Australian Journal of Basic and Applied Sciences, 6(12): 134-140, 2012

ISSN 1991-8178

Prediction of Heat Affected Zone (HAZ)Corrosionin CO2Environment by Artificial

Neural Network
1

Saeid Kakooei, 1M.C. Ismail, 2Khaled Alawadhi, 3Mahmood Bataee

Centre for Corrosion Research, UniversitiTeknologi PETRONAS, Bandar Seri Iskandar, 31750
Tronoh, Perak, Malaysia.
2
Department of Automotive and Marine Technology, The Public Authority for Applied Education and
Training, Shuwaikh, Kuwait.
3
Petroleum Engineering Department, UniversitiTeknologi PETRONAS, Bandar Seri Iskandar, 31750
Tronoh, Perak, Malaysia.
Abstract: The corrosion is the natural process of the metal deterioration. Regardless of the process
time, major cost is expended for the corrosion prohibition. One of the important corrosion parts in
industrial consideration is the corrosion of heat affected zone (HAZ) for the welded carbon steel in
CO2 environment. The experimental study has been done for determination of the corrosion in this
part. Nowadays, artificial intelligent systems (ANNs) are very popular as the tool for recognizing the
systems and preparing a network to simulate the conditions. A neural network has been chosen to
predict the corrosion of HAZ. Variable parameters such as temperature and flow velocity were used for
data collection in both inhibited and non-inhibited CO2 containing solution. The ANN model showed
very good agreement with experimentally obtained corrosion rates in this study, as it would be applied
widely in corrosion studies.
Key word: Artificial neural network; Heat affected zone (HAZ);WeldedCarbon steel; CO2 corrosion
INTRODUCTION
Generally, CO2 corrosion predictionmodelling can be categorized under three groups: empirical, semiempirical and mechanistic models as appeared in various publications (Nesic et al., 2001, Fatah et al., 2008,
Ismail and Skeldon, 2005). However, in non-homogenous material condition such as that of welded
pipeline;predictionmodelling is complicated due to metallurgical variation of welded region such as weld metal
(WM), heat affected zone (HAZ) and parent material (PM) (Alawadhi and Robinson, 2011).One salient point to
be considered in such prediction is the contribution of galvanic corrosion of the WM and HAZs with the parent
material(Lee and Woollin, 2005).Galvanic effect between WM and PM can arise if WM is selected to be
slightly noble to the PM. In this case WM will remain cathodic which means WM will be protected and PM will
be corroded. Another factor is micro-structuralvariation between HAZ and PM. For example, although HAZ and
PM have the same composition, heat modified microstructure of HAZ can be more susceptible to corrosion. It
has been shown that increasing the hardness, the grain size and the proportion of aligned second phase in the
WM causespreferentialweldcorrosion(PWC) in welded carbon steels(Turgoose and Palmer, 2005, Lee and
Woollin, 2005).
Modelling the nonlinear behaviour of corrosion currents with time during corrosion process requires a tool
such as ANNs; which is a better approach than other statistical methods (Topu et al., 2009).Neural networks
consist of paralleldistributed units known as neuron that process the relation between parameters. The neurons
acts as a realbiological neuron(Kamrunnahar and Urquidi-Macdonald, 2011). Neural networkcanlearn a system
andthen, recognize, and classify it. These networks are strictly driven by the dataand false data could affect the
result. The more data trained, the better results will be given and somehow the error data would be less
effective.Thethree layers of neurons are the major parts of a neural network; which are input, hidden and output
layer as shown in Figure 1(Cassidy and McCoy, 2002). The procedure of the network starts when the neuron
gets the input value. Then, neuron multiplies it by a value which is called weight. The proceeding neuron,
correct it with a value (bias). The first iteration ends as this value transfers by the transfer function to produce
results. As the iteration procedure continues, the bias and weights (which were random at the start) changes, but
the input and target value never changes.
The disadvantages of neural networks are the training and the high process time of operations in the large
networks(Kamrunnahar and Urquidi-Macdonald, 2011).

Corresponding Author: Saeid Kakooei, Centre for Corrosion Research, UniversitiTeknologi PETRONAS, Bandar Seri
Iskandar, 31750 Tronoh, Perak, Malaysia.
E-mail: skakooei59@hotmail.com

134

Aust. J. Basic & Appl. Sci., 6(12): 134-140, 2012

Fig. 1: Arrangement of the layers in the ANN network(Cassidy and McCoy, 2002)
2. Experimental Detail:
Electrochemical corrosion rate measurements were carried out using a rotating cylinder electrode (RCE)
apparatus with the samples produced from the different regions of the weld. An advantage of using the RCE is
that the hydrodynamic conditions are very well defined, and it is feasible to simulate conditions field condition
in laboratory tests(Gabe et al., 1998, Alawadhi, 2009).Linear Polarization Resistance approach was employed
for corrosion rate calculation. The experiments were done at three temperatures 20, 50 and 70 C in a cell
containing artificial seawater, saturated with carbon dioxide. Two experimental conditions were used: 1)
uninhibited condition, 2) inhibited solution containing 30 ppm by volume of an oilfield corrosion inhibitor. The
concentration of corrosion inhibitor was fixed at 30 ppm by volume, which is typical of oilfield practice
(Gulbrandsen and Dugstad, 2005).In each test, the RCE was held stationary for the first 10 hours, where a stable
dark grey film developed on the electrode surfaces. The RCE was then rotated for 20 minutesat 500 rev.min-1
while LPR technique was employed for corrosion rate measurement. The measurements were repeated with the
speed increased to 1000 rev.min-1 and then in steps of 1000 up to 4000 rev.min-1.
Samples were machined from welded X65 pipeline steel with a composition as shown in Table 1. More
detail ofsample preparation are mentioned by Alawahdiet al.(2011)(Alawadhi and Robinson, 2011).
Table1. Chemical Composition of X65 Pipeline Steel
Chemical
C
Ni
Cr
Mo
composition

Si

Mn

Al

Cu

(%)

0.3

1.6

0.04

0.009

0.02

0.05

0.006

0.08

0.04

0.02

0.01

Fig. 2: Rotating cylinder electrode components: a)Cylindrical working electrodes

b) Diagram showing the weldment sectioning c) Welded cylindrical electrode sections
d)Standard glass cell

135

Aust. J. Basic & Appl. Sci., 6(12): 134-140, 2012

In the ANNmodel, input layer consists of the three parameters; inhibitor concentration, temperature and
rotational speed. The target (output) layer consists of the corrosion rate of the HAZ region in Welded X65
Pipeline Steel as shown in figure 3. The input and target layers data were obtained from the laboratory
experiments. Several main functions have been tested to train the network. The feedback results are compared
and the proper function has been selected.

Fig. 3: The system used in the neural network corrosion simulationof HAZ in Welded X65 Pipeline Steel.
The neural network program has been applied to the data.Some data should be chosen randomly to check
the accuracy and the validity of the trained data. The data are distributed randomly and the 60% of them are
used for the training, 20% for the validation and 20% asthe test for the network. After assuming the error limits,
the model is developed. Using this network, the different situation for the corrosion phenomena could be
simulated within the range of the experience. So the results and the diagrams for temperature, velocity and
inhibitor could be achieved, correlated to each other.
RESULTS AND DISCUSSION
The neural network provides a model that creates a correlation between input and target. There are some
characteristics of the trained network. Different network types had been tested and feed forward back
propagation had selected as showed better results. The trainlm and tansig were selected as the training and
neuron transfer functions respectively, regarding to their more accurate results.
The result of the training network shows that all data are successfully modelled. The progress of learning
the network and performance of the ANN model is shown in Figure 4.

Fig. 4:The progress of network learning

136

Aust. J. Basic & Appl. Sci., 6(12): 134-140, 2012

Fig. 5: Error of (a) training (b) validation (c) test and (d) total data in the network
Table 2: Experimental and ANN simulated result ofcorrosion rate of HAZ (*: 0 used for uninhibited solution and 1 used for inhibited
solution by 30 ppm inhibitor)
Experimental Corrosion Rate of Simulated
Inhibitor*
Temperature (C)
Rotational Speed HAZ (mm/year)
Corrosion Rate of HAZ
(RPM)
(mm/year)
1
1
1
1
1
1
0
0
0
0
0
0
1
1
1
1
1
1
0
0
0
0
0
0
1
1

25
25
25
25
25
25
25
25
25
25
25
25
50
50
50
50
50
50
50
50
50
50
50
50
70
70

0
500
1000
2000
3000
4000
0
500
1000
2000
3000
4000
0
500
1000
2000
3000
4000
0
500
1000
2000
3000
4000
0
500

0.017
0.008
0.008
0.008
0.01
0.01
0.971
0.719
0.882
1.241
1.42
2.069
0.116
0.093
0.116
0.127
0.131
0.2
1.758
1.923
2.119
2.105
2.333
2.277
0.029
0.084

137

0.018759
0.019246
0.019711
0.020222
0.018019
0.013673
0.977845
0.742652
0.689292
1.186953
1.428698
1.548251
0.126523
1.918372
0.098043
0.131152
0.168843
0.189227
1.76193
1.76193
2.00138
2.121931
2.290321
2.325816
0.032093
0.041641

Aust. J. Basic & Appll. Sci., 6(12): 134

4-140, 2012

1
1
1
1
0
0
0
0
0
0

70
70
70
70
70
70
70
70
70
70

10
000
20
000
30
000
40
000
0
50
00
10
000
20
000
30
000
40
000

0.06
0.063
0.195
0.158
1.034
1.125
1.125
1.003
0.912
1.155

0.057409
0.09352
0.147168
0.178382
1.007879
1.121365
1.15438
0.999721
0.820689
1.216127

0
was chhosen as a corrrelation betweeen
Also, the best ANN model, whichh had a total rooot square of 0.99107
a output layeer. The Figure 5 shows the errror of trainingg, validation, teest and total daata in the modeel.
the input and
Looking att the R2=0.9910
07 in Figure 5,, the ANN moddels were highlly correlated w
with the observaation data.
Table 2 shows expeerimental desig
gn, experimenttal and predictted results for this study. Chhanging of thrree
factors wass investigated and
a result will be discussed inn forthcoming figures.
Figuree 6 shows a co
omparison betw
ween experim
mental and simuulated corrosioon rate of heatt affectedarea in
welded carrbon steel in uninhibited
u
CO
O2environment for threedifferrentvelocities. According to this figure, it is
clear that ANN model was
w accurate in
i estimating oof corrosion rate in HAZ arrea in welded carbon steel in
d condition. In
n next effort, 30 ppm by vvolume of corrrosion inhibitoor was added to experimenttal
uninhibited
solution.Thheexperimentaal result recordded in this conndition was used for ANN m
model. A compparison betweeen
experimenttal and simulaated corrosion rates
r
is shownn in Figure 7. It
I can be seen that the predicctedANN moddel
fits well with
w
the experrimental data. The developped ANN moddel displays good performan
nce in differeent
temperaturre and velocityy, and the netw
workresults espeecially in low velocities
v
valuues are in good
d agreement wiith
experimenttal result.

mulated and expperimentalresuult of corrosionn rate of HAZ versus

v
temperaature at differennt velocity andd
Fig. 6: Sim
uniinhibited condiition

Aust. J. Basic & Appll. Sci., 6(12): 134

4-140, 2012

Fig. 7: Sim
mulated and expperimental resu
ult of corrosionn rate of HAZ versus temperaature at differeent velocity andd
inh
hibited conditio
on
e
of velociity on corrosionn rate of HAZ area in weldedd carbon steel in
i three differeent
Figuree 8 shows the effect
temperaturres. A good fitt is observed between
b
simulaated and experrimentaldata.A
At room temperrature, corrosioon
rate increaase by increasing velocity. Whereas
W
at hiigh temperaturre, corrosion mechanism
m
changesdue to thhe
formation of corrosion product
p
film. The
T detail aboout this mechan
nism is discusssed elsewheree (Alawadhi annd
Robinson, 2011).
EDX analysis
a
in Figgure 9 revealed
d the presence of significant iron level withh carbon in HA
AZ weld regionns
correspondded to the preseence of cementtite (Fe3C). Thhis Fe3C film iss formed in unninhibited soluttion at 70 C annd
maximum velocity condiition.
pplication of ANN
A
model fo
or prediction of
o HAZ corrossion rate in CO
O2 environmen
nt was effectivve.
The ap
Influence of
o factors such as temperaturee, and flow vellocities were coonsidered in ANN
A
model.

Fig. 8: Sim
mulated and expperimentalresuult of corrosionn rate of HAZ versus
v
velocityy at different teemperature andd
uninnhibited condittion

Aust. J. Basic & Appl. Sci., 6(12): 134-140, 2012

Fig. 9: EDX spectra and analysis of the corrosion product film on the HAZ under flowing conditions at 70 C
Conclusions:
Corrosion rate of HAZ was investigated in this work due to its higher corrosion rate in comparison with
other area in welded carbon steel such as WM and PM. The ANN model in this study was successful to model
the corrosion rate of HAZ region in welded X65 carbon steel under different velocities and temperatures in both
uninhibited and inhibited condition. The preliminary results have approved the validity of the artificial neural
network tool with values of R2 for the correlation between experimentally and simulated determined corrosion
rates of 0.99.ANN simulation result showed a good agreement with experimentally obtained corrosion rates for
the test system which was used in this study which confirmed its effective approach for corrosion prediction.
REFERENCES
ALAWADHI, K., 2009. Inhibition of weld corrosion in flowing brines containing carbon dioxide. PhD
THESIS, CRANFIELD UNIVERSITY.
ALAWADHI, K. & M. ROBINSON, 2011. Preferential weld corrosion of X65 pipeline steel in flowing
brines containing carbon dioxide. Corrosion Engineering, Science and Technology, 46(4): 318-329.
CASSIDY, J. & T.H. MCCOY, 2002. A Neural Network for Predicting Corrosion of Grade N80 Downhole
Tubulars Exposed to Stimulation Fluids. CORROSION 2002. NACE International.
FATAH, M., M. ISMAIL, & B. KURNIAWAN, 2008. Comparison of CO2 Corrosion Prediction Models
based on Field and Experiment Data, Paper No. International Conference on Plant and Reliability Management
(ICPER). Universiti Teknologi PETRONAS, Malaysia.
GABE, D.R., G. WILCOX, J. GONZALEZ-GARCIA, & F. WALSH, 1998. The rotating cylinder
electrode: its continued development and application. Journal of applied electrochemistry, 28(8): 759-780.
GULBRANDSEN, E. & A. DUGSTAD, 2005. Corrosion loop studies of preferential weld corrosion and its
inhibition in CO2 environments. CORROSION 2005. Houston, TX, USA: NACE International.
ISMAIL, M.C. & P. SKELDON, 2005. Prediction of C02 Corrosion with the Presence of Acetic Acid, PhD
THESIS,University of Manchester.
KAMRUNNAHAR, M. & M. URQUIDI-MACDONALD, 2011. Prediction of corrosion behaviour of
Alloy 22 using neural network as a data mining tool. Corrosion science, 53(3): 961-967.
LEE, C.M. & P.M. WOOLLIN, 2005. Preferential weld corrosion: Effects of weldment microstructure and
composition. CORROSION 2005, 2005 Houston, TX, USA. NACE International, pp: 05277.
NESIC, S., M. NORDSVEEN, N. MAXWELL, & M. VRHOVAC, 2001. Probabilistic modelling of CO 2
corrosion laboratory data using neural networks. Corrosion science, 43(7): 1373-1392.
TOPU, .B., A.R. BOGA, & F.O. HOCAOGLU, 2009. Modeling corrosion currents of reinforced
concrete using ANN. Automation in Construction, 18(2): 145-152.
TURGOOSE, S. & J.W. PALMER, 2005. Preferential Weld Corrosion of 1% Ni Welds: Effects of Solution
Conductivity and Corrosion Inhibitors. CORROSION 2005. Houston, TX,USA: NACE International.

140