See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/380890185

Predicting the mechanical properties of stainless steels using Artificial Neural

Networks

Article in Production Engineering Archives · May 2024

DOI: 10.30657/pea.2024.30.21

CITATIONS READS

0 37

6 authors, including:

Djordje Ivković Dusan Arsic

University of Kragujevac University of Kragujevac
11 PUBLICATIONS 10 CITATIONS 155 PUBLICATIONS 569 CITATIONS

SEE PROFILE SEE PROFILE

Dragan Adamovic Ružica R. Nikolić

University of Kragujevac University of Žilina, Slovakia
69 PUBLICATIONS 407 CITATIONS 261 PUBLICATIONS 885 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Dusan Arsic on 27 May 2024.

The user has requested enhancement of the downloaded file.

 PRODUCTION ENGINEERING ARCHIVES 2024, 30(2), 225-232

PRODUCTION ENGINEERING ARCHIVES

ISSN 2353-5156 (print) Exist since 4th quarter 2013
ISSN 2353-7779 (online) Available online at https://pea-journal.eu

Predicting the mechanical properties of stainless steels

using Artificial Neural Networks
Djordje Ivković1 , Dušan Arsić1* , Dragan Adamović1 , Ružica Nikolić2* , Andjela Mitrović1 ,
Otakar Bokuvka3
1 Faculty of Engineering, University of Kragujevac, Sestre Janjić 6, 34000 Kragujevac, Serbia; [email protected] (DI);
[email protected] (DA); [email protected] (DA); [email protected] (AM)
2 Research Centre, University of Žilina, Univerzitna 8215/1, 010 26 Žilina, Slovakia; [email protected]
3 Faculty of Mechanical Engineering, University of Žilina, Univerzitna 8215/1, 010 26 Žilina, Slovakia; [email protected]

*Correspondence: [email protected]; [email protected]

Article history Abstract

Received 29.02.2024 Knowing the material properties is of a crucial importance when planning to manufacture some struc-
Accepted 15.04.2024 ture. That is true for the steel structures, as well. Thus, for the proper planning of a certain steel part
Available online 31.05.2024 or a structure production, one must be aware of the properties of the material, to be able to make a
Keywords qualified decision, which material should be used. Considering that the manufacturing of steel prod-
Stainless steel, ucts is constantly growing in various branches of industry and engineering, the problem of predicting
Yield stress, the material properties, needed to satisfy the requirements for the certain part efficient and reliable
Tensile strength, functioning, becomes an imperative in the design process. A method of predicting four material prop-
Hardness, erties of the two stainless steels, by use of the artificial neural network (ANN) is presented in this
ANN. article. Those properties were predicted based on the particular steels’ known chemical compositions
and the corresponding material properties available in the Cambridge Educational System EDU PACK
2010 software, using neural network module of MathWorks Matlab. The method was verified by com-
paring the values of the material properties predicted by this method to known values of properties for
the two stainless steels, X5CrNi18-10 (AISI 304), X5CrNiMo17-12-2 (AISI 316). The difference be-
tween the two sets of values was below 5% and, in some cases, even negligible.
DOI: 10.30657/pea.2024.30.21

1. Introduction problems, such as material loss, as well as sudden compo-

nents’ failures that can have catastrophic consequences, not
Application of steels in technics began to grow with the dis- only regarding financial losses, but endangering human lives,
covery of Bessemer’s process, back in 1856. From then on, as well (Jovanović et al., 2017).
almost all new structures and machines began to incorporate The solution to this problem was found in 1913 with discov-
steel parts, extending the service life of parts in that way. ering of steels that are resistant to corrosion, the so-called
Nowadays, in the so-called conventional engineering, such as stainless steels. The resistance to corrosion comes as a conse-
mechanical, civil, automotive etc., steel is still the most ap- quence of those steels’ chemical composition, consisting of
plied material. One of the reasons for such a situation is the minimum of 10.5 to 12% of chromium dissolved in steel solid
significantly lower price of steels, as compared to other metal solution. At present, there exist thousands of stainless steels
materials, such as aluminum or titanium, or composite mate- grades, carefully engineered to meet the specific requirements
rials, new-nano materials and others. However, there exist of the parts that are to be made of them, all regulated by spe-
some disadvantages that are limiting their even wider applica- cific standards, (EN 10088-1:2005; EN 10088-2:2005; EN
tion. The first and definitely the biggest is phenomenon of cor- 10088-3:2005; EN 10088-4:2009; EN 10088-5:2009). Be-
rosion, i.e., reacting of steel with oxygen, forming the corro- sides chromium, the stainless steels also contain other chemi-
sive layer on the components surface, and, in the more severe cal elements, such as nickel, molybdenum, vanadium and oth-
cases, penetrating deeper into the material bulk. The corro- ers. Stainless steels with adequate addition of niobium are
sion, to put it simple, is an irreversible process of material de- known as austenite (or austenitic) stainless steels, since their
struction. Problems related to corrosion could lead to greater austenitic structure remains even at room temperatures. The
© 2024 Author(s). This is an open access article licensed
under the Creative Commons Attribution (CC BY)
License (https://creativecommons.org/licenses/by/ 4.0/). 225 ARCHIWUM INŻYNIERII PRODUKCJI
 DJORDJE IVKOVIĆ ET AL. / PRODUCTION ENGINEERING ARCHIVES 2024, 30(2), 225-232

reason for that is the action of Ni i Mn, which expand the re- (Mukherjee et al., 1995), mechanical properties of steels (Do-
gion of stable austenite within the Fe-Fe3C phase diagram. brzanski and Sitek, 1999; Knap et al., 2008; Knap et al., 2014),
The microstructure is the main criterion for classification of influence of alloying elements on steel properties (Sitek et al.,
the stainless steels: ferritic, austenitic, austenitic-ferritic (the 2022), modelling of high-speed steels‘ properties (Sitek et al.,
so-called duplex steels), martensitic, precipitation hardening. 2004), etc.
Differences between mechanical and physical properties of
carbon and stainless steels are important for the structural 2. Literature review
safety. They include principal material properties (Young
modulus, yield stress and ultimate strength), as well as stain- Qamar and Ali Zardari (Qamar and Ali Zardari, 2023) have
less steel behavior (the stress-strain curve i.e., the σ–ε dia- recently summarized all the basic data on artificial neural net-
gram, behavior during the cold working or at elevated temper- works, explaining the ANN’s basics, the fundamental neuron
atures). Those differences must be considered when the and the artificial computer model. They described the net-
ultimate and serviceability limit states are being formulated work’s 3-layer structure (input layer, hidden layer and output
during the structures design (Tylek and Kuchta, 2014). layer), and network learning and training methods. They em-
When designing any steel structure, one must obey adequate phasized the basic advantages of the ANN applications, adap-
standards (EN 1990:2002; EN 1993-1-1:2005; EN 1993-1- tive learning, self-organization, real-time operation and the
2:2005; EN 1993-1-3:2006; EN 1993-1-4:2006), respecting fault tolerance.
the material properties. The designer must be, in advance, (Sorić et al., 2023) presented a review of machine learning
aware of properties of different steels, so that he can decide methods employing the neural network algorithm, and dis-
which of the steels is the best for manufacturing the certain cussed the most commonly used neural networks, such as
part or a structure. One of the methods for predicting some feedforward NN, including deep learning, the convolutional
material properties of the two stainless steels is presented in NN, the recurrent NN, and the physics-informed NN, with
this paper. The method is based on application of the artificial special emphasis on their applications in solid mechanics.
intelligence, precisely the artificial neural networks. Based on application of ANNs on two simple examples au-
Artificial Neural Networks (ANN) are created from artifi- thors concluded that the NN approaches show lower compu-
cial neural cells (artificial neurons), which are modelled based tational costs, as compared to the finite element method
on the principle of functioning of the biological neural cells (FEM), simultaneously preserving the high accuracy of com-
(bio-neurons). Interpretation of the bio-neuron’s working puting.
principle is the following: all the signals received on dendrites (Ciocan et al., 2000) have considered the recognition of dif-
are summed in the neuron’s body. If a sum of all the signal ferent types of austenitic steels with an ultrasonic system that
values exceeds the critical value, the signal is sent through provided the necessary data for the two different neural net-
axon to further layers and cells. Similar is the interpretation of works. The input vector for the first ANN contained processed
working principle of artificial neurons. The difference be- data (propagation velocity and ultrasonic attenuation), while
tween the two is that in the case of artificial cells, numbers for the second ANN it contained the amplitude of digitized
values are used instead of signals. The artificial neuron’s radio-frequency signal and its numerical Fourier transform.
working principle is as follows: if the value of number re- Two thirds of acquired data for three kinds of steels were used
ceived from other cells, multiplied by value of a weight coef- in the learning process, while the last third was used in the
ficient, and added to a bias value, exceeds some critical value, testing process. Both neural networks gave similar results on
the transfer function is activated and it sends the obtained input data, above 98% of steels classification probability.
value to the next layer. (Kusiak and Kuziak, 2020) presented research of applying
It is of a great importance to emphasize that the artificial the ANN for predicting the volume fraction and mean size of
neural network’s ability to predict some value is based on its the phase constituents in a steel subjected to the thermome-
training. The training of ANN means that adequate set of data chanical processing and cooling. The network was trained on
is supplied to the network. Each data set has two parts, input the data obtained in the laboratory tests, and next validated us-
and output. For each input there is a known output. Based on ing the data from the industrial measurements. Authors pre-
that, the ANN arranges weight and bias values so that the input sented results which show that the prediction of the micro-
data, after summing and multiplying through ANN, gives al- structure and mechanical properties of the considered steel is
ready known output. The smaller the difference (the error) be- in a good agreement with the experimental data. They claimed
tween the known output and calculated output, the better is the that the ANN based model could predict, with good accuracy,
ANN, i.e., the better it is constructed, (Lisjak, 2004). such microstructural features as the ferrite grain size, ferrite
Application of the ANNs covers a wide spectrum of areas fraction, as well as the steel’s yield stress and ultimate tensile
(Qamar and Ali Zardari, 2023), such as medicine (Basheer and strength. The accuracy of values evaluated by the ANN model
Hajmeer, 2000), text recognition and classification (Kim, was much higher than that obtained from calculations using
2023), railway transport (Bursać et al., 2022), geology the classical, experimental models.
(Varenina et al., 2018), solid mechanics (Sorić, Stanić and Le- (Sitek et al., 2004) were modelling the relationship between
sičar, 2023), material types (Ciocan et al., 2000), material the chemical composition and hardenability of structural alloy
composition and microstructure (Kusiak and Kuziak, 2002), steels, using the artificial neural network and multiple regres-
prediction of mechanical behavior of metal matrix composites

ARCHIWUM INŻYNIERII PRODUKCJI 226

 DJORDJE IVKOVIĆ ET AL. / PRODUCTION ENGINEERING ARCHIVES 2024, 30(2), 225-232

sion models, and using large set of experimental data contain- software default number of neurons for the second layer and it
ing required information on the chemical compositions and could be changed. However, for these purposes adequate pre-
corresponding Jominy hardenability curves for over 400 data cision was obtained, thus this number has not been changed.
steel sheets with variety of chemical compositions. Authors Number of neurons in the third (output) layer is equal to num-
demonstrated the full practical usefulness of the developed ber of properties predicted by neural network. In all the cases
model in selection of materials for particular applications with that number was one, as only one property’s values were pre-
intended performance in the area of application. They also dicted.
concluded that the developed ANN model can be used for sim-
Table 1. Chemical composition of the X5CrNi18-10 steel, wt%
ulations of the relationship between hardness at a given dis-
tance from the Jominy bar specimen face and the chemical 1 2 3 4 5 6 7 8 9
composition of the steel, as well. C Cr Mn Ni N P Si S Fe
(Knap et al., 2008) were studying the possibility to predict 0.08 19.0 0 9.0 0.10 0.04 0.7 0.02 Bal.
the hardness profile – hardenability of the constructional steel, Table 2. Chemical composition of the X5CrNiMo17-12-2 steel,
based on its chemical composition, using the ANN model. The wt %
database consisted of hardness profile measurements from the
1 2 3 4 5 6 7 8 9
Jominy test samples and it contained almost twenty thousand
C Cr Mn Ni N P Si S Fe
data vectors, with extensive range of steels chemical compo- 0.025 19.5 1.98 11.4 8.0 0.037 0.72 0.02 Bal.
sitions. The authors reported that, in spite of the great varia- Note: since the first layer has 18 neurons and the chemical composi-
tions of the chemical compositions of each steel grade and the tion contains 9 elements, the neurons 10 to 18 have been assigned the
data base itself, the ANN made very accurate predictions of value of 0
each steel’s hardenability.
In continuation of their research, (Knap et al., 2014) pre- 3.2. Prediction of the yield stress values of austenitic
sented an attempt to model the effect of differences in chemi- stainless steels
cal composition on material hardenability within one steel
grade. They used very broad and heterogeneous database. The neural network was created according to previously ex-
They concluded that, if the database is big enough, predictions plained construction, three layers, 18 neurons in the input
of hardenability would be accurate and of high quality. How- layer, 10 neurons in the hidden layer and 1 neuron in the out-
ever, for a less comprehensive database there appeared differ- put layer – for the yield stress value. The activation function
ences in hardness predictions for various chemical composi- between the layers was chosen to be the same for activation
tions of the considered steel grade. between the first and the second layer, as well as for activation
Application of artificial neural network has been proven to between the second and the third layer, the tan sigmoid tan
be reliable for use in diagnostics of various technical systems function. The Bayesian Regularization algorithm was chosen
(Menasri and Aimeur, 2023). as the training algorithm. Structure of this network is shown
in Figure 1, where W is notation for the weight coefficient and
3. Prediction of stainless steels mechanical prop- b for the bias value.
erties using ANN

3.1. Input and output data preparation

Four properties, the yield stress, tensile strength, elongation
and hardness, of the two stainless steels, X5CrNi18-10 (AISI
304), X5CrNiMo17-12-2 (AISI 316), were predicted using
ANN. Input and output data for training of the artificial neu- Fig. 1. Structure of the neural network for predicting the yield
rons were created using the Cambridge Education Software stress values
Edu Pack 2010 (CES EDU PACK 2010). As four properties The created network was trained using parameters given in
were investigated, four input and four output data sets were Table 3. Application of given parameters resulted with follow-
needed. In all the four cases, properties were predicted based ing regression displayed in Figure 2. Entering the chemical
on chemical composition of 57 other stainless steels available composition values of steels X5CrNi18-10 and X5CrNiMo17-
in the used software, so the input set of data in all four cases 12-2, predicted values of the yield stress, shown in Table 4
was common. The four output sets consisted of values of each were obtained.
investigated property for those 57 other stainless steels. The
complete process of properties prediction was conducted in
MathWorks Matlab software using its Neural Network Mod- Table 3. Values of parameters used for network training
ule. Parameter Value
For the prepared data, four neural networks with three layers Maximum number of epochs 1000
were prepared. The first (input) layer consisted of 18 neurons, Time Infinite
each representing one element from the steel’s chemical com- Goal 0
position. The second layer had 10 neurons. This value was Min gradient 0.0000001

227 ARCHIWUM INŻYNIERII PRODUKCJI

 DJORDJE IVKOVIĆ ET AL. / PRODUCTION ENGINEERING ARCHIVES 2024, 30(2), 225-232

Maximum number of fails 0

Initial momentum value 0.005 Yield stress, MPa
Incline momentum 0.1
280
Decline momentum 10

Yield stress, MPa

258
Maximum momentum value 10000000000 260 247.5
240
240 228
220
200
X10CrNi18-10 X5CrNiMo17-12-2

CES EDU PACK 2010 ANN

Fig. 3. Graphical display of predicted yield stress values

Error, %
6.00% 5.00%
5.00% 4.00%
4.00%

Error, %
3.00%
2.00%
1.00%
0.00%
X10CrNi18-10 X5CrNiMo17-12-2

Fig. 4. Error values for the yield stress values prediction

Fig. 2. Display of obtained regression values r = 0.933 3.3. Prediction of the tensile strength values of aus-
The predicted values of the yield stress are graphically dis- tenitic stainless steels
played in Figure 3. For the X5CrNi18-10 steel obtained value A special neural network was constructed for the tensile
for the yield stress is 247.57 MPa, which is 10.43 MPa smaller strength values predicting. The input data set was the same as
than the experimental value (found in the CES software). As for the case of yield stress prediction. The output data set was
for the steel X5CrNiMo17-12-2, the predicted value of the prepared from the tensile strength values for other stainless
yield stress is for almost 12 MPa smaller. Regarding the error steels available in the CES software.
percentage, for the first steel it is 4% , and for the second it is The neural network structure was built from three layers, as
somewhat bigger, little below 5% Figure 4. Since the error in the previous case. The activation functions between the lay-
values are lower than 5% for both steels, it could be concluded ers were selected differently, with respect to the previous case.
that the results predicted by the ANN are close to the real val- The tan sigmoid function was chosen as the activation func-
ues. Thus, the constructed ANN provided for the reliable re- tion between the first and the second layer, same as in the first
sults and it could be used for predicting the yield stress values network. The linear function was selected for activation be-
of other austenitic stainless steels. tween the second and the third layer. The Levenberg-Mar-
Table 4. Predicted values of the yield stress, MPa quardt algorithm was selected as the training algorithm for this
neural network, the structure of which is presented in Figure
Steel X5CrNi18-10 X5CrNiMo17-12-2 5.
CES EDU PACK 258 240
ANN 247.5739 228.3207

Fig. 5. Structure of the neural network for predicting the tensile

strength values

ARCHIWUM INŻYNIERII PRODUKCJI 228

 DJORDJE IVKOVIĆ ET AL. / PRODUCTION ENGINEERING ARCHIVES 2024, 30(2), 225-232

Created network was trained using parameters given in Ta- bigger than the experimentally obtained value. Regarding the
ble 5. Application of given parameters resulted in the regres- error percentage of the predicted values, for the first steel it
sion presented in Figure 6. The values of the tensile strength was little below 3% and for the second steel it was approxi-
for the two considered steels were predicted by entering chem- mately 0.3%, Figure 8. Since in this case the error values are
ical composition values for steels X5CrNi18-10 and practically negligible, it can be concluded that this ANN is
X5CrNiMo17-12-2 and using values of tensile strength from suitable for predicting the values of tensile strength for other
the CES EDU PACK 2010 software. Results are shown in Ta- austenitic stainless steels, as well.
ble 6.
Table 5. Values of parameters used for network training Tensile strength, MPa
Parameter Value 580 570 572
565

Tensile strength, MPa

Maximum number of epochs 1000
Time Infinite 560 550
Goal 0
Min gradient 0.0000001 540
Maximum number of fails 6 520
Initial momentum value 0.001
X10CrNi18-10 X5CrNiMo17-12-2
Incline momentum 0.1
Decline momentum 10 CES EDU PACK 2010 ANN
Maximum momentum value 10000000000

Fig. 7. Graphical display of predicted tensile strength values

Error, %
4.00%
3.00%
3.00%
Error, %

2.00%

1.00% 0.30%
0.00%
X10CrNi18-10 X5CrNiMo17-12-2

Fig. 8. Error values for prediction of the tensile strength values

3.4. Prediction of the elongation values of austenitic

stainless steels
The neural network for predicting the elongation values of
stainless steels was created in a similar manner as for the first
two cases: three layers with 18, 10 and 1 neurons, respectively.
The input data set, as mentioned, is the same in all the cases.
Fig. 6. Display of obtained regression values r = 0.815 The output data set is prepared from elongation values for
other stainless steels from the CES software. The network
Tensile strength values from the CES software and values structure, regarding the activation functions between the lay-
predicted by the ANN are shown in Figure 7. For the ers, is the same as for the case of predicting the yield stress
X5CrNi18-10 steel the predicted value for tensile strength is values, in both cases the activation function is tan sigmoid
549.7284 MPa, which is almost 17 MPa smaller than the value function, Figure 1. The difference with respect to that case is
obtained by experiment. that the training algorithm is selected to be the Levenberg-
Table 6. Predicted values of the tensile strength, MPa Marquardt algorithm.
The created network was trained using parameters given in
Steel X5CrNi18-10 X5CrNiMo17-12-2 Table 7. Application of given parameters resulted with the re-
CES EDU PACK 565 570 gression values displayed in Figure 9. After entering the chem-
ANN 549.7284 571.825 ical composition values for steels X5CrNi18-10 and
X5CrNi17-12-2 in the neural network, elongation values were
obtained, which are given in Table 8.
On the other hand, the predicted value for the steel
X5CrNiMo17-12-2 is 571.825 MPa, which is almost 2 MPa

229 ARCHIWUM INŻYNIERII PRODUKCJI

 DJORDJE IVKOVIĆ ET AL. / PRODUCTION ENGINEERING ARCHIVES 2024, 30(2), 225-232

Table 7. Values of parameters used for network training

Elongation, %
Parameters Value
Maximum number of epochs 1000 60 53 48

Elongation, %
Time Infinite 40 42
40
Goal 25
Min gradient 0.00000001 20
Maximum number of fails 6
Initial momentum value 0.01 0
Incline momentum 6 X10CrNi18-10 X5CrNiMo17-12-2
Decline momentum 6
Maximum momentum value 10000000 CES EDU PACK 2010 ANN

Fig. 10. Graphical display of predicted elongation values

Error, %
10.00% 9.00%
8.00%

Error, %
6.00% 5.00%
4.00%
2.00%
0.00%
X10CrNi18-10 X5CrNiMo17-12-2

Fig. 11. Error values for elongation values prediction

3.5. Predicting the hardness values of austenitic

stainless steels
The neural network for predicting the hardness values of
stainless steels was created in the same way as in the previous
cases, with the same set of input data and the output data set
prepared from hardness values for other stainless steels from
the CES software.
Fig. 9. Display of obtained regression values r = 0.923 The neural network structure was the same as in the previ-
ous cases. Activation functions between layers were selected
Table 8. Predicted values of elongation, % in the same way as in the case of predicting the tensile strength
Steel X5CrNi18-10 X5CrNiMo17-12-2 values, Figure 5. Between the first and the second layer the
activation function was the tan sigmoid function, while be-
CES EDU PACK 52.5 40 tween the second and the third layer the activation function
ANN 47.9 42.1652 was a linear function. The training algorithm for this case was
the Bayesian Regularization algorithm.
Graphical presentation of the predicted elongation values of The created network was trained using parameters given in
the two stainless steels is shown in Figure 10. The predicted Table 9.
elongation value for steel X5CrNi18-10 is 47.9%, which is al- Table 9. Values of parameters used for network training
most 4.6% smaller than the value obtained by experiment. For
the steel X5CrNiMo17-12-2, the predicted elongation value is Parameters Value
42.17% and it is 2.17% bigger than the experimentally ob- Maximum number of epochs 1000
Time Infinite
tained value. The error value for the first steel is little below
Goal 100
10% (9.6%) and for the second steel it is 5%. Since the values Min gradient 0.000001
predicted by this ANN have errors between 5% and 10%, it Maximum number of fails 0
could be concluded that the predicted values are slightly off Initial momentum value 0.1
with respect to values obtained by experiments. Results ob- Incline momentum 10
tained by this ANN need to be further investigated. The ob- Decline momentum 10
tained error values are displayed in Figure 11. Maximum momentum value 1000000000

ARCHIWUM INŻYNIERII PRODUKCJI 230

 DJORDJE IVKOVIĆ ET AL. / PRODUCTION ENGINEERING ARCHIVES 2024, 30(2), 225-232

Application of given parameters resulted in the regression

values displayed in Figure 12. Entering the chemical compo- Hardness, HV
210
sition values for steels X5CrNi18-10 and X5CrNi17-12-2 in 205

HARDNESS, HV
the neural network, hardness values were predicted for these
two stainless steels that are presented in Table 10. 200 196 195
190
190

180
X10CrNi18-10 X5CrNiMo17-12-2
CES EDU PACK 2010 ANN

Fig. 13. Graphical display of predicted hardness values

Error, %
6.00%
5.00%

Error, %
4.00% 3.00%

2.00%

0.00%
X10CrNi18-10 X5CrNiMo17-12-2

Fig. 14. Error values for hardness values prediction

4. Conclusions
Stainless steels were discovered in 1913 with idea to over-
come one of the greatest disadvantages of structural steels –
Fig. 12. Display of obtained regression values r = 0.799 corrosion. It was found that addition of minimum 12% of
Table 10. Predicted values of hardness, HV chromium to steels solid solution enables steel to become cor-
rosion resistant. Besides Cr, other alloying elements could be
Steel X5CrNi18-10 X5CrNiMo17-12-2 found in various steels chemical composition (Ni, Mo, V etc.).
CES EDU PACK 190 205 One of the artificial intelligence methods for replicating hu-
man thinking process are the Artificial Neural Networks
ANN 195.529 195.234
(ANN). They are made of multiple artificial neuron cells,
which are connected in layers. The most frequent application
The predicted hardness values are shown in Figure 13. For of the ANN is for prediction of values based on known input
steel X5CrNi18-10 the predicted value is 195.529 HV, which values.
is for 5.529 HV bigger than the experimental value. As for the The topic of this paper was to investigate if the ANN could
steel X5CrNiMo17-12-2, the predicted hardness value is predict four mechanical properties (yield stress, tensile
195.234 HV, which is for 9.76 HV smaller than the experi- strength, elongation and hardness) of austenitic stainless
mentally obtained value. Error value for the first steel is 3% steels, with sufficient accuracy. The data sets for the ANNs
and for the second steel it is somewhat below 5% (4.76%), were created from CES EDU PACK software, in which infor-
Figure 14. Both error values are lower than 5%. Thus, it could mation about various steels could be found. The input data set
be concluded that this ANN is suitable for predicting the hard- (with 18 neurons) consisted of chemical composition of con-
ness values of austenitic stainless steels. sidered steels and the output data sets (wit 1 neuron for each
network) were the investigated properties of those steels. The
hidden layer consisted of 10 neurons, which was the default
property of the used software and it was not changed since it
provided for sufficient precision of the network(s).
After the data sets were prepared, the four neural networks
were created and trained. Two stainless steels X5CrNi18-10
(AISI 304) and X5CrNiMo17-12-2 (AISI316) were used to
validate results obtained from the ANNs. Based on obtained

231 ARCHIWUM INŻYNIERII PRODUKCJI

 DJORDJE IVKOVIĆ ET AL. / PRODUCTION ENGINEERING ARCHIVES 2024, 30(2), 225-232

results it can be concluded that for all properties obtained re- https://standards.iteh.ai/catalog/standards/cen/5da77ead-c665-4c16-a063-
1b086a1543c2/en-10088-2-2005
sults were close to previously known experimentally obtained EN 10088-3:2005. Stainless steels – Part 3: Technical delivery conditions for
values, where the error for predicted values of considered semi-finished products, bars, rods, wire, sections and bright products of
properties were under 5% and in some cases negligible. corrosion resisting steels for general purposes.
The slight exception was the result for elongation where the https://standards.iteh.ai/catalog/standards/cen/4e6c80d2-c72d-42b3-aeae-
2564ae23eb38/en-10088-3-2005
error values were between 5 and 10%, which are also low, but
EN 10088-4:2009. Stainless steels – Part 4: Technical delivery conditions for
they cannot be ignored. If one would use this ANN to predict sheet/plate and strip of corrosion resisting steels for construction pur-
elongation of unknown steel it would be necessary to confirm poses
results experimentally. This problem should be further inves- https://standards.iteh.ai/catalog/standards/cen/a9506eec-011c-47b3-9313-
6ec0bb544a7b/en-10088-4-2009
tigated. Adding other input parameters besides the chemical EN 10088-5:2009. Stainless steels – Part 5: Technical delivery conditions for
composition of considered steels, would probably increase the bars, rods, wire, sections and bright products of corrosion resisting steels
precision of predicted values. As for the other ANNs, one for construction purposes
could state that they are able to predict mechanical properties, https://standards.iteh.ai/catalog/standards/cen/affdda76-226f-42d3-b4c7-
5d00eba17733/en-10088-5-2009
with great precision, as the error values are low. Certainly, Ivković, Dj., Arsić, D. Adamović, D., Nikolić, R., Mitrović, A., Bokuvka, O.,
precision would increase if the number input variable types is 2024. Predicting the yield stress and tensile strength of two stainless
increased. steels using artificial intelligence. Proceedings of The 27th International
Seminar of Ph.D. students - SEMDOK 2024, 05-07.02.2024, Western Ta-
tras - Zuberec, Slovakia, 57-62.
Acknowledgements Jovanović, M., Lazić, V., Arsić, D., 2017. Material Science, Faculty of Engi-
neering. University of Kragujevac, Kragujevac, Serbia, ISBN 978-86-
Research presented in this paper was partially financially supported
6335-042-7. (in Serbian)
by the project of Operational Programme Integrated Infrastructure Kim, D., 2023. Text Classification Based on Neural Network Fusion. Tech-
“Support of research and development capacities to generate ad- nical Journal, 17(3), 359-366, DOI: 10.31803/tg-20221228154330
vanced software tools designed to increase the resilience of economic Knap, M., Lamut, J., Rozman, A., Falkus, J., 2008. The prediction of harden-
entities against excessive volatility of the energy commodity mar- ability using neuronal networks. Archives of Metallurgy and Materials,
ket”, ITMS2014+ code 313011BUK9, co-funded by European Re- 53(3), 761-766, DOI: 10.2478/amm-2014-0021
gional Development Fund, and by the project TR35024 of the Min- Knap, M., Falkus, J., Rozman, A., Konopka, K., Lamut, J., 2014, The Predic-
istry of Education, Science and Technological Development of tion of Hardenability using Neural Networks. Archives of Metallurgy and
Materials, 59(1), 133-136, DOI: 10.2478/amm-2014-0021
Republic of Serbia.
Kusiak, J., Kuziak, R., 2002. Modelling of microstructure and mechanical
Note: The shorter version of this research results was presented at properties of steel using the artificial neural network. Journal of Materials
the SEMDOK 2024 conference (reference Ivković et al., 2024). Processing Technology, 127(1), 115–121, DOI: 10.1016/S0924-
0136(02)00278-9
Lee, J-G., Jun, S., Cho, -W., Lee, H., Kim, G. B., Seo, J. B., Kim, N., 2017.
References Deep Learning in Medical Imaging: General Overview. Korean Journal
of Radiology, 18(4), 570-584, DOI: 10.3348/kjr.2017.18.4.570
Basheer, I. A., Hajmeer, M., 2000. Artificial neural networks: fundamentals, Lisjak, D., 2004. Application of various artificial intelligence methods in ma-
computing, design, and application. Journal of Microbiological Methods, terial selection. Doctoral dissertation, Faculty of Mechanical Engineer-
43, 3–31, DOI: 10.1016/S0167-7012(00)00201-3 ing, University of Zagreb, Zagreb, Croatia.
Bursać, M., Jevtić, S., Tričković, G., 2021. Application of artificial neural Menasri, N., Aimeur, N., 2023. Faults diagnostics of cement draft fan using
networks for predictions of failure of railway signaling devices. Proceed- artificial neural network (ANN). Structural Integrity and Life, 23(1), 23-
ings of Third International Conference “Transport for Today’s Society “, 29. http://divk.inovacionicentar.rs/ivk/ivk23/023-IVK1-2023-NM-
14-16.10.2021, Bitola, North Macedonia, Faculty of Technical Sciences NA.pdf
Bitola, 194-197. https://ttsconf.org/wp-content/uploads/2022/04/p45.pdf Mukherjee, A., Schmauder, S., Ruhle, M., 1995. Artificial neural networks
Ciocan, R., Petulescu, P., Ciobanu, D., Roth, D. J., 2000. The use of the neural for the prediction of mechanical behavior of metal matrix composites.
networks in the recognition of the austenitic steel types. NDT&E Inter- Acta Metall. Mater. 43(11), 4083–4091, https://edisciplinas.usp.br/
national 33, 85-89, DOI: 10.1016/S0963-8695(99)00032-8 pluginfile.php/5791436/mod_resource/content/1/Artigo%204.pdf
Dobrzanski, L.A., Sitek, W., 1999, The modelling of hardenability using neu- Qamar, R., Zardari, B. A., 2023. Artificial Neural Networks: An Overview.
ral networks. Journal of Materials Processing Technology, 92-93, 8-14, Mesopotamian journal of Computer Science, 2023, 130-139, DOI:
DOI: 10.1016/S0924-0136(99)00174-0 10.58496/MJCSC/2023/015
EN 1990:2002. Eurocode – Basis of structural design. Sitek, W., Dobrzanski, L. A., Zacłona, J., 2004. The modelling of high-speed
https://www.phd.eng.br/wp-content/uploads/2015/12/en.1990.2002.pdf steels’ properties using neural networks. Journal of Materials Processing
EN 1993-1-1:2005. Eurocode 3 – Design of steel structures – Part 1-1: Gen- Technology 157-158, 245-249, DOI: 10.1016/j.jmatprotec.2004.09.037
eral rules and rules for buildings Sitek, W., Trzaska, J., Gemechu, W. F., 2022. Modelling and Analysis of the
https://www.phd.eng.br/wp-content/uploads/2015/12/en.1993.1.1.2005.pdf Synergistic Alloying Elements Effect on Hardenability of Steel. Archives
EN 1993-1-2:2005. Eurocode 3 – Design of steel structures – Part 1-2: Gen- of foundry engineering, 4, 102-108, DOI: 10.24425/ afe.2022.143957
eral rules – structural fire design Sorić, J., Stanić, M., Lesičar, T., 2023. On neural network application in solid
https://www.phd.eng.br/wp-content/uploads/2015/12/en.1993.1.2.2005.pdf mechanics. Transactions of FAMENA, 47(2), 45-66, DOI:
EN 1993-1-3:2006. Eurocode 3 – Design of steel structures – Part 1-3: Gen- 10.21278/TOF.472053023
eral rules – Supplementary rules for cold-formed members and sheeting Tylek, I., Kuchta, K., 2014. Mechanical properties of structural stainless
https://www.phd.eng.br/wp-content/uploads/2015/12/en.1993.1.3.2006.pdf steels. Technical Transactions Civil Engineering, 4-B(12), 59-80,
EN 1993-1-4:2006. Eurocode 3 – Design of steel structures – Part 1-4: Gen- https://www.ejournals.eu/Czasopismo-Techniczne/2014/Budownictwo-
eral rules – Supplementary rules for stainless steels Zeszyt-4-B-(12)-2014/art/5743/
https://www.phd.eng.br/wp-content/uploads/2015/12/en.1993.1.4.2006.pdf Varenina, A., Malvić, T. Režić, M., 2018. Improved neural network model of
EN 10088-1:2005. Stainless steels – Part 1: List of stainless steels assessment for interpretation of Miocene lithofacies in the Vukovar for-
https://standards.iteh.ai/catalog/standards/cen/952db42f-8160-4518-8932- mation. Northern Croatia, Materials and Geoenvironment, 68(3), 145-
c51bc76f8715/en-10088-1-2005 156, DOI: 10.2478/rmzmag-2018-0029
EN 10088-2:2005. Stainless steels – Part 2: Technical delivery conditions for
sheet/plate and strip of corrosion resisting steels for general purposes

ARCHIWUM INŻYNIERII PRODUKCJI 232

View publication stats