VOLUME 17
ISSUE 1-2
of Achievements in Materials July-August
and Manufacturing Engineering 2006
An artificial intelligence approach in designing
new materials
W. Sitek*, J. Trzaska, L.A. Dobrzański
a Division of Materials Processing Technology and Computer Techniques
in Materials Science, Institute of Engineering Materials and Biomaterials,
Silesian University of Technology, ul. Konarskiego 18a, 44-100 Gliwice, Poland
* Corresponding author: E-mail address: wojciech.sitek@polsl.pl
Received 15.03.2006; accepted in revised form 30.04.2006
Analysis and modelling
Abstract
Purpose: The paper presents the computer aided method of chemical composition designing the metallic
materials with a required property.
Design/methodology/approach: The purpose has been achieved in two stages. In the first stage a neural
network model for calculating the Jominy curve on the basis of the chemical composition has been worked out.
This model made possible to prepare, in the second stage, a representative set of data and to work out the neural
classifier that would aid the selection of steel grade with the required hardenability.
Findings: Obtained results show that AI tools used are effective and very useful in designing new metallic
materials.
Research limitations/implications: The presented models may be used in the ranges of mass concentrations of
alloying elements presented in the paper. The methodology presented in the paper makes it possible to add new
grades of steel to the models.
Practical implications: The worked out models may be used in computer systems of steel selection and
designing for the heat-treated machine parts.
Originality/value: The use of the artificial intelligence method, particularly the neural networks as a tool for
designing the chemical composition of steels with the required properties.
Keywords: Design; Computational material science; Artificial intelligence methods
manufacturing and forming method of engineering materials. One can
1. Introduction
1. Introduction observe gradual leaving the empirical way of mastering/learning of
the reality for the advantage of new, mainly computerised, methods
Since many years the investigations in the area of materials using mathematical model of the object of investigations.
science have been carrying out to develop new materials, among them The paper presents the computer aided methods of chemical
tool ones with higher working properties and fulfilling economical composition designing the metallic materials with a required property.
and ecological requirements. The main aim of this activity is the
development of materials for suitable application having a few of the
following required features: hardness, heat resistance, high- 2. D esigning
temperature creep resistance, wear resistance, corrosion resistance, 2. composition
Designing thethe chemical
chemical composition
of steel with
impact strength, elasticity, strength, life, ecological manufacturing
technology, recycling, low cost and other. Fulfilling this demand is of steel with required
required hardenability
hardenability
extorted by the changes still speeded up in global free-market
economy where the role of inexpensive series/mass production grows The new method is dedicated for the conventional carburising
up significantly. The evolution of science at the end of the 20th and heat-treatable steels. Carbon, silicon, manganese, chromium,
century take effect in the development of new research, nickel and molybdenum are the main alloying elements used in
© Copyright by International OCSCO World Press. All rights reserved. 2006 Short paper 277
Journal of Achievements in Materials and Manufacturing Engineering Volume 17 Issue 1-2 July-August 2006
the carburising and heat-treatable groups of conventional alloy framework of each steel class for which the required hardenability
constructional steels considered. The chemical composition curve is within the experimental hardenability band and the
calculations were assumed to be made basing on the given relevant chemical compositions of the actual heats are included in
hardenability curve shape, presented as the successive hardness Table 1. Figure 1 presents the graphical comparison of the
values at 15 fixed distances from the Jominy specimen face. required hardenability curve and the experimental ones for the
Initial classification of steels was done to obtain a high steel heats with the designed chemical composition. Basing on
conformity of the computational results with the experimental such calculations made for about 550 testing industrial heats it
data. The basis of the classification is the value of the alloy factor was found out that the neural network model developed secures
(AF) describing digitally the fraction of alloying elements in steel the satisfactory adequacy with experimental data since in each
according to standard ASTM A255. The problem is discussed in case the calculated coefficient of adequacy assessment s is smaller
detail in [1]. Basing on the investigations carried out, it was found than its critical value 2.5 HRC.
out that classification of steels into three classes within the
framework of each group of carburising and heat-treatable steels Table 1
is sufficient to obtain a good conformity of the calculations with Comparison of the calculated and the relevant chemical compositions
the experimental data of the chemical composition. of the actual heats
For designing of the chemical composition of the steel with
Mass concentration, %
the required hardenability, unidirectional multilayer neural
networks were employed with the learning method based on the Required Alloying
calculated 1 calculated 2 calculated 3
error backpropagation algorithm. Fifteen input nodes and 6 output curve element
ones assumed in the network structure are the consequence of the actual 1 actual 2 actual 3
assumption that the hardenability of the steels’ analysed is
affected mainly by the concentration of six basic alloying 0.20 0.24 0.22
C
elements. And additionally, hardenability curve is plotted by the 0.18 0.23 0.26
values of hardness measured at fifteen successive points in fixed 0.91 0.80 0.59
Mn
distances from Jominy specimen face. Finally, after preliminary 0.95 0.78 0.60
tests, the 15-30-6 network model was assumed for the 0.29 0.26 0.23
Si
calculations, with the learning coefficient Ș=0.15 and momentum 0.28 0.29 0.20
parameter Į=0.3. Networks with such structures were trained I
0.93 0.59 1.01
individually for each steel class, using a data set prepared basing Cr
0.95 0.53 1.06
on the results of the experiments carried out. The neural networks 0.12 0.53 0.18
developed were experimentally verified, which consisted in the Ni
0.12 0.45 0.16
evaluation of the conformity of the computational results 0.25 0.32 0.22
(obtained by using the network models) with the experimental Mo
0.23 0.32 0.21
data. As a criterion of the evaluation a coefficient of assessment
0.41 0.40 0.42
of the computation method adequacy s was employed. The C
coefficient defines the difference between the required 0.41 0.40 0.41
hardenability and the one obtained for an actual heat. As a result 0.60 0.77 0.79
Mn
of the investigations performed, the limiting value 2.5 HRC of the 0.68 0.72 0.69
coefficient s was assumed [2]. 0.25 0.29 0.26
Si
Verification procedure for such a model consists in 0.28 0.31 0.36
II
calculating the chemical composition of the steel with the 0.75 1.01 1.02
Cr
required Jominy curve shape and in making the heat of the steel 0.74 1.03 1.06
with the chemical composition calculated. Then, the relevant 1.29 1.29 0.23
hardenability investigation is carried out and the actual Ni
1.35 1.35 0.26
experimental hardenability curve of the heat is compared to the 0.16 0.18 0.07
required Jominy curve shape. For experimental verification Mo
0.16 0.17 0.07
hardenability curves with the assumed and distinctly different
shapes were selected. Calculations of the chemical composition
for the curves with the required shape were made within the 3. Designing the chemical
framework of a particular steel grade only when the required 3. Designing
compositionthe chemical
of steel composition
with of
hardenability curve was within the experimental hardenability
band for the class considered. Then, investigations of
steel with the assumed hardness after
the assumed hardness
after cooling from the
hardenability of the heats with the actual chemical compositions cooling from the austenitising
austenitising temperature
temperature
the nearest to the calculated ones were made. Hardenability
curves’ shapes, the required and the actual ones, were compared The method presented in the paper makes it possible to
afterwards. As an example of the calculations made, the results determine the mass concentrations of the alloying elements for
for two of the required shapes of hardenability curves are steels with the required curve of hardness changes versus cooling
presented (curve I for carburising steel, curve II for heat-treatable rate obtained during the continuous cooling from the austenitising
steel). The chemical compositions calculated within the temperature.
278 Short paper W. Sitek, J. Trzaska, L.A. Dobrzański
Analysis and modelling
50 60
40 50
Hardness, HRC
Hardness, HRC
30 40
20 30
10 20
0 10
1 1
50 50
Distance from quenched-end, mm Distance from quenched-end, mm
required curve I experimental curve 1 required curve II experimental curve 1
experimental curve 2 experimental curve 3 experimental curve 2 experimental curve 3
Fig. 1. Comparison of the required hardenability curves and the experimental ones of the steels’ with designed chemical composition
Designing the optimum chemical composition is carried out in training parameters were specified analyzing the effect of these
three stages: quantities on the network performance coefficient values for the
x preparing the database containing information on mass test set. The number of training epochs was determined by
concentrations of the elements, observing the network forecast error for the training and
x calculating the austenitising temperature (temperature validating sets. The model developed was subjected to the
Ac3+50°C) using the neural network model described in [3], numerical verification using the data that were not used in its
x calculating hardness of steel cooled continuously from the development. The network with two hidden layers and numbers of
austenitising temperature for various cooling rates using the neurons in these layers as twenty and two was assumed to be
model developed employing the neural networks, optimal. Table 3 presents error values and correlation coefficients
x selecting the chemical composition of steel meeting the for calculated hardness. Training method was used based on the
assumed criterion. conjugate gradient algorithm. The detailed problem description
The data set was developed basing on literature data, was presented in [4,5].
including chemical compo-sitions, austenitising temperature (TA) To prepare the database containing the information about the
and the CCT diagrams of the constructional and engineering randomly selected chemical compositions of steel, taking into
steels. The obtained curves were worked out, assuming mass account limitations presented in Table 2, the computer program
fractions of the alloying elements as the criterion. Basing on the was developed generating random chemical compositions of steel
collected data it was assumed in addition that total of the mass basing on user specified parameters:
fractions of manganese, chromium, nickel, and molybdenum does x range of mass concentrations for each element,
not exceed 5%. The ranges of the assumed mass fractions of x number of cases,
elements and austenitising temperature are included in Table 2. x maximum sum of the selected elements’ concentrations,
To develop the relationship between the chemical composition, x additional parameter (cooling time from the austenitising
austenitising temperature, and cooling rate, and hardness of the temperature to the ambient temperature).
constructional steel the feedforward neural network (MLP) was Austenitising temperature was determined as the Ac3+50°C
used. The data was divided into four sets: training, validating, test, temperature for the prepared set of 6500 various chemical
and verifying one. The training set was used for development of the compositions of steel and next hardness was calculated for ten
neural network model, the validating set was used for checking the assumed average cooling rates. Three numerical procedures were
model during establishing the values of weights, and the verifying developed making it possible selection of the optimum chemical
set was used for verifying the model when the network training was composition in respect to one of the three search criteria:
completed. Allocation of data to the particular subsets was done x required hardness for the assumed cooling rate,
randomly. The number of vectors was determined in the particular x required hardness in the assumed range of cooling rate changes,
sets: 1582, 791, 790, and 369. The activation level of the successive x required curve of hardness change in the entire range of
fourteen network input nodes depended on: mass concentration of cooling rate changes.
elements (C, Mn, Si, Cr, Ni, Mo, V, Cu), austenitising temperature, In the first method, the chemical composition of steel is
cooling rate, and structure type. The type of structure developed searched, for which the module value of difference between the
after cooling the steel at a particular rate was specified using four calculated and expected hardness (for the assumed cooling rate) is
binary nominal variables. the smallest. In the second method, the chemical compositions of
Hardness was determined basing on the activation level of steels are searched, for which hardness is within the range defined
a single neuron in the network output layer. The number of hidden by specifying the minimum and maximum values at the assumed
layers and number of nodes in these layers, and also method and cooling rate range.
An artificial intelligence approach in designing new materials 279
Journal of Achievements in Materials and Manufacturing Engineering Volume 17 Issue 1-2 July-August 2006
Table 2.
Ranges of mass concentrations of elements and austenitising temperature for the analysed steels
Mass concentrations of elements, % Austenitising
Range
C Mn Si Cr Ni Mo V Cu temperature TA, qC
min 0.08 0.13 0.12 0 0 0 0 0 770
max 0.77 2.04 1.90 2.08 3.65 1.24 0.36 0.3 1070
Table 3.
Error values and correlation coefficients for hardness calculated for data from the training/validating/testing /verifying data sets
Standard deviation of the Quotient of standard Pearson correlation
Data set Error EHV, HV
error, HV deviations coefficient
Training 28.7 27.2 0.24 0.97
Validating 34.8 35.1 0.29 0.96
Testing 36.1 37.3 0.31 0.95
Verifying 38.4 38.5 0.32 0.95
Table 4.
Chemical composition of steel calculated for the assumed versus cooling time from the austenitising temperature
Cooling time, s 3 20 50 100 250 500 1000 5000 2103 105
Predetermined hardness, HV 550 540 480 420 360 320 290 230 210 200
Calculated hardness, HV 560 552 473 417 353 312 277 240 215 207
Error, HV 9.8 12.0 7.4 2.9 7.2 8.2 12.6 10.2 5.1 7.1
Ferrit No Yes Yes Yes Yes Yes Yes Yes Yes Yes
Pearlit No No No No No No No Yes Yes Yes
Bainit Yes Yes Yes Yes Yes Yes Yes No No No
Martensit Yes Yes Yes Yes Yes Yes Yes No No No
Predicted mass concentration of alloying elements, %
C Mn Si Cr Ni Mo V Cu TA
0.2 0.63 0.43 0.53 0.21 0.24 0.29 0.03 912
In the third method, calls for specifying the expected hardness
for the successive ten values of cooling time from the austenitising
4. Final
4. Finalremarks
remarks
temperature to 100°C. The paper presents some examples of application of artificial
The chemical composition of steel is searched, for which the intelligence tools, i.e. neural networks and genetic algorithms in
sum of the absolute differences between the calculated and expected designing of new materials with required properties. Results of
hardness, for the successive cooling time values, is the smallest. numerical simulation show that AI tools used are effective and
Figure 2 present comparison of the required curve of hardness very useful in designing new metallic materials.
changes versus cooling time with the curve calculated for the
assumptions presented in Table 4.
600
References
References
Pred ete rmin ed hardne ss
[1] L.A. DobrzaĔski, W. Sitek: “Application of neural network in
Calculated hardness
500 modelling of hardenability of constructional steels”, Journal of
Materials Processing Technology, 78 (1998) 59-66.
Hardness, HV
400
[2] L.A. DobrzaĔski, W. Sitek: “The modelling of hardenability
using neural networks”, Journal of Materials Processing
Technology, 92-93 (1999) 8-14.
300 [3] L.A. DobrzaĔski, J. Trzaska: “Application of neural networks
for prediction of critical values of temperatures and time of
200 the supercooled austenite transformations”, Journal of
Materials Processing Technology, 155-156 (2004) 1950.
[4] J. Trzaska PhD Thesis, Silesian University of Technology,
100 Poland, 2002.
1 10 100 100 0 100 00 100000
[5] L.A. DobrzaĔski, J. Trzaska: “Application of neural networks for
Time, s prediction of hardness and volume fractions of structural
components in constructional steels cooled from the austenitizing
Fig. 2. Comparison of the assumed and calculated curves temperature”, Material Science Forum, 437–438 (2003) 359-362.
280 Short paper READING DIRECT: www.journalamme.org
