See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/287507996

Computational intelligence based design of

age-hardenable aluminium alloys for different
temperature regimes

Article · February 2016

DOI: 10.1016/j.matdes.2015.12.076

CITATION READS

1 115

5 authors, including:

Swati Dey M. S. Kaiser

Indian Institute of Engineering Science and T… Bangladesh University of Engineering and Te…
15 PUBLICATIONS 80 CITATIONS 59 PUBLICATIONS 109 CITATIONS

SEE PROFILE SEE PROFILE

Partha Dey Shubhabrata Datta

Academy of Technology SRM University
15 PUBLICATIONS 81 CITATIONS 100 PUBLICATIONS 704 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Nanoparticles added lubricants View project

Image analyses of micrographs View project

All content following this page was uploaded by M. S. Kaiser on 03 August 2017.

The user has requested enhancement of the downloaded file.

 Materials and Design 92 (2016) 522–534

Contents lists available at ScienceDirect

Materials and Design

journal homepage: www.elsevier.com/locate/matdes

Computational intelligence based design of age-hardenable aluminium

alloys for different temperature regimes
Swati Dey a,⁎, Nashrin Sultana a, Md Salim Kaiser b, Partha Dey c, Shubhabrata Datta d
a
Indian Institute of Engineering Science and Technology, Shibpur, Howrah 711103, India
b
Bangladesh University of Engineering and Technology, Dhaka 1000, Bangladesh
c
Academy of Technology, Hooghly 712121, India
d
Bankura Unnayani Institute of Engineering, Bankura 722146, India

a r t i c l e i n f o a b s t r a c t

Article history: Computational intelligence based approaches are used in tandem to design novel age-hardenable aluminium
Received 29 June 2015 alloy, which would utilize the effect of all precipitate forming elements together, crossing the limit of the compo-
Received in revised form 12 December 2015 sitions defined within different series. A pool of data is created from the tensile properties of age-hardenable al-
Accepted 14 December 2015
uminium alloys in the 2XXX, 6XXX and 7XXX series. Based on the testing temperature the data is segregated, and
Available online 17 December 2015
different models for the tensile properties in the different temperature regimes are developed using Artificial
Keywords:
Neural Network (ANN). The inherent relation between the composition and processing variables with the me-
Age-hardenable aluminium alloy chanical properties are explored using sensitivity analysis (SA). In order to design alloys with the conflicting ob-
Tensile properties jectives of high strength and adequate ductility, Multi-Objective Genetic Algorithm (MOGA) is used to search
Artificial Neural Network optimum solutions using the ANN models as the objective functions. The Pareto solutions from MOGA and the
Multi-objective optimization SA results are used along with prior knowledge of the alloy systems to design age-hardenable aluminium alloys
Genetic algorithm with improved mechanical properties at different temperature regimes. The designed composition, which is be-
Alloy design yond any of the age-hardenable series, has been developed experimentally, with encouraging results and inter-
esting observations.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction The performance of an alloy might be improved only if the bound-

Aluminium alloys that respond to heat treatment are the age-
hardenable or precipitation hardened alloys. At elevated temperature,
the second phase dissolves in the solid solution, which precipitates
upon quenching and ageing at a lower temperature. For an aluminium
alloy to be age hardened the second phase must be soluble at elevated
temperatures, but must show decreasing solubility with decreasing
temperature. This second condition of decreasing solubility with tem-
perature puts a limit on the number of useful precipitation-hardened
alloy systems [1]. With proper alloying and heat treatment, hardness
in precipitation hardened alloy can be increased to nearly 40 times as
compared to pure aluminium alloys. Therefore, it is one of the most im-
portant strengthening mechanisms in aluminium [2]. The common age
hardenable alloys are the Al–Cu (2XXX), Al–Mg–Si (6XXX) and Al–Zn–
Mg (7XXX) series of alloys, which have specific stable and metastable
precipitates and their fixed precipitation sequences for each system.
To improve the properties of these alloys, several attempts have been
made for years through minor additions [3–5] or through thermal
[6–8] or thermomechanical processing [9–14].
⁎ Corresponding author.
E-mail address: swatidey@yahoo.com (S. Dey).
aries of the series of alloys could be crossed and if the effects of the pre-
cipitates of the different series could be incorporated. But such effort has
never been reported by any previous worker. In this work, attempts
have been made to design alloys with improved mechanical properties
having such composition, which can have precipitates of all the three
age hardenable alloy series. Experimental trial-and-error method to
search for suitable chemical composition and processing parameters
that will lead to the desired material properties is tedious, time consum-
ing and costly, with no warranted results. On the other hand, unveiling
the mathematical interrelationships between the composition and pro-
cessing parameters with the materials properties of the alloy will make
it possible to computationally design an alloy possessing the optimal
combination of strength and ductility [15]. The complex correlation
being difficult to describe through any physical model, data-driven
models have been used extensively in the materials engineering domain
[16–18] to successfully find such complex correlations leading to
effective materials design. Applications of this computational approach
to the design of Al alloys and composites have also been reported
[19–25]. While computational intelligence techniques, particularly
using rough and fuzzy set theories have seen wide application in map-
ping the complex composition–processing–property relationships
[26–28], Artificial Neural Network (ANN) [29,30] seems to be the most
widely used paradigm in this domain to satisfactorily extract non-

http://dx.doi.org/10.1016/j.matdes.2015.12.076
0264-1275/© 2015 Elsevier Ltd. All rights reserved.
 S. Dey et al. / Materials and Design 92 (2016) 522–534 523

Table 1
List of input and output variables with their minimum, maximum, mean and standard deviation values at low, room and high testing temperature respectively.

Variables Symbols Low testing temperature Room testing temperature High testing temperature

Min Max Mean Std dev Min Max Mean Std dev Min Max Mean Std dev

Silicon (wt.%) Si 0.1 0.9 0.43 0.22 0.1 1.4 0.60 0.35 0.18 0.9 0.48 0.19
Iron (wt.%) Fe 0.12 1.1 0.51 0.26 0.12 1.1 0.51 0.24 0.3 1.1 0.58 0.22
Copper (wt.%) Cu 0.1 6.3 2.35 1.98 0.1 6.3 1.95 2.13 0.1 6.3 2.58 2.15
Manganese (wt.%) Mn 0 0.8 0.29 0.23 0 0.8 0.31 0.26 0 0.8 0.29 0.23
Magnesium (wt.%) Mg 0.02 2.7 1.38 0.83 0 2.8 1.07 0.77 0 2.7 1.17 0.84
Chromium (wt.%) Cr 0 0.3 0.12 0.09 0 0.4 0.10 0.10 0 0.3 0.12 0.09
Nickel (wt.%) Ni 0 2 0.12 0.43 0 2 0.13 0.47 0 2 0.14 0.47
Zinc (wt.%) Zn 0.1 6.8 1.84 2.64 0.05 6.8 1.11 2.11 0.1 6.8 1.33 2.38
Zirconium (wt.%) Zr 0 0.18 0.01 0.05 0 0.18 0.02 0.05 0 0.18 0.02 0.05
Titanium (wt.%) Ti 0 0.2 0.12 0.06 0 0.2 0.11 0.07 0 0.2 0.12 0.06
Testing temperature (°C) Ttest −269 −28 −105.5 74.2 100 371 229.8 92.7
Solutionizing temperature (°C) Tsoln 468 540 508.4 21.8 468 565 517.5 23.7 468 540 507.9 22.9
Ageing temperature (°C) Tage 24 205 149.6 51.2 24 240 123.8 70.4 24 205 140.7 60.8
Ageing time (hours) tage 1 72 31.2 26.7 1 72 36.2 30.0 1 72 31.2 26.6
Cold work cw 0 1 0.21 0.41 0 1 0.18 0.39 0 1 0.24 0.43
Yield strength (MPa) YS 97 745 402.8 145.2 90 540 300.5 118.1 10 470 124.5 118.9
Ultimate tensile strength (MPa) UTS 165 876 483.6 150.7 152 605 371.2 116.4 16 505 151.0 136.2
%Elongation %El 5 44 15.3 6.9 6 24 14.5 4.6 8 125 48.7 30.8

trivial relationships between the properties and processing routes of adequate ductility. The objectives of maximizing strength and ductility
different materials systems. are however mutually conflicting, which can be computationally
The objective of the present paper is to design novel age hardenable approached as a multi-objective optimization problem [31,32] and
Al alloys with optimum mechanical properties; i.e. good strength and solved through genetic algorithm (GA) [33], a well-established

Fig. 1. Scatter plot for low temperature showing target vs achieved values for (a) YS, (b) UTS and (c) %El as predicted by ANN models.
524 S. Dey et al. / Materials and Design 92 (2016) 522–534
Fig. 2. Sensitivity plots done on trained ANN models for YS, UTS and %El using connection weight method for low temperature.
evolutionary algorithm for optimization. GA is a population based optimi-
zation scheme that mimics the Darwin's law of natural selection, where
the optimum solution evolves through several generations of population.
In this way, it minimizes the possibility of being stuck in the local optima.
GA has been effectively used for designing materials with improved per-
formance in the field of Al alloys [34,35], composites [36,37], as well as
other metallic and non-metallic materials [38]. There are several exam-
ples of tandem application of ANN and GA for materials design, where
ANN is used as objective functions for the GA based optimization [39,
40]. These two tools are chosen here also due to their proven performance
in dealing such problems. The present study proposes a unique design of
Al alloys with a superior combination of strength and ductility for applica-
tion in different temperature conditions. This novel effort is intended to
design such alloy, which cross the border of the defined series of age
hardenable Al alloys. While Multi-Objective Genetic Algorithm (MOGA)
provides acceptable resolutions of the conflicting objectives, ANN formu-
lates the objective functions describing the composition–processing–
property correlation in three different test temperature regimes.
2. Problem formulation
As no accepted physical relation exists till date that can describe the
strength and ductility of Al alloys in terms of their composition and pro-
cessing variables, ANN has been used to frame such a relation. These
relations act as separate objective functions for strength (YS or UTS)
and ductility (%Elongation) that are to be optimized through MOGA.
The objective functions are thus formulated as:
hX n X o i
max Wj tanh wji xi þ b j þb
subject to xLB UB
i ≤xi ≤xi
Fig. 3. Sensitivity plots done on trained ANN models for YS, UTS and %El using connection weight method for room temperature.
where,
wji is the input to the hidden layer weight matrix, Wj represents the
weight vector for the jth hidden layer, bj and b denote the bias values
for the ith input layer to the hidden layer and the hidden layer to the out-
put layer respectively for the trained network. xiis the variable for which
the maximization search is attempted, however xLB i and xi
UB
are the
th
lower and upper bounds defining the search space for the i variable.
The Multi-Objective Genetic Algorithm code NSGA-II [41] has been
used in this work.
3. Database
259 numbers of data on the age hardenable aluminium alloys, i.e.
2XXX, 6XXX and 7XXX were collected from standard source [42].The
data consists of the chemical composition and the processing parame-
ters as the input variables and the mechanical properties as the output
variables. This data is used for the development of ANN models to find
the correlation between the input and output variables and these
models serve as objective functions for further optimization studies.
Separate ANN models are developed for the data in the low, room and
high temperatures respectively. Table 1 gives the list of input and output
variables, the symbols used for each, the minimum and maximum
values, the mean and the standard deviations of each of the variables
in the three different temperature ranges.
4. Computational procedure
4.1. Modelling using Artificial Neural Network (ANN)
This work uses ANN with a feed forward multi-layered perceptron
type architecture trained with scale conjugate gradient backpropagation
 S. Dey et al. / Materials and Design 92 (2016) 522–534 525

Fig. 4. Sensitivity plots done on trained ANN models for YS, UTS and %El using connection weight method for high temperature.

algorithm. The inputs and outputs are normalized within the range of −1
to 1 for ANN work using the following relation:
ðx−x min Þðb−aÞ
xN ¼ a þ
x max −x min
where a here is −1, b is +1, xN is the normalized value of a variable
x, xmax and xmin are the maximum and minimum values of x
respectively.
The weighted combination of the normalized inputs (xN i ) is operated
on by a hyperbolic tangent transfer function as shown in Eq. (2) for
every hidden unit (hj), which ensures that each input contributes to
every hidden unit. The suffix ‘j’ is used to denote the hidden node num-
ber and ‘i’ is used to denote the input number.
X  tion. But when more than one conflicting objective functions comes in
h j ¼ tanh wji xNi þ b j :
The output neuron then calculates a linear weighted sum of the
outputs of the hidden units, as given in Eq. (3):
X set, in terms of all the objectives, and strictly better in terms of at least
y¼ W j h j þ b:
In the above equations, wji and Wj are the weights and b is the bias.
Different outputs can be obtained by changing the weights wji in
Eqs. (2) and (3) [43,44].The optimum values of these weights are deter-
mined by “training” the network on a set of normalized input–output
data. The network is trained by adjusting the weights (wij) so as to min-
imize an error function, which is basically a regularized sum of square
errors. This ultimately gives to an optimal input–output relationship.
network model. Several methods are available and tried to find the rel-
ative importance of the input variables with respect to the output vari-
ables [45,39]. Similar trends were obtained in the results obtained using
the various methods. Therefore only one of them, the connection weight
ð1Þ
method [45] is reported in the results section. This approach uses input-
hidden and hidden-output connection weights to calculate the variable
importance.
4.3. Multi-objective optimization using genetic algorithm
Genetic algorithm is influenced by the principles of natural selection
and natural genetics. When there are multiple objectives, the genetic
search is performed following the theory of Pareto-optimality [33]. For
single objective situation, we search for unique global optimum solu-
ð2Þ the scenario then the aim becomes different. Here we try to find a set
of solutions that provides the best possible compromises between the
objectives, which is known as the Pareto set. The very definition the
Pareto-optimality necessitate that no other solution could exist in the
feasible range that is at least as good as some member of the Pareto
ð3Þ one [39].
Simple Genetic Algorithm (SGA) emulates three major biological
processes [38]:
i. Selection operation for identifying the candidates for the next gener-
ation
ii. Crossover operation for the probabilistic exchange of genetic infor-
mation between two randomly picked parents, facilitating a global
search for the optimum in the process

4.2. Sensitivity analysis

For designing a new alloy it is pertinent to have clear understanding

of the relative influence of the alloying elements in the final properties
of the alloys. But it is difficult to identify influences of the input variables
on the output of a system, particularly in case of ANN models, due the
complex hidden relationships generated. Sensitivity analysis provides
way to automatically identify all relevant parameters from a set of po-
tential parameters. The sensitivity analysis is done on a trained neural

Table 2
Genetic algorithm parameters.

Crossover probability 0.95

Mutation probability 0.05
Parental Selection Strategy Tournament selection
Fig. 5. Pareto front of %El vs YS showing the optimized combination of generation and pop-
Random seed value 0.19
ulation size for low temperature developed by multi-objective GA.
526 S. Dey et al. / Materials and Design 92 (2016) 522–534
iii. The mutation operation for inducing a small, probabilistic change in
the genetic makeup, resulting in a local search
The Pareto set offers a number of equivalent optimum solutions, out
of which a decision maker can easily pick and choose the most suitable
ones. To find the solution to the model system taken into account for Al
alloys, a non-dominated sorting Multi-Objective Genetic Algorithm i.e.
NSGA-II code [41], as mentioned in Section 2 is utilized, to obtain the
optimal values of strength and ductility.
5. Design of alloys
5.1. ANN modelling
For each of the three temperature zones, two separate ANN models
have been developed. While the first model maps all the input variables
to the strengths (YS and UTS), the second model maps all input vari-
ables to the elongation (%El) as output.
For the low testing temperature models, the optimum number of
hidden nodes of the model for YS and UTS, as found through experi-
mentation, is 9 and that for %El is 15. Ttest is constant for the room
temperature dataset, so T test is omitted from the input variables,
resulting in 14 input variables instead of 15. The number of hidden
nodes for both the ANN models in this temperature regime is 7. In
the high testing temperature zone, the number of hidden nodes of
the model for YS and UTS is 6 and that for %El is 5. Fig. 1(a–c) gives
a representative scatter plot for YS, UTS and %El respectively for
low testing temperature dataset.
Similarly, the scatter plots are obtained for the room temperature
and high temperature data subsets. As in case of Fig. 1, all those plots
show that ANN is able to map the inputs with the outputs quite
Fig. 6. Variation of (a) Fe, (b) Cu, (c) Mg and (d) Ni respectively in Pareto solutions with YS in ascending order in the low temperature regime.
successfully. Figs. 2, 3 and 4 presents the results of sensitivity analysis
of trained ANN models for YS, UTS and %Elongation using connection
weight method for the low temperature, room temperature and high
temperature respectively. The sensitivity plot for low temperature
shows that most of the alloying elements except Si and Ni have positive
effect on strength and negative effect on ductility (Fig. 2). The alloying
additions in Al alloys cause strengthening either through solid solution
hardening or through precipitation hardening. Both these phenomena
have negative effect on the ductility.
The sensitivity plots for room temperature, Fig. 3, also show that
most of the alloying elements have a positive effect on strength. Here
also Si has positive effect on ductility. The behaviour of Si is found to
have different effect compared to other alloying additions. This needs
to be investigated further.
It is to be noted here that the sensitivity graphs for high temper-
ature is scaled on the y-axis to a minimum–maximum limit in such a
way that the significance of the input variables can be understood.
As Ttest has a very large significance compared to the other inputs,
so the other variables appeared insignificant and also the relative
significance could not be understood if not scaled properly for this
temperature range. Here also the sensitivity plots have similar
trends (Fig. 4), except the fact that in this temperature regime test-
ing temperature has a significant role. The ageing precipitates could
not resist the softening, due to their coarsening. In this situation, in-
termetallic compounds like Al3Ti act as a potent hindrance for dislo-
cation movement [46], which is seen in the strengthening effect of Ti
in the figure. But such effect is not shown by Zr, which is known to
form similar compound with Al. In all cases it is evident that im-
provement of strength and ductility depends mostly on different
variables, expectedly. This also justifies our approach of formulating
the issue of design the alloys as a multi-objective optimization
problem.
 S. Dey et al. / Materials and Design 92 (2016) 522–534 527

Fig. 7. Contour plot showing the variation of Fe–Cu with (a) UTS, (b) %El and variation of Mg–Ni with (c) UTS and (d) %El in the low temperature regime.

5.2. Constraint formulation for the optimization • Similarly ageing temperature model is also used as a constraint and a
tolerance of ±25 °C is given so that the Tage chosen by GA should fall
It is known that solution treatment should be carried out at temper- between this range of the Tage calculated from the model.
atures within the single phase region so that most of the alloying ele- • The search space for low testing temperature data was kept between
ments dissolve in this temperature. However, this temperature should −101 °C and −99 °C and for high testing temperature was kept be-
never be above the solidus temperature, which can cause melting of tween 149 °C and 151 °C.
the alloy. Consequently, depending on the composition of the alloy,
the temperature should be kept within a narrow range. The ageing tem-
perature and time also depends on alloy systems where the proper pre-
5.3. Multi-objective optimization towards designing
cipitate in adequate quantity can be formed. Therefore, predictive
models for Tsoln and Tage are developed using ANN, which can be put
The multi-objective optimization for maximization of strength and
as constraints during optimization. The compositional variables are
ductility using the ANN models as the objective functions, and the
the inputs for the Tsoln model, which uses the entire dataset for model-
ling. The best ANN model had 15 hidden nodes, with R value 0.857. Like-
wise, the compositional variables and Tsoln act as inputs for the Tage
model. Here only the artificial ageing data is used. The best ANN
model had 19 hidden nodes, with R value 0.783. The detail of the pro-
cess is described elsewhere [47].
Finally the following constraints are applied while carrying out the
optimization studies:

• The sum of the compositional parameters, i.e. weight percent of

alloying elements, is kept at 10 wt.% as a constraint keeping in mind
that higher alloy content can decrease the toughness of the alloy.
• Solutionizing temperature model is used as a constraint. A tolerance of
±25 °C is given so that the Tsoln chosen by GA should fall within this
range of the Tsoln calculated from the model. This is done to make
the solutions within a range of temperature suitable for the alloy sys- Fig. 8. Pareto front of %El vs YS showing the optimized combination of generation and pop-
tem. ulation size for room temperature developed by Multi-Objective GA.
528 S. Dey et al. / Materials and Design 92 (2016) 522–534

above described constraint is executed using GA for several times vary-
ing GA parameters, so that the solutions are not stuck in the sub-optimal
region. The genetic algorithm parameters that are kept constant in all
the optimization studies reported in this work are presented in Table 2.
Fig. 5 shows the optimized Pareto front for low testing temperature
data for a population size of 500 and generation 500. It is to be noted
here that as three objectives are optimized simultaneously, the Pareto
solutions form a surface, and 2 dimensional projection of that surface
is shown in Fig. 5. All the solutions shown in the figure are non-
dominated.
The non-dominated solutions generated from the above optimiza-
tion show that the variation in the YS is mainly due to variation in Fe,
Cu, Mg and Ni. The other variables almost remained constant. For
most of the solutions, Si was fixed at 0.86 (towards the maximum
value in the search space). This is expected as it does not hamper the
ductility of the alloy. Mn was fixed at 0.79 (very near to the maximum
value in the search space), Cr was nearly 0, Zn had a small variation be-
tween 0.6 and 1.7, Zr also had a small variation between 0.11 and 0.14
and Ti was fixed at nearly 0.19. The variation of Fe, Cu, Mg and Ni
with YS (arranged in ascending order and plotted as the population
size) is shown in Fig. 6(a–d). It is known that elements like Zr and Ti re-
fine Al grains through formation of precipitates like Al3Zr and Al3Ti [46].
Ductility is least compromised when strengthening is done through
grain refinement, and may be the most suitable proposition for
strengthening an alloy used at low temperature, where the ductility is
generally low. It can be generally stated that those alloying elements
are preferred which improves one objective without much affecting
the other.

Fig. 9. Variation of (a) Cu, (b) Mn, (c) Mg (d) Zr and (e) Ti respectively in Pareto solutions with YS in ascending order in the room temperature regime.
 S. Dey et al. / Materials and Design 92 (2016) 522–534 529

Fig. 10. Contour plot showing the variation of Cu–Mn with (a) UTS, (b) %El and variation of Mg–Zr with (c) UTS and (d) %El in the room temperature regime.

Fig. 7(a–d) show low temperature contour plots for Fe–Cu against The solutions generated from the above optimization show that the
UTS and %El and for Mg–Ni against UTS and %El respectively. These fig- variation in the YS is due to variation in Si, Cu, Mg and Zn. The other var-
ures are developed from the simulation study of the ANN models, to iables almost remained constant. For most of the solutions, Fe is fixed at
study the effect of variations of the elements near the optimum solu- 0.34, Mn at 0.002, Cr at 0.081, Ti at 0.029 (all towards the minimum
tions. This gives an idea about the robustness of proposed solutions. value in the search space), Ni is nearly at 1.99 (towards the maximum
The figures clearly show that Cu and Mg do not have any positive effect value in search space), Zr is fixed at nearly 0.153. The variation of Si,
either on strength or on ductility at low temperature. Fe is detrimental Cu, Mg and Zn with YS (arranged in ascending order and plotted as
to ductility. But Ni seems to improve the strength as well as ductility. the population size) is shown in the Figs. 12(a–d). It is evident from
The optimized Pareto front for room temperature data having the the results that the preferred alloy for high temperature applications
number of generation and population size both 500 is shown in Fig. 8. is an alloy having good amount of Cu, Mg, Zn and Ni, with significant
The non-dominated solutions generated from the above optimiza-
tion show that the variation in the YS is mainly due to variation in Cu,
Mn, Mg, Zr and Ti. The other variables almost remained constant. For
most of the solutions, Si was fixed at 1, Fe was fixed at 1.09 (towards
the maximum value in the search space), Cr and Ni was nearly 0, Zn
was fixed at nearly 4 for most of the solutions. The variation of Cu,
Mn, Mg, Zr and Ti with YS (arranged in ascending order and plotted as
the population size) is shown in the Fig. 9(a–e). Most of the solutions
are found to have Cu around 2.0 wt.%. Mn varied near the upper
range. Low values for Mg, Zr and Ti are preferred. The trends for Mn
and Mg are found to be similar to that of the preferred composition
for low temperature applications. But Zr and Ti are not preferred for
the present alloy meant for room temperature applications.
Fig. 10(a–d) below show room temperature contour plots for Cu–
Mn against UTS and %El, and for Mg–Zr against UTS and %El respectively.
Cu and Mn seem to have positive effect on both strength and ductility.
Zr has incremental effect only on strength and Mg only on ductility.
Cu seems to be the most important element in this temperature regime.
Finally the optimized Pareto front for high temperature data is Fig. 11. Pareto front of %El vs YS showing the optimized combination of generation and
shown in Fig. 11 for 100 generations and population size 500. population size for high temperature developed by Multi-Objective GA.
530 S. Dey et al. / Materials and Design 92 (2016) 522–534
Fig. 12. Variation of (a) Si, (b) Cu, (c) Mg and (d) Zn respectively in Pareto solutions with YS plotted in ascending order in the high temperature regime.
minor addition of Zr. As maintaining the strength at high temperature is
difficult, high amount of alloy additions are suggested.
Fig. 13(a–d) show the high temperature contour plots of Si–Cu
against UTS and %El and for Mg–Zn against UTS and %El respectively.
Here it is revealed that Cu has incremental effect both on strength and
ductility at high temperature applications of Al alloys. Si does not have
any significant positive effect on strength, but detrimental for ductility.
A combination of high Mg along with medium amount of Zn seems to
achieve good combination of strength and ductility.
Table 3 gives a tabular representation of the contribution of the input
variables that are significantly responsible for the determination of the
final optimum properties.
The optimum solutions for low service temperature zones show that
Si, Mg and Ni are the most significant additions. Among these elements,
Si and Mg provide direct precipitation and Ni provides solid solution
strengthening. Thus the mode of strengthening is suggested to be
mixed for achieving improved strength and ductility. In case of solutions
for ambient temperature applications, Fe, Cu, Mg, Zn and Si are found to
be important. But one observation could be made here that presence of
almost all the elements responsible of ageing are used in this tempera-
ture regime. Though Fe and Ni are also used, but definitely the ageing is
the preferred mode of strengthening here.
The preferred elements for alloying Al suitable for high temperature
applications are Fe, Mg and Ni along with Si, Cu and Zn, all three potent
precipitators. Al–Cu precipitates soften significantly at high tempera-
tures, and thus the cutting of the precipitates during dislocation move-
ment becomes easy [48]. Fe rich phases like Al8Fe2Si and Al13Fe4 have
the capacity to stabilize sub-grain structures and thus improving the
strength at high temperature [49].
5.4. Decision making
For designing an alloy suitable for applications in the room temper-
ature as well as high temperature condition, normal requirement for in-
dustrial conditions, the above findings are followed. All precipitate
forming elements are taken into consideration. In the designed alloy
targeted additions of both Cu and Mg are made to be 2.0 wt.%. Cu is
found to have adequately high strengthening effect, but it deteriorate
the ductility significantly. To achieve good strength-ductility balance,
it may be preferable to add a good amount of Cu for the alloy to be
used at high temperatures. But Zn is targeted at higher amount (4.0
wt.%). It may be noted that higher Zn is preferred in most of the solu-
tions at all three temperature regimes, presumably due to its lesser det-
rimental effect on ductility. In the optimum composition Si, Fe and Ni
are also preferred, as some of the Pareto solutions contained these ele-
ments in moderate amount. The final target composition is given in
Table 4.
6. Experimental procedure
Melting is carried out in a resistance heating pot furnace under the
suitable flux cover (coveral etc.). In the process of preparation of the al-
loys the commercially pure aluminium (99.5% purity) is taken as the
starting material. The alloying elements that are in powder form are
compacted. The Zn granules put in a cover of aluminium foil is added
and then finally magnesium ribbon is added into solution. The final tem-
perature of the melt is always maintained at 800 ± 15 °C. Casting is
done in cast iron metal moulds preheated to 200 °C. Mould size is
 S. Dey et al. / Materials and Design 92 (2016) 522–534 531

Fig. 13. Contour plot showing the variation of Si–Cu with (a) UTS, (b) %El and variation of Mg–Zn with (c) UTS and (d) %El in the high temperature regime.

12.5 mm × 12.5 mm × 102 mm. The cast samples are first ground prop-
erly to remove the oxide layer from the surface. Portions of different al-
loys are kept into furnace at the temperature of 500 °C for 24 h for
homogenization of the alloys. The chemical compositions of all the al-
loys are analysed using atomic absorption spectroscopic method. The
chemical composition of the alloy is given in Table 4.
Hot rolling of homogenized samples is carried out in a laboratory
scale two-high rolling mill of 10 HP capacity at temperature around
420 °C. The cast as well as hot rolled samples are solution treated at
520 °C for 60 min and quenched in ice-brine. This temperature is deter-
mined using the ANN model developed for predicting solutionizing
temperature and reported earlier in the present work. Both natural

Table 3
Testing temperature wise contribution of input variables towards final optimum property determination.

Testing Important Si Fe Cu Mg Ni Zn
temp variables

Low Variable 0.59–0.87 0.12–1.1 0.1–1.5 0.18–2.22 0.18–2 0.61–1.73

temp range after
optimization
Remarks Most of the values are Maximum values Maximum values are The values are distributed The values are Maximum values
towards the upper limit, i.e. are towards the towards the lower limit, from lower to upper limit distributed from are towards the
0.87 lower limit, i.e. i.e. 0.1 obtained after optimization lower to upper lower limit, i.e. 0.61
0.12 limit
Room Variable 0.93–1.08 1.1–1.1 0.59–2.7 0.27–2.1 0–0.03 3.59–4
temp range after
optimization
Remarks Most of the values are Most of the values A few values in lower Most of the values towards Most of the values Most of the values
towards the upper limit, i.e. are towards the limit, i.e. 0.59, maximum the lower limit, i.e. 0.27 towards the lower are towards the
1.08 higher range in upper limit, i.e. 2.7 limit, i.e. 0 upper limit, i.e. 4
High Variable 0.18–0.36 0.32–0.89 1.41–4.08 0.05–2.34 0.73–2 0.45–2.79
temp range after
optimization
Remarks The values are distributed Most of the values Most of the values are in The values are concentrated Most of the values The values are
from lower to upper limit towards the lower the range 0.35–4 at two levels, one at 1.5, are towards the concentrated at 3
obtained after optimization limit, i.e. 0.32 another at 2.34 upper limit, i.e. 2 levels, at 0.5, 0.8
and 2
532 S. Dey et al. / Materials and Design 92 (2016) 522–534

Table 4
Target and achieved composition of the designed alloy.

Elements (wt.%) Zn Cu Mg Si Fe Ni

Target alloy 4.0 2.0 2.0 0.7 0.7 0.7

Developed alloy 5.14 2.10 1.96 1.06 1.01 0.89

ageing and artificial ageing is done on the as-cast and hot-rolled sample.
Artificial ageing is performed on cast and hot rolled samples both as per
the schedule given in Table 5.
Hardness of different alloys processed with different schedules and
aged at different temperatures is measured in LECO LM 100 Micro-
Vickers hardness testing machine with 300 g load with dwell time
15 s for assessing the age-hardening effect of the alloys. An average of
three consistent readings is accepted as the representative hardness
value of an alloy. Room temperature tensile testing is carried out using
Fig. 15. Isochronal ageing curve of the alloy in as cast-solutionized and hot rolled-
Instron4204 with crosshead velocity of 0.2 mm/s. The test specimen is solutionized condition.
prepared as per ASTM Standard (ASTM: Vol. 03.01: E8M — 96) from
solutionized as well as aged specimens of the cast alloys. The yield
strength (YS), ultimate tensile strength (UTS) and percent total elonga- also the precipitation strengthening has occurred in two stages punctu-
tion (%El) are determined from the machine output. Metallographic ated by a softening, maybe due to formation of stable coarse phases.
studies of the selected samples are carried out by a Scanning Electron Isothermal ageing studies for the cast and wrought alloys are con-
Microscope (Model: Hitachi S3400N). ducted for 150, 200, 250 and 300 °C (Figs. 16 and 17 respectively).The
rationale behind selecting these four temperatures is in the isochronal
ageing, where the peak hardness is achieved at 200 °C. So in isothermal
7. Experimental results study, temperature around the peak hardness temperature is consid-
ered in order to get an in depth understanding of the ageing phenome-
Fig. 14 shows the natural ageing hardness curve obtained for the na. Ageing at 300 °C is considered to find the effect of over-ageing.
alloy in cast-solutionized condition and hot rolled-solutionized condi- As expected in both cast and wrought alloys there are significant age-
tion. As during natural ageing only GP zones are formed, primarily ing effects when aged at 150, 200 and 250 °C, and in case of ageing at
from the nature of both the curves it seems that different GP zones are 300 °C the softening effect due to overageing is predominant. But it should
formed during different time. Presence of all the GP zone forming ele-
ments together might have resulted in such feature. This needs to be
studied in depth.
The artificial ageing started with isochronal ageing (60 min), and the
hardness curve is shown in Fig. 15 for various ageing temperatures as
mentioned in Table 5. It is interesting to note that the age hardening
trend is similar in artificial ageing as in case of natural ageing. Here

Table 5
Ageing schedule.

Ageing temperature (°C) Ageing time (min)

100 30 60 90 120 240

150 30 60 90 120 240
200 30 60 90 120 240
250 30 60 90 120 240
300 30 60 90 120 240
350 60
400 60
Fig. 16. Artificial ageing curves for as cast samples for various ageing temperatures.

Fig. 14. Natural ageing curve of the alloy in as cast-solutionized and hot rolled-
solutionized condition. Fig. 17. Artificial ageing curves for hot rolled samples for various ageing temperatures.
 S. Dey et al. / Materials and Design 92 (2016) 522–534
Table 6
Tensile properties of the cast alloys in solutionized and peak aged (200 °C for 60 min)
conditions.
Condition YS (MPa) UTS (MPa) %Elongation
Solutionized 295 309 11.5
Peak aged 333 354 11.9
be noted here that in both cases the pattern of ageing is different at differ-
ent ageing temperatures. At 150 °C the ageing curve rises continuously
without any softening at higher ageing time. After ageing at 200 °C the
hardness values become almost constant with time. In case of ageing at
250 °C the increase in hardness is sharp followed by significant softening.
This phenomenon clearly indicates that the ageing precipitates are differ-
ent at different ageing temperatures. Thus the presence of different age
hardening elements has presumably developed a complex precipitation
sequence. The complex interaction between vacancies and different sol-
ute elements resulting in competitive and/or synergistic precipitation
can be an important area for further study of the present alloy.
Table 6 shows the tensile properties of the cast alloys in solutionized
as well as peak aged conditions. The results are encouraging, but it
seems that the strength ductility combination could be further im-
proved through designing proper ageing sequence. The complexity in
the precipitation, as discussed earlier, needs to be understood in a better
way to design the heat treatment schedule.
The microscopic studies on the cast and wrought alloys in solutionized
and peak aged conditions are performed using SEM. The microstructure of
Fig. 18. SEM micrograph for an aluminium alloy that is (a) cast-solutionized, (b) cast-solutionized-aged at 20 °C—60 min, (c) hot-rolled and (d) hot-rolled-aged at 200 °C—60 min.
533
the alloy is quite fine in the both cast and wrought condition (Fig. 18). But
it is not translated to its strength due to two factors. Primarily there is
grain coarsening during ageing, and also significant grain boundary thick-
ening. These two factors could not allow the matrix to strengthen during
ageing. This may be avoided with proper heat treatment.
8. Conclusion
The study undertakes designing alloys using computational intelli-
gence tools, and experimentally developing the alloy. The following
conclusions may be reached regarding the models built and related Pa-
reto optimized results, along with the experiments performed.
i. The ANN models are found to be good enough to act as objective
functions for multi-objective optimization using GA.
ii. The Pareto solutions suggest a trend of preferred alloying elements,
which provides cues for designing the alloy.
iii. Almost all the alloying elements contributing for the age hardening
are preferred for addition, along with Fe and Ni.
iv. The preferred composition for applications in different temperature
regimes varied. The reason for such variation needs to be found out
through systematic experimentation.
v. The experimental finding of the designed alloy produces encourag-
ing results, which allows scope for further experimentation.
vi. The sequence of GP zone formation and precipitation, and their
competitive effects needs to be investigated through proper experi-
mentation. Understanding the precipitation sequence will lead to
534 S. Dey et al. / Materials and Design 92 (2016) 522–534
formulation of the exact heat treatment schedule to exploit the ef-
fect precipitation hardening to its fullest extent.
Acknowledgements
Authors Swati Dey and Nashrin Sultana are highly grateful to the
Council of Scientific and Industrial Research (CSIR), Govt. of India, for
providing them the financial support through Senior Research Fellow-
ships (CSIR Award Nos. 08/3(0056)/2008-EMR-I and 08/003(0067)/
2010-EMR-I, respectively) to carry out this research work.
References
[1] I.J. Polmear, Light Alloys — From Traditional Alloys to Nanocrystals, fourth ed.
Butterworth-Heinemann, Burlington, MA, 2006.
[2] V. Raghavan, Physical Metallurgy: Principles and Practice, Prentice-Hall of India Pri-
vate Limited, New Delhi, 2004.
[3] G. Wang, Q. Sun, L. Feng, H. Luo, C. Jing, Influence of Cu content on ageing behavior
of AlSiMgCu cast alloys, Mater. Des. 28 (2007) 1001–1005.
[4] N.Q. Chinh, J. Lendvai, D.H. Ping, K. Hono, The effect of Cu on mechanical and precip-
itation properties of Al–Zn–Mg alloys, J. Alloys Compd. 378 (2004) 52–60.
[5] J. Hernandez-Sandoval, G.H. Garza-Elizondo, A.M. Samuel, S. Valtiierra, F.H. Samuel,
The ambient and high temperature deformation behavior of Al–Si–Cu–Mg alloy
with minor Ti, Zr, Ni additions, Mater. Des. 58 (2014) 89–101.
[6] G. Mrowka-Nowotnik, J. Sieniawski, Influence of heat treatment on the microstruc-
ture and mechanical properties of 6005 and 6082 aluminium alloys, J. Mater. Pro-
cess. Technol. 162–163 (2005) 367–372.
[7] S. Pogatscher, H. Antrekowitsch, H. Leitner, A.S. Sologubenkoand, P.J. Uggowitzer, In-
fluence of the thermal route on the peak-aged microstructures in an Al–Mg–Si alu-
minum alloy, Scr. Mater. 68 (2013) 158–161.
[8] M.N. Ilman, Kusmono, P.T. Iswanto, Fatigue crack growth rate behaviour of friction-
stir aluminium alloy AA2024-T3 welds under transient thermal tensioning, Mater.
Des. 50 (2013) 235–243.
[9] L.-h. An, C. Yang, W. Liu, S.-j. Yuan, S.-q. Zhu, F.-c. Meng, Effect of pre-deformation on
microstructure and mechanical properties of 2219 aluminum alloy sheet by
thermomechanical treatment, Trans. Nonferrous Metals Soc. China 22 (2012)
s370–s375.
[10] L.-w. Quan, G. Zhao, S. Gao, B.C. Muddle, Effect of prestretching on microstructure of
aged 2524 aluminium alloy, Trans. Nonferrous Metals Soc. China 21 (2011)
1957–1962.
[11] M.H. Goodarzy, H. Arabi, M.A. Boutorabi, S.H. Seyedein, S.H. Hasani Najafabadi, The
effects of room temperature ECAP and subsequent aging on mechanical properties
of 2024 Al alloy, J. Alloys Compd. 585 (2014) 753–759.
[12] B. Adamczyk-Cieślak, J. Mizera, K.J. Kurzydłowski, Microstructures in the 6060 alu-
minium alloy after various severe plastic deformation treatments, Mater. Charact.
62 (2011) 327–332.
[13] F. Rajabi, A. Zarei-Hanzaki, M. Eskandari, S. Khoddam, The effects of rolling parame-
ters on the mechanical behavior of 6061 aluminum alloy, Mater. Sci. Eng. A 578
(2013) 90–95.
[14] S.-y. Chen, K.-h. Chen, L. Jia, G.-s. Peng, Effect of hot deformation conditions on grain
structure and properties of 7085 aluminum alloy, Trans. Nonferrous Metals Soc.
China 23 (2013) 329–334.
[15] H.K.D.H. Bhadeshia, R.C. Dimitriu, S. Forsik, J.H. Pak, J.H. Ryu, Performance of neural
networks in materials science, Mater. Sci. Technol. 25 (4) (2009) 504–510.
[16] S. Datta, P.P. Chattopadhyay, Soft computing techniques in advancement of structur-
al metals, Int. Mater. Rev. 58 (8) (2013) 475–504.
[17] A. Sinha, S. Sikdar(Dey), P.P. Chattopadhyay, S. Datta, Optimization of mechanical
property and shape recovery behavior of Ti–(~49 at.%)Ni alloy using artificial neural
network and genetic algorithm, Mater. Des. 46 (2013) 227–234.
[18] T. Bhattacharyya, S.B. Singh, S. Sikdar(Dey), S. Bhattacharyya, W. Bleck, D.
Bhattacharjee, Microstructural prediction through artificial neural network (ANN)
for development of transformation induced plasticity (TRIP) aided steel, Mater.
Sci. Eng. A 565 (2013) 148–157.
[19] A. Patra, S. Ganguly, M.S. Kaiser, P.P. Chattopadhyay, S. Datta, Effect of quaternary
zirconium addition on mechanical properties of Al–6Mg–Sc (0.2–0.6%) alloy studied
by ANN technique, Int. J. Mechatron. Manuf. Syst. 3 (2010) 144–154.
View publication stats
[20] N. Haghdadi, A. Zarei-Hanzaki, A.R. Khalesian, H.R. Abedi, Artificial neural network
modeling to predict the hot deformation behavior of an A356 aluminum alloy,
Mater. Des. 49 (2013) 386–391.
[21] M.H. Shojaeefard, R.A. Behnagh, M. Akbari, M.K.B. Givi, F. Farhani, Modelling and Pa-
reto optimization of mechanical properties of friction stir welded AA7075/AA5083
butt joints using neural network and particle swarm algorithm, Mater. Des. 44
(2013) 190–198.
[22] R.-G. Song, Q.-Z. Zhang, M.-K. Tseng, B.-J. Zhang, The application of artificial neural
networks to the investigation of aging dynamics in 7175 aluminium alloys, Mater.
Sci. Eng. C 3 (1995) 39–41.
[23] A.A. Tofigh, M.O. Shabani, Applying various training algorithms in data analysis of
nano composites, Acta Metall. Slovaca 19 (2013) 94–104.
[24] A. Mazahery, M.O. Shabani, M.R. Rahimipour, M. Razavi, The numerical modeling of
abrasion resistance in casting aluminum–silicon alloy matrix composites, J. Compos.
Mater. 46 (2012) 2647–2658.
[25] M. Ali, M.O. Shabani, A. Mehdi, A.A. Tofigh, Concurrent fitness evaluations in
searching for the optimal process conditions of Al matrix nanocomposites by linear-
ly decreasing weight, J. Compos. Mater. 47 (2013) 1765–1772.
[26] S. Dey, S. Datta, P.P. Chattopadhyay, J. Sil, Modeling the properties of TRIP steel using
AFIS: a distributed approach, Comput. Mater. Sci. 43 (2008) 501–511.
[27] S. Dey, P. Dey, S. Datta, J. Sil, Rough set approach to predict the strength and ductility
of TRIP steel, Mater. Manuf. Process. 24 (2009) 150–154.
[28] P. Dey, S. Dey, S. Datta, J. Sil, Dynamic discreduction using rough sets, Appl. Soft
Comput. 11 (2011) 3887–3897.
[29] J.A. Anderson, An Introduction to Neural Networks, MIT Press, Cambridge MA, 1995.
[30] S. Kumar, Neural Networks—A Classroom Approach, Tata McGraw-Hill Publishing
Company Limited, New Delhi, 2004.
[31] K. Miettinen, Nonlinear Multiobjective Optimization, Kluwer Academic Publishers,
Boston, 1999.
[32] K. Deb, Multi-Objective Optimization Using Evolutionary Algorithms, John Wiley &
Sons Ltd., Chichester, 2001.
[33] D.E. Goldberg, Genetic Algorithmsin Search, Optimization and Machine Learning,
Pearson-Education, New Delhi, 2002.
[34] R.G. Song, Q.Z. Zhang, Heat treatment optimization for 7175 aluminum alloy by ge-
netic algorithm, Mater. Sci. Eng. C 17 (2001) 133–137.
[35] S.S. Sahay, R. Mehta, K.K. Krishnan, Genetic-algorithm-based optimization of an in-
dustrial age-hardening operation for packed bundles of aluminum rods, Mater.
Manuf. Process. 22 (2007) 615–622.
[36] A. Mazahery, M.O. Shabani, Process conditions optimization in Al–Cu alloy matrix
composites, Powder Technol. 225 (2012) 101–106.
[37] A. Mazahery, M.O. Shabani, Development of the principle of simulated natural evo-
lution in searching for a more superior solution: proper selection of processing pa-
rameters in AMCs, Powder Technol. 245 (2013) 146–155.
[38] N. Chakraborti, Differential evolution: the real-parameter genetic algorithm applied
to materials and metallurgy, Int. Mater. Rev. 49 (2004) 259–260.
[39] I. Mohanty, D. Bhattacharjee, S. Datta, Designing cold rolled IF steel sheets with op-
timized tensile properties using ANN and GA, Comput. Mater. Sci. 50 (2011)
2331–2337.
[40] P. Das, S. Mukherjee, S. Ganguly, B.K. Bhattacharyay, S. Datta, Genetic algorithm
based optimization for multi-physical properties of HSLA steel through hybridiza-
tion of neural network and desirability function, Comput. Mater. Sci. 45 (2009)
104–110.
[41] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective genetic
algorithm: NSGA-II, IEEE Trans. Evol. Comput. 6 (2002) 182–197.
[42] ASM Handbook, Properties and Selection: Nonferrous Alloys and Special-Purpose
Materials, vol. 2ASM International Handbook Committee, 1990.
[43] H.K.D.H. Bhadeshia, Neural networks in materials science, ISIJ Int. 39 (1999)
966–979.
[44] H.K.D.H. Bhadeshia, Neural networks and information in materials science, Stat.
Anal. Data Min. ASA Data Sci. J. 1 (2009) 296–305.
[45] J.D. Olden, M.K. Joy, R.G. Death, An accurate comparison of methods for quantifying
variable importance in artificial neural networks using simulated data, Ecol. Model.
178 (2004) 389–397.
[46] I.J. Polmear, Role of trace elements in aged aluminium alloys, Mater. Sci. Forum 13
(14) (1987) 195–214.
[47] S. Dey, S. Ganguly, S. Datta, In silico design of high strength aluminium alloy using
multi-objective GA, Lect. Notes Comput. Sci 8947 (2015) 316–327.
[48] A. Deschamps, Y. Brechet, Influence of predeformation and ageing of an Al–Zn–Mg
alloy—II. Modeling of precipitation kinetics and yield stress, Acta Mater. 47 (1998)
293–305.
[49] M.J. Starink, N. Gao, J.L. Yan, The origins of room temperature hardening of Al–Cu–
Mg alloys, Mater. Sci. Eng. A 387–389 (2004) 222–226.