1st National Iranian Drilling Industry Congress.

Using Neural Network System for Casing Collapse

Occurrence and its Depth Prediction in a Middle
Eastern Carbonate Field

Saeed Salehi , Keivan Khademi Dehkordi , Seyed Farzad Shariatpanahi , Hojatollah Moradi
Petropars Ltd

G.Hareland
University of Calgary

Abstract :
A large carbonate oil field in Iran is suffering from severe casing collapses. 48
casing collapses have been found to be reservoir compaction and poro-elastic effects
and corrosion.
The application of neural networks for predicting casing collapses using complex
multi-dimensional field data has been undertaken. This paper shows how a neural
network (ANN) system can be trained based on the parameters affecting casing
collapse to estimate the potential of collapse of wells to be drilled as well as the
current wells producing in the field. The potential use of this type of analysis is
large in that it can be linked as a critical risking parameter in future field
development analysis. Being able to quantify the potential for collapse of a well in
the future can give management the foundation for a better financial decision
making on what wells and where to drill them with the potential for the larger net
return on the investment. The estimated collapse and corresponding depth could also
benefit in the type of casing design and completion method to be selected as well as
workover designs. Interpretation of the neural network results, together with
engineering judgment, allowed us to conclude that using this method is technically
feasible for predicting casing collapses in this field.

Key Words:
ANN (Artificial Neural Network), PCF (Pound per cubic feet), MW (Mud Weight),
BPNN (Back Propagation Neural Network), GRNN (General Regression Neural
Network)

Introduction :
The field analyzed has been produced since the early 1950’s, but the first casing
collapse was not observed until after 1974. Since then the collapses have increased
in numbers until today. Figure 1 gives an overview of the collapse grouped over the
active failure periods. In this field 48 of totally 267 wells drilled have collapsed or
more then 18 %. This is seen as a serious problem and a predictive tool for
estimating this occurrence is sought. A cross section of the field along the short axis
seen from NE is shown in Figure 2. Based on cross section map it shows that a fault
is present which intersect the G formation members from 2 to 4. This means that
there is a major plane of weakness with a low dip angle present across the entire

www.petroman.ir
 1st National Iranian Drilling Industry Congress.

field. The fault goes nearly to the surface at South East side where there are seen
quite deformed and crushed rocks. From the casing collapse data analysis it is seen
that the casing collapse occur mainly in G formation member 2 to 4. Table 1
summarizes the date of casing completion and casing collapse. From previous
studies there is no single mechanism for casing collapse in this field but rather
combinations of mechanism contributing to the casing collapses. Of these the most
prominent one is reservoir compaction. A field that is depleted may also undergo
reservoir compaction, even if the reservoir rock is relatively stiff. The reservoir
compaction results in increased horizontal stresses in the crestal area of the field,
while in the flank there will be horizontal unloading. This will result in slip along
weak bedding planes as the formation tries to adjust to the reservoir compaction
taking place. Another mechanism contributing to casing collapse is internal and
external corrosion of the casing. The possible cause of corrosion in this field could
be contact with saline water in the G formations due to poor cement jobs, bringing
oxygen to the casing creating a corrosive environment outside the casing. The main
objective of this work was to be able to predict the potential for collapse
occurrences. Data from 20 wells in the field was collected and analyzed using the
neural network method approach. The results from the analysis were potential for
collapse and corresponding collapse depth at different locations around the field on
future wells.

Field Overview :
This field produces from the Asmari reservoir formation, and the structure is oblong,
63km long in the NE-SW direction and 7km wide. The reservoir thickness is on
average 1300 ft of Asmari formation in the crestal area, but due to the dipping strata
the vertical thickness of the reservoir is approximately 2600-2700 ft, depending on
the cut off value used for the reservoir rocks. The initial oil in place of the Asmari
was 43 MMMSTB, with an initial pore pressure of 5887 psi at datum depth of 10991
ft SS. In the initial years the reservoir was drilled with a MW of 72 PCF, while
(1)
around 2002 the reservoir could be drilled with a MW of 53 PCF . That is a
depletion of approximately 1300 psi over a period of 50 years, referenced to the
10991 ff SS datum depth. Current depletion of the reservoir is estimated to 1500-
1800 psi. The average porosity is 12 % in the mainly carbonaceous Asmari
formation.

Rock Mechanical Failure Mechanisms :

Reservoir compaction results in 4 different collapse mechanism; buckling, bending,
traction and shear. In the context of this field it is believed that only buckling and
shear may be relevant.
The creep phenomenon may be important in the relation to the salt layers in the G
formation. Shear due to compaction will occur due to differential loading across
lithology interfaces, especially if it is an interface between hard and soft formations.
The reservoir depletion will directly affect the horizontal stresses due to poro-elastic
response of the reservoir rock. This phenomenon will result in a reduction of the
horizontal stresses in the reservoir rock and these stresses must be taken by the

www.petroman.ir
 1st National Iranian Drilling Industry Congress.

surrounding rocks above and below the reservoir. Rock mechanical properties were
( 2)
used from reference .
Corrosion :
External Corrosion :
External corrosion can be caused by several factors, such as:
¾ Exposure to water zones (salt water in G formation)
¾ Formation differences (e.g. salt, carbonates, sand etc)
¾ The casing acting as an offer anode for surface equipment or other wells
Corrosion due to exposure to water and different formations can be effectively
prevented by a proper cement job, which will protect the casing towards the
exposure and thereby exclude any external corrosion. In the G formation there is
high pressure saltwater which may be creating an electro-potential external
corrosion, especially if there is a poor cement job. This type of corrosion would be
evenly distributed across the field as we expect that the cementing practices are
equally distributed across the field. Corrosion should therefore be a problem which
is found in most wells across the field, also wells in the crestal area. But according
to the analysis corrosive failures are only found at the flanks of the field not in the
crestal area which indicates that there is not just corrosion alone that is fatal but in
conjunction with other parameters more dominant at the flanks like possibly
stresses. The third effect can normally be avoided in on future well by cathodic
protection schemes.

Internal and Annulus Corrosion :

When a well is completed there will normally be a mixture of fluids in the annulus
between the casing and tubing. Typically, there will be a mixture of mud, brine and
sometimes produced oil or gas in the annulus, and the corrosion will be a function of
the composition of this fluid. To prevent such internal corrosion it is normal to use
packer fluids that are specially designed to avoid corrosion. One example to avoid
this could be to use oil based mud drilling the last section before the completion is
put in place.
In case the completion is set without a packer the annulus will be filled with
produced fluid up the fluid level and above there will be wet gas. If the H2S is
present in the gas this will cause corrosion.

Neural Network Approach :

Artificial neural networks have proven to be excellent predictive tools in various
petroleum-engineering applications. Such applications include the prediction of fluid
properties, well logging, well testing and horizontal drilling. The excellent
predictive capability of artificial neural networks comes from the fact that neural
networks have large degrees of freedom that allows them to capture the non-linearity
of the system being studied better than conventional regression techniques. Artificial
neural networks also have the ability to learn and adapt themselves to new situations
in which, if additional data becomes available, the neural networks can be further
trained and refined to include this new data. Furthermore, neural networks can map
a multiple-input multiple-output behavior system. These are some of the qualities
that make neural networks superior to conventional regression techniques. The

www.petroman.ir
 1st National Iranian Drilling Industry Congress.

theory behind artificial neural networks has been the subject of numerous studies. In
an artificial neural network, a training set comprising input and output data is
entered and neural-network algorithms attempt to map the process by which the
inputs become outputs. It is desired that the difference between the predicted and the
observed (actual) outputs be as small as possible. To obtain best results, the two
neural network models that have shown to yield accurate results in various
applications are tested in this study. These neural network algorithms are the Back
Propagation Neural Network (BPNN) and the General Regression Neural Network
(GRNN).
The network architecture affects the optimal training and its ability to generalize.
The general criterion in designing neural networks is to start from the simplest
structure that can provide the essential consistency and adequacy. Following is short
description of these architectures.
BPNN are the most widely used type of artificial neural networks. Typically BPNN
consists of many simple processing elements called neurons grouped in layers and
connected by interconnections called synapses. Figure 3 illustrates a five-layer feed
forward neural network. Many types of the back propagation algorithms
(Levenberg-Marquard, Resilient Back propagation, Gradient Descent, Conjugate
Gradient, Quasi-Newton, etc.) are available. In this study, a three-hidden-layer back
propagation network was utilized. Choosing an appropriate number of hidden
neurons is extremely important aspect in the back propagation networks. However,
there is no exact method for determining the number of hidden layer neurons. Using
too many will increase the training time and may cause the over fitting problem
(Over-fitting is the name given to the case where neural networks memorize the
training pattern rather than generalizing the prediction) On the other hand, using
fewer hidden neurons often increases the likelihood of the learning algorithm to
become trapped in local minima. Therefore, it is imperative that we use absolute
minimum number of hidden neurons that will perform adequately. The best
approach to finding the optimal number of hidden neurons is to start with a small
number, and then slightly increase the number of hidden neurons until no significant
improvement is noted.

Technical Approach :
Preliminary studies indicate that the failed casings are found at the SE and NW
flanks of the long axes of the field, and also at the flanks of the short field axes. No
casing collapse is observed along the crest of the field. Based on these observations,
the candidate wells, “possible future wells”, for neural network analysis were
selected at six different field locations (Figure 4-9);
1. SW flank of field and SW-3 on NE flank,
2. NE-1 & 2 at NW flank & SW-3 on NE flank,
3. NE-2 at NW flank and SW-2 on NE flank,
4. Middle of field NW flank and middle of field SE flank,
5. NE flank of field
6. NE flank of field and middle of SW flank
The five inputs required for the neural network analysis are listed in Table 2. The
corrosion weight factor refers to a normalized value (0-1) that was selected

www.petroman.ir
 1st National Iranian Drilling Industry Congress.

according to internal and external corrosion situations in the possible future selected
wells. Zone factor refers to the selected zone or locations for the future drilling.
(these zones are shown in figure 4-9)
Failure time factor refers to the time interval that the well is to be drilled in and is
discussed in the next section.

Year of Failure :
The first collapse occurred in 1974, according to the provided information, while the
field started producing in the early 1950’s. In other words it took approximately 20
years for the problem to manifest itself. Analyzing the result shows that failures
have taken place in 3 active periods, with quiet periods between them. The first
period covers the interval 1974-1979, the second period covers 1982-1989 while the
third period started in 1991 and lasts to present.
This shows that the first active period is 5 years, the next period increases to 7 years
while the last period is at least 8 years (to 1999) ,but likely longer.
This indicates that it is time dependant processes acting. A process like reservoir
compaction, related to depletion of the reservoir, resulting in stress transfer to the
overburden will comply well with this time dependant behavior.
In the first period also wells across the field collapsed. This holds true for the other
two periods as well, so there is apparently no graphic pattern, except for the
tendency of the failure at the flanks of the short axis of the field. The wells from the
SE flank that collapsed did so in the 3rd active period, i.e. after 1991.
Another conclusion is that number of collapses seems to increase as we go from
period 1 to 3. This is a sign that the problem is escalating over time. This basically
means that the number of casing collapses to be expected in the coming years will
continue to increase. These notions have been considered in neural network analysis
regarding the failure time factor.

Neural Network Results :

Considering the nature of our analysis a simple five layer neural network structure
was chosen for the ANN model to predict the possibility for casing collapse
occurrence and depth. In order to make the process and analysis more effective with
a higher degree of precision, noisy data sets were removed in pre-processing stage.
Using this structure for the previous collapsed wells resulted in only a 5 percent
(4)
error in the real to predicted casing collapse depth .
During this analysis, different values for possibly future wells were analyzed as
different sets of input parameters and the results are shown in table 4 and 5. As
indicated in the tables, probability of collapse and expected depth are changing
according to other input parameters. Nine possible wells with different depths were
analyzed, having different drilling and completion conditions that affect factors like
corrosion, time and zone/location.

Discussion and results :

This study utilizes a neural network model for prediction of casing collapse
occurrence and depth. A spreadsheet program was used for implementation of the

www.petroman.ir
 1st National Iranian Drilling Industry Congress.

neural network model. In this research, three-hidden-layer back propagation

networks were used. The results indicate that;
1-This model provides a competent approximation of casing collapse depth and its
occurrence in future wells.
2-Using this approach is feasible but its accuracy needs to be evaluated in future
works.
3-There is a good match between the real and predicted depth using neural network
in existing wells.
For the future works, if the range of given input or output parameter is not suitable
normalization of data should be performed. For example the data on longitudes and
altitudes of the wells in the field were too close to each other. This prevented the
neural network from being able to find a good correlation between the longitude and
latitude and the output.
It is also suggested in future studies that different sets and combinations of input
parameters should be considered and performance of the neural network using those
different sets and combinations be compared to the current study.
In this analyis only one data setoff input/output variables was chosen and this should
be seen as a first attempt on solving a very complex problem. Even as this is a first
attempt at a complex problem it shows that the method used herein gives reasonable
predictions.

Conclusions :
1-An approach to predicting casing collapse and depth using neural networks is
proposed.
2-The model calibration using offset wells from a carbonate field in the middle-east
generated a model match to collapses and collapse depths within +/- 5 percent of
existing failures.
3-The current ANN model uses five inputs and generates 2 outputs which are
potential for collapse and depth of collapse.
4-This kind of fine tuned model can have a large impact on risking wells for future
field developments. The value of this risk parameter can clearly dictate where the
most optimal economical field development should take place.

Acknowledgement :
The authors would like to thank NISOC (National Iranian South Oil Company) for
their help and permission to publish the results presented in this paper.

References :
1-NISOC, “NISOC R&D Solutions Project #1, Casing Collapse (well integrity);
Phase 1-Concept & Feasibility Study “, June 2005, Ahwaz, Iran.

www.petroman.ir
 1st National Iranian Drilling Industry Congress.

2-Salehi, S., Hareland , G., SPE, University of Calgary, “Wellbore Stability

Analysis in UBD Wells of Iranian Fields “, SPE 105155, 2007.
3-Sadiq, T and Nashavi, I.S., “Using Neural Network for Prediction of Formation
Fracture Gradient “, SPE/ Petroleum Society of CIM 65463, 2000.
4-Siruvuri, C, Haliburton Digital and Consulting Solutions; Nagarakanti, S., Nabors
Industries; Samuel, R., Haliburton Digital and Consulting Solutions, “Stuck Pipe
Prediction and Avoidance: A Convolutional Neural Network Approach “,
IADC/SPE 98378, 2006.

Figure 1: Illustration of collapses over time

Figure 2: Simplified cross section M field-seen from NE, with typical casing designs

Table 1: Review of wells with collapsed casings and liners

www.petroman.ir
1st National Iranian Drilling Industry Congress.

Table 3: Output parameters for neural network analysis

Input Parameters

www.petroman.ir
1st National Iranian Drilling Industry Congress.

Latitude and longitude of the well

Total depth of the well

Corrosion weight factor
Failure time factor
Zone factor

Table 4: Neural network results for different areas

Output parameter
Expected collapse depth
Probability of casing collapse in the next five
years(0-1)

www.petroman.ir
 1st National Iranian Drilling Industry Congress.

Table 5: Neural network results for different areas

assumed De Corrosi Zone Failure Expect Probabili

Assumed Depth Corrosion Zone Failure Expected Probability
well number pth on factor factor (Area) time factor ed collapse ty of collapse
well number (ft) factor factor time factor collapse of collapse
(ft) (0-1) (0-1) depth(ft) (0-1)
(0-1) (Area) (0-1) depth(ft) (0-1)
1 32 0.2 NE 0.6 1950 0.35
56 flank of
field
2 45 0.4 middle 0.7 3241 0.6
21 of field SE
flank
3 38 0.5 middle 0.7 2221 0.7
24 of field SE
flank
4 52 0.5 moving 0.5 2843 0.5
34 NE-2 at NW
flank
5 43 0.4 moving 0.6 3220 0.5
21 NE-2 at NW
flank
6 62 0.4 moving 0.4 4297, 0.2 ,0.4
34 NE-2 at NW 3241
flank
7 54 0.3 middle 0.4 3224 0.4
21 of field SE
flank
8 44 0.6 middle 0.4 2792 0.7
35 of field SE
flank
9 32 0.3 moving 0.3 2654 0.2
56 NE-2 at NW
flank

Table 5: Neural network results for different areas

www.petroman.ir
 1st National Iranian Drilling Industry Congress.

1 3256 0.2 moving 0.6 2432 0.1

NE-2 at NW
flank
2 4521 0.4 SW-3 0.7 3098 0.4
on NE flank
3 3824 0.5 SW-3 0.7 2700 0.7
on NE flank
4 5234 0.5 moving 0.5 3400 0.4
NE-2 at NW
flank
5 4321 0.4 moving 0.6 2987 0.5
NE-2 at NW
flank
6 6234 0.4 SW-3 0.4 3400 0.3
on NE flank
7 5421 0.3 SW-3 0.4 2800 0.2
on NE flank
8 4435 0.6 SW-3 0.4 3200 0.6
on NE flank
9 3256 0.3 SW-3 0.3 1872 0.2
on NE flank

Figure 3:A five layer neural network

www.petroman.ir
 1st National Iranian Drilling Industry Congress.

Figure 4: Position: SW flank of field & SW-3 on NE flank

Figure 5 : Position: Moving NE-1&2 at NW flank & SW-3 on NE flank

Figure 6 : Position :Moving NE-2 at NW flank & SW-2 on NE flank

Figure 7 : Position :Middle of field NW flank & middle of field SE flank

www.petroman.ir
1st National Iranian Drilling Industry Congress.

Figure 8 : Position :NE flank of field

Figure 9 : Position : NE flank of field and middle of SW flank

www.petroman.ir