IADC/SPE 98378

Stuck Pipe Prediction and Avoidance: A Convolutional Neural Network Approach

C. Siruvuri, Halliburton Digital and Consulting Solutions; S. Nagarakanti, Nabors Industries; and R. Samuel, Halliburton
Digital and Consulting Solutions

Copyright 2006, IADC/SPE Drilling Conference

This paper was prepared for presentation at the IADC/SPE Drilling Conference held in Miami,
Florida, U.S.A., 2123 February 2006.
This paper was selected for presentation by an IADC/SPE Program Committee following
review of information contained in a proposal submitted by the author(s). Contents of the
paper, as presented, have not been reviewed by the International Association of Drilling
Contractors or Society of Petroleum Engineers and are subject to correction by the author(s).
The material, as presented, does not necessarily reflect any position of the IADC, SPE, their
officers, or members. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the International Association of Drilling
Contractors and Society of Petroleum Engineers is prohibited. Permission to reproduce in print
is restricted to an abstract of not more than 300 words; illustrations may not be copied. The
abstract must contain conspicuous acknowledgment of where and by whom the paper was
presented. Write Librarian, SPE, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A.,
fax 1.972.952.9435.

Abstract
Refining advanced technologies for successful completion of
wells is paramount in high-risk, high-cost environments such
as ultra-deep water. Challenges are still very much associated
with extreme depths, high temperatures, and with the small
differences between formation pore pressure and pressures
that will unintentionally frac a formation. Drilling mud
pressures outside this narrow range can readily lead to
additional problems such as stuck pipe and wellbore
instability.
Comparatively little research has been done to predict
stuck pipe, and there are few references for quantifying
estimates of force needed to free stuck pipe. Traditionally,
stuck pipe problems are solved by using available standard
methods and by post-sticking, ad hoc techniques. However,
the real key to savings and success is in avoiding the risks
leading to stuck pipe occurrences. If these risks are identified
in advance, procedures can be set up to reduce the possibility
of stuck pipe and, therefore, to reduce the costs associated
with freeing stuck pipe.
This paper presents an application of neural network
methods for understanding the causes of differentially stuck
pipe. The method is proposed as a more effective prognostic
tool than are currently available procedures. The methodology
enables drilling industry personnel to estimate the risk of
occurrence of stuck pipe not only during well planning
procedure but also during drilling. A three layer feed-forward
network with full connection topology between layers is used.
Sigmoid type non-linear activation functions and an error,
back propagation learning rule are used. The convolutional
analysis of the model is based on the constraints of different
drilling variables. The trained neural net is shown to be able to
automatically detect variables of concern.
Extensive simulations have been carried out. The paper
also includes three case studies of stuck pipe incidents from

Gulf of Mexico where the neural network methodology

successfully predicted pipe sticking and also suggested the
procedure and forces necessary to free the pipe.
Introduction
Over several years oil industry is facing troubles associated
with the stuck pipe. Differential pipe sticking is one of the
stuck pipe mechanisms that have had a major impact on
drilling efficiency and well costs1-3. These occurrences are
common everywhere in the world and are estimated to cost the
industry hundreds of millions of dollars annually. In some
areas, events related to differentially stuck pipe can be
responsible for as much as 40% of the total well cost.
Differential pipe sticking problems generally result in the well
cost overruns and time overruns as a non-productive time in
terms of loss of rig days either due to stopping of drilling
operations or an attempt to free the stuck pipe. This huge loss
is always accounted in the well budget cost as a contingency
factor for the risks associated with the stuck pipe problems in
the well planning and drilling performance approach.
Differential pipe sticking occurs when a part of the drill
string, casing, or logging tool becomes embedded in a mud
solids filter cake and is held there by a significant amount of
differential pressure. This differential pressure is the pressure
difference between the hydrostatic pressure of mud and the
formation pore pressure. Usually, because of the excessive
differential pressure, the sticking takes place across porous
and permeable formations such as sandstone or limestone,
where a mud filter cake builds up during drilling. It does not
occur in very low permeability formations such as shales,
where mud filter cakes normally do not form. Stuck pipe is
identified as an impedance of drilling mud flow in the annular
space and the difficulty of the pipe movement either in the
upward or downward direction. In a complete stuck pipe
situation, neither circulation nor pipe movement are possible.
Methodology
Traditionally, differential stuck pipe situations were solved
using standard tools and techniques developed in oil field
practice. There have also been recommendations to study the
causes of drill pipe sticking and thereafter to suggest changes
in operating practices to minimize stuck pipe incidents. In
early 1980s and 1990s, various researchers1-5 believed that
the real key to savings and success in a well planning was in
the avoidance of the risks associated with the stuck pipe
incidences. Several studies5-8 in the past have been performed

to identify important parameters associated with the

occurrence of differentially stuck pipe. Among them are
multivariate statistical analysis techniques and simulated
sticking tests using different drilling fluids. Both kinds of
studies have been used to identify and modify variables that
lead to stuck pipe in order to minimize or prevent sticking.
This paper introduces a more powerful and an efficient
application called neural network modeling, a technique that
can provide better and more accurate solutions for the
problems associated with differential pipe sticking events.
Neural networks, with their remarkable ability to derive
meaning from complicated or imprecise data, can be used to
extract patterns and detect trends that are too complex to be
noticed by either humans or other computer techniques. A
trained neural network can be thought of as an "expert" in the
category of information it has been given to analyze. This
expert can then be used to provide projections given new
situations of interest and answer "what if" questions. The
current neural network model has a large capacity to
incorporate a great many drilling parameters and relatable
sticking fluid characteristics for the better prediction of stuck
pipe. This artificial neural network model has an interesting
feature of assigning some precise values for the missing
parameters and assigning some uncertainties in predicting the
desired output. The main objectives of this paper is to provide
the user a general workflow procedure involved in the better
prediction of the stuck pipe conditions using a neural network.
More details on artificial network/neural network can be found
elsewhere.
Following a background review, development of the
database model and the relevant drilling and the fluid mud
parameters required for the study are described. Further, the
description of the neural network along with the results is
presented with an example application.
Database Assembly
Database preparation for the training of the neural network
represents a crucial step in the neural network modeling. The
performance of the network model rests solely upon the input
training pattern. The data used in this study are selected in a
random manner, according to the specifications given in the
literature, to predict differentially stuck pipe conditions. Ten
input parameters belonging to drilling and fluid mud
characteristics are analyzed in an effort to recognize the
environments conducive to a high risk of differentially stuck
pipe. The database model is constructed by formulating 200
datasets representing 120 differentially stuck pipe incidents
and 50 non-stuck pipe incidents. Each of these datasets has
been classified either as stuck or not-stuck and used as training
sets for the neural network model preparation. Thirty-five
rows of data sets were used for cross validation purposes. The
remaining dataset rows were used for testing purpose.
Two separate database models were prepared to conduct
studies for water-based and oil-based sticking phenomenon.
The water-based mud database model includes drilling
parameters such as differential pressure, hole depth, and water
based mud properties such as API fluid loss, MBT, chlorides,
total hardness, PV of drilling mud, YP, gels, inhibitor
concentration, and ph. The oil-based mud database model
includes the drilling parameters, such as hole depth, and oil

IADC/SPE 98378

based mud properties such as PV, YP, gels, emulsion ability,

HPHT fluid loss, lime, chlorides, oil-water ratio.
All these parameters have been developed in such a way
that they are largely independent of one another. The
independence in the variable selection makes it possible to
predict the mechanisms for differentially stuck pipe
mechanism. Commercial NN software was used to analyze
this large database model. The average mean squared error
(MSE) is given in the results section. The parameter selection
in this model is designed to represent the actual physical
phenomenon of stuck pipe.
Construction of Neural Network
Artificial Neural Networks (ANNs) are not new; they were
introduced half a century ago, creating together with
knowledge-based expert systems (KBESs) the field of
artificial intelligence (AI). NNs loosely mimic in structure and
behavior the human brain. Essentially, ANNs consist of a
number of simple processing elements (PEs) or artificial
neurons linked with a set of interconnections representing,
respectively, neural cells or neurons and their axons or
semiconductors in the human brain. The processing elements
are arranged and organized in different forms, depending on
the type of network and its paradigm. These neurons are
organized in the form of layers. In the simplest architecture of
an ANN, there is an input layer, one or more hidden layers,
and an output layer. The Neural Network model (Fig. 1) can
be as shown as follows.

Fig. 1Flow of Information in the Network.

Neural networks have the capability of learning from a

number of input patterns (representing different problem
encounters) and their associated output patterns (representing
the conclusions and decisions). See Fig. 2. During the process
called training, the network generalizes the knowledge and
becomes capable of providing solutions to the new problems.
Once a network is trained using an adequately representative
training set, it can be used to classify or to predict the output
of the modeled system for a given input pattern. ANNs are
very useful in providing the meaningful answers even when
the data to be processed include errors or are incomplete.
ANNs can also process information extremely rapidly when
applied to real world problems.

IADC/SPE 98378

are tested by taking a set of known inputs and pertinent

outputs and calculating the difference between the known
output of the training sets and the output generated by the
neural model. It is denoted as follows:

E =T A

Fig. 2Topology of a Neural Network.

In an ANN, each neuron receives information from other

neurons; process it through an activation function, and
produces output to other neurons. The output of a neuron k,
Yk, is given by

Yk = F Wkj x j + bk
j =1

(1)

Where,
F = activation function
xj = inputs to the neuron
wkj = weights
bk = extremely applied bias
Before an ANN can be used to perform its task, it should
be trained to do so. This training or learning process is simply
to determine the weights and biases using an appropriate
learning algorithm. A neural network derives its computing
power through its massive parallel-distributed structure and its
ability to learn and generalize. Generalization refers to the
ability of ANNs to produce reasonable outputs for inputs not
encountered during training.
Weights and biases are used to control the influence of a
certain input parameter. The weight represents the correlation
between the input and the output. An input parameter that will
have a large influence of the result is assigned a large weight
magnitude. Weights can also be negative denoting a negative
correlation. Biases are used to shift the origin of the influence.
For example, imagine an input that gives a negative influence
if the magnitude is below 35 and a positive influence above
35. Adding a bias, -35 will then move the origin of influence
to 0. By using biases the origin of influence is same for all
neurons and the same transfer functions are used for all
neurons. Transfer functions are used to transform the input of
a neuron to an output. Two common transfer functions are
used for transformation (1) non sigmoid and (2) linear
function. The selection depends on the complexity of the
problem. A neural network program is designed in such a way
that this is performed automatically. The training algorithm is
generated by assigning arbitrarily weights and biases. Results

(2)

Where, E contains the error of all output parameters for all

training sets. T contains the targets (known output
parameters), and A contains the output predicted by the neural
network. The error between the predicted outputs and actual
outputs is used to fine-tune the model and evaluate the
effectiveness of the model to accurately identify the key
differential sticking mechanisms. In this particular case, A
Generalized Feed Forward/Back Propagation training model
(as shown in Fig. 3) is used to reduce the error by adjusting
the weights and biases incorporated in the model. Reduction is
achieved by adding this error to the old weights with new
weights calculated as follows:

w = O + Prev

(3)

Where

= learning rate; = momentum; both between 0 and 1.

O = output from the given layer
= the difference between the output and the expected
output
Prev = is the previous change in weights
The number of the nodes in the input layers and output
layer are determined according to the requirement of the
problem. The number of the nodes in the hidden layer is
maintained atleast at three, and is increased according to the
weight difference in order to get the error difference to a
minimum value.

Fig. 3Generalized Feed Forward Back Propagation.

In order to account for some of the unknown input data

and to minimize the error, an additional regularization form is
introduced which is given as:

C = E + ( w)
Where,
C = regularization parameter

(4)

IADC/SPE 98378

= weighting function
E = error function
Neural networks method of analysis has been used in
various research fields including petroleum; however, the
usage has been limited in the drilling industry.
Neural Network Architecture
Considering the nature of our problem, a simple four-layer
Generalized Feed Forward Neural Network structure was
chosen for the Artificial Neural Network (ANN) model for
analyzing the stuck pipe mechanisms. The first layer, an input
layer, consists of ten Processing Elements (PE). The number
of the PEs in the input is dictated by the problem considered.
Different input data parameters are considered corresponding
to the two database models: oil or water.
The second layer is the hidden layer, and the number of
PEs is automatically assigned according to the strength of the
data.
Finally, the output layer consists of two PEs: Differentially
Stuck or No-stuck. The data relating to these parameters in the
input and output layer is gathered and stored in a spreadsheet
application before being applied to neural network model. For
this case study, Neuro solutions software add-in is
incorporated in the spreadsheet and the Neural Network model
is created. The data in the output layer is normalized to values
between 0 and 1 and later transformed into symbolic output
conditions. Preprocessing is necessary for ensuring the
convergence of neural nets.
Neural Network methods have the ability to represent
complex stuck pipe situations which involve several variables.
The output layer is fully connected to all the units in the
hidden layers as shown in Fig. 4.

Fig. 5Real Time Pre & Post Processing.

In order to make the process and analysis more effective

with a high degree of precision, noisy incoming data is
removed in the pre-processing stage. Furthermore, the real
time feed can be filtered at this stage to a desired end sampling
rate. The real time data coupling architecture is shown in Fig.
5.
Training and Learning
The data in the database spreadsheet model is divided into
training, cross validation, and testing subsets. In this study, the
training subset constitutes 90% of the database. The cross
validation and the testing data make up the remaining 10%,
with each 5% assigned from the dataset. These are prepared
using the tag options available in the Neuro solutions software
network add-in. Later, a supervised training network model is
built using the Create Network option in the add-in. In our
work, a default value of 5000 epochs was factored into the
training of the dataset. An epoch is one sweep through all the
records in the training set. In order to increase the training set
as large as possible, we added distorted data to the training
dataset. The weights were then adjusted during the supervised
run by auto performing gradient descent in weight space. This
particular error minimization criterion implicitly maximizes
the margins for predicting a stuck condition with large
interrelated input variables.
In order to train the network fully for robust reliability,
several runs were made with just a few input variables. Even
though this type of extended training required lot of training
time, it resulted in improved recognition and prediction. Thus,
after the successful training of data set, the neural network
model was released to test the production dataset.

Fig. 4Neural Network Architecture.

In Fig. 4
f(x1.xn)
= goal function
n(x1,x2xi) = activation function in the hidden
layer of n units
x1,x2,x3.xi = input units
wi
= weight of the basis function
The suggested neural network model for the two databases
takes all the data parameters related to the oil-based and
water-based mud properties as input data to classify and
predict differentially stuck-pipe conditions.

Evaluation Criteria and Limitations

To assess the model prediction performance, two quantitative
measures for predicted accuracy were calculated and
examined for training datasets: the minimum training error
(MSE), and the training error at last epoch (Final MSE).
The MSE and Final MSE errors for the oil-based database
model were as high as 0.3 and 0.9, and 0.5.
The training model was cross validated and tested with the
testing dataset, and the percentage errors and coefficient of
variation were observed. Seven quantitative accuracy
measurements were calculated and examined for testing
datasets: Mean Squared Error (MSE), Normalized Mean

IADC/SPE 98378

Squared Error (NMSE, MSE/Variance of desired output),

Mean Absolute Error (MAE), Minimum Absolute Error (Min
Abs Error), Maximum Absolute Error (Max Abs Error),
Linear correlation coefficient (r), and percentage of output
classified correctly (percent correct).All these values for the
database model are listed in the table below.

training and cross validations MSE versus a default number of

epochs. Attempts were made to increase and decrease the
number of hidden layers in the neural network during the
training of the system so as to assign the best weights to the
input data parameters and to obtain a good fit to the data.
Minor changes were observed during these attempts. All the
values calculated during training and testing of the network
model are listed in Table 1. A 3D view of the selected wells is
shown in Fig. 7.

Table 1 Test results for oil-based mud database model

Emean =

1 n
( fapp ( x1....xn ) fact ( x1....xn ) )
n i =1

(5)

And the root mean square (rms) is calculated as below:

1 n
fapp ( x1....xn ) f act ( x1....xn ) E 2mean ) (6)
(

n i =1

Erms =

Where n is the number of input variables. The plot as

shown in Fig. 6 gives the rms error for the tested field data.
This figure gives an uncertainty of the model outcome since
some of the data input from the daily drilling reports are
incomplete.
MSE versus Epoch
0.35
0.3
0.25

MSE

0.2

Training MSE
Cross Validation MSE

0.15
0.1
0.05
0
1

500

999 1498 1997 2496 2995 3494 3993 4492 4991

Epoch (# of runs)

Fig. 6Oil-based mud model.

This uncertainty error can be minimized by coupling the

calculation to the real time feed. A sample plot showing some
of the input variables, the stuck pipe prediction and the root
mean square error tracks are shown.
It is observed that the proposed network was able to
approximate the error within + 5%. Fig. 6 shows the picture of

Fig. 7- 3-D View of the wells with stuckpipe Incidents.

The knowledge attachments in the 3D view shown as

rectangular boxes indicate the stuck pipe incidents that
occurred along the path of the wells during various drilling
events.
Several cases were run with actual data from Gulf of
Mexico to test the stability of the model. Several input
variables from the daily drilling report are used and the few
input and output variables that were used in the neural
network were plotted with depth as the reference as shown in
Fig. 8. The first four columns are the inputs and the last
column is the output in which 0 indicates a no-stuck condition
and 1 indicates a stuck condition. It can be seen that the model
predicted a stuck condition at depth of 14700ft and 17900ft.
which is close to the actual stuck depth of 14685ft and 17913
ft respectively.
Data availability is a potential problem when carrying out
an analysis to predict stuck-pipe. Insufficient amount of data
to train/test and cross validate a network can lead to a poor
generalization. Proper limitations of the hidden layers are
necessary to consider during the construction of the network
model as either less/large number of layers can underestimate
or overestimate the desired output required for a project.
Further research work in the use of different algorithms for
neural network model construction can better enhance the
prediction of differential stuck pipe conditions.

IADC/SPE 98378

Nomenclature
PV = Plastic Viscosity (CP)
YP = Yield Point (lb/100 ft2)
PH = Measurement of the acidity or alkalinity of any
solution,
Lime= Excess lime content (lb/bbl)
Chlorides = Water Phase Salinity (mg/l)
API fluid loss = Filtration characteristic (cc)
Total Hardness = Concentration of Ca & Mg ions in mud
(mg/l)
MBT= level of bentonite equivalent solids in polymer muds
(mg/l)
Emulsion stability = Mud stability (Volts)
HPHT fluid loss = Filtration characteristic (cc)
Acknowledgments
The authors would like to thank the management of
Halliburton for their support of this study and permission to
publish this paper.

Fig 8Plot of the Input and Output for the network.

Conclusions
The results of our study lead to the following conclusions:

Accurate prediction of stuckpipe incidents using the

proposed method
A robust database model yields quite accurate and
feasible results
The accuracy of the predictive model depends on the
size of the database and the variables selected for the
analysis.
A log viewer tracks coupled with WITSML data feed
provides a potential means of monitoring the stuck
condition in real time during drilling.

There is always a huge demand for oil and gas in this

world. As many reservoirs deplete, a significant number of
wells will be drilled; and a large numbers of issues will arise,
differential pipe sticking as one among them. Challenging
drilling will not diminish and will still continue in the future
when there will be more deviated and extended reach wells
through depleted reservoirs. To avoid these losses, the oil
industry must recognize and focus on innovative technologies
used in other scientific areas and try to understand their
applicability to drilling engineering in order to provide better
solutions to problems such as stuck pipe. This paper addresses
the use and applicability of one innovative technology: the use
of simple neural networks to understand the causes of
differentially stuck pipe. This use and application should
enable, in principle, drilling personnel, contractors, and
service companies to understand, in advance, the risks
associated with stuck pipe, understand its possible prevention,
and, if successful, save money related to the issue.

References
1. Neal Adams, Member SPE-AIME, Prentice and
Records Enterprises, Inc. A Field Case Study of
Differential-Pressure Pipe Sticking, SPE 6716,
1977.
2. R.R. Weakley, Chevron Services Inc. Use of Stuck
Pipe Statistics to Reduce the Occurrence of Stuck
Pipe, SPE 20410, 1990.
3. A.P. Wisnie, Conoco Inc., and Zheiwei Zhu,
University. Of Southwestern Louisiana. Quantifying
Stuck Pipe Risk in Gulf of Mexico Oil and Gas
Drilling, SPE 28298, 1994.
4. J.A. Howard, Enertech Engineering & Research Co.,
and S.B. Glover, Enertech Europe. Tracking Stuck
Pipe Probability While Drilling, SPE 27528, 1994.
5. M.W.Biegler, Exxon Production Research Co., and
G.R. Kuhn, Exxon Co USA.., Advances in
Prediction of Stuck Pipe Using Multivariate
Statistical Analysis, SPE 27529, 1994.
6. Commercial Neural Network Software, Neuro
Solutions Version 4.0.http://www.nd.com.
7. Robert J. Boomer, Texaco Exploration and
Production Inc., Predicting Production Using a
Neural Network, SPE 30202, 1995.
8. E&P Exchange, Neural Network: What It Can Do
For Petroleum Engineers.