Journal of Materials Processing Technology 192193 (2007) 582587

Applications of the artificial intelligence methods for modeling of the

ACAlSi7Cu alloy crystallization process
L.A. Dobrzanski a, , R. Maniara a , J. Sokolowski b , W. Kasprzak a ,
M. Krupinski a , Z. Brytan a
a

Institute of Engineering Materials and Biomaterials, Konarskiego Str. 18A, 44-100 Gliwice, Poland
b NSCER/Ford/University of Windsor, Windsor, Onatario, Canada

Abstract
In this work the derivation analysis techniques is applied for the preparation of ACAlSi7Cu alloy thermal characteristic, and assessed influence
of cooling rate for characteristic transformations during the crystallizations. Alloys were cooled with three cooling rate: 0.2, 0.5 and 1 C/s. Rise of
the cooling rate is influenced on a dendrite phase nucleation point which has an effect on a solid fraction. In this case alloys are much more uniform.
The changes of crystallization temperature at the end of the process are not observed during the rise of the cooling rate. The neural networks were
used for modeling of the ACAlSi7Cu alloy crystallization process cooled with different cooling rate.
2007 Elsevier B.V. All rights reserved.
Keywords: Neural network; Derivation analysis; Thermal modification; Quantities assessment of the crystallization process

1. Introduction
Cast aluminum alloys are important construction material,
which are used in various fields of technology. Because of their
low density, relatively low melting point, good heat and electrical
conduction, low thermal expansion coefficient, good castability
and low casting shrinkage, they are mainly used in car manufacturing as: piston castings, cylinder head castings, engine blocks,
structural supporting reinforcements and elements absorbing
crash impact. Aluminum alloys are also widely used in production of household goods, as well as in telecommunication
and information technology.
In order to determine various physical and metallurgical
parameters of solidified aluminum alloys (e.g. number of fractions of solid in any solidification time, degree of superfusion
of liquid, solidification front velocity and microstructures) it is
necessary to find out the precise the temperature of phase transformation of solidified alloys of certain chemical content. As
it has been proved by many research centers, the temperatures
change not only along with the change of chemical composition but also with the change of thermodynamic conditions of
solidified alloys.

Corresponding author.
E-mail address: leszek.dobrzanski@polsl.pl (L.A. Dobrzanski).

0924-0136/$ see front matter 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.jmatprotec.2007.04.022

The findings of research presented in articles [19] show that

the changes of solidification rate results in changes of temperature of phase transformation of the examined alloy. That leads
to increasing or decreasing of range of temperatures in which
alloy undergoes solidification, therefore to change in proportion
of fraction of solid and liquid occurring while commencing successive phase transformations, which triggers directly change
of all alloy microstructural features including grain size, dendrite arm spacing and intermetallic phases and improves silicon
modification level.
High purity AlSiCu hypoeutectic alloys exhibit three
main solidification reactions during the solidification process. Starting with the formation of aluminum dendrites
followed by the development of two main eutectic phases.
The presence of alloying and impurity elements such as: Cu,
Mg, Mn, Fe leads to more complex constituents (including
intermetallic) that are characterized by metallographic techniques. Backerud et al. [79] identified the reactions in AC
AlSiXCuX, (3XX) alloys and listed four solid-state phases
(Table 1).
The solidification path for AC AlSiXCuX, (3XX) alloy can
be described as follows:
1. A primary -aluminum network forms between 620 and
580 C. The exact temperature depends mainly on the amount
of Si and Cu concentration in the alloy.

L.A. Dobrzanski et al. / Journal of Materials Processing Technology 192193 (2007) 582587
Table 1
Reactions occurring during the solidification of the AC AlSiXCuX (3XX) alloys
according to Refs. [3,4,79]
No.

Reaction

Temperature [ C]

1
2
3
4

Development of dendritic network

L + +AlFe5 Si
L + + Al8 FeMg3 Si6 + Mg2 Si
L + Al2 Cu + + Al5 Cu2 Mg8 Si6

620580
570555
540500
500470

583

Table 2
Range of mass concentration of elements and cooling rates (CR) in investigated
AlSi hypoeutectic alloys
Range

Min
Max

CR ( C/s)

Mass concentration of elements (wt%)

Si

Fe

Cu

Mn

Mg

Zn

Ti

5
11.9

0.08
1.35

0.006
4.64

0
0.6

0.02
1.28

0
2.9

5
11.9

0.03
2.5

%Sr + %Ni + %Sn + %Pb + %Na%Ca 0.3.

2. Within the 570555 C the eutectic mixture of Si and aluminum forms, leading to a further localized increase in
Cu content of the remaining liquid. The Fe rich phases can
also precipitate in this temperature range.
3. At approximately 540 C the Mg2 Si and Al8 Mg3 FeSi6
phases begin to precipitate.
4. A reduction in the temperature allows for precipitation of
Al2 Cu and Al15 Mg8 Cu2 Si6 phases between 500 and 470 C.
The main characteristic of neural networks is the ability
to generalize knowledge for new data not yet provided during learning. It is not necessary for neural networks to gather
data and have access to entire database containing knowledge
on predicted issue. They also show tolerance to discontinuity,
random disturbances or shortage in learning set. Such characteristics allow their application where there are problems with
data transformation, analysis, sorting and classification along
with prediction or controlling on their basis of a certain process.
Research carried out by Prof. L.A. Dobrzanskis team [10,11]
has shown that neural network is a useful tool to predict materials engineering phenomena. They have facilitated formation of
numerical models describing phases kinetic transformations in
working steel. Results obtained from neural networks turned out
more precise than the ones obtained from empirical calculations.
This article presents the application of neural networks for
calculation of cooling rate (CR) influence on: aluminum nucleation temperature (TND ), dendrite coherency point temperature
(TDCP ), + nucleate temperature (TE(Al+Si)N ) and end of solidification process temperature (TSOL ) for the AC AlSi7Cu alloy.
Additionally in this work neural networks applied to the calculation of grain size (GS), secondary dendrite arm spacing (SDAS),
yield strength (YS), apparent elastic limit (EL), hardness (HRF)
and matrix microhardness (HV0.025).
2. Materials and experimental procedure

has been divided into two subsets: the learning set and the validation set. The
data have been divided in proportion of 75% for the learning set and 25% for
the validation set. The division into the sets has been made at random and
the arithmetic mean and standard deviation from the assumed value have been
observed.
In order to control the learning correctness of neural networks data obtained
from a metallurgical experiment have been used. The results of the experiment
have been used to make a test-verification set. The data used in the process of
learning and testing have been normalized by means of the minimax function
that transforms the domains of variables to range 0, 1.

2.2. Materials for testing neural networks

The data obtained from the metallurgical experiment have been used for
testing neural networks. For the sake of research in the metallurgical laboratory of the University of Windsor there had been made six AlSi experimental
hypoeutectic alloys cooled with three different cooling rates.
The alloys of chemical content shown in Table 3 were casted in an electrical
resistance furnace with a ceramic melting crucible. The process of melting took
place in the nitrogen protective atmosphere to prevent the liquid metal from
the diffusion of oxygen and hydrogen. Directly before casting of ingots, melted
metal was degassed for 20 min in order to eliminate from the alloy hydrogen
atoms as one of the learning vectors which influenced the liquidus and solidus
temperatures and was undesirable in the investigated alloys. The melting and
soaking temperature of the alloy was 850 C. Ingots of 45 mm in diameter and
50 mm in height were cast into thin-walled cups made of hot-work tool steel.
The ingot mass was 330 10 g.

2.3. Thermal analysis procedure

In order to determine exactly the phase transformation temperatures of the
investigated alloys thermoanalysis (TA) has been made by means of the UMSA
device equipped with data-measuring and recording system made by National
Instruments Company, attached to a PC computer [14].
The device has been calibrated according to American Regulations of
National Institute of Standards and Technology. Openings of 2 mm in diameter and 25 mm in length from the upper surface of the ingot have been centrally
made in the samples in order to install thermocouple (Fig. 1). The thermocouple
of the K type made by OMEGA Company has been installed for measuring of
the temperature. The samples have been placed in the heatingcooling system
and the lower and upper parts of the cups have been isolated by ceramics in order

2.1. Materials for training neural networks

Formation of numerical model that allows calculation of the characteristically temperature, microstructural features and mechanical properties depending
on chemical composition and cooling rate has been determined by working out
a suitable set of experimental data.
The set of representative data including: mass concentration of elements,
TND , TDCP , TE(Al+Si)N TSOL , SDAS, GS, EL, YS, HRF and HV0.025 has been
elaborated through own research, the results made available by and specialist publications [3,79,13]. The range of concentration of alloy elements and
cooling rates have been presented in Table 2.
The gathered set of data designed for formation of a numerical model determining the solidus temperature in relation to chemical composition, cooling
rate, TND , TDCP , TE(Al+Si)N TSOL , SDAS, GS, EL, YS, HRF and HV0.025

Table 3
Chemical content of alloys used for neural networks testing
Alloy label

71
72
74
91
92
94

Mass concentration of elements (wt%)

Si

Fe

Cu

Mn

Mg

Zn

Ti

7.17
6.98
7.45
9.09
9.03
9.27

0.14
0.17
0.34
0.72
0.19
0.17

0.99
1.91
3.60
1.05
2.25
4.64

0.11
0.01
0.25
0.36
0.01
0.01

0.27
0.26
0.28
0.27
0.19
0.28

0.05
0.43
0.05
0.14
0.45
0.05

0.08
0.09
0.13
0.07
0.10
0.09

%Sr + %Ni + %Sn + %Pb + %Na%Ca 0.95.

584

L.A. Dobrzanski et al. / Journal of Materials Processing Technology 192193 (2007) 582587
where E is the mean network error, N the number of data in test set, Xmi the i-times
calculated value and Xpi is the i-times calculated value by neural network.
Standard deviation quotient for errors and data have been adopted as a quality
quotient of a numerical model made by neural networks. The numerical model
of relationship between the phase transformation temperature, microstructure
feature and mechanical properties in relation to chemical concentration and
cooling rate calculated by neural networks can be accepted as a correct one
provided that the output values given by the network contain smaller errors than
a simple calculation of unknown output value.
The simplest method for calculation of output value is still assumption of
an average value of output values for it learning and testing sets. In that case,
mean error is equal of standard deviation for output value in the learning set,
whereas standard deviation quotient equals one. The smaller network error is,
the smaller values for the standard deviation quotient become, finally reaching
zero for the ideal forecast [1012].

3. Results and discussion

Fig. 1. Schematic diagram of ingot cross section used for preparing compression
test specimens with dimensions of the tensile test specimen (all dimensions are
in mm).
to enforce the direction of solidification process from the cup walls towards the
centre.
To ensure exact measurement of temperature changes in time during the
cooling process, the investigated alloys have been melted and held in the temperature of 750 C for 15 min.The samples have been cooled at appropriate rate
by means of compressed air provided through nozzles of the heatingcooling
system. In the experiment three types of cooling rates have been applied:
=0.2,

=1 and
=2.5 C/s.

2.4. Compression, hardness and microhardness testing

Samples for compression testing were machined from center of the thermal
analyses specimen ingots (Fig. 1). The machined samples were polished with fine
sandpaper to remove any machining marks from the surface. Compression tests
were conducted at room temperature using a Zwick universal testing machine.
Prior to testing, a precision alignment device was used to minimize framebending strains. Compression specimens were tested corresponding to each of
the three cooling rate.
Rockwell F-scale hardness tests were conducted at room temperature using
a Zwick HR hardness-testing machine. Vickers microhardness tests were conducted using a DUH 202 microhardness-testing machine. Load of indenter was
set at 25 g.

This paper presents the technique of determination of phase

transformation temperature in the crystallization process, structural properties and basic mechanical properties depending on
the cooling rate and chemical composition with the application
of neural networks. To solve this problem the computer system
was designed. It undertakes the issue of predication of chemical
composition and the cooling speed on the characteristic temperatures of phase transitions occurring during the solidification
process, structural properties and mechanical properties. Fig. 2
shows the flow chart of the realized calculation.
The calculation of phase nucleation temperature (TDN ),
phase dendrites coherence temperature (TDKP ), + eutectic
nucleation temperature (TE(Al+Si)N ) and end-of-crystallization
temperature (TSol ) was possible thanks to the application of the
radial basis functions neural network with net k-mean and knearest neighbour learning algorithms with the search option of
the optimal ones in input parameter subset where the input data
vector is composed of six parameters for chemical composition
and cooling rate. The designed neural network consists of four
outputs for the calculated temperatures. The standard deviation
quotient for the applied neural network resulting from the training set calculations is TDN = 0.21, TDKP = 0.08, TE(Al+Si)N = 0.22
and TSol = 0.46. The values of errors, standard deviation ratios

2.5. Image analyses

Microstructural changes (GS, SDAS) were quantified using a Leica QWinTM and a Simagis ResearchTM Image Analysis System, image analyzer in
conjunction with a Leica optical microscope. In each case, 21 fields were analyzed over the entire sample surface. From these readings, the average GS and
SDAS was determined for each chemical composition and each cooling rate.

2.6. Evaluate of the neural networks

Variety of neural networks used in the experiment makes it necessary to
apply number indicators, which allow the evaluation of the learning process.
The following indicators have been regarded as the essential ones: mean network
error, standard deviation quotient for errors and data, standard deviation for the
error and the Pearson correlation coefficient (R2 ).
The mean network error has been calculated according to the following:
1
(|Xmi Xpi |)
n
n

E=

i=1

(1)

Fig. 2. Flow chart for calculation of phase transformation temperature,

microstructure features and mechanical properties.

L.A. Dobrzanski et al. / Journal of Materials Processing Technology 192193 (2007) 582587

585

Table 4
Quality assessment coefficients for applied neural networks for calculate of phase transformation
Calculated temperature ( C)

Average of tested
population

Absolute mean
error

Standard error
deviation

Standard deviation
quotient

Pearson correlation
coefficient

TDN
TDCP
TE(Al+Si)
TSol

606.22
584.21
568.14
471.36

2.21
1.87
1.09
2.82

2.94
1.18
1.37
3.44

0.21
0.08
0.22
0.46

0.98
1.00
0.98
0.89

Fig. 3. Comparison of experimental determination TDN and TDN calculated by

applied neural network.

Fig. 5. Comparison of experimental determination TE(Al+Si)N with the TE(Al+Si)N

calculated by applied neural network.

and Pearson correlation coefficients (R2 ) for the values in this

research are shown in Table 4.
Figs. 36 show the comparison of real temperatures characteristic to the crystallization process for testing alloys with

temperatures calculated by neural network. The biggest error

in the predication of chemical composition and cooling speed
on the temperatures characteristic to the crystallization process
occurs for the determination of solidus temperature for experimental alloys. In 30% of cases the difference between the real
temperature and the temperature calculated by the neural nets

Fig. 4. Comparison of experimental determination TDCP with TDCP calculated

by applied neural network.

Fig. 6. Comparison of experimental determination TSol with the TSol calculated

by applied neural network.

586

L.A. Dobrzanski et al. / Journal of Materials Processing Technology 192193 (2007) 582587

Table 5
Quality assessment coefficients for applied neural networks for calculate of microstructure features

SDAS (m)
GS (m)

Average of tested
population

Absolute mean error

Standard error deviation

Standard deviation
quotient

Pearson correlation
coefficient

54.69
3181.12

4.37
352

6.03
407.3

0.27
0.32

0.96
0.94

Table 6
Quality assessment coefficients for applied neural networks for calculate of mechanical properties
Mechanical properties

Average of tested
population

Absolute mean error

Standard error
deviation

Standard deviation
quotient

Pearson correlation
coefficient

HV0.025
HRF
EL (MPa)
YS (MPa)

84.241
76.915
138.79
203.46

5.066
3.227
12.54
9.05

2.94
1.18
1.37
3.44

0.628
0.342
0.32
0.28

0.94
0.97
0.96
0.97

exceeds 8 C. However, this error can be acceptable when taking into account the range of measurement of the calculated
temperature and while comparing the error to the experimentally
measurable error value.
Radial basis functions neural network with k-nearest neighbour learning algorithms was applied to calculate of SDAS and
GS. The values of errors, standard deviation quotient and Pearson correlation coefficients (R2 ) for the values in this research
are shown in Table 5.
Radial basis functions neural network with k-nearest neighbour learning algorithms as well as pseudoinversion were
applied for the predication of the mechanical properties in the
functions of chemical composition and cooling speed. The standard deviation ratio calculated for the training set is: 0.628 for
microhardness; 0.343 for hardness, 0.32 for elastic limit and 0.28
for yield strength. Table 6 shows the values of errors, standard
deviation ratios and Pearson correlation coefficients (R2 ) for the
calculated values of microhardness, hardness, elastic limit and
yield strength.
Synergy of alloying elements influence and cooling rate
on the phase transformation occurring during alloy solidification with constant specified cooling rate afford the influence
of single elements do not fully present their real performance.
Also it is quite difficult to determine the alloy liquidus temperature and solidus temperature depending on the applied
chemical composition and cooling rate. The designed methodology enables the evaluation of the influence of the chemical
composition and the change of the cooling rate on the change
of mechanical and structural properties and enables the determination of phase transformation temperature of solidifying
alloy.
Thanks to the designed computer system enabling the predication of chemical composition and cooling rate on phase
transformation temperatures occurring in the crystallization process of the hypoeutectic AlSiCu alloys it was possible to
perform simulation of the influence of cooling rate on: TDN ,
TDKP , TE(Al+Si)N , and TSol for the experimental AC AlSi7Cu
alloy. The assumed chemical composition for the tested alloy
is presented in Table 7. The calculation results are presented in
Fig. 7.

Table 7
Chemical content of alloys (AC AlSi7Cu) used for prediction of the cooling rate
influence on the phases transformation temperature, microstructure feature and
mechanical properties
Mass concentration of elements (wt%)
Si

Fe

Cu

Mn

Mg

Zn

Ti

7.00

0.15

1.00

0.10

0.2

0.05

0.08

Fig. 7. Predication of the cooling rate on phase transition temperatures for the
AC AlSi7Cu alloy.

Fig. 8. Predication of the cooling rate on mechanical properties for the AC

AlSi7Cu alloy.

L.A. Dobrzanski et al. / Journal of Materials Processing Technology 192193 (2007) 582587

Fig. 8 shows the influence of the cooling speed on HRF, HV,

EL, YS changes of the tested alloy.
4. Conclusion
This paper presents the application of neural networks of the
basic parameters characterising the crystallization process of the
hypoeutectic AlSi casting alloys. The neural networks models
applied in this research enable the determination of phase transition temperatures depending on the chemical composition and
cooling rate which is the basic parameter affecting the structure
quality and mechanical properties in the result.
The designed computer system enables also the reverse
task. It is possible to determine the chemical composition and
mechanical properties on the basis of the temperatures characterising the crystallisation process of the AlSi casting alloys
determined by means of the thermal derivative analysis in
the production process. Thanks to this implementation we can
reduce control time of metal quality applied in the casting process. Also, in a small degree, it reduces the costs related to quality
control. This solution is currently tested in one of the aluminium
casting production companies.
At the same time the works connected with the development
of the designed system continue by adding a module enabling
the determination of time when characteristic phase changes
take place. Thanks to this model it will be possible to calculate
the cooling curve and CTPc graphs for aluminium alloys.
References

[2]
[3]

[4]

[5]

[6]

[7]
[8]
[9]
[10]

[11]

[12]

[13]
[14]

[1] L.A. Dobrzanski, R. Maniara, M. Krupinski, W. Kasprzak, J. Sokoowski,

Applications of the derivation analysis for assessment of the ACAlSi7Cu

587

alloy crystallization process cooled with different cooling rate, COMMENT 2005, in: Proceedings of the 13th International Scientific
Conference, Gliwice, 2005.
J. Gilbert Kaufman, E.L. Rooy, Aluminium Alloy Castings, AFS, 2005.
M.B. Djurdjevic, W.T. Kierkus, J.H. Sokolowski, Analysis of the Solidification Path of the 3XX Family of Aluminum Alloys, Technical
Report Submitted to the NEMAK Canada Corporation, Windsor, October
2002.
R. MacKay, M. Durdjevic, J. Sokolowski, The effect of cooling rate on
the fraction solid of the metallurgical reaction in the 319 alloy, AFS Trans.
(2000).
S.G. Shabestari, H. Moemeni, Effect of copper and solidification conditions
on the microstructure and mechanical properties of AlSiMg alloys, J.
Mater. Process. Technol. (2004).
J.M. Boileau, J.W. Zindel, J.E. Allison, The Effect of Solidification Time
on the Mechanical Properties in a Cast A356-T6 Aluminum Alloy, Society
of Automotive Engineers, Inc., 1997.
L. Backerud, E. Krol, J. Tamminen, Solidification Characteristics of Aluminum Alloys, vol. 1, Universitetsforlaget, Oslo, 1986.
L. Backerud, G. Chai, J. Tamminen, Solidification Characteristics of Aluminum Alloys, vol. 2, AFS, 1992.
L. Backerud, G. Chai, Solidification Characteristics of Aluminum Alloys,
vol. 3, AFS, 1992.
J. Trzaska, L.A. Dobrzanski, Application of neural networks for designing
the chemical composition of steel with the assumed hardness after cooling from the austenitising temperature, in: 13th International Scientific
Conference AMME 2005, Gliwice, 2005.
J. Trzaska, L.A. Dobrzanski, Modeling of transformations occurring during
quenching in engineering steels, in: 3rd International Scientific Conference
MMME 2005, Gliwice, 2005.
S. Malinov, J.J. McKeown, W. Sha, Modelling the correlation between processing parameters and properties in titanium alloys using artificial neural
network, Comput. Mater. Sci. vol. 21 (2001).
W.T. Kierkus, M.B. Durdjevic, J.H. Sokolowski, Analysis of the Solidification of the 3XX Family of Aluminium Alloys, 2002, data not published.
http://www.uwindsor.ca/umsa.