International Journal of Heat and Mass Transfer 121 (2018) 663–679

Contents lists available at ScienceDirect

International Journal of Heat and Mass Transfer

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

Experimental and numerical investigation of a mini channel forced air

heat sink designed by topology optimization
Shi Zeng, Bugra Kanargi, Poh Seng Lee ⇑
Department of Mechanical Engineering, National University of Singapore, 9 Engineering Drive 1, 117576, Singapore

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

Article history: This work presents a method of designing an air heat sink with forced convection by topology optimiza-
Received 18 August 2017 tion. Both pressure drop and heat transfer performances are evaluated. To reduce computational cost, a
Received in revised form 10 January 2018 2D two-layer model is first developed and implemented in COMSOL Multiphysics to represent three-
Accepted 10 January 2018
dimension fully conjugate heat transfer modeling. It has been shown to be accurate in temperature field
Available online 7 March 2018
prediction and able to capture trends of pressure drop variation. Through a multi-stage optimization pro-
cess, a non-conventional fin structure is created. The optimized structure is then manufactured and
Keywords:
experimentally validated. Compared to a conventional straight channel heat sink, the topology optimized
Topology optimization
Forced convection
heat sink can achieve lower junction temperatures with the same pumping power or requires lower
Air heat sink pumping powers for maintaining the same junction temperature. Furthermore, full 3D numerical analysis
Multi-stage optimization by ANSYS Fluent is performed to study the detailed characteristics of the topology optimized heat sink. It
2D two-layer model shows that the non-conventional layout of the fins introduces strong mixing effect, continuous boundary
Experimental validation layer interruption and local high speeds, which all contribute to heat transfer enhancement.
Ó 2018 Elsevier Ltd. All rights reserved.

1. Introduction heat transfer. Following this idea, wavy channels were numerically
studied by Sui et al. [5] and proved to increase the heat transfer
As the power density of electronic devises is rapidly increasing, coefficient by 153% while increasing the friction factor only by
thermal management has become a major challenge today [1]. The 54% at Re = 800. Furthermore, the wavy channel concept was com-
heat flux of modern microprocessors could easily go up to the bined with secondary branches, which leads to a performance
order of 10,000 W=m2 . For example, Intel Xeon Processor E5- boost as high as 190% [6]. Similar secondary flow concept was
2600 v4 processors have to dissipate around 100 W of heat within applied in a new sectional oblique fin design proposed by Lee
a substrate of 52:5
45 mm [2]. A widely adapted solution is et al. [7]. It was shown to increase heat transfer performance by
attaching chip packages onto heat sinks with straight fins and pas- 50–60% with almost no additional pressure drop penalty compared
sively or actively cooling them using air. In research, the micro- to conventional microchannels.
and mini-channel heat sinks have gained its popularity as a Inspired by natural structures, Bejan and Errera [8] proposed
promising and reliable technique for high density power dissipa- the tree-shaped channel network based on constructal theories.
tion. In the pioneering work of Tuckerman and Pease [3], a 50 Similar fractal-like flow network was studied by Pence [9] and
lm wide and 300 lm deep microchannel heat sink was capable Chen and Cheng [10]. Except for fin structures, there were also sev-
of dissipating a power density of 790 W=cm2 with a junction tem- eral reports about designing the whole heat sinks including mani-
perature of 71  C. However, the initial microchannel concept had folds. Sharma et al. [11] demonstrated a novel concept for energy
several evident drawbacks, such as large pressure drop penalty efficiency-hotspot-targeted liquid cooling of multicore micropro-
and significant lateral temperature gradient. To circumvent these cessors. Through the introduction of the flow-throttling zones,
disadvantages and further enhance the thermal performance, higher flow rates were guaranteed over hotspots. As a result, chip
many different fin and channel structures have been investigated. temperature non-uniformities were greatly reduced. Other system
As highlighted by Steinke and Kandlikar [4], channel curvature level designs can also be found in [12–16].
could skew the traditional parabolic velocity profile to enhance The designs mentioned above were all initiated from intuition
or experience. With the increase in computational power, it has
become a popular practice to optimize designs through numerical
⇑ Corresponding author. simulations. In general, there are three levels of optimization: size,
E-mail addresses: [email protected] (S. Zeng), [email protected] (P.S. Lee).

https://doi.org/10.1016/j.ijheatmasstransfer.2018.01.039
0017-9310/Ó 2018 Elsevier Ltd. All rights reserved.
664 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679

Nomenclature

A area (m2 ) Greek symbols

cp specific heat (J=kg  K) D gradient (difference)
dz thickness (m) a friction coefficient
D diameter (m) l dynamic viscosity (Pa  s)
h convection coefficient (W=m2  K) q density (kg=m3 )
H height (m) c local design variable
I current (A) g fin efficiency
k thermal conductivity (W=m  K)
m _ mass flow rate (kg=s) Subscripts
P pressure (Pa) ave average
q_ heat transfer rate (W) b base
€
q heat flux (W=m2 ) bt bottom layer
qf =qk =qh convexity parameter f fin
Q power (W) fl fluid
R thermal resistance ( C=W) HS heat sink
SC straight channel in inlet
TO topology optimized min minimum
T temperature ( C) max maximum
u velocity vector out outlet
V voltage (V) s solid
v velocity (m/s) top top layer
Dh hydraulic diameter (m) th thermal
Re Reynolds number wet wet surface
Nu Nusselt number sec sectional area
f friction factor
L length (m)

shape and topology optimization. Size optimization requires well- equation as an additional friction force term. For c ¼ 0, a  1,
predefined geometric parameters such as length, width and depth. resulting in a reduction of local velocity to 0; for c ¼ 1, a ¼ 0,
It aims to find a set of dimensions that will maximize the overall retaining the original momentum equation. Extreme values of c, 0
performance. As in [6], various aspect ratios were examined in and 1, represented solid phase and fluid phase, respectively. The
order to find the optimum values of the key dimensions. In shape initial problem of optimization was then transferred to the opti-
optimization, on the other hand, geometric boundaries are nor- mization of values of c in each element of the design domain. With
mally defined by curve functions. Boundaries are adjusted as the minimizing pressure drop as the objective, the method was suc-
functions change. Hilbert et al. [17] presented a successful imple- cessful implemented in the design of a diffuser, a pipe bend, a rugby
mentation on the blade shape of a heat exchanger. Finally, in topol- ball and a double pipe. Soon after this pioneering work, Gersborg-
ogy optimization, size, shape and topology are optimized Hansen et al. [22] extended it from Stokes flow to incompressible
simultaneously [18]. It does not require predefined geometric laminar viscous flow at low-to-moderate Reynolds numbers. Sev-
parameters and allows the formation of new boundaries. There- eral other studies also reported successful implementation of this
fore, it is the highest level of optimization. The difference between method in designing various flow structures [23–25]. As a real
these three optimization methods is shown in Fig. 1: in the opti- application, a microfluidic mixer was designed with a 70% increase
mization of a fin and tube heat exchanger, the diameter of the tube in mixing accompanied by a 2.5 times increase in the pressure drop
is the parameter to be defined and it is adjusted in size optimiza- penalty [26].
tion; in shape optimization, the tube may not remain as a circle The method is also applicable to cases of heat conduction where
but morph to other geometries; in topology optimization, no defi- the distribution of high conductivity material is to be optimized. As
nition of prescribed tube geometry is required and novel geome- reported in [27], this method generated a leaf like conductance
tries may be generated. path for a volume to point heat dissipation problem. At the system
Topology optimization was first introduced by Bendsøe and design level, distribution of insulation material and thermal con-
Kikuchi [19] in 1988 for the design of mechanical elements that nection between a thermoelectric cooler and a structural chassis
can withstand given loads. An overview of this method was pro- were optimized to ensure sufficient cooling for temperature sensi-
vided in the monograph by Bendsøe and Sigmund [20]. In the tive electronic components in a downhole oil well intervention tool
monograph, topology optimization was described as a material dis- [28]. The optimized result was then validated by experiments.
tribution method in which a design domain was discretized and In cases of natural convection where fluid flow and heat transfer
each element of the domain was decided to be occupied as either are combined together, it is worth mentioning the work by Alexan-
void or material. In 2003, this method was extended by Borrvall dersen et al. [29]. In their work, various novel heat sink structures
and Petersson [21] to include physics of fluid flow for the first time, were generated under different Grashof numbers. However, due to
a technique termed as topology optimization of fluids. In this the three-dimensional simulation of conjugate heat transfer, the
method, each element of the discretized design domain was computational cost was considerably high and required a cluster
assigned a local design variable c which varied continuously from for implementation.
0 to 1. The authors then introduced a friction coefficient a; which Finally, there were also several attempts in utilizing topology
was a function of the local design variable c. The product of a optimization in heat sink design with forced convection. Oevelen
and velocity u (au) was then introduced to the momentum and Baelmans [30] first introduced a 2D numerical model to solve
 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679 665

(a) (b) (c)

Fig. 1. Illustration of (a) size, (b) shape and (c) topology m optimization.

the conjugate heat transfer problem in water heat sinks with the flow and heat transfer characteristics of the TO heat sink are
assumption of fully developed flow in the direction of height and described based on 3D numerical simulations.
constant heat sink temperature. Designs with channel width as This paper is organized into the following sections: Section 2
small as 0.1 mm were presented and dead ends of channel introduces and verifies the simplified model for simulating heat
branches were observed in their results. Koga et al. [31] generated sinks; Section 3 describes the formation and implementation pro-
several water heat sink designs by topology optimization in Stokes cess of topology optimization; Section 4 describes the numerical
flow regime with uniform heat source through the design domain. and experimental study of the TO heat sink design; discussion is
Using the same uniform heat source formulation, a multipass presented in Section 5, followed by conclusion in Section 6.
branching microchannel heat sink was created [32]. The design
was then numerically and experimentally tested [33]. However, 2. A simplified 2D two-layer model
the studies mentioned above are more focused on presenting the
concept of topology optimization and lack performance validation Full 3D simulation of conjugate heat transfer is highly computa-
and analysis of the optimized designs. An exception is another tionally expensive. Topology optimization may run for hundreds of
work by Dede et al. [34], in which the topology optimized air heat iterations and each iteration requires solving the conjugate heat
sink was manufactured, experimentally tested and compared with transfer model once. Without the computation power like a cluster
other designs to analyze its performance. However, during opti- used in [29], it is almost impossible to implement topology opti-
mization process, only heat conduction was considered since the mization. Thus, a 2D two-layer model is first developed to reduce
convection coefficient was predetermined. the simulation domain from 3D to 2D. Different from the simplified
The scope of this study is to design an air heat sink with forced models in [34–38], the interaction between air and fins is captured
convection by topology optimization, and verify its performance in our model.
with both numerical simulations and experimental tests. A simpli-
fied 2D model is developed to represent 3D conjugate heat transfer 2.1. Development of the simplified model
model to reduce the computational cost. A multi-stage optimiza-
tion strategy is employed to decrease the inaccuracy caused by A heat sink is composed of two parts: fins and channels at the
the treatment of fictitious porous design domain and gradually top and a solid base at the bottom as shown in Fig. 2. Assuming
improve the design’s performance. The topology optimized (TO) that the temperature of each part in the direction of height is uni-
heat sink is compared with a conventional straight channel (SC) form, fins and channels can be represented by the top layer and the
heat sink in experimental tests. Finally, detailed analysis of the solid base can be represented by the bottom layer in the 2D model.

Fin structure

Air
Fin & Channel Top layer

Base
̇ Bottom layer

Fig. 2. A simplified 2D two-layer model for a heat sink simulation.

666 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679

Fig. 3. Single straight channel 3D simulation domain (adapted form [41]).

hfl ðT bt  T top Þ
qfl cfl u  rT top ¼ r  ðkfl rT top Þ þ ð3Þ
dztop

where u is the velocity vector, P is the pressure field, Ttop is the top
layer temperature field, Tbt is the bottom layer temperature field,
dztop is the top layer thickness, which is the fin height, hfl is the con-
vection heat transfer coefficient between the fluid and the solid
Fin and channel base, and qfl ; cfl ; kfl are the density, specific heat capacity and con-
ductivity of air, respectively.
For the solid fin domains in the top layer, only heat conduction
Base occurs and the energy equation is expressed as:

hs ðT bt  T top Þ
r  ðks rTtop Þ þ ¼0 ð4Þ
dztop

z
where ks is the heat conductivity of the heat sink material, and hs is
y the nominal convection heat transfer coefficient between fins and
the solid base to account for the conduction heat transfer.
. The terms with hfl and hs in Eqs. (3) and (4) are the convective
boundary conditions through which the top and bottom layers are
Fig. 4. Thermal resistance model of an air heat sink. thermally coupled. The convection coefficients, hfl and hs , are dif-
ferent for channel domains and fin domains due to the nature of
different heat transfer mechanism and will be discussed in detail
in the next section. Due to this difference, the energy equation in
Inside the bottom layer, only heat conduction occurs. The constant
the bottom layer also has to take two forms:
heat flux boundary condition that is applied on the bottom surface
In domains beneath fluid domains of the top layer:
of the heat sink base in the 3D model can be represented with a
uniform heat source throughout the bottom layer. At the top layer,
convective heat transfer between fluid and fins is fully captured by €bt
q hfl ðT bt  T top Þ
r  ðks rT bt Þ þ  ¼0 ð5Þ
solving conjugate heat transfer model. Finally, a convective heat dzbt dzbt
transfer boundary condition is applied to combine these two layers In domains beneath solid fin domains of the top layer:
together.
As such, the governing equations for fluid domains in the top
€bt
q hs ðT bt  T top Þ
layer include continuity, momentum and energy equations: r  ðks rT bt Þ þ  ¼0 ð6Þ
dzbt dzbt
ru¼0 ð1Þ
€ bt is the heat flux, and dzbt is the height of the bottom layer,
where q
qfl ðu  rÞu ¼ rP þ lr2 u ð2Þ
which is the heat sink base thickness.
 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679 667

Table 1
Simulation results comparison of 3D and simplified 2D models for the SC heat sink.

Re vin (m/s) T max ( C) T av e ( C) DP (Pa)

2D 3D 2D 3D 2D 3D
150 0.287 95.6 94.9 94.8 94.0 0.59 1.23
200 0.431 73.7 73.5 73.0 72.8 0.90 1.75
250 0.574 63.3 63.6 62.7 63.0 1.23 2.32
400 1.149 48.1 49.1 47.7 48.6 2.85 5.15
600 1.723 43.8 43.8 43.4 43.4 4.73 8.73
800 2.298 40.3 40.8 39.9 40.3 7.33 13.00
1000 2.872 38.8 38.7 38.5 38.2 9.48 17.90

2.2. Determination of hfl and hs The heat transfer rate from the bottom surface to the fin could
be expressed as:
Similar two-layer treatment has been seen in literature. For
example, in [39], the author defined a temperature dependent heat T bt  T top
q_ ¼ ð10Þ
source term which is similar to the convection heat transfer term R1 þ R2
in Eqs. (3) and (4). In [40], the authors used a very similar formu- If expressed by nominal convection heat transfer coefficient, it
lation as in this study. However, the convection heat transfer coef- is:
ficients were not well defined, instead, randomly set. This study
takes a step further to ensure the accuracy of the 2D two-layer q_ ¼ hs Ab ðT bt  T top Þ ð11Þ
model by determining the coefficients based on 3D full scale
numerical simulations and a thermal resistance model. where Ab is the fin base area.
To determine hfl , the convection heat transfer coefficient So hs is derived as:
between fluid and the heat sink base, 3D simulation results of a
SC heat sink from [41] is used. For completeness, we briefly intro- 1 1
hs ¼ ¼ ð12Þ
duce the work related here. ðR1 þ R2 ÞAb dzbt þ ð1gÞAb
ks gAf hav e
The heat sink studied was 120 mm long and 100 mm wide with
a fin width of 1 mm and a channel width of 4 mm. As shown in The fin efficiency is further assumed to be 1 considering that the
Fig. 3, to save computation cost, only a single channel was selected conduction resistance through the fin is much lower than the con-
and periodic boundary condition was set on both sides of the sim- vection resistance. In this study, the heat sink material is alu-
ulation domain. The problem was solved in ANSYS Fluent 15.0 with minum alloy 6061-T6 with a thermal conductivity of 168
laminar viscous flow model for Reynolds number from 150 to W=ðm  KÞ and base thickness dzbt is fixed as 17 mm. Thus, hs is cal-
1000. The average convection heat transfer coefficient was evalu- culated as 9882 W=ðm2  KÞ.
ated as: hs is found to be 365 times higher than hfl , which means most of
the heat is conducted into fins then absorbed by air through con-
¼ q_ vection rather than directly taken away through convection on
h ð7Þ
ðT av e;wet  T av e;air Þ  Awet base unfinned areas.
where T av e;wet is the average wetted surface temperature, T av e;air is
the arithmetic average of air inlet and outlet temperatures, q_ is 2.3. Validation of the 2D two-layer model
the heat transfer rate, and Awet is the wetted surface area, consisting
of both fin surface area and bottom unfinned base area. A commercial software COMSOL Multiphysics 5.2a, which per-
 is an overall evaluation reflecting how strong the heat transfer
h forms numerical simulations based on the finite element method,
is between air and the heat sink, and it will increase with the is used to solve the 2D two-layer model. A SC heat sink with an
increasing Reynolds number. A h  of 27 W=m2  K, at Re ¼ 400 and aspect ratio of 2 (channel height to width ratio) was studied. The
a heat flux of 1666.7 W=m2 was selected as hfl in the 2D two- boundary conditions are the same as the 3D simulations in [41]:
layer model. constant heat flux of 1666.7 W=m2 , inlet temperature of 20  C, uni-
On the other hand, hs determines the amount of heat trans- form velocity at the inlet and 0 Pa gage pressure at the outlet, and
ferred from the solid base to fins. In fact, this part of heat transfer no-slip conditions on the channel walls. The results compared with
is heat conduction. To calculate hs , a thermal resistance model of the 3D simulations under various flow rates are in Table 1.
the whole heat sink is considered as in Fig. 4. T max is the maximum heat sink temperature, T av e is the average
Different from [42,43], heat conduction in the base is further bottom surface temperature, and DP is the pressure drop. As
assumed to happen only in the height direction, which means shown in the table, the difference of the results from two models
the spanwise heat conduction is neglected. As such, R1 is expressed is less than 1.1  C for both T max and T av e under all flow rates, which
as: proves the validity of the simplified 2D two-layer model in predict-
ing the heat transfer performance of finned heat sinks.
dzbt For pressure drop, the results from 3D simulations are always
R1 ¼ ð8Þ
Ab ks higher than the results from the simplified model, and the discrep-
ancy gradually increases as the inlet velocity increases. It is not
where Ab is the base finned area. Based on the fin efficiency theory,
surprising since 2D flow is treated as a channel of infinite height,
R2 could be expressed as:
therefore the friction forces at the bottom and top walls are
1g neglected. However, similar as the results from 3D simulations,
R2 ¼ ð9Þ pressures drops are also higher for higher inlet velocities in 2D
gAf hav e
simulations. Thus, for different fin structures, this method is cap-
where g is the fin efficiency, Af is the fin surface area, and hav e is the able of determining a topologically optimum design, for which
average heat transfer coefficient on the fin surface. low pressure drop or pumping power are set as the objectives.
668 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679

Top design domain

Air inlet
= 1.149 /
= 20℃
8 mm
17 mm

̈ = 1666.7 / 2

(a)

(b)
Fig. 5. Illustration of the problem setup (a) 3D overview; (b) reduced 2D simulation domain.

The design variable is included into governing equations to impose

the influence of different material phases. In the Navier-Stokes
equation, a friction force term is added [21,22]:

qfl ðu  rÞu ¼ rP þ lr2 u  au ð13Þ

where a, the friction force coefficient or inverse permeability, is

dependent on the local design variable c by an interpolation
function.
In the heat transfer phase, the introduction of design variable
differs in literature. Some work [31,44,45] interpolated all the
material properties, including density, specific heat capacity and
conductivity, as well as the convective heat transfer coefficient if
a convection boundary was applied. More recent work [32,34]
Fig. 6. Design variable field before (left) and after (right) density filtering.
mostly only interpolated the conductivity and the convective heat
transfer coefficient if a convection boundary was applied. In this
In summary, a simplified 2D two-layer model is developed study, the second approach is used. The conductivity and convec-
based on the assumption that the heat sink base and fins hold their tion heat transfer coefficient are interpolated in a way that:
own uniform temperature distribution in height direction. The two


layers are coupled by a convective boundary condition in which ks ; c ¼ 0; solid phase hs ; c ¼ 0; solid phase
kðcÞ ¼ hðcÞ ¼
heat transfer coefficients for unfinned and solid finned areas are kfl ; c ¼ 1; fluid phase hfl ; c ¼ 1; fluid phase
determined from 3D simulations and a thermal resistance model,
respectively. This model accurately predicts the temperatures With the interpolation, the governing energy equations in the
and the pressure drop characteristics. Thus, it could be used in simplified 2D two-layer model can be unified as one equation for
topology optimization of forced convection heat sinks with accept- each layer:
able accuracy and greatly reduced computational cost.
The top layer:
3. Topology optimization   hðcÞðT bt  T top Þ
qfl cfl u  rT top ¼ r  kðcÞrT top þ ð14Þ
dztop
3.1. Modified governing equations
The bottom layer:
The working principle of topology optimization of cases includ- €bt
q hðcÞðT bt  T top Þ
ing fluid flow and heat transfer is to find a pattern of solid material r  ðks rT bt Þ þ  ¼0 ð15Þ
dzbt dzbt
distribution which regulates flow conditions, such that certain
objectives can be achieved. As explained in the introduction, a At the top layer, when c ¼ 0, a takes a large value to ensure 0
design variable c, which can vary between 0 and 1 is assigned to local velocity through the friction force term in the modified N-S
each element of the discretized design domain. The value of the equation (13). Thus, the term on the left hand side of the energy
design variable represents the material phase of the local element. equation will also be reduced to 0. In the meantime, the
 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679 669

conductivity and convection heat transfer coefficient both take the For optimization objectives, both pressure drop and heat trans-
values of the solid phase when c ¼ 0. Then Eq. (14) is reduced to fer performances are taken into consideration. Some previous work
Eq. (4), the governing equation for the top layer finned domain. [31,44] combined both into a single objective function. However,
On the other hand, if c ¼ 1, a ¼ 0, additional friction force term magnitudes of the two objective values may greatly differ from
in the modified N-S equation will vanish and the conductivity each other. The introduction of weighting ratios also increases
and heat transfer coefficient will take the values of the fluid phase, the complexity of topology optimization. As such, in this work,
so that Eq. (14) is reduced to Eq. (3), the governing equation for the minimizing pressure drop is set as the sole objective while heat
top layer fluid domain. Similar for bottom layer Eq. (15), the con- transfer performance is set as a constraint.
vection heat transfer coefficient will take the corresponding values There are two ways to impose the heat transfer performance
for solid and fluid phases automatically according to the interpola- constraint. The first one is to define a constant heat flux on the bot-
tion rules. tom layer, and set an upper bound of the bottom layer tempera-
ture. The second method is to define a constant bottom layer
3.2. Interpolation function temperature, and set a lower bound of the amount of heat
absorbed by air based on application requirement. This assumes
The introduced inverse permeability a, conductivity and con- that the temperature is uniform throughout the bottom layer,
vection heat transfer coefficient are all dependent on the local which is reasonable according to the 3D simulation results of the
design variable c through interpolation functions. Rational SC heat sink in Table 1 where the difference between T max and
Approximation of Material Properties (RAMP)-style function [18] T av e are within 1  C under the entire flow rates range considered.
is applied in this study: The second method is found to be easier for implementation and
used throughout this study. Bottom layer temperature T bt is set
1c as 314 K and minimum absorbed heat is set as 2 W, which corre-
a ¼ amin þ ðamax  amin Þ ð16Þ
1 þ qf c sponds to a heat flux of 1666.7 W=m2 .
In summary, the formulation of topology optimization is as
 
c kfl ð1 þ qk Þ  ks þ ks follows:
kðcÞ ¼ ð17Þ
1 þ qk c
Minimize: DP
  Subject to: 0 6 cðrÞ 6 1
c hfl ð1 þ qh Þ  hs þ hs R
hðcÞ ¼ ð18Þ 
Volume constraint: X cðrÞdX 6 V
1 þ qh c R
Minimum heat absorbed: X hðcÞðT bt  T top ÞdX P q_
where amin is the minimum inverse permeability, which is set as 0 Continuity Equation (1)
throughout this work. amax is the maximum inverse permeability, Modified N-S Equation (6)
and qf , qk and qh are parameters that controls the convexity of the Energy Equation (14) & (15)
interpolation function, termed as convexity parameters.
 is the
where DP is the pressure, r is the local coordinates, and V
3.3. Implementation of topology optimization volume upper bound.

3.3.1. Problem setup

The considered optimization problem is to design an air-cooling 3.3.2. Density filter
heat sink. The overall dimensions and boundary conditions are To avoid checker board problem and govern the minimum
shown in Fig. 5(a): the heat sink base is 120 mm long, 100 mm structure size, a Helmholtz-type PDE density filter is applied
wide, and 17 mm high; the finned section is 108 mm long and 8 [40,44]:
mm high, which is also the fin height; the inlet velocity is
1:149 m=s; the inlet temperature is 20  C; the bottom heat flux is r 2 r2 c ~¼c
~þc ð19Þ
1666.7 W=m2 . To further reduce the computational cost, only a
10 mm wide domain is considered for optimization with symmetry where r controls the size of filtering, c is the original design vari-
boundary condition on both sides. The reduced computational ~ is the filtered design variable. When density filter is
able, and c
domain is shown in Fig. 5(b). ~ will replace c in all the governing equations.
turned on, c

Fig. 7. (a) Design variable field after 1st stage optimization; (b) extracted structure from 1st stage optimization.
670 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679

Table 2
Performance of design variable fields and extracted structures in different stages.

Design vin (m/s) Tave ( C) q_ (W) DP (Pa)

Straight fin 1.149 47.7 2 2.84
Design variable field after 1st stage optimization 1.149 40.9 2 8.88
Extracted structure from 1st stage optimization 1.149 48.3 2 3.78
Design variable field after 2nd stage optimization 1.149 40.9 2 5.47
Extracted structure from 2nd stage optimization 1.149 45.6 2 4.82
Design variable field after 3rd stage optimization 1.149 40.9 2 11.91
Extracted structure from 3rd stage optimization 1.149 42.9 2 7.10
Design variable field after 4th stage optimization 1.149 40.9 2 13.84
Extracted structure from 4th stage optimization 1.149 42.2 2 7.82

isfying the constraints or lowering the objective function. Then the

design variable field is updated again based on sensitivity analysis.
This process is repeated iteratively. Finally, a design variable field
is generated with all the constraints satisfied and the objective
function minimized.
The final result is controlled to converge to a discrete 0/1
(solid/fluid) field by finely tuning the convexity parameters qf , qk ,
qh . We apply the same continuation strategy as in [29] except that
qk and qh are first increased, then decreased. Because the design
domain is treated as porous media, non-physical results may be
generated from the optimization process. For example, a whole
band of 0 design variable would block the flow channel and
the fluid is forced to penetrate through the ‘‘solid area”. This struc-
ture is not physically possible while viewed as a high performance
structure in the optimization algorithm. This blockage problem
was explicitly reported and discussed in [47]. There were also
suspicious blocking structures in results of [30,32], where channel
dead ends and layers of solid material before the outlet portion are
observed. To mitigate this problem, qh and qk are first gradually
increased to 500; once the blockage problem shows up, they are
decreased to alleviate the penalization on intermediate material
and the blockage structure disappears.
For density filter, filtering size r is first set as 0.2 mm, then
decreased to 0.1 mm and once the basic topology is formed, the
density filter is turned off to avoid bands of intermediate material
between 0 and 1 regions.
Fig. 8. Multi-stage optimization process. Finally, the design variable field is optimized as in Fig. 7(a). It is
noticed that there is still a small band of intermediate material
between 0 and 1 regions which could not be eliminated with fur-
ther iteration. Solid fins are then extracted by a level curve of
As shown in Fig. 6, after filtering, the original clearly separated 0
c ¼ 0:5 and smoothed as shown in Fig. 7(b). Then the extracted
and 1 design variable field is blurred and bands of intermediate
structure is studied by the simplified 2D two-layer model to eval-
material appear. However, intermediate material has to disappear
uate its real performance. A heating power of 2 W is applied to the
in the final optimized result since it could not be interpreted as
bottom layer and governing energy equations for both layers are
either solid or liquid phase. Thus, the density filter is turned off
solved. The results in Table 2 shows that under the same heating
once basic topology is formed.
power, the bottom layer temperature has increase from 40.9  C
for the design variable field to 48.3  C for the extracted structure.
3.3.3. Multi-stage optimization process The deterioration of performance is mainly due to the manual
The geometry, design domain, boundary conditions, objectives elimination of intermediate material from the design variable field.
and constraint have been defined in previous sections. In this sec- To improve the performance of the extracted structure, a multi-
tion, we will discuss the following process of implementing topol- stage optimization strategy is applied as shown in Fig. 8.
ogy optimization. The fluid domain of the extracted structure from the first stage
The simulation domain is discretized by triangular elements. optimization is set as design domain while the extracted fin struc-
Then modified governing equations are first solved to find the flow tures remain solid. Following the same topology optimization for-
and temperature field based on the initial design variable value, mulation and procedure except no volume constraint and density
which is set as 0.5. The objective function and constraints are then filter, the design domain is optimized as Fig. 9(a). Then the new
evaluated and their sensitivities with respect to the design variable fin structure is extracted by a level curve of v ¼ 0:25 m=s. As can
is calculated [18], based on which the design variable filed is be seen, additional solid fins are formed to enhance the heat trans-
updated by the globally convergent version of Method of Moving fer performance as the average bottom layer temperature
Asymptotes (GCMMA) [46]. With the adjusted design variable decreased from 48.3  C of the 1st optimization result to 45.6  C
field, modified governing equations are first solved again to find of the new structure. The pressure drop has also slightly increased
the new flow and temperature fields, which should be closer to sat- from 3.78 Pa to 4.82 Pa.
 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679 671

Fig. 9. (a) Design variable field after 2nd stage optimization; (b) extracted structure from 2nd stage optimization; (c) design variable field after 3rd stage optimization; (d)
extracted structure from 3rd stage optimization; (e) design variable field after 4th stage optimization; (f) extracted structure from 4th stage optimization; (g) design variable
field after 5th stage optimization.

However, the heat sink temperature is still 5  C higher than the from the 4th stage optimization is taken as the final design, and
set value of 40.85  C (314 K), so the obtained 2nd stage structure is its velocity and temperature fields are shown in Fig. 10.
subjected to the 3rd stage optimization which generates a design
variable field as in Fig. 9(c). It is then extracted by a level curve
of v ¼ 0:2 m=s, giving an average temperature of 42.9  C. Another 4. Numerical and experimental study of the TO heat sink
stage of optimization is performed and the design is extracted by
a level curve of v ¼ 0:2 m=s as shown in Fig. 9(f). It should be noted 4.1. 3D numerical study
that all the extracted structures are slightly modified to make sure
the smallest gap is bigger than 1 mm for manufacturability. (This Since the pressure drop prediction of the 2D two-layer model is
limitation is according to equipment available in this study and not accurate, 3D simulation is necessary to verify the performance
the size could definitely be smaller with better manufacturing of the optimized structure. The simulation domain is shown in
equipment and techniques.) Fig. 11. As defined in the optimization problem setup, the fin
The multi-stage optimization process is terminated if the height is 8 mm and heat sink base thickness is 17 mm. The inlet
improvement of Tave of extracted structures is less than 1  C and portion is extended by 40 mm to get a developing velocity profile
no new fin pieces would be generated with further optimization. in the direction of the channel height as in the experimental set
Tave of the extracted structure from 4th stage is 0.7  C lower than up. The outlet portion is also extended by 40 mm to avoid reverse
the 3rd stage one, and no fin pieces are added in the 5th optimiza- flow problem. The numerical analysis is carried out in ANSYS Flu-
tion process as shown in Fig. 9(g). Thus, the extracted structure ent 18 with the assumption of steady state, incompressible flow.
672 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679

Fig. 10. (a) Temperature field and (b) velocity field of the final design.

Fig. 11. 3D model of final design.

Table 3 k ¼ 168 W=m  C, q ¼ 2700 kg=m3 and cp ¼ 896 J=kg  C [48,49],

Mesh dependence study.
respectively. For air, its properties are treated as temperature
v in ðm=sÞ € (W=m2 )
q DP (Pa) T min ð CÞ T max ð CÞ Mesh elements dependent with a 3rd degree polynomial functions [48].
5.744 1666.7 127.14 26.16 26.83 6383356
5.744 1666.7 127.18 26.13 26.83 10213136
q ¼ 3:2894  0:010286T 2 þ 0:000010653T 3
5.744 1666.7 127.05 26.18 26.75 16489807

k ¼ 0:0007626595 þ 0:0000965192T 2  0:0000000371T 3

Due to the complexity of the flow channels, a turbulence model, cp ¼ 1033:78  0:217226807T 2 þ 0:000421911T 3
RNG  ke, with enhanced wall treatment is used.
The boundary conditions are the same as that in the optimiza-
tion problem setup: uniform velocity inlet, zero pressure outlet,
l ¼ 0:0000012568043  0:00000006698901T 2
and 20  C inlet temperature. Except for the case of constant bottom  0:0000000000313T 3
surface heat flux of 1666.7 W=m2 , an extra case of 3333.3 W=m2 is
also studied. Both side surfaces are set as symmetry and others are where T is in unit of Kelvin.
set as adiabatic walls. In the solver phase, SIMPLEC is selected as the pressure-
The heat sink material is aluminum alloy 6061-T6. Its velocity coupling algorithm with the second order upwind as the
thermal conductivity, density and specific heat values are discretization scheme. A residual value of 106 is set as the
 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679 673

sink: Omega PX653-0.25D5V is used to measure pressure drops

below 70 Pa and Unik 5000 is used to measure higher pressure
drops. Finally, a 1 m long circular duct is designed to ensure a fully
developed velocity profile at the outlet. A probe type velocimeter
(Velocicalc 9565) is placed at the center of the outlet to measure
the maximum velocity.

4.2.2. Data acquisition and reduction

The data from the 8 thermocouples inserted into the heat
sink base, 3 thermocouples before the test section and the pres-
sure transducers are taken by a National Instrument cDAQ 9178
USB chassis and displayed in a LABView program. If all the tem-
Fig. 12. Manufactured (a) TO heat sink and (b) SC heat sink. perature variations are within 0:1  C in 5 min, the experiments
are said to have reached steady state. Data in the last 5 min are
recorded and averaged to obtain the final set of experimental
convergence criteria for continuity, x-y-z velocities, k, e, x and data points.
1010 for energy. Though the test section is covered by an insulation layer of low
Mesh dependence study is conducted under the highest velocity thermal conductivity foam, the heat loss to the environment is still
considered, 5.744 m/s. As shown in Table 3, the difference of pres- not negligible. As the environment temperature is constant
sure drop, minimum and maximum heat sink base temperatures (around 20  C), the heat loss can be assumed to be a function of
are all smaller than 1% among the three meshes; therefore the sec- heat sink temperature. To retrieve this function, the test section
ond mesh, comprising approximately 10 million mesh elements, is is heated under various powers, and steady state heat sink temper-
selected for further case studies under different flow rates. atures are measured without fluid flow. The relationship between
the heating power and the average heat sink temperature T HS;av e
4.2. Experiment validation

The TO heat sink is manufactured by CNC machining for exper-

imental test as shown in Fig. 12a. Fig. 12b shows the manufactured
SC heat sink with dimensions stated in Section 2.2.

4.2.1. Experimental setup

A desktop wind tunnel experiment setup, similar to the one
used in [41], is utilized in this study, and a schematic is shown in
Fig. 13. Air is driven by a centrifugal fan (EBM-Papst rg 148) with
variable speed control. Before the test section, there are 3 T-type
thermocouples measuring the inlet air temperature. A plate heater
is attached to the bottom of the test piece through a thin layer of
high thermal conductivity interface material, then the whole test
section is fully covered by a Teflon case. To reduce heat loss to
the environment, the Teflon case is wrapped by another layer of
low thermal conductivity foam. 8 holes are drilled half way into
the heat sink base for inserting T-type thermocouples to measure
the heat sink base temperature at the middle height of the heat
sink base. The flat heater is powered by the Sorensen XTR-150-
5.6 power supply module, with voltage and current display and
adjustment function. Two differential pressure transducers are Fig. 14. The relationship between the heat loss and the average heat sink
used to measure the pressure drop of the flow through the heat temperature.

Fig. 13. Schematic of the experimental setup with a side view of the test section.
674 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679

Fig. 15. Comparison of experimental and numerical (a) average heat sink base temperature and (b) pressure drop.

is presented in Fig. 14, where T HS;av e is the average of the readings where T air;mean is the arithmetic average of T air;out and T air;in . Here, air
of the 8 thermocouples inside the heat sink base. density in Eq. (26) is first taken as the density at inlet, then air mass
The data is fitted into a linear equation. flow rate is calculated. From Eq. (27), by taking cp at the inlet tem-
perature, T air;out is calculated. Then air density in Eq. (26) takes the
q_ loss ¼ 0:2413T HS;av e  5:08 ð20Þ
value at the newly calculated T air;out to calculate a new m _ air . Simi-
Then, the power absorbed by air flow is calculated as larly, cp in Eq. (27) takes the corresponding value at T air;mean calcu-
q_ air ¼ V heater Iheater  ð0:2413T HS;av e  5:08Þ ð21Þ lated from previous T air;out to get a new T air;out . This iteration
process is repeated until the difference of T air;out between two iter-
The heater voltage V heater and current Iheater are read from its ations are less than 1%.
power supply display and manually input into a LABView program Then the inlet velocity is calculated as
to calculate real time q_ air with the measured T HS;av e . Then, the
m _ air
heater voltage is finely tuned to make q_ air as the set heating power v in ¼ ð28Þ
(20 W or 40 W) with a variation less than ±0.1 W. q@T air;in Aair;in
The average junction temperature and average wetted wall
The whole heat sink thermal resistance is defined as
surface temperature are calculated from T HS;av e assuming 1D heat
conduction within the base of the heat sink from the junction to T junction;av e  T air;in
Rth ¼ ð29Þ
the wetted walls: q_ air
q_ air HHS Finally, pumping power is calculated as
T junction;av e ¼ T HS;av e þ ð22Þ
2kHS AHS _ air DP
m
Q pump ¼ ð30Þ
q_ air HHS
q@T air;mean
T wall;av e ¼ T HS;av e  ð23Þ
2kHS AHS where DP ¼ Pin  Pout .
As the long circular duct ensures a fully developed velocity pro- Due to the irregular geometry of the TO heat sink, it is not suit-
file, the velocity profile at the outlet is given as [50] able to calculate hydraulic diameter as
 r 1=n 4Asec
v r ¼ v max 1 ð24Þ Dh ¼
wetted perimeter
ð31Þ
R
The average velocity is calculated by integrating the velocity Instead, an expression used in [51] is adapted:
profile over the cross sectional area:
4fluid v olume
Dh ¼ ð32Þ
2n2 wetted surface area
v av e ¼ v max ð25Þ
ðn þ 1Þð2n þ 1Þ
Thus, the inlet Reynolds number is defined as
where n was taken as 7 as suggested by [50]. The mass flow rate is
q@T air;in v in Dh
then calculated as Re ¼ ð33Þ
l@T air;in
_ air ¼ q@T
m air;out
v av e Aair;out ð26Þ
The average heat transfer coefficient is
Due to the fin structures, the air flow after the heat sink is quite
non-uniform, which makes the direct measurement of air outlet q_ air
hav e ¼ ð34Þ
temperature inaccurate. Thus, the air density evaluated at the mea- Awet ðT wall;av e  T air;mean Þ
sured outlet temperature would also be inaccurate, resulting in an
The average Nusselt number is then calculated as
inaccurate air flow rate evaluation in Eq. (26). To solve this prob-
lem, an iterative calculation is required and energy balance equa- hav e Dh
tion has to be included: Nu ¼ ð35Þ
k@T air;mean
q_ air ¼ m
_ air cp@T air;mean ðT air;out  T air;in Þ ð27Þ
Finally, the friction factor is calculated as
 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679 675

Fig. 16. Comparison of experimental (a) thermal resistance, (b) pumping power, (c) Nu and (d) friction factor of the SC heat sink and TO heat sink.

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2 2
@f @f @f
df ¼ dx þ dy þ dz þ    ð37Þ
@x @y @z

where f is the derived parameter, x, y, z, . . . are the measured

parameters, and d means the absolute error. The uncertainties of
the sensors could be found in [41].

5. Results and discussion

5.1. Numerical results validation

Numerical simulations and experiments are conducted under

the same range of mass flow rates and heating powers. As shown
in Fig. 15, the average heat sink temperatures from numerical sim-
ulations match the experimental results reasonably well. For the
case of 40 W heating power, the deviation of the numerical and
experimental pressure drops is less than 7%. Though a larger devi-
ation of 15% is seen at high flow rates under 20 W heating power,
Fig. 17. Comparison of average junction temperature under various pumping the trend of increasing pressure drop with increasing mass flow
powers.
rate is still captured. Thus, the numerical results are trustworthy
for detailed analysis of the flow and heat transfer characteristics
of the TO heat sink.
2DPDh
f ¼ ð36Þ
q@T air;mean v 2in LHS 5.2. Heat transfer and pressure drop performances comparison

A standard uncertainty analysis [52] is performed. The error for The SC heat sink is used as a benchmark case for comparing the
a derived parameter is defined as performance of the TO heat sink. Fig. 16 shows their performance
676 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679

Fig. 18. Velocity fields of (a) the SC heat sink and (b) TO heat sink.

Fig. 19. Temperature fields of (a) the SC heat sink and (b) TO heat sink.

130 considering both pressure drop and heat transfer, a chart of aver-
age junction temperature versus pumping power is drawn.
120 SC
As shown in Fig. 17, the TO heat sink can achieve lower junction
TO
110 temperatures with the same pumping power or requires less
pumping power to get the same junction temperature. Quantita-
100
tively speaking, for 40 W heating power, the junction temperature
90 6th piece of TO heat sink is 5.5  C lower than the SC heat sink under the same
of fin
80 pumping power of 0.065 W. Alternatively, to maintain the same
junction temperature of 44.5  C, the SC heat sink requires a pump-
70 ing power of 0.144 W while the TO heat sink only requires 0.065
60 W, achieving a pumping power saving of 54.9%.
50
40 5.3. Flow and thermal characteristics of the TO heat sink

30 Through the experimentally-validated numerical simulation

0 20 40 60 80 100 120
results, more details can be revealed about the characteristics of
Streamwise position (mm ) the TO heat sink.
Fig. 20. Local heat transfer coefficient along the streamwise position.
Fig. 18 shows the velocity field of inlet velocity as 1.723 m/s at
the middle height of the channel. The boundary layer thickness of
straight channel gradually increases along the streamwise direc-
comparison in terms of heat sink thermal resistance, pumping tion. Instead, the distributed sectional fins of the TO heat sink
power, Nu and friction factor under 20 W and 40 W heating power. interrupt the growth of the boundary layer continuously, and form
The enhancement in heat transfer is always accompanied by a several secondary flow channels, e.g. the channel between 7th and
penalty in pumping power. To illustrate the overall performance 8th pieces of fins, introducing strong mixing effect. In addition,
 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679 677

Fig. 21. Smallest periodically repeated portions of the (a) OSF and (b) SPF heat sink.

65 with downstream fins, which helps to enhance heat transfer. It is

TO HS Exp Q = 20 W also noticed that the temperatures of the fins are almost uniform,
60 around 36  C for the TO heat sink and 43  C for the SC heat sink.
SC HS Exp Q = 20 W Fig. 20 shows the local heat transfer coefficient along the
55
streamwise direction. The computational domain is axially divided
OSF HS Num Q = 20 W
50 into 11 sections and the heat transfer coefficient is computed as a
SPF Num Q = 20 W local average based on the average wetted surface temperatures,
45 average air temperatures and total heat absorbed in each section.
Initially, heat transfer is strong in the SC heat sink. Then it
40
decreases rapidly as the boundary layers develop and remains
35 low in the downstream region. Similar rapid decrease is seen in
the TO heat sink in the upstream portion, but it increases up at
30 around 40 mm due to the introduction of the 6–8th pieces of fins
and maintained higher than the SC heat sink’s value along the
25
whole downstream section. As shown in the enlarged view of
20 streamlines in Fig. 18(b), 6–8th pieces of fins interrupt the straight-
0.001 0.01 0.1 1 ened flow from upstream and branch air into small channels,
which contributes to the enhancement of heat transfer in this local
section. A similar phenomenon is observed in the last section of the
Fig. 22. Comparison of average junction temperatures under various pumping finned area, where the center piece of fin branches air into two
powers for various fin configurations.
channels. The effective cooling in the downstream part is the
key factor that makes the TO heat sink superior than the SC heat
sink.
several local high speed points are seen in the small gaps between
fins, which contribute to the increase of local heat transfer
coefficients. 5.4. Performance comparison with other fin configurations
Fig. 19 shows the temperature field at the middle height of the
channel. In the SC heat sink, thermal boundary layers grow gradu- Except for benchmark with the basic SC heat sink, two more
ally on wall surfaces before merging at the end of channels. While commonly seen fin configurations, namely offset strip fin (OSF)
in the TO heat sink, this pattern of boundary layer growth is and staggered pin fin (SPF), are studied for comparison. The overall
avoided by the non-conventional distribution of fins. Moreover, dimensions of the whole heat sink are the same as the topology
strong mixing effect decreases air temperature that is in contact optimized one, and their detailed geometries are shown in Fig. 21.

Table 4
Results comparison of the 2D and 3D model for the TO heat sink.

vin (m/s) T max ( C) T av e ( C) DP (Pa)

2D 3D 2D 3D 2D 3D
0.574 59.7 59.1 58.9 58.4 2.97 4.72
1.149 42.7 42.2 42.2 41.7 7.82 11.19
1.723 37.1 36.3 36.7 35.9 14.61 19.65
2.298 34.2 33.2 33.8 32.8 22.54 30.00
2.872 32.4 31.2 32.1 30.8 32.15 42.15
678 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679

The results of these two heat sinks are obtained through the 4. 3D conjugate numerical simulation is performed to further
experimentally verified 3D numerical simulation method. As can study the characteristics of the TO heat sink. The simulation
be seen in Fig. 22 of pumping power-junction temperature below, results are validated experimentally. The flow and temperature
the TO heat sink is still slightly better than the OSF and SPF heat fields show the optimized fin structure could interrupt the
sinks. development of boundary layers. The local heat transfer coeffi-
cient at the downstream part of the TO heat sink also is higher
5.5. Further validation of the 2D two-layer model compared to that of the SC heat sink.
5. Topology optimization is a powerful tool because of its highest
Previously, the accuracy of the 2D two-layer model is verified level of design freedom. However, the treatment of porous
with 3D simulations of the SC heat sink. With the TO heat sink, media may lead to unphysical structures. The decision of con-
comprising non-conventional structures, the accuracy of the sim- vexity parameters is the key to resolve the difficulties encoun-
plified model is studied, as well. For hfl and hs , we still use 27 tered in implementing topology optimization, which calls for
and 9882 W=m2  K as previous values. further research.
As shown in Table 4, the maximum discrepancy between 2D
and 3D results of T max and T av e is 1.2  C and 1.3  C, respectively. Conflict of interest
Pressure drop is lower than 3D simulation values but follows the
same trend of increase with increasing inlet velocities. The authors of this submission have no conflicts of interest to
The convection heat transfer coefficient between air and heat declare.
sink hfl varies with various inlet velocities and structures. For
example, at 1.723 m/s in the TO heat sink, hfl is 53 W=m2  K,
References
around 2 times of 27 W=m2  K as set in the 2D model. However,
when compared to hs ; it is still a rather small value, which causes [1] International Technology Roadmap for Semiconductors 2.0, 2015.
most of the heat to be transferred to the fins then absorbed by air [2] Intel Ò Xeon Ò Processor E5 v4 Product Family Thermal Mechanical
and small amount of heat is directly transferred to air through the Specification and Design Guide, 2016.
[3] D.B. Tuckerman, R.F.W. Pease, High-performance heat sinking for VLSI, IEEE
channel bottom surface. In the TO heat sink with inlet velocity as Electron Device Lett. 2 (1981) 126–129, https://doi.org/10.1109/
1.723 m/s, 86.5% of heat is transferred through fins and 13.5% is EDL.1981.25367.
transferred through channel bottom surface. The convection heat [4] M.E. Steinke, S.G. Kandlikar, Single-phase heat transfer enhancement
techniques in microchannel and minichannel flows, in: ASME 2nd Int. Conf.
transfer between fins and air is the dominant mechanism and it Microchannels Minichannels, 2004, pp. 141–148. http://doi.org/10.1115/
is numerically computed in the 2D model, which ensures the tem- ICMM2004-2328.
perature prediction accuracy of the 2D model. [5] Y. Sui, C.J. Teo, P.S. Lee, Y.T. Chew, C. Shu, Fluid flow and heat transfer in wavy
microchannels, Int. J. Heat Mass Transf. 53 (2010) 2760–2772, https://doi.org/
10.1016/j.ijheatmasstransfer.2010.02.022.
[6] Z. Lin, P. Seng, P. Kumar, N. Mou, Investigation of fluid flow and heat transfer in
6. Conclusions wavy micro-channels with alternating secondary branches, Int. J. Heat Mass
Transf. 101 (2016) 1316–1330, https://doi.org/10.1016/j.ijheatmasstransfer.
This work utilizes topology optimization to design a forced air 2016.05.097.
[7] Y.J. Lee, P.S. Lee, S.K. Chou, Enhanced thermal transport in microchannel using
heat sink. A 2D two-layer model is developed to reduce the oblique fins, J. Heat Transf. 134 (2012) 101901, https://doi.org/10.1115/
computational load of a full scale 3D conjugate heat transfer 1.4006843.
numerical simulation model during the optimization process. A [8] A. Bejan, M.R. Errera, Convective trees of fluid channels for volumetric cooling,
Int. J. Heat Mass Transf. 43 (2000) 3105–3118, https://doi.org/10.1016/S0017-
non-conventional fin structure is generated through topology 9310(99)00353-1.
optimization, and its high performance is validated in full 3D scale [9] D. Pence, Reduced pumping power and wall temperature in microchannel heat
numerical simulations and experimental tests. The following sinks with fractal-like branching channel networks, Microscale Thermophys.
Eng. 6 (2003) 319–330, https://doi.org/10.1080/10893950290098359.
conclusions could be drawn from this work:
[10] Y. Chen, P. Cheng, Heat transfer and pressure drop in fractal tree-like
microchannel nets, Int. J. Heat Mass Transf. 45 (2002) 2643–2648, https://
1. The developed 2D two-layer model considers the heat sink base doi.org/10.1016/S0017-9310(02)00013-3.
[11] C.S. Sharma, G. Schlottig, T. Brunschwiler, M.K. Tiwari, B. Michel, D. Poulikakos,
and fin channel part as two layers, respectively, and combines
A novel method of energy efficient hotspot-targeted embedded liquid cooling
them with a convection heat transfer condition. Heat transfer for electronics: an experimental study, Int. J. Heat Mass Transf. 88 (2015) 684–
coefficients for the fin region and the unfinned area are different 694, https://doi.org/10.1016/j.ijheatmasstransfer.2015.04.047.
and are determined from the thermal resistance model and 3D [12] P.R. Parida, S.V. Ekkad, K. Ngo, Experimental and numerical investigation of
confined oblique impingement configurations for high heat flux applications,
conjugate heat transfer simulation, respectively. This model is Int. J. Therm. Sci. 50 (2011) 1037–1050, https://doi.org/10.1016/j.ijthermalsci.
able to predict the heat sink temperature accurately with an 2011.01.010.
error smaller than 1.3  C for both the SC and TO heat sink under [13] C. Green, A.G. Fedorov, Y.K. Joshi, Fluid-to-fluid spot-to-spreader (F2/S2)
hybrid heat sink for integrated chip-level and hot spot-level thermal
different inlet velocities. Though its pressure drop prediction is management, J. Electron. Packag. 131 (2009) 25002, https://doi.org/10.1115/
always smaller than the results from 3D simulations, the sim- 1.3104029.
plified model also get higher pressure drops for higher inlet [14] J.F. Maddox, R.W. Knight, S.H. Bhavnani, Liquid Jet impingement with an
angled confining wall for spent flow management for power electronics
velocities. cooling with local thermal measurements, J. Electron. Packag. 137 (2015)
2. A multi-stage optimization strategy is used to improve the per- 31015, https://doi.org/10.1115/1.4030953.
formance of the extracted structures from topology optimiza- [15] J. Barrau, M. Omri, D. Chemisana, J. Rosell, M. Ibañez, L. Tadrist, Numerical
study of a hybrid jet impingement/micro-channel cooling scheme, Appl.
tion step by step. New fin structures are added to the design
Therm. Eng. 33–34 (2012) 237–245, https://doi.org/10.1016/j.
domain to enhance heat transfer at each stage and finally a applthermaleng.2011.10.001.
non-conventional structure is obtained. [16] F. Zhou, Y. Liu, Y. Liu, S.N. Joshi, E.M. Dede, Modular design for a single-phase
manifold mini/microchannel cold plate, J. Therm. Sci. Eng. Appl. 8 (2015)
3. The optimized structure is manufactured and experimental
21010, https://doi.org/10.1115/1.4031932.
tests are performed. Compared to a conventional SC heat sink, [17] R. Hilbert, G. Janiga, R. Baron, D. Thévenin, Multi-objective shape optimization
it shows better heat transfer ability with an acceptable pressure of a heat exchanger using parallel genetic algorithms, Int. J. Heat Mass
drop penalty. Quantitatively speaking, the TO heat sink could Transf. 49 (2006) 2567–2577, https://doi.org/10.1016/j.ijheatmasstransfer.
2005.12.015.
save 54.9% pumping power while maintaining the heat sink [18] J. Alexandersen, Topology Optimisation for Coupled Convection Problems
temperature at 44.5  C with a heating power of 40 W. (Master Thesis), Technical University of Denmark, Lyngby, 2013.
 S. Zeng et al. / International Journal of Heat and Mass Transfer 121 (2018) 663–679 679

[19] M.P. Bendsøe, N. Kikuchi, Generating optimal topologies in structural design [36] T.E. Bruns, Topology optimization of convection-dominated, steady-state heat
using a homogenization method, Comput. Methods Appl. Mech. Eng. 71 (1988) transfer problems, Int. J. Heat Mass Transf. 50 (2007) 2859–2873, https://doi.
197–224, https://doi.org/10.1016/0045-7825(88)90086-2. org/10.1016/j.ijheatmasstransfer.2007.01.039.
[20] M.P. Bendsøe, O. Sigmund, Topology Optimization-Theory, Methods and [37] A. Iga, S. Nishiwaki, K. Izui, M. Yoshimura, Topology optimization for thermal
Applications, Springer, Berlin, 2003. conductors considering design-dependent effects, including heat conduction
[21] T. Borrvall, J. Petersson, Topology optimization of fuids in Stokes flow, Int. J. and convection, Int. J. Heat Mass Transf. 52 (2009) 2721–2732, https://doi.org/
Numer Methods Fluids 107 (2003) 77–107. 10.1016/j.ijheatmasstransfer.2008.12.013.
[22] A. Gersborg-Hansen, O. Sigmund, R. Haber, Topology optimization of channel [38] M. Zhou, J. Alexandersen, O. Sigmund, C.B.W. Pedersen, Industrial application
flow problems, Struct. Multidiscip. Optim. 30 (3) (2005) 181–192. of topology optimization for combined conductive and convective heat
[23] L.H. Olesen, F. Okkels, H. Bruus, A high-level programming-language transfer problems, Struct. Multidiscip. Optim. (2016), https://doi.org/
implementation of topology optimization applied to steady-state Navier- 10.1007/s00158-016-1433-2.
Stokes flow, Int. J. Numer. Meth. Eng. 65 (2006) 975–1001. [39] T. Matsumori, T. Kondoh, A. Kawamoto, T. Nomura, Topology optimization for
[24] C. Othmer, T. Kaminski, R. Giering, Computation of topological sensitivities in fluid – thermal interaction problems under constant input power, Struct.
fluid dynamics: cost function versatility, in: Proceedings of ECCOMAS CFD Multidiscip. Optim. 47 (4) (2013) 571–581, https://doi.org/10.1007/s00158-
2006, 2006. 013-0887-8.
[25] N. Aage, T.H. Poulsen, A. Gersborg-Hansen, O. Sigmund, Topology optimization [40] J.H.K. Haertel, K. Engelbrecht, B.S. Lazarov, O. Sigmund, Topology optimization
of large scale stokes flow problems, Struct. Multidiscip. Optim. 35 (2008) 175– of thermal heat sinks, in: Proceedings of COMSOL Conference 2015, 2015.
180. [41] O.B. Kanargi, P.S. Lee, C. Yap, A numerical and experimental investigation of
[26] C.S. Andreasen, A.R. Gersborg, O. Sigmund, Topology optimization of heat transfer and fluid flow characteristics of a cross-connected alternating
microfluidic mixers, Int. J. Numer. Meth. Fluids 61 (5) (2009) 498–513. converging—diverging channel heat sink, Int. J. Heat Mass Transf. 106 (2017)
[27] A. Gersborg-Hansen, M.P. Bendsøe, O. Sigmund, Topology optimization of heat 449–464, https://doi.org/10.1016/j.ijheatmasstransfer.2016.08.057.
conduction problems using the finite volume method, Struct. Multidiscip. [42] H.C. Kang, Evaluation of fin efficiency and heat transfer coefficient for finned
Optim. 31 (2006) 251–259. tube heat exchanger, in: Int. Refrig. Air Cond. Conf., 2012, pp. 1–5.
[28] S. Soprani, J.H.K. Haertel, B.S. Lazarov, O. Sigmund, K. Engelbrecht, A design [43] N. Mou, P.S. Lee, S.A. Khan, Coupled equivalent circuit models for fluid flow
approach for integrating thermoelectric devices using topology optimization, and heat transfer in large connected microchannel networks – the case of
Appl. Energy 176 (2016) 49–64, https://doi.org/10.1016/j.apenergy. oblique fin heat exchangers, Int. J. Heat Mass Transf. 102 (2016) 1056–1072,
2016.05.024. https://doi.org/10.1016/j.ijheatmasstransfer.2016.06.069.
[29] J. Alexandersen, O. Sigmund, N. Aage, Large scale three-dimensional topology [44] X. Qian, E.M. Dede, Topology optimization of a coupled thermal-fluid system
optimisation of heat sinks cooled by natural convection, Int. J. Heat Mass under a tangential thermal gradient constraint, Struct. Multidiscip. Optim.
Transf. 100 (2016) 876–891, https://doi.org/10.1016/j.ijheatmasstransfer. (2016), https://doi.org/10.1007/s00158-016-1421-6.
2016.05.013. [45] G.H. Yoon, Topological design of heat dissipating structure with forced
[30] T.V. Oevelen, M. Baelmans, Numerical topology optimization of heat sinks, in: convective heat transfer, J. Mech. Sci. Technol. 24 (6) (2010) 1225–1233,
Proceedings of IHTC-15, 2014. https://doi.org/10.1007/s12206-010-0328-1.
[31] A.A. Koga, E. Comini, C. Lopes, H.F. Villa, C.R. De Lima, E. Carlos, N. Silva, [46] K. Svanberg, MMA and GCMMA – Fortran Versions March 2013, Royal Institute
Development of heat sink device by using topology optimization, Int. J. Heat of Technology, 2013, pp. 1–23.
Mass Transf. 64 (2013) 759–772, https://doi.org/10.1016/j.ijheatmasstransfer. [47] K. Lee, Topology Optimization of Convective Cooling System Designs (PhD
2013.05.007. Thesis), University of Michigan, MI, 2012.
[32] E.M. Dede, Optimization and design of a multipass branching microchannel [48] VDI-Gesellschaft Verfahrenstechnik und Chemieingenieurwesen (Ed.), VDI
heat sink for electronics cooling 041001-1 041001-10 J. Electron. Packag. 134 Heat Atlas, second ed., Springer, Berlin Heidelberg, 2010.
(2012), https://doi.org/10.1115/1.4007159. [49] J.R. Davis, Aluminum and aluminum alloys, ASM Int. (1993).
[33] E.M. Dede, Y. Liu, Experimental and numerical investigation of a multi-pass [50] Y.A. Cengel, J.M. Cimbala, Fluid Mechanics: Fundamentals and Applications,
branching microchannel heat sink, Appl. Therm. Eng. 55 (2013) 51–60, https:// third ed., McGraw-Hill Education, 2014.
doi.org/10.1016/j.applthermaleng.2013.02.038. [51] C. Li, R.A. Wirtz, Development of a high performance heat sink based on
[34] E. Dede, S.N. Joshi, F. Zhou, Topology optimization, additive layer screen-fin technology, IEEE Trans. Compon. Packag. Technol. 28 (2005) 80–87,
manufacturing, and experimental testing of an air-cooled heat sink, ASME J. https://doi.org/10.1109/TCAPT.2004.843171.
Mech. Des. 137 (2015), https://doi.org/10.1115/1.4030989. [52] J.R. Taylor, An Introduction to Error Analysis: The Study of Uncertainties in
[35] L. Yin, G.K. Ananthasuresh, A novel topology design scheme for the multi- Physical Measurements, second ed., University Science Books, 1997.
physics problems of electro-thermally actuated compliant micromechanisms,
Sensors Actuators A 97–98 (2002) 599–609.