You are on page 1of 34

Accepted Manuscript

Geomechanical parameter estimation from mechanical specific energy using artificial


intelligence

Mohammad Anemangely, Ahmad Ramezanzadeh, Mohammad Mohammadi


Behboud

PII: S0920-4105(18)31158-6
DOI: https://doi.org/10.1016/j.petrol.2018.12.054
Reference: PETROL 5614

To appear in: Journal of Petroleum Science and Engineering

Received Date: 6 April 2018


Revised Date: 17 December 2018
Accepted Date: 20 December 2018

Please cite this article as: Anemangely, M., Ramezanzadeh, A., Mohammadi Behboud, M.,
Geomechanical parameter estimation from mechanical specific energy using artificial intelligence,
Journal of Petroleum Science and Engineering (2019), doi: https://doi.org/10.1016/j.petrol.2018.12.054.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Geomechanical Parameter Estimation from Mechanical Specific Energy using
Artificial Intelligence
Mohammad Anemangely1; Ahmad Ramezanzadeh1*; Mohammad Mohammadi Behboud1

School of Mining, Petroleum and Geophysics engineering, Shahrood University of Technology, Shahrood,
Iran
*
Corresponding author’s E-mail: aramezanzadeh@gmail.com

PT
Abstract
A comprehensive knowledge of geomechanical characteristics of the formations above a
reservoir (cap rock, in particular) can prevent a large spectrum of problems during drilling

RI
operation. Furthermore, access to such information during drilling operation greatly contributes
to proper adjustment of controllable parameters and enhanced rate of penetration (ROP).
Unfortunately, due to the high costs associated with the acquisition of petrophysical logs, these

SC
sets of data are commonly acquired within the reservoir interval only, i.e. at most 20% of total
drilled length. As such, the present contribution is an attempt to use the concept of mechanical
specific energy (MSE) to estimate geomechanical parameters of rock. For this purpose, data was
collected from two vertical wells drilled into an oilfield in southwest of Iran. Firstly, based upon

U
available drilling reports and Tukey method, outlier data points were detected and omitted.
Then, MSE was evaluated using different models proposed for this purpose previously. The
AN
evaluation results indicated that, the results obtained from Dupriest and Keoderitz (2005) model
were in good agreement with actual conditions at the considered wells. Therefore, the obtained
MSE from this model along with flow rate (FR), bit tooth wear (CT), and Depth logs were used
M

as input for multivariate nonlinear regression (MNLR) method as well as multi-layer perceptron
neural network combined with cuckoo optimization algorithm (MLP-COA) and also with particle
swarm optimization (MLP-PSO) to estimate confined compressive strength (CCS), uniaxial
D

compressive strength (UCS), internal friction angle (ϕ), and Poisson’s ratio (ν). Models were
trained on data from one well and then validation-tested on the other well’s data. Results of
TE

adopting these models indicated that, as far as the estimation of geomechanical parameters was
concerned, intelligent models were of higher accuracy and reliability than the regression model.
A comparison between the results of MLP-COA and MLP-PSO models showed that, COA
outperforms PSO algorithm in achieving a model of higher accuracy and reliability. Results of
EP

the three models indicated that, the presented method in this research possesses large potentials
for estimating CCS, UCS, and ϕ parameters, and it can be stipulated with certainty that, the
proposed method can be used to estimate the parameters at other wells across the field under
C

study provided further data is available from more wells and even the formations overlying the
reservoir. However, application of this method for estimating ν shall be practiced with care.
AC

Keywords: mechanical specific energy, Geomechanical parameters prediction, Artificial


Intelligence, Outlier detection

1. Introduction
The concept of MSE was first introduced by Teal in 1965. Teal defined MSE as the
mechanical work done to remove unit volume of rock. In drilling operation, MSE is proportional
to input energy with ROP of drilling tool serving as the coefficient of proportionality (Dupriest
et al., 2005). As such, in drilling industry, this concept has been utilized as a tool for evaluating
performance of drilling operation, because maximum efficiency (maximum ROP) can be
achieved at the point where MSE is at its lowest value. The lowest value of MSE in drilling
operation corresponds to confined compressive strength (CCS) of the rock. This MSE is spent to

1
ACCEPTED MANUSCRIPT
break the rock into fragments of the largest that can be carried by the drilling fluid. However, if
the fragments become any smaller than this size, the amount of consumed energy will increase,
while fragmentation of the rock into sizes larger than what can be carried by the drilling fluid
requires re-fragmentation of those before those can be carried out of the well; and this re-
fragmentation process will require even more energy. These two sub-optimal cases end up with
reduced efficiency of drilling operation.

In rotary drilling, the required energy for breaking the rock is provided by two forces, namely
weight on bit (WOB) and bit RPM. On this basis, the deal of energy spent to break the rock can

PT
be calculated using WOB, RPM, and drilling torque. The most significant relationships for
calculating MSE are presented on Table 1. The first relationship for calculating MSE for a rotary
system was proposed by Teal (1965) based on experimental results. Bit torque is easy to measure

RI
in laboratory, while most of field data is acquired using measurement-while-drilling (MWD)
system at surface (Dupriest and Koederitz, 2005). As such, evaluation of bit torque has been a
challenge when applying the model on field data; when the value of this parameter is unknown,

SC
the calculated MSE will be associated with large errors. Considering this problem Pessier and
Fear (1992) developed a relationship for estimating bit torque using WOB and measured torque
at surface, and devised it to improve the Teal’s model. Given that the Teal’s model was based on
experimental data, Dupriest and Koederitz (2005) showed that the calculated value of MSE by

U
this model based on field data was three times as large as CCS of rock. As such, they suggested
to consider 35% of the MSE value obtained from Teal’s model for field data. Cherif and Bits
AN
(2012) used the model developed by Dupriest and Koederitz to design and optimize the needed
PDC bit for drilling parts of two wells. Results of implementing the model showed that, actual
values of MSE were not 35% of the MSE value obtained from Teal’s model, but rather 26% and
64% of the MSE value obtained from Teal’s model. Given that mechanical efficiency of bit
M

depends on many factors including well type, drilling depth, bit type, and downhole tool, so that
its value shall be determined according to real conditions into the well (Chen et al., 2014).
Knowing that WOB is a primary factor for determining MSE and considering the large
D

difference between actual WOB and measured WOB at surface in deviated and horizontal wells
(due to the friction between drilling string and borehole wall, Chen et al. (2014) developed
TE

relationships for calculating MSE in these wells. A comparison between the results of this model
and those of previous models showed that the calculated value of MSE using the model
proposed by Chen et al. was so close to the rock CCS. This model can be applied not only in
EP

horizontal and deviated wells, but also in vertical ones. It should be noted that, bit condition
imposes direct impacts on the energy transferred to rock in drilling operation. The further the
inserts on a bit are abraded, the greater will be the loss of mechanical energy, and this increase in
energy loss occurs at higher rates in lower abrasions (Al-Sudani, 2017). Al-Sudani (2017)
C

developed a new analytic relationship to investigate the effect of abrasion of bit inserts on
transmission of mechanical energy from bit to the rock. This relation is way different from the
AC

previously presented relationships for such a purpose, and considers a larger weight for the
effect of ROP, rather than WOB, bit RPM, and torque, in MSE calculation.

2
ACCEPTED MANUSCRIPT

Table 1. The most significant relationships proposed for calculating MSE.

120 . .
No. Model Model Formula
= +
1 Teal (1965)
.
1 13.33. .
= .( + )
2 Pessier and Fear (1992)
.
= 36
.

PT
120 . .
= 0.35( + )
3 Dupriest and Koederitz
.
4 480 .
(2005)

RI
= ( + )
4 Cherif and Bits (2012)
.
1 13.33. .
= . .( + )
5 Chen et al. (2016)
.

SC
= . !"# $%

= 36
. . !"# $%
6 Al-Sudani (2017)
U &
. (120 . ) .
AN
g ( . )12*
=
M
D

Due to high costs associated with the acquisition of petrophysical logs (Anemangely et al.,
2017a), the petrophysical logs are commonly acquired within reservoir interval only. This is
TE

while; reservoir thickness comprises utmost 20% of drilled depth. In addition, many of drilling
problems occur within the layers above reservoir, particularly cap rock. As such, it is equally
important to identify rock properties in the layers overlaying a reservoir. Since the energy spend
to break a rock in drilling process is linked to rock properties, it can serve as an appropriate
EP

criterion for estimating rock properties. Meanwhile, compared to hydraulic energy, mechanical
energy plays a greater role in rock breakage (Chen et al., 2016). Therefore, the present research
uses the MSE calculated using drilling parameters is used to estimate rock properties.
C
AC

2. Methodology
Figure 1 demonstrates the flow of steps taken to estimate rock properties from MSE. For this
purpose, we begin with collecting the required data from the wells understudy. The data was
then checked for identification and elimination of outliers. In the next step, MSE was calculated
at studied wells using previous relationships, and the values of maximum consistency with actual
well data were selected for the rest of analysis. MLP neural network combined with COA and
PSO algorithms were used to estimate rock properties. All of these steps are described in the
following subsections.

3
ACCEPTED MANUSCRIPT

PT
RI
U SC
Figure 1. Flow of steps taken in this research.
AN
2.1. Research data
For the sake of this study, data from two vertical wells drilled into an oilfield in southwest of
Iran was used. Due to confidentiality reasons, the wells are herein referred to as Wells A and B.
Logging-while-drilling (LWD) data was acquired all along the wells, but costly petrophysical
M

logs were only run within reservoir interval along the well column. Therefore, only the data
within reservoir intervals along Wells A and B are used in this study.
D

2.1.1. Input data


As was mentioned before, MSE plays a great role in breaking rocks, so that it can well
TE

represent rock properties. However, inconsistency between flowrate and downhole drill cuttings
can end up increasing the value of MSE. On the other hand, bit inserts condition is an important
factor in energy transmission to the rock, and eroded inserts may inhibit effective delivery of
energy to the rock, thereby enhancing the value of MSE. In addition to MSE, depth of the rock
EP

also impacts its physical and mechanical properties. Therefore, inputs into a geomechanical
parameter estimator model included not only MSE, but also WOB, RPM, torque (Tor), ROP,
and bit size (BS), which were collected from the considered wells. PDC bits were used in both
C

wells.
Figures 2 and 3 shows curves of the drilling parameters along Wells A and B. Sampling
AC

interval of LWD data along Wells A and B were 0.2 and 1 m, respectively. Ranges of variations
of these parameters along the studied wells are reported in Table 2. An investigation on the
values of these parameters shows that, the data are contaminated with noise. For instance, it can
be observed that the flowrate at a depth of 3380 m into Well A increases to 20,000 gal/min
suddenly, which is practically impossible considering the pumping equipment used for drilling
fluid in this well. This necessitates some sort of pre-processing before we can proceed to the
calculation of MSE and data input into the estimator algorithm.

4
ACCEPTED MANUSCRIPT

PT
RI
SC
Figure 2. Profile of variations of the required drilling parameters for this research within the depth interval studied along Well A.

U
AN
M
D
TE
EP

Figure 3. Profile of variations of the required drilling parameters for this research within the depth interval studied along Well B.
C

Table 2. Ranges of variations of the required drilling parameters along the studied wells.
Well A Well B
AC

Parameters Minimum Mean Maximum Minimum Mean Maximum


value value value value value value
Depth (m) 2873 3132 3391 2404 2716 3028
Bit size (in) 6.125 6.125 6.125 8.5 8.5 8.5
Rotation speed
47.715 236.195 3725.300 45.62 135.10 187.80
(rpm)
Weight on bit (lbf) 0 7478.337 32780.120 1367.30 6926.03 15424.38
Torque (daN.m) 0 113.539 222.789 56.74 90.62 108.78
Penetration rate
0.582 9.611 151.319 2.20 12.02 19.17
(ft/h)
Flow rate (gal/min) 68.507 291.723 20687.334 232.93 332.54 412.35
Tooth wear (%) 0 1.370 3.008 0 1.21 3.55

5
ACCEPTED MANUSCRIPT

2.1.2. Geomechanical parameters


Application of petrophysical logs and execution of laboratory tests on core samples are the
two primary methods for determining mechanical and physical properties of rocks within
reservoir interval. Coring is much more costly than acquiring petrophysical logs. On the other
hand, the data obtained from well logs are continuous; therefore, we used petrophysical log data
for estimating geomechanical properties of the studied formations in this research. For this
purpose, Gamma Ray (GR), density (RHOB), neutron porosity (NPHI), effective porosity

PT
(PHIE), compressive wave slowness (DTCO), and shear (DTSM) logs were collected from the
considered wells. Figures 4 and 5 show variations of the values of petrophysical logs within the
studied depth interval along the two wells. Sampling interval was 0.1524 cm in both wells. Table
3 presents average value and range of variations of the petrophysical logs acquired along the two

RI
wells.

U SC
AN
M
D
TE

Figure 4. Profile of variations of the required petrophysical logs for this research within the depth interval studied along Well A.
C EP
AC

Figure 5. Profile of variations of the required petrophysical logs for this research within the depth interval studied along Well B.

6
ACCEPTED MANUSCRIPT
Table 3. Ranges of variations of the required petrophysical logs along the studied wells.
Well A Well B
Parameters Minimum Mean Maximum Minimum Mean Maximum
value value value value value value
GR (GAPI) 1.987 32.811 146.802 7.859 49.772 128.104
RHOB (G/C3) 2.256 2.575 2.848 2.310 2.622 2.832
NPHI (V/V) 0.002 0.118 0.325 0.003 0.116 0.360
PHIE (V/V) 0 0.060 0.256 0 0.040 0.218
DTCO (us/ft) 46.511 68.637 165.850 49.106 64.617 93.464
DTSM (us/ft) 93.313 129.416 467.408 96.508 121.165 179.363

PT
Using the petrophysical logs, geomechanical rock properties were calculated along the
studied depth intervals in each well. Flow of calculations and used relationships for this

RI
estimation are presented in Appendix. Figures 6 and 7 present variation curves of geomechanical
parameters of interest in this research along Wells A and B, respectively. Among other
geomechanical parameters of rocks, these parameters were selected for this research because of

SC
their widespread applications in reservoir modeling, well bore stability assessment, and drilling
analysis (Chatterjee and Mukhopadhyay, 2002a, 2002b, Das and Chatterjee, 2018, 2017a,
2017b; Singha and Chatterjee, 2015). These parameters will be considered as outputs of the

U
estimator model. Table 4 presents average value and range of variations of this parameters.
AN
M
D
TE
C EP
AC

Figure 6. Profile of variations of petrophysical parameters of interest along Well A.

7
ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
Figure 7. Profile of variations of petrophysical parameters of interest along Well B.

Table 4. Ranges of variations of the considered petrophysical parameters along the studied wells.
M

Well A Well B
Parameters Minimum Mean Maximum Minimum Mean Maximum
value value value value value value
Confined compressive
D

27.516 131.734 260.208 67.15 117.82 186.37


strength (Mpa)
Uniaxial compressive
9.635 71.144 121.019 43.47 77.77 108.73
TE

strength (Mpa)
Internal friction angle
20.869 31.720 49.215 20.56 30.18 40.96
(deg)
Poisson’s ratio 0.294 0.303 0.428 0.23 0.30 0.32
EP

2.1.3. Scale-equalization of data


Since the scale of mud logging data is different from that of petrophysical logs at both wells,
C

one cannot proceed to develop a single database and other steps of the study. There are two
methods for scale-equalization of data, namely upscaling and downscaling. Using the
AC

downscaling method on Well B in this method, one should scale each meter of mudlogging data
to 0.15 m. On this basis, two data points (in the scale of one 1 m) were used to interpolate 5-6
data points in mud logging dataset corresponding to the geomechanical data. It is obvious that
the use of only two points to interpolate as large as 5-6 data points is not reasonable. This is
while, when applying downscaling to Well A, two mud logging data points are used to
interpolate a maximum of two mud logging data points corresponding to the geomechanical
data. However, it shall be noticed that, accuracy and validity of the obtained values via this
method will be lower than those of upscaling method. Therefore, in the next step, upscaling was
used for scale-equalization of data to upscale geomechanical data to the scale of mud logging
data at each well.
In order to upscale geomechanical data at Well A, considering the small difference between
the scales of geomechanical mud logging data, ordinary kriging method was used to interpolate
the values of geomechanical parameters at depth points corresponding to the mud logging data.

8
ACCEPTED MANUSCRIPT
Equation (1) presents general form of ordinary kriging method, where +, depends on the fitted
model to the acquired data points, distance to prediction location, and spatial relationship among
the measured values in the vicinity of the prediction location.
3

- ∗ (/0 ) = 2 +, -(/, ) (1)


,45

In order to determine upscaled values of geomechanical parameters at well B, considering the

PT
large difference between the scales of geomechanical parameters (0.1524 m) and mud logging
data (1 m), simple averaging method was used. For this purpose, for each geomechanical
parameter, available data points along each meter were averaged and the result was assumed as

RI
representative of that meter of drilling.
Once finished with scale-equalization of data, databases were formed for each well
separately. The databases of the Wells A and B contained 2591 and 625 data points,

SC
respectively. Considering the large number of data points at Well A and more extensive ranges
of variation of most of the studied parameters at Well A, rather than Well B, the data from Well
A was used to train the estimator model, while those from Well B were devised to test the
trained model.

2.2. Data preprocessing


U
AN
Serving as inputs into processing algorithms, real data are affected by numerous factors, of
which the presence of noise is a key factor (Wang et al., 1995). Various definitions have been
proposed for noise in literature on statistics and machine learning. Most of the definitions agree
that noisy data can affect learning process (Quinlan, 1986) and extent the training period. In
M

addition, the presence of noise in data can raise problems in extracting rules from data in
machine learning models, which may end up with inappropriate performance of the models in
the process of predicting new data points (Garćia et al., 2013; Lorena and de Carvalho, 2004).
D

However, noise is an unavoidable issue in data measurement. Some research have predicted that,
even under controlled conditions, a minimum of 5% error exists in any dataset (Maletic and
TE

Marcus, 2000; Wu, 1995). There are two major solutions for addressing the noise in data: (1)
omitting the data points which are highly immersed in noise (outliers), and (2) noise attenuation.
Omission of outliers seems to be way easier than attenuating the effect of noise within real data,
EP

because exact level and characteristics of the noise in real data are not known. A very important
challenge in the elimination of outliers is how to recognize a data point as outlier.
In this study, two approaches were considered to identify and remove outliers. Firstly, daily
drilling reports of the studied well were checked and the data related to the depths wherein
C

drilling problems were reported were omitted from the dataset. For this purpose, we considered
wellbore collapse and intense vibration of drilling string as drilling problems. In the next step,
AC

Tukey’s method was used to eliminate outliers. To this end, the first and third quarter were
calculated for each of the considered parameters. Then, the values of LIF and UIF were obtained
for Equations (2), (3), and (4). For each parameter, the data points with values below LIF or
above UIF were identified as outliers and eliminated from the rest of research.

67 = 8 × (7: − 75 ) (2)

<6= = 75 − 67 (3)

>6= = 7: + 67 (4)

Applying the above-described outlier emission method on the databases of the Wells A and
B, some 935 and 175 data points were eliminated from the corresponding datasets, respectively.
9
ACCEPTED MANUSCRIPT
2.3. Calculation of MSE
For the wells under study, the value of MSE was calculated using the relationships mentioned
in Table 1. Range of variation of MSE using each of these models is reported in Table 5.
Investigations indicate that, the models proposed by Cherif and Bits (2012) and Chen et al.
(2014) had estimated a MSE value smaller than CCS at majority of depth points along the wells.
This was while Al-Sudani’s (2017) model estimated MSE values much larger than those
estimated by other models and also than CCS of the rock along both wells. MSE values obtained
from the models proposed by Teal (1965) and Pessier and Fear (1992) were highly similar to one
another, but still higher than the corresponding CCS values at both wells. Meanwhile, the values

PT
obtained from Dupriest and Koederitz (2005) model were so close to CCS values, although the
calculated MSE from this model was smaller than CCS of the rock at limited depth points. A
comparison between the values of rock CCS and MSE obtained from the model proposed by

RI
Dupriest and Koederitz (2005) along the studied wells is demonstrated in Figure 8. Even though
the main objective of this study has not been to verify the results of MSE estimator models, but
it seems that the considered Em coefficient by the Dupriest and Koederitz (2005) model is not

SC
suitable for the entire length of the studied depth interval. However, it should be kept in mind
that, in order to select the most appropriate MSE estimator model for this study, one should pay
attention to the resultant correlation coefficient between model-resulted MSE value and CCS of
the rock. An investigation on the correlation coefficient between model-resulted MSE values and

U
CCS of the rock indicated higher correlation coefficients for the models proposed by Tear (1965)
and Dupriest and Koederitz (2005), when correlated to the CCS of the rock. Considering what
AN
was mentioned above, the MSE value obtained from Dupriest and Koederitz (2005) model was
used to predict geomechanical parameters of the considered formation in this study.
M

Table 5. Ranges of variation of MSE values obtained using different models introduced in Table 1.
Well A Well B
Model Minimum Mean Maximum Minimum Maximum
Mean value
value value value value value
D

Teal (1965) 11659.670 49471.451 194679.9 29377.693 79146.396 141687.957


Pessier and Fear
11650.072 49459.384 194631.198 29370.641 79126.952 141652.897
TE

(1992)
Dupriest and
4080.884 17315.01 68137.95 10282.193 27701.238 49590.785
Koederitz (2005)
Cherif and Bits (2012) 1356.267 6456.479 18007.057 2942.858 7577.484 13365.842
EP

Chen et al. (2016) 3251.746 13806.233 56443.056 8517.486 22946.816 41079.340


Al-Sudani (2017) 84362.45 4006040 29041235 286569.450 884186.584 1868313.181
C
AC

10
ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D

(a) (b)
Figure 8. Depth profiles of changes in CCS and MSE along (a) Well A and Well B (b).
TE

2.4. Combined neural network with COA and PSO


Artificial neural networks (ANNs) are inspired by the function of biologic neurons in
human’s brain. The most significant advantage of these networks is their ability to find complex
EP

nonlinear relationships among variables (Ali, 1994), and the fact that those can generalize a
model trained for a particular problem to other similar problems. Fundamental steps toward
constructing a neural network model include data preparation, selection of input variables, and
C

selection of type and structure of neural networks, transform function, and training algorithm
(Maimon and Rokach, 2005). Neural networks have a diverse spectrum of structures, of which
AC

multi-layer perceptron (MLP) networks are more popular because of their ability to estimate any
function at an acceptable accuracy provided enough number of hidden units is considered
(Bishop, 2007). Furthermore, autoregressive multi-layer neural networks have been widely
studied and applied to practical problems. Ideally, these neural networks are appropriate for
modeling relationships between a set of input variables (i.e. predictors) and one or more output
variable (i.e. response) (Maimon and Rokach, 2005; Ashrafi et al., 2018). As such,
autoregressive MLP was used to estimate ROP at the considered wells.
Number and scattering of data points and learning process contribute to generalizability of
trained models to other problems. The larger the number and the wider the scattering of available
data points, the higher will be the generalizability of the resulting trained model to other similar
problems. By scattering we mean the presence of information from various cases. Learning
process refers to the process wherein network parameters (including weighting factors and
biases) are modified through a continuous process to achieve optimal value of an objective

11
ACCEPTED MANUSCRIPT
function. In most cases, the objective function is the minimization of associated error with
estimator or classifier model. The most popular and significant learning algorithm classically
used for MLP neural networks is Levenberg-Marquardt algorithm; however, application of this
algorithm to complicated problems have shown that, in such problems, it can do no more than
returning local optima. Considering limitations of classical algorithms in determining optimal
values of the parameters of neural networks, COA and PSO algorithm were selected to have the
selected neural network trained. In order to determine the structure of the neural network, that is
the number of hidden layers and neurons in each layer, a trial and error approach was followed.
On this basis, a MLP neural network with three hidden layers containing 4, 4, and 3 neurons in

PT
the first, second, and third layer, respectively, was selected. In order to train the model, data
from Well A were divided into two subsets, namely training and testing subsets. The training
subset included 70% of the whole dataset while the remaining 30% of the whole dataset was

RI
considered as testing subset.
Substitution of metaheuristic and communal intelligence algorithms instead of conventional
learning algorithms in MLP neural network showed that, the metaheuristic algorithms have large

SC
potentials in finding a global optimal solution, as compared to classic optimization algorithms
used with neural networks. Cuckoo optimization is a metaheuristic algorithm recently proposed
by Yang and Deb, 2009). This algorithm is inspired by the lifestyle of a bird called cuckoo.
Rajabioun (2011) proposed an optimized version of this algorithm. A comparison between this

U
algorithm and other optimization algorithms indicated that the results of this algorithm were
much better than those of other (Kawam and Mansour, 2012) while convergence rate of this
AN
algorithm was higher (Anemangely et al., 2017b; Anemangely et al., 2018).
Figure 9a shows a flowchart of COA. Similar to other evolutionary algorithms, this algorithm
begins with an initial population of cuckoos (Npop). These initial cuckoos lay a number of eggs (5
M

– 20 eggs) in nests of a number of host birds. A number of the eggs, which are most similar to
those of the host bird, find an opportunity to grow and turn into a mature bird. Other eggs,
however, are detected and then killed by the host bird. The remaining eggs indicate
D

appropriateness of the nests within the corresponding area. The objective is to maximize the
number of surviving eggs by finding the most appropriate position for laying eggs. Cuckoos are
in search for finding appropriate locations for laying eggs in an attempt to increase the rate of
TE

surviving eggs. Once the remaining eggs grew up and turned into mature cuckoos, they form
societies. Each society has its own habitat. The best habitat of all societies will be the destination
of cuckoos in other habitats. They then migrate toward this best habitat and reside somewhere
EP

close to the best habitat. Considering the number of eggs laid by each cuckoo, total number of
eggs, maximum Egg Laying Radius (ELR) adjustment factor (α), and lower (Varlow) and higher
(Varhi) bounds for each decision variable, a number of ELRs are defined for each cuckoo
C

(Equation (5)). This process is then continued until the best location with maximum profit is
found, wherein all cuckoos gather.

@/AB C D/ EF D/DG H I JJI


AC

< =?× × (MK − MK OPQ )


FKL E/AB C JJI N, (5)

In order to solve an optimization problem, the array of decision variables shall be formed. In
an optimization algorithm with Nvar decision variables, a habitat is defined as a 1 × Nvar array

corresponding array to that habitat (R5 , R , … , RUVWX ). Given that COA is originally used to
which indicates current habitat of the cuckoo. Profit of a habitat is determined by evaluating the

maximize an objective function, then the following objective function was used for this study:

CYF = −D IF(ℎKBYFKF) = −D IF[\R5 , R , … , RUVWX ]^ = − (6)

12
ACCEPTED MANUSCRIPT
Performing a sensitivity analysis on COA parameters, number of cuckoos were set to 100,
number of iterations were set to 120, and ELR adjustment factor was set to 1.
Figure 9b provides flowchart of PSO algorithm. PSO algorithm finds the best solution within
the search space using a population of particles called swarm. The algorithm begins with
randomly initializing the position of population within the search space (lower and upper bounds
of decision variables) and also initializing particle velocities between a lower (- Vmax) and upper
(Vmax) limits. The positions specified for the particles are stored as their personal best positions
(Pb). Then, positions of all particles are evaluated using the objective function, and the particle
with the lowest value of objective function is selected as the one with the global best position

PT
(Gb). At each iteration, a new velocity (Vi (t + 1)) is defined for each particle (i) based on its
previous velocity (Vi (t)) and distance between its current position (xi (t)) and the best personal
and global positions (Equation (7)). Subsequently, new position of the particle (xi (t + 1)) is

RI
evaluated based on its previous position and the new velocity (Equation (8)). Then, the new
position of the particle is reevaluated using the objective function. The personal and global best
values are updated by comparing the new values of the objective function against previous

SC
personal and global best values, respectively. Iterations are continued until stoping criterion is
reached.

M, (F + 1) = _M, (F) + D5 5 [ B, (F) − R, (F)^ + D (` (F) − R, (F))


R, (F + 1) = R, (F) + M, (F + 1)
(7)

U
(8)
AN
where i = 1, 2, …, n, n is the number particles in the sward, w is inertial weight (controls
recurrence of the particle velocity; Pedersen and Chipperfield, 2010), c1 and c2 are positive
coefficients called personal and communal learning factors, respectively, and r1 and r2 are
M

random numbers in the interval [0, 1] (Coello et al., 2007).


According to the results of the sensitivity analysis performed on the parameters of this
algorithm, number of particles, number of iterations, and personal and communal learning
D

factors were set to 160, 400, 2, and 2 respectively.


TE
C EP
AC

(a) (b)
Figure 9. Flowchart of (a) COA and (b) PSO algorithm (Anemangely et al., 2018a).

13
ACCEPTED MANUSCRIPT
3. Results and discussion
An investigation on the relationship between CCS, in one hand, and UCS and ϕ of the rock
along the studied wells showed a relatively consistent linear relationship between them. On the
other hand, given that drilling MSE shall principally be equal to the rock CCS, a linear
relationship is expected between MSE and CCS. As such, the presence of a linear relationship
between each of the studied geomechanical parameters and MSE was investigated. Figure 10
shows the relationship between the considered geomechanical parameters and MSE. As can be
observed on the figure, it can be firmly stipulate that, there is a linear relationship between the
considered geomechanical parameters and MSE, so that MSE increased with increasing the

PT
geomechanical parameters of CCS, UCS, ϕ, and ν. Principally, with increasing strength
parameters of rock (e.g. CCS, UCS, and ϕ), the consumed mechanical energy for breaking the
rocks increases accordingly. Therefore, it makes sense to expect a linear relationship between

RI
these parameters and MSE. Given that the stronger a rock, the lower will be its ν value, it seems
reasonable to consider an inverse relationship between ν and MSE. This is while, observations
indicated a direct relationship between these parameters at Well A. Obtained values of the

SC
coefficient of determination for the linear relationship between MSE and geomechanical
parameters indicate that, MSE is well related to compressive strength of the rock. It further
seems that, MSE is well related to ϕ, while the relationship between ν and MSE was much
weaker than that of other parameters in this well. As such, one may expect lower accuracy for

U
the estimator models developed to estimate the value of ν. An investigation on the scattering of
data points across the corresponding space to the relationship between MSE and each of the
AN
geomechanical parameters along Well A in the same figure suggests a linear or parabolic
relationship between MSE and rock strength parameters (CCS, UCS, and ϕ). Distribution of data
points across the corresponding space to the relationship between ν and MSE indicates that no
M

consistent relationship can be identified between them.


D
TE
C EP
AC

14
ACCEPTED MANUSCRIPT

PT
RI
(a) (b)
R-Square = 0.41 R-Square = 0.35

SC
50 0.335

0.33
45
0.325

U
40 0.32

0.315
AN
35
0.31

30 0.305

0.3
M

25
0.295

20 0.29
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
D

4
MSE(psi) 10 MSE(psi) 104

(c) (d)
TE

Figure 10. Relationship between MSE and geomechanical parameters at Well A: (a) CCS, (b) UCS), (c) internal friction angle,
and (d) Poisson’s ratio.

Figure 11 shows the relationship between the considered geomechanical parameters and MSE
EP

at Well B. As can be observed on the figure, the values of coefficient of determination obtained
for the relationships between the considered geomechanical parameters and MSE are lower in
this well, as compared to those at Well A. In contrary to Well A, CCS of the rock exhibits a
better relationship to MSE, as compared to other rock properties. Furthermore, MSE exhibits a
C

better relationship to ϕ rather than UCS. Similar to the Well A, Well B also indicated a weak
relationship between ν and MSE. Considering the obtained coefficients of determination,
AC

estimator models for geomechanical parameters are expected to be capable of estimating rock
strength parameters (CCS, UCS, and ϕ) at Well B well accurately. Similar to Well A, direct
relationships were identified between MSE and rock strength parameters at Well B. Even though
it seems MSE is directly related to ν also at this well, but the low value of the obtained
coefficient of determination indicates that probability of such a direct relationship is very low.

15
ACCEPTED MANUSCRIPT

PT
RI
(a) (b)

U SC
AN
M
D
TE

(c) (d)
Figure 11. Relationship between MSE and geomechanical parameters at Well B: (a) CCS, (b) UCS), (c) internal friction angle,
and (d) Poisson’s ratio.
EP

Figure 12 shows cross plot of training testing phases of MLP-COA model when applied on
the data from Well A. As can be observed on this figure, the MLP-COA model produces higher
coefficients of determination in estimating CCS, UCS, and ϕ, rather than ν. Results of
implementing MLP-PSO model on the data from Well A are presented in Figure 13, indicating
C

better capabilities of this model estimating rock strength parameters rather than Poisson’s ratio.
As such, it can be stipulated that the presented method is of high accuracy in estimating rock
AC

strength parameters. On the other hand, a comparison between associated errors with the used
models for training and testing phases on the data from Well A depicts that the MLP-COA
model enjoys higher accuracy than that of MLP-PSO model. Furthermore, comparing errors
incurred at training and testing phases of these models, a small difference was noticed between
the error at training stage and that at testing stage with MLP-COA model, as opposed to the
situation with the other model. Therefore, this model was found to be more reliable than the
other model.

16
ACCEPTED MANUSCRIPT

(a)

PT
RI
Observed UCS(Mpa)

SC
(b)

U
AN
RMSE = 1.355 & R-Square = 0.97473 RMSE = 1.3885 & R-Square = 0.97388
M

45 45
Data point Data point
Best linear fit Best linear fit
Y=T Y=T
40 40
D

35 35

(c)
TE

30 30

25 25
EP

25 30 35 40 45 25 30 35 40 45

Predicted (deg) Predicted (deg)


C
AC

Observed

(d)

Figure 12. Cross plot of estimated values versus observed data at training (left) and testing (right) phases with MLP-COA model:
(a) CCS, (b) UCS), (c) internal friction angle, and (d) Poisson’s ratio.

17
ACCEPTED MANUSCRIPT

Observed CCS(Mpa)
Observed CCS(Mpa)

(a)

PT
RI
Observed UCS(Mpa)

Observed UCS(Mpa)

SC
b)
(

U
AN
M
D
(deg)

(deg)
TE
Observed

Observed

(c)
C EP
AC

Observed
Observed

d)
(

Figure 13. Cross plot of estimated values versus observed data at training (left) and testing (right) phases with MLP-PSO model:
(a) CCS, (b) UCS), (c) internal friction angle, and (d) Poisson’s ratio.

18
ACCEPTED MANUSCRIPT

Figure 14 compares error reduction rates through various iterations of COA and PSO
algorithm. As can be seen on this figure, error reduction rate of MLP-COA model was higher
than that of MLP-PSO model, i.e. the MLP-COA model could achieve lower errors within less
iterations. These results suggest superiority of COA in achieving global optimal solution, while
PSO algorithm ended up with a solely local optimal solution. However, it is worth noticing that,
PSO algorithm converged to its solution within shorter processing time than that of COA,
because of the values of the parameters considered in these algorithms.

PT
35 26
MLP-COA model MLP-COA model
24 MLP-PSO model
MLP-PSO model

30 22

20

RI
25 18

16

20 14

SC
12

15 10

U
10 6
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Iteration Iteration
AN
(a) (b)
16 0.7
MLP-COA model MLP-COA model
14 MLP-PSO model MLP-PSO model
0.6

12
M

0.5

10
0.4
8
0.3
D

0.2
4
TE

2 0.1

0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Iteration Iteration
EP

(c) (d)
Figure 14. A comparison between convergence rates of COA and PSO algorithm for training MLP neural network using data
from Well A for estimating the following parameters: (a) CCS, (b) UCS), (c) ϕ, and (d) ν.
C

Figure 15 presents a comparison between the values of geomechanical parameters obtained


from petrophysical logs and those estimated using MLP-COA and MLP-PSO models within the
AC

depth interval considered along Well A. As can be observed in this figure, both models could
well predict general trend of changes in geomechanical parameters. Even though the values
obtained from the two estimator models match at some points, but it is obvious that the values
obtained from MLP-COA model are much closer to the measured values of the corresponding
geomechanical parameters. In this figure, it is further observed that, the obtained values from
these two models well reproduce actually measured values of CCS, UCS, and ϕ, but exhibit
significant differences when it comes to ν. An investigation on the results shows that, compared
to MLP-PSO model, MLP-COA model results were closer to actually measured values.

19
ACCEPTED MANUSCRIPT

Depth(m)

PT
RI
SC
Figure 15. Comparison between observed values and estimated results of rock parameters using MLP-COA and MLP-PSO
models within the studied depth interval along Well A.

U
AN
4. Comparing MLP-COA, MLP-PSO, and MNLR models
In order to validate and verify performances of MLP-COA and MLP-PSO models,
multivariate nonlinear regression (MNLR) model was used – stepwise multivariate regression
where quadratic mode was defined as model specification. In this case, the model would include
M

intercept, linear terms, interactions, and squared terms. In order to select input parameters, five
different criteria were chosen, namely a change in SSE, an increase in R-square, an increase in
adjusted R-square, a change in AIC, and a change in BIC. Comparing the results of these criteria
D

for this problem showed that, adoption of SSE minimization as a criterion produces a model of
higher accuracy, as compared to the models developed based on other criteria.
TE

Table 6 reports the selected variables and corresponding weighting factors when MNLR
method is applied to the data from Well A to estimate CCS. Upper and lower bounds
(confidence interval) of the weighting factors in this table were determined at 95% confidence.
EP

All of the selected terms had p-values below 5%, indicating significant relationships between
these predictive variables and the independent variable. Equation (9) expresses the estimator
model obtained by applying NMLR method for estimating CCS using the data from Well A. The
values of coefficient of determination and error for this model, when applied to the data from
C

Well A, are presented in Table 7. The obtained p-value for this model indicates the significance
of this relationship.
AC

Table 6. Selected variables and obtained coefficients for each term using MNLR method in estimating CCS based on the data
from Well A.
Coefficient
Added Estimated confidence Standard
T-statistics P-Value
variables Coefficient intervals error
Lower Upper

a5
Intercept 715.85 -379.973 1811.663 558.69 1.2813 0.02

a 2.05 × 10!b0
-0.6474 -1.3834 0.0886 0.3752 -1.7254 0.0084

a: 2.59 × 10!5d
0.0202 0.0173 0.0231 0.0015 13.677

ab
-404.75 -500.669 -308.825 48.905 -8.2762

−6.87 −4.97 4.827


-0.0021 -0.0049 0.0007 0.0014 -1.4594 0.0145
a5 a −5.921 × 10!d 3.66 × 10!::
× 10!d × 10!d × 10!f
-12.267

20
ACCEPTED MANUSCRIPT
a5 a: 1.02 × 10!5b
3.98 2.37 5.024
0.1355 0.1015 0.1696 0.0174 7.8086
a ab 1.383 × 10!f
× 10!g × 10!f × 10!g
2.7531 0.006
6.303
a5
× 10!h
0.00014 0.00002 0.00027 2.2878 0.0223
−2.29 −1.12 2.985
a −1.702 × 10!g 1.4 × 10!g
× 10!g × 10!g × 10!i
-5.7021
a: -10.245 -15.1027 -5.3868 2.477 -4.1363 3.71 × 10!h

jj = 715.85 − 0.6474a5 + 0.0202a − 404.75a: − 0.0021ab − 5.921


× 10!d a5 a + 0.1355a5 a: + 1.383 × 10!f a ab + 0.00014a5 − 1.702

PT
× 10!g a − 10.245a:
(9)

RI
Table 7. Results of applying NMLR method on the data from Well A to estimate CCS.
RMSE R-square Adjusted R-square P-Value
MNLR model 16.8 0.907 0.906 0

SC
The terms selected for estimating UCS based on MNLR method using the data from Well A
are presented in Table 8. Associated p-values with these terms indicated significant relationships

U
between the terms and UCS. The results in this table show that, there was no significant
relationship between FR and MSE, so that the parameter was neglected when estimating UCS.
AN
Equation (9) sets out the general form of the model obtained from MNLR method for estimating
UCS using the data from Well A. In contrary to the expectation inferred from Figure 10,
obtained value of R-squared for this model was lower than that of the regression model
developed for estimating CCS. In the meantime, the value of this coefficient was remarkable.
M

Associated p-value with Equation (10) indicates significance of the developed regression model.

Table 8. Selected variables and obtained coefficients for each term using MNLR method in estimating UCS using the data from
D

Well A.
Coefficient
Added Estimated confidence Standard
TE

T-statistics P-Value
variables Coefficient intervals error
Lower Upper
8.6101 × 10!5
a5 3.5344 × 10!5
Intercept 1744.8 1247.173 2242.335 253.69 6.8776
EP

a 3.7599 × 10!5g
-1.1938 -1.528 -0.86 0.1704 -7.007

a: 3.5268 × 10!d
0.0061 0.005 0.008 0.0007 8.7875

−2.15 −1.3 2.2703


-107.95 -153.452 -62.449 23.198 -4.6534
a5 a −1.7023 × 10!d 1.0505 × 10!5:
× 10!d × 10!d × 10!f
C

-7.4983
a5 a: 1.8436 × 10!h
2.8617
0.0354 0.019 0.052 0.0082 4.2956
a5 4.2389 × 10!5:
AC

× 10!h
0.0002 0.00015 0.00027 7.3069
−9.87 −4.22 1.4414
a −7.042 × 10!i 1.1314 × 10!d
× 10!i × 10!i × 10!i
-4.8856
a: -1.9981 -4.340 0.343 1.1938 1.6737 0.0094

>j = 1744.8 − 1.194a5 + 0.006a − 107.95a: − 1.702 × 10!d a5 a + 0.035a5 a:


+ 0.0002a5 − 7.042 × 10!i a − 1.998a:
(10)

Table 9. Results of applying NMLR method on the data from Well A to estimate UCS.
RMSE R-square Adjusted R-square P-Value
MNLR model 8.15 0.891 0.891 0

21
ACCEPTED MANUSCRIPT
The terms selected for estimating ϕ based on MNLR method using the data from Well A are
presented in Table 10. Associated p-values with these terms indicated significant relationships
between the terms and ϕ. Equation (11) expresses the regression model obtained by applying
MNLR method on the data from Well A. Obtained R-squared for this model (Table 11) is higher
than those of previous regression models, indicating that changes in independent parameters
explain variations in ϕ better than other geomechanical parameters.

Table 10. Selected variables and obtained coefficients for each term using MNLR method in estimating ϕ using the data from
Well A.
Coefficient

PT
Added Estimated confidence Standard
T-statistics P-Value
variables Coefficient intervals error
Lower Upper
1.36 × 10!::

RI
a5 7.56 × 10!::
Intercept -833.932 -966.338 -701.527 67.505 -12.354

a 8.45 × 10!5f
0.553 0.4641 0.6419 0.0453 12.202

a: 4.62 × 10!5db
0.0015 0.0011 0.0018 0.00018 8.414

SC
ab 2.18 × 10!g
-187.444 -199.419 -175.469 6.105 -30.702

−5.852 −3.653 5.61


-0.0020 -0.0027 -0.0013 0.00035 -5.624
a5 a −4.752 × 10!f 5.02 × 10!5f
× 10!f × 10!f × 10!g
-8.478
a5 a: 6.89 × 10!5d0
1.856

U4.429 6.56
0.0655 0.0613 0.0698 0.0022 30.209
a ab 3.143 × 10!g 1.8 × 10!d
× 10!g × 10!g × 10!i
4.793
AN
−1.039 −7.409 7.61
a5 −8.901 × 10!h 1.98 × 10!:0
× 10!: × 10!h × 10!d
-11.699
a: 2.83 × 10!55d
1.98 1.203 2.56
-7.741 -8.352 -7.131 0.3111 -24.881
ab 7.005 × 10!g
M

× 10!g × 10!f × 10!g


2.734 0.0063

k = −833.932 + 0.553a5 + 0.0015a − 187.444a: − 0.002ab − 4.752 × 10!f a5 a


+ 0.0655a5 a: + 3.143 × 10!g a ab − 8.901 × 10!h a5 − 7.741a:
D

+ 7.005 × 10!g ab
(11)
TE

Table 11. Results of applying NMLR method on the data from Well A to estimate ϕ.
RMSE R-square Adjusted R-square P-Value
EP

MNLR model 2.12 0.939 0.938 0

The terms selected for estimating ν using MNLR method based on the data from Well A are
C

presented in Table 12. An investigation on the results of applying this method in this table shows
that, FR has no significant effect on estimated ν. Equation (12) sets out the general form of the
AC

regression model obtained for estimating ν. Even though the obtained p-value for this model
(Table 13) is indicative of its significance, but lower value of R-squared of this model compared
to other regression models implies that, car should be taken for applying this model onto other
datasets.

Table 12. Selected variables and obtained coefficients for each term using MNLR method in estimating ν using the data from
Well A.
Coefficient
Added Estimated confidence Standard
T-statistics P-Value
variables Coefficient intervals error
Lower Upper
1.981 × 10! b

9.344
Intercept 1.442 1.170 1.715 0.1392 10.365
a5 2.205 × 10!5b
× 10!h
-0.0007 -0.0009 -0.00054 -7.707

22
ACCEPTED MANUSCRIPT
−6.929 −5.426 3.832
a −6.177 × 10!d 2.102 × 10!hb
× 10!d × 10!d × 10!f
-16.120
a:
1.718 2.190 1.203
-0.035 -0.0611 -0.0094 0.0132 -2.678 0.0075
a5 a 1.954 × 10!i 4.063 × 10!hh
× 10!i × 10!i × 10!50
16.237
4.789 2.318 4.687
a5 a: 1.398 × 10!h
× 10!d × 10!h × 10!d
2.983 0.0029
8.200 1.435 1.569
a5 1.128 × 10!f 9.856 × 10!5:
× 10!g × 10!f × 10!g
7.189
a: -0.0028 -0.0042 -0.0015 0.00068 -4.153 3.452 × 10!h

PT
l = 1.442 − 0.0007a5 − 6.177 × 10!d a − 0.035a: + 1.954 × 10!i a5 a + 1.398
× 10!h a5 a: + 1.128 × 10!f a5 − 0.0028a:
(12)

RI
Table 13. Results of applying NMLR method on the data from Well A to estimate ν.
RMSE R-square Adjusted R-square P-Value

SC
MNLR model 0.00469 0.733 0.732 0

Table 14 presents a comparison among MLP-COA, MLP-PSO, and MNLR models based on
RMSE and R-squared values obtained from the estimation of geomechanical parameters using

U
the data from Well A. Looking at RMSE and R-squared values for different models used to
estimate geomechanical parameters at Well A (as reported in this table) indicate superiority of
AN
intelligent models. In addition, MLP-COA algorithm had higher accuracy and reliability than the
other two models in estimating all of the considered geomechanical parameters. Figure 16
presents a comparison between the values of geometrical parameters estimated using MNLR and
M

MLP-COA models within the studied depth range along Well A. As can be observed in this
figure, Results of MLP-COA model (rather than regression model) were much closer to the
measured values.
D

Table 14. A comparison among MLP-COA, MLP-PSO, and MNLR models based on values of error and coefficient of
determination obtained from the estimation of geomechanical parameters using the data from Well A.
TE

Model Parameters R-Square RMSE


CCS 0.943 13.125

k
UCS 0.925 6.769
MLP-COA
l
0.975 1.371
EP

0.750 0.00453
CCS 0.924 15.140

k
UCS 0.908 7.478
MLP-PSO
l
0.940 2.085
C

0.736 0.00465
CCS 0.907 16.8
AC

k
UCS 0.891 8.15
MNLR
l
0.939 2.12
0.733 0.00469

23
ACCEPTED MANUSCRIPT

PT
RI
SC
Figure 16. A comparison between measured data and estimated values of tock parameters using MLP-COA and MNLR models

U
within the depth interval of interest along Well A.
AN
5. Validating the models in the test well
Application of trained models to other wells drilled into the same field can identify whether
the trained models are generalizable to other wells. Accordingly, in this section, the trained
M

model at Well A was used to estimate geomechanical parameters of interest at Well B. Table 15
presents a comparison among MLP-COA, MLP-PSO, and MNLR models based on RMSE and
R-squared values obtained from the estimation of geomechanical parameters using the data from
D

Well B. As can be seen in this table, the two models built based on intelligent methods exhibit
higher accuracy than those of regression model. In addition, MLP-COA outperformed MLP-
TE

PSO in estimating all of the considered geomechanical parameters. As expected, the trained
models exhibited higher accuracy in estimating CCS, UCS, and ϕ rather than ν. As such, care
should be taken in using ν estimator model. Figure 17 compares curves of estimated values
EP

obtained from MLP-COA, MLP-PSO, and MNLR versus measured values of the considered
geomechanical parameters within the depth interval along Well B. Also in this figure, it is
clearly observed that, the models built based on intelligent methods tend to produce results
which are closer to the measured value.
C

Table 15. A comparison among trained MLP-COA, MLP-PSO, and MNLR models for estimating the considered geomechanical
AC

parameters at Well B.
Model Parameters R-Square RMSE
CCS 0.821 7.708

k
UCS 0.806 4.934
MLP-COA
l
0.817 1.475
0.617 0.0028
CCS 0.773 9.480

k
UCS 0.738 6.014
MLP-PSO
l
0.756 1.950
0.601 0.0030
CCS 0.650 11.786

k
UCS 0.610 7.360
MNLR
l
0.635 2.348
0.543 0.0035

24
ACCEPTED MANUSCRIPT

PT
RI
SC
Figure 17. A comparison between measured and estimated values of the considered geomechanical parameters using MLP-COA,
MLP-PSO, and MNLR models at Well B.

6. Conclusion
U
AN
In this paper, the concept of MSE was used to present a solution for estimating
geomechanical parameters across the formations overlying reservoir. To this end, an MLP neural
network combined with COA or PSO as well as MNLR method were used to estimate CCS,
M

UCS, ϕ and ν. This was based on the data from two oil wells in southwestern Iran.
A comparison between training and testing phases of the two models, namely MLP-COA and
MLP-PSO, when applied to estimate geomechanical parameters at Well A showed that, the
D

MLP-COA model is of higher accuracy and reliability than the other model. In addition, an
investigation on the error reduction rate through successive iterations of PSO and COA indicated
TE

higher convergence rate of COA, although processing time of this algorithm was longer than
PSO. A comparison between the two intelligent models and MNLR model a Well A showed that
the intelligent models were of higher accuracy. Validation of the trained models for estimating
geomechanical parameters at test well (i.e. Well B) revealed that the intelligent models tended to
EP

return lower error than the regression model. In addition, MLP-COA model had higher accuracy
than the other two models.
Results of adopting intelligent algorithms and the regression method for estimating
C

geomechanical parameters using the concept of MSE suggested that, this concept can serve as an
appropriate tool for estimating geomechanical parameters of the rock. In addition, an
AC

investigation on the accuracy of estimator models in estimating the considered geomechanical


parameters showed that, application of this concept is of higher accuracy and reliability in
estimating CCS, UCS, and ϕ parameters rather than ν. It can be stipulated firmly that, the
presented method in this paper can be generalized to estimate geomechanical parameters in
formations overlying the reservoir at other wells across the field under study provided further
data is available, particularly the data from the formations above the reservoir.
7. Data limitation
To the best of our knowledge, this is the first model that is developed for estimating
geomechanical properties of rock using MSE. Utilized for this study was the data from a vertical
well, which was measured at surface. Surface-measured data is associated with large deals of
uncertainty. As such, a pre-processing step was performed on the data to omit outliers as well as
the data pertaining to the depth intervals where drilling problems occurred. Although the pre-
processing of the data led to a high-accuracy model, but researchers are recommended to take

25
ACCEPTED MANUSCRIPT
much care when applying this methodology for the data measured at surface, and it is generally
better to use bottom-hole MSE data instead.

8. Acknowledgement
The authors express their gratitude toward National Iranian South Oil Company (N.I.S.O.C)
for their support and the information they provided.
9. Nomenclature
An Bit area (YE ) E Young module (GPa)
Dpqr

PT
Bottom hole differential pressure using
ECD Equivalent Circulating Density
Dn
Skempton relation (MPa)

Dp
Bit diameter (in) FR Flow Rate (gal/min)

Etuv
Bottom hole differential pressure (MPa) GR Gamma ray log (GAPI)

RI
Ew
Dynamic Young module (GPa) IQR Interquartile Range
Mechanical efficiency of new bit (ratio) LIF Lower Inner Fence
Exyz Static Young module (GPa) MLP Multi-Layer Perceptron
O {.|}
g(

SC
O ~ .• €
Mass conversion unit = 32.174 MNLR Multivariable Non-Linear Regression
GR wzƒ Maximum value of gamma ray log (GAPI) MSE Mechanical Specific Energy (psi)
GR w„v Minimum value of gamma ray log (GAPI) N The number of measured values
Vx

U
Shear wave velocity (km/s) NPHI Neutron porosity log (ratio)
V†‡zˆ‰ Shale/clay volume (V/V) OB Over Burden pressure
X5
AN
Coded factor for Depth PHIE Effective Porosity log (ratio)
X Coded factor for MSE PP Pore Pressure (MPa)
X: Coded factor for CT PSO Particle Swarm Optimization
Xb
The significance of a regression
M

Coded factor for FR P-Value


γn
coefficient
Inclination of the bottom hole (rad) RHOB Rock density (G/C3)
μn
Bit-specific coefficient of sliding friction
RMSE Root Mean Square Error
D

μx
(dimensionless)
Coefficient of friction of drill string ROP Rate of Penetration (ft/h)
νtuv
TE

Dynamic Poisson’s ratio RPM Rotary Speed (rpm)


νxyz Static Poisson’s ratio R-Square Coefficient of determination
75 First quartile Tor Torque (daN.m)
7:
T- The coefficient divided by its
Third quartile
EP

Statistics standard error



Uniaxial Compressive Strength
Compressional wave velocity (km/s) UCS
/0
(MPa)
The prediction location UIF Upper Inner Fence
+,
C

An unknown weight for the measured at the


WOB Weight on Bit (psi)
ith location
Ν
AC

Adjusted
Adjusted coefficient of determination Poisson’s ratio
ϕ
R-Square
BS Bit Size (in) Internal friction angle (deg)
CCS“” -(/, )
The measured value at the ith
location
CCS†•
Confined compressive strength using
Skempton relation (MPa)

10. Appendix
Elastic properties of rock, including Young’s modulus, Poisson’s ratio, shear modulus, and
bulk modulus can be obtained from three logs, namely density log and compressive and shear
wave velocities (Fjaer et al., 2008). In order to determine static elastic properties of rock from

26
ACCEPTED MANUSCRIPT
petrophysical logs, one should begin with calculating dynamic rock properties. Then, using
proper empirical relationships for the specific field under study, the static elastic properties can
be calculated from the dynamic properties. Equations (13) and (14) were used to calculate
dynamic Young’s modulus and Poisson’s ratio, respectively.

3MŽ − 4M•
= ˜ ×
–—3
MŽ − M•
(13)

MŽ − 2M•
l–—3 =

PT
2[MŽ − M• ^
(14)

RI
Given that numerous empirical relationships have been proposed for converting dynamic rock
moduli to static once for various rock types, determination of the proper relationship for the
specific field in question is one of the most important challenges in this scope. The choice of the

SC
proper empirical relationship requires a knowledge of actual static elastic moduli of the rock
from laboratory experiments on a number of samples taken from different depth points. On the
other hand, such experimental data can be devised to develop specific correlations for the
particular field in question. Equations (15) and (16) were obtained for determining static

U
Young’s modulus and Poisson’s ratio from their corresponding dynamic values based on
experimental data on core samples taken from wells drilled into the studied field (Anemangely et
AN
al., 2017c; Anemangely et al., 2018b).

•}™ = 0.7 –—3 (15)


M

l•}™ = l–—3 (16)


D

A number of relationships have been proposed for estimating uniaxial compressive strength
(UCS) of rock using well logs or elastic properties based on studies performed in different fields
TE

(Chang et al., 2006). In order to obtain a continuous profile of UCS, one shall use empirical
relationships developed using experimental data. Based on the experimental results obtained for
determining UCS, it was found that, the values of this parameter can be approximated with good
accuracy based on static Young’s modulus (Anemangely et al., 2017).
EP

>j = 2.27 •}™ + 4.74 (17)


C

Rare efforts have been made to find a relationship between internal friction angle and
geophysical characteristics, because even weakest rocks exhibit relatively high values of internal
AC

friction angle, and there is a relatively complex relationship between internal friction angle and
micromechanical properties of rock, such as rock hardness which is largely dependent on
cementation and rock porosity (Cholami et al., 2014). Nevertheless, experimental evidences
show that, shales with higher values of Young’s modulus commonly exhibit higher values of
internal friction angle (Lama and Vutukuri, 1978). Among various relationships for the
estimation of internal friction of rock, the one proposed by Plumb (1994) exhibited good
consistency with the conditions across the field (Equation (18)).

k = 26.5 − 37.4(1 − @ ˜6 − MšN™O› ) + 62.1(1 − @ ˜6 − MšN™O› ) (18)

` −`
MšN™O› =
,3
` ™œ − ` ,3
(19)

27
ACCEPTED MANUSCRIPT

Equation (20) is a widely accepted relationship in the scope of rock mechanics and has been
extensively used to calculate CCS of rock (Caicedo et al., 2005; Shi et al., 2015). This
relationship was used to estimate CCS along the studied wells in this research.

sin k
jj = >j + +2
•ž Ÿ Ÿ
1 − sin k
(20)

Ÿ = j − (21)

PT
In cases where the considered rock is impermeable and the well is vertical, CCS is calculated
from the relationship developed by Skempton (1984) (Equation (22); Calhoun and Ewy, 2009).

RI
sin k
jj = >j + +2
š£ Ÿ¤¥ Ÿ¤¥
1 − sin k
(22)

SC
− j
= j −( − )
Ÿ¤¥
3
(23)

U
In order to estimate CCS values at an acceptable accuracy, one should select appropriate
relationship for bottom-hole pressure difference based on effective rock porosity. On this basis,
AN
Table 16 was used to select appropriate relationship for estimation of CCS based on effective
rock porosity.
M

Table 16. Selection of appropriate relationship for estimation of CCS based on effective rock porosity.

jj = jj •ž
Condition Relationship

( ˜6 − 0.05) (0.20 − ˜6 )
If porosity is greater than 0.2

jj = jj + jj
D

•ž
0.15 š£
0.15
If porosity is between 0.2 and 0.5
jj = jj š£
TE

If porosity is smaller than 0.05

Figures 18 and 19 show pore pressure and ECD pressure within the considered depth interval
along Wells A and B, respectively.
C EP
AC

28
ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
Figure 18. Variation profiles of pore pressure and equivalent current density (ECD) pressure within the studied depth interval
along Well A.
M
D
TE
C EP
AC

Figure 18. Variation profiles of pore pressure and equivalent current density (ECD) pressure within the studied depth interval
along Well B.

11. References
Al-Sudani, J.A., 2017. Real-time monitoring of mechanical specific energy and bit wear using

29
ACCEPTED MANUSCRIPT
control engineering systems. J. Pet. Sci. Eng. 149, 171–182.
Ali, J.K., 1994. Neural networks: a new tool for the petroleum industry? In: European Petroleum
Computer Conference. Society of Petroleum Engineers.
Anemangely, M., Ramezanzadeh, A., Tokhmechi, B., 2017a. Shear wave travel time estimation
from petrophysical logs using ANFIS-PSO algorithm: A case study from Ab-Teymour
Oilfield. J. Nat. Gas Sci. Eng. 38, 373–387.
Anemangely, M., Ramezanzadeh, A., Tokhmechi, B., 2017b. Determination of Constant

PT
Coefficients of Bourgoyne and Young drilling rate Model Using a Novel Evolutionary
Algorithm. J. Min. Environ.

RI
Anemangely, M., Ramezanzadeh, A., Tokhmechi, B., 2017c. Safe Mud weight window
determination using log based methodology. In: 79th EAGE Conference and Exhibition
2017.

SC
Anemangely, M., Ramezanzadeh, A., Tokhmechi, B., Molaghab, A., Mohammadian, A., 2018a.
Development of a new rock drillability index for oil and gas reservoir rocks using punch
penetration test. J. Pet. Sci. Eng. 166.

U
Anemangely, M., Ramezanzadeh, A., Tokhmechi, B., Molaghab, A., Mohammadian, A., 2018b.
AN
Drilling rate prediction from petrophysical logs and mud logging data using an optimized
multilayer perceptron neural network. J. Geophys. Eng. 15.
Anemangely, M., Ramezanzadeh, A., Amiri, H. and Hoseinpour, S.A., 2018. Machine learning
M

technique for the prediction of shear wave velocity using petrophysical logs. J. Pet. Sci.
Eng. 174.
Ashrafi, S.B., Anemangely, M., Sabah, M. and Ameri, M.J., 2018. Application of hybrid
D

artificial neural networks for predicting rate of penetration (ROP): A case study from
Marun oil field. J. Pet. Sci. Eng (In press).
TE

Bishop, C., 2007. Pattern Recognition and Machine Learning (Information Science and
Statistics), 1st edn. 2006. corr. 2nd printing edn. Springer, New York.
EP

Caicedo, H.U., Calhoun, W.M., Ewy, R.T., 2005. Unique ROP predictor using bit-specific
coefficient of sliding friction and mechanical efficiency as a function of confined
compressive strength impacts drilling performance. In: SPE/IADC Drilling Conference.
C

Society of Petroleum Engineers.


AC

Calhoun, W.M., Ewy, R.T., 2009. Method for estimating confined compressive strength for rock
formations utilizing skempton theory.
Chang, C., Zoback, M.D., Khaksar, A., 2006. Empirical relations between rock strength and
physical properties in sedimentary rocks. J. Pet. Sci. Eng. 51, 223–237.
Chatterjee, R., Mukhopadhyay, M., 2002a. Effects of rock mechanical properties on local stress
field of the Mahanadi basin, India—results from finite element modelling. Geophys. Res.
Lett. 29.
Chatterjee, R., Mukhopadhyay, M., 2002b. Petrophysical and geomechanical properties of rocks
from the oilfields of the Krishna-Godavari and Cauvery Basins, India. Bull. Eng. Geol.
Environ. 61, 169–178.

30
ACCEPTED MANUSCRIPT
Chen, X., Fan, H., Guo, B., Gao, D., Wei, H., Ye, Z., 2014. Real-Time Prediction and
Optimization of Drilling Performance Based on a New Mechanical Specific Energy Model
8221–8231.
Chen, X., Gao, D., Guo, B., Feng, Y., 2016. Real-time optimization of drilling parameters based
on mechanical specific energy for rotating drilling with positive displacement motor in the
hard formation. J. Nat. Gas Sci. Eng. 35, 686–694.
Coello, C. a C., Lamont, G.B., Veldhuizen, D. a. Van, 2007. Evolutionary Algorithms for
Solving Multi-Objective Problems, second. ed. Springer Science & Business Media.

PT
Das, B., Chatterjee, R., 2017a. Mapping of pore pressure, in-situ stress and brittleness in
unconventional shale reservoir of Krishna-Godavari basin. J. Nat. Gas Sci. Eng.

RI
Das, B., Chatterjee, R., 2017b. Wellbore stability analysis and prediction of minimum mud
weight for few wells in Krishna-Godavari Basin, India. Int. J. Rock Mech. Min. Sci. 93,

SC
30–37.
Das, B., Chatterjee, R., 2018. Well log data analysis for lithology and fluid identification in
Krishna-Godavari Basin, India. Arab. J. Geosci. 11, 231.

U
Dupriest, F.E., Koederitz, W.L., 2005. Maximizing drill rates with real-time surveillance of
AN
mechanical specific energy. In: SPE/IADC Drilling Conference. Society of Petroleum
Engineers.
Dupriest, F.E., Koederitz, W.L., Totco, M.D., Company, V., 2005. SPE / IADC 92194
M

Maximizing Drill Rates with Real-Time Surveillance of Mechanical Specific Energy.


Fjaer, E., Holt, R.M., Horsrud, P., Raaen, a M., Risnes, R., 2008. Petroleum Related Rock
Mechanics, Marine Environmental Research. Elsevier.
D

Garćia, L.P.F., de Carvalho, A.C., Lorena, A.C., 2013. Noisy data set identification. In:
TE

International Conference on Hybrid Artificial Intelligence Systems. Springer, pp. 629–638.


Gholami, R., Moradzadeh, A., Rasouli, V., Hanachi, J., 2014. Practical application of failure
criteria in determining safe mud weight windows in drilling operations. J. Rock Mech.
EP

Geotech. Eng. 6, 13–25.


Kawam, A.A.L., Mansour, N., 2012. Metaheuristic optimization algorithms for training artificial
neural networks. Int. J. Comput. Inform. Technol 1, 156–161.
C

Lama, R.D., Vutukuri, V.S., 1978. Handbook on mechanical properties of rocks-testing


AC

techniques and results-volume iii.


Lorena, A.C., de Carvalho, A.C., 2004. Evaluation of noise reduction techniques in the splice
junction recognition problem. Genet. Mol. Biol. 27, 665–672.
Maimon, O., Rokach, L., 2005. Data mining and knowledge discovery handbook, second. ed,
Data Mining and Knowledge Discovery Handbook. Springer.
Maletic, J.I., Marcus, A., 2000. Data Cleansing: Beyond Integrity Analysis. In: IQ. Citeseer, pp.
200–209.
Pedersen, M.E.H., Chipperfield, A.J., 2010. Simplifying particle swarm optimization. Appl. Soft
Comput. 10, 618–628.

31
ACCEPTED MANUSCRIPT
Plumb, R.A., 1994. Influence of composition and texture on the failure properties of clastic
rocks. In: Rock Mechanics in Petroleum Engineering. Society of Petroleum Engineers.
Quinlan, J.R., 1986. The effect of noise on concept learning. Mach. Learn. An Artif. Intell.
approach 2, 149–166.
Rajabioun, R., 2011. Cuckoo optimization algorithm. Appl. Soft Comput. 11, 5508–5518.
Shi, X., Meng, Y., Li, G., Li, J., Tao, Z., Wei, S., 2015. Confined compressive strength model of
rock for drilling optimization. Petroleum 1, 40–45.

PT
Singha, D.K., Chatterjee, R., 2015. Geomechanical modeling using finite element method for
prediction of in-situ stress in Krishna–Godavari basin, India. Int. J. Rock Mech. Min. Sci.

RI
73, 15–27.
Skempton, A.W., 1984. The pore-pressure coefficients A and B. In: SELECTED PAPERS ON
SOIL MECHANICS. Thomas Telford Publishing, pp. 65–69.

SC
Wang, R.Y., Storey, V.C., Firth, C.P., 1995. A framework for analysis of data quality research.
IEEE Trans. Knowl. Data Eng. 7, 623–640.

U
Wu, X., 1995. Knowledge acquisition from databases. Intellect books.
AN
Yang, X.-S., Deb, S., 2009. Cuckoo search via Lévy flights. In: Nature & Biologically Inspired
Computing, 2009. NaBIC 2009. World Congress on. IEEE, pp. 210–214.
M
D
TE
C EP
AC

32
ACCEPTED MANUSCRIPT

Highlights

Mechanical Specific Energy (MSE) concept in drilling was used for determination of
geomechanical parameters.
Data was acquired from two wells in southwest of Iran.
Multi-Layer Perceptron (MLP) neural network with Cuckoo Optimization Algorithm (COA) and
Particle Swarm Optimization (PSO) as training algorithms was used for prediction of Confined

PT
Compressive Strength (CCS) Uniaxial Compressive Strength (UCS), friction angle ( ) and
Poisson’s ratio ( ).
Built models showed that they are more capable in prediction of CCS, UCS and than prediction

RI
of .
MLP-COA model is superior to MLP-PSO and multivariate non-linear regression models.

U SC
AN
M
D
TE
C EP
AC

You might also like