You are on page 1of 17

Journal of Petroleum Science and Engineering 214 (2022) 110573

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering


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

A novel approach by integrating the core derived FZI and well logging data
into artificial neural network model for improved permeability prediction
in a heterogeneous gas reservoir
Nasser Alizadeh a, *, Negin Rahmati a, Adel Najafi a, Edmund Leung b, Pooria Adabnezhad a
a
Petro Pars Oil and Gas Company, Tehran, Iran
b
RESMAN, Abu Dhabi, United Arab Emirates

A R T I C L E I N F O A B S T R A C T

Keywords: Predicting and extrapolating the permeability between wells to obtain the 3D distribution for the geological
Permeability prediction model, is a crucial and challenging task in reservoir simulation. Permeability is influenced by both digenetic
Heterogeneity characteristics and depositional factors like sorting and grain size. Hence, a reliable model should consider these
Artificial neural networks
characteristics for prediction of permeability. Grouping the rocks into different hydraulic flow units (HFU) or
Polynomial regression models
Discrete rock types
discrete rock types (DRT), improves the identity of the reservoir characteristics and provide a more accurate
Flow zone indicator permeability prediction. Multi variable regression models and Artificial Neural Networks (ANN) were applied in
this study to correlate core permeability and porosity with well logs to predict permeability logs. It was observed
that the accuracy of the models diminished in heterogeneous reservoirs, where there is a wide permeability
distribution.
In this study, we are presenting a novel approach to predict permeability in heterogeneous oil and gas res­
ervoirs. In this method the core permeability and porosity data are categorized using the concept of DRT and the
probability density functions are used to investigate the relationships between the logs and DRT groups. The ANN
model is applied to correlate the core derived flow zone indicator (FZI) with wire-line logging data with a single
key well to predict K-logs. In this approach one single well, which contains all DRT groups is considered as a key
well to develop and train the ANN model. It was observed that ANN model exhibits better prediction perfor­
mance in heterogeneous reservoirs when it is developed and trained on single well data containing all DRT
groups. This approach can capture heterogeneity in the reservoirs where it has been applied successfully to
predict permeability in an actual heterogeneous carbonate gas reservoir.

Credit author statement challenging task. A poor prediction of permeability will result in inef­
ficient and unreliable dynamic models, thus, reducing the accuracy of
Nasser Alizadeh: conceptualization, methodology, supervision, these models for describing and modeling the past, current and future
original draft preparation, validation. Negin Rahmati: data prepara­ performance of oil and gas reservoirs. The classical approach in the oil
tion, visualization, data curation, software, investigation. Adel Najafi: industry was to predict permeability from a single well log attribute. The
visualization, software, writing original draft, validation and data Kozeny Carman or Wyllie and Rose equations (Wyllie and Rose, 1950)
curation. Edmund Leung: review and editing. Pooria Adabnezhad: were often used to predict permeability from a porosity log. These
software. equations have adjustable variables with values depending on the
sorting and geometry of rock grains (Rose and Bruce, 1949). Wendt et al.
1. Introduction (1986) reported that by including other well log attributes than porosity
and using multiple linear regression approach, the correlation coeffi­
Permeability is a crucial parameter in reservoir characterization and cient between the estimated and actual permeability increases. Helle
predicting its 3D distribution in heterogeneous reservoirs is a et al. (2001) have made similar suggestions. They suggested that instead

* Corresponding author.
E-mail address: Alizadeh.Na@ppars.com (N. Alizadeh).

https://doi.org/10.1016/j.petrol.2022.110573
Received 6 January 2022; Received in revised form 22 April 2022; Accepted 23 April 2022
Available online 27 April 2022
0920-4105/© 2022 Elsevier B.V. All rights reserved.
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Table 1
The statistical parameters of the core porosity (fraction) and permeability (mD).
Parameter Core plugs Min Max Average Variance Std. Deviation

Porosity 10,387 0.00005 0.423 0.094 0.007 0.085


Permeability 10,387 0.0003 1849.64 23.72 10,455.66 102.25

of using a single well log, a group of several logs provide a better available on prediction of permeability in tight carbonate gas conden­
permeability prediction because of the existing correlations between sate reservoirs, which are among the most challenging reservoirs to
permeability and other well log parameters. Lacentre and Carrica develop in the world (Vafaie et al., 2021; Rashid et al., 2015). Since
(2003); Yao and Holditch (1993) introduced a method to estimate developing new methodologies to accurately predict permeability is
permeability based on well logs and core data. Many other researchers important, further analysis and examination of machine learning and
also attempted to establish mathematical relations between perme­ regression methods to find reliable, fast and accurate models are of great
ability and other petrophysical parameters and well logs by using sta­ importance.
tistical methods such as multivariate regression analysis and proxy In this study, we are presenting a novel approach based on the
modeling (Balan et al., 1995; Xue et al., 1997). These researches were artificial neural networks (ANN) to predict the 3D distribution of
helpful in identifying influential parameters on permeability, but were permeability in heterogeneous tight gas condensate reservoirs using the
not successful to provide a general workflow for accurate prediction of petrophysical log data and FZI from core data. In this approach, the core
permeability in all reservoirs and their applicability should also be permeability and porosity data are categorized using the concept of
checked in other field studies. Although these techniques have provided DRT. The ANN method is then applied to correlate the core derived FZI
good results in some cases, they require the existence of many core with wire-line logging data to predict permeability. It was observed that
measurements, which is expensive to achieve in practice. In addition, the current ANN models provide good results in homogeneous sandstone
the accuracy of these correlations diminishes when dealing with het­ reservoirs but they don’t deliver satisfactory results in heterogeneous
erogeneous reservoirs. carbonate reservoirs. Contrary to the existing ANN models, the novel
Recently, intelligent approaches including artificial neural networks approach presented in this paper applies to a single well, which contains
(ANNs), fuzzy clustering and least square support vector machines all DRT groups to develop and train the ANN model. In this study, it is
(LSSVM) have been proven as effective tools for prediction of reser­ demonstrated that the model delivers satisfactory results in heteroge­
voirpermeability. Finol et al. (2001) developed a rule based fuzzy model neous reservoirs when it is trained on the data of a single well containing
for the prediction of petrophysical rock parameters. Lim and Kim (2004) all DRT groups. This approach enables the ANN model to capture het­
used neural network and fuzzy logic modeling for estimation of porosity erogeneity in oil and gas reservoirs.
and permeability in a reservoir, revealing that these methods have po­
tential for future use and implementation. Zhu et al. (2017) have suc­ 2. Rock typing: FZI method
cessfully applied machine learning methods for permeability prediction
and achieved acceptable results. Khalifah et al. (2019) compared ANN Rock typing is a process in which reservoir rocks are classified into
with seven conventional empirical equations for prediction of perme­ distinct units. One of the frequently used approaches for rock typing is
ability in a tight carbonate reservoir. They also used the genetic algo­ the FZI method. Amaefule et al. (1993) introduced this method for the
rithm to develop a new empirical equation for prediction of permeability first time. This technique identifies existing trends between permeability
based on their dataset. Their results showed that the machine learning and porosity using the FZI values of reservoir rocks. The general
methods provide better predictions in comparison with empirical Kozeny-Carman relation for calculating permeability is given by Eq. (1).
equations. Matinkia et al. (2022a, 2022b) predicted permeability from [ ]
well logs using a new hybrid machine learning algorithm. Their research ∅3e 1
K = 1014 (1)
combined Social Ski-Driver (SSD) algorithm with the multilayer (1 − ∅e )2 Fs τ2 Sgv
2

perception (MLP) neural network and presented a new hybrid algorithm


to predict the rock permeability. Their results indicated that the hybrid where, K is core permeability in mD, Øe is effective core porosity, Fs
models can deliver satisfactory results. denotes shape factor, τ is tortuosity and Sgv represents the surface area
Models developed based on hydraulic flow units (HFU) provide a per grain volume. Calculation of permeability by this equation is not an
better permeability prediction compared to other ones. Flow Zone In­ easy task because it is difficult to measure the Fs, τ and Sgv parameters
dicator (FZI) which is calculated according to the rock porosity and for reservoir rocks. Amaefule et al. (1993) defined the FZI as the square
permeability is strongly related to HFU. Rock samples, which belong to root of the 1/Fsτ2S2gv term and derived Eq. (2) to obtain FZI from core
the same HFU, have similar FZI ranges. While HFU is a parameter used data:
for classification, FZI is a quantitative parameter, which makes it suit­
RQI
able for developing prediction models. In the past, different researchers FZI = (2)
∅n
(Baker et al., 2013; Amraei and Falahat, 2021; Abbaszadeh et al., 1996)
developed models to predict FZI, and then applied the predicted FZI to √̅̅̅̅̅̅
K
calculate permeability. Soto et al. (2001) implemented a fuzzy model for RQI = 0.0314 (3)
Øe
estimation of permeability by grouping the data into different HFUs
using the FZI values, and managed to predict permeability accurately. A Øe
neural network model was proposed by Aminian et al. (2003) to identify Øn = (4)
1 − Øe
various flow units and estimate the permeability of a reservoir. Kharrat
et al. (2009) developed an ANN model for prediction of permeability in where, RQI is a parameter called reservoir quality index (μm) and Øn is
different rock types of a heterogeneous carbonate reservoir. Ali et al. the normalized porosity. The calculated FZI is then used to group the
(2013) utilized the support vector machine (SVM) and proposed an rocks into discrete rock types (DRTs) according to the following relation:
approach to predict the FZI values in a carbonate reservoir by using well
log data as input parameters. Although there are different research DRT = Round(2 Log(FZI) + 10.6) (5)
works in the literature on application of regression and machine The X field is a natural gas field, which covers an area of 3700 square
learning approaches for prediction of permeability, but few papers are kilometers. This tight gas field is 3000 m below the seabed at a water

2
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Fig. 1. Histogram for core porosity across the field.


Fig. 4. Representative relations between core permeability and porosity for
each DRT group.

depth of 65 m and it is mainly composed of limestone and dolomite. The


field consists of four main gas bearing reservoirs, namely L1, L2, L3 and
L4 with similar rock properties. These four reservoirs are isolated by
impermeable barriers but based on the observed pressure analysis, they
are communicating through faults. Gas production from the X field
commenced in 2002.
The core data of eighteen wells were used in this study for rock
typing based on the FZI approach. The statistical parameters of the core
porosity and permeability are summarized in Table 1.
From the histogram for porosity shown in Fig. 1, a unimodal porosity
distribution is observed with a minimum and maximum porosity values
of 0.00005 and 0.423 (m3/m3) respectively. Fig. 2 shows a histogram of
the permeability across the field with the data range of 0.0003–1849.64
(mD). The wide range of permeability distribution on this histogram
Fig. 2. Histogram for core permeability across the field.
indicates that the understudy gas field is very heterogeneous and for an
accurate permeability prediction, clustering the core data into appro­
priate DRT groups is essential.
By applying the FZI rock typing approach ten different rock types
were identified in the X gas field. The plot of reservoir quality index
(RQI) against normalized porosity (Øn) on a log-log scale results in a unit
slope straight line for each rock type. The intercept of the straight line at
Øn = 1 gives the average FZI value for the specific rock type. Based on
the reservoir quality index vs. normalized porosity plot shown in Fig. 3,
ten different DRT groups (rock types) can be identified in this gas field
study.
For each rock type, a representative curve describing the correlation
between core permeability and porosity was derived by replacing the 1/
Fsτ2S2gv term with FZI2avg in Eq. (1) as follows:

∅3e
K = 1014 FZI 2avg (6)
(1 − ∅e )2
Fig. 3. A log-log plot of reservoir quality index vs. normalized porosity, which
classifies 10 discrete rock types in the X gas field. where, FZIavg is the average FZI for each DRT group given in Table 2. The
plot of representative correlations (log K vs. porosity) for each DRT
group is shown in Fig. 4.
Table 2 The FZI range, average FZI, average porosity and average perme­
Average values of FZI, porosity (fraction) and permeability (md) for each rock
ability of each rock type (DRT group) are summarized in Table 2. This
type.
table shows that, there are notable differences in average permeability
DRT group FZI range avg. FZI avg. Porosity avg. Permeability and porosity of various rock types, which indicate the differences in
1 0.01–0.05 0.04 0.25 0.05 their fluid capacity. In other words, good rock types with higher FZI
2 0.05–0.158 0.11 0.19 0.28 values have better reservoir quality and percolation capacity compared
3 0.158–0.50 0.32 0.14 1.04
to those with lower FZI values (poor rock types).
4 0.50–1.58 0.95 0.11 5.40
5 1.58–5.01 2.87 0.08 30.98
6 5.01–15.82 8.79 0.04 70.60 3. The regression approach: polynomial regression model
7 15.82–50 27.09 0.02 45.16
8 50–153 83.23 0.01 44.00 In this study we are aiming to correlate FZI from core data with wire-
9 153–4465 686.61 0.003 17.38
line logging data by means of multiple linear regression models to pre­
10 4465–118,753 46,237.80 0.0002 10.91
dict the permeability-logs in the logged intervals of the wells.

3
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Fig. 5. The density functions of different logs for all rock types. (a) Dolomite volume log, (b) Neutron porosity log, (c) Shale volume log, (d) Effective porosity log, (e)
Bulk density log, (f) Effective water saturation log, (g) Calcite volume log and (h) Anhydrate volume log.

4
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Regression analysis is utilized for estimating the relationship be­ and responses of dolomite volume log in the X field and this log can be a
tween a response (dependent variable) and regressors (independent good indicator of rock typing. The density functions of dolomite volume
variables). It can be used for modeling the future relationship between log are shown in Fig. 5 (a). The density functions of all rock types show 4
variables. clusters. The first cluster, which has the lowest volume of dolomite in­
The relationship between a dependent variable and an independent cludes DRT1 and DRT2, which are poor quality rock types. DRT3 and
variable can be assessed by a simple linear regression model. It can be DRT10 are falling into the second cluster, while DRT4 and DRT9 are
expressed by the following equation (Eq. (7)): forming cluster three. The fourth cluster, which has the highest dolomite
volume, includes good quality rock types, namely DRT5 through 8. The
y = a + bx + ε (7)
wide variance intervals of all clusters are an indication of heterogeneous
where, y is dependent variable (response), a is intercept, b is the slope, x formations. The dolomite volume log was considered as a variable (re­
is independent variable (regressor) and ε is error. gressor) in the polynomial regression model.
Multiple linear regression model is similar to the simple linear Neutron tool is a nuclear device which is very sensitive to the amount
model, with the exception that it has multiple independent variables. of hydrogen atoms in a formation and less sensitive to other elements.
The response variable can also be approximated by a polynomial Neutron porosity log (NPHI) measures the hydrogen concentration in
regression model such as in Eq. (8). It is just another variant of the the formation and its output should be calibrated to read the true
general linear regression method. porosity assuming that the pores are saturated with fresh water and for a
given matrix like limestone, sandstone or dolomite. The responses of the
y = b0 + b1 x1 + b2 x2 + b3 x1 x2 + b4 x21 + b5 x22 + error (8) neutron porosity log depend on the porosity, the type of the fluid in the
pores and the composition of the minerals. Determining the porosity of
where, y is dependent variable (response), b0 is intercept, b’s are the formation is the main usage of the neutron porosity log. There is no
regression coefficients and x’s are independent variables (regressors). relationship between responses of neutron porosity log and FZI because
neutron porosity log can give two different responses to the formation
3.1. Lithology and sensitivity analysis intervals with the same FZI values. An example of this contradictory
relationship between neutron porosity log and FZI is the low response of
The giant X gas field is composed of Anhydrate, Calcite, Dolomite the neutron porosity log in the gas cap of a permeable and porous for­
and clay minerals with a significant vertical heterogeneity in terms of mation where FZI is high, while in the oil-bearing interval with the same
lithology, porosity and permeability. This field is under development FZI the response of neutron porosity log is high. The density functions of
with limited log data available. At first, we selected eight conventional neutron porosity logs for all rock types are shown in Fig. 5 (b). As can be
open hole well logs from eighteen wells to develop the polynomial seen an overlapping is almost obtained and all density functions have
regression model. These include effective porosity log (PHIE), effective the same mean value. The density functions of neutron porosity log are
water saturation log (SWE), bulk density log (RHOB), neutron porosity forming just one cluster and they are very similar and with no distinct
log (NPHI), calcite volume log (VCALC), dolomite volume log (VDOL), differences for all rock types. Hence, we can conclude that this log is not
shale volume log (WCS) and anhydrate volume log (VANH). It should be a good regressor and can therefore be discarded.
highlighted that the effective porosity log is obtained by combining the The same results were obtained for the shale volume log (WCS)
neutron porosity, bulk density and sonic logs. Effective water saturation which can be seen in Fig. 5 (c). For the same reasons mentioned for the
log is obtained from resistivity log, while the calcite volume, dolomite neutron porosity log, the shale volume log was also discarded and was
volume, shale volume and anhydrate volume logs are obtained from not considered as regressor in the polynomial regression model. Shale
bulk density and photo electric factor logs. The bulk density and neutron volume log represents the volume of the shale. Shale is a fine-grained
porosity are raw logs. sedimentary rock that forms from the compaction of silt and clay-size
The initial polynomial regression model is given by Eq. (9). mineral particles. Shale reduces the porosity and permeability of a
∑8 ∑7 ∑8 ∑8 rock. Therefore, the high shale volume indicates the poor rock quality
FZIcore = λ0 + i=1
λi Xi + i=1 j=i+1
λij Xi Xj + i=1
λi Xi2 (9) and low FZI. It should be highlighted that the field being studied has a
very low volume of shale.
where, FZIcore is the core derived FZI, λ0 is intercept, λ′ s are regression Effective porosity log (PHIE) measures the total porosity by
coefficients, X1 is PHIE, X2 is SWE, X3 is RHOB, X4 is NPHI, X5 is VCALC, excluding shale water. This log can be a good indicator of rock typing.
X6 is VDOL, X7 is WCS and X8 is VANH. The density functions of the effective porosity log are shown in Fig. 5 (d).
In this study the probability density function (PDF) was used in order The density functions of all DRT groups show three distinct clusters. The
to investigate the relationships between the selected log responses and first cluster includes DRT6 through 10 with the same mean values. DRT4
DRT groups. The probability density functions of the log responses for and DRT5 fall into cluster two and finally the third cluster includes
each DRT group are plotted and compared. If they are overlapping and DRT1, DRT2 and DRT3. The first cluster has the lowest mean value for
forming just one cluster with almost the same mean value for all rock porosity. This cluster represents the good rock types, which are partially
types, it can be concluded that the log is not a good regressor and can fractured. The third cluster has the highest mean value and it has a wide
therefore be discarded. On the other hand, if distinct clusters can be range of porosity (variance interval) which is an indication of hetero­
distinguished, it indicates that the log is a good regression variable and geneous formation. This log was considered as a dependent variable in
can be used in multiple regression model. the regression model.
Dolomite is formed when magnesium ions replace calcium ions in The bulk density log (RHOB) measures the bulk density of a forma­
calcite. This geological process is called dolomitization. During this tion. The formations are bombarded with gamma rays, which are
process the volume reduction occurs, which results in larger pore spaces. emitted from a radioactive source. The radiation response is used for
Therefore, dolomitization modifies the rock quality of a formation. The determining the bulk density. The main use of the bulk density log is to
dolomite volume log (VDOL) represents the volume of the dolomite of a determine the total porosity of a formation, and it is also used to detect
formation. The relationship of the dolomite volume log with FZI can be the gas bearing formations. There is a contradicting relationship be­
very contradicting. The FZI is low when the large pore spaces created tween the response of density log and FZI. For example, the response of
during dolomitization are filled with cement or closed by tectonic forces, the density log is low for a permeable and porous sandstone where FZI is
and it is high when the pore spaces are unfilled and open, but for both high, while this log would also give low response for the gas bearing
cases the dolomite volume log gives high responses. Based on seismic intervals with very low FZI. The density functions of the density log for
stratigraphy studies, there is no contradictory relationship between FZI all DRT groups are shown in Fig. 5 (e). Different clusters can be

5
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Table 3 distinguished by the histograms of the density log responses. The density
The coefficients of the polynomial regression model for FZI prediction. functions can be grouped into 3 clusters. In the first cluster DRT6
Term Coefficients Value through 10 can be grouped as high-density rock types. DRT4 and DRT5
are falling into the second cluster, while DRT1 through 3 are forming the
Intercept λ0 − 3889.9
PHIE (X1) λ1 4579.6 third cluster with a low bulk density and wide variance intervals, which
SWE (X2) λ2 − 382.93 is an indication of a heterogeneous formation. This log was considered
RHOB (X3) λ3 2910.3
VCALC (X4) λ4 320.4
VDOL (X5) λ5 − 790.2
PHIE × SWE λ6 197.51
PHIE × RHOB λ7 − 2136.5
PHIE × VCALC λ8 433.12
PHIE × VDOL λ9 1455.9
SWE × RHOB λ10 − 2.4663
SWE × VCALC λ11 286.65
SWE × VDOL λ12 403.05
RHOB × VCALC λ13 − 326.85
RHOB × VDOL λ14 − 63.867
VCALC × VDOL λ15 754.85
PHIE2 λ16 − 47.427
SWE2 λ17 137.65
RHOB2 λ18 − 489.51
VCALC2 λ19 239.55
VDOL2 λ20 604.28

Fig. 7. A cross plot of the predicted K by regression model against core K on a


log-log scale.

Fig. 6. A comparison between the predicted K by regression model and core K for test wells. (a) test well#5, (b) test well#7, (c) test well#8, (d) test well#9.

6
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Table 4
The regression coefficients of the DRT-based regression models for DRT groups 1–5.
Term Coefficients DRT1 DRT2 DRT3 DRT4 DRT5

Intercept λ0 − 10793 − 2.9371 0.65401 3.2659 − 13.111


PHIE (X1) λ1 21,704 1.9082 − 2.0131 1.896 − 12.025
SWE (X2) λ2 − 14.992 0.21536 0.69633 0.49196 − 1.7747
RHOB (X3) λ3 − 8.0824 2.1639 − 0.74535 − 2.9663 12.308
VCALC (X4) λ4 21,604 1.0418 0.3971 2.3312 2.124
VDOL (X5) λ5 21,606 − 0.43961 2.2333 4.2531 3.9929
PHIE × SWE λ6 14.971 − 0.4271 − 0.74052 − 0.09514 − 0.22708
PHIE × RHOB λ7 7.4044 − 1.0724 0.91074 1.1357 − 2.2261
PHIE × VCALC λ8 − 21719 0.56706 0.5719 − 4.7873 10.082
PHIE × VDOL λ9 − 21720 1.8006 − 0.62486 − 5.2118 10.935
SWE × RHOB λ10 − 0.16629 − 0.08404 − 0.22658 − 0.21744 0.094058
SWE × VCALC λ11 15.539 0.094237 − 0.03095 − 0.13506 1.3404
SWE × VDOL λ12 15.488 0.11486 0.002774 − 0.08858 0.60197
RHOB × VCALC λ13 9.1607 − 0.47104 0.033854 − 0.70382 − 1.6617
RHOB × VDOL λ14 9.2022 − 0.04535 − 0.45619 − 1.6732 − 2.0596
VCALC × VDOL λ15 − 21626 0.35301 − 1.0043 0.46414 1.0322
PHIE2 λ16 − 10910 0.2502 1.0424 − 3.614 29.796
SWE2 λ17 − 0.06215 − 0.05081 0.003987 0.12236 0.3449
RHOB2 λ18 − 0.17021 − 0.35308 0.18873 0.75673 − 2.1443
VCALC2 λ19 − 10812 − 0.02652 − 0.21349 − 0.21244 1.0313
VDOL2 λ20 − 10814 0.26382 − 0.72149 0.61847 0.83259

Table 5
The regression coefficients of the DRT-based regression models for DRT 6–10.
Term Coefficients DRT6 DRT7 DRT8 DRT9 DRT10

Intercept λ0 − 23.477 − 22.82 4233.4 − 95865 − 91,480,000


PHIE (X1) λ1 108.66 304.05 − 7587.8 117,000 137,280,000
SWE (X2) λ2 1.6332 27.946 287.8 − 13825 − 39,478,000
RHOB (X3) λ3 22.942 39.891 − 2929.8 70,379 95,740,000
VCALC (X4) λ4 5.9112 − 5.0134 − 1007.3 10,978 − 31,034,000
VDOL (X5) λ5 − 8.045 2.8917 − 530.39 − 1020.3 − 53,460,000
PHIE × SWE λ6 23.688 − 56.069 − 160.22 4831.8 − 68,357,000
PHIE × RHOB λ7 − 48.792 − 158.91 2317.2 − 36743 128,500,000
PHIE × VCALC λ8 5.9138 137.44 1446.9 − 26861 − 156,640,000
PHIE × VDOL λ9 4.6476 180.38 1228 − 16317 178,420,000
SWE × RHOB λ10 − 1.0007 − 7.6807 − 74.794 3733.8 181,710,000
SWE × VCALC λ11 − 0.02198 − 6.5685 − 71.803 3258.4 196,280,000
SWE × VDOL λ12 − 1.5071 2.7925 10.303 3061.1 8,354,600
RHOB × VCALC λ13 − 1.615 4.5268 386.71 − 4460.2 12,409,000
RHOB × VDOL λ14 2.0261 0.64715 170.04 − 915.02 10,750,000
VCALC × VDOL λ15 2.0147 − 27.7 99.729 772.09 11,010,000
PHIE2 λ16 − 17.036 − 374.21 4873.1 − 31790 5,741,100
SWE2 λ17 1.3605 − 3.5179 − 35.428 281.65 15,467,000
RHOB2 λ18 − 3.9729 − 7.5642 508.29 − 12562 21,908,000
VCALC2 λ19 − 2.2628 − 7.0159 30.703 − 893.33 1,167,000
VDOL2 λ20 3.4545 − 10.276 76.329 1179.7 − 601470

Fig. 8. A comparison between the predicted K by regression model and core K Fig. 9. A comparison between the predicted K by regression model and core K
for DRT1. for DRT2.

7
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

effective water saturation log measures the effective pore spaces which
are occupied with water. This log can be used to determine the irre­
ducible water saturation, which can be a good indicator of rock typing.
The calcite volume log represents the volume of the calcite with the
chemical formula CaCO3. This mineral is the major component of car­
bonate reservoirs.
The density functions of anhydrate volume log (VANH) are shown in

Fig. 10. A cross plot of the predicted K by the DRT-based regression models
against core K.

as a variable in the regression model.


The density functions of effective water saturation (SWE) and calcite
volume (VCALC) logs (Fig. 5 (f) and (g)) are very similar to those of
dolomite volume log. The probability density functions of these logs are
not overlapping and distinct clusters can be distinguished. Therefore,
effective water saturation and calcite volume logs were also considered
as regressors and used to develop the polynomial regression model. The Fig. 12. A cross plot of the predicted K by ANN against core K across the field.

Fig. 11. A comparison between the predicted K by ANN model and core K for test wells. (a) test well#5, (b) test well#7, (c) test well#8 and (d) test well#9.

8
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Fig. 13. A comparison between the predicted K by ANN model and core K
for DRT1.

Fig. 14. A comparison between the predicted K by ANN model and core K
for DRT2.

Fig. 16. The flowchart of the novel ANN approach.

3.2. The final regression model

Based on the results of probability density function analysis that we


conducted for each log type in Section 3.1, five logs were selected to
develop the polynomial regression model, namely effective porosity
(PHIE), effective water saturation (SWE), bulk density (RHOB), calcite
volume (VCALC) and dolomite volume (VDOL) logs. Three logs were
excluded from the initial model (Eq. (9)), namely neutron porosity
(NPHI), shale volume (WCS) and anhydrate volume (VANH) logs. The
final polynomial regression model is given by Eq. (10).
Fig. 15. A cross plot of the predicted K by the DRT-based ANN models against ∑5 ∑
4 ∑
5 ∑5
core K. FZIcore = λ0 + i=1
λi Xi + λij Xi Xj + i=1
λi Xi2 (10)
i=1 j=i+1

Fig. 5 (h). As can be seen in this Figure the histograms of anhydrate


where, FZIcore is the core derived FZI, λ0 is intercept, λ′ s are regression
volume log responses are almost overlapping and no distinct clusters can
coefficients, X1 is PHIE, X2 is SWE, X3 is RHOB, X4 is VCALC and X5 is
be distinguished, consequently this log was excluded from the regression
VDOL. It should be highlighted that the mathematical regression models
model. Anhydrite is an evaporite mineral with the chemical formula
were created and solved using MATLAB 8.4 (R2014b) programming
CaSO4. It forms where large volumes of sea water have been evaporated
language.
and salt deposition has occurred. This mineral has a very low neutron
The polynomial regression model given by Eq. (10) is correlating the
porosity and gamma ray responses. There is a strong relationship be­
core derived FZI with the log responses to predict the FZI values for the
tween anhydrate volume log and FZI. The presence of anhydrite de­
uncored but logged intervals of the wells. The predicted FZI-logs are
creases the rock quality and FZI.
then converted to K-logs using Eq. (2).
For regression analysis the core and log data of eighteen wells were

9
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

calculated regression coefficients are summarized in Tables 4 and 5.


The predicted permeability for DRT1 and DRT2 are plotted and
compared to core permeability in Figs. 8 and 9. The results for DRT
groups 3 to 10 are illustrated in Figs. S1 to S8, Supplementary Material.
It can be observed that good matches have been achieved for all DRT
groups.
Fig. 10 shows a cross plot of the predicted permeability by the DRT-
based regression models against core permeability on a log-log scale for
all DRT groups, which gives a correlation coefficient of 0.7218.
Based on the achieved results, it can be concluded that the DRT-
based regression models can accurately predict Permeability. Howev­
er, the developed regression model for the entire field (all DRT groups)
did not achieve a good match and is unable to capture heterogeneity in
the X gas condensate field.

4. Artificial neural networks approach

The second approach in this study used artificial neural networks


(ANN) to predict K-logs. They were used to better correlate the core
derived FZI and well logging data. ANN are one of the main tools used in
machine learning. Artificial intelligence methods have been applied
widely in the petroleum industry. Artificial intelligence methods to es­
timate different reservoir properties as referenced by Anemangely et al.
(2019), Rashidi et al. (2021) and Matinkia et al. (2022a, 2022b). Their
results show that these techniques are particularly useful and provide
acceptable reservoir property predictions.
In this study, one-hidden layer feed forward ANN model was trained
with 17 hidden neurons. The number of hidden layers and neurons were
determined based on a trial and error approach.
The mathematical models were created using MATLAB 8.4 (R2014b)
programming language. The error back propagation algorithm was
employed for adjusting the weights and biases of the network to mini­
mize the error for the validation data set. The mean square error was
Fig. 17. The predicted K-log by the trained ANN and core K-log for the single considered for evaluating the performance of the network. The
key well. Levenberg-Marquardt numerical optimization technique was used for
training the network, which operated in batch mode. Batch training of
used, out of which fourteen were considered as key wells (wells# 1–4, 6, the network proceeds by making weight and bias changes based on an
10–18) and four, for testing or used as blind wells (wells# 5, 7, 8 and 9). entire set (batch) of input vectors.
The key wells were used for modeling purposes and the test wells for In order to keep the model from being over-trained, the error stop­
checking the accuracy of the developed regression models. It should be ping method was employed. The maximum number of epochs for
highlighted that all eighteen wells were logged and cored as well. At training the network was determined to be 1000. The linear and Tan-
first, the outcome variable, namely the core derived FZI (FZIcore) was Sigmoid transfer functions were considered for the output and hidden
regressed on PHIE, SWE, RHOB, VCALC and VDOL logs to find the layers of the network, respectively.
weight factors or regression coefficients. The results are summarized in Based on the statistical analysis conducted in section 3.1, five logs
Table 3. were considered as input for the ANN model. They include effective
Having determined the regression coefficients, the regression model porosity, effective water saturation, bulk density, calcite volume and
was quality controlled by predicting the FZI-logs in the cored intervals of dolomite volume logs and the core derived FZI was considered as the
the test wells. The predicted FZI-logs were then converted to K-logs, response of the model. The core and log data of fourteen key wells,
which are plotted and compared to core permeability in Fig. 6 (a) to (d). namely wells# 1–4, 6, 10–18 were used for modeling purposes and four
It can be observed that no good match is achieved and the regression wells, namely wells# 5, 7, 8 and 9 were considered as test wells.
model is unable to match most points of very low and high Having trained the ANN model, it was used to predict the FZI-logs in
permeabilities. the cored intervals of the test wells to quality control the trained model.
A cross plot of the predicted permeability by regression model The predicted FZI-logs were then converted to K-logs, which are shown
against core permeability on a log-log scale is shown in Fig. 7, which in Fig. 11 (a) to (d). It can be observed that better results have been
gives a correlation coefficient of 0.054. achieved by ANN model, but like regression model, it is also unable to
Different polynomial regression models of different orders with capture heterogeneity in the field.
different combinations of logs were selected and new regressions were Fig. 12 shows a cross plot of the predicted K by ANN model against
conducted but no improvement was observed. Furthermore, four core K on a log-log scale across the field, which gives a correlation factor
different regression models for L1-L4 reservoirs were developed but no of 0.07.
significant changes were observed and all developed regression models In order to investigate the effects of heterogeneity on the ANN per­
were unable to capture heterogeneity in the X field. formance, ten DRT-based ANN models were developed (an ANN model
In order to investigate the effects of heterogeneity on the results of for each DRT group). The predicted permeability for DRT1 and DRT2 are
the polynomial regression model, the core derived FZI and the related shown and compared to core permeability in Figs. 13 and 14. The results
log data for each DRT group were regressed and ten new DRT-based for DRT groups 3 to 10 are shown in Figs. S9 to S16, Supplementary
regression models (one for each DRT group) were developed. The Material. It can be observed that good matches have been achieved for
all DRT groups.

10
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Fig. 18. A comparison between the predicted K by the new approach and core K for blind test wells. (a) blind test well#2, (b) blind test well#3, (c) blind test
well#14 and (d) blind test well#18.

Fig. 15 shows a cross plot of the predicted permeability by DRT- and porosity data are categorized using the concept of discrete rock
based ANN models against core permeability on a log-log scale for all types (DRT). The ANN method is then applied to correlate the core
DRT groups across the field, which gives a correlation coefficient of derived flow zone indicator (FZI) with wire-line logging data to predict
0.826. permeability.
The achieved results by the developed ANN models in this study Contrary to the current ANN models, the novel approach presented
indicated that the DRT-based ANN models predicted Permeability in this paper applies to a single well, which contains all DRT groups
accurately. However, applying the ANN model for the entire field (all (rock types) to develop and train the ANN model. It was observed that
DRT groups) did not deliver satisfactory results, since it was unable to ANN model exhibits better prediction performance in heterogeneous
capture heterogeneity. reservoirs when it is trained on the data of a single key well containing
all DRT groups. This data reduction strategy, which is the novelty of this
5. Artificial neural networks, a novel approach approach enables the ANN model to capture heterogeneity in oil and gas
reservoirs.
ANN models have been used by different researchers to date to The flowchart of the novel approach, which is presented in this paper
predict permeability. The existing ANN models provide good results in is shown in Fig. 16.
homogeneous sandstone reservoirs but their accuracy diminishes in Well#7 was selected as the key well to develop and train the ANN
heterogeneous reservoirs, such as carbonates. model, since it contained all DRT groups. The predicted FZI-log by the
In the fields, where permeability has a wide range of distribution and trained ANN model was converted to K-log, which is shown and
multiple DRT groups (rock types) are available, ANN models are unable compared to core K-log in the following Fig. 17.
to capture heterogeneity. That’s why clustering the core data into In order to check the prediction performance and accuracy of the
appropriate DRT groups can yield better results. developed ANN model, it was used to predict K-logs in the cored in­
In this study, we are presenting a novel approach to predict perme­ tervals of the blind-test wells, and was observed that good matches were
ability in heterogeneous reservoirs. In this method the core permeability achieved for all these wells. The results for the wells# 2, 3, 14 and 18 are

11
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Fig. 19. The predicted K-log by the new approach for prediction well#13.

12
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Fig. 20. The predicted K-log by the new approach for prediction well#14.

13
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Fig. 21. The predicted K-log by the new approach for prediction well#15.

14
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Fig. 22. The predicted K-log by the new approach for prediction well#16.

15
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

Fig. 23. 3D distribution of permeability on top of X gas field.

training instances in the total data points. In other words, when all data
points are used, the ANN model provides better predictions for larger
permeability values, but it fails to provide acceptable prediction results
for intermediate and low permeability values. This means that the ANN
can’t select appropriate training instances of the input and output data
for intermediate and low permeability values when it is trained on all
data points, instead, it prefers to learn existing patterns within the
training instances of high permeability data and tends to be trained on
these values. However, reducing the training points and training the
ANN model on a single well which contains all DRT groups, enables the
model to better identify and learn the existing patterns between input
and output data for the whole range of permeability variation. This al­
lows the ANN model to utilize a smaller dataset that represents the entire
training data of all wells, which increases the accuracy of the predictions
of the ANN model trained on the data of a single key well compared to
the model, which is trained on the entire data set.
Finally, the developed ANN model by the new approach was used to
predict the K-logs in the uncored but logged intervals of the wells. The
predicted K-logs for the prediction wells#13, 14, 15 and 16 are shown in
Figs. 19–22.
Having predicted the K-logs, they were populated in the geological
model. Appropriate variogram models were set and Sequential Gaussian
Simulation method (SGS) was used for this purpose. The generated 3D
permeability model is shown in Fig. 23, and an East-West cross section of
the permeability is shown in Fig. 24.

7. Conclusions
Fig. 24. An E-W cross section of permeability.
Empirical equations, regression models and machine learning
shown in Fig. 18 (a) to (d). methods have been used by different researchers to date, to predict
permeability in oil and gas reservoirs. Clustering the core data into
6. Results and discussion appropriate DRT groups, enables the models to provide more accurate
permeability predictions. However, it was observed that their accuracy
Based on the achieved results by the proposed approach, it can be diminished in heterogeneous reservoirs.
concluded that the ANN model exhibits better prediction performance In this study the polynomial regression models and artificial neural
when it is developed and trained on the core and log data from a single networks were used to correlate the core derived FZI with wire-line
well containing all DRT groups rather than with all data points. The core logging data to predict K-logs. It was observed that both models were
permeability data, which are used in the present work, are highly het­ unable to predict permeability satisfactorily in heterogeneous reser­
erogeneous, noisy and exhibit considerable variation range versus voirs, where there was a wide permeability distribution.
depth. In some cases, the core permeability data in two adjacent depths A novel approach is presented in this paper to predict permeability in
of a well differ by a fourth order of magnitude. When the data points of a heterogeneous oil and gas reservoirs. This methodology is applying ANN
well with all DRT groups are considered to train the ANN model, it was models to correlate the core derived FZI with well logs to predict
found to better capture heterogeneity. This data reduction strategy en­ permeability in the uncored but logged intervals of the wells. This
ables the ANN model to select instances of input and output data for all approach is applicable in reservoirs, which are subdivided into different
DRT groups from a single key well, which are representatives of similar hydraulic flow units or discrete rock types. In this approach, contrary to

16
N. Alizadeh et al. Journal of Petroleum Science and Engineering 214 (2022) 110573

the existing ANN models, one single well, which contains all DRT groups and Predict Permeability in Uncored Intervals/wells, SPE Annual Technical
Conference and Exhibition. Society of Petroleum Engineers, Houston, Texas.
is considered as a key well to develop the ANN model. It is demonstrated
Aminian, K., Ameri, S., Oyerokun, A., Thomas, B., 2003. Prediction of Flow Units and
that ANN model exhibits better prediction performance in heteroge­ Permeability Using Artificial Neural Networks, SPE Western Regional/AAPG Pacific
neous reservoirs when it is developed and trained on the data of a single Section Joint Meeting. Society of Petroleum Engineers, Long Beach, California.
well containing all DRT groups. This data reduction strategy, which is Amraei, H., Falahat, R., 2021. Improved ST-FZI method for permeability estimation to
include the impact of porosity type and lithology. J. Petrol. Explor. Prod. 11 (1),
the novelty of this approach enables the ANN model to capture het­ 109–115.
erogeneity in heterogeneous oil and gas reservoirs. This approach was Anemangely, M., Ramezanzadeh, A., Amiri, H., Hoseinpour, S.A., 2019. Machine
successfully applied in an actual carbonate gas reservoir to predict learning technique for the prediction of shear wave velocity using petrophysical logs.
J. Petrol. Sci. Eng. 174, 306–327.
permeability logs. The permeability logs were then populated in the Baker, H.A., Al-Jawad, S.N., Murtadha, Z.I., 2013. Permeability prediction in carbonate
geological model by applying geostatistical methods to obtain the 3D reservoir rock using FZI. Iraqi J. Chem. Petrol. Eng. 14 (3), 49–54.
permeability distribution. Balan, B., Mohaghegh, S., Ameri, S., 1995. State-of-the-art in permeability determination
from well log data: Part 1-A comparative study, model development. In: In SPE
Eastern Regional Meeting. OnePetro.
Declaration of competing interest Finol, J., Guo, Y.K., Jing, X.D., 2001. A rule based fuzzy model for the prediction of
petrophysical rock parameters. J. Petrol. Sci. Eng. 29 (2), 97–113.
Helle, H.B., Bhatt, A., Ursin, B., 2001. Porosity and permeability prediction from wireline
The authors declare that they have no known competing financial logs using artificial neural networks: a North Sea case study. Geophys. Prospect. 49
interests or personal relationships that could have appeared to influence (4), 431–444.
the work reported in this paper. Kharrat, R., Mahdavi, R., Bagherpour, M.H., Hejri, S., 2009. Rock Type and Permeability
Prediction of a Heterogeneous Carbonate Reservoir Using Artificial Neural Networks
Based on Flow Zone Index Approach, SPE Middle East Oil and Gas Show and
Nomenclature Conference. Society of Petroleum Engineers, Manama, Bahrain.
Lacentre, P.E., Carrica, P.M., 2003. A Method to Estimate Permeability on Uncored Wells
HFU hydraulic flow unit Based on Well Logs and Core Data, SPE Latin American and Caribbean Petroleum
Engineering Conference. Society of Petroleum Engineers, Port-of-Spain, Trinidad and
DRT discrete rock type Tobago.
ANN artificial neural networks Lim, J.-S., Kim, J., 2004. Reservoir porosity and permeability estimation from well logs
FZI flow zone indictor using fuzzy logic and neural networks. In: SPE Asia Pacific Oil and Gas Conference
and Exhibition. Society of Petroleum Engineers, Perth, Australia.
LSSVM least square support vector machine Matinkia, M., Amraeiniya, A., Behboud, M.M., Mehrad, M., Bajolvand, M.,
SVM support vector machine Gandomgoun, M.H., Gandomgoun, M., 2022a. A novel approach to pore pressure
RQI reservoir quality index modeling based on conventional well logs using convolutional neural network.
J. Petrol. Sci. Eng. 110156.
PHIE effective porosity log Matinkia, M., Hashami, R., Mehrdad, M., Hajsaeedi, M., Velayati, A., 2022b. Prediction
SWE effective water saturation log of permeability from well logs using a new hybrid machine learning algorithm.
RHOB bulk density log Petroleum. https://doi.org/10.1016/j.petlm.2022.03.003.
Rashid, F., Glover, P.W.J., Lorinczi, P., Hussein, D., Collier, R., Lawrence, J., 2015.
NPHI neutron porosity log Permeability prediction in tight carbonate rocks using capillary pressure
VCALC calcite volume log measurements. Mar. Petrol. Geol. 68, 536–550.
VDOL dolomite volume log Rashidi, S., Mehrad, M., Ghorbani, H., Wood, D.A., Mohamadian, N., Moghadasi, J.,
Davoodi, S., 2021. Determination of bubble point pressure & oil formation volume
WCS shale volume log
factor of crude oils applying multiple hidden layers extreme learning machine
VANH anhydrate volume log algorithms. J. Petrol. Sci. Eng. 202, 108425.
PDF probability density function Rose, W., Bruce, W., 1949. Evaluation of capillary character in petroleum reservoir rock.
K permeability, mD J. Petrol. Technol. 1 (5), 127–142.
Soto, B.R., Garcia, J., Torres, F., Perez, G., 2001. Permeability prediction using hydraulic
Ø porosity, fraction flow units and hybrid soft computing systems. In: SPE Annual Technical Conference
Øe effective porosity, fraction and Exhibition. Louisiana, New Orleans. OnePetro.
Øn normalized porosity, dimensionless Vafaie, A., Kivi, I.R., Moallemi, S.A., Habibnia, B., 2021. Permeability prediction in tight
gas reservoirs based on pore structure characteristics: a case study from South
Western Iran. Unconvent. Resour. 1, 9–17.
Appendix A. Supplementary data Wendt, W., Sakurai, S.t., Nelson, P., 1986. Permeability Prediction from Well Logs Using
Multiple Regression, Reservoir Characterization. Elsevier, pp. 181–221.
Wyllie, M., Rose, W.D., 1950. Some theoretical considerations related to the quantitative
Supplementary data to this article can be found online at https://doi. evaluation of the physical characteristics of reservoir rock from electrical log data.
org/10.1016/j.petrol.2022.110573. J. Petrol. Technol. 2 (4), 105–118.
Xue, G., Datta-Gupta, A., Valko, P., Blasingame, T., 1997. Optimal transformations for
multiple regression: application to permeability estimation from well logs. SPE
References Form. Eval. 12 (2), 85–94.
Yao, C., Holditch, S., 1993. Estimating permeability profiles using core and log data. In:
Abbaszadeh, M., Fujii, H., Fujimoto, F., 1996. Permeability prediction by hydraulic flow SPE Eastern Regional Meeting. Society of Petroleum Engineers, Pittsburgh,
units-theory and applications. SPE Form. Eval. 11 (4), 263–271. Pennsylvania.
Al Khalifah, H., Glover, P., Lorinczi, P., 2020. Permeability prediction and diagenesis in Zhu, L.-q., Zhang, C., Wei, Y., Zhang, C.-m., 2017. Permeability prediction of the tight
tight carbonates using machine learning techniques. Mar. Petrol. Geol. 112, 104096. sandstone reservoirs using hybrid intelligent algorithm and nuclear magnetic
Ali, S.S., Nizamuddin, S., Abdulraheem, A., Hassan, M.R., Hossain, M.E., 2013. Hydraulic resonance logging data. Arabian J. Sci. Eng. 42 (4), 1643–1654.
unit prediction using support vector machine. J. Petrol. Sci. Eng. 110, 243–252.
Amaefule, J.O., Altunbay, M., Tiab, D., Kersey, D.G., Keelan, D.K., 1993. Enhanced
Reservoir Description: Using Core and Log Data to Identify Hydraulic (Flow) Units

17

You might also like