassic age. This field of The prediction of absolute permeability is a key element in
reservoir descriptions and has a direct impact on, among
others, [14], effective completion designs, successful water
injection programs and more efficient reservoir management.
Traditionally simple linear regressions of core porosity – per-
meability data or empirical formulae using various log
responses (usually porosity, clay volume and water saturation)
have been used to predict permeability. These empirical meth-
ods only apply locally and ignore the fact that there is no the-
oretical basis for a relationship between porosity and
permeability. In addition the scatter of the data about the
regression line is explicitly ignored and implicitly attributed
to measurement error or lower order variability in reservoir
characteristics. Amaefule et al. [1] state that a practical and
theoretically correct methodology is proposed for identifica-
tion of hydraulic flow units based on a modified Kozeny–Car-
man equation [6,5,2,12,1]. In doing so they make extensive use
of core data to describe the complex variations in pore geom-
etry and the identification of ‘‘flow zone indicators” (FZI)
from hydraulic units, which are, in turn, implicitly linked too
much improved permeability estimations [24].
This article investigates the usefulness of the mathematical
concepts and the predictive capabilities of multivariate analysis
[23] with a view to better estimation of facies using the tree
regression methods and to improving permeability estimates
for the Hassi R’Mel Field in Triassic Formations.
In this context, we utilize the stepwise algorithm method to
further improve permeability estimates. Stepwise regression is
a kind of stepwise algorithm used in linear regression. We first
demonstrate the availability of the stepwise algorithm by
applying stepwise regression [10,11] to Hassi R’Mel Southern
(Triassic). A method combining the stepwise algorithm and
ACE is proposed and applied to the Hassi R’Mel Field (Trias-
sic). Results are compared with those from regression without
variable selection.

2. Methodology

The applications of the results obtained during the facies anal-

ysis, starting from the available log data for the Triassic forma-
tions of Hassi R’Mel Southern made it possible, first, to define
ten electrofacies (sandstone, shale, dolomite, evaporate, ande-
Figure 1 Field location of Hassi R’Mel. Sands (yellow), lime- site and clay), in the wells of Triassic formations. Manual
stone (blue), shale (green), dolomite (Magenta), Silts (light facies analysis of Triassic well log data was carried out in ten
brown), Sebkha (dark brown).
Prediction of permeability and porosity from well log data 765

surveys [3]. This allowed us, in a first stage, to define ten This approach was developed in two stages for the predic-
lithologies. Next, after developing a model on a well using tion of permeability using the non-parametric regression in
the ‘‘Petrolog” software, and checking it on other wells, a conjunction with multivariate statistical analysis [17]. In a first
semi-automatic data processing was carried out on seven other step, a classification of log data in many electrofacies types is
wells. The facies analysis in the Hassi R’Mel Field is shown in made in accordance with the unique characteristics of well
Fig. 2. log data and measures reflecting minerals and lithofacies in

Figure 2 Facies analysis in Triassic of Hassi R’Mel [3].

766 B. Rafik, B. Kamel

the interval studied. A combination of principal component Table 1a and b compares regression errors for the three models
analysis, cluster analysis based on a model and discriminant used for developing the correlations. These errors are summa-
analysis [16] is used to characterize and identify the types of rized in terms of mean squared error (MSE) and mean abso-
electrofacies. Second, we apply nonparametric regression tech- lute error (MAE) during regression. The GAM appears to fit
niques to predict permeability using well logs in each electrofa- the data better compared to the other models.
cies. Three non-parametric approaches are examined:ACE,
GAM, and NNET and the relative advantages and disadvan- 4. Analyses
tages are explored.
4.1. Electrofacies analysis
3. Data available
Recent methods used in electrofacies classification are mainly
The data presented in this analysis have been gathered from 5 based on techniques for identifying groups from responses log-
wells. We left out two cored wells (HRS-7 and HRS-8) to ver- ging data with the same characteristics. (Fig. 3). For this pur-
ify our correlations using blind tests. The data presented in this pose, the process results of principal component analysis can
analysis have been gathered from 7 wells. The consideration of be shown in Table 2. First principal component (PC1) appears
the data quality and field-wide availability, a suite of well logs to indicate porosity (NPHI) of the formation while second
is selected for the analysis. In this field we have 7 well logs: cali- principal component (PC2) shows a stronger correlation with
per, gamma ray (GR), and three different resistivity (LLD, density (RHOB) readings. The eigenvectors of the covariance
P
LLS, and MSFL), acoustic transit-time (DT), neutron (NPHI), matrix ( ) provides coefficients of the principal component
and density (RHOB), and photoelectric logs (PEF). Among transformation.
the 7 well logs, only 6 logs (GR, RLLD, DT, NPHI, RHOB,
and SW) are chosen for characterizing the electrofacies groups. 4.1.1. Principal component analysis
After partitioning of well log responses into electrofacies Principal component analysis (PCA) is probably the oldest and
groups, statistical regression techniques are applied to model best known of the techniques of multivariate analysis. It was
the correlation between permeability and well log responses first introduced by Pearson [19], and developed independently
within the partitioned groups. In this study, three non- by [13]. Like many multivariate methods, it was not widely
parametric techniques are examined using ACE, GAM, and used until the advent of electronic computers, but it is now well
NNET and their relative predictive performances are assessed. entrenched in virtually every statistical computer package.
In neural network modeling, the 927 sample data set from 7 Principal components constitute an alternative form of dis-
wells was divided into two subsets for training and supervising. playing the data [15], thereby allowing better knowledge of its
The supervising data set is used for testing whether the neural structure without changing the information. In addition,
net can generalize from the training data set. For each electro- because the total variance in a data set can be defined as the
facies group, we modeled the networks containing the optimal sum of the variances associated with each principal compo-
number of nodes in the hidden layer which produce the least nent, the first few principal components that explain most of
mean square error for the supervising data set. the variation in the original variables are often useful to reveal

Table 1a Shows the Mean Square Errors (MSE) and Mean Absolute Errors (MAE) for porosity.
GR (API) DT (ls/ft) Rhob (cc) RLLD (X m) NPhi (%) SW (%)
MSE 1420.377 4485.003 177.4659 349071.5 230.4 224.171
MAE 33.329 66.021 11.2626 109.9 13.4 13.136
MRSE 0.496 0.696 34.1510 2.5 630535.9 4185.298
MRAE 0.681 0.829 4.8987 0.9 212.8 45.430
R2 0.275 0.343 0.4992
0.1 0.5 0.494
Erreurs quadratiques moyennes (MSE) et erreurs absolues moyennes (MAE) pour la porosité. MSE: Erreur quadratique moyenne (Mean
square error); MAE: Erreur absolue moyenne (Mean absolute error).
Erreurs quadratiques moyennes (MSE) et erreurs absolues moyennes (MAE) pour la perméabilité. Mêmes notations qu’au Table 7a.

Table 1b Shows the Mean Square Errors (MSE) and Mean Absolute Errors (MAE) for permeability. Same notations as Table 7a.
GR (API) DT (ls/ft) Rhob (cc) RLLD (X m) NPhi (%) SW (%)
MSE 193258.4 171866.0 223655.5 514328.3 225,150 225,010
MAE 317.8 289.0 355.7 415.0 358 358
MRSE 110.7 24.7 45558.6 5268.2 474,475,956 4,928,654
MRAE 7.5 3.5 156.8 32.1 4785 1295
R2 0.1 0.4
0.1
0.1
0
0
Erreurs quadratiques moyennes (MSE) et erreurs absolues moyennes (MAE) pour la porosité. MSE: Erreur quadratique moyenne (Mean
square error); MAE: Erreur absolue moyenne (Mean absolute error).
Erreurs quadratiques moyennes (MSE) et erreurs absolues moyennes (MAE) pour la perméabilité. Mêmes notations qu’au Table 7a.
Prediction of permeability and porosity from well log data 767

Figure 3 Electrofacies – Well Hassi R’Mel – Triassic. (Red) paleosols, (Blue) Sebkha, (Green) Floodplain, (Magenta) Fluvial, (Black)
Dolerites.

Table 2 First principal component analysis.

Variables PC1 PC2 PC3 PC4 PC5 PC6 PC7
GR 0.810222
0.221781 0.112247
0.326747 0.254331 0.332007 0.008781
DT 0.928652
0.085377
0.060521 0.064263 0.149098
0.298393 0.106090
RHOB 0.802908 0.415180
0.075180 0.373197 0.141656
0.053781
0.122791
RLLD
0.771005 0.191981
0.209594 0.454785 0.309703 0.135997 0.059369
NPHI
0.017782
0.613081 0.664005 0.427526 0.008350
0.002590
0.007632
SW 0.431749 0.779061 0.310636 0.177623
0.216682 0.163703 0.069734
CPOR 0.525146
0.521029
0.491840 0.356037
0.239513 0.162650 0.016158
Eigen value 3.220466 3.220466 3.220466 3.220466 3.220466 3.220466 3.220466
Contribution (%) 46.0067 46.0067 46.0067 46.0067 46.0067 46.0067 46.0067
Total variance (%) 46.00666 46.00666 46.00666 46.00666 46.00666 46.00666 46.00666

the structure in the data. This can reduce the dimensionality of 4.1.3. Discriminant analysis
the problem and complexity in the cluster and discriminant It is a statistical analysis to predict a categorical dependent
analysis. variable (called a grouping variable) by one or more continu-
ous or binary independent variables (called predictor vari-
4.1.2. Cluster analysis ables). Discriminant analysis is used when groups are known
This method can provide better performance compared to tra- a priori (unlike in cluster analysis). Each case must have a
ditional methods such as the nearest neighbor and k-mean score on one or more quantitative predictor measures, and a
clustering methods, which are often not capable of identifying score on a group measure. In simple terms, discriminant func-
the groups (sebkha, river channel, paleosols, floodplains and tion analysis is classification the act of distributing things into
dolerite) that are either overlapping or of varying sizes and groups, classes or categories of the same type.
shapes. This technique is based on the assumption that an individ-
Another advantage of the model-based approach is that ual sample arises [8] from one of g populations or groups
there is an associated Bayesian criterion for assessing the (P1, . . . Pg, g > 2). If each group is characterized by a
model [22]. Bayesian criterion is associated with this method group-specific probability density function fc(x) and the prior
to evaluate the model [22] where the number of clusters with- probability of the group/c is known, then according to the
out the subjective judgments necessary in other conventional Bayes theorem, the posterior distribution of the classes given
cluster analysis techniques can be estimated. the observation x is as follows (Eq. (1)):
768 B. Rafik, B. Kamel
!!
pc pðx=cÞ pc fc ðxÞ X
nh X
ni
pðc=xÞ ¼ ¼ apc fc ðxÞ ð1Þ yj ¼ fo aj þ whj fh ah þ wih xi ð4Þ
pðxÞ pðxÞ
h¼1 i¼1

where y is the vector of output variables, x is the vector of

4.1.4. Permeability correlation
input variables, 6 is the bias term, wij are the connection
The non-parametric regression techniques which do not weights on the link from i to j node, nh is the number of hidden
require a priori assumptions about the functional forms to nodes, and ni is the number of input variables.
model the data in this case may be applied if electrofacies The original activation function proposed by McCulloch
are known. In this case correlations between permeability and Pitts [18] is a step function, the so called threshold func-
and log responses can be established for each kind of electro- tion (with f(x) = I (x > 0)). However, the artificial neuron
facies. There are generally three different nonparametric model has been expanded to include other functions such as
approaches that address this issue: GAM, ACE, and Nnet. the sigmoid, piecewise linear, and Gaussian. In this study, a
logistic function is used as the activation function, which gives
4.2. The Generalized additive models (GAM) values that range from 0 to 1 as (Eq. (5)):
expðxÞ
Regression model has the general form (Eq. (2)): fðxÞ ¼ ð5Þ
1 þ expðxÞ
  X
p
E y=x1 ; x2 ; . . . ; xp ¼ a þ u1 ðxl Þ þ e ð2Þ A critical aspect in a neural network is the learning process
i¼1 of forcing a network to yield a particular output (response) for
where vL is the predictor and uL is function of predictors. Thus a specific input (signal). For multilayer feed forward neural
additive models replace the problem of estimating a function networks, a more powerful supervised learning algorithm,
of a p-dimensional variable v by one of estimating p separate called back propagation, can be employed to recursively adjust
one-dimensional functions, uL . Such models are attractive if the connection weights so that the difference between the pre-
they fit the data since they are far easier to interpret than a dicted and the observed outputs is as small as possible.
p-dimensional multivariate surface.
Algorithm: The response transformation models generalize the 4.2.2. Permeability in uncored wells
additive model by allowing transformation of the response vari- Many of the Petrophysical models used for simulation studies
able y. The models have the following general form as (Eq. (3)): are based on the classical approach of cross plotting the loga-
rithms of permeability versus porosity (Fig. 8) and then, by fit-
X
p
hðyÞ ¼ a þ /l ðxÞ þ e ð3Þ ting a regression line on this plot, predicting the permeability
l¼1 through the reservoir rock with correlation coefficient
(R2 = 0.029). This approach is critical when used to model
The ACE algorithm, originally proposed by Breiman and
permeable rocks, as it implies two misleading concepts. First,
Friedman [4], provides a method for estimating optimal trans-
it considers the relationship between the logarithms of core
formations for multiple regressions that result in a maximum
permeability versus core porosity as linear. Secondly, using
correlation between a dependent (response) random variable
log porosities on this plot to predict the permeabilities would
and multiple independent (predictor) random variables. Such
imply a scaling agreement between the macroscopic level (core
optimal transformations can be derived by minimizing the vari-
plug) and the megascopic level (log data). Discretizing the
ance of a linear relationship between the transformed response
reservoir into units such as layers and blocks, and assigning
variable and the sum of transformed predictor variables.
values of all pertinent physical properties to these blocks will
give a better agreement with the reservoir heterogeneity.
4.2.1. Feed-forward neural networks (NNET)
A feed forward neural network is an artificial neural network 4.3. Classification based on stratigraphical zones
where connections between the units do not form a directed
cycle. This is different from recurrent neural networks. The feed
The Saharan Triassic consisted of varied continental environ-
forward neural network was the first and simplest type of artifi-
ments, namely, fluvial, flood plain, lake, Sebkha or Evaporate,
cial neural network devised. In this network, the information
and Eolian [3]. This Triassic is composed of three Formations:
moves in only one direction, forward, from the input nodes,
I, II and III. At the top of Formation I, the lower series of strata
through the hidden nodes (if any) and to the output nodes.
of the Triassic, there are intercalations of volcanic rocks repre-
Usually, the feed forward neural net is arranged in multiple
sented by dolerites. Sedimentation interspersed with periods of
layers as shown in Fig. 7: Six input layers, two output layers
no deposition is thought to have occurred during the Triassic
and more hidden layers. Each layer contains a number of
continental; resulting in a deposition of ground-level strata and
nodes (also called neuron’s processing units) which are con-
the development of a more or less intense pedogenesis. At the
nected to each node in the preceding layer by simple weighted
top of Formation II, the eruptive on the level of the lower mem-
links. Except for nodes in the input layer, each node multiplies
ber IIa at the base and constituted primarily of dolerites but the
its specific input value by the corresponding weight and then
roof is primarily by a channel in a fine sandy filling. The member
sums up all the weighted inputs. Sometimes a constant (the
IIb is characterized by fluvial facies of channels prevailing and
‘bias’ term) can be involved in the summation. The final output
evolving to the top in a complex of playa or evaporates.
from the node is calculated by applying an activation function
Formation III is characterized primarily by an evaporate
(transfer function) to the sum of the weighted inputs as shown
facies of sebkha type halite.
in Eq. (4):
Prediction of permeability and porosity from well log data 769

4.4. Classification based on Lithofacies: Error rate from all nent (PC1) appears to indicate porosity of the formation while
classifications second principal component (PC2) shows a stronger correla-
tion with gamma ray readings.
The facies studies started with a core analysis, and was carried Fig. 5a shows the projection of the well log variables on the
out on the Triassic formations of the Southern Hassi R’ Mel. factors plane (1x2), where the contribution of the factor 1 is of
This required a core analysis in the laboratory of Sonatrach 46.01% and 21.71% for the second factor.
(Algiers), as well as the combination of the different well log data: For the PC1 and PC2 (Table 3a), with porosity, the compo-
Gamma Ray, neutron and sonic log, their fittings in sequences nents are given by,
and their vertical evolution. The study of the sedimentary figures, PC1 ¼ 0:3994GR
0:2456LLD þ 0:4688NPHI
representing the hydrodynamic and physical conditions of the

0:4329RHOB
0:5153DT
0:1417SW
sedimentation environment, was also taken into account.
Well log responses can be also partitioned using lithofacies þ 0:3328CPOR
that are usually determined by core descriptions and thin sec-
tion analysis for all the wells (2, 4, 7, 8, 9, 11 and 12). The Field PC2 ¼ 0:3426GR
0:4122LLD þ 0:1544NPHI
of Hassi R’Mel can be described by 10 different lithofacies
þ 0:3402RHOB þ 0:0746DT þ 0:6803SW
types as shown in (Table 6).
The crossval function (Matlab 12a, 2012) can estimate the
0:3156CPOR
misclassification error for both LDA (Linear Discriminant Anal- Fig. 5b shows the projection of the well log variables on the
ysis) and QDA (Quadratic Discriminant Analysis) using the factors plane (1  2), where the contribution of the factor 1 is
given data partition for the two wells HRS-7 and HRS-8, where of 48.22% and 35.61% for the second factor.
the QDA has a slightly larger cross-validation error than LDA. For the PC1 and PC2 (Table 3b), with permeability KH,
the components are given by,
5. Results
PC1 ¼ 0:4818GR
0:2642LLD þ 0:3283NPHI
5.1. Principal component analysis þ 0:4406RHOB
0:1455DT þ 0:5331SW þ 0:3009KH

In the scatter plot (Fig. 4), the relationship between reservoir PC2 ¼ 0:0930GR
0:4067LLD þ 0:4906NPHI
properties and the 3 major principal components generated
from the 5 well logs can be explored. First principal compo-
0:3862RHOB þ 0:6197DT þ 0:0280SW
0:2259KH

Figure 4 Scatter plot of GR, DT, RLLD, RHOB, RHOCMA and NPHI, SW.
770 B. Rafik, B. Kamel

Table 3b Eigenvector spreadsheets of Hassi R’Mel Upper

Triassic for core Permeability (KH)
Eigenvector spreadsheet Number of components is 2
Variable number Component 1 Component 2
DT (ls/ft) 1
0.145534 0.619764
GR (API) 2 0.481838 0.093089
LLD (X mm) 3
0.264245
0.406760
NPHI (cc) 4 0.328384 0.490641
RHOB (g/cc) 5 0.440670
0.386256
SW (cc) 6 0.533144 0.028047
CPERM (mD) 7
0.300927 0.225976

basis of a measure of similarity or dissimilarity between the

groups (Fig. 6a, b). In this study, the grouping based on a
model, a hierarchical clustering technique by agglomeration
is used.

5.3. Discriminant analysis

The observations should be allocated into the group with the

maximal posterior probability p(c/x). The discriminant analy-
sis used in this study requires prior classification (Table 4) of
the data into relatively homogenous subgroups whose charac-
teristics can be described by the statistical distributions [9] of
the grouping variables associated with each subgroup.

5.4. Permeability correlation

The optimal transformations are derived solely based on the

data sets and can be shown to result in a maximum correlation
in the transformed space. The transformations do not require a
priori assumptions of any functional form for the response or
predictor variables and thus, provide a powerful tool for
exploratory data analysis and correlation. The correlations
of vectors in matrix representing the predictor values (well
logs) and the predicted value (core permeability KH or CPerm
and core porosity CPor) are summarized in Table 5a and b.
Figure 5 (a) Projection of the variables on the factor-plane
(1  2) with porosity CPor. (b) Projection of the variables on the
factor-plane (1  2) with permeability KH.
5.5. Feed forward neural network

In this study, we used a typical feed forward model that con-

5.2. Cluster analysis sists of one input layer, one hidden layer and one output layer.
In addition, the optimum number of hidden nodes is deter-
The goal of the cluster analysis is to establish classes of data mined by trial and error.
groups, homogeneous and isolated from the outside on the This technique is applied to highly shaly sand reservoirs in
the Triassic formations of Hassi R’Mel, Sahara Algeria. Con-
ventional multiple regressions did not provide satisfactory per-
Table 3a Eigenvector spreadsheets of Hassi R’Mel Upper meability and porosity predictions in this field. The superior
Triassic for core Porosity (CPor). predictive ability of our proposed approach is verified using
Eigenvector spreadsheet Number of components is 2 blind tests. We also, compared our results with the neural net-
work method for permeability and porosity prediction based
Variable number Component 1 Component 2
lithofacies characterization and the results are better (Fig. 7).
GR (API) 1 0.399453 0.342616
DT (ls/ft) 2 0.491891 0.074640
5.6. Permeability in uncored wells
RHOB (g/cc) 3
0.432947 0.340221
RLLD (X mm) 4
0.245687
0.418269
NPHI (cc) 5 0.468838 0.154481 Some flow units in Well HR-7 and the permeability porosity
SW (cc) 6
0.141704 0.680331 relationship constructed for the unique HFUs are illustrated
CPOR (cc) 7 0.332846
0.315642 in Figs. 8 and 9. The samples in the scatter plot are grouped
Prediction of permeability and porosity from well log data 771

Figure 6 (a) Cluster analysis of facies environments (Root 1 vs Root 2). (b) Cluster analysis of facies environments (Root 1 vs Root 3).

by HFUs after calculating FZI in uncored well using Eqs. (6)– this well had been logged only and did not have core data,
(11) from well log data, permeability can be determined for while in reality core data were available at well. This was done
each HFU (mean FZI value) using Eq. (12). only to check how accurately the HFU method would predict
A result of calculated permeability versus core permeability permeability in these well if they had not been cored. As
with depth in uncored wells is shown in Fig. 10. These pre- shown, the permeability profiles of the log-derived HFU agree
dicted permeability profiles were obtained by assuming that with core data.

Table 4 Classification matrix rows: observed classifications columns: predicted facies classifications. Well: (HRS-7).
ldaCVerr qdaCVerr Perf Tr_Perf Val_Perf Test_Perf
HRS-7 HFUs 0.4052 0.4125 0.7425 0.5266 1.0002 0.9255
LITHO 0.4444 0.3919 0.7535 0.6496 1.0285 0.9622
Facies 0.4450 0.3871 0.7337 0.6894 0.9048 0.7693
HRS-8 HFUs 0.4129 0.4135 0.7140 0.4828 0.7797 0.8358
LITHO 0.4444 0.3919 0.7535 0.6496 1.0285 0.9622
Facies 0.4189 0.3221 0.1954 0.2181 0.1415 0.1439
772 B. Rafik, B. Kamel

5.7. Classification based stratigraphy 5.8. Classification based Lithofacies

The estimation of the permeability by application of the corre- The Means with error plot of core porosity and core perme-
lation equation established in each area can be applied directly ability, associated with the lithofacies can be shown in
to the log responses in corresponding predefined area without Fig. 12a and b. An examination of this plot will show that vari-
the use of discriminant analysis. All the Permeability and ables evaporate, shale, shale dolomite, shale sand and low sand
porosity predictions based on the zonal classification for the shale are the least influential in determining the first principal
well HRS-7 as: KH (core permeability) can be shown in component while sands plays the most significant important
Fig. 10: KH-NN (Neural permeability) and PERMC (FHU role. The predictive performance of different classifiers is
permeability). The same results were obtained for the well defined as an error rate which represents the fraction of train-
HRS-8 (Figs. 10 and 11). ing or a test set that have been misclassified. The error rates

Table 5a Correlations of vectors in matrix representing the predictor values (well logs) and the predicted value (core permeability KH
(CPERM).
DT (ls/ft) GR (API) RLLD (X mm) NPHI (cc) RHOB( g/cc) SW (cc) KH (mD)
DT 1.000000
0.089607
0.339964 0.571517
0.758730
0.224298 0.309288
GR
0.089607 1.000000
0.330645 0.580730 0.563389 0.731145
0.256745
RLLD
0.339964
0.330645 1.000000
0.530421
0.011256
0.457333 0.047765
NPHI 0.571517 0.580730
0.530421 1.000000 0.046107 0.554921
0.090472
RHOB
0.758730 0.563389
0.011256 0.046107 1.000000 0.708240
0.495770
SW
0.224298 0.731145
0.457333 0.554921 0.708240 1.000000
0.380541
KH 0.309288
0.256745 0.047765
0.090472
0.495770
0.380541 1.000000

Table 5b Correlations of vectors in matrix representing the predictor values (well logs) and the predicted value core porosity (CPor).
GR (API) DT (ms/ft) RHOB (g/cc) (X mm) NPHI (cc) SW (cc) CPOR (%)
GR 1.000000 0.619490 0.138185
0.313039 0.153794 0.150801 0.275141
DT 0.619490 1.000000 0.075433
0.327866 0.102853 0.084975 0.342660
RHOB 0.138185 0.075433 1.000000 0.002034 0.998923 0.999061 0.499198
RLLD
0.313039
0.327866 0.002034 1.000000
0.006388
0.014186
0.120998
NPHI 0.153794 0.102853 0.998923
0.006388 1.000000 0.998692 0.515081
SW 0.150801 0.084975 0.999061
0.014186 0.998692 1.000000 0.493719
CPOR 0.275141 0.342660 0.499198
0.120998 0.515081 0.493719 1.000000

Figure 7 Feed forward neural net in multiple layers.

Prediction of permeability and porosity from well log data 773

Figure 8 Core porosity/core permeability vs. Depth, HRS-7, in triassic formations of HRS. (R2 = 0.717 and Y = 0.314X2.602).

valuated for this context are summarized in (Tables 7 and 6) sents the expected prediction error on an independent set. To
for the wells HRS-7 and HRS-8. It shows that a simpler model this end, the discriminant analysis based on lithofacies data
may get comparable or better performance than a more com- performs worse than classifiers based on the electrofacies
plicated model of other classifications (Table 7). The error rate information.
associated with the classification of HFUs is the highest among Among the three non-parametric methods, the ACE model
all techniques considered in this study. appears to be the best models in terms of their predictive of
The quadratic discriminant analysis (QDA) for the data porosity and permeability. The Neural Network model tends
used shows the computed ‘‘ldaCVerr” and the ‘‘qdaCVerr” to overestimate or underestimate in some intervals of the for-
errors for linear and quadratic discriminant analysis and repre- mation. Especially the poor predictive performance in the
blind well HRS-8 indicates that a trained neural network will
not necessarily provide accurate predictions for the data that
were not used during training. Therefore, the GAM model
produces good results for most intervals in the blind wells.

6. Comparison for other techniques for permeability prediction

The Hydraulic Unit concept [1] was selected for subdividing

the reservoir into distinct petrophysical types. Each distinct
reservoir type has a unique Flow Zone Indicator (FZI) value.
According to [20], a hydraulic flow unit is a continuous
body over a specific reservoir volume that practically possesses
consistent petrophysical and fluid properties, which uniquely
characterize its static and dynamic communication with the
wellbore.
This technique is based on a modified Kozeny–Carman [5]
(cited in Amaefule et al. [1]) and the concept of mean hydraulic
radius (Eq. (6)):
( ) !
1 u3e
k¼ ð6Þ
2s2 S2g ð1
ue Þ2
Sgv: may also be defining as the surface area of grains exposed
to fluid per unit volume of solid material.
Flow zone indicator depends on geological characteristics
Figure 9 Scatter plot of log RQI versus log PHIZ when one
of the material and various pore geometry of a rock mass;
HFU is used to fit the whole data. Well: HRS-7 (R2 = 0.738 and
hence, it is a good parameter for determining hydraulic flow
Y = 0.262X0.684).
774 B. Rafik, B. Kamel

units (HFUs). Flow zone indicator is a function of reservoir Taking the logarithm of both sides of (Eq. (11)) yields:
quality index and void ratio. LogRQI ¼ LogFZI þ LoguZ ð11Þ
Amaefule et al. [1] addressed the variability of Kozeny’s
constant by dividing Eq. (1) by the effective porosity, ue and On a log–log plot of RQI versus uz, all samples with similar
taking the logarithm: FZl values will lie on a straight line with unit slope, in Fig. 9,
Defining the flow zone indictor FZI (lm) as (Eq. (7)): the value of the FZI constant can be determined from the
intercept of the unit slope straight line at uz = 1. Samples that
1
FZI ¼ pffiffiffiffiffi ð7Þ lie on the same straight line have similar pore throat attributes
Sg s Fs and, thereby, constitute hydraulic unit. The permeability of a
Reservoir quality index RQI (lm) as (Eq. (8)): sample point is then calculated from a pertinent HFU using
sffiffiffiffiffi the mean FZI value and the corresponding sample porosity
K using the following (Eq. (12)):
RQI ¼ 0:0314 ð8Þ
ue
ue
K ¼ 1014  FZI2 ð12Þ
Normalized porosity uz (fraction) as (Eq. (9)): ð1
ue Þ2
 
ue However, it is worth-mentioning that given the true poros-
uZ ¼ ð9Þ
1
ue ity and true HFUs (based on core data), the predicted perme-
Therefore (Eq. (10)) becomes: ability shows almost perfect agreement with the true
permeability. So, the principal difficulty appears to be the iden-
RQI ¼ FZI  uZ ð10Þ tification of hydraulic flow units in uncored wells. In Table 7,

Figure 10 Permeability and porosity predictions based on the zonal classification for the well HRS-7 as: KH (core permeability).
Legend: Composite log of well log predictions for HR-167. U1, U2, M1, M2, L1, L2 and L3: in-situ formations; GR [API]: gamma ray;
Rhob [g/cm3]: density; Nphi [%]: neutron porosity; DT [ms/ft]: sonic log; RT [X m]: true resistivity, CPerm [mD]: permeability; subscript
nn relates to predicted values from neural network. The horizontal arrows represent the levels of samples used in the training. Fuzzy logic
Bin number distributions within the Hassi R’Mel Triassic reservoir for the associated (a) core porosity (CorPor or CPor) and (b) core
permeability (CPerm); gamma ray (GR), sonic (DT), neutron porosity (Nphi), density (Rhob), water saturation (Sw) and true resistivity
(RT); Bin number: predicted values. For notations see Appendix A and Nomenclature. VGLAUC-CPX: Volume of Glauconite in
complex lithology, VSND–CPX: Volume of Sandstones in complex lithology, PHSW-CPX: Water volume in complex lithology, VBW-
CPX: Volume of Bulk Water in complex lithology.
Prediction of permeability and porosity from well log data 775

Figure 11 Permeability and porosity predictions based on the zonal classification for the well HRS-8 as: KH (core permeability).
Legend: Composite log of well log predictions for HR-167. U1, U2, M1, M2, L1, L2 and L3: in-situ formations; GR [API]: gamma ray;
Rhob [g/cm3]: density; Nphi [%]: neutron porosity; DT [ms/ft]: sonic log; RT [X m]: true resistivity, CPerm [mD]: permeability; subscript
nn relates to predicted values from neural network. The horizontal arrows represent the levels of samples used in the training. Fuzzy logic
Bin number distributions within the Hassi R’Mel Triassic reservoir for the associated (a) core porosity (CorPor or CPor) and (b) core
permeability (CPerm); gamma ray (GR), sonic (DT), neutron porosity (Nphi), density (Rhob), water saturation (Sw) and true resistivity
(RT); Bin number: predicted values. For notations see Appendix A and Nomenclature. VGLAUC-CPX: Volume of Glauconite in
complex lithology, VSND–CPX: Volume of Sandstones in complex lithology, PHSW-CPX: Water volume in complex lithology, VBW-
CPX: Volume of Bulk Water in complex lithology.

the predictive performances of the four approaches are com- of Fit for uncored wells indicates that data classification
pared using linear and quadratic errors for both the two wells based on electrofacies characterization look more robust
HRS-7 and HRS-8. compared to other methods.
All the results of Permeability and porosity prediction are 2. The simplest approach uses flow zones or reservoir layer-
plotted in the Fig. 12a and b where the three methods used ing. Other approaches have used lithofacies information
in this work shows the well log data correlation with porosity identified from cores, electrofacies derived from well logs,
and permeability calculated from the HFU (PERMC) and and the concept of HFUs.
others porosity/permeability as (KH-NN) from FFNET with 3. Non-parametric regression offer better technical means to
core data (KH). better approximate the permeability in the shaly sand reser-
voirs, even in intervals where there is no relationship
7. Discussion and conclusion between permeability and logs.
4. The best performance of permeability prediction result of
the discrimination ability of the pattern recognition and
The use of nonparametric regression methods in the Electrofa-
cies Characterization study allows more significant results in reliability correlation models and this has, in advance;
the model clay sandstone oil reservoir. Therefore, explained the methods for predicting porosity and perme-
ability calculated from partitioning and correlation of data.
1. The predicted permeability profiles obtained from the data The identification of errors from HFU and lithofacies can
recorded in the depth, by assuming that this well had been thus lead to incorrect results in the predictions of porosity
logged only and did not have core data, while in reality core and permeability.
To improve further the permeability and porosity
data were available at well. An examination of the error
rates associated with Discriminant Analysis for core poros- predictions,
ity and core Permeability showing the summary Goodness  The results of treatment should be checked carefully.
776 B. Rafik, B. Kamel

200
(a)
0

-200

-400

CPHI
-600

-800

-1000

-1200

EVAPORATE

SANDS

SHALE

SHALE-SANDS
ANDESITES

LOW SD SHALE
SHALE-DOLOM
LITHO

Mean
Mean±0,95 Conf. Interval
LITHO

200

(b)
0

-200

-400
KH

-600

-800

-1000

-1200
LITHO

EVAPORATE

ANDESITES
SANDS

SHALE

SHALE-SANDS

LOW SD SHALE
SHALE-DOLOM

Mean
Mean±0,95 Conf. Interval
LITHO

Figure 12 (a) Means with error plot of porosity versus lithology. (b) Means with error plot of permeability versus lithology.

Table 6 Facies description represented by 10 different litho-  The number of the Hydraulic flow units must be opti-
facies types in Hassi R’Mel Triassic. mum in the current study.
 Calculations FZI and HFU settings must first be made.
Facies number (Facino) Facies description
 We need to ensure that a better quality of permeabilities
1. Halite is obtained practically and good effect compaction cor-
2. Shale halite rections have been introduced into the petrophysical a-
3. Sand
nalysis stage.
4. Low shale sand
5. Shale sand
6. Dolomitic shale
Acknowledgements
7. Dolomitic shale sand
8. Low dolomitic shale sand
9. Sandy shale I would like to take this opportunity to express my deepest
10. Dolerites gratitude and appreciation to the people who have given me
their assistance throughout my studies and during the prepara-
Prediction of permeability and porosity from well log data 777

Table 7 Predictive performances of different classifiers are compared using linear and quadratic errors for the two wells HRS-7 and
HRS-8.
ldaCVerr qdaCVerr Perf Tr_Perf Val_Perf Test_Perf
HRS-7 HFU 0.4052 0.4125 0.7425 0.5266 1.0002 0.9255
LITHO 0.4444 0.3919 0.7535 0.6496 1.0285 0.9622
Facies 0.4450 0.3871 0.7337 0.6894 0.9048 0.7693
HRS-8 HFU 0.4129 0.4135 0.7140 0.4828 0.7797 0.8358
LITHO 0.4444 0.3919 0.7535 0.6496 1.0285 0.9622
Facies 0.4189 0.3221 0.1954 0.2181 0.1415 0.1439

tion of this paper. I would especially like to thank SPE mem- PERMC: Permeability obtained from FHU (mD)
bers for their helpful comments. Matlab codes developed for KH-NN: Core permeability neural network (mD)
this study to simulate and predict porosity and permeability VGLAUC-CPX: Volume of Glauconite in complex
are available upon request. lithology
VSND–CPX: Volume of Sandstones in complex lithology
Appendix A PHSW-CPX: Water volume in complex lithology
VBW-CPX: Volume of Bulk Water in complex lithology
K: is the permeability (lm2), X, Y: Coordinates (UTM)
ue: is the effective porosity (fraction), HRS: Hassi R’Mel Southern
FZI: is the Flow Zone Indicator (lm), K: Permeability (mD)
ACE: Alternating Conditional Expectation, Kv: Vertical permeability (mD)
GAM: General Additive Models, Kh: Horizontal permeability (mD)
DT: Sonic travel time (Delta Time, lsec/ft), U: Porosity (%)
NNET: Neural Network, Vsh: Shale fraction (%)
PHIT: Total Effective Porosity Swi: Initial water saturation (%)
RHOCMA: Matrix Density l: Viscosity (cp)
SUM: Summation Vap: Apparent velocity (m/s)
LdaCVerr: Linear discriminant CV error Vac: Actual velocity (m/s)
QdaCVerr: Quadratic discriminant CV error P: Pressure (psi)
Perf: Performance L: Length
Tr_Perf: Train Performance DLL: Dual Laterolog (X m)
Val_Perf: Validation Performance LDT: Litho density tool (Pe)
Test_Perf: Test Performance Rhob: Bulk density (g/cc)
SS: Sandstones CNL: Compensated neutron log (%)
LS: Limestones GR, SGR: Gamma ray (API)
Dol: Dolomite SGR: Spectrometry Gamma ray (API)
Dtf = Fluid transit time (ls/ft) MSFL: Micro Spherically Focused Log (X m)
Dt = Transit time (ls/ft) BHC: Borehole compensated (ls/ft)
qb = Bulk density (g/cc) CAL: Caliper (in.)
qf = Fluid density (g/cc) nnt: Neural network
UNf = Neutron fluid (l.p.u.%)) mlr: Multi linear regression
UN = Neutron (l.p.u.%)) MSE: Mean square error
Root1 = Principal Composant Axe 1 (PC1) MAE: Mean absolute error
Root2 = Principal Composant Axe 2 (PC2) MRSE: Mean root Square error
Root3 = Principal Composant Axe 3 (PC3) MRAE: Mean root absolute error
DtMAA = Apparent matrix transit time – including shale R2: correlation coefficient
(ls/ft)
qMAA = Apparent matrix density – including shale (g/cc)
UTAA = Apparent total porosity – including shale (%) References
CorPor (Cpor): Core Porosity (%)
CPerm: Core Permeability (mD) [1] J.O. Amaefule, M. Altunbay, D. Taib, D.G. Kersey, D.K.
Keelan, Enhanced reservoir description: using core and log data
KH: Core Permeability (mD)
to identify hydraulic (flow) units and predict permeability in
PHIR: Porosity obtained from FHU (%)
uncored intervals/wells, Paper SPE 26436 Prepared for
NET-CPX: Net Volume in complex lithology Presentation at the 68th Annual Technical Conference and
CPHI-NN: Core porosity neural network (%) Exhibiton of SPE Held in Houston, Texas, 3–6 October, 1993.
PHIS: Porosity sonic (%) [2] J.W. Amyx, D.M. Bass Jr., R.L. Whiting, Petroleum Reservoir
PHIT-CPX: Total porosity in complex lithology (%) Engineering, McGraw-Hill Book Co, New York City, 1960.
778 B. Rafik, B. Kamel

[3] R. Baouche, A. Nedjari, Africa Geosci. Rev. 17 (2) (2010) 151– [14] J.L. Jensen, L.W. Lake, Optimization of regression-based
219. porosity-permeability predictions, in: CWLS 10th Symposium,
[4] L. Breiman, J. Friedman, D. Pregibon, Y. Vardi, A. Buja, R. Calgary, Alberta, Canada, 1985.
Kass, E. Fowlkes, J. Kettenring, J. Am. Stat. Assoc. 80 (391) [15] Jiang Xie, Improved Permeability Prediction Using Multivariate
(1985) 580–619. Analysis Methods, Thesis Master of Science, Submitted to the
[5] P.C. Carman, Flow of Gases Through Porous Media, Academic Office of Graduate Studies of Texas A&M University, 2008.
Press Inc, New York City, 1956. [16] Jude O. Amaefule, Mehmet Altunbay, Djebbar Tiab, David G.
[6] P.C. Carmen, Trans. AIChE. 15 (1937) 150–166. Kersey, Dare K. Keelan, Enhanced reservoir description: using
[7] A. Datta-Gupta, G. Xue, S.H. Lee, Nonparametric core and log data to identify hydraulic (flow) units and predict
Transformations for Data Correlation and Integration: From permeability in uncored intervals/ wells, Paper SPE 26436
Theory to Practice, in: Reservoir Characterization: Recent Presented at the 68th Annual Technical Conference and
Advances, AAPG Datapages, 1999. Exhibition of the Society of Petroleum Engineers held in
[8] J. Friedman, A. Barron, X. Xiao, G. Golubev, R. Hasminskii, L. Houston, Texas, 3–6 October, 1993.
Breiman, A. Bujaz, D. Duffy, T. Hastie, R. Tischirani, Ann. [17] S.H. Lee, K. Arun, A. Datta-Gupta, SPE Reservoir Eval. Eng. 5
Stat. 19 (1) (1991) 1–141. (3) (2002) 237–248.
[9] J. Friedman, C. Roosen, An introduction to multivariate [18] W.S. McCulloch, W. Pitts, Bull. Mathe. Biophys. 5 (1943) 115.
adaptive regression splines, Stat. Methods Med. Res. 4 (3) [19] K. Pearson, Phil. Mag. 2 (6) (1901) 559–572.
(1995) 197–217. [20] D. Tiab, Advances in Petrophysics, Vol. 1-Flow Units, Lecture
[10] Guoping Xue, A. Datta-Gupta, Peter Valko, T. Balsingame, Notes Manual, University of Oklahoma, 2000.
Optimal transformations for multiple regression: application to [21] A. Timur, An Investigation of Permeability, Porosity, & Residual
permeability estimation from well logs, SPE 35412 Presented at Water Saturation Relationships For Sandstone Reservoirs, The
the Improved Oil Recovery Symposium, Tulsa, Ok, 21 April, Log Analyst IX (4), SPWLA-1968-VIXn4a2, 1968.
1996. [22] D. Wang, M. Murphy, J. Data Sci. 2 (4) (2004) 329–346.
[11] T. Hastie, R. Tibshirani, Generalized Additive Models, [23] W.A. Wendt, S. Sakurai, P.H. Nelson, Permeability prediction
Chapman & Hall/CRC, New York, 1990. from well logs using multiple regression, in: L.W. Lake, H.B.J.
[12] J.R. Hearst, P.H. Nelson, F.L. Paillet, Well Logging for Physical Caroll (Eds.), Reservoir Characterization, Academic Press, New
Properties, John Wiley & Sons, New York City, 2000. York, 1985.
[13] H. Hotelling, J. Educ. Psychol. 24 (1933) 417–441, 498–520. [24] G. Xue, A. Datta-Gupta, P. Valko, T. Blasingame, SPE
Reservoir Eval. Eng. 12 (2) (1997) 85–94.