DC175238 DOI: 10.
2118/175238-PA Date: 6-March-17 Stage: Page: 1
Intelligent Tool To Design Drilling, Spacer,
Cement Slurry, and Fracturing Fluids by
Use of Machine-Learning Algorithms
Arash Shadravan, ReservoirFocus; Mohammadali Tarrahi, Shell; and
Mahmood Amani, Texas A&M University at Qatar
Summary
Design of drilling fluids, spacers, cement slurries, and fracturing
fluids is often done by trial and error in the laboratory. In the first
step, the required properties of these fluids are categorized and
then efforts will be started with a rough idea of the optimal com-
position. This first guess usually depends on the experience of the
laboratory analyst or fluid engineer. Afterward, the trial-and-error
testing starts, and it continues until the fluid design moves closer
to the desired fluid criteria. There are several test data that would
not be used in this method, and it is difficult to digest a large
amount of information by the user. Trial and error could be time-
consuming, very costly, and misleading. Today, there is a need
for an intelligent system that uses all the available data (big data),
even if the data sets are not close to the desired goal, and offers
insights for fluid designs.
This paper conducted a study on the application of machine-
learning-based methodologies, including Gaussian-process regres-
sion (GPR) and artificial neural networks (ANNs), to reduce the
costs of testing, integrate available experimental data, and elimi-
nate the need for personnel supervision. These practical nonlin-
ear-regression methods empower efficient and fast prediction
tools that do not require including complex physics of the under-
lying system while integrating all available data from different
sources. GPR, which is also known as Kriging in geostatistics lit-
erature, has exceptional advantages over traditional regression
methods because it does not require a known form for regression
function and also has the capability of determining the estimation
error and the confidence interval. This machine-learning-based
tool offers insights for intelligent fluid design and could
reduce costs.
Introduction
Machine-learning methodologies and expert systems have proved
to be conducive to digesting big data sets and making intelligent
decisions for agile productivity and more-accurate measures. These
algorithms have been successfully applied to petroleum-
engineering fields such as production forecasting, history matching,
reservoir characterization, and enhanced-oil-recovery screening.
The proposed work flow in this paper includes the following tools.
Data-Preparation Tool. The tool prepares the training data
through guided experiments.
• Quality assurance/quality control for the rheological-proper-
ties experiment
* Calibrate the equipment according to the user manual
* Operate and assemble the equipment according to the
user manual
n Maintain 150 rev/min for fluid conditioning
n Maintain and record the data during the experiment in
one mode (manual or automatic)
Copyright V
C 2017 Society of Petroleum Engineers
This paper (SPE 175238) was accepted for presentation at the SPE Kuwait Oil and Gas
Show and Conference, Mishref, Kuwait, 11–14 October 2015, and revised for publication.
Original manuscript received for review 24 August 2015. Revised manuscript received for
review 19 December 2016. Paper peer approved 10 January 2017.
2017 SPE Drilling & Completion
ID: jaganm Time: 15:57 I Path: S:/DC##/Vol00000/170004/Comp/APPFile/SA-DC##170004
Total Pages: 10
n Repeat experiments with more than 10% dial-reading
fluctuation; inspect the jewel, pivot, rotor, and belt, and
then recalibrate
n Repeat experiments with a modified fluid formulation if
fluid-particulate settling is observed
• Determining the most sensitive and effective design parame-
ters (such as temperature and density)
• Design of experiment (DOE) or Latin hypercube sampling
to produce the most-representative and wide-enough sample
cases (Shadravan et al. 2015a, 2015b).
Prediction Tool. This is an intelligent data-driven prediction
tool.
• Build a reliable prediction tool (regressor) through machine-
learning algorithms
• Predict (and assess the uncertainty of) any new samples’
properties (e.g., rheological properties)
Design Tool. This is an intelligent fluid-design tool.
• Designing the desired fluid system depending on the
requested sample properties
• Performing the opposite of prediction tool and building a
regressor to predict corresponding parameters of a suitable
fluid system (through machine-learning algorithms)
• Applying inverse problem methods to prediction tool and
obtaining the desired fluid system
In the machine-learning literature, some of the most-well-
known regression methods include linear regression (ordinary
least squares), nonlinear regression or curve fitting (e.g., polyno-
mial regression), logistic regression, ridge regression, ANN,
Bayesian network, radial-basis-function network, support-vector
machine, GPR, and kernel regression (Theodoridis et al. 2010;
Duda et al. 2012). In this paper, we choose to investigate the appli-
cation of the ANN and GPR algorithms to fluid-characteristics
estimation and fluid design and build an intelligent multipurpose
prediction and design tool. Fig. 1 lists various types of fluid-design
experiments that usually are pursued by a trial-and-error method.
There have been some studies regarding the applications of
ANN in drilling and fluid design (Osman and Aggour 2003; Gidh
et al. 2012). Shadravan and Amani (2012) and Amani et al.
(2012) studied the rheological properties on various drilling fluids
under extreme high-pressure/high-temperature (HP/HT) condi-
tions. Lee et al. (2012) investigated how various HP/HT rheome-
ters could produce different results under HP/HT conditions,
assuming each data set is repeatable and each rheometer is well-
calibrated. In this study, one HP/HT rheometer was used to pre-
vent exposing the data sets to such errors. The complexity and
variations of downhole conditions, geology, additives (polymers,
cement types, weighting agents, crosslinker, or breaker), testing
procedures, and experiment schedule could all be taken into
account properly to construct a useful laboratory database for an
intelligent fluid-design tool. The “mixing” water quality should
also not be ignored. Proper equipment calibration and competent
laboratory personnel are valuable elements of successful fluid
design. Fig. 2 shows traditional trial-and-error fluid design (Bland
et al. 1993; Diarra et al. 2016), and Fig. 3 demonstrates the intelli-
gent machine-learning-fluid-design work flow.
1
 DC175238 DOI: 10.2118/175238-PA Date: 6-March-17 Stage: Page: 2 Total Pages: 10

Fluid design

Drilling fluids Spacers Cement Fracturing fluids

Rheology Rheology Rheology
Fluid loss and free fluid Rheology
Gel strength Stability
Thickening time Turbidity
Fluid loss Compatibility
Dynamic setting Apparent density
Resistivity Contamination
Mechanical properties Acid solubility
Lubricity Spacer/surfactant screening
Barite sag Annular-seal durability
Interaction analysis Thermal expansion
Gel strength
Heat of hydration
Permeability

Fig. 1—Applications of the intelligent design tool for various fluid-design experiments.

to predict the rheological properties of a fluid as a simple example

Trial-and-
of the breadth of the application of machine-learning methods.
Desired
error fluid Fluid-design Laboratory
fluid Predicting such properties can be viewed as a regression analysis
design criteria tests (a data-driven approach), where the goal is to design the best lin-
design
ear- or nonlinear-regression function (or rule).
In this study, we propose to solve this problem with machine-
Fig. 2—Traditional arbitrary and trial-and-error-based fluid- learning methods that are well-established in computer-science
design procedure. literature. The proposed algorithms eliminate an arbitrary
approach in making decisions, and provide accuracy and fast
computation. This approach substitutes for the need for days and
Machine-Learning Methods for weeks of testing in the laboratory with an intelligent and agile
Nonlinear Regression fluid design that could require only a few tests to confirm the pro-
Investigation of the rheological properties of fracturing, drilling, posed compositions (additives) for the desired properties.
spacer, and cement-slurry fluids is conducted on a daily basis. In the parametric-regression methods, a known regression
This paper delineates the benefits of machine-learning algorithms function (such as polynomial or exponential) is defined in terms

Data Detecting DOE Most-representative

Preparation most-effective experimental
Latin hypercube
parameters samples
sampling

Fluid Fluid characteristics

Intelligent properties and Machine-learning (e.g., rheology),
data-driven experiment algorithm associated uncertainty
prediction condition
tool

Machine Confirm Desired

Fluid-design learning fluid
laboratior
criteria algorithm design
test
Intelligent
fluid-design
tool
Intelligent
data-driven Fluid characteristics
Candidate-fluid
prediction (e.g., rheology)
properties
tool

Inverse
Modeling

Fig. 3—Schematic of the proposed intelligent integrated tools for guided experiment design, characteristic prediction, and fluid-
system design.

2 2017 SPE Drilling & Completion

ID: jaganm Time: 15:57 I Path: S:/DC##/Vol00000/170004/Comp/APPFile/SA-DC##170004

 DC175238 DOI: 10.2118/175238-PA Date: 6-March-17
Table 1—Pros and cons in use of machine learning.
of a finite number of unknown parameters that are estimated from
the training data. In general, because the nature of the underlying
physical process (e.g., relation of rheological properties and fluid
characteristics) is very complex, assuming a known and closed
form for the regression function is a rigorous task. Therefore, non-
parametric-regression methods are proposed that have no (or very
little) a priori knowledge of the form of the function that is being
estimated. These methods allow for the class of functions that the
model can represent to be very broad. Here we implement and
apply ANNs and GPR, which are among the most-practical non-
parametric-regression methods (Bishop 2006; Theodoridis et al.
2010; Duda et al. 2012).
Recently data-driven machine learning, initially a field of com-
puter science, has become a widespread and powerful tool for pre-
diction, decision making, and data analysis (Schakel and Mesdag
2014; Mijnarends et al. 2015; Cao et al. 2016). Machine-learning
algorithms function by constructing a mathematical model from
example or training inputs (e.g., laboratory measurements) to per-
form data-driven predictions or decisions. Machine-learning
methods are successfully used in various applications such as bio-
informatics, medical diagnosis, computational finance, marketing,
online advertising, and computer vision (Bishop 2006; Rasmussen
and Williams 2006; Theodoridis et al. 2010).
Cement and Spacer Design
To accomplish long-term zonal isolation, several factors, includ-
ing the cement-spacer/drilling-fluid rheological properties and
friction-pressure hierarchy, must be carefully engineered. Design-
ing the cement and spacer fluids with the desired rheological
properties traditionally is achieved by a trial-and-error approach
given the complications and high costs of performing pressurized
(especially cement) rheology tests at high temperatures (greater
than 220 F). Shadravan et al. (2016) showed three case studies
where well integrity was maintained, benefiting from the applica-
tion of DOE in designing spacer fluids. The proposed model
worked well for specific formulations, but changes in some of the
ingredients or the range of their concentrations would necessitate
developing another DOE model, which could be time-consuming.
Limited inventories of the ingredients for the fluid design or their
cost could encourage operating companies or service providers to
use different ingredients, therefore making DOE models futile. In
fact, DOE models lack dynamic learning and progression, and the
pursuit of a more-comprehensive approach is vital.
Shadravan and Tarrahi (2016) criticized the trial-and-error
fluid-design approach and discussed what is missing in stepping
toward intelligent fluid design. They showed how machine learn-
ing could save hours of laboratory testing, integrate sizable exper-
imental databases, and optimize the operational cost and delivery.
Tarrahi and Shadravan (2016a, 2016b) and Shadravan et al.
(2016) demonstrated the application of machine learning and
inverse-modeling theory to design a spacer fluid with desired
rheological properties for optimal cement/spacer/drilling fluids.
The use of machine learning as opposed to DOE can yield
improved zonal isolation, but it requires establishing appropriate
infrastructures to record and quality control the data. It is evident
that there are pros and cons in use of machine learning as well.
Table 1 illustrates some of the positive and negative attributes of
machine-learning algorithms.
2017 SPE Drilling & Completion
ID: jaganm Time: 15:57 I Path: S:/DC##/Vol00000/170004/Comp/APPFile/SA-DC##170004
Stage: Page: 3 Total Pages: 10
ANNs
ANN models are nonparametric-regression methods that have
been used progressively in different areas of science (Rafieisakhaei
et al. 2016a, 2016b). Although analyzing data by use of ANNs is
usually more complex than traditional regression approaches,
ANN models are more flexible and efficient when the main goal is
estimation of dependent output values by use of different explana-
tory independent variables (Theodoridis et al. 2010; Duda et al.
2012; Rafieisakhaei et al. 2016a, 2016b).
In our current experiment, there are three input parameters
(fluid density, Ingredient A content, and temperature) and seven
output values (300-, 200-, 100-, 60-, 30-, 6-, and 3-rev/min viscos-
ities), and we have nine data points (training samples)
Training data: fXi ; Yi g; i ¼ 1; 2; …; N; . . . . . . . . . . . ð1Þ
Xi ¼ ith input vector ðn  1Þ; . . . . . . . . . . . . . . . . . . . ð2Þ
Yi ¼ ith output vectorðm  1Þ; . . . . . . . . . . . . . . . . . . ð3Þ
where N represents the number of experiments or the number of
data points. In this study, n, m, and N are 3, 7 and 9, respectively.
Our ANN in the input and output layers contains four (three inputs
plus bias term) and seven neurons, respectively. We also consider
a hidden layer, and its number of neurons will be determined by
cross validation. The most commonly used ANN is multiple-layers
perceptron, where the learning (finding network weights) is per-
formed dependent on minimization of the mean square error of the
output and by use of a backpropagation-learning algorithm.
As a preprocessing step, it is required to scale input and output
variables to have them in a similar range. Input parameters are
transformed to be standard normal variables (zero mean and unit
variance), and output values are shifted and scaled to a [–1,1]
interval to be consistent with the range of the activation function
(tanh function). The original data transformation and normaliza-
tion is performed in the following way:
XX
Provided the training data Xsn ¼ ;
rX
2ðY  Ymin Þ
Ys ¼  1; . . . . . . . . . . . . . . . . . . . . . . . . ð4Þ
Ymax  Ymin
where X and rX represent input mean and standard deviation
(SD), respectively; Xsn is the standard normal input that is being
used in the machine-learning algorithms; Ymin and Ymax are mini-
mum and maximum of the output or target values, respectively;
and Ys is the scaled output that will be used to train the network.
After ANN learning and finding the network weights W
and U, now we can estimate the output for a new set of inputs
(where the activation function is tanh and the hidden layer has
q neurons):
( )!
 1 
Y0 ¼ tanh U T 1 ; . . . . . . . . . . . . ð5Þ
tanh W T
X0
W ¼ Input layer to hidden-layer weights ½ðn þ 1Þ  q;
                   ð6Þ
3
 DC175238 DOI: 10.2118/175238-PA Date: 6-March-17
300 rev/min
70
60
Viscosity (cp)
50
40
30
20
10
0 400
5 20
4 20
3 200
2 15
10
1 100 0
0 5
Ingredient A (gal/bbl) Fluid Density (lbm/gal)
(a)
Fig. 4—Nonlinear estimator resulting from ANN (pink asterisks show the training data).
and
U ¼ Hidden layer to output layer wights ½ðq þ 1Þ  m:
                   ð7Þ
The activation function at neurons is chosen to be a hyperbolic
tangent function. We used the backpropagation method in sequen-
tial mode to train the network and obtain the weights W and U
matrices (Bishop 2006; Theodoridis et al. 2010). To start the
learning process, we initialize the weights as random numbers.
Estimation error for the training data is not a promising perform-
ance measure for ANN because it may raise the overfitting issue,
where the constructed regression tool reproduces the training
data almost perfectly but fails to properly estimate the new data
(Hawkins 2004). The constructed regression tool should fit the
training data properly and at the same time should have the gener-
alization power to predict any unknown and new cases. Therefore,
to find the optimal ANN, we should split the data to the new train-
ing and test data; design ANN by use of training data; and per-
form cross validation on the test data. To obtain the best ANN
estimator, we tuned the ANN parameters (maximum error and
maximum iteration in the backpropagation-learning algorithm and
the number of hidden layers) through leave-one-out cross valida-
tion (LOOCV). Each time we leave out one aspect of the training
data and design ANN with the remaining data. We repeat this pro-
cedure N times (N is number of training data). After tuning, the
best-obtained average relative cross-validation error is 26% and
its corresponding training data error is 13%. Depending on
LOOCV, the optimum number of hidden-layer neurons is six.
For simplicity, we plotted the viscosity with respect to only two
input parameters and kept the third one constant. Fig. 4a shows the
300-rev/min viscosity with respect to fluid density and Ingredient
A content while temperature is constant (80 F). Fig. 4b illustrates
the 60-rev/min viscosity with respect to temperature and Ingredient
A content while fluid density is constant (12 lbm/gal). The plotted
surface is the nonlinear regression resulting from ANN. With this
nonlinear-regression function, we can predict the viscosity of the
new sample without performing extra laboratory experiments.
Gaussian Process Regression. GPR is a nonlinear-regression or
interpolation technique that models the new estimated (interpolated)
values derived from the Gaussian process determined by a covari-
ance function (Williams and Rasmussen 2006). GPR is a statistical
machine-learning method that is also known as Kriging and is well-
established in geostatistics and computer-science literature (Deutsch
and Journel 1992; Shi and Choi 2011). To estimate the correspond-
ing value for the new input, GPR calculates the weighted average
(linear combination) of the known values (training data) depending
on the correlation (or proximity) of the new input and the training
data governed by the covariance function (Rasmussen and Williams
2006). Providing the training data and the prior assumptions (e.g.,
4 2017 SPE Drilling & Completion
ID: jaganm Time: 15:57 I Path: S:/DC##/Vol00000/170004/Comp/APPFile/SA-DC##170004
Stage: Page: 4 Total Pages: 10
60 rev/min at Constant Fluid Density
55
50 40 30
45 30 25
Viscosity (cp)
40
20
35 20
30 10
15
25 0
5 10
300 4
3
15 2
1 5
Temperature (°F) Ingredient A (gal/bbl)
(b)
covariance function) are suitable (Deutsch and Journel 1992), GPR
or Kriging is the best linear unbiased estimator. In geostatistics lit-
erature, the spatial dependencies are represented by a variaogram
function instead of a covariance function. In signal-processing lit-
erature, this estimation procedure is known as Kalman filtering,
which has been also applied to history-matching problems (Tarrahi
et al. 2013, 2015). GPR by nature does not need any a priori
assumption on the form of regression functions. It is a purely data-
driven methodology that extracts the relationship between varia-
bles from the information provided by the training data.
It is certainly the inherent correlation and relationship of the
parameters and measured properties that guide or train the regres-
sion tool to follow a specific trend. Data-driven machine-learning
methods are trained by a set of measurements (training samples).
Therefore, if there is no correlation or dependency between a set
of parameters and the measured outputs, then the machine-learn-
ing tool cannot extract any meaningful relationship. Consequently,
the designed regression tool will be insensitive to that parameter,
so it is recommended to remove the independent or less-effective
parameters from the parameter set. Sensitivity analysis is a proce-
dure that can be used to identify the most-effective parameters.
GPR has been applied in a variety of applications, such as environ-
mental science, geoscience, mining, and remote sensing, where it is
also known as Kriging (Journel and Huijbregts 1978). Other fields of
GPR application are hydrology, neuroscience, real-estate appraisal,
and image processing. In geostatistics, a subfield of geoscience and
mining, GPR or Kriging is used to populate the rock-property meas-
urements (that is, from core at the wells), such as permeability and
porosity, over the whole subsurface resource or reservoir extent.
It should be noted that the underlying physical process of the
system (i.e., the governing equations, such as the diffusion equa-
tion in case of fluid flow in porous media) is not incorporated
directly into the GPR formulation. However, GPR infers the inter-
relationship of the parameters from the provided training data set.
In addition, GPR as a data-driven method requires a sample data
set in advance. A covariance model that shapes the correlation
between parameters is also needed to be determined beforehand,
which depends on the physical system under investigation. The
form of the covariance function is usually predetermined, such as
an exponential, spherical, or Gaussian function (Journel and
Huijbregts 1978), and its parameters can be inferred from the
given data set in Appendix A.
Methodology
In this study, the fluid was weighted up by adding particulates and
not salt. Because there are three input parameters and plotting
each of the dial readings with respect to three independent input
parameters is not very informative, for illustration purposes, we
plotted the dial reading with respect to only two input parameters
and kept the third one constant. In Figs. 5 and 6, all the output
 DC175238 DOI: 10.2118/175238-PA Date: 6-March-17 Stage: Page: 5 Total Pages: 10

3 rev/min at Constant T (°F) 3 rev/min at Constant Fluid Density

18
18 16
16 20
20 14
14
12

Viscosity (cp)
12 15

Viscosity (cp)
15 10
10
10 8
10 8
6
6
5 4
5 4
2 2
0 0
5 400
4 20 300 5
3 4
2 15 200 3
10 2
1 1
0 5 100 0
Ingredient A (gal/bbl) Fluid Density (lbm/gal) Temperature (°F) Ingredient A (gal/bbl)
(a) (b)

6 rev/min at Constant T (°F) 6 rev/min at Constant Fluid Density

22 22
20 20
18 18
25 25
16 16
20 14 14
Viscosity (cp)

Viscosity (cp)
20
12 12
15 10 15
10
10 8 8
10
6
6
5 4 5
4
2 0
0
5 400
4 20 5
3 300 4
2 15 3
200 2
1 10 1
100 0
Ingredient A (gal/bbl) 0 5 Fluid Density (lbm/gal) Temperature (°F) Ingredient A (gal/bbl)
(c) (d)

30 rev/min at Constant T (°F) 30 rev/min at Constant Fluid Density

30
30
25
25 35
35
30 30 20
20
Viscosity (cp)

25 25
Viscosity (cp)

20 15 20 15
15 15
10 10
10 10
5 5 5 5
0 0
5 400
4 20
3 300 5
15 4
2 3
1 10 200 2
1
Ingredient A (gal/bbl) 0 5 Fluid Density (lbm/gal) 100 0
Temperature (°F) Ingredient A (gal/bbl)
(e) (f)
60 rev/min at Constant T (°F) 60 rev/min at Constant Fluid Density
35
30 35
40
35 25 40 30
Viscosity (cp)

Viscosity (cp)

30 20 30 25
25
20 15
20 20
15 10
10 10 15
5 5
0 0
400 10
5
4 20 300 5
3 4 5
15 200 3
2 2
1 10 1
100 0
Ingredient A (gal/bbl) 0 5 Fluid Density (lbm/gal) Temperature (°F) Ingredient A (gal/bbl)
(g) (h)

Fig. 5—Nonlinear estimator resulting from GPR (pink asterisks show the training data).

2017 SPE Drilling & Completion 5

ID: jaganm Time: 15:57 I Path: S:/DC##/Vol00000/170004/Comp/APPFile/SA-DC##170004

 DC175238 DOI: 10.2118/175238-PA Date: 6-March-17 Stage: Page: 6 Total Pages: 10

100 rev/min at Constant T (°F) 100 rev/min at Constant Fluid Density

45 45
40 40
50 35 50 35
30 30
40 40

Viscosity (cp)
Viscosity (cp)

25 25
30 20 30 20
20 15 20 15
10 10
10 10
5 5
0 0
5 400
4 20 300 5
3 4
2 15 200 3
10 2
1 1
100 0
Ingredient A (gal/bbl) 0 5 Fluid Density (lbm/gal) Temperature (°F) Ingredient A (gal/bbl)
(a) (b)

200 rev/min at Constant T (°F) 200 rev/min at Constant Fluid Density 55

55 50
50 45
45 60
60 40
40 50 35
50 35

Viscosity (cp)
Viscosity (cp)

40 30
40 30
30 25
25
30 20
20 20
20 15 15
10 10
10 10
0 0
5 400
4 20 300 5
3 4
2 15 3
200 2
1 10 1
100 0
Ingredient A (gal/bbl) 0 5 Fluid Density (lbm/gal) Temperature (°F) Ingredient A (gal/bbl)
(c) (d)

300 rev/min at Constant T (°F) 300 rev/min at Constant Fluid Density

60

70 60 70 50
60 60
Viscosity (cp)

40
Viscosity (cp)

50 50 50
40 40 30
30 40 30
20 20 20
10 30 10
0 10
0
5 400
4 20 20
3 300 5
15 4
2 3
1 10 200 2
10 1
0 5 100 0
Ingredient A (gal/bbl) Fluid Density (lbm/gal) Temperature (°F) Ingredient A (gal/bbl)
(e) (f)

Fig. 6—The predictor function to estimate viscosity resulting from GPR method (pink asterisks show the experimental data
points).

values are plotted vs. two pairs of the input parameters. For
instance, Fig. 6e shows the 300-rev/min viscosity with respect to
fluid density and Ingredient A (polymer) content while tempera-
ture is constant ( F). The pressure was held constant at 7,500 psi.
Pressure affects the rheological properties of oil-based fluids more
than water-based fluids and its effects should not be ignored, espe-
cially across abnormally pressured zones. The plotted surface is
the nonlinear regression resulting from GPR, and, as shown, it
goes through the training data (pink asterisks). With this nonlin-
ear-regression function, we can predict the viscosity of the new
sample without performing laboratory experiments. Fig. 7 also
shows how the predicted viscosity changes vs. Ingredient A at
constant rev/min, temperature, and fluid density.
In addition, to show the estimation error or confidence interval
of estimation, in Fig. 8 the nonlinear-regression function or non-
linear estimator of 60-rev/min viscosity vs. Ingredient A content
(fluid density and temperature are constant) is plotted along with
the associated confidence interval of 68%, corresponding to the
estimated value 6 SD.
Conclusions
Trial-and-error fluid design is not always a cost-effective option
and lacks long-term vision to accumulate knowledge on fluid sys-
tems. There is a need for a tool that is empowered by machine-
learning methodologies. This paper investigated an intelligent
tool that is equipped with flexible machine-learning algorithms.

6 2017 SPE Drilling & Completion

ID: jaganm Time: 15:57 I Path: S:/DC##/Vol00000/170004/Comp/APPFile/SA-DC##170004

 DC175238 DOI: 10.2118/175238-PA Date: 6-March-17 Stage: Page: 7 Total Pages: 10

At 100 rev/min (Constant Fluid Density At 60 rev/min (Constant Fluid Density

and Temperature) and Constant Temperature)
50 50

45 45

40 40

35 35

Viscosity (cp)
Viscosity (cp)

30 30

25 25

20 20

15 15

10 10

5 5

0 0
0 1 2 3 4 5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Ingredient A (gal/bbl)
Ingredient A (gal/bbl)
Fig. 8—Nonlinear-regression function resulting from GPR and
Fig. 7—2D plot of viscosity vs. Ingredient A at constant temper-
the corresponding confidence interval (the highlighted enve-
ature (2008F) and fluid density (12 lbm/gal) for a fixed rev/min.
lope around the nonlinear-regression curve shows the estima-
tion error), with temperature of 2008F and fluid density of 12
lbm/gal.

ANN is a fast and wide-ranging regression and prediction

approach that can be trained and built in real time as the training
data become available. This study also introduced GPR as a reli-
able regression tool that has shown promising applicability in dif-
ferent computational-fluid designs. The presented results, which
are derived from driven training data from experiments, demon-
strate the advantage of GPR over ANN as one of the most-popular
machine-learning methods. GPR, unlike ANN, honors the training
data (i.e., GPR-estimated output is exactly equal to the training-
sample output, in contrast to the ANN-estimated output). GPR
also offers a prior-free-regression implementation (i.e., it does not
need any a priori assumption on the form of regression functions).
Furthermore, GPR is unique in the sense of providing the estima-
tion error and uncertainty assessment while predicting the fluid
characteristics and designing a fluid system. The accuracy of esti-
mated output is a byproduct of the GPR method, unlike with
ANN. The developed intelligent prediction and design tool can
extract and relate any set of input parameters to output variables
depending on the provided training samples, which can be both
experiment and simulation results.
Nomenclature
C ¼ covariance matrix
D ¼ training data set
g ¼ transformation matrix
h ¼ transformed correlation length
k ¼ the dimension of function f
L ¼ scaling matrix
m ¼ number of outputs (dimension of output)
n ¼ number of input properties (dimension of input)
N ¼ number of training samples
Nðl; RÞ ¼ Gaussian distribution with mean and covariance of l
and R
p ¼ power parameter
pð f Þ ¼ probability of estimation function f
q ¼ number of neurons in hidden layer
R ¼ rotation matrix
U ¼ hidden layer weights for neural networks
v ¼ total variance of training-data output
W ¼ input-layer weights for neural networks
X ¼ input vector
Xi ¼ input vector for ith sample
Xsn ¼ normalized input vector
X ¼ input mean
yi ¼ output value for ith sample
Y ¼ output vector
Ys ¼ standardized output vector
c ¼ covariance function
h ¼ rotation angle
rX ¼ input SD
Acknowledgment
The authors would like to thank the management of Reservoir
Focus for the permission to publish this work.
References
Amani, M., Al-Jubouri, M., and Shadravan, A. 2012. Comparative Study
of Using Oil-Based Mud Versus Water-Based Mud in HPHT Fields.
Adv. Petrol. Explor. Develop. 4 (2): 18–27. https://doi.org/10.3968/
j.aped.1925543820120402.987
Bishop, C. M. 2006. Pattern Recognition and Machine Learning. New
York City: Springer.
Bland, R. G., Clapper, D. K., Fleming, N. M. et al. 1993. Biodegradation
and Drilling Fluid Chemicals. Presented at the SPE/IADC Drilling
Conference, Amsterdam, 22–25 February. SPE-25754-MS. https://
doi.org/10.2118/25754-MS.
Cao, Q., Banerjee, R., Gupta, S. et al. 2016. Data Driven Production Fore-
casting Using Machine Learning. Presented at the SPE Argentina Ex-
ploration and Production of Unconventional Resources Symposium,
Buenos Aires, 1–3 June. SPE-180984-MS. https://doi.org/10.2118/
180984-MS
D’Orangeville, C. and Lasenby, A. N. 2003. Geometric Algebra for Physi-
cists. Cambridge, UK: Cambridge University Press.
Deutsch, C. V. and Journel, A. G. 1992. Geostatistical Software Library
and User’s Guide. New York City: Oxford University Press.
Diarra, R., Carrasquilla, J., Phelan, D. et al. 2016. New Approach In Lift-
ing Cement In Highly Depleted And Naturally Fractured Formations.
Presented at the IADC/SPE Drilling Conference and Exhibition, Fort
Worth, Texas, 1–3 March. SPE-178772-MS. https://doi.org/10.2118/
178772-MS
Duda, R. O., Hart, P. E., and Stork, D. G. 2012. Pattern Classification.
Hoboken, New Jersey: John Wiley & Sons.
Gidh, Y. K., Purwanto, A., and Ibrahim, H. 2012. Artificial Neural Net-
work Drilling Parameter Optimization System Improves ROP by Pre-
dicting/Managing Bit Wear. Presented at the SPE Intelligent Energy
International, Ultrecht, The Netherlands, 27–29 March. SPE-149801-
MS. https://doi.org/10.2118/149801-MS.

2017 SPE Drilling & Completion 7

ID: jaganm Time: 15:57 I Path: S:/DC##/Vol00000/170004/Comp/APPFile/SA-DC##170004

 DC175238 DOI: 10.2118/175238-PA Date: 6-March-17
Hawkins, D. M. 2004. The Problem of Overfitting. J. Chem. Inf. Comp.
Sci. 44 (1): 1–12. https://doi.org/10.1021/ci0342472.
Journel, A. G. and Huijbregts, C. J. 1978. Mining Geostatistics, London:
Academic Press.
Lee, J., Shadravan, A., and Young, S. 2012. Rheological Properties of
Invert Emulsion Drilling Fluid Under Extreme HPHT Conditions. Pre-
sented at the IADC/SPE Drilling Conference and Exhibition, San
Diego, California, 6–8 March. SPE-151413-MS. https://doi.org/
10.2118/151413-MS.
Mijnarends, R., Frolov, A., Grishko, F. et al. 2015. Advanced Data-Driven
Performance Analysis For Mature Waterfloods. Presented at the SPE
Annual Technical Conference and Exhibition, Houston, 28–30 Sep-
tember. SPE-174872-MS. https://doi.org/10.2118/174872-MS.
Osman, E. A. and Aggour, M. A. 2003. Determination of Drilling Mud
Density Change with Pressure and Temperature Made Simple and
Accurate by ANN. Presented at the Middle East Oil Show, Bahrain,
9–12 June. SPE-81422-MS. https://doi.org/10.2118/81422-MS
Rafieisakhaei, M., Barazandeh, B., and Tarrahi, M. 2016a. Analysis of
Supply and Demand Dynamics to Predict Oil Market Trends: A Case
Study of 2015 Price Data. Presented at the SPE/IAEE Hydrocarbon
Economics and Evaluation Symposium, Houston, 17–18 May. SPE-
179976-MS. https://doi.org/10.2118/179976-MS.
Rafieisakhaei, M., Barazandeh, B., Moosavi, A. et al. 2016b. Modeling
Dynamics of the Carbon Market: A System Dynamics Approach on
the CO2 Emissions and its Connections to the Oil Market. Oral presen-
tation given at the 34rd International Conference of the System Dy-
namics Society, Delft, The Netherlands, 17–21 July.
Rasmussen, C. E. and Williams, C. K. I. 2006. Gaussian Processes for
Machine Learning. Cambridge, Massachusetts: MIT Press.
Schakel, M. D. and Mesdag, P. R. 2014. Fully Data-Driven Quantitative
Reservoir Characterization by Broadband Seismic. SEG Technical
Program Expanded Abstracts 2014: 2502–2506. https://doi.org/
10.1190/segam2014-0671.1.
Shadravan, A. and Amani, M. 2012. HPHT 101: What Every Engineer or
Geoscientist Should Know about High Pressure High Temperature
Wells. Presented at the SPE Kuwait International Petroleum Confer-
ence and Exhibition, Kuwait City, Kuwait, 10–12 December. SPE-
163376-MS. https://doi.org/10.2118/163376-MS.
Shadravan, A. and Tarrahi, M. 2016. Machine Learning Leads Cost Effec-
tive Intelligent Fluid Design: Fluid Engineering Perspective. Presented
at the SPE Bergen One Day Seminar, Bergen, Norway, 20 April. SPE-
180033-MS. https://doi.org/10.2118/180033-MS.
Shadravan, A., Narvaez, G., Alegria, A. et al. 2015a. Engineering the
Mud-Spacer-Cement Rheological Hierarchy Improves Wellbore Integ-
rity. Presented at the SPE E&P Health, Safety, Security and Environ-
mental Conference-Americas, Denver, 16–18 March. SPE-173534-
MS. https://doi.org/10.2118/173534-MS.
Shadravan, A., Tarrahi, M., and Amani, M. 2015b. Intelligent Cement
Design: Utilizing Machine Learning Algorithms to Assure Effective
Long-term Well Integrity. Presented at the Carbon Management Tech-
nology Conference, Sugar Land, Texas, 17–19 November. CMTC-
440236-MS. https://doi.org/10.7122/440236-MS.
Shadravan, A., Tarrahi, M., and Amani, M. 2016. Agile Data-Driven Fluid
Design: Predicting the Properties of Drilling, Spacer and Cement
Slurry Fluids. Oral presentation given at the 2016 AADE Fluids Tech-
nical Conference and Exhibition, Houston, 12–13 April. AADE-16-
FTCE-07.
Shi, J. Q. and Choi, T. 2011. Gaussian Process Regression Analysis for
Functional Data. Boca Raton, Florida: CRC Press.
Subrahmanya, N., Xu, P., El-Bakry, A. et al. 2014. Advanced Machine
Learning Methods for Production Data Pattern Recognition. Presented
at the SPE Intelligent Energy Conference and Exhibition, Ultrecht,
The Netherlands, 1–3 April. SPE-167839-MS. https://doi.org/10.2118/
167839-MS.
Tarrahi, M. and Shadravan, A. 2016a. Inverse Modeling for Fluid System
Characterization through Machine Learning Algorithms. Presented at
the SPE Bergen One Day Seminar, Bergen, Norway, 20 April. SPE-
180034-MS. https://doi.org/10.2118/180034-MS.
Tarrahi, M. and Shadravan, A. 2016b. Advanced Big Data Analytics
Improves HSE Management. Presented at the SPE Bergen One Day
8 2017 SPE Drilling & Completion
ID: jaganm Time: 15:57 I Path: S:/DC##/Vol00000/170004/Comp/APPFile/SA-DC##170004
Stage: Page: 8 Total Pages: 10
Seminar, Bergen, Norway, 20 April. SPE-180032-MS. https://doi.
org/10.2118/180032-MS.
Tarrahi, M., Jafarpour, B., and Ghassemi, A. 2013. Assimilation of Micro-
seismic Data into Coupled Flow and Geomechanical Reservoir Models
with Ensemble Kalman Filter. Presented at the SPE Annual Technical
Conference and Exhibition, New Orleans, 30 September–2 October.
SPE-166510-MS. https://doi.org/10.2118/166510-MS.
Tarrahi, M., Jafarpour, B., and Ghassemi, A. 2015. Integration of Micro-
seismic Monitoring Data Into Coupled Flow and Geomechanical Mod-
els With Ensemble Kalman Filter. Water Resour. Res. 51 (7):
5177–5197. https://doi.org/10.1002/2014WR016264.
Theodoridis, S., Pikrakis, A., Koutroumbas, K. et al. 2010. Introduction to
Pattern Recognition: A Matlab Approach. Burlington, Massachusetts:
Academic Press.
Williams, C. K. and Rasmussen, C. E. 2006. Gaussian Processes for
Machine Learning. Cambridge, Massachusetts: The MIT Press.
Appendix A—GPR
To perform GPR, we consider one output at a time [i.e., we con-
struct seven GPR models for seven dial-reading outputs (Shi and
Choi 2011)]. Therefore, the training data for each GPR-model
design is
Training data DfXi ; yi g; i ¼ 1; 2; …; N; . . . . . . . . ðA-1Þ
where yi is a scalar and represents one of the dial-reading outputs.
Each nonlinear-regression function is a random sample (para-
metrized by the input variable) drawn from a joint Gaussian prob-
ability function given the training data set D. Our goal in GPR is
to train a function f from data D. A Gaussian process is a set of
random estimation functions, f , each with probability pð f Þ that
formally can be expressed as Bayesian formulation while a set of
training samples D is given:
pð f ÞpðDj f Þ
pð f jDÞ ¼ ; . . . . . . . . . . . . . . . . . . . . . ðA-2Þ
pðDÞ
where pð f jDÞ shows the posterior probability of all the regression
functions given the training data. pð f Þ represents the prior proba-
bility of all the possible regression functions and pðDj f Þ is the
likelihood function. In Bayesian inference context, pðDÞ is only a
scaling parameter that does not affect the intended posterior prob-
ability density function (PDF) pð f jDÞ.
The nonlinear-regression function in GPR is a set of random
variables indexed by a continuous parameter (e.g., time, space, or
temperature), which is also called random function f ðXÞ. The
assumption in GPR is that any set of regression functions has a
jointly Gaussian distribution with zero mean:
k 1 1
   f CðXÞ1 f
pð f jXÞ ¼ N½0; CðXÞ ¼ ð2pÞ 2 jCðXÞj 2 e 2 ; . . . ðA-3Þ
where CðXÞ is the covariance matrix and k represents the dimen-
sion of function f with respect to the continuous parameter. Esti-
mating the unknown target value y0 given the set of training y
presents a conditional-probability function:
pðy0 jyÞ ¼ Nðy0 ; ry0 Þ: . . . . . . . . . . . . . . . . . . . . . . ðA-4Þ
The GPR estimation value is the mean of the above condi-
tional PDF (Eq. A-4). Basically GPR proposes a Gaussian PDF
for the new estimated value, which can be fully characterized by a
mean and a SD value.
As a necessary preprocessing step to avoid scaling issues from
different input and output ranges, we apply data normalization
and scaling to input and output values. Target or output values are
assumed to be normal standard variables (resulting from a Gaus-
sian distribution with zero mean and unit variance). Moreover,
the input values (input parameters) are shifted and scaled to a
[0,1] interval.
 DC175238 DOI: 10.2118/175238-PA Date: 6-March-17 Stage: Page: 9 Total Pages: 10

X  Xmin 2 3
Xs ¼ . . . . . . . . . . . . . . . . . . . . . . . . ðA-5Þ CðX1 ; X1 Þ  CðX1 ; XN Þ
Xmax  Xmin 6 .. .. .. 7
Cnn ¼4 . . . 5
YY CðXN ; X1 Þ … CðXN ; XN Þ
Ysn ¼ : . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-6Þ
rY ¼ Covariance matrix of training data; . . . . . . . ðA-18Þ
To define the spatial relation of the samples in the input space 2 3
(in this study, a 3D space), a covariance function is established: CðX1 ; X0 Þ
6 .. 7
Cnu ¼6
4 .
7
5
CðXi ; Xj Þ ¼ vexpðhp Þ; . . . . . . . . . . . . . . . . . . . . . ðA-7Þ
CðXN ; X0 Þ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
h ¼ gT L1 g; . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-8Þ ¼ Cross covariance of training data and new input:
                   ðA-19Þ
g ¼ RðXi  Xj Þ; . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-9Þ
2 3 The estimated output is basically a linear combination of the
l1 0 0 training outputs where the coefficients are determined by the spa-
L¼40 l2 0 5 ¼ Scaling matrix; . . . . . . . . . . . ðA-10Þ tial configuration of the input (with respect to the training data)
0 0 l3 and the covariance-function shape. One of the unique advantages
of GPR compared with ANN is the ability to obtain the estimation
Xi ; Xj ¼ Experiment input values ð3  1Þ; . . . . . . . ðA-11Þ error. GPR not only provides us with the estimated output at the
2 3 new input value, but it is also able to calculate its associated error
1 0 0 or variance. With this capability, we can specify how good each
R1 ðh1 Þ ¼ 4 0 cosðh1 Þ sinðh1 Þ 5; . . . . . . . . . . . ðA-12Þ estimation is or determine the confidence of each estimation and
0 sinðh1 Þ cosðh1 Þ decide if we need to perform new laboratory experiments.
2 3
cosðh2 Þ 0 sinðh2 Þ
Estimation variance ofy0 ¼ v  CTnu Cnn Cnu : . . . . . . ðA-20Þ
R2 ðh2 Þ ¼ 4 0 1 0 5; . . . . . . . . . . . . ðA-13Þ
sinðh2 Þ 0 cosðh2 Þ
2 3 This provides us with the estimation SD and consequently the
cosðh3 Þ sinðh3 Þ 0 confidence interval associated with each estimation. It should be
R3 ðh3 Þ ¼ 4 sinðh3 Þ cosðh3 Þ 0 5; . . . . . . . . . . ðA-14Þ noted that GPR is an absolute estimator; that is, estimated values
0 0 1 for the training data are equal to true outputs, and unlike ANN,
the estimation error for training data is zero. After tuning by use
and of LOOCV, the average minimum cross-validation error is 20%.
We obtained different hyper parameters (different covariance
R ¼ R3 ðh3 ÞR2 ðh2 ÞR1 ðh1 Þ; . . . . . . . . . . . . . . . . . ðA-15Þ functions) for different dial-reading outputs. For all the output
dial readings, the obtained power parameter p through cross vali-
where l1 ; l2 , and l3 are correlation lengths in three input directions; dation is 2, so the preferred covariance function is the squared-
the matrix R represents the rotation matrix in 3D, which can be exponential-covariance function.
calculated by use of the rotation angles around different axes
(D’Orangeville and Lasenby 2003); and v represents the variance
parameter. In specific cases, where the parameter p ¼ 2, the result- Arash Shadravan is a director at ReservoirFocus in Houston. He
ing covariance is the squared-exponential-covariance function (or manages various projects related to enhancing the econom-
Gaussian), and if p ¼ 1, the exponential-covariance function is ics of upstream operations in the US. Shadravan has previously
reproduced. The parameters involved in defining the covariance worked for Schlumberger, Baker Hughes, Occidental Petro-
function are called hyper parameters. By changing the hyper pa- leum, and Superior Energy Services. He has been a journal
reviewer and technical contributor as a committee member
rameters, the GPR estimator quality will change. In this study, we
in SPE conferences and workshops. Shadravan has published
tune seven hyper parameters (three correlation lengths, three rota- more than 30 peer-reviewed journal and technical confer-
tion angles, and the power) through LOOCV to obtain the optimal ence papers related to the application of machine learning in
GPR estimator. In LOOCV, we assume one of the training data production optimization, cementing, drilling fluids, and under-
points is unknown; use the other eight training data points to balanced drilling. He has been a SPE member since 2006.
design the GPR estimator; estimate the unknown data point; and Shadravan holds a master’s degree in petroleum engineering
then compare the estimated value with the true value. This proce- from Texas A&M University and a bachelor’s degree in petro-
dure is repeated for all the outputs (seven dial readings) and for leum engineering.
all the training data (nine samples). In this cross-validation proce- Mohammadali Tarrahi is a senior reservoir engineer in the Subsur-
dure, the performance measure of the GPR estimator is the aver- face Modeling and Optimization Department at Shell Global
age relative error. Solutions in Houston. He has been part of several research
The following is the relation between the semivariogram (pop- and development groups at Shell, Occidental Petroleum Corpo-
ular spatial-correlation representation in geostatistics) and the co- ration, and Microseismic. Tarrahi is the author of more than
variance function: 30 peer-reviewed-journal and technical-conference papers
regarding numerical reservoir simulation, geostatistics, reservoir
characterization, uncertainty assessment, big-data analytics,
cðXi ; Xj Þ ¼ v  CðXi ; Xj Þ: . . . . . . . . . . . . . . . . . . . ðA-16Þ and statistical machine learning. He holds a PhD degree in petro-
leum engineering from Texas A&M University, a master’s degree
To estimate the corresponding output y0 of the new input pa- in petroleum engineering, and bachelor’s degrees in electrical
rameter X0 , we use the Gaussian process interpolation (also called engineering and petroleum engineering.
simple Kriging):
Mahmood Amani is an associate professor of petroleum engi-
2 3T neering at Texas A&M University at Qatar (TAMUQ), and has
y1 been on the faculty of Texas A&M University for more than 13
6 .. 7 1 XN
y0 ¼ 4 . 5 Cnn Cnu ¼ ki yi ; . . . . . . . . . . . . . . ðA-17Þ years. He moved from the main campus of Texas A&M Univer-
i¼1 sity in College Station, Texas, to Qatar to help start the Petro-
yN leum Engineering Program at TAMUQ. Amani is the founding

2017 SPE Drilling & Completion 9

ID: jaganm Time: 15:57 I Path: S:/DC##/Vol00000/170004/Comp/APPFile/SA-DC##170004

 DC175238 DOI: 10.2118/175238-PA Date: 6-March-17 Stage: Page: 10 Total Pages: 10

faculty member and the first program coordinator and
chair of the Petroleum Engineering Program at TAMUQ. He
considers his role in establishing the petroleum-engineering
program in Qatar as his most-remarkable professional
achievement. Before joining academia as a faculty member
focused on petroleum engineering, Amani worked as a
research scientist with the Texaco Exploration and Production
Technology Department in Houston. He was awarded the
2016 SPE Middle East Regional Distinguished Achievement
Award for Petroleum Engineering Faculty as well as the 2013
SPE Middle East Regional Service Award, and he holds two US
patents. Amani holds a bachelor’s degree in mechanical en-
gineering from Wichita State University, a master’s degree in
natural-gas engineering from Texas A&M University-Kingsville,
and a PhD degree in petroleum engineering from Texas A&M
University.

10 2017 SPE Drilling & Completion

ID: jaganm Time: 15:57 I Path: S:/DC##/Vol00000/170004/Comp/APPFile/SA-DC##170004