Society of Petroleum Engineers

SPE 26106

Complexities of Using Neural Network in Well Test Analysis of

Faulted Reservoirs
I.A. Juniardi and Iraj Ershaghi, U. of Southern California'
SPE Members

Copyright 1993, Society of Petroleum Engineers, Inc.

This paper was prepared for presentation at the Western Regional Meeting held in Anchorage, Alaska, U.S.A., 26-28 May 1993.

This paper was selected for presentation by an SPE Program Committee folloWing review of information contained in an abstract submitted by the author(s). Contents of the paper,
as presented, have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the aut~or(s). The ma~erlal, as presen~ed: does no! necessarily ref!ect
any posnion of the Society of Petroleum Engineers, its officers, or members. Papers presented at SPE '!'eellngs are sUbJ9clto publication revIew by EdItorial Commillees of the SOCIety
of Petroleum Engineers. Permission to copy is restricted to an abstract of not more than 300 words. illustrations may not be copIed. The abstract should contain conspicuous acknowledgment
of where and by whom the paper is presented. Write Librarian, SPE, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A. Telex, 163245 SPEUT.

Abstract rameter estimation methods can be used to obtain quantita-

Neural networks models bring a significant dimension to
the capability of computer aided approaches for well test
analysis. This study shows, however, that when dealing with
complex reservoirs, these models need augmentation from an
expert system containing two types of information; indepen-
dent field data and tables containing frequency of occurence
of non-related models. A Monte Carlo Simulation method is
proposed to provide the augmented statistical informations
required for the expert system.
Introduction
Industry wide application of pressure transient test data
has been limited in terms of number, data quality and in-
tegrity of interpretation. Considerable advances in analyt-
ical and numerical modeling of flow problems in reservoirs
have allowed analysts to study the response of an idealized
reservoir geometry of flow regime under the conditions of
specified wellbore flow. Because of many possibilities that
may exist in real situations, continued efforts are underway
to formulate and to obtain the pressure response in reservoirs
of increasing complexity.
The use of computer-aided interpretation technique has
enhanced the capability of well-test analyst in terms of accu-
racy and speed to process and analyze the well-test data.
Many possible analytical reservoir models can be investi-
gated to find the best interpretation model. Once the proper
reservoir model has been determined, various non-linear pa-
oReferences and illustrations at end of paper.
tive description of reservoirs properties.
However, as documented by Ershaghi and Woodbury 1, the
problem of non-uniqueness always creates confusion about
model selection. In particular, the same type of response
may exist among the performance of completely different
systems. The correct estimation of the reservoir parame-
ters, consequently, depends on the integrity of the selected
well test model. Thus, the selection of the reservoir interpre-
tation model and parameter estimation are two major issues
in well-test analysis.
Ramey 2 and Horne3 recently reviewed various diagnostic
methods commonly used in the model selection process. In
the traditional analysis, this task was carried out by present-
ing the observed pressure response in log-log or semilog plots;
Le., Horner plot. As indicated by Horne3 , this approach has
limitations with increasing complexity of the models and
number of controlling parameters. Application of pressure
derivative curves in well-test analysis, proposed by Bourdet
et al. 4, enhances the capability of the analyst to recognize
the reservoir model. Ershaghi et al. 5 indicated the impor-
tance of derivative plots as ISDP (Industry Standard Di-
agnostic Plots) for pattern recognition. Pressure derivative
method requires high-precision data to filter noises associ-
ated with the tool or the differentiation process. Application
of Expert Systems in well-test analysis has been investigated
by Allain and Horne 6 and AI-Kaabi et al. 7 In this approach,
imitating the reasoning process performed by a human ex-
pert, symbolic representation is used to diagnose the im-
age of the reservoir pressure response. Segmentation and

711
2 COMPLEXITIES OF USING NEURAL NETWORK IN WELL TEST ANALYSIS OF FAULTED RESERVOIRS
knowledge- based system are the main parts of the diagnosis
process to identify well-test interpretation model. AI-Kaabi
and Lees studied the application of back-propagation neural
networks to identify the well-test interpretation model. The
main advantage of this approach is the elimination of the
rule based system and its tolerance to the noise.
Once the proper model is selected, parameters estimation
of reservoir, such as permeability, skin, etc., is obtained by
matching the modeJ response to that observed. In traditional
approaches, Horner plot and appropriate type curves are
commonly used in the matching process. Rosa and Borne2 ,
used the non-linear regression in Laplace space to match the
field data and provide confidence interval of the result. As it
has been reviewed by Horne 3 , the advantages of using nonlin-
ear regression are the possibility to have variable rates "nd
the ability to match complex models with more unknown
reservoir parameters. The objective of this method is to
minimize the sum of squares of the differences between the
measured data and the calculated model by adjusting the
unknown reservoir parameters. Newton's method, optimal
control theory, and direct search method are some of the
approaches to achieve this goal.
The purpose of this paper is to focus on the model se-
lection process and to extend previous studies. Specifically
the strength and weaknesses of neural network models in
well-test interpretation technique are examined and a new
approach is proposed to safeguard against selection of inap-
propriate models.
Neural Networks
In recent years, application of neural networks in pattern
recognition relative to petroleum engineering problems has
been addressed by some investigators. Baldwin and OtteS
studied the application of neural networks to well log inter-
pretation. Other studies include Arehart's9 analysis of drill
bits and Rogers et al. 10 proposed application in beam pump-
ing. For well test analysis, AI-Kaabi et al,u indicated that
from the application of back propagation method, neural net-
work can help in identification of well test models.
Formalized first by Werbos and later by Parker, as dis-
cussed by Rumelhart and McClelland,12 and Freeman and
Skapural3 , back-propagation neural network has primary ap-
plications in addressing problems requiring recognition of
complex patterns, fault tolerance or coping with noise, and
performing nontrivial mapping functions.
In this study, a back propagation neural network model
was developed to handle complex pattern recognition prob-
lem associated from well test responses in faulted reservoirs.
The process consists of training the computer with char-
acteristic patterns of known well test models, and generating
the appropriate weight factors for perfect recognition of field
data. For well test studies, the ISDP used was the derivative
712
SPE 026106
plots, normalized and digitized to a set of (x,y) pairs. The
training model consists of three layers of nodes as shown in
Figure 1.
The i level nodes receive the input signal, the k level nodes
generate activation numbers. The two levels are connected
via an intermediate level designated here as the j level. The
training consists of an iterative process where optimal weight
factors among three levels of nodes are generated such that
the activation numbers correspond to the appropriate mod-
els.
As an example, consider Table 1 and 2 where the objective
is to train the computer to recognize any of the ten indicated
patterns. A portion of weight factors generated are shown
in Figure 2 and Table 3. If a pattern similar to model 3 is
tested for the recognition purpose, the program will generate
an activation number of pattern similar to column 3 where
the 3rd entry is the most significant number.
Before the training process, the dimensionless derivative
type curves are normalized for x-axis and y-axis respectively
0.1 ~ (tD)N ~ 0.9
0.1 ~ (Pb.tD)N ~ 0.9
The entire pattern is then represented as a model with
30 pairs of x and y. The normalization process is shown in
Figure 3.
The iterative process distributes the error using the back
propagation algorithm. A sigmoid function is used as de-
scribed below to show the calculation process in a single node
1
Output· = --~:------:­
J 1 + exp(-sumj )
where:
OutputF the output of a node in j level
sumj = the summation of incoming signal to
to node j from all nodes in the i level
A similar process is used for the k nodes receiving signal
from the j nodes.
The output of sigmoid function will approach 1 at a very
large positive argument and 0 at very small argument. These
values represent the activation level of the node based on the
input received.
In our studies we experimented with the j level nodes at
130 and 200. The number of k level nodes correspond to the
SPE 026106 IRWAN R.JUNIARDI AND IRAJ ERSHAGHI
number of models under study. To examine the application
of neural networks as a pattern recognition tool in well test
analysis, we focused on the faulted reservoirs. These reser-
voirs as indicated by Ershaghi et al. 14 exhibit some of the
most complicated and confusing shapes on ISDP. Figure 4
shows various fault configurations studied for both homoge-
neous and dual porosity reservoirs. Type curves generated
for 10 models covering 20 patterns within limited range of
reservoir and wellbore parameters are shown in Figures 5a -
5 j. Table 4 shows the equation representing the 10 selected
models. In spite of significant differences among equations,
similar patterns are obtained attributed to more than one
well test model.
Experimentation
The neural network model was applied to the 20 patterns
described above. Weight factors were generated for various
levels of j nodes. In all experiment a total error of 0.01 was
imposed as a tolerance level. Figures 6 - 9 show synthetic
patterns with known models which were tested against the
network model and activation codes were obtained indicating
perfect recognition as shown in Table 5 (Cases 1 and 2) for
Figures 6 - 7 and multiple answers for Figures 8 - 9, and
wrong answer for Figure 9.
Because oflimited number of training patterns, ifthe range
of parameters used for the generation of test cases were out-
side the range corresponding to the input models, at time
we faced non unique solutions. For example, an input pat-
tern similar to Figure 10 was recognized by the network as
either model 3 and 8. This obviously casts uncertainties in
the ability of the network to perfectly identify the correct
model.
Statement of the Problem
From the above experiments, it is clear that a reliable neu-
ral network for well test application must include in its train-
ing set considerably more patterns. The patterns must in-
clude all possible combinations of reservoir parameters. As
such, the required number of patterns will make the training
difficult if not impossible.
Proposed Solution
To alleviate the requirements for generating a wide range
of type curves, a Monte Carlo Simulation approach was em-
ployed to test the recognition capability of the network. With
the random parameter selection approach, a significant num-
ber of patterns for a particular well test model can be gen-
erated with expected ranges of minimum and maximum of
various parameters. Consider well test model 7. Figures
lla - llb show eight typical normalized type curves gener-
ated for the model using different input parameters. Obvi-
ously, the network model will have a different and at time
713
3
wrong interpretation for the cases.
Suppose that the response of the same model is now sim-
ulated via a Monte Carlo approach for parameter selection
within the ranges of parameters shown in Table 6. In one ex-
periment with 50 iterations, we generated 50 patterns which
were recognized by the network according to the frequency
plot shown in Figure 12. Note that model 7 is the pick in
only 28% of the cases. Other potential models such as model
2 and 3 are selected with high frequencies. This is because
of unusual shapes caused by certain combination of param-
eters. This type of frequency table must become part of the
data base assisting the neural network.
From the above studies we have observed that the neural
network models need augmentation from an expert system.
Such an expert system is required to assist in the selection
of the correct model in two different ways. First, the type
of the frequency data similar to the one determined above,
can provide a data base of potential models which can fit a
particular shape. For example, for the base patterns studied
here, if a new pattern is recognized by the neural network
with an activation number similar to model 2, the expert
system must have adequate statistical data to also identify
model 7 as a potential answer because of its high frequency of
occurence in the category 2 Model. Experiments such as the
one proposed above can be repeated for other models with
even a wider range and the information for construction of
the frequency data base can be generated for the expert sys-
tem. Second, information in the expert system can include
checks and balances on other field data; i.e., independent
evidence as to fracturing, mud loses, etc.
Thus, the neural network approach for pattern recognition
has a limited degree of reliability unless either the system is
trained on almost infinite number of patterns or the system
is augmented with a database including the statistical fre-
quencies of non-related well test models resembling a given
pattern and other field related information.
Conclusions
The complexities of neural network pattern recognition ap-
proach for well test models have been pointed out through
numerous case studies. A hybrid method is proposed where
from a Monte Carlo simulation approach significant number
of patterns can be generated and tested against a network
model with essential well test patterns. Statistical data gen-
erated from such comparison can, in a form of an expert
system, augment the neural network ability for model selec-
tion processes.
Nomenclature
CD = wellbore storage,dimensionless
d = distance to fault, ft
4 COMPLEXITIES OF USING NEURAL NETWORK IN WELL TEST ANALYSIS OF FAULTED RESERVOIRS
dD = distance to fault,dimensionless, d/rw
/(s) =
defined by equation of Model 7
I<O(x) = modified lJessel function, 2nd kind,
zero order
I<l(X) = modified Bessel function, 2nd kind,
first order
= wellbore pressure drop, dimensionless
== radial distance, ft
= radial distance, dimensionless, r/rw
= wellbore radius, ft
= Laplace parameter
= dimensionless skin factor
= time, dimensionless
= matrix-fracture interporosity parameter,
dimensionless
w = oil zone fracture storativity with respect
D = denotes dimensionless
/ =
denotes fracture
w = denotes wellbore
Acknowledgements
This research was supported by The Gas Research Insti-
tute. We acknowledge the cooperation of Mr. Yusuf Shikari
during the course of this study and thank GRI for permis-
sion to publish the paper. We also acknowledge Pertamina
for supporting Mr. Juniardi's scholarship.
References
1. Ershaghi, I. and Woodbury, J.J. : "Examples of Pitfalls
in Well Test Analysis", J.Pet.Tech (February 1985) 335-
341.
2. Ramey, H.J. Jr. : "Advances in Practical Well Test
Analysis, "J .Pet.Tech(June 1992) 650-659.
3. Horne, R.N.: "Advances in Computer-Aided Well Test
Interpretation, "paper SPE 24731 presented at the 1992
Annual Meeting, Washington DC, October 4-7.
4. Bourdet, D., Ayoub, J.A., and Pirard, Y.M. :"Use of the
Pressure Derivative in Well Test Interpretation, "SPE
Formation Evaluation(June 1989) 293-302.
5. Ershaghi, I., Chang J. and Khachatoorian, R.A. : "The
Importance of Pattern Recognition for the Analysis of
ISDP for Gas Storage Reservoirs," paper presented at
the 1992 International Gas Research Conference, Or-
lando, November 16-19.
714
SPE 026106
6. Allain, O.F. and Horne, R.N. : "Use of Artificial Intel-
ligence in Well Test Analysis, "J .Pet.Tech(March 1990)
342-349.
7. AI-Kaabi, A.U. , McVay, D.A. , and Lee, W.J. : "Using
an Expert System to Identify a Well Test Interpretation
Model" J .Pet.Tech(May 1990).
8. Baldwin, J.L and Otte, D.N.: "Computer Emulation of
Human Mental Process: Application of Neural Network
Simulators to Problems in Well Log Interpretation," pa-
per SPE 19619 presented at 1989 Annual Conference,
San Antonio, October 8-11.
9. Arehart, R.A. : "Drill Bit Diagnosis Using Neural Net-
works," paper SPE 19558 presented at the 1989 Annual
Meeting, San Antonio, October 8-11.
10. Rogers, J.D., Guffey, C.G., and Oldham, W.J.B. : "Ar-
tificial Neural Networks for Identification of Beam Pump
Dynamometer Load Cards, "paper SPE 20651 presented
at the 1990 Annual Conference,New Orleans,September
23-26.
11. AI-Kaabi, A.U. and Lee, W.J. : "Using Artificial
Neural :Networks to Identify the Well Test Interpreta-
tion Model, "paper SPE 20332 presented at the 1990
Petroleum Computer Conference, Denver, June 25-28.
12. Rumelhart, D.E. and McClelland, J.L.: "Paral-
lel Distributed Processing: Explorations in the Mi-
crostructures of Cognition, VoU, "The MIT Press,
Cambridge, 1986,318-362.
13. Freeman, J .A. and Skapura, D.M.: "Neural Net-
works: Algorithms, Applications, and Programming
Techniques, "Addison-Wesley Publishing Co.Inc., Oc-
tober 1991.
14. Ershaghi, I., Khachatoorian, R.A. and Y.Shikari :
"Complexities in the Analysis of Pressure Transient Re-
sponse in Faulted Naturally Fractured Reservoirs, "pa-
per SPE 23424 presented at the 1991 Eastern Regional
Meeting, Lexington, October 22-25.
 Tlible 1

I~J::: : ShoItT"""Teat
fM~~':.:II~Infinite Acting

I~,~, :,II~ lnIiniteAding + Ang...... F'"

Fig.l Input of a derivative plot into the

neural network (After AI-Kaabi l l )

TIIble2

node 1 node 2 node 3 node 4 node 5 node 6 node 7 nodeS node 9 node 10

:':IiiJ:::::::::" 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1

0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1

0.1 0.1 !i¥) 0.1 0.1 0.1 0.1 0.1 0.1 0.1

0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1

0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1

0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1

0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1

0.1 0.1 0.1 0.1 0.1 0.1 0.1 M:} ::: 0.1 0.1
0.1 0.1 0.1 0.1 0.1 0.1 I 0.1 I 0.1 ::<~~ ~{: 0.1
0.1 0.1 0.1 0.1 0.1 0.1 I 0.1 I 0.1 0.1 ?#:\t)

Tlible 3

irc>ut hidden :,::~!'ift hidden ouIpUl }~t

j j k
1 0.1951 1 1 -G.3076
2 -G.0545 1 2 0.6587
3 -G.4537

• • -G.4871
2
2
1
2
~0610
0.7137
• •
2
2
1
2 1.1758
2 3 0.5024 3 0.05111

Fig.2 Set of weights obtained after

training the neural network

715
0.9 10 1

0.8

0.7
100
0.6

O.S

0.4
10-1

0.3

0.2

0.1 10-2
102 103 1(}4 lOS 106 107 108 109 1010

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Fig.3 Normalization of a derivative plot

• •
linear perpendicular

•
• well

I fault

angular parallel

FigA Fault configurations in the reservoirs

716
 0.9......--~--~--~--~--~-~--~--~ 0'9,---~--_--~-~--_-_~-~-____,

0.8 0.8
...•..........
0.7 ...................
~'.-.,
- .

~/~~'~:-
0.2 a
0.3 ..

0.2
Model D : Homogcaeous Inrinlle Acting

b
0.1 ' - _......._ _-:'-_ _ _ _-'-_ _ ......._ _-'-_ _...J
O'L.L.I--..,.0'~2--..,.0.3'::---0.~4--~0.5'::-- .......
0.~6--~0.~7--~0.'::"8--JO.9
~ ~ _

0.1 0.2 0.3 Q.4 0.5 0.6 0. 7 0.8 0.9

o.9......--~--~--~-~--~-_~-~-__, 0.9,---~--~--~-~--_-_~-~-__,

0.8 0.8
close
0.7 0.7
/ .
0.6 0.6
..... _-- ... _- .... -- ..
far
0.5 far 0.5

0.4 0.4

0.3 .. Model ill : Homogcaeous + Unear Faull 0.3 .. Model IV : Homogeneous + Perpendicular Fault

0.2 c 0.2 d
0.1 L--~::----:-,O __ ~ __ ~_~~_~ _ _-:'-_ _...J O.I'---~--~--~--~- ......._ -.......- -.......- - - '
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 ~ U U Q.4 ~ U U M M

0.9r--~--~--~-~--~--~-~--__, 0.9,..---~--~--~--~--_-~--~----.

0.8 _----==----4 0.8

\
close
~.~ -_.-
......\ ..\. .. / . close
0.7 0.7
...."'/" .............
0.6 0.6
........... _- .. _--_ ....
far
0.5 far

l' 0.4 ././.... Model VI : Homogeneous + Par2lIel F.ult

0.3 .. Model V ~ Homogeneous + Angular Fault 0.3

0.2 e 0.2 f
0.1 ';---::'::----::;---:'-:---::':::----:'-:"-----:'::---::':::----:' 0·b.LI---0~.2--~0.~3--..,0~.4--..,0~.S---0~.-6---0.~7---0~.8---0-'.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

0'9r--~--_--~-~--_--_-~-__, 0.9......---_--_--~--_- __- __ -~--__,

0.8

0.7

0.6

0.5 / ...... 0.5 close

close

0.4 0.4

0.3
Model vn : NFR + Unear Fault 0.3
Model VID : NFR + Peq>endicular F.uk

0.2 g 0.2 h
0.1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

( continued )

717
0.9'r--~--~-~--~-~--~-~-----' 0.9r--~--~--~-~--~--~-~----'

0.8 0.8

0.7

0.6

o.s
0.4

MocIeJ IX : NfR + Aacular Fouk

Model X : NFR + PanUeI FIUIt

i j
O'Il.;----::-::--:'':---:-c---::'::-----:~-__:,:O-__:'::_____.,J
0.1 D.2 0.3 0.4 O.S 0.6 0.7 0.8 0.9

Fig.Sa.-Sj Patterns of the models used for training the network

Table 5

node 1 node 2 node 3 node .. node 5 node 6 node 7 node 8 node 9 node 10
0.0898 0.0896 0.0000 0.1642 0.1342 0.1175 0.1117 0.0897 :::"'~~::: 0.0959
0.1905.. ..~:!~\.. 0.1282 ..

O.llM .. .. .. .. :~{ .. 0.1282 0.1323

:: negligible

~r:::)):{ NFR + Angu.... Faull (Model 9) MOdel 9

~~:: NFR + u.- Fault (Model 7) Model 7
~~:::::):: :::: NFR + P"'l*lCficu18t Faull (Model 8) Model 7 I 8
:~~:): <:::::::::: NFR + PlIlplIIlCIic:uIar Faull (Model 8) Model 7

0.9 0.9r--~--~--~-~--~--~-~--_,

0.8 0.8

0.7

0.6

o.s
0.4 0..

0.3 0.3

0.2 Case I 0.2 Casell

O.IL....-~--~---'--~,___-~-____,~-~--- 0.1 L--~:__-~-__,~-__c'::__-,,-'-:--__,"'::_-_:_'":______}

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 O.S 0.6 0.7 0.8 0.9

Fig.6 Pattern of synthetic data case 1 Fig.7 Pattern of synthetic data case 2
tested againts the neural network model tested againts the neural network model

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

O.S o.S

0.4 0..

0.3 0.3

0.2 Case UI 0.2 Case IV

0.1 '--~--~--~-~--~--~-~-.- 0.1

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.. 0.5 0.6 0.7 0.8 0.9
Fig.8 Pattern of synthetic data case 3 Fig.9 Pattern of synthetic data case 4
tested againts the neural network model tested againts the neural network model
718
 Table 4

- SPD +S - Ko(rD../i)
Model 1 : Short Term Test
PWD = s[1 + SCD(SPD + 8)] PD = s-{iK1(../i)

Model 2 : Homogeneous Reservoir Infinite Acting The same equation 88 of Model 1

- SPD +S PD = Ko(rD../i) + Ko(dD../i)

Model 3 : Homogeneous Reservoir + Linear Fault PWD = s[1 + SCD(SPD + 8)] s-{iK l(../i)

Po _ Ko(rDVS) + Sy'SK 1(v'i) + A _

Model 4 : Homogeneous Reservoir + Perpendicular Fault WD - S{VSK1(v'i) + CDs[Ko(rDv'i) + SVSK1(JS) + An
3

A = EKo(dDiVS) (3 images)
i=l

Model 5 : Homogeneous Reservoir + Angular Fault The same equation 88 of Model 4 but instead of 3 images put 7 images.
"'......co"
Model 6 : Homogeneous Reservoir + Parallel Fault The same equation as of Model 4 but instead of 3 images put 00 images.

Model 7 : Naturally Fractured Reservoir + Linear Fault The same equation 88 of Model 1 where

I'D _ Ko(rDVif(8) + Ko(dDVif(8) !(s) = w(1 - w)s +,x

- sVS !(S)Kl(VS!(S» (l-w)s+'x

Model 8 : Naturally Fractured Reservoir + Perpendicular Fault

The same equation as of Model 4 but instead of y'S put vs/(s).

Model 9 : Naturally Fractured Reservoir + Angular Fault The same equation as of Model 4 but instead of VS put vs/(s)
and instead of 3 put 7 images.

Model 10 : Naturally Fractured Reservoir + Parallel Fault The same equation as of Model 4 but instead of y'S put vs/(s)
and instead of 3 put 00 images.