International Journal of Heat and Mass Transfer 67 (2013) 118–130

Contents lists available at ScienceDirect

International Journal of Heat and Mass Transfer

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

An investigation of effects of heights and a length on natural convection

of three dimensional parallel plates with a heated bottom surface by a
CUDA computation platform
Wu-Shung Fu ⇑, Shang-Hao Huang, Wei-Hsiang Wang
Department of Mechanical Engineering, National Chiao Tung University, Hsinchu 30010, Taiwan, ROC

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

Article history: Effects of heights and a length on natural convection of three dimensional parallel plates with a heated
Received 21 February 2013 bottom surface are investigated numerically. The compressibility of fluid induced by a high temperature
Received in revised form 9 July 2013 difference is taken into consideration for simulating more realistic situations, and then the Boussinesq
Accepted 9 July 2013
assumption is not necessary to be adopted. Methods of the Roe scheme, preconditioning and dual time
Available online 29 August 2013
stepping matching the method of LUSGS are simultaneously used to solve the compressible flow problem.
Non-reflecting boundary conditions are held on apertures to prevent reflections from apertures induced
Keywords:
by pressure waves. A CUDA computational platform is used to economize massive computing time. Under
Natural convection
Horizontal parallel plates
a low height case, the flow field is restricted by the shear force caused by the low height and displays a
Heated bottom surface stable situation. Oppositely, under a high height case the flow field indicates active behaviors and an
CUDA irregular situation is observed.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction examined in detail in the above literature, and remarkable results

were achieved.
A subject of natural convection of three dimensional parallel A heated bottom wall in parallel plates with open-ended aper-
plates with a heated bottom surface has attracted lots of attention tures which is another important geometry of this subject is often
in both practical applications and academic studies due to its assumed as the physical model as well, and was investigated by
importance to wide kinds of physical phenomena. In practical numerous studies [5–14]. However, because the existence of the
applications, this subject can be observed in many thermal equip- opened-ended apertures causes fluids to flow into and out off the
ment, such as energy storage equipment, solar collector, liquid channel continuously that results in an unsteady phenomenon eas-
crystal growth chamber. As for in the academic study, except un- ily occurring, but the phenomenon of the Rayleigh–Benard convec-
steady behaviors, another interesting characteristic of the Ray- tion being relatively less found out and discussed. As for the
leigh–Benard convection can also be found out. Studies of a treatment of boundary conditions at the opened-ended apertures,
numerical simulation [1,2] and a model analysis [3,4] were then the enough length for satisfying a fully developed flow is adopted
investigated. An interaction between forces of a viscous force and [15–17], and the enlarged computational domain is usually used
the buoyancy force is recognized to dominate the occurrence of by the far field boundary conditions [18–23]. However, existence
Rayleigh–Benard convection. Also, that boundary conditions have of a pressure difference at the apertures would lead acoustic waves
deep influence on phenomena of the Rayleigh–Benard convection caused by the compressibility of fluid to reflect from the aperture,
is well-known. In the above literature, an enclosure which is one and then the low speed compressible flow in parallel plates will be
of important geometries of this subject was assumed the physical polluted by the reflections of acoustic waves. As well, setting the
model. The Boussinesq assumption is adopted to substitute the enlarged computational domain outside the opened-ended aper-
variation of the density of the fluid. Due to the situation of the tures causes extra computational time and memory to be added.
enclosure completely surrounded by walls, phenomena of the var- A preliminary study [24] of this work adopted the model of parallel
iation of the Rayleigh–Benard convection were very active and plates with non-reflecting boundary conditions at open-ended
apertures to investigate the subject. In this study, in an initial sta-
tus cooling fluids were sucked from surroundings and flowed into
⇑ Corresponding author. Address: Department of Mechanical Engineering, the channel, and a pair of T-shape roll cells which are like the Ray-
National Chiao Tung University, 1001 Ta Hsueh Road, Hsinchu 30056, Taiwan, leigh–Benard convection and a kind of intermediate product were
ROC. Tel.: +886 3 5712121x55110; fax: +886 3 5735065. observed. Finally the flow became unsteady when the bottom wall
E-mail address: [email protected] (W.-S. Fu).

0017-9310/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ijheatmasstransfer.2013.07.030
 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130 119

Nomenclature

A area (m2) Ra⁄ modified Rayleigh number defined in Eq. (27)

d width of the plate (m) Ra ¼ Ra  hl
e internal energy (J/kg) Re Reynolds number
g acceleration of gravity (m/s2) t time (s)
h height of the plate (m) T temperature (K)
k thermal conductivity (W/m K) T0 initial temperature in the physical model (298.05 K)
k0 surrounding thermal conductivity (W/m K) Tc temperature of surroundings (298.05 K)
l length of the plate (m) Th temperature of heat surface (K)
Nux local Nusselt h number
i defined in Eq. (28) u, v, w velocities in x, y and z directions (m/s)
Nux ¼ k0 ðT hhT c Þ kðTÞ @T U, V, W uffi
dimensionless velocities U ¼ pffiffiffiffiffi v ffi ; W ¼ pffiffiffiffiffi
; V ¼ pffiffiffiffiffi wffi
@y Ra ah Ra ah Ra ah
Nux time average local Nusselt number defined in Eq. (29) x, y, z Cartesian coordinates (m)
R X, Y, Z dimensionless Cartesian coordinates X ¼ lx3 , Y ¼ ly3 and
Nux ¼ 1t t Nux dt
NuA area average Z ¼ lz3
1
R R local Nusselt number defined in Eq. (31)
 NuA ¼ dl d l Nux dxdz
Nu average Greek symbols
 R RNusselt number defined in Eq. (32)
1
Nu ¼ dl d l Nux dxdz
a thermal diffusivity rate (m2/s)
P pressure (Pa) q density (kg/m3)
P0 surrounding pressure (Pa) q0 surrounding density (kg/m3)
Pr Prandtl number l viscosity (N s/m2)
R gas constant (J/kg/K) l0 surrounding viscosity (N s/m2)
R⁄ length ratio R ¼ hl c specific heat ratio
g q2 ðT T Þh3
Ra Rayleigh number defined in Eq. (26) Ra ¼ Pr 0 h 2c
T 0 lðTÞ

was completely installed by a heated surface. However, as the 2. Physical model

length of the heated surface was shorter than a half of the bottom
surface length, the T-shape roll cell was no longer observed, and fi- A physical model of three dimensional horizontal parallel plates
nally the flow became a quasi-steady or a steady flow. The reason is indicated in Fig. 1. The length and height of the physical model
is suggested as that the shorter heated bottom surface provides the are l and h, respectively. The bottom plate is a heated wall and
less buoyancy force and it is difficult to compete with the horizon- the temperature of the heated wall is Th and equal to 700 k. The
tal viscous force. As for the influences of different heights and a top plate is adiabatic. The width of the physical model is d, and
longer heated bottom surface on the subject mentioned above the direction of gravity is downward. Boundary conditions at both
are scarcely investigated. apertures are non-reflecting conditions and both sides of the width
Therefore, two different heights which are separately lower and are periodic conditions. The pressure and temperature of the out-
higher than the height of [24] and a longer length which is two side of parallel plates are 101,300 Pa and 298 K, respectively.
times as long as the heated bottom surface length of [24] are taken For facilitating the analysis, the following assumptions are
into consideration to examine the interaction of the viscous and made.
buoyancy forces and the growth of the T-shape roll cell, respec-
tively. For simulating the subject more realistically, the Boussinesq 1. The work fluid is ideal gas and follows the equation of state
assumption is not adopted, and instead, the compressibility of the of an ideal gas.
fluid is considered. The buoyancy force is then to vary with the 2. Magnitudes of gradients of density and pressure on the
heating process and exerts on the fluid from weakly to strongly. whole surfaces in the normal direction are zero.
The time needed for computation and data acquisition is extremely 3. Radiation heat transfer is neglected.
huge, and then an effective method of a CUDA computation [24] is
needed for executing parallel computations. Solution methods of The governing equations are expressed as follows.
the Roe [25], preconditioning [26] and dual time stepping [27]
for solving a low speed compressible fluid flow in a transient state @U @F @G @H
þ þ þ ¼S ð1Þ
are combined together. As well, a modified method of the non- @t @x @y @z
reflecting boundary condition [28] is used to treat the reflection
of pressure wave at the apertures of parallel plates. The results and the equation of state of an ideal gas is used.
show that under a low height situation the parallel plates are like P ¼ qRT ð2Þ
narrow parallel plates, and then the viscous force dominates the
The terms included in U, F, G, H and S are separately shown in the
flow field that easily causes the flow field to become a steady state.
following equations.
Oppositely, under a high height situation the parallel plates are like
wide parallel plates, and then the more active buoyancy force nat- 2 3
urally dominates the flow field that causes the flow field to be an
q
6 7
unsteady state. Unexpectedly, the T-shape roll cell cannot be ob- 6 qu 7
6 7
served. As for the results of the longer heated bottom surface, U¼6 7
6 qv 7 ð3Þ
two pairs of the T-shape roll cells are found out during a transient 6 7
4 qw 5
status. The height adopted in [24] seems to be an appropriate
qE
height to produce the T-shape roll cells.
120 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130

Y
h
X d
z
l

∂T
P0 = 1.013 × 105 Pa =0
∂y
T0 = 300 K u=v=0

aperture aperture
(non-reflecting g h (non-reflecting
boundary condition) boundary condition)
y
Th = 700 K
x
l2 l1 l2

Fig. 1. Physical model.

2 3 9
qu >
> 3. Numerical method
6 qu2 þ P  sxx 7 >>
6 7 >>
6 7 >>
F¼6 quv  sxy 7 >> The Roe scheme [25] and preconditioning method [26] are used
6 7 >>
4 quw  sxz 5 >>
>
>
to solve the governing equations of a compressible flow shown in
>
qEu þ Pu  k @x  usxx  v sxy  wsxz >
@T >
> Eq. (1). Also, the method of dual time stepping is added to calculate
2 3>>
> transient states. And Eq. (8) can be obtained.
qv >
>
>
6 q v s 7>>
6 u  yx 7>>
>
> C
@U p @U @F @G @H
þ þ þ þ ¼S ð8Þ
6 7
qv þ P  syy 7>
2
G¼6 > @s @t @x @y @z
6 7>>
4 qv w  syz 5>>
>
>
> where s is an artificial time, t is a physical time. C is a precondition-
qEv þ Pv  k @y  usyx  v syy  wsyz =
@T
ing matrix proposed by Weiss and Smith [26] and Up is a primitive
2 3 ð4Þ
qw >
> form of [q, qu, qv, qw, qe]T. Discretization of Eq. (8) is executed. The
>
>
6 quw  szx 7> @U
6 7>>
> terms of @ sp and @U
@t
are differentiated by a first-order forward differ-
6 7 >
H¼6 qv w  szy 7>>
> ence and a second-order backward difference, respectively, and the
6 7 >
>
4 qw þ P  szz
2 5 >
> terms of @F , @G, and @H are differentiated by a central difference, the
>
>
@x @y @z

qEw þ Pw  k @z  uszx  v szy  wszz >

@T > following equation can be obtained.
>
>
2 3 >
>
0 >
>
>
> U kþ1
p  U kp 3U nþ1  4U n þ U n1 1  kþ1 
6
6 0 7
7
>
>
> C þ þ F iþ1;j;k  F kþ1
i12;j;k
6 7 >
> Ds 2Dt Dx 2
ð9Þ
S ¼ 6 ðq  q0 Þg 7 >
> 1  kþ1  1  
6 7 >
>
4 0 5 >
> þ Gi;jþ1;k  Gkþ1 1;k þ H kþ1
1  Hkþ1 1 ¼S
>
> Dy 2
i;j2 Dz i;j;kþ 2
i;j;k 2
;
ðq  q0 Þg v
The terms of Uk+1 and Fk+1 in Eq. (9) are necessary to be linearized
Sutherland’s law is adopted to evaluate the viscosity and thermal and expressed as follows.
conductivity shown as follows.
 23 U kþ1 ¼ U k þ M DU p ð10Þ
T T 0 þ 110
lðTÞ ¼ l0 ð5Þ @U
in which M ¼ @U @F
and Ap ¼ @U
k
T0 T þ 110 p p

F kþ1 ¼ F k þ Ap DU p ð11Þ
lðTÞcR
kðTÞ ¼ ð6Þ
ðc  1ÞPr where Ap ¼ @U @F
is the flux Jacobian and the same methods for
p

in which
@G
Bp ¼ @U p
and @H
C p ¼ @U p
are used in linearization of Gk+1 and Hk+1,
respectively.
P 1 To substitute Eqs. (10) and (11) into Eq. (9), the following equa-
E¼ þ ðu2 þ v 2 þ w2 Þ ð7Þ
qðc  1Þ 2 tion is derived.
  
I 3
q0 ¼ 1:1842 kg=m3 ; g ¼ 9:81 m=s2 ; þ C1 M þ C1 dx Akp þ dy Bkp þ dz C kp DU p ¼ C1 Rk ð12Þ
Ds 2Dt
l0 ¼ 1:85  105 Ns=m2 ; T 0 ¼ 298:05 K;
c ¼ 1:4; R ¼ 287 J=kg=K;Pr ¼ 0:72: where dx, dy, and dz are central-difference operators and
k n
þU n1
Rk ¼ S  ð3U 4U
2Dt
Þ  ðdx F k þ dy Gk þ dz Hk Þ
 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130 121

The solver of the following Eq. (13) is the LUSGS implicit meth- Pði; 0; kÞ ¼ Pði; 1; kÞ
od proposed by Yoon and Jamesont [29]. uði; 0; kÞ ¼ uði; 1; kÞ
Ap ¼ C1 Akp v ði; 0; kÞ ¼ v ði; 1; kÞ ð21Þ
wði; 0; kÞ ¼ wði; 1; kÞ
Bp ¼ C1 Bkp ð13Þ Tði; 0; kÞ ¼ 2T h  Tði; 1; kÞ

where Th is the wall temperature.

C p ¼ C1 C kp 0 indicates the ghost cell and 1 indicates the cell closest to the
wall.
Eq. (12) can be rearranged as follows.
As for the boundary conditions at the apertures of parallel
plates, in order to avoid a low speed compressible flow in parallel
ðL þ D þ UÞDU p ¼ C1 Rk ð14Þ
plates to be polluted by the reflections of acoustic waves, the non-
where reflecting boundary conditions are then necessarily used at the
apertures of parallel plates.
        9
L ¼  D1x Aþp þ D1y Bþp þ D1z C þp >
>
>
In a high speed compressible flow condition, the method of
i1;j;k i;j1;k i;j;k1 >
>
              
>
>
= LODI (local one dimensional inviscid relations) proposed by Poin-
þ 
D ¼ DIs þ C1 M 23Dt þ D1x Ap  Ap þ D1y Bþp  Bp þ D1z C þp  C p sot and Lele [31] was suitably adopted for determining the non-
i;j;k i;j;k i;j;k i;j;k i;j;k i;j;k >
>
        >
>

>
>
> reflecting boundary conditions at the apertures of the channel.
U ¼ D1x Ap þ D1y Bp þ D1z C p ;
iþ1;j;k i;jþ1;k i;j;kþ1 However, a preconditioning matrix is not used in the above meth-
ð15Þ od that causes the method to be not available for determining the
non-reflecting boundary conditions at the apertures under a low
As for the computation of
    speed compressible flow. As a result, the modification of the
k 3U k 4U n þU n1 k k k
R ¼S 2Dt
 dx F þ dy G þ dz H in the right hand side above method [28] is adopted for resolving the non-reflection
(RHS) of Eq. (12), the terms included in F shown in Eq. (4) based on boundary conditions under an extremely low speed compressible
Cartesian coordinates can be divided into two parts. One is the flow.
inviscid term Finviscid and the other is the viscous term Fviscous. Along the Z direction, periodic conditions are used, and then the
boundary conditions of surfaces of ABCD and EFGH shown in Fig. 1
F ¼ F inv iscid þ F v iscous can be expressed as the following equations, respectively.
8 9 8 9
>
> qu >
> > 0 >
>
> >
> >
> >
> Pði; j; 0Þ ¼ Pði; j; nzÞ
>
> qu2 þ p >
> >
> s >
>
>
< >
= < > xx >
= uði; j; 0Þ ¼ uði; j; nzÞ
¼ quv  sxy ð16Þ
>
> >
> >
> >
> v ði; j; 0Þ ¼ v ði; j; nzÞ ð22Þ
>
> quw >
> >
> sxz >
>
>
>   >
> >
> >
> wði; j; 0Þ ¼ wði; j; nzÞ
> 2
: q e þ V u þ pu  k @T ; > : ;
2 @x þusxx þ v sxy þ wsxz Tði; j; 0Þ ¼ Tði; j; nzÞ
The upwind difference scheme developed by Roe [25] is employed
in discretization of the term of Finviscid at the interface cells ði þ 12Þ Pði; j; nz þ 1Þ ¼ Pði; j; 1Þ
and expressed as follows at a low Mach number situation. uði; j; nz þ 1Þ ¼ uði; j; 1Þ
v ði; j; nz þ 1Þ ¼ v ði; j; 1Þ ð23Þ
1 1 wði; j; nz þ 1Þ ¼ wði; j; 1Þ
F inv iscid;iþ1 ¼ ðF R þ F L Þ  jC1 AMjDU P ð17Þ
2 2 2 Tði; j; nz þ 1Þ ¼ Tði; j; 1Þ
The third-order MUSCL scheme proposed by Abalakin et al. [30] is
used to compute the term of Finviscid. And a second central difference 0 indicates the cell at the ABCD surface and nz + 1 indicates the cell
is adopted to calculate the magnitudes of the viscous terms and the at the EFGH surface.
related derivations are indicated as follows. A procedure calculating the equations mentioned above is
briefly described as follows.
@u ui1  uiþ1
¼ þ oðDxÞ ð18Þ
@x 2Dx (1) Assign the initial conditions of pressure, velocities and
And the same methods for G and H are used, respectively. temperatures.
On the adiabatic surface, the boundary conditions are (2) Use the MUSCL method to calculate Eq. (12) to obtain the
magnitude of DUp.
Pði; 0; kÞ ¼ Pði; 1; kÞ (3) Substitute the magnitude of DUp into Eq. (17) and use the
Roe method to calculate the magnitudes of inviscid terms
uði; 0; kÞ ¼ uði; 1; kÞ
of Finviscid.
v ði; 0; kÞ ¼ v ði; 1; kÞ ð19Þ (4) Calculate Eq. (18) to obtain the magnitudes of viscous terms
wði; 0; kÞ ¼ wði; 1; kÞ and substitute in Eq. (16).
Tði; 0; kÞ ¼ Tði; 1; kÞ (5) Solve U pkþ1 .

Pði; ny þ 1; kÞ ¼ Pði; ny; kÞ U kþ1

p ¼ U kp þ DU kp ð24Þ
uði; ny þ 1; kÞ ¼ uði; ny; kÞ (6) Calculate Eq. (9) and examine the convergence of the itera-
U kþ1
p U p
k
v ði; ny þ 1; kÞ ¼ v ði; ny; kÞ ð20Þ tional computation of U kþ1
p . Repeat (2)–(5) until Ds
< e,
3
wði; ny þ 1; kÞ ¼ wði; ny; kÞ e = 10 .
Tði; ny þ 1; kÞ ¼ Tði; ny; kÞ
The time of computation and date acquisition is extremely
On the heated surface, the boundary conditions are huge, and an effective method of a CUDA computation platform
122 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130

Fig. 2. Schematic diagram of CUDA platform [32].

is then used for executing parallel computation. The main function 4. Results and discussion
of the CUDA computation platform [14], which is a kind of inte-
grated technique developed by NVIDIA company, is to integrate The height of parallel plates is usually regarded as a character-
the performance of a GPU (Graphics Processing Unit) arrayed in a istic length when the Rayleigh number is used in a parallel plates
video card installed in a host computer to a numerical parallel problem and defined as follows.
computation processor executed in the host computer. 3
A brief description of usage of the CUDA is indicated as follows. g q20 ðT h  T c Þh
Ra ¼ Pr ð26Þ
Three steps have been executed and Fig. 2 is the schematic dia- T 0 lðTÞ2
gram. The first step is that the data are moved from the memory
However, the length of the heated bottom wall affects heat
of the host to the memory in the graphic hardware. Then the
transfer phenomena remarkably. In order to highlight the influence
GPU is utilized to compute. The general method of assigning num-
of the length of the heated wall, a modified Rayleigh number Ra⁄ is
bers in three dimensional grids is a three dimensional matrix (i, j,
newly defined in the following equation.
k). However, it will be very complicated to move the data from
the host to the graphic hardware if the three dimensional matrix l
Ra ¼ Ra  ð27Þ
is adopted. Therefore, it is necessary to transport the three dimen- h
sional matrix to the one dimensional matrix. Then, the data will be In this study, relative to the scales of the previous study [24] two
easy to be moved. The equation of transporting is showed in the different heights and a different length of the heated wall are con-
following process. sidered, and three different magnitudes of R ¼ hl are shown as fol-
lows, respectively.
Device ðassign numbersÞ ¼ hostði  ny  nz þ j  nz þ kÞ ð25Þ
11 11 22
ðRa ¼ 2:18  106 Þ; ðRa ¼ 1:95  107 Þ and ðRa ¼ 1:73  107 Þ:
2 6 4
where nx, ny and nz are the numbers of grids in the x, y and z direc-
tion, respectively. The grid distribution of 250  160  10 is selected according to
The second step is to parallel the program. The best advantage the previous study [24]. The definitions of local and time-averaged
of the graphic hardware is that there are multiple cores in it to exe- local Nusselt numbers of Nux and Nux are expressed as follows,
cute the computation. Then the program paralleling is absolutely respectively.
necessary.  
h @T
The third step is to move the results computed by GPUs from Nux ¼ kðTÞ ð28Þ
k0 ðT h  T c Þ @y
the graphic hardware to the host and transform the one dimen-
sional matrix to the original three dimensional matrix. Perfor- Z
 1
mance tests of CUDA computation platforms are similar to those Nux ¼ Nux dt ð29Þ
t t
of [24].
 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130 123

(a) t = 1.99 ×10−3

(b) t = 1.56 ×10−2

(c ) t = 2.34 ×10−2

(d ) t = 4.68 ×10−2

( e) t = 5.66 ×10−2

(f) t = 7.41×10−2
thermal field streamlines
Fig. 3. Histories of development of thermal field and streamlines under Ra⁄ = 2.18  106.

In Fig. 3, histories of development of thermal fields and stream- cooling fluids sucked from surroundings begin to increase and con-
lines are separately indicated under Ra⁄ = 2.18  106 situation. The tract the space filled with the heated fluids. Finally, the ascending
height is a half of [24]. The darker the color is, the lower the tem- heated fluids concentrate in the central region, and the circulations
perature displays. At the initial stage (t = 1.99  103), heat con- are completely eliminated. The thermal and flow fields become a
duction mode dominates the heat transfer phenomenon, and symmetrically steady situation shown in Fig. 3(f). Due to the low
then thermal layers distributed on the whole heated bottom sur- height of parallel plates, the strength of viscous shear force is rel-
face are uniform. Fluids close to the heated bottom surface are atively apparent, and then the strength of natural convection
heated and expanded, so via apertures fluids are discharged into seems to have difficulty to produce the T-shape roll cell which
surroundings. In the next stage (t = 1.56  102), because the found out in [24].
height of parallel plates is low, light heated fluids close to the aper- Shown in Fig. 4, the height of parallel plates is increased and
tures are first raised and flow out off the parallel plates. Meanwhile two times as high as that of [24]. This situation means that natural
cooling fluids are sucked from surroundings and flow into the par- convection plays a more active role and the strength of it becomes
allel plates, but they are restricted in the outside region of the stronger. In Fig. 4(a), at an initial stage heat conduction mode is
ascending fluids mentioned above. Therefore, close to the aper- dominant, and then fluids which are expanded by the heat from
tures there are two dark zones surrounded by two light belts to of the heated bottom wall are discharged to surroundings. Thermal
be observed. This phenomenon causes the temperatures of the flu- layers are uniformly distributed on the heated bottom surface
ids in the region between the two light belts to be increased grad- which is similar to that shown in Fig. 3(a). However, as the time
ually. Accompanied with the increment of the temperatures of passes and the strong strength of natural convection directly
fluids, the fluids start to form circulations by themselves. As the causes heated fluids in the central region to ascend and flow out-
time passes, the light region in which the temperature is increased ward. The T-shape roll cell is unable to be formed in time and
is enlarged gradually because of the activity of natural convection not observed. The T-shape roll cell shown in [24] indicated the
mode. The circulations merge into two big pairs of circulations ascending fluids separately to flow outward and inward that
shown in Fig. 3(d). Afterward, the heated fluids ascend and im- means the strength of cooling fluids which are sucked from sur-
pinge on the top wall and start to flow out off parallel plates. Then roundings due to natural convection to be not strong enough to
124 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130

(a) t = 6.54 × 10−4

(b) t = 3.04 × 10−3

(c ) t = 3.9 × 10−3

(d ) t = 2.25 × 10−2

(e) t = 2.47 × 10−2

(f) t = 2.62 × 10−2

thermal field streamlines
Fig. 4. Histories of development of thermal field and streamlines under Ra⁄ = 1.95  107.
 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130 125

(a) t = 4.97 ×10−4

(b) t = 7.32 ×10−3

(c ) t = 9.27 ×10−3

(d ) t = 1.66 ×10−2

( e) t = 1.9 ×10−2

(f) t = 2.19 ×10−2

(g) t = 2.88 ×10−2

thermal field streamlines

Fig. 5. Histories of development of thermal field and streamlines under Ra⁄ = 1.73  107.

lead the ascending fluids to flow outward exclusively. Except the big circulation shown in Fig. 4(e). Due to easy accumulation of heat
appearance of two large circulations caused by the ascending fluids energy on the central bottom surface, sometimes the two small
mentioned above, two small closed circulations in the low central suppressed circulations are suddenly ejected from the bottom sur-
region induced by the two large circulations are observed. After- face like a shape of mushroom.
ward, the two big circulations continuously grow and suppress In Fig. 5, the length of the heated bottom surface of this situa-
the two small closed circulations that forces the two new small tion is two times as long as that of the previous study [24]. The
closed circulations to be released and flow outward along the height of both situations is the same. Since the heated bottom sur-
edges of the big circulations. Finally, the two big circulations con- face is long, cooling fluids sucked from surroundings are more dif-
tact with each other and mutually squeeze. Although the two small ficult to flow into the central region directly in initial stages. As a
closed circulations are suppressed by the two big circulations, they result, shown in Fig. 5(b) based on the central region two pairs of
still exist on the heated bottom surface. No small closed circulation high temperature convexes are observed symmetrically in the
existed and only two big swing circulations were observed at the thermal field on both sides. Corresponding to the convexes, several
final stage of [24]. The reason of the different phenomenon be- circulations are formed. Afterward, at the corresponding locations
tween this work and [24] is suggested as that the strength of nat- of the convexes T-shape roll cells which were observed in [24] are
ural convection of this work is stronger than that of [24]. The active indicated. The number of the convexes is larger than that of [24],
strength of natural convection of this work in the central region because of the longer length of this study. As well, circulations
easily produces ascending fluids, but the ascending fluids have grow bigger accompanied with the growth of T-shape roll cells.
no ability to push away the upper big circulations formed by the These phenomena are not observed in the above situations shown
whole regions. As a result, the ascending fluids are permitted to in Figs. 3 and 4. It implies that an appropriate ratio of the height to
form small closed circulations only. Because of the mutual squeeze the length will cause the indication of the T-shape roll cell to ap-
of the two big circulations, the situation of natural convection be- pear in parallel plates. The Rayleigh–Benard convection indicated
comes unsteady and the strength of the closed small circulation in the situation of the enclosure [1–4] is not found out in this
varies inversely with the variation of the strength of the upper study. It is reasoned as that the situation with open-ended
126 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130

7(10, 75,5) 6(55, 73,5) 5(105, 67,5)

4(108, 40,5)

1(10, 6,5) 2(55,8,5) 3(105, 23,5)

(a)

1(10, 6,5)

2(55,8,5)

3(105, 23,5)

4(108, 40,5)
Vmag

5(105, 67,5)

6(55, 73,5)

7(10, 75,5)

t
(b)
Fig. 6. Variations of local resultant velocity Vmag at different locations of a time average streamline under Ra⁄ = 2.18  106.

apertures leads fluids to exchange interruptedly between the in- In duration of an initial stage, fluids in parallel plates are dis-
side of parallel plates and surroundings. Then the permanent charged into surroundings by heat conduction mode, and right
detention of fluids in their own circulations mentioned above be- after fluids are sucked from surroundings and flows into parallel
comes impossible, and the indication of the Rayleigh–Benard con- plates by natural convection mode. As a result, the variation of
vection in parallel plates is hardly found out. The growth of the T- the resultant velocity is apparent even though the magnitude of
shape roll cell proceeds with a continuously heating process, and a the resultant velocity is small. Accompanied with increment of
mutual interference and mergence among the T-shape roll cells be- time, natural convection gradually dominates the variations of
gin to be observed. The number of circulations begins to be de- thermal and flow fields. Due to the low height of parallel plates,
creased. Finally two big circulations are left, and two small the space for the development of natural convection is limited,
closed circulations hide beneath the two big circulations to be and then the situation of natural convection becomes steady men-
found out that causes the situation to become an unsteady state. tioned above. Therefore, after the initial stage the variation of the
The height and length are separately lower and longer than those resultant velocity at each location is hardly found out.
of Fig. 4, and then the two small closed circulations have no possi- In Fig. 7, the height of parallel plates is increased. The space for
bility to be ejected from the central bottom surface. the development of natural convection is enlarged that causes ac-
In Fig. 6, a time-averaged streamline close to both the top and tive behaviors of natural convection to be induced. The location
bottom surfaces is indicated under the Ra⁄ = 2.18  106 situation. of point 1 is close to the aperture and fluids begin to be heated as
Points of 1–7 from the inlet to outlet present corresponding loca- they flow through this location. Then the influence of natural con-
tions on the streamline. Computational grids have difficulty to vection just begins that causes the variation of the resultant veloc-
match points of 1–7 exactly, and then the closest computational ity to be slight. At the location of point 2, a vertical velocity induced
grids corresponding to points of 1–7 are selected. Variations of by the buoyancy force becomes more remarkable than that at the
resultant velocities with time at each point are shown in location of point 1 that easily results in the variation of the resul-
Fig. 6(b). The definition of the resultant velocity Vmag is expressed tant velocity at this location being more drastic than that at the
as follows. location of point 1. The location of point 3 is in the central region
1 that leads the resultant velocity to have an upward trend and nat-
V mag ¼ ðU 2 þ V 2 þ W 2 Þ2 ð30Þ urally the variation of the thermal field to be slight. Consequently,
 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130 127

7(10, 233,5) 6(55, 226,5) 5(103,195,5)

4(105,120,5)

1(10,3,5) 2(55,11,5) 3(105, 26,5)

(a)

1(10,3,5)

2(55,11,5)

3(105, 26,5)

Vmag 4(105,120,5)

5(103,195,5)

6(55, 226,5)

7(10, 233,5)

t
(b)
Fig. 7. Variations of local resultant velocity Vmag at different locations of a time average streamline under Ra⁄ = 1.95  107.

the variation of the resultant velocity at this location is more stable the instantaneous area-averaged Nusselt number is expressed as
than that at the location of point 2. follows.
From point 4 to 7, the heating process is almost completed, then Z Z
 1
fluids possess more active behaviors. Except the indication of large NuA ¼ Nux dxdz ð31Þ
magnitudes of resultant velocities, variations of resultant velocities dl d l

at these locations are naturally apparent than those at the front Naturally, the higher the height is, the larger the instantaneous
locations mentioned above. The observation of the relatively small area-averaged Nusselt is achieved. Under the situation of
variation of the resultant velocity at the location of point 7 is Ra⁄ = 2.18  106, it is stable except the initial stage, then no varia-
caused by a viscous effect because of the short distance between tion of the instantaneous area-averaged Nusselt number is ob-
the location and top surface. served. Oppositely, under the situation of Ra⁄ = 1.95  107, it is
In Fig. 8, due to the longer length of the heated bottom surface, reveals an unstable state, then an irregular variation of the instan-
cooling fluids sucked from surroundings absorb more heat energy taneous area-averaged Nusselt number appears.
that leads behaviors of fluids to become more active. Restrictive ef- Shown in Fig. 10, there are variations of time-averaged local
fects of the shear force caused by the top and bottom walls on Nusselt numbers on the central cross section of the xy plane. A
behaviors of fluids are no longer dominant. Variations of the resul- more larger space for the development of natural convection can
tant velocity at each point are then more apparent than those be used under a higher height situation. As a result, large time-
shown in Fig. 6. But magnitudes of resultant velocities are gener- averaged local Nusselt numbers are achieved with the situation
ally smaller than those indicated in Fig. 7 reasoned by the smaller of the large Rayleigh number. In Fig. 4(e), two small closed circula-
height adopted in this situation. tions exist that causes two convexes of the time-averaged local
In Fig. 9, variations of instantaneous area-averaged Nusselt Nusselt number to be indicated in the central region of the
numbers on the xz plane with time are shown. The definition of situation of Ra⁄ = 1.95  107. Under the stable situation of
128 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130

7(10,150,5) 6(50,149,5) 5(92,150,5)

4(101,80,5)

1(10,8,5) 2(50,10,5) 3(92,16,5)

(a)

1(10,8,5)

2(50,10,5)

3(92,16,5)

Vmag 4(101,80,5)

5(92,150,5)

6(50,149,5)

7(10,150,5)

t
(b)
Fig. 8. Variations of local resultant velocity Vmag at different locations of a time average streamline under Ra⁄ = 1.73  107.

Ra* = 2.18 × 106 Ra* = 2.18 × 106

Ra* = 8.66 ×106 [24] Ra* = 8.66 ×106 [24]
Ra* = 1.73 × 107 Ra* = 1.73 × 107
Ra = 1.95 ×10
* 7
Ra* = 1.95 ×107

Nu x
Nu A

X
t
Fig. 10. Distributions of time-averaged local Nusselt numbers on the central cross
Fig. 9. Variations of instantaneous area-averaged Nusselt numbers with time. section of the xy plane.
 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130 129

definition of time and area averaged Nusselt number is shown as

follows.
Z Z
 1 
Nu ¼ Nux dxdz ð32Þ
dl d l

The reasonable difference between the experimental results

[11] and the previous study [24] was reasoned in [24]. The Ray-
leigh numbers of this work are larger than those in [24], larger time
and area averaged Nusselt numbers are then achieved. The slope of
(a) thermal field (present results) the distribution of the present results is slightly larger than the
slope of the extension line of the results of [24]. The reason is sug-
gested as that unstable phenomena easily happen in situations of
the large Rayleigh numbers.

5. Conclusions

An investigation of effects of heights and a length on natural

convection of three dimensional parallel plates with a heated bot-
tom surface is studied numerically. Phenomena of this work are
(b) streamlines (present results) deeply affected by the height and length, and different from those
observed in the previous study [24]. Several conclusions are drawn
as follows.

(1) The space used for developing natural convection plays an

important role to dominate phenomena of natural convec-
tion. In a narrow space, phenomena trend to a stable situa-
tion, and vice versa.
(c) existing results [14] (2) A T-shape roll cell is a kind of intermediate product and
observed in an earlier stage of the growth of natural convec-
Fig. 11. Comparison of flow fields of present results and existing results [14]. tion under a certain height in this work.

Ra⁄ = 2.18  106, the appearance of the smallest time-averaged lo-

cal Nusselt number in the central region is observed and Acknowledgments
reasonable.
Flow visualization of existing results of [14] compared with The authors gratefully acknowledge the support of the Natural
present results are shown in Fig. 11. Near the central line, a region Science Council, Taiwan, ROC under Contact NSC99-2221-E-009-
where fluids turn flowing directions is found out. Both results have 058.
a similar tread.
A comparison of time and area averaged Nusselt numbers of References
present results and existing results is revealed in Fig. 12. The
[1] L.Q. Tang, T.T.H. Tsang, Temporal, spatial and thermal features of 3-D
Rayleigh–Bénard convection by a least-squares finite element method,
Comput. Methods Appl. Mech. Eng. 140 (1997) 201–219.
[2] H. Yang, Z. Zhu, Numerical simulation of turbulent Rayleigh–Benard
convection, Int. J. Heat Mass Transfer 33 (2006) 184–190.
Turgut and Onur [11] [3] B.A. Puthenveettil, A. Jaywant H, Plume structure in high-Rayleigh-number
⎡⎛ 2 h 2 ⎞ * ⎤
0.3362 convection, J. Fluid Mech. 000 (2005) 1–33.
Nu exp = 0.1686 ⎢⎜ ⎟ Ra ⎥ [4] M. van Reeuwijk, H.J.J. Jonker, K. Hanjalić, Wind and boundary layers in
⎣⎝ w × l3 ⎠ ⎦ Rayleigh–Bénard convection. I. Analysis and modeling, Phys. Rev. E 77 (2008)
0.24
⎡⎛ 2 h 2 ⎞ * ⎤ 036311.
Nu num = 0.16 ⎢⎜ ⎟ Ra ⎥ [ 24] [5] W. Aung, L.S. Fletcher, V. Sernas, Developing laminar free convection between
⎣⎝ w × l3 ⎠ ⎦ vertical flat plates with asymmetric heating, Int. J. Heat Mass Transfer 15
Present result (1972) 2293–2308.
[6] P.H. Oosthuzien, A numerical study of laminar free convective flow through a
vertical open partially heated plane duct, ASME HTD 32 (1984) 41–48.
[7] S. Kim, N.K. Anand, W. Aung, Effect of wall conduction on free convection
Nu between asymmetrically heated vertical plates: uniform wall heat flux, Int. J.
Heat Mass Transfer 33 (1990) 1013–1023.
[8] D. Naylor, J.M. Floryan, J.D. Tarasuk, A numerical study of developing free
convection between vertical parallel plates, Trans. J. Heat Mass Transfer ASME
113 (1991) 620–626.
[9] A. La Pica, G. Rodono, R. Volpes, An experimental investigation on natural
convection of air in a vertical channel, Int. J. Heat Mass Transfer 36 (1993) 611–
616.
[10] K. Kitamura, T. Asakawa, Fluid flow and heat transfer of natural convection
over upward-facing, horizontal, heated plate shrouded by a parallel insulated
plate, Heat Transfer Asian Res. 29 (2000) 333–346.
[11] Oğuz Turgut, Nevzat Onur, An experimental and three-dimensional numerical
study of natural convection heat transfer between two horizontal parallel
Ra* plates, Int. J. Heat Mass Transfer 34 (2007) 644–652.
[12] H. Yang, Z. Zhu, J. Gilleard, Numerical simulation of thermal fluid instability
Fig. 12. Comparison of averaged average Nusselt numbers of present results and between two horizontal parallel plates, Int. J. Heat Mass Transfer 44 (2001)
existing results [11,24]. 1485–1493.
130 W.-S. Fu et al. / International Journal of Heat and Mass Transfer 67 (2013) 118–130

[13] O. Manca, B. Morrone, S. Nardini, Experimental analysis of thermal instability [23] A. Andreozzi, O. Manca, Numerical investigation on the steady state natural
in natural convection between horizontal parallel platesuniformly heated, J. convection in a horizontal open-ended cavity with a heated upper wall,
Heat Transfer 122 (2000) 50–57. Numer. Heat Transfer A 57 (2010) 453–472.
[14] O. Manca, S. Nardini, Experimental investigation on natural convection in [24] W.S. Fu, W.H. Wang, S.H. Huang, An investigation of natural convection of
horizontal channels with the upper wall at uniform heat flux, Int. J. Heat Mass three dimensional horizontal parallel plates from a steady to an unsteady
Transfer 50 (2007) 1075–1086. situation by a CUDA computation platform, Int. J. Heat Mass Transfer 55 (2012)
[15] S. Baskaya, H. Kauli, Effects of plate separation, plate length and temperature 4638–4650.
difference on buoyancy driven heat transfer in finite horizontal parallel plates, [25] P.L. Roe, Approximation Riemann solver, parameter vectors, and difference
Heat Mass Transfer 40 (2004) 293–330. schemes, J. Comput. Phys. 43 (1981) 357–372.
[16] J.F. Hinojosa, C.A. Estrada, R.E. Cabanillas, G. Alvarez, Numerical study of [26] J.M. Weiss, W.A. Simth, Preconditioning applied to variable and constants
transient and steady-state natural convection and surface thermal radiation in density flows, AIAA 33 (1995) 2050–2056.
a horizontal square open cavity, Numer. Heat Transfer A 48 (2005) 179–196. [27] X.F. Xu, J.S. Lee, R.H. Pletcher, A compressible finite volume formulation for
[17] T. Icoz, Y. Jaluria, Numerical simulation of boundary conditions and the onset large eddy simulation of turbulent pipe flows at low Mach number in
of instability in natural convection due to protruding thermal sources in an Cartesian coordinates, J. Comput. Phys. 203 (2005) 22–48.
open rectangular channel, Numer. Heat Transfer A 48 (2005) 831–847. [28] W.S. Fu, C.G. Li, C.C. Tseng, An investigation of a dual-reflection
[18] F. Penot, Numerical calculation of two-dimensional natural convection in phenomenon of a natural convection in a three dimensional horizontal
isothermal open cavities, Numer. Heat Transfer 5 (1982) 421–437. channel without Boussinesq assumption, Int. J. Heat Mass Transfer 53
[19] K. Vafai, J. Ettefagh, The effect of sharp corners on buoyancy-driven flows with (2010) 1575–1585.
particular emphasis on outer boundaries, Int. J. Heat Mass Transfer 33 (1990) [29] S. Yoon, A. Jamesont, Lower–upper symmetric-gauss–Seidel method for the
2311–2328. Euler and Navier–Stokes equations, AIAA 26 (1988) 1025–1026.
[20] K. Khanafer, K. Vafai, Buoyancy-driven flows and heat transfer in open-ended [30] I. Abalakin, A. Dervieux, T. Kozubskaya, A vertex centered high order MUSCL
enclosures: elimination of the extended boundaries, Int. J. Heat Mass Transfer scheme applying to linearised Euler acoustics, INRIA No 4459 (2002).
43 (2000) 4087–4100. [31] T.J. Poinsot, S.K. Lele, Boundary conditions for Navier–Stokes, J. Comput. Phys.
[21] K. Khanafer, K. Vafai, Effective boundary conditions for buoyancy-driven flows 101 (1992) 104–129.
and heat transfer in fully open-ended two-dimensional enclosure, Int. J. Heat [32] NVIDIA CUDA Compute Unified Architecture 3.0 programming Guide, NVIDIA
Mass Transfer 45 (2002) 2527–2538. Corporation, 2010.
[22] A. Andreozzi, Y. Jaluria, O. Manca, Numerical investigation on transient natural
convection in a horizontal channel heated from upper wall, Numer. Heat
Transfer A 51 (2007) 815–842.