Energy and AI 2 (2020) 100031

Contents lists available at ScienceDirect

Energy and AI
journal homepage: www.elsevier.com/locate/egyai

Data–driven decision–making for lost circulation treatments: A machine

learning approach
Husam H. Alkinani∗, Abo Taleb T. Al-Hameedi∗, Shari Dunn-Norman
Missouri University of Science and Technology, 1201 N State St, Rolla, MO 65409, USA

h i g h l i g h t s g r a p h i c a l a b s t r a c t
• After examining multiple machine learn-
ing methods, quadratic SVM was se-
lected to train the model since it gave
the highest accuracy.
• 5-fold cross-validation was used to avoid
overfitting and poor generalization.
• QC/QA was implemented in the process
of selecting the lost circulation treat-
ments to limit the error due to incorrect
classification.
• The main decision-making criteria were
based on whether the treatment will be
successful or not. The cost is a pivotal pa-
rameter in the decision-making process.
• Two treatments recommended to stop
partial loss, four for severe loss, and
seven for complete loss, depending on
the reason of loss.

a r t i c l e i n f o a b s t r a c t

Article history: Lost circulation is an expensive and critical problem in the drilling operations. Millions of dollars are spent every
Received 31 July 2020 year to mitigate or stop this problem. In this work, data from over 3000 wells were collected from multiple
Received in revised form 1 October 2020 sources. The data went through a processing step where all outliers were removed and decision rules were set up.
Accepted 3 October 2020
Multiple machine learning methods (support vector machine, decision trees, logistic regression, artificial neural
Available online 15 October 2020
networks, and ensemble trees) were used to create a model that can predict the best lost circulation treatment
Keywords: based on the type of loss and the reason of loss. 5-fold cross-validation was conducted to ensure no overfitting in
Machine learning the created model. After using all the aforementioned machine learning methods to train models to choose the
Lost circulation best lost circulation treatment, overall, the results showed that support vector machine had the highest accuracy
Data-driven
among the other algorithms. Thus, it was selected to train the model. The created model went through quality
Classification
control/quality assurance (QC/QA) to limit the results of incorrect classification. Two treatments were suggested
to treat partial loss, four to treat severe loss, and seven for complete loss, based on the reason of loss. In addition,
a formalized methodology to respond to lost circulation was provided to help the drilling personnel handling lost
circulation in the field.

1. Introduction more drilling fluids, time, and extra cost. It can, also, bring on issues like
formation damage caused by plugging of pore throats by mud particles,
Lost circulation is an expensive and critical problem at any point in unsuccessful production tests, borehole instability, well control issues,
the drilling operations [1–5]. Improper or untimely responses cause loss substandard hydrocarbon production after well completion, and stuck
pipe [6–8]. A ton of money is spent every year due to this issue. An
∗ average of 10% to 20% of the price of drilling under extreme tempera-
Corresponding authors.
E-mail addresses: hhappf@umsystem.edu (H.H. Alkinani), ata2q3@umsystem.edu (A.T.T. Al-Hameedi).

https://doi.org/10.1016/j.egyai.2020.100031
2666-5468/© 2020 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
H.H. Alkinani, A.T.T. Al-Hameedi and S. Dunn-Norman
tureandand and pressure wells is spent on mud losses according to the
United States Department of Energy in 2010 [9]. Drilling fluid materials
are very costly. The global market for drilling fluids reveals an annual
maximize of 10.13% which indicates that in 2018, the cost of drilling
fluid materials will reach 12.31 billion dollars [10]. On average 10% of
the total well costs are the cost of drilling mud. The drilling fluid can
greatly affect the overall cost as well [11]. Losing drilling fluids into the
formation is called lost circulation events, are one of the greatest chal-
lenges to be stopped throughout the drilling process. The seriousness of
the outcomes depends on loss severity. The outcome could range any-
where from losing drilling fluid to ending up in a blowout [12]. Lost
circulation is a major drilling challenge facing the oil and gas industry.
Large progress has been made to learn how to combat this issue. How-
ever, many items and instructions are available for countering this issue
are biased towards announcement for a service company.
There are a couple of conditions required for lost circulation to hap-
pen downhole. First, is that the pressure in the wellbore needs to be
greater than the pore pressure, and second, when mud loss happens,
there needs to be a flow pathway [13,14]. Lost circulation can result
from either natural or induced causes and can range from a couple of
barrels per hour to hundreds of barrels in minutes. For natural reasons,
lost circulation events are often encountered when drilling into cav-
ernous, vugular, high permeability, and naturally fractured formations
due to both the geology and the drilling conditions. Drilling fluid loss
into natural fractures, cavernous, vugular, and high permeability forma-
tions are initiated as long as the drilling mud pressure exceeds the pore
pressure (formation pressure), and in some cases even applying under-
balanced drilling phased can lead to facing mud loss because the tensile
strength for naturally fractured formations is very small. When lost cir-
culation is encountered in naturally fractured formation, the practical
practice of is utilizing blind drilling technique to penetrate the entire
section of the loss zone, then corrective remedies will be pumped to
plug the thief zone. On the other hand, lost circulation will be faced due
to induced fractures as a result of drilling-induced tensile failures. In
the same context, mud losses into induced fractures are initiated when
drilling mud pressure exceeds fracture pressure, particularly in the de-
pleted and weak formations. In most cases, induced mud losses are gen-
erated from the creation and extension of fractures by errors in drilling
practice and incorrect design of the mud weight window. To sum it
all up, induced fractures are created when the effective mud weight of
drilling mud surpasses the formation strength, opening a fracture and
paths for the drilling fluid to be lost. Lost circulation is grouped by the
quantity of mud or fluid lost per hour [13,15,16]:
• Complete loss (no return)
• Severe loss (loss rate goes up to 15 m3 /h)
• Partial loss (loss rate goes up 1–10 m3 /h
• Seepage loss (loss up to 1 m3 /h)
Treating drilling fluid with filtration control agents or standard lost
circulation material (LCM) as concentrated pills or background treat-
ments is a practice to diminish seepage or partial losses. Different so-
lutions that need a greater length of time for placement and prepara-
tion can be used for severe or complete losses such as cement [17–19],
cross-linked cement [20], dense sand slurries [21], chemically initiated
cross-linked pills [22,23], gunk squeezes [23,24], deformable-viscous-
cohesive systems [25–27], and nanocomposite gel [28].
Deciding which lost circulation treatment to use is based on differ-
ent factors such as type of loss zone, drilling fluid characteristics, the
amount of lost circulation, and drilling operation parameters [13,29].
Discovering an analytical solution to this problem is difficult because
of the complexity and nonlinear behavior. Due to these complications,
machine learning methodologies are an appealing replacement to model
this elaborate physical procedure. Computational intelligence methods
2
Energy and AI 2 (2020) 100031
inquire from past experiences with sets of different learning data can
help the decision-making process.
The aim of this work is to use various classification algorithms (sup-
port vector machine, decision trees, logistic regression, artificial neural
networks, and ensemble trees) to choose the best lost circulation treat-
ment based on the type of loss and reason of lost circulation using his-
torical data from 3000 wells collected from many sources. In addition,
this work will serve as a practical guideline for lost circulation treatment
strategies based on data-driven models.
2. Machine learning
An easy way to solve a problem is by having a simple and well-
defined equation. This is very difficult in real life with so many prob-
lems. For example, it is easy for a person to differentiate family members
by their looks, but there is no simple equation that can do this [30]. Ar-
tificial intelligence is focused on the ability of computers/machines to
act like humans when making decisions. Machine learning is a subset
of artificial intelligence. There are many algorithms used to implement
machine learning. Examples include; support vector machine, artificial
neural network (ANN), nearest neighbor, decision tree, regression, en-
semble learning, and Naive Bayes [30,31]. Certain machine learning
algorithms have a better learning accuracy, but other factors such as
the speed and the difficulty of interpreting results are all critical for
considering which machine learning algorithm to use. The following
sub-sections will briefly explain the machine learning methods used in
this study.
2.1. Support vector machine
Support vector machine (SVM) represents a powerful technique in
machine learning due to its ease of use and performance [32]. The clas-
sification in SVM is performed by creating a decision boundary that has
the largest distance from the training points. Data that are not linearly
separable can be classified in SVM by creating a higher-dimensional hy-
perspace [33,34]. Wang et al. [35] used SVM and neural networks as
well as analytical models to predict the drilling fluid density in high
temperature and pressure environment. The findings showed that the
SVM did better than the other algorithms. SVM was utilized for lithol-
ogy classification and permeability estimation by Al-Anazi and Gates
[36]. They used linear discriminant analysis and probabilistic neural
networks as well as general regression neural networks alongside SVM.
The results showed that SVM outperformed other methods. Deng et al.
[37] used SVM to classify lithology of crystalline rocks within an accept-
able margin of error.
2.2. Decision Tree (DT)
DT is a method used for classification by creating tree structures for
the training data. The tree structure can consist of a few nodes for a sim-
ple tree, to hundreds for a complex tree. DT was utilized by Perez et al.
[38] to estimate the hydraulic flow units and lithofacies from well logs.
Perez et al. [38] showed that the more nodes the tree has, the better
the estimation. However, too many nodes can lead to overfitting and
complex interpretation for the influence of the inputs on the outputs.
The benefit of DT is that it can present a model that can easily show
the influence of the inputs on the outputs. DT can be combined with
neural networks to have the benefit of neural networks and DT by form-
ing a neural decision tree. Sabah et al. [39] used multiple algorithms,
including DT, to predict lost circulation with a case history from Marun
oil field, Iran. Sabah et al. [39] showed that DT outperformed other
algorithms to predict lost circulation. Alkinani et al. [40] provided a
summary of the applications of DT in the oil and gas industry.
H.H. Alkinani, A.T.T. Al-Hameedi and S. Dunn-Norman
2.3. Ensemble learning (ESL)
ESL has proven to be successful in image classification, face recog-
nition, and medical image analysis. This is due to its ability to com-
bine multiple classifiers which will lead to mitigating the weaknesses
in those classifiers [41]. There are many algorithms used for ESL, in-
cluding but not limited to; boosted tree, bagged tree, and random un-
dersampling boosting (RUSBoost). More information about these algo-
rithms and other ESL algorithms can be found in the literature [42].
Anifowose et al. [43] used ESL in reservoir characterization modeling.
Anifowose et al. [43] showed that the incorrect decision from a classi-
fier is a binomial distribution. Thus, Anifowose et al. [43] showed that
the uncorrelated errors from the classification can be minimized by av-
eraging.
2.4. Logistic regression
A very reliable classification technique is logistic regression. A com-
monly used instrument by most statisticians. It has an S-shaped distri-
bution which is seen in many areas such as banking, marketing, de-
mographics, psychology, and epidemiology. Because of its qualities, the
technique is being used in supervised classification problems [44]. It
is used to show categorical dependent variables that acquire discrete
outcomes. Independent variables can be quantitative or qualitative. It
can have qualitative variables with more than two responses. When the
response variable has 2, ≥3 nominal, ≥3 ordered categories, it can be
categorized as ordinal logistic regression, multinomial and binary. Lo-
gistic regression was used for EOR applications, drilling, and rock type
identification [45–47].
2.5. Neural Networks (NNs)
An artificial neural network is “an information-processing system
that has certain performance characteristics in common with biolog-
ical neural network” [48]. The idea of neural networks is to mimic
the biological neurons. A neural network consists of three layers; one
input layer where the inputs are fed, one or multiple hidden layers
where the information is processed and features are extracted from the
data, and one output layer where the results are given. NNs can con-
sist of one hidden layer, or very complex with many hidden layers.
More hidden layers require more computational power. Muojeke et al.
[49] utilized NNs, binary classifier, and monitoring of the drilling flow
parameters to detect kicks during drilling operations. Adedigba et al.
[50] presented NNs model to a process accident. NNs were used to as-
sess the risk of blowouts on offshore platforms [51]. Chamkalani et al.
[52] used multiple pattern recognition methods for drilling optimiza-
tion, including NNs. Prediction of lost circulation for naturally fractured
reservoirs using neural networks was also explored [53]. Furthermore,
Li et al. [54] used long short-term memory (LSTM), a neural-network-
based model, to generate synthetic nuclear magnetic resonance (NMR)
logs using “easy-to-acquire” fluid saturation and formation mineral logs.
Onalo et al. [55] used dynamic neural networks to estimate shear and
compressional wave travel time. He et al. [56] used multiple machine
learning methods, including neural networks, to predict shear and com-
pressional travel times. Alkinani et al. [57] provided a summary of NNs
applications in the oil and gas industry.
3. Data and methods
In this section, the process of data collection and processing, as well
as the SVM algorithm, will be explained in details since it is the algo-
rithm that was selected among the others. The other algorithms tested
in this study (decision tree, ensemble learning, NNs, and logistic regres-
sion) will not be explained since that is beyond the scope of the paper.
3
Energy and AI 2 (2020) 100031
3.1. Data collection
The data used in this study were collected from many resources in-
cluding daily drilling reports, mud logging reports, final drilling reports,
etc. Most of the data were in pdf formats. Each data point was collected
manually which was a very time-consuming process. In addition, data
from case studies in the literature were collected [15,57–92]. More than
3000 wells were used in this study. Fig. 1 shows a map with the red dots
being the locations of data collections. Lost circulation treatment names
used in this study are summarized in the Appendix. It is worth mention-
ing that the data used for this study are up to the beginning of 2019.
3.2. Data processing, algorithms testing, and input data selection
The data were cleaned from any outliers [29]. Once the data were
cleaned, the data were classified based on the type of loss as partial,
severe, and complete loss. In addition, the data were classified based on
the reason for lost circulation as induced fractures and natural fractures
as well as vugs and caves.
The total number of treatments will influence the outcome of the
model. As an example, it is not reasonable to compare the probability
of a treatment used 10 times with a treatment used 100 times, the re-
sults will be biased in this case. The cut off of less than 50 was set after
carefully looking at the whole database to avoid any biases in proba-
bility calculations. In other words, any lost circulation treatment that
was utilized less than 50 times was ignored. In addition, the classifica-
tion problem was set up to predict success or fail based on the type of
loss and reason of lost circulation. The probability cut off for success
was set to be ≥75%. Thus, when the model predicts success, it means
the probability is ≥75%. Otherwise, the treatment is considered a failed
treatment. The inputs for the models were selected based on trial and
error to meet two goals, the first one is to have the highest accuracy, and
the second one is to minimize the number of predictors. Table 1 summa-
rizes the predictors (inputs) and the response (output). The algorithms
used for this study are summarized in Table 2.
The accuracy (ACC) of each classifier was measured by the following
Eq. (1) [93]:
𝑇𝑃 + 𝑇𝑁
𝐴𝐶𝐶 (%) = 100 × (1)
𝑇𝑃 + 𝐹𝑃 + 𝑇𝑁 + 𝐹𝑁
where TP is the true positive, TN is the true negative, FP is the false
negative, and FN is the false negative.
3.3. SVM algorithm
Support vector machine (SVM) is a powerful machine learning al-
gorithm. SVM performs classification by creating a decision boundary
with the largest distance to the data point [34]. Fig. 2 shows the theory
of the SVM algorithm. Assume there are two classes "×" and "−". The
goal is to maximize the margin between these two classes.
Assume w is a vector perpendicular to the optimal hyperplane and u
is an unknown vector. Then u classified as "×" if (this will be the decision
rule) [33,34]:
𝒇 (𝒖 ) = 𝒘 ⋅ 𝒖 + 𝒃 > 0 (2)
Assume x× is a known vector classified as "×" class. Then, constrain
for all "×" samples vectors is:
( )
𝒇 𝒙× = 𝒘 ⋅ 𝒙 × + 𝒃 ≥ 1 (3)
Assume 𝒙_ is a known vector classified as "−" class. Then, constrain
for all "−" samples vectors is:
( )
𝒇 𝒙− = 𝒘 ⋅ 𝒙− + 𝒃 ≤ −1 (4)
Assume 𝒚 𝒊 = +1 for the "×" class, and 𝒚 𝒊 = −1 for the "−" class. Mul-
tiplying y i by both Eqs. (3) and (4) and doing some algebra will result
in one equation as follows:
( )
𝒚 𝒊 𝒘 ⋅ 𝒙𝒊 + 𝒃 − 1 ≥ 0 (5)
H.H. Alkinani, A.T.T. Al-Hameedi and S. Dunn-Norman Energy and AI 2 (2020) 100031

Fig. 1. Data Collection Locations.

Table 1
Predictors and Response.

Predictors Response

Reason Treatment Number (NO) Type of Loss Well Type Results

Vugs and Caves 1–46 Partial Vertical Fail

Natural Fractures N/A Severe Deviated Success
Induced Fractures N/A Complete N/A N/A

Table 2
Summary of the Classifiers Used in This Study [42].

Classifier Classifier Type Prediction Speed Memory Usage

Tree Fine Fast Small

Medium
Coarse
SVM Linear Binary: Fast Multiclass: Medium Medium
Quadratic Binary: Fast Multiclass: Slow Binary: Medium Multiclass: Large
Cubic
Fine Gaussian
Medium Gaussian
Coarse Gaussian
ANN Vary depends on the training Vary depends on the training
algorithm algorithm
Logistic Regression (LR) Fast Medium
Ensemble Boosted Tree Fast Low
Bagged Tree Medium High
RUSBoosted Tree Fast Low

For xi in the hyperplane, the results will be: be minimized:

( ) 1
𝒚 𝒘 ⋅ 𝒙𝒊 + 𝒃 − 1 = 0 (6) ‖𝒘‖2 (9)
2
The distance between the decision lines (width of the hyperplane) The Lagrange’s multipliers can be written as:
can be expressed as: ∑𝒍
[ ( ) ]
1
𝒘 ( ) 𝑳= ‖𝒘‖2 − 𝜶 𝒊 𝒚 𝒊 𝒘 ⋅ 𝒙𝒊 + 𝒃 − 1 (10)
𝒘𝒊𝒅 𝒕𝒉 = ⋅ 𝒙× − 𝒙− (7) 2 𝒊=1
‖𝒘‖
where 𝜶 i is the multipliers for the constrains. The derivative of
After honoring the constrains in Eq. (6), the results will be as follow:
Eq. (10) with respect to w will give:
𝒘 ( ) 2 ∑𝒍
⋅ 𝒙× − 𝒙− = (8) 𝜹𝑳
‖𝒘‖ ‖𝒘‖ =𝒘− 𝜶 𝒊 𝒚 𝒊 𝒙𝒊 = 0 (11)
𝜹𝒘 𝒊=1
To maximize the optimal margin (distance between the lines),
Then,
Eq. (8) has to be maximized (length of vector w should be minimized),
𝒍
∑
while still honoring constrains (Eq. (6)). A possible solution is to use La-
𝒘= 𝜶 𝒊 𝒚 𝒊 𝒙𝒊 (12)
grange’s method, to maximize the optimal margin, the following should
𝒊=1

4
H.H. Alkinani, A.T.T. Al-Hameedi and S. Dunn-Norman Energy and AI 2 (2020) 100031

Table 3
Values of 𝛾 for Radial Basis Kernel [42].

Radial Basis Kernel 𝛾

√
Fine Gaussian SVM 𝑃 ∕4, where P is the number of predictors
√
Medium Gaussian SVM 𝑃
√
Coarse Gaussian SVM 𝑃 × 4

To use and recognize the classes, the following is needed:

( )
∅ 𝑥𝑖 ⋅ ∅(𝑢) (17)

Then, introducing the Kernel function (K) which will provide the dot
product of the two vectors in another space as the following:
( ) ( ) ( )
𝐾 𝑥𝑖 , 𝑥𝑗 = ∅ 𝑥𝑖 ⋅ ∅ 𝑥𝑗 (18)

The most common Kernel functions are:

Fig. 2. SVM Algorithm. ( ) ( )𝑛
𝑃 𝑜𝑙𝑦𝑛𝑜𝑚𝑖𝑎𝑙 𝐾 𝑥𝑖 , 𝑥𝑗 = 𝑥𝑖 ⋅ 𝑥𝑗 + 1 (19)

( )
Taking the derivative of Eq. (10) with respect to b (which is a con- 𝑅𝑎𝑑𝑖𝑎𝑙 𝐵𝑎𝑠𝑖𝑠 (𝐺𝑎𝑢𝑠𝑠𝑖𝑎𝑛) 𝐾 𝑥𝑖 , 𝑥𝑗 = 𝑒−𝛾 ‖𝑥𝑖 − 𝑥𝑗 ‖2
(20)
stant) will give:
where n is the degree of the polynomial function (linear=1, quadratic=2,
𝒍
𝜹𝑳 ∑ cubic=3), and 𝛾 sets the width of the bell shaped curve. The values of 𝛾
= 𝜶𝒊 𝒚𝒊 = 0 (13) used for fine, medium, and coarse Gaussian are summarized in Table 3.
𝜹𝒃 𝒊=1

Then, 3.4. Cross-validation

𝒍
∑
𝜶𝒊 𝒚𝒊 = 0 (14) Cross-validation is important to ensure the robustness of the created
𝒊=1 model. The problem associated with dividing the data to training and
After doing some algebra and using Eqs. (12) and (14), the following testing is that some representative samples may be missing during the
formula will be obtained: training process. Cross-validation ensures a better representation of the
𝑙 𝑙 𝑙 data in the training set. In this study, 5-fold cross-validation was utilized
∑ 1 ∑∑
𝑀𝑎𝑥𝑖𝑚𝑖𝑧𝑒 𝐿 = 𝛼𝑖 − 𝛼 𝛼 𝑦 𝑦 𝑥 ⋅ 𝑥𝑗 (15) to ensure a robust model. The idea of 5-fold cross-validation is that the
𝑖=1
2 𝑖=1 𝑗=1 𝑖 𝑗 𝑖 𝑗 𝑖 data will be randomly divided into 5 equal-sized sets (5-folds). Then, in
the training process, one of the pairs will be used for testing (the other
The maximization of Eq. (15) depends on the dot product. To transfer
4 will be used for training). This is done five times, each time leaving
into another space, where data are separable, a transformation function
another fold out of the training and used for testing. Finally, the average
(Ø(x)) can be utilized. Since Eq. (15) depends on the dot product, then
of those iterations will be taken to calculate the accuracy of the model
the following will have to be maximized:
[94]. This will ensure that every data point is represented in the training
( ) ( )
∅ 𝑥𝑖 ⋅ ∅ 𝑥𝑗 (16) process.

Fig. 3. Comparison between Different Classification Methods.

5
H.H. Alkinani, A.T.T. Al-Hameedi and S. Dunn-Norman
Fig. 4. Quadratic SVM Training Scatter Plot Based on (a) Reason of Loss and (b) Type of Loss.
4. Results
After using multiple machine learning methods to train models to
choose the best lost circulation treatment, the model with the highest
accuracy (ACC) was selected. Fernández-Delgado et al. [95] evaluated
179 classifiers and used 121 data sets and found out that SVM and ran-
dom forest provided the best results overall compared to the other al-
gorithms. Overall, the results showed that the support vector machine
had the highest accuracy among the other algorithms. Fig. 3 shows the
accuracy of all models created to choose the best lost circulation treat-
ment. Quadratic SVM showed the best performance among the other al-
gorithms with an accuracy of 74%. Thus, the quadratic SVM model was
selected to choose the best lost circulation treatment. Only the quadratic
SVM model results will be shown in this paper, the results from the other
algorithms will not be shown.
Fig. 4 (a) and (b) show a scatter plot example of the classification
of the treatments based on reason for loss and type of loss, respectively.
6
Energy and AI 2 (2020) 100031
Orange points indicate successful results and blue points show failed
treatments. In the same vein, dots indicate successful classification and
crosses show incorrect classification. Fig. 4 can be utilized to select the
best treatments based on the type of loss and reason for loss. For ex-
ample, treatment 23 was correctly classified as a successful treatment
used to treat naturally fractured formations (Fig. 4a). In the same vein,
treatment 23 was correctly classified as a successful treatment used to
treat partial loss (Fig. 4b). Thus, it can be concluded that treatment 23
can be used to treat partial loss in naturally fractured formations.
Fig. 5 shows the confusion matrix for the quadratic SVM. Confu-
sion matrix green boxes show successful classification while red boxes
show incorrect classification. The model shows a 79% successful clas-
sification rate for the fail class, and 66% successful rate for the success
class. Fig. 5 also shows the true positive rate and the false negative rate.
Figs. 6 and 7 show the receiver operating characteristics (ROC) curve
for the positive class being the success class and fail class being a posi-
tive class, respectively. A perfect ROC curve will have an area under the
H.H. Alkinani, A.T.T. Al-Hameedi and S. Dunn-Norman
Fig. 6. ROC Curve (Positive Class =Success Class).
curve to be 1. The area under the ROC curve of this model is 0.77 which
is considered good.
5. Discussion
Quadratic SVM gave the best accuracy (74%) among the other mod-
els. The closest model to the quadratic SVM was the NNs model with
an accuracy of 70%. However, the quadratic SVM was selected since it
gave the highest accuracy. The model was cross-validated using 5-fold
cross-validation to avoid overfitting and poor generalization. Thus, the
model is reliable and can be used for new data.
An argument can be made about the false negative rate which is
34% for the success class and 21% for the fail class. However, quality
control/ quality assurance (QC/QA) in the process of selecting the treat-
ments that will prevent or minimize this false negative rate. This is done
using the scatter plot in Fig. 4. The way the treatment selection work
is by looking at the correctly classified treatments (shown in dots) and
7
Energy and AI 2 (2020) 100031
Fig. 5. Confusion Matrix.
Fig. 7. ROC Curve (Positive Class =Fail Class).
recommend these treatments based on the type of loss and reason for
the loss. On the other hand, the incorrectly classified treatments (shown
in crosses) will be avoided since the model was unable to classify them
correctly. By doing this, the false negative rate can be minimized. Thus,
QC/QA is very important in the process of selecting the best lost circula-
tion treatments and avoiding the false negative rate. Using this criterion,
Fig. 8 was created that shows the results of the model after applying the
QC/QA process.
Starting with treatments of partial loss. For natural fractures, the
results showed that treatment 23 (High Fluid Loss Squeeze Resilient
Graphitic Carbon + Blend of (Granules, Flakes and Fibers) (Low Concen-
tration)) is the best treatment for partial loss due to natural fractures. On
the other hand, in case of induced fractures, two treatments resulted in
success; treatments 6 (Cement Plug) and treatment 12 (H.V Mud + Blend
of LCM (High Concentration)). In this case, the decision will be made
based on the cost. Cement plug treatments are pumped through open
end drill pipe (OEDP), not through bottom hole assembly (BHA). The
H.H. Alkinani, A.T.T. Al-Hameedi and S. Dunn-Norman
typical procedure in the field is to trip out of the hole, and take off the
BHA, and trip in again with OEDP and the treatment is pumped through
the OEDP to avoid bit nozzles plugging. Hence, this will add additional
non-productive time (NPT) which means more cost; while treatment 12
is pumped through the BHA and there is no need to trip out of the hole
to take off the BHA. The average total cost of any treatment includes
the cost of the materials, waiting period, and trip and out time. The cost
of treatments varies depending on the materials and the volume of the
treatment. After pumping the selected treatment, a waiting period has
to be allowed for this treatment to be effectively stabilized in the thief
zone. For example, treatment 6 (Cement Plug) has a waiting period of
10 hours after it is pumped in the hole; while treatment 29 (Pill of LCM
(High Concentration)) and treatment 30 (Pill of LCM (Low Concentra-
tion)) have a waiting period of approximately 3 hours. Trip in and out
of the hole is also an additional NPT for some of the treatments that are
pumped through OEDP. For instance, the total average cost of treatment
6 is $75,900, and the total average cost of treatment 12 is $12,000 [8].
In this case, the cost difference is significant. Thus, treatment number
12 (H.V Mud + Blend of LCM (High Concentration)) is recommended
for partial loss due to induced fractures since it has the required criteria
of success and lower cost.
For severe loss due to vugs and caves, treatment 3 (Blend of Fibers
in Cement) resulted in success. For induced fractures, there were three
treatments showed successful results; treatment 12 (H.V Mud + Blend
of LCM (High Concentration)), treatment 18 (High Concentration Acid
Soluble LCM), and treatment 21 (High Fluid Loss Squeeze Resilient
Graphitic Carbon + Blend of (Granules, Flakes and Fibers) (High Con-
centration)). All these treatments have proven successful results and can
8
Energy and AI 2 (2020) 100031
Fig. 8. Treatments Selection.
be used to treat severe mud loss due to induced fractures. It is very hard
to obtain cost data from the companies and even if it is available, it
will be limited. The data used for this study had a total average cost for
treatment 12 that was applied in Iraq, close to $13,000. The other two
treatment costs are not available [8]. However, it will not be significant
which treatment to choose since the cost will be close. It is recommended
to use any of them and if one fails for some reason, then the other two
treatments should be given priority in the application after the first one
fails.
In the case of complete loss, three treatments resulted in success for
complete loss due to vugs and caves; treatment 9 (Diesel Oil Bentonite
(DOB) Plug), treatment 10 (Diesel Oil Bentonite Cement (DOBC) Plug),
and treatment 11 (Gilsonite Cement). Treatments 9 and 10 are difficult
to be applied and they need a specialized crew to conduct the treat-
ments. Cost-wise, the cost difference is not significant comparing to the
entire drilling operations. For instance, treatment 9 total average cost
is $49,500, and treatment 9 total average cost is $53,000, while treat-
ment 11 total average cost is $48,000. A couple of thousand dollars is
not significant in comparison with the total cost of the drilling operation
[8]. On the other hand, the results showed complete loss due to induced
fractures has four successful treatments; treatment 10 (Diesel Oil Ben-
tonite Cement (DOBC) Plug), treatment number 11 (Gilsonite Cement),
treatment number 41 (Right Angle Set Polymers Cement), and treatment
number 42 (Soft and Hard Plugs Special Fibers Cement). Once again,
the cost difference is not significant and all four treatments have shown
successful results. Table 4 summarizes the treatments recommended for
each type of loss.
H.H. Alkinani, A.T.T. Al-Hameedi and S. Dunn-Norman Energy and AI 2 (2020) 100031

Table 4
Summary of the Recommended Treatments.

Type of Loss Reason of Loss Treatment

Partial Induced Fractures H.V Mud + Blend of LCM (High Concentration)

Natural Fractures High Fluid Loss Squeeze Resilient Graphitic Carbon + Blend of (Granules, Flakes and Fibers) (Low Concentration)
Severe Vugs and caves Blend of Fibers in Cement
Induced H.V Mud + Blend of LCM (High Concentration)
Fractures High Concentration Acid Soluble LCM
High Fluid Loss Squeeze Resilient Graphitic Carbon + Blend of (Granules, Flakes and Fibers) (High Concentration)
Complete Vugs Diesel Oil Bentonite (DOB) Plug
and Diesel Oil Bentonite Cement (DOBC) Plug
caves Gilsonite Cement
Induced Diesel Oil Bentonite Cement (DOBC) Plug
Fractures Gilsonite Cement
Right Angle Set Polymers Cement
Soft and Hard Plugs Special Fibers Cement

Table 5
Treatments Used in This study.

Treatment NO

Attapulgite Plug 1
Blend of Fibers (Low Concentration) 2
Blend of Fibers in Cement 3
Blend of Flaky (Low Concentration) 4
Blend of Granular (Low Concentration) 5
Cement Plug 6
Cement Plug + HV Mud 7
Crosslink Polymers Plug 8
Diesel Oil Bentonite (DOB) Plug 9
Diesel Oil Bentonite Cement (DOBC) Plug 10
Gilsonite Cement 11
High Viscosity (H.V) Mud + Blend of LCM (High Concentration) 12
H.V Mud Patch (Bentonite and Polymer) 13
H.V Mud Patch (Bentonite Material + Resilient Graphitic Carbon) 14
H.V Mud Patch (Bentonite Material) 15
H.V Mud Patch (Lime Material) 16
H.V Mud Patch (Salt Clay Material) 17
High Concentration Acid Soluble LCM 18
High Filtration Spot Pills, High Filtration Mixtures (200-400 cc API) 19
High Filtration Spot Pills, Very High Filtration Mixtures (> 600cc API) 20
High Fluid Loss Squeeze Resilient Graphitic Carbon + Blend of (Granules, Flakes and Fibers) (High Concentration) 21
High Fluid Loss Squeeze Resilient Graphitic Carbon + Blend of (Granules, Flakes and Fibers) (High Concentration) + Cement 22
High Fluid Loss Squeeze Resilient Graphitic Carbon + Blend of (Granules, Flakes and Fibers) (Low Concentration) 23
InstandSeal 24
Late Material + Blend of LCM (High Concentration) 25
Late Material + Blend of LCM (Low Concentration) 26
Late Material + Cement Plug 27
Magnesia Cross-linked Cement (MCC), 28
Pill of LCM (High Concentration) 29
Pill of LCM (Low Concentration) 30
Plugging Materials Nut Plug (Low Concentration) 31
Plugging Materials, Blend of CaCO3 (High Concentration) 32
Plugging Materials, Blend of Mica (High Concentration) 33
Plugging Materials, Blend Magma Fiber (High Concentration) 34
Plugging Materials, Blend of Nut Plug (High Concentration) 35
Plugging Materials, Fine and Medium Mica (Low Concentration) 36
Plugging Materials, Medium and Coarse CaCO3 (Low Concentration) 37
Regular Cross-Linked Cement (RCC) 38
Resin-Coated Sand Pill 39
Reverse Diesel Oil Bentonite (RDOB) 40
Right Angle Set Polymers Cement 41
Soft and Hard Plugs Special Fibers Cement 42
Super Stop Material 43
Thermoset Rubber-Medium with Ground Walnut Shells-Coarse and a Blend of Fibers, Flakes and Granules (Coarse Grade) 44
Various Size and Types of Conventional LCM’s (Granules, Flakes and Fibers) (Low Concentration) 45
Waiting Method 46

6. Conclusion sources. Machine learning was implemented to assist in the decision-

Lost circulation is a difficult problem that has been encountering
drilling operations. Choosing the best treatment for lost circulation is
not a straightforward process and requires a high level of knowledge
and experience. Large data set of 3000 wells was collected from many
making process. Based on this study, the following conclusions were
made:
1 After applying multiple machine learning methods, quadratic SVM
was selected to train the model since it gave the highest accuracy.
SVM has a history of great performance in classification problems.

9
H.H. Alkinani, A.T.T. Al-Hameedi and S. Dunn-Norman
5-fold cross-validation was used to avoid overfitting and poor gen-
eralization.
2 QC/QA was implemented in the process of selecting the lost circu-
lation treatments to limit the error due to incorrect classification.
This ensures more accuracy in selecting the best treatment for lost
circulation.
3 The main decision-making criteria were based on whether the treat-
ment will be successful or not. The cost is a pivotal parameter in the
decision-making process. However, a couple of thousand dollars is
not significant compared to the cost of the entire drilling operations
to choose between the treatments.
4 Two treatments recommended to stop partial loss, four for severe
loss, and seven for complete loss, depending on the reason of loss.
5 With the large historical data available in the oil and gas industry,
machine learning can be utilized to make better future decisions that
will optimize the drilling operations and save time and money.
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
Acknowledgment
The authors would like to thank Basra Oil Company from Iraq for
providing us with real field data.
Appendix
Table 5 shows the treatments used in this study with their reference
number (NO).
References
[1] Al-Hameedi ATT, Alkinani HH, Dunn-Norman S, Flori RE, Hilgedick SA, Amer AS,
et al. Mud loss estimation using machine learning approach. J Pet Explor Prod Tech-
nol 2019. https://doi.org/10.1007/s13202-018-0581-x.
[2] Al-Hameedi ATT, Alkinani HH, Dunn-Norman S, Flori RE, Hilgedick SA,
Alkhamis MM, et al. Predictive data mining techniques for mud losses mitigation
Soc. pet. eng. – SPE Kingdom Saudi Arab. annu. tech. symp. exhib. 2018, SATS
2018; 2018.
[3] Al-Hameedi ATT, Alkinani HH, Dunn-Norman S, Flori RE, Hilgedick SA, Alsaba MT,
et al. Mud losses estimation using partial least squares algorithm Soc. pet. eng. –
Abu Dhabi int. pet. exhib. conf. 2018, ADIPEC 2018; 2019.
[4] Al-Hameedi ATT, Alkinani HH, Dunn-Norman S, Flori RE, Hilgedick SA,
Alkhamis MM, et al. Data analysis of lost circulation events in the Hartha forma-
tion, Rumaila field, Iraq Soc. pet. eng. – SPE Kingdom Saudi Arab. annu. tech. symp.
exhib. 2018, SATS 2018; 2018.
[5] Xu C, Yan X, Kang Y, You L, You Z, Zhang H, et al. Friction coefficient: a significant
parameter for lost circulation control and material selection in naturally fractured
reservoir. Energy 2019. https://doi.org/10.1016/j.energy.2019.03.017.
[6] Yang M, Li X, Deng J, Meng Y, Li G. Prediction of wellbore and formation tem-
peratures during circulation and shut-in stages under kick conditions. Energy 2015.
https://doi.org/10.1016/j.energy.2015.09.001.
[7] Nasiri A, Ghaffarkhah A, Keshavarz Moraveji M, Gharbanian A, Valizadeh M.
Experimental and field test analysis of different loss control materials for
combating lost circulation in bentonite mud. J Nat Gas Sci Eng 2017.
https://doi.org/10.1016/j.jngse.2017.04.004.
[8] Alkinani HH, Al-Hameedi ATT, Dunn-Norman S, Flori RE, Hilgedick SA,
Amer AS, et al. Economic evaluation and uncertainty assessment of lost cir-
culation treatments and materials in the hartha formation, Southern Iraq Soc.
pet. eng. – SPE Asia pacific oil gas conf. exhib. 2018, APOGCE 2018; 2018.
https://doi.org/10.2118/192097-ms.
[9] Mansour AK, Taleghani AD, Li G. Smart lost circulation materials for wellbore
strengthening. 51st US rock mech. / geomech. symp. 2017; 2017.
[10] Transparency Market Research. Drilling fluids market (Type: water-based muds, oil-
based muds, and synthetic-based muds; and application: onshore and offshore) –
global industry analysis, size, share, growth, trends, and forecast 2018–2026. 2013.
[11] Caenn R, Darley HCH, Gray GR. Composition and properties of drilling and comple-
tion fluids. 2011. https://doi.org/10.1016/C2009-0-64504-9.
[12] Messenger JU. Lost circulation. PennWell Publishing Company; 1981.
[13] Alkinani HH, Al-Hameedi ATT, Dunn-Norman S, Flori RE, Alsaba MT, Amer AS,
et al. Using data mining to stop or mitigate lost circulation. J Pet Sci Eng 2019.
https://doi.org/10.1016/j.petrol.2018.10.078.
10
Energy and AI 2 (2020) 100031
[14] Nayberg TM, Petty BR. Laboratory study of lost circulation materials for use in
oil-base drilling muds. Soc Pet Eng AIME SPE 1986.
[15] Basra Oil Company. Various daily reports, Final Reports, and Tests for 2007, 2008,
2009, 2010, 2011, 2012, 2013, 2014, 2015, 2016, 2017, and 2018. Several Drilled
Wells, Basra’s Oil Fields, Iraq.
[16] Howard GC, Scott PP Jr. An analysis and the control of lost circulation. J Pet Technol
1951;3:171–82. https://doi.org/10.2118/951171-G.
[17] Messenger JU, McNiel JS. Lost circulation corrective: time-setting clay cement. J Pet
Technol 1952. https://doi.org/10.2118/148-g.
[18] Morita I, Black AD, Fuh GF. Theory of lost circulation pressure. In: Proc. – SPE annu.
tech. conf. exhib.; 1990 https://doi.org/10.2523/20409-ms.
[19] Fidan E, Babadagli T, Kuru E. Use of cement as lost-circulation ma-
terial: best practices. Can. int. pet. conf. 2004, CIPC 2004; 2004.
https://doi.org/10.2118/2004-090.
[20] Mata F, Veiga M. Crosslinked cements solve lost circulation problems. In: Proc. –
SPE annu. tech. conf. exhib.; 2004 https://doi.org/10.2118/90496-ms.
[21] Saasen A, Wold S, Ribesen BT, Tran TN, Huse A, Rygg V, et al. Permanent aban-
donment of a North Sea well using unconsolidated well-plugging material. SPE Drill
Complet 2011. https://doi.org/10.2118/133446-PA.
[22] Caughron DE, Renfrow DK, Bruton JR, Ivan CD, Broussard PN, Bratton TR,
et al. Unique crosslinking pill in tandem with fracture prediction model
cures circulation losses in deepwater gulf of Mexico. Proc Drill Conf 2002.
https://doi.org/10.2523/74518-ms.
[23] Bruton JR, Ivan CD, Heinz TJ. Lost circulation control: Evolving tech-
niques and strategies to reduce downhole mud losses. Proc Drill Conf 2001.
https://doi.org/10.2523/67735-ms.
[24] Collins N, Kharitonov A, Whitfill D, Miller M, Kulakov K. Comprehensive approach
to severe lost circulation problems in Russia. Soc. pet. eng. – SPE Russ. oil gas tech.
conf. exhib. 2010, RO G 10; 2010. https://doi.org/10.2118/135704-ms.
[25] Wang H, Sweatman R, Engelman B, Deeg W, Whitfill D, Soliman M, et al. Best prac-
tice in understanding and managing lost circulation challenges. SPE Drill Complet
2008. https://doi.org/10.2118/95895-pa.
[26] Wang H, Sweatman R, Engelman B, Deeg W, Whitfill D. The key to successfully
applying today’s lost-circulation solutions. In: Proc. – SPE annu. tech. conf. exhib.;
2005 https://doi.org/10.2523/95895-ms.
[27] Whitfill D, Wang H. Making economic decisions to mitigate lost circulation. In: Proc.
– SPE annu. tech. conf. exhib.; 2005.
[28] Lécolier E, Herzhaft B, Rousseau L, Néau L, Quillien B, Kieffer J. Development of
a nanocomposite gel for lost circulation treatment. SPE – Eur. form. damage conf.
proceedings. EFDC; 2005. https://doi.org/10.2523/94686-ms.
[29] Alkinani HH, Al-Hameedi ATT, Dunn-Norman S, Flori RE, Hilgedick SA, Al-
Maliki MA, et al. Examination of the relationship between rate of penetration and
mud weight based on unconfined compressive strength of the rock. J King Saud Univ
– Sci 2018. https://doi.org/10.1016/j.jksus.2018.07.020.
[30] Russell S, Norvig P. Artificial intelligence a modern approach. 3rd ed; 2010
https://doi.org/10.1017/S0269888900007724.
[31] Chio C, Freeman D. Machine learning and security: protecting systems with data and
algorithms. O’Reilly Media; 2018.
[32] Haykin S. Neural networks and learning machines. 2008. https://doi.org/978-
0131471399.
[33] Cortes C, Vapnik V. Support-vector networks. Mach Learn 1995;20:273–97.
https://doi.org/10.1007/BF00994018.
[34] Winston P. 6.034 artificial intelligence. Massachusetts Inst Technol MIT Open-
CourseWare 2010. https://ocw.mit.edu.
[35] Wang G, Pu X-L, Tao H-Z. A support vector machine approach for the prediction
of drilling fluid density at high temperature and high pressure. Pet Sci Technol
2012;30:435–42. https://doi.org/10.1080/10916466.2011.578095.
[36] Al-Anazi A, Gates ID. A support vector machine algorithm to classify litho-
facies and model permeability in heterogeneous reservoirs. Eng Geol 2010.
https://doi.org/10.1016/j.enggeo.2010.05.005.
[37] Deng C, Pan H, Fang S, Konaté AA, Qin R. Support vector machine as an alternative
method for lithology classification of crystalline rocks. J Geophys Eng 2017;14:341–
9. https://doi.org/10.1088/1742-2140/aa5b5b.
[38] Perez HH, Datta-Gupta A, Mishra S. The role of electrofacies, lithofacies, and
hydraulic flow units in permeability predictions from well logs: a compara-
tive analysis using classification trees. SPE Reserv Eval Eng 2005;8:143–55.
https://doi.org/10.2118/84301-PA.
[39] Sabah M, Talebkeikhah M, Agin F, Talebkeikhah F, Hasheminasab E. Application
of decision tree, artificial neural networks, and adaptive neuro-fuzzy inference sys-
tem on predicting lost circulation: a case study from Marun oil field. J Pet Sci Eng
2019;177:236–49. https://doi.org/10.1016/j.petrol.2019.02.045.
[40] Alkinani HH, Al-Hameedi ATT, Dunn-Norman S, Alsaba MT, Amer AS. Re-
view of the applications of decision tree analysis in petroleum engineer-
ing with a rigorous analysis. SPE Middle East Oil Gas Show Conf 2019:9.
https://doi.org/10.2118/194795-MS.
[41] Mao S, Chen JW, Jiao L, Gou S, Wang R. Maximizing diver-
sity by transformed ensemble learning. Appl Soft Comput J 2019.
https://doi.org/10.1016/j.asoc.2019.105580.
[42] MATLAB. Statistics and machine learning toolbox release notes. Matlab 2019.
[43] Anifowose FA, Labadin J, Abdulraheem A. Ensemble machine learning: an untapped
modeling paradigm for petroleum reservoir characterization. J Pet Sci Eng 2017.
https://doi.org/10.1016/j.petrol.2017.01.024.
[44] Tufféry S. Data mining and statistics for decision making. 2011.
https://doi.org/10.1002/9780470979174.
[45] Aldhaheri M, Wei M, Bai B, Alsaba M. Development of machine learning method-
H.H. Alkinani, A.T.T. Al-Hameedi and S. Dunn-Norman
ology for polymer gels screening for injection wells. J Pet Sci Eng 2017.
https://doi.org/10.1016/j.petrol.2016.12.038.
[46] Hegde C, Millwater H, Gray K. Classification of drilling stick slip severity using ma-
chine learning. J Pet Sci Eng 2019. https://doi.org/10.1016/j.petrol.2019.05.021.
[47] Klyuchnikov N, Zaytsev A, Gruzdev A, Ovchinnikov G, Antipova K, Ismailova L, et al.
Data-driven model for the identification of the rock type at a drilling bit. J Pet Sci
Eng 2019. https://doi.org/10.1016/j.petrol.2019.03.041.
[48] Mohaghegh S. Virtual-intelligence applications in petroleum engineering: part I –
artificial neural networks. JPT J Pet Technol 2000.
[49] Muojeke S, Venkatesan R, Khan F. Supervised data-driven approach to early
kick detection during drilling operation. J Pet Sci Eng 2020;192:107324.
https://doi.org/10.1016/j.petrol.2020.107324.
[50] Adedigba SA, Khan F, Yang M. Dynamic failure analysis of process sys-
tems using neural networks. Process Saf Environ Prot 2017;111:529–43.
https://doi.org/10.1016/j.psep.2017.08.005.
[51] Shi J, Zhu Y, Khan F, Chen G. Application of Bayesian Regularization Artificial Neu-
ral Network in explosion risk analysis of fixed offshore platform. J Loss Prev Process
Ind 2019;57:131–41. https://doi.org/10.1016/j.jlp.2018.10.009.
[52] Chamkalani A, Zendehboudi S, Amani M, Chamkalani R, James L, Dusseault M.
Pattern recognition insight into drilling optimization of shaly formations. J Pet Sci
Eng 2017;156:322–39. https://doi.org/10.1016/j.petrol.2017.05.017.
[53] Jahanbakhshi R, Keshavarzi R, Jalili S. Artificial neural network-based pre-
diction and geomechanical analysis of lost circulation in naturally frac-
tured reservoirs: a case study. Eur J Environ Civ Eng 2014;18:320–35.
https://doi.org/10.1080/19648189.2013.860924.
[54] Li H, Misra S, He J. Neural network modeling of in situ fluid-filled pore size distri-
butions in subsurface shale reservoirs under data constraints. Neural Comput Appl
2020;32:3873–85. https://doi.org/10.1007/s00521-019-04124-w.
[55] Onalo D, Oloruntobi O, Adedigba S, Khan F, James L, Butt S. Dynamic data driven
sonic well log model for formation evaluation. J Pet Sci Eng 2019;175:1049–62.
https://doi.org/10.1016/j.petrol.2019.01.042.
[56] He J, Li H, Misra S. Data-driven in-situ sonic-log synthesis in shale reser-
voirs for geomechanical characterization. SPE Reserv Eval Eng 2019;22:1225–39.
https://doi.org/10.2118/191400-PA.
[57] Alkinani HH, Al-Hameedi ATT, Dunn-Norman S, Flori RE, Alsaba MT, Amer AS.
Applications of artificial neural networks in the petroleum industry: a review. In:
SPE middle east oil gas show conf; 2019. p. 12.
[58] Alberty MW, McLean MR. A physical model for stress cages. In: SPE annu tech conf
exhib; 2004. p. 8.
[59] Fidan E, Babadagli T, Kuru E. Use of cement as lost circulation material – field case
studies. In: IADC/SPE asia pacific drill technol conf exhib; 2004. p. 9.
[60] Friedheim JE, Young S, De Stefano G, Lee J, Guo Q. Nanotechnology for oilfield
applications – hype or reality?. In: SPE int oil nanotechnol conf exhib; 2012. p. 7.
[61] Gallardo Falero H, Cassanelli JP, Barrett S, Romero PD, Mufarech Rey A. Casing–
Drilling Technology (CwD) mitigates fluid losses in Peruvian jungle. In: SPE Lat Am
Caribb pet eng conf; 2010. p. 11.
[62] Hoelscher KP, De Stefano G, Riley M, Young S. Application of nanotechnology in
drilling fluids. In: SPE int oilf nanotechnol conf exhib; 2012. p. 7.
[63] Jiao D, Sharma MM. Mud induced formation damage in fractured reservoirs. In: SPE
drill complet, 11; 1996. p. 11–16.
[64] Kumar A, Savari S, Whitfill D, Jamison DE. Wellbore strengthening: the less-studied
properties of lost-circulation materials. In: SPE annu tech conf exhib; 2010. p. 13.
[65] Nwaoji CO, Hareland G, Husein M, Nygaard R, Zakaria MF. Wellbore strengthening-
nano-particle drilling fluid experimental design using hydraulic fracture apparatus.
In: SPE/IADC drill conf; 2013. p. 12.
[66] Pilehvari AA, Nyshadham VR. Effect of material type and size distribution on per-
formance of loss/seepage control material. In: Int symp exhib form damage control;
2002. p. 13.
[67] Ramirez M, Diaz A, Luna E, Figueroa Y, Per-Bjarte T. Successful application of
synthetic graphite to overcome severe lost circulation problem in the troublesome
foothills of Colombia. AADE Natl. Tech. Conf. Exhib. Houston, Texas; 2005.
[68] Al Menhali S, Kashwani G, Sahwani A. Construction material, cement-
ing, lost circulation, light weight cement; construction material, cement-
ing, lost circulation, light weight cement. Int J Mater Eng 2015;2015:46–9.
https://doi.org/10.5923/j.ijme.20150503.02.
[69] Riley M, Young S, Stamatakis E, Guo Q, Ji L, De Stefano G, et al. Wellbore stability
in unconventional shales – the design of a nano-particle fluid. In: SPE oil gas India
conf exhib; 2012. p. 8.
[70] Salehi S, Nygaard R. Evaluation of new drilling approach for widening operational
window: implications for wellbore strengthening. In: SPE prod oper symp; 2011.
p. 16.
11
Energy and AI 2 (2020) 100031
[71] Salehi S, Nygaard R. Numerical modeling of induced fracture propagation: a novel
approach for Lost Circulation Materials (LCM) design in borehole strengthening ap-
plications of deep offshore drilling. In: SPE annu tech conf exhib; 2012. p. 12.
[72] Sanders MW, Scorsone JT, Friedheim JE. High-fluid-loss, high-strength lost circula-
tion treatments. In: SPE deep drill complet conf; 2010. p. 6.
[73] Savari S, Whitfill DL, Kumar A. Resilient Lost Circulation Material (LCM): a signifi-
cant factor in effective wellbore strengthening. In: SPE deep drill complet conf; 2012.
p. 7.
[74] Soroush H, Sampaio JHB, Nakagawa EY. Investigation into strengthening methods
for stabilizing wellbores in fractured formations. In: SPE annu tech conf exhib; 2006.
p. 8.
[75] El-Hassan HI, Abbas R, Jarouj H, Munk T. Using a novel fiber cement system to
control lost circulation: case histories from the middle east and the far east. In:
SPE/IADC middle east drill technol conf exhib; 2003. p. 7.
[76] Tessari RM, Warren TM, Jo JY. Drilling with casing reduces cost and risk. SPE Russ
oil gas tech conf exhib 2006:9. https://doi.org/10.2118/101819-RU.
[77] Tare UA, Whitfill DL, Mody FK. Drilling fluid losses and gains: case histories and
practical solutions. In: SPE annu tech conf exhib; 2001. p. 8.
[78] van Oort E, Friedheim JE, Pierce T, Lee J. Avoiding losses in depleted and weak
zones by constantly strengthening wellbores. In: SPE annu tech conf exhib; 2009.
p. 14.
[79] Arshad U, Jain B, Ramzan M, Alward W, Diaz L, Hasan I, et al. Engineered solution
to reduce the impact of lost circulation during drilling and cementing in Rumaila
field, Iraq. In: Int pet technol conf; 2015. p. 17.
[80] Wang H. Is it really possible to efficiently form a strong seal inside subterranean
openings without knowing their shape and size?. AADE natl. tech. conf. exhib. Hous-
ton, Texas; 2011.
[81] Whitfill D. Lost circulation material selection, particle size distribution and fracture
modeling with fracture simulation software. In: IADC/SPE asia pacific drill technol
conf exhib; 2008. p. 12.
[82] White RJ. Lost-circulation materials and their evaluation. In: Drill prod pract; 1956.
p. 8.
[83] Wang H, Soliman M, Whitfill DL, Towler BF. Case histories show wellbore strength-
ening as a cost-effective option for drilling with narrow mud weight windows. ADE
fluids conf. exhib.; 2008.
[84] Wang H, Soliman MY, Towler BF, Shan Z. Strengthening a wellbore with multiple
fractures: further investigation of factors for strengthening a wellbore. In: 43rd US
rock mech symp 4th US - Canada rock mech symp; 2009. p. 9.
[85] Alkinani HH, Al-Hameedi ATT, Dunn-Norman S, Al-Alwani MA, Mutar RA, Al-Baz-
zaz WH. State-of-the-art review of lost circulation materials and treatments – part I:
general trends and uses. In: Abu Dhabi int pet exhib conf; 2019. p. 29.
[86] Alkinani HH, Al-Hameedi ATT, Dunn-Norman S, Al-Bazzaz WH. State-of-the-art re-
view of lost circulation materials and treatments – part II: probability and cost anal-
yses. In: Int pet technol conf; 2020. p. 20.
[87] Ali A, Kalloo CL, Singh UB. A practical approach for preventing lost circulation
in severely depleted unconsolidated sandstone reservoirs. In: SPE drill complet, 9;
1994. p. 32–8.
[88] Bourgoyne AT, Millheim KK, Chenevert M, Young FS. Applied drilling engineering.
Society of Petroleum Engineers; 1986.
[89] Canson BE. Lost circulation treatments for naturally fractured, vugular, or cavernous
formations. In: SPE/IADC drill conf; 1985. p. 12.
[90] Carlton LA, Chenevert ME. A new approach to preventing lost returns. In: Fall meet
soc pet eng AIME; 1974. p. 15.
[91] Contreras Puerto O. Wellbore strengthening by means of nanoparticle-based drilling
fluids. University of Calgary; 2014.
[92] Davison JM, Leaper R, Cauley MB, Bennett B, Mackenzie A, Higgins CJ, et al. Extend-
ing the drilling operating window in BRENT: solutions for infill drilling in depleting
reservoirs. In: IADC/SPE drill conf; 2004. p. 10.
[93] Pomares A, Martínez JL, Mandow A, Martínez MA, Morán M, Morales J. Ground
extraction from 3D LIDAR point clouds with the classification learner app. In: 2018
26th mediterr. conf. control autom.; 2018. p. 1–9.
[94] Alpaydin E. Introduction to machine learning (Adaptive Computation and Machine
Learning). The MIT Press; 2004.
[95] Fernández-Delgado M, Cernadas E, Barro S, Amorim D. Do we need hundreds of clas-
sifiers to solve real world classification problems? J Mach Learn Res 2014;15:3133–
81. https://doi.org/10.1117/1.JRS.11.015020.