Journal of Materials Processing Technology 170 (2005) 536–544

Artificial neural networks for modelling the mechanical properties

of steels in various applications
Z. Sterjovski ∗ , D. Nolan, K.R. Carpenter, D.P. Dunne, J. Norrish
Faculty of Engineering, University of Wollongong, Northfields Avenue, NSW 2522, Australia

Received 16 November 2004; received in revised form 26 November 2004; accepted 26 May 2005

Abstract

The application of artificial neural networks (ANNs) in predicting some key properties of steels is discussed in detail. This paper reports on
the effectiveness of three back-propagation artificial neural network models that predict (i) the impact toughness of quenched and tempered
pressure vessel steel exposed to multiple postweld heat treatment (PWHT) cycles, (ii) the hardness of the simulated heat affected zone in
pipeline and tap fitting steels after in-service welding and (iii) the hot ductility and hot strength of various microalloyed steels over the
temperature range for strand or slab straightening in the continuous casting process. Predicted and actual experimental values for each model
are well matched and highlight the success of applying ANNs in predicting mechanical properties. The capability of ANNs in predicting
multiple outputs (hot ductility and hot strength) is also demonstrated.
The sensitivity, which is a measure of the response of an output across the range of an individual input variable, of key input variables
(individual alloys and/or process steps) for each model is shown to be in agreement with findings of both the experimental investigation and
reports in the literature. Although this paper shows that ANNs can be employed for optimizing steel and process design parameters, some
difficulty can arise when inter-relationships exist between input variables. An understanding of the inter-relationships between input variables
is essential for interpreting the sensitivity data and optimizing design parameters. It is argued that artificial neural network models can be
developed that have the capacity to eliminate the need for expensive experimental investigation in areas, such as welding (new and repair),
inspection and testing, and manufacturing processes.
© 2005 Elsevier B.V. All rights reserved.

Keywords: Artificial neural network modelling; Steels; Impact toughness; Hardness; Hot ductility; Continuous casting; In-service welding

1. Introduction in analysing the tensile properties of welds in power plant

Artificial neural networks (ANNs) are computational net-
works that attempt to simulate the processes that occur in
the human brain and nervous system during pattern recog-
nition, information filtering and functional control [1]. Alter
[2] describes ANNs as systems that recognize patterns based
on the data used to train them, where each training case is
characterized by a set of inputs and a result (output or multi-
ple outputs). The use of these networks to make predictions
of the mechanical properties of materials is a relatively new
concept, but one that has received considerable interest in
recent years. ANNs have been reported to be very effective
∗ Corresponding author. Tel.: +61 2 42214842.
E-mail address: zoran@uow.edu.au (Z. Sterjovski).
steels [3], the effect of carbon content on the hot strength of
steels [4] and the impact toughness of welds based on com-
mon welding parameters [5].
This paper introduces three different back-propagation
neural network models which can predict the (i) impact
toughness of quenched and tempered steels exposed to var-
ious postweld heat treatment (PWHT) cycles, (ii) simulated
heat affected zone toughness of pipeline steels resulting from
in-service welding and (iii) hot ductility (reduction of area
(ROA)) and hot tensile strength of microalloyed steels. A
back-propagation neural network is the most common type of
neural network and it undergoes its learning phase by calcu-
lating an error between the predicted and actual output. This
ANN incorporates a hidden layer that is used to establish the
inter-relationships between the input variables and their rela-

0924-0136/$ – see front matter © 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.jmatprotec.2005.05.040
 Z. Sterjovski et al. / Journal of Materials Processing Technology 170 (2005) 536–544 537

tionship with the output to minimize the error between the
actual and predicted output.
The testing of QT steels to qualify them for pressure vessel
use is a time consuming and expensive process. Hence, the
development of a reliable model to predict relevant mechan-
ical properties, such as impact toughness as a function of
alloy content (steel design) and heat treatment conditions
(process parameters), is a worthwhile goal for the pressure
vessel industry. Modelling impact toughness can be difficult
due to the well-known scatter that is associated with Charpy
V notch impact testing. However, ANNs automate the mod-
elling process and this improves the possibility of success.
Accurate prediction of hardness of the heat affected zone
of a weldment as a function of composition and cooling time
can make a valuable contribution to the avoidance of crack-
prone microstructures during in-service welding. The second
model presented in this paper demonstrates the accuracy with
which hardness values can be predicted in simulated HAZ
structures, which in turn can be used to avoid crack-prone
structures (>350 HV). This neural network model incorpo-
rates materials characteristics (chemical composition and
as-received hardness), peak temperature, holding time and
cooling rate as key input variables for predicting hardness.
The third model predicts hot ductility and hot strength
of microalloyed steels based on chemical composition and
treatment conditions. The purpose of this model is to assess
the likelihood of steels developing transverse cracking during
the straightening operation in the continuous casting process.
This model is designed to predict two outputs, and in doing so
it provides an insight of the functional behaviour of artificial
neural networks in cases of multiple outputs.
As well as predicting specific mechanical properties as a
function of process and material parameters, all three models
allow the optimization of these parameters through a sensitiv-
ity analysis of the input variables. Sensitivity is a measure of
the response of output across the entire range of an individual
input variable [6]. In short, the aim of this paper is to underline
the usefulness and effectiveness of artificial neural networks
to predict, with relatively high accuracy, various properties
of steels and discuss the advantages and limitations of ANNs
in steel and process design.
2. Experimental methods
2.1. Artificial neural network modelling
Neuralworks Professional II/Plus Software was used to
build the three back-propagation neural networks. As men-
tioned, the models were designed to predict the: (i) impact
toughness of quenched and tempered steels exposed to
repeated PWHT cycles (Model 1), (ii) simulated HAZ hard-
ness of pipeline and tap fitting steels after in-service welding
(Model 2) and (iii) reduction of area (hot ductility) and hot

Table 1
Input variable data ranges used for the three models
Input variable Model 1 Model 2 Model 3

Minimum value Maximum value Minimum value Maximum value Minimum value Maximum value
Thickness (mm) 11 20
Test typea T L M S
Test temperature −90 20 700 1100
Timeb 0 200 0 2
Coolingc Slow Fast 2 10 100 200
Normalised HV 159 234
%C 0.155 0.180 0.07 0.29 0.155 0.165
%Mn 1.10 1.43 0.44 1.63 0.63 1.23
%S 0.0020 0.0035 0.005 0.024
%Cr 0.016 0.2 0.012 0.31
%B 0.0005 0.0013
%Si 0.005 0.39
%Cu 0.010 0.28
%Ni 0.009 0.1
%Mo 0.005 0.033
%V 0.005 0.06
%P 0.008 0.02
%Ti 0.005 0.021 0.003 0.018
%Al 0.005 0.039 0.027 0.036
%Nb 0.005 0.053 0.001 0.022
%N 0.0021 0.0033
a ‘T’ denotes the direction of rolling is transverse to the length of Charpy sample, ‘L’ denotes the direction of rolling is longitudinal to the length of Charpy

sample, ‘M’ denotes the samples were melted and solidified in situ to simulate direct casting conditions and ‘S’ denotes the samples were solution treated to
simulate re-heated conventional type cast structures.
b In Model 1 ‘time’ is PWHT time (min) and in Model 2 time is holding time (s) at 1400 ◦ C.
c In Model 1 ‘slow’ is cooling from 570 ◦ C at 250 ◦ C/h to 400 ◦ C and then still air, ‘fast’ cooling is cooling in still air from 570 ◦ C. In Model 2 ‘cooling’ is

t8/5 cooling time (s) and in Model 3 cooling is the cooling rate (◦ C/min) of the samples.
538 Z. Sterjovski et al. / Journal of Materials Processing Technology 170 (2005) 536–544

Table 2
Key neural network model parameters for the three back-propagation models
Model 1 Model 2 Model 3
Hidden layers 1 1 1
PEs in hidden layer 5 12 4
PEs in input layer 10 16 9
PEs in output layer 1 1 2
Learn rule Delta rule Delta rule Delta rule
Fig. 1. Schematic representation of a back-propagating ANN. Six and three Learn sequence Random Random Random
processing elements are shown, respectively, in the input and hidden layers. Learn cycles 150000 500000 500000
Transfer mode Sigmoid Hyperbolic tangent Sigmoid
strength of microalloyed steels during straightening in the Cases in training file 73 63 177
Cases in testing file 13 12 44
continuous casting process (Model 3). The input variables Epoch size 73 63 170
and ranges used for the models are shown in Table 1. For Learning coefficient (HL) 0.80 0.80 0.75
all models the input variables can be classified into two Learning coefficient (OL) 0.40 0.60 0.45
major areas: original material characteristics (chemical com- Learning coefficient ratio 0.85 0.85 0.80
position, plate thickness and/or hardness) and process/test Transition point 10000 10000 10000
Momentum 0.35 0.35 0.35
conditions (temperature, cooling rate and/or holding time).
Three data files were used in the modelling process,
namely the training file, testing file and the sensitivity file. The reached. The network type back-propagation derives its name
data are normalised between 0 and 1, or −1 and 1 depending from this method of calculating the error term [7].
on the intended transfer function, which modifies the sum of


the weightings into an output. For sigmoid transfer the data  
1
are normalised between 0 and 1, and for hyperbolic tangent RMSi =  (P(ij) − Tj )2 (1)
transfer the data are normalised between −1 and 1. The sen- n
j=1
sitivity file is generated by the neural network software upon
request. where P(ij) is the value predicted by the individual program
The initial step in building the models was to set the num- i for sample case j (from n sample cases) and Tj is the target
ber of processing elements (Fig. 1). Individual processing value for sample case j [8].
elements in the input layer correspond to an input variable Random weightings are assigned to each processing ele-
(listed in Table 1), and the processing element in the out- ment as an arbitrary starting point in the training process
put layer corresponds to a desired output. The role of the and the weightings are progressively altered on exposure to
hidden layer, which is a layer that contains a systematically numerous repetitions of training examples. Hence, the neural
determined number of processing elements, is to establish network is in the process of learning and constant adjustment
the relationships and inter-relationships between processing of weightings between processing elements leads to a reduc-
elements in the input and hidden layers that lead to the best tion in the RMS error of the overall training data. In the three
estimates of output. The latter stages in building the model models, 150,000–500,000 cycles were used to minimize and
include defining other parameters, such as selecting the learn stabilize the RMS error. The model reaches its optimum when
rule (the mathematical method the program uses to corre- the RMS error has stabilized. The model is verified against
late the processing elements in the entire model), learning cases in the test data file, which are independent of cases in
coefficients (values assigned to the output and hidden layers, the train data file. The predicted results are plotted versus the
and also the overall model) and transfer mode (transferring experimental (or actual) results.
the weighted sum of the processing elements into an output Finally, sensitivity data were used to further validate the
prediction) (see Table 2). Table 2 also lists the number of model and to interrogate the model to try to gain a physical
training cases used for each model and the epoch size, which understanding of the predicted trends. The sensitivity factor
is the number of training cases the model goes through before (SF) measures the contribution of each input variable to the
it readjusts the weightings of the processing elements that desired output, and it differs for each individual case in both
determine the value of the output. the train and test data files. Therefore, the SF presented in this
The learn rule used in each model is the delta rule and paper is an average of the contribution of an individual input
the error in the output layer is the difference between the variable towards the specified output. The SF for each data
desired output and the actual output. The error calculated case is generated by the Neuralworks Professional II/Plus
for the model was the root mean square (RMS) (Eq. (1)). Software, and this is then averaged to give an indication of the
This error, transformed by the derivative of the transfer average SF of an input variable. For further insight, the SF for
function, is “back-propagated” to prior layers where it is temperature for each prediction of ROA in the training file is
accumulated. This back-propagated and transformed error plotted versus test temperature in Model 3, and the sensitivity
becomes the error term for that prior layer. The process of factors for carbon and niobium for each case in the training
back-propagating the errors continues until the input layer is file are plotted versus the SF for carbon in Model 2.
 Z. Sterjovski et al. / Journal of Materials Processing Technology 170 (2005) 536–544
2.2. PWHT simulation and impact toughness of
quenched and tempered pressure vessel steels for the
data in Model 1
PWHT of 11 mm 700PV grade, 12 mm 700 grade and
20 mm 700PV grade quenched and tempered steels was car-
ried out in a box furnace. Heat treatment conformed to
AS4458-1997-Pressure Equipment Manufacture, and was
conducted in an argon atmosphere. The temperature of
treatment was 570 ± 5 ◦ C (validated by two thermocouples
attached to each plate), and the holding time was 30 min for
the 11 and 12 mm plates and 50 min for the 20 mm plate. The
ramp-up rate was 200 ◦ C/h and samples were then cooled in
argon at 250 ◦ C/h or in still air.
Charpy V notch samples were machined and tested to
AS1544.2-1989 from the middle of the plate. Parent plate
impact test data, transverse and longitudinal to the direction
of rolling, were collected to build an artificial neural network
capable of predicting impact toughness as function of PWHT
time, chemical composition and test conditions. Five samples
were used for each group of tests.
2.3. HAZ simulation and hardness testing of in-service
welding of pipeline and tap fitting steels for the data in
Model 2
The simulation of the HAZ was carried out in a quench
dilatometer. Tubular dilatometer samples with 5 mm outside
diameter, 3.5 mm inside diameter and 10 mm in length were
held between two quartz tubes and heated by a water-cooled
induction coil under vacuum. A Type ‘S’ thermocouple spot-
welded to the sample was used to monitor and control the
temperature of the sample. Eleven samples of various grades
(up to X70 grade) of Australian and USA pipeline and tap
fitting steels were heated to a peak temperature of 1400 ◦ C and
cooled at various rates (t8/5 of 2–10 s) to simulate the coarse
grained HAZ resulting from in-service welding. The t8/5 (s)
is a commonly used parameter in the pipeline industry and it
is defined as the time taken for a weldment to cool from 800
and to 500 ◦ C during in-service welding, and the equivalent
linear cooling rate was maintained down to 150 ◦ C. Although
typical arc-welding process thermal cycles ramp-up rates are
at least 200 ◦ C/s [9], the ramp-up rate for HAZ simulation
was 50 ◦ C/s due to the limitations of the dilatometer. Holding
times were varied (0–2 s) to observe the effect of holding time
on hardness. Finally, Vickers hardness testing was carried
out using loads of 5 kg (as-received material) or 10 kg (HAZ
simulated material). Five indentations were averaged for each
type of test.
2.4. Hot ductility and hot strength of microalloyed steels
for the data in Model 3
Transverse cracking in microalloyed steels during
straightening in the continuous casting process is a problem
often encountered in industry, especially for thin slab casters.
539
Conditions for hot tensile tests were chosen to simulate both
conventional casting (200–250 mm thick slabs) and thin slab
casting (50–100 mm thick slabs) conditions. A Gleeble 3500
thermomechanical simulator was used to perform laboratory
tensile tests, which have been shown to correlate well with
the problem of transverse cracking [10,11].
Four microalloyed steels (Nb, Nb–Ti, Ti and C–Mn–Al) of
various chemical compositions were examined and the ten-
sile test specimens were cylindrical with a 10 mm diameter
and a gauge length of 95 mm. Some samples were melted in
situ, under a compressive strain and quartz tube containment
to maintain the cross-sectional area of the melted sample.
Other samples were solution treated at 1330 ◦ C. In both cases,
cooling rates of either 100 or 200 ◦ C/min down to the test tem-
perature range of 1100–700 ◦ C, corresponding to typically
used straightening temperatures, were used. Samples were
held for 1 min at the test temperature and then strained to
failure at approximately 7.5 × 10−4 s−1 . After failure, sam-
ples were immediately water quenched. A minimum of four
measurements were made on each fracture half to determine
the reduction of area. The ultimate tensile strength (MPa) is
referred to as the hot tensile strength in this paper.
3. Results
3.1. Model 1
Fig. 2 shows a plot of predicted impact energy versus
experimental impact energy for both the train and test data for
Model 1. There is no overlap between the two sets of data and
good predictions occur in both cases. The RMS error of the
predictions for the entire data (train and test data combined)
is 6.23 J.
3.2. Model 2
Fig. 3 shows a plot of predicted HAZ hardness versus
experimental HAZ hardness for both the train and test data
for Model 2. There is no overlap between the two sets of data
and excellent predictions occur in both cases. The RMS error
of the predictions for the complete data is 11.69 HV.
Fig. 2. Predicted impact energy vs. measured impact energy for Model 1
(RMS error of 6.23 J).
540 Z. Sterjovski et al. / Journal of Materials Processing Technology 170 (2005) 536–544

Fig. 3. Predicted HAZ hardness vs. measured HAZ hardness for Model 2
(RMS error of 11.69 HV).

3.3. Model 3

Figs. 4 and 5, respectively, show a plot of predicted ROA

versus experimental ROA, and a plot of predicted hot tensile
strength versus experimental hot tensile strength, for both the
train and test data for Model 3. There is no overlap between
the two sets of data and reasonable predictions occur in both
cases, particularly for hot tensile strength (Fig. 5). The RMS
errors for the predictions for the entire data set are 7.79% for
ROA and 8.68 MPa for hot tensile strength.

3.4. Sensitivity results

Fig. 6 is a plot of the average SF for selected key input

variables in each ANN model. Positive values of the SF in

Fig. 6. SF vs. key input variables in the train data for (a) Model 1, (b) Model
2 and (c) Model 3. The average SF values error bars indicating a scatter of
±1 S.D.

Fig. 6 indicate an increase in property value with an increas-

Fig. 4. Predicted ROA vs. measured ROA for Model 3 (RMS error of 7.79%).
Fig. 5. Predicted hot tensile strength vs. measured hot tensile strength for
Model 3 (RMS error of 8.68 MPa).
ing value of the input variable. Negative values of SF predict
a decrease in property value with an increasing value of the
input variable. Fig. 7(a) is a plot of actual ROA versus test
temperature and Fig. 7(b) is a plot of the SF for test tem-
perature for each data case in the prediction of ROA and hot
strength of Model 3, which has a large standard deviation
for the prediction of ROA (see Fig. 6(c)) due to negative and
positive sensitivity factors. The calculated sensitivity value
varies for each data case and the scatter range is also shown
in Fig. 6 for selected variables from the three models. Fig. 8
is a plot of the SF in each data case for carbon versus the
SF for either original hardness or niobium in Model 2. This
graph illustrates a strong inter-relationship between the sen-
sitivity factors for carbon level and as-received hardness, and
a non-definitive relationship between the factors for niobium
and carbon.
 Z. Sterjovski et al. / Journal of Materials Processing Technology 170 (2005) 536–544
Fig. 7. (a) Typical ROA (%) vs. test temperature and (b) test temperature SF
for ROA and hot strength (obtained from train data) vs. actual test tempera-
ture (Model 3).
Fig. 8. Plot of sensitivity factors of carbon (C), as-received hardness (HV)
and niobium (Nb) in relation to predicted hardness for the train data of
Model 2. The SF for carbon, SF(C), is plotted against the sensitivity factors
for hardness, SF(HV), and niobium, SF(Nb).
4. Discussion
4.1. Model 1
ANN models have, to varying degrees, been successfully
used to predict various properties of steels in relation to mate-
rials characteristics and process/test conditions. In Model 1,
a back-propagation neural network using the delta learn rule
and sigmoid transfer function has been employed to success-
fully predict impact energy in relation to a number of input
parameters. The input parameters for this model included
material thickness, cumulative PWHT time, PWHT cooling
type, C content, Mn content, S content, Cr content and B
content, test orientation and test temperature (Table 1). The
levels of agreement for the training data and test data sets
are shown in Fig. 2. The errors in predictions for both sets of
data are relatively low, with the RMS error for the combined
data set at 6.23 J. This error is lower than the average over-
541
Fig. 9. Comparison of the predicted and measured HAZ hardness values for
the ANN model and the regression equation proposed by Yurioka et al. [14].
all error of the entire Charpy data, which was 7.5 J (standard
deviation) [12].
4.2. Model 2
A back-propagation ANN using the delta learn rule and
hyperbolic tangent transfer function was used to accurately
predict the simulated HAZ hardness values of a number of
pipeline and tap fitting steels with respect to 16 input param-
eters. The input parameters for this model included various
alloying elements, cooling time (t8/5 ) and holding time at
peak temperature (see Table 1). The levels of agreement for
the training data and test data sets are shown in Fig. 3. The
RMS error for the predictions was quite low, 11.69 HV, even
more apparent when the results of this model are compared
to results of other models published in the literature [13,14].
For example, a regression equation developed by Yurioka
et al. [14] is relatively poor in predicting the current data
(Fig. 9). This equation is based on Tekken y-groove welding
tests and therefore different cooling conditions to those asso-
ciated with the dilatometer data. Consequently, the predicted
hardnesses are systematically lower. However, the scatter in
predicted hardness is much higher than that of the neural
network model.
4.3. Model 3
In the last case, a back-propagation neural network model
using the delta learn rule and sigmoid transfer function was
successfully used to predict two outputs, ROA (hot ductility)
and hot tensile strength, with respect to a number of input
parameters as listed in Table 1. The levels of agreement for
the training data and test data sets are shown in Figs. 4 and 5.
The RMS error for the prediction of ROA is significantly
higher than that for hot tensile strength, with the RMS error
for the complete data set at 7.79% for ROA (15.6% error
on average ROA of 50%) and 8.68 MPa for the hot tensile
strength (8.7% error on an average strength of 100 MPa). The
scattered appearance in the ROA results in Fig. 4 is predom-
inantly due to the method used to measure reduction of area
(vernier calliper measurements of diameter) and the reduced
542 Z. Sterjovski et al. / Journal of Materials Processing Technology 170 (2005) 536–544
confidence in the accuracy of the melt samples due to the
difficult nature of the test.
4.4. Sensitivity analysis
Sensitivity analysis was used to further validate each
model and examine the contribution of an input variable to the
output for each individual case in the train data. Fig. 6 shows
the average SF of selected input variables for each model.
As mentioned, positive values of the SF in Fig. 6 indicate an
increase in the value of the output with an increasing value
of the input variable. Negative values of the SF predict a
decrease in the value of the output with an increasing value
of the input variable.
Fig. 6 shows that carbon content is a significant input vari-
able for all of the models. Carbon is known to have a strong
influence on hardness, strength and ductility [15]. Increasing
the carbon content of steel leads to increased hardness and
reduced ductility [16], and hence reduced impact energy. The
carbon content in neural network Model 1 ranged from 0.15
to 0.19%, and the average sensitivity calculated across this
range was −8.6% (Fig. 6(a)), which implies that an increase
in C content decreases the impact energy (over the specified
C range).
The carbon content values used in Model 2 ranged from
0.07 to 0.29 wt%, and the SF calculated along this range was
85 (Fig. 6(b)). This predicted effect of carbon on hardness in
the model is confirmed by Irvine and Pickering [17]. Their
work shows a significant increase in hardness with increas-
ing carbon content of various steels that have martensitic and
bainitic microstructures, which are the dominant microstruc-
tures of the simulated HAZ samples in the current work.
Increasing carbon content increases the hardenability of steel
[18], thus reducing the capacity to form softer transformation
products (such as upper bainite and ferrite) and promoting the
formation of a predominantly martensitic structure, which
is significantly harder than ferrite or upper bainite. In fully
martensitic structures, increases in hardness with increasing
carbon content are mainly due to the increased concentration
of carbon in supersaturated solution [19]. Carbon is inter-
related with the hardness of the as-received (normalised)
steels (Fig. 8) and the consequences of this relationship are
discussed later.
Temperature and holding time at temperature typically
have a significant impact on the mechanical properties of
steels. Fig. 6(a) shows that impact energy decreases with
cumulative PWHT time, and this is attributed to significant
coarsening of second phase carbide particles, which accel-
erates high strain rate fracture [20]. In Model 2, increasing
cooling time, t8/5 (s), from 2 to 10 s (or decreasing cooling
rate from 150 to 30 ◦ C/s) results in a SF of −17 (Fig. 6(b)).
This suggests that slowing down the cooling rate (increasing
cooling time) will result in a significant drop in hardness,
although the effect is not as significant as decreasing carbon
content. The reduction in hardness as a result of decreasing
cooling rate is due predominantly to martensitic microstruc-
tures undergoing autotempering during cooling through the
Ms-Mf range, or the autotempering of bainitic structures
formed at higher temperatures. At still lower cooling rates
softer transformation products will be produced (such as fer-
rite and bainite formed at higher temperatures).
Test temperature has the most significant effect on impact
energy of all the input variables in Model 1 (Fig. 6(a)). In
its learning phase, the model was able to predict a ductile to
brittle transition temperature curve [12], hence further val-
idating the model. A decrease in hot tensile strength with
an increase in test temperature is also an expected result in
Model 3 (Fig. 6(c)), due to increased mobility of dislocations
at higher temperatures associated with decreased resistance
to dislocation glide and climb, and the generation of a failure
mechanism at a lower stress level [21].
Test temperature has a complex effect on the ROA, evi-
dent by the large value for standard deviation in the SF for
test temperature (Fig. 6(c)). This large scatter is due to the
presence of different phases at temperatures below ∼900 ◦ C
compared to temperatures above ∼900 ◦ C. Below ∼900 ◦ C,
an increase in test temperature in the presence of dual phases
(ferrite and austenite) results in a decrease in ROA (indicated
by the predominantly negative SF values in Fig. 7(b)). As test
temperature moves away from 900 ◦ C and towards 700 ◦ C
the sensitivity factors in Fig. 7(b) indicate an increased neg-
ative effect on ROA. At temperatures where austenite is the
only phase present, above ∼900 ◦ C, the sensitivity factors on
average show that an increase in test temperature results in
an increase in ROA (indicated by the positive SF values in
Fig. 7(b)). This trend is well documented and is schemati-
cally shown in Fig. 10 [10]. The high sensitivity of SF to test
temperatures between 900 and 700 ◦ C reflects data associated
with variable volume fractions of ferrite and austenite, and
indicates that caution should be exercised in predicting trends
in ROA from the model over this range of temperatures.
An in-depth understanding of ANNs is required to use
them effectively as design tools and a comprehensive under-
standing of the subject being modelled is also required.
Analysis of the neural network model can become difficult
Fig. 10. Schematic diagram showing the three characteristic ductility
regions on a hot ductility curve for a low carbon steel [10].
 Z. Sterjovski et al. / Journal of Materials Processing Technology 170 (2005) 536–544
in the presence of inter-related input variables. Optimizing
parameters (process or design) when input variables are not
inter-related is simply a matter of determining the SF for a
particular variable and appropriately reducing or increasing
the magnitude of this variable to achieve the desired out-
put. However, interdependent input variables can introduce a
higher degree of complexity in the interpretation of sensitiv-
ity data.
Fig. 8 shows the SF for wt% carbon associated with
the train data of Model 2 plotted against the SF for as-
received steel hardness, and also against the SF for niobium.
Although the variables of C and Nb contents are independent,
an inter-relationship exists between carbon and as-received
steel hardness. As-received hardness is a property that relates
quantitatively to composition, microstructure, grain size and
segregation effects. In particular, the hardness of these nor-
malised low alloy steels will largely depend on carbon content
(and hence volume fraction of pearlite). The presence and
absence of inter-relationships between input variables are
indicated by the relationships between the sensitivity fac-
tors in Fig. 8: direct proportionality for SF(C) and SF(HV),
essentially no relationship for SF(C) and SF(Nb). Although
the incorporation of both as-received hardness and C content
as input variables might appear to be redundant and unnec-
essary, the as-received hardness was incorporated into the
model to provide a more complete set of input variables that
relate to the starting material characteristics and hence are rel-
evant to the accurate prediction of HAZ hardness. Therefore,
the quantitative contribution of carbon to HAZ hardness was
considered by using both the carbon content and the start-
ing (normalised) hardness as input variables, with the aim
of gaining improved insight into the carbon levels required
to achieve an acceptable HAZ hardness value. This exer-
cise demonstrates that the interdependence of input variables
does not compromise the capacity of the ANN model to pre-
dict HAZ hardness, nor to establish, for the purpose of steel
design, the carbon level required to limit the HAZ hardness
level, for a given cooling rate.
In summary, ANNs can eliminate the need for detailed
experimental investigation into processes and even other
areas, such as inspection and field trials, through their abil-
ity to accurately predict the mechanical properties of steels.
Model 1 shows that if the history of relevant data for in-
service pressure vessels (PWHT temperature, cumulative
PWHT time, chemical composition, wall thickness, etc.)
is recorded, then the need for inspection and testing may
only be required if predicted impact toughness values reach
or approach non-conformance limits in the relevant Aus-
tralian standards. Similarly, Model 2 shows that through
prior knowledge of material composition and the relation-
ship between weld heat input and weldment cooling rates
(also dependent on pipe wall thickness and internal fluid flow
rates), neural networks can predict HAZ hardness to provide
knowledge of susceptibility of microstructure to cracking.
This can therefore eliminate the need for extensive in-service
welding trials and/or hardness testing of each weld after in-
543
service welding. Finally, Model 3, which is based on data
generated from experimental simulations, can assist in deter-
mining the optimal temperature to avoid transverse cracking
in microalloyed steels during straightening. The potential
for transverse cracking, in this situation, can be minimized
by maximizing hot ductility (increasing ROA) and optimiz-
ing the temperature of straightening in the ANN, which is
also effected by the composition of the microalloyed steel.
From a practical viewpoint, ANNs can be used to predict the
likelihood of transverse cracking in the actual straightening
process. For example, a characterization model could be used
to predict the likelihood of transverse cracking based on the
input variables used for Model 3 (materials composition, test
temperature and cooling rate) and other casting variables,
such as slab thickness, mould exit temperature and casting
speed.
5. Conclusions
For the ranges of the input variables examined the three
back-propagation neural network models successfully pre-
dicted impact toughness, HAZ hardness, ROA and hot tensile
strength for variations in as-received material characteristics
and process/test parameters. It was also shown that ANNs
could successfully predict multiple mechanical properties.
The results of the sensitivity analysis were in agreement
with both findings of the experimental investigations and
reports in the literature. Manipulation of the sensitivity data
showed the complex nature of ANNs and their ability to solve
complex patterns. The presence of inter-related variables does
not compromise model predictions, but interpretation of the
predicted trends in sensitivity data must be approached with
caution to effectively optimize design or process parameters.
ANNs have the potential to minimize the need for expen-
sive experimental investigation and/or inspection of steels
used in various applications, hence resulting in large eco-
nomic benefits for organisations.
Acknowledgements
The current work was conducted as part of projects spon-
sored by the Cooperative Research Centre for Welded Struc-
tures (CRC-WS) and its core partner, the Australian Pipeline
Industry Association (APIA). CRC-WS was established and
is supported by the Australian Government Cooperative
Research Centres Program. The authors would also like to
thank the Australian Research Council and Bluescope Steel
for their support.
References
[1] D. Graupe, Principles of Artificial Neural Networks—Advanced
Series on Circuits and Systems, vol. 3, World Scientific Publish-
ing, Singapore, 1997.
544 Z. Sterjovski et al. / Journal of Materials Processing Technology 170 (2005) 536–544
[2] S. Alter, Information Systems—A Management Perspective, third
ed., Addison Wesley Longman, 1999.
[3] T. Cool, H.K.D.H. Bhadeshia, D.J.C. MacKay, The use of an arti-
ficial neural network to investigate the yield and ultimate tensile
strength of steel welds, Mater. Sci. Eng. A A223 (1–2) (1997)
186–200.
[4] L.X. Kong, P.D. Hodgson, D.C. Collinson, ISIJ Int. 38 (10) (1998)
1121–1129.
[5] H. Tsuei, Analysis and Modelling of Weld Metal Mechanical Prop-
erties in Flux Cored Arc Welded Steels, PhD Thesis, University of
Wollongong, 2000.
[6] Z. Sterjovski, D. Nolan, D. Dunne, J. Norrish, Predicting the HAZ
hardness of pipeline and tap fitting steels with artificial neural net-
works, in: Proceedings of the Fourth International Conference on
Pipeline Technology, vol. 3, Ostend, Belgium, May 9–13, 2004, pp.
1223–1233.
[7] Neural Ware Incorporated, Neural Computing Guide, Technical Pub-
lications Group, Pittsburgh, U.S.A., 2003.
[8] http://www.gepsoft.com/gepsoft/APS3KB/Chapter09/Section1/SS04.
htm, November 2004.
[9] J.F. Lancaster, Metallurgy of Welding, sixth ed., Abington Publish-
ing, Cambridge, 1999.
[10] B. Mintz, S. Yue, J.J. Jonas, Hot ductility of steels and its relationship
to the problem of transverse cracking during continuous casting, Int.
Mater. Rev. 36 (5) (1991) 187–217.
[11] K.R. Carpenter, B.A. Parker, C.R. Killmore, Hot ductility of Nb,
Ti and Nb–Ti microalloyed C–Mn–Al steels, in: Conference Pro-
ceedings: Thermomechanical Processing: Microstructure, Modelling
and Control, IMMPETUS, University of Sheffield, UK, June 23–26,
2002, pp. 270–274.
[12] Z. Sterjovski, D. Dunne, S. Ambrose, Modelling of repeatedly post-
weld heat-treated quenched and tempered steel, Australas. Weld. J.
48 (2nd Quarter) (2003) 41–46.
[13] M. Painter, In-Service Welding of Gas Pipelines—Project 96:34 Final
Report, Cooperative Research Centre for Welded Structures, Wollon-
gong, Australia, June 2000.
[14] N. Yurioka, H. Suzuki, S. Oshita, Weld. J. 62 (6) (1983) 147–153.
[15] American Society for Metals, Metals Handbook, vol. 6, ninth ed.,
1983.
[16] C.A. Siebert, D.V. Doane, D.H. Breen, The Hardenability of
Steels: Concepts, Metallurgical Influences and Industrial Applica-
tions, American Society for Metals, Metals Park, OH, 1977.
[17] K.J. Irvine, F.B. Pickering, J. Iron Steel Inst. 187 (1957) 292.
[18] R.W.K. Honeycombe, Steels: Microstructures and Properties, Edward
Arnold, London, 1981, p. 131.
[19] D.R. Askeland, The Science and Engineering of Materials, second
ed., Chapman and Hall, London, 1990, p. 263.
[20] Z. Sterjovski, D. Dunne, S. Ambrose, Multiple PWHT of submerged
arc welded quenched and tempered pressure vessel steels, in: Con-
ference Proceedings: International Institute of Welding (IIW) Asian
Pacific International Congress, Suntec Centre, Singapore, October
2002.
[21] R.E. Reed-Hill, R. Abbashian, Physical Metallurgy Principles, third
ed., PWS-KENT, Boston, 1992.