International Journal of Heat and Mass Transfer 101 (2016) 295–302

Contents lists available at ScienceDirect

International Journal of Heat and Mass Transfer

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

Multi-objective optimization of a honeycomb heat sink using Response

Surface Method
Abdussamet Subasi a,⇑, Bayram Sahin b, Irfan Kaymaz b
a
Department of Mechanical Engineering, Istanbul Technical University (ITU), Istanbul 34437, Turkey
b
Department of Mechanical Engineering, University of Atatürk, Erzurum 25240, Turkey

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

Article history: The aim of this study is to find optimum values of design parameters of a heat sink having hexagonal
Received 10 August 2015 aluminum honeycomb fins by using the Response Surface Method (RSM) and the Pareto based
Received in revised form 18 April 2016 multi-objective optimization. In this context, fin height ðHÞ, fin thickness ðtÞ, longitudinal pitch ðSy Þ, angle
Accepted 4 May 2016
of attack ðhÞ and Reynolds number ðReÞ are selected as design variables while Nusselt number ðNuÞ and
Available online 30 May 2016
friction factor ðf Þ are chosen as objective functions. Firstly, the RSM with the face centered central
composite design (FCCCD) has been employed to construct mathematical models required in multi-objective
Keywords:
optimization problem definition. In the next step, accuracy and validity of these mathematical models
Heat sink
Metallic honeycomb
are proved both statistically and experimentally. Finally, a Pareto based multi-objective optimization
Response Surface Method (RSM) study has been conducted to determine optimum values of the design parameters that maximize Nu
Multi-objective optimization and minimize f . It is concluded that Pareto solution set obtained provides important insights into the
Pareto front design parameters and allows designers freedom to make a selection among the optimal solutions.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction the literature, numerous studies are dedicated to optimization of

The failure of an electronic device depends on a few parameters
such as temperature, vibration, humidity and dust [1]. The operat-
ing temperature with a value of 55% is the most affecting factor
among these parameters [1]. The relationship between the operat-
ing temperature and failure rate factor defined as the ratio of
failure rate at any temperature to failure rate at 75  C is almost
exponential [2]. That means reliability of an electronic device
decreases with increasing operating temperature [2]. Therefore,
heat arising from electronic devices must be dissipated to improve
reliability of electronic devices, and to have a long life and high
performance electronic device. Heat sinks are generally used to
overcome these overheating problems that lead to corruption of
electronic devices. Furthermore, miniaturized sizes of heat sinks
in parallel to advancement of packaging technologies [3], and
increase of power density at integrated circuits have made thermal
system engineering applications as a key factor for development of
microelectronic technologies. Aforementioned facts necessitate
realization of heat transfer mechanism at optimum conditions as
it should be required to find optimum conditions to maximize effi-
ciency of any engineering system considered [4–6]. Therefore, in
⇑ Corresponding author. Tel.: +90 212 293 13 00/2453; fax: +90 212 245 07 95.
E-mail addresses: subasiab@itu.edu.tr, sametsubasi@gmail.com (A. Subasi),
bsahin@atauni.edu.tr (B. Sahin), ikaymaz@atauni.edu.tr (I. Kaymaz).
heat sinks having various fin geometries such as rectangular, dia-
mond, square, cylindrical, annular, tapered or pin fins [7,8]. In
addition to above mentioned fin types, with their superior proper-
ties, metallic honeycomb structures stand out as a prospective can-
didate for heat sink applications. The various applications of
honeycomb structures are reported in [9]. According to the paper,
in many engineering applications, honeycomb structures have an
extensive usage area such as structural load support, impact
energy absorption and thermal management. Due to their struc-
tural morphologies and thermal characteristics, they have large
surface-area-to-volume ratio, low pressure drop and high conduc-
tivity walls. These properties make them suitable for the use of
compact heat exchangers and heat sink applications [9–13]. How-
ever, in the literature, studies related to heat sink applications of
honeycomb structures are rather limited.
The literature review related to the optimization of various
heat sinks, the Response Surface Method (RSM), the Pareto based
multi-objective optimization and heat transfer applications of
honeycomb structures are reported below. Lu [12] analyzed the
heat transfer performance of micro cell aluminum honeycombs
in compact heat exchangers by using analytical models, and
found optimal cell morphology. Dempsey et al. [13] examined
heat transfer and pressure drop characteristics of the Linear
Cellular Alloys (LCA) heat sinks having square cells under laminar

http://dx.doi.org/10.1016/j.ijheatmasstransfer.2016.05.012
0017-9310/Ó 2016 Elsevier Ltd. All rights reserved.
296 A. Subasi et al. / International Journal of Heat and Mass Transfer 101 (2016) 295–302
Nomenclature
A heat transfer area (m2)
Dh hydraulic diameter of channel (m)
F objective functions
f friction factor ðDP=ððLt =Dh Þðq u2 =2ÞÞÞ
h mean heat transfer coefficient (W/m2 °C)
H fin height (mm)
k thermal conductivity of air (W/m °C)
L length of base plate (m)
Lt length of test chamber (m)
Nu Nusselt number ðh Dh =kÞ
Q_ heat transfer rate (W)
Re Reynolds number [u Dh/m]
Sy longitudinal pitch (mm)
t fin thickness (mm)
T steady state temperature ð CÞ
u mean inlet velocity of air (m/s)
W width of base plate (m)
xi independent design parameters
y desired response or yield
Greek letters
DP pressure drop (Pa)
e fitting error
h angle of attack ð Þ
t kinematic viscosity of air (m2/s)
q density of air (kg/m3)
flow conditions, both experimentally and numerically. In the
experimental part of the study, they measured pressure drop
along the honeycomb channels and total heat transfer rate of
the LCA heat sinks. Then, they compared these experimental
measurements with numerical results obtained from both a finite
difference code and commercially available finite volume method
based software. Wen et al. [10] conducted an analytical study to
find the optimal design of 2D cellular metallic sandwiches for
various cell topologies, under laminar forced convection condi-
tions at constant pumping power. They checked the accuracy
and validity of their analytical model by comparing the predic-
tions with Computational Fluid Dynamics (CFD) results. Liu
et al. [11] proposed an analytical model that accomplishes the
assumptions in the corrugated wall model, to examine thermal
performance of sandwich metallic honeycomb structures sub-
jected to forced convection conditions. They investigated accu-
racy of the new method proposed by comparing the results
with those obtained from the corrugated wall model, the effective
medium model and numerical simulation. They reported that
their method is accurate and gives close results to numerical
results. In their numerical study, Park et al. [3] employed the
RSM and CFD to find the optimum levels of various design
parameters of a plate-fin heat sink. Chiang [14] carried out an
experimental study to investigate the effects of selected design
parameters on thermal performance of a parallel-plain fin heat
sink and to find optimal values of these parameters subject to
mass and space constraints using RSM and the sequential approx-
imation optimization method. In their experimental study, Chiang
et al. [7] used the RSM to predict and optimize thermal resistance
and pressure drop of a pin–fin heat sink. They found the opti-
mum levels of the selected design variables of a pin fin heat sink
under the defined space and mass limitations. Chen et al. [15]
conducted a multi-objective optimization study to determine
structural design of a serpentine-channel heat sink. They adopted
channel width, fin width, channel height and inlet velocity as
Superscripts
L lower limit
U upper limit
Subscripts
b bulk
cond conduction
conv convection
in inlet
loss losses
n number of independent design parameters
out outlet
s surface
rad radiation
Abbreviations
ANOVA Analysis of Variance
CCD Central Composite Design
CFD Computational Fluid Dynamics
DOE Design of Experiment
FCCCD Face Centered Central Composite Design
LCA Linear Cellular Alloys
NBI Normal Boundary Intersection Method
RSM Response Surface Method
TLBO Teaching–Learning-based Optimization
design variables and total thermal resistance and pressure drop
as objective functions. They employed a multi-objective artificial
swarm fish algorithm with a variable population size to get Par-
eto optimal solutions. In their study, Horiuchi et al. [16] worked
on the Pareto based multi-objective optimization of pin fin heat
sinks. They used semi-analytical equations to find the relation-
ship between objective functions (heat transfer rate and pressure
drop) and four independent design parameters (height and diam-
eter of the pin fins, longitudinal and transverse pitches). Then,
they employed genetic algorithm for the multi-objective opti-
mization. Rao et al. [17] investigated the performance of a
plate-fin heat sink equipped with flow-through air cooling sys-
tems and impingement-flow system. In multi-objective optimiza-
tion definition, they selected entropy generation rate and
material cost with five constraints as objective functions while
they consider number of fins, height of fins, spacing between
two fins and oncoming air velocity as design parameters. They
obtained dynamic heat dissipation performance of the heat sink
by a commercial software. Then, they obtained Pareto fronts by
Teaching–Learning-based Optimization (TLBO) algorithm to com-
pare the performance of two cooling systems.
To the authors’ knowledge, there is no experimental investi-
gation dealing with the optimization of design parameters of a
heat sink having hexagonal aluminum honeycomb fins under
turbulent forced convection conditions. This paper reports,
therefore, the results of an experimental study that has been
conducted to determine optimal values of the design
parameters of a honeycomb heat sink using the RSM and the
Pareto based multi-objective optimization. The organization of
this paper is as follows. Section 2 explains the experimental
facilities and summarizes the analysis of experimental data.
Then, Section 3 includes application of the RSM and Pareto
based multi-objective optimization study, and combines
evaluation and interpretation of the results. Finally, Section 4
summarizes the findings.
 A. Subasi et al. / International Journal of Heat and Mass Transfer 101 (2016) 295–302
2. Experimental details and procedure
2.1. Experimental set-up
In the experiments, an open-loop wind tunnel operating in suc-
tion mode was used. The experimental set-up was schematically
illustrated in Fig. 1. The test chamber of the wind tunnel had an
internal cross section of 112  60 mm and a length of 440 mm.
The outer surfaces of the test chamber were insulated against heat
losses. A rectangular prism-shaped aluminum flat plate coupled
with a plate-type electric heater was placed in the test chamber
and used as base plate of the heat sink. The dimensions of this plate
were 242 mm in length, 111 mm in width and 6 mm in thickness.
The uniform and constant wall heat flux boundary condition
throughout the base plate was provided with a controllable power
supply. The electrical power input was adjusted with a Hioki 3333
brand wattmeter and set as 30 W constant during the experiments.
At the steady state conditions, average surface temperature of the
base plate and mean temperature of the air at the inlet and outlet
of the test chamber and the environment temperature were deter-
mined via 12 T-type thermocouples in total. The mean inlet velocity
of the air was measured by a TESTO 400 brand anemometer placed
just before the inlet of the test chamber. A KIMO CP100 brand digital
differential pressure transducer was employed to measure the pres-
sure drop throughout the test chamber. More detailed information
about the experimental set-up can be found in [18].
The aluminum hexagonal honeycombs having 0:05 mm wall
thickness were used as test elements (fins). The following steps
were carried out in preparation of the honeycombs for the
(a)
(b)
Fig. 1. (a) Schematic illustration of experimental set-up, (b) cross-section view of heating unit. (1) Effuser, (2) anemometer, (3) differential pressure transducer, (4) heating
unit, (5) test section, (6) mixer, (7) diffuser, (8) fan, (9) inlet and outlet thermocouples, (10) data acquisition card (DAQ), (11) computer, (12) variac.
297
experiments. First, the honeycombs were cut to desired sizes. Then,
aluminum layers having 0:7 mm wall thickness were sliced consid-
ering the honeycombs’ length and thickness. Then, these stripes
were polished with a thin sanding paper to minimize the effects
of surface roughness. Subsequently, polished stripes were glued
at the top and bottom surfaces of each honeycomb’s using an adhe-
sive having high thermal conductivity to obtain a more rigid struc-
ture and to ensure a full contact with the base plate. Finally, a
silicone heat transfer compound was applied between the obtained
structure and the base plate in order to reduce the contact resis-
tance. Fig. 2 shows the various views of a honeycomb heat sink
configuration.
2.2. Data reduction
This section summarizes the analysis of experimental data.
At steady state conditions, one can conclude that power input of
the electric heater is equal to heat transfer rate by convection con-
sidering that heat losses and radiation are in negligible level [8].
Therefore, the mean heat transfer coefficient ðhÞ is calculated as:
h ¼ Q_ conv =As ðT s  T b Þ ð1Þ
in which,
Q_ conv ¼ Q_ heater  Q_ cond  Q_ rad  Q_ loss ð2Þ
Q_ conv ¼ Q_ heater since ðQ_ cond þ Q_ rad þ Q_ loss Þ 6 0:5% of Q_ heater ð3Þ
T b ¼ ðT in þ T out Þ=2 ð4Þ
298 A. Subasi et al. / International Journal of Heat and Mass Transfer 101 (2016) 295–302

3. Results and discussion

3.1. Application of Response Surface Method (RSM)

The main focus of the present study is to optimize the design

parameters of a heat sink having hexagonal aluminum fins. Fin
height ðHÞ, fin thickness ðtÞ, longitudinal pitch ðSy Þ, angle of attack
ðhÞ and Reynolds number ðReÞ are selected as five independent
design parameters considering that they would affect heat transfer
and friction characteristics of the heat sink, and schematically
illustrated in Fig. 2. Furthermore, Nusselt number ðNuÞ and friction
factor ðf Þ are chosen as performance criterion of the heat sink,
namely objective functions. These objective functions need to be
expressed as mathematical expressions to use them in the opti-
mization problem formulation. Therefore, the relationship
between the objective functions and the independent design vari-
ables is given as a closed function in the following form
F ¼ f fH; t; Sy ; h; Reg ð9Þ
and the RSM introduced by Box and Wilson [20] has been
employed to construct polynomial models for the objective func-
tions. The RSM used in modeling and analysis of problems is a
method that combines mathematical and statistical techniques.
The procedure of the RSM is summarized in Fig. 3. The method
uses design of experiment (DOE) and applies regression analysis
to develop an adequate functional relationship between a
response (yield) of interest, y, and a number of associated
independent input variables which have an effect on system
response denoted as x1 ; x2 ; . . . ; xi . The accuracy and validity of
these functional relationships (i.e. mathematical models) are
evaluated by statistical tests and confirmation experiments. The
authors recommend [21] to readers who want to get more
Fig. 2. (a) Structure of a honeycomb fin, (b) top view and (c) perspective view of a information about the RSM.
heat sink configuration. In this study, a second-order polynomial model were selected
for both Nu and f . These polynomial models are then used in
place of the objective functions in the optimization problem def-
and inition after having determined the goodness of the models
As ¼ W L ð5Þ using statistical and experimental tests. A second-order polyno-
mial model can be expressed for a defined response y in Eq.
where As is projection area of the base plate, T s is mean temperature (10) as:
of the base plate surface and T b is the mean bulk temperature of the
X
n X
n X
n
air. y ¼ b0 þ b i xi þ bii x2i þ bij xi xj  e ð10Þ
In the present study, Nusselt number Nu and friction factor f are i¼1 i¼1 i<j
regarded as a measure of thermal and hydraulic performance and
calculated as follows, respectively: where b0 is a constant while bi , bii and bij define the coefficient of
linear, second degree and mixed terms of the model, respectively.
Nu ¼
h Dh
ð6Þ The e is the fitting error, and the independent design variables are
k denoted as xi . The three levels of the independent design variables
are given in Table 1. An experimental plan to determine the coeffi-
2 DP cients of a response function is usually defined using one of the DOE
f ¼ ð7Þ
Lt =Dh qu2 methods such as full factorial, fractional factorial, Box-Behnken
where Dh is the hydraulic diameter of the channel, DP is pressure Design or Central Composite Design (CCD). In this study, however,
drop along the test chamber, Lt is length of the test chamber and the face centered central composite design (FCCCD) has been used
u is mean inlet velocity of the air. The Reynolds number based on as the experiment plan [22]. Thirty-two experiments were con-
hydraulic diameter of the channel is expressed via Eq. (8). ducted to obtain values of the response parameters Nu and f accord-
ing to experimental plan given in Table 2. Then, the results of these
u Dh experiments are used to determine coefficients of the mathematical
Re ¼ ð8Þ
t models using regression analysis. The second-order polynomial
The thermo-physical properties of the air are determined models constructed for Nu and f are given in Eqs. (11) and (12),
according to bulk temperature given in Eq. (4). respectively.
Nu ¼ 89:027 þ 0:300  H  0:096  t  1:124  Sy  0:968  h
2.2.1. Uncertainty analysis
An uncertainty analysis is conducted according to the method þ 4:148  103  Re þ 0:0464  Ht  0:0244  HSy
proposed by Kline and McClintock [19], and the maximum relative þ 0:0159  Hh þ 4:151  105  HRe þ 0:1111  tSy
uncertainties for Re, Nu and f are determined as 4.15%, 7.8% and
17.1%, respectively.  4:121  105  Sy Re þ 4:192  105  hRe ð11Þ
 A. Subasi et al. / International Journal of Heat and Mass Transfer 101 (2016) 295–302 299

Table 2
Experimental plan and corresponding results.

Exp. No. Design parameters Experimental results

H t Sy h Re Nu f
1 20 6 20 0 24905.85 179.93 0.41875
2 20 6 20 30 7934.82 110.24 0.43180
3 20 6 40 0 7938.65 89.80 0.32374
4 20 6 40 30 24899.86 162.12 0.39678
5 20 15 20 0 7994.98 142.50 0.50754
6 20 15 20 30 25150.68 225.60 0.50347
7 20 15 40 0 25166.73 200.67 0.32821
8 20 15 40 30 8043.35 125.28 0.56636
9 60 6 20 0 7910.74 128.91 1.19002
10 60 6 20 30 24408.73 258.09 2.14031
11 60 6 40 0 24739.07 166.12 0.50932
12 60 6 40 30 7939.92 103.78 1.07901
13 60 15 20 0 24795.58 278.44 0.84013
14 60 15 20 30 7901.05 178.80 2.81522
15 60 15 40 0 7908.14 151.48 0.64925
16 60 15 40 30 25084.09 269.25 1.65574
17 20 10 30 15 16022.17 145.82 0.32169
18 60 10 30 15 15874.96 188.26 1.07925
19 40 6 30 15 15866.66 144.90 0.53982
20 40 15 30 15 16033.99 190.90 0.72342
21 40 10 20 15 16065.29 190.99 0.77596
22 40 10 40 15 16042.43 138.71 0.40175
23 40 10 30 0 16069.40 155.02 0.48154
24 40 10 30 30 16109.56 155.16 0.66766
25 40 10 30 15 8024.52 126.22 0.32131
26 40 10 30 15 25036.19 208.58 0.60389
27 40 10 30 15 16031.62 175.85 0.64311
28 40 10 30 15 16013.06 166.78 0.64363
29 40 10 30 15 16024.01 170.31 0.64332
30 40 10 30 15 16081.66 173.27 0.64172
31 40 10 30 15 16092.72 166.30 0.64142
32 40 10 30 15 16064.55 168.75 0.64219

3 and 4, respectively. It is concluded that the second-order polyno-

mial models formed for Nu and f are considered as significant by
taking into account that values of (Prob > F) of the models are less
than 0:05 (i.e. a ¼ 0:05 or 95% confidence interval). The other sta-
tistical criteria to evaluate statistical significance of the models
obtained are the R2 and Adj  R2 values. According to Tables 3
and 4, the R2 values for Nu and f are 98:3% and 96:6% while the
Adj  R2 values for Nu and f are 97:2% and 94:4%, respectively.
Fig. 3. Flowchart of multi-objective optimization procedure using RSM [7,25–27].
Since these values are close to unity, and the difference between
the R2 and Adj  R2 values for both Nu and f are small enough, it
Table 1
The design parameters and their values.
is concluded that the polynomial models obtained for Nu and f suf-
ficiently represent relationship between the independent design
Design parameters Levels
variables and responses. Moreover, in addition to statistical analy-
Low Medium High sis, confirmation experiments are conducted to determine whether
–1 0 +1 the models formed fully represents the process. The configurations
Fin height H (mm) 20 40 60 of these experiments are randomly selected considering that val-
Fin thickness t (mm) 6 10 15 ues of the independent design parameters should be in the studied
Longitudinal pitch Sy (mm) 20 30 40
Angle of attack h (°) 0 15 30
interval, and selected configurations shouldn’t be the same with
Reynolds Number Re (u Dh/t) 8000 16500 25000 the experimental plan given in Table 2. Then, Nu and f values
obtained from these experiments are compared to the results cal-
culated from the second-order polynomial models formed. The rel-
f ¼ 0:4753  0:0181  H þ 0:0420  t þ 5:481  103  Sy  0:0191
ative error between the experimental and predicted results are
 h  3:416  106  Re  8:851  104  HSy þ 8:702  104 presented in Table 5. According to Table 5, the maximum relative
 Hh þ 1:536  103  th  2:761  106  tRe  4:400  104 error for Nu and f are 4:02% and 5:20%, respectively. It is obvi-
ous that these errors are within the acceptable margin when con-
 Sy h þ 9:714  107  Sy Re þ 6:777  104  H2 ð12Þ
sidering the similar studies [7,14].
The goodness of these polynomial models obtained are deter- With aforementioned statistical and experimental analysis, we
mined both statistically and experimentally by applying the anal- proved that the polynomial models given in Eqs. (11) and (12)
ysis of variance (ANOVA) method and confirmation experiments are well fitted and valid within the range of the parameters given
before using them as objective functions. in Table 1. Thus, they can be used to predict the values of Nu and f .
The statistical significance of the terms in a mathematical The effects of the design parameters on Nu and f can be deter-
model and a model itself can be determined using a variance anal- mined by means of the ANOVA results given in Tables 3 and 4,
ysis. The results of the ANOVA for Nu and f are presented in Tables respectively. The design parameters effecting Nu are, in sequence,
300 A. Subasi et al. / International Journal of Heat and Mass Transfer 101 (2016) 295–302

Table 3
ANOVA table for Nu.

Source Sum of squares Degrees of freedom Mean square f-value Prob. > F
Model 59128.10 12 4927.34 90.84 <0.0001 Significant
H 6884.51 1 6884.51 126.92 <0.0001
t 9513.26 1 9513.26 175.38 <0.0001
Sy 4684.90 1 4684.90 86.37 <0.0001
h 520.15 1 520.15 9.59 0.0059
Re 35016.21 1 35016.21 645.54 <0.0001
Ht 279.76 1 279.76 5.16 0.0350
H  Sy 381.64 1 381.64 7.04 0.0157
Hh 362.43 1 362.43 6.68 0.0182
H  Re 791.91 1 791.91 14.60 0.0012
t  Sy 400.68 1 400.68 7.39 0.0137
Sy  Re 195.15 1 195.15 3.60 0.0732
h  Re 454.45 1 454.45 8.38 0.0093
Residual 1030.61 19 54.24
Cor. total 60158.71 31
Standard deviation 7.36 R2 0.9829
Mean 169.90 Adj.-R2 0.9720
Coefficient of variation 4.33 Predicted R2 0.9101
Predicted residual of sum of squares (PRESS) 5408.45 Adequate precision 39.2970

Table 4
ANOVA table for f.

Source Sum of squares Degrees of freedom Mean square f-value Prob. > F
Model 8.745 12 0.729 44.399 <0.0001 Significant
H 3.6898 1 3.689 224.785 <0.0001
t 0.139 1 0.139 8.446 0.0091
Sy 0.756 1 0.756 46.067 <0.0001
h 1.408 1 1.408 85.779 <0.0001
Re 0.0137 1 0.014 0.840 0.3710
Ht 0.5013 1 0.501 30.547 <0.0001
H  Sy 1.090 1 1.090 66.436 <0.0001
Hh 0.172 1 0.172 10.491 0.0043
H  Re 0.178 1 0.178 10.826 0.0038
t  Sy 0.0697 1 0.070 4.247 0.0533
Sy  Re 0.1084 1 0.108 6.606 0.0187
h  Re 0.5767 1 0.577 35.137 <0.0001
Residual 0.31 19 0.016
Cor. total 9.06 31
Standard deviation 0.13 R2 0.9656
Mean 0.75 Adj.-R2 0.9438
Coefficient of variation 16.99 Predicted R2 0.8604
Predicted residual of sum of squares (PRESS) 1.26 Adequate precision 29.7610

Reynolds number ðReÞ, fin thickness ðtÞ, fin height ðHÞ, longitudinal
pitch ðSy Þ and angle of attack ðhÞ. The most effective design param-
eters on f , in sequence, are fin height ðHÞ, angle of attack ðhÞ, lon-
gitudinal pitch ðSy Þ, fin thickness ðtÞ, and Reynolds number ðReÞ.
Furthermore, the effects of the mixed and second order terms in
f are even higher than that of the design variables while effect of
the mixed terms in Nu is less than that of the design variables.
3.2. Multi-objective optimization
The thermal performance of a heat sink can be improved
by enhancing heat transfer rate and/or reducing the pressure
drop [3]. Additionally, it is a well-known fact that usage of
extended surfaces results in increasing heat transfer rate and
pressure drop, so Nu and f too. In such studies, on the other
hand, Nu is desired to be as high as possible while f is
desired to be as low as possible to achieve a high perfor-
mance heat sink. Therefore, in this study, Nu and f are
adopted as objective functions and an optimization study has
been conducted to determine optimum values of the design
parameters corresponding to maximum Nu and minimum f
under the constraints of the investigated values of the design
parameters. The design optimization problem can be described
as follows:

Table 5
Results of confirmation experiments and their comparison with predicted values.

Exp. No. Design parameters Nusselt number, Nu Friction factor, f

H t Sy h Re Exp. Predicted Error (%) Exp. Predicted Error (%)
1 60 15 20 30 11249.69 209.96 208.03 0.92 2.660 2.643 0.63
2 60 15 40 30 11266.20 183.50 180.42 1.67 1.735 1.645 5.20
3 20 10 30 15 19678.25 166.25 164.58 1.00 0.413 0.408 1.14
4 40 15 30 15 20104.65 222.96 214.01 4.02 0.667 0.637 4.59
 A. Subasi et al. / International Journal of Heat and Mass Transfer 101 (2016) 295–302 301

The optimization formulation defined in Eq. (13) contains more

than one objective function. Thus, this type of optimization prob-
lems are called ‘‘multi-objective optimization problem”, and all
objective functions have to be minimized/maximized simultane-
ously. Contrary to optimization problems having only one objec-
tive function, multi-objective optimization problems such as
defined in Eq. (13) result in a ‘‘Pareto optimal set” which contains
infinite number of ‘‘Pareto optimal solutions”, and objective func-
tions corresponding to this set form a ‘‘Pareto front” [23]. In this
study, second-order polynomial models (Eqs. (11) and (12)) of
the objective functions obtained previously by the RSM has been
employed in the optimization formulation, and the Normal Bound-
ary Intersection Method (NBI) has been utilized to obtain the Par-
eto optimal solutions. The details of the NBI can be found in [24].
The result of optimization problem defined by Eq. (13) is pre-
sented in Fig. 4. Pareto front shown in Fig. 4 consists of infinite
Fig. 4. Pareto front and Pareto optimal set.
number of points, namely Pareto optimal solutions, and fifteen of
these points and their corresponding values of design parameters
have been tabulated in Table 6. In Pareto optimal approach, to
Table 6
improve an objective function can only be provided by sacrificing
Pareto optimal solutions. from other objective functions, and each Pareto optimal solution
is equally satisfy the multi-objective optimization problem. That
No. Design parameters Objective functions
means each point in Table 6 has no superiority over one another,
H t Sy h Re Nu f and a designer/decision maker can select any solution on the
1 39.46 15 40 0 25000 213.54 0.1199 Pareto front. Therefore, a designer/decision maker has a flexibility
2 29.35 15 20 0 25000 226.91 0.1135 in decision phase [23]. Considering to Table 6, one can conclude
3 36.20 15 20 0 25000 237.52 0.1728
that Re and t are the same for all optimal solutions, so a
4 41.97 15 20 0 25000 246.43 0.2720
5 47.05 15 20 0 25000 254.28 0.3966 designer/decision maker can only change H, Sy and h. Fig. 5 given
6 51.63 15 20 0 25000 261.37 0.5391 to explain this situation shows the configurations of three optimal
7 55.85 15 20 0 25000 267.88 0.6953 solutions marked as Design A, B and C in Fig. 4. It is clear from Fig. 5
8 59.77 15 20 0 25000 273.95 0.8622
that Design A has shorter and less honeycombs compared to the
9 60 15 20 4.06 25000 278.49 1.0649
10 60 15 20 8.38 25000 282.96 1.2697
other designs. For example, if one has a limited space to install
11 60 15 20 12.71 25000 287.42 1.4745 the heat sink, in such a case, selection of Design A among the
12 60 15 20 17.03 25000 291.88 1.6793 optimal solutions appears more appropriate in terms of material
13 60 15 20 21.35 25000 296.34 1.8841 saving and space constraints provided that it satisfies the
14 60 15 20 25.68 25000 300.81 2.0889
requirements.
15 60 15 20 30.00 25000 305.27 2.2936

4. Conclusion
8
>
> Find H; t; Sy ; h; Re An experimental study has been conducted to find optimum
>
>
>
>
> E v aluate Nu ¼ f fH; t; Sy ; h; Reg and f ¼ f fH; t; Sy ; h; Reg values of the design parameters of a honeycomb heat sink. The
>
>
>
> Minimize f ¼ Nu and f 2 ¼ f RSM is employed to express relationship between the objective
>
> 1
>
> functions ðNu and f Þ and the design parameters ðH; t;
>
< Subject to :
Sy ; h and ReÞ, as polynomial functions. The variance analysis and
20 6 H 6 60
>
> confirmation experiments are performed to show validity of these
>
> 6 6 t 6 15
>
> functions. By utilizing these polynomial functions obtained, a
>
>
>
> 20 6 Sy 6 40 multi-objective optimization study is carried out to determine
>
>
>
> 0 6 h 6 30 optimal values of the design parameters that maximize Nu and
>
>
: minimize f , subject to the range of design parameters studied.
8000 6 Re 6 25000
The Pareto optimal solutions are obtained by the NBI. The main
ð13Þ results can be summarized as follows.

Design A Design B Design C

Fig. 5. Configurations of three optimal solutions marked as Design A, B and C in Fig. 4 (correspond to Nos. 1, 8 and 15 in Table 6, respectively).
302 A. Subasi et al. / International Journal of Heat and Mass Transfer 101 (2016) 295–302
 The ANOVA results showed that the second-order polynomial
models for Nu and f are significant and adequately represent
the experimental results since the (Prob > F) value of the models
are less than 0:05 and the R2 and Adj  R2 values of the models
are close to 1.
 The confirmation experiments also verified that the models
obtained can be used to predict the values of Nu and f within
studied range of the design parameters because the maximum
relative error between the experimental and predicted results
are small enough and fall in the acceptable range.
 The optimization result given as Pareto solution set offers
designers important insights into the design parameters and
allows designers freedom to make a selection among the Pareto
optimal solutions since each of these solutions is an optimal
solution.
References
[1] H. Kristiansen, Thermal management in electronics, Chalmers University of
Technology, Göteborg, Sweden, 2001. [Online]. Available: <http://www.ppd.
chalmers.se/edu/mpr235/mpr235_thermgmnt.pdf>.
[2] Y.A. Cengel, Heat and Mass Transfer: A Practical Approach, McGraw-Hill,
Boston, 2007.
[3] K. Park, S. Moon, Optimal design of heat exchangers using the progressive
quadratic response surface model, Int. J. Heat Mass Transfer 48 (11) (2005)
2126–2139.
[4] M.M. Rashidi, O.A. Beg, A.B. Parsa, F. Nazari, Analysis and optimization of a
transcritical power cycle with regenerator using artificial neural networks and
genetic algorithms, Proc. Inst. Mech. Eng. Part A, J. Power Energy 225 (6) (2011)
701–717.
[5] A.G. Munoz, V. Ayala-Ramirez, J.a. Alfaro-Ayala, B.M. Toledo Acosta,
Optimization of the transition piece applying genetic algorithms, Appl.
Therm. Eng. 31 (16) (2011) 3214–3225.
[6] M.M. Rashidi, N. Galanis, F. Nazari, A. Basiri Parsa, L. Shamekhi, Parametric
analysis and optimization of regenerative Clausius and organic Rankine cycles
with two feedwater heaters using artificial bees colony and artificial neural
network, Energy 36 (9) (2011) 5728–5740.
[7] K.T. Chiang, C.C. Chou, N.M. Liu, Application of response surface methodology
in describing the thermal performances of a pin–fin heat sink, Int. J. Therm. Sci.
48 (6) (2009) 1196–1205.
[8] B. Sahin, A. Demir, Performance analysis of a heat exchanger having perforated
square fins, Appl. Therm. Eng. 28 (5–6) (2008) 621–632.
[9] Q. Zhang, X. Yang, P. Li, G. Huang, S. Feng, C. Shen, B. Han, X. Zhang, F. Jin, F. Xu,
T.J. Lu, Bioinspired engineering of honeycomb structure – Using nature to
inspire human innovation, Prog. Mater Sci. 74 (2015) 332–400.
[10] T. Wen, F. Xu, T.J. Lu, Structural optimization of two-dimensional cellular
metals cooled by forced convection, Int. J. Heat Mass Transfer 50 (13–14)
(2007) 2590–2604.
[11] S. Liu, Y. Zhang, P. Liu, New analytical model for heat transfer efficiency of
metallic honeycomb structures, Int. J. Heat Mass Transfer 51 (25–26) (2008)
6254–6258.
[12] T.J. Lu, Heat transfer efficiency of metal honeycombs, Int. J. Heat Mass Transfer
42 (11) (1999) 2031–2040.
[13] B.M. Dempsey, S. Eisele, D.L. McDowell, Heat sink applications of extruded
metal honeycombs, Int. J. Heat Mass Transfer 48 (3–4) (2005) 527–535.
[14] K.T. Chiang, Modeling and optimization of designing parameters for a parallel-
plain fin heat sink with confined impinging jet using the response surface
methodology, Appl. Therm. Eng. 27 (14–15) (2007) 2473–2482.
[15] Y. Chen, B. Peng, X. Hao, G. Xie, Fast approach of Pareto-optimal solution
recommendation to multi-objective optimal design of serpentine-channel
heat sink, Appl. Therm. Eng. 70 (1) (2014) 263–273.
[16] K. Horiuchi, A. Nishihara, K. Sugimura, Multi-objective optimization of water-
cooled pinfin heatsinks, Int. J. Heat Mass Transfer 81 (2015) 760–766.
[17] R.V. Rao, G.G. Waghmare, Multi-objective design optimization of a plate-fin
heat sink using a teaching–learning-based optimization algorithm, Appl.
Therm. Eng. 76 (2015) 521–529.
[18] A. Subasi, Heat exchanger optimization using response surface method (MSc
thesis), Department of Mechanical Engineering, Ataturk University, Erzurum/
Turkey, 2010.
[19] S.J. Kline, F.A. Mcclintock, Describing uncertainties in single sample
experiments, Mech. Eng. 75 (1953) 3–8.
[20] G.E.P. Box, K.B. Wilson, On the experimental designs for exploring response
surfaces, Ann. Math. Stat. 13 (1951) 1–45.
[21] R.H. Myers, D.C. Montgomery, Response Surface Methodology: Process and
Product in Optimization using Designed Experiments, first ed., John Wiley &
Sons Inc, New York, NY, USA, 1995.
[22] D.C. Montgomery, Design and Analysis of Experiments, John Wiley & Sons,
2006.
[23] O. Koksoy, G. Hocaogu, Multi-objective optimization solutions to the Taguchi’s
problem, G.U. J. Sci. 18 (4) (2005) 613–626.
[24] I. Das, J. Dennis, Normal-boundary intersection: an alternate method for
generating pareto optimal points in multicriteria optimization problems, Soc.
Ind. Appl. Math. J. Optim. 8 (1998) 631–657.
[25] H.K. Kansal, S. Singh, P. Kumar, Parametric optimization of powder mixed
electrical discharge machining by response surface methodology, J. Mater.
Process. Technol. 169 (3) (2005) 427–436.
[26] K. Foli, T. Okabe, M. Olhofer, Y. Jin, B. Sendhoff, Optimization of micro heat
exchanger: CFD, analytical approach and multi-objective evolutionary
algorithms, Int. J. Heat Mass Transfer 49 (5–6) (2006) 1090–1099.
[27] T. Goel, R. Vaidyanathan, R.T. Haftka, W. Shyy, N.V. Queipo, K. Tucker, Response
surface approximation of Pareto optimal front in multi-objective optimization,
Comput. Methods Appl. Mech. Eng. 196 (4–6) (2007) 879–893.