hr. J. Heor Mar TransJ~r. Vol. 27. No. 5, pp. 755~~772. 1984 0017 Y310/84.$300+0.

00
Printed in Great Britam tc! ,984 Pergamon Pres Ltd

LAMINAR AND TURBULENT NATURAL CONVECTION

IN AN ENCLOSED CAVITY

N. C. MARKATOS* and K. A. PERICLEOUS

Concentration Heat and Momentum Limited, 40 High Street, Wimbledon, London SW19 SAU, U.K.

(Received 9 May 1983 and in revisedform 3 October 1983)

Abstract-The paper presents a computational method used to obtain solutions of the buoyancy-driven
laminar and turbulent flow and heat transfer in a square cavity with differentially heated side walls. A series of
Rayleigh numbers, ranging from lo3 to 1Or6was studied. Donor-cell differencing is used. and mesh-refinement
studies have been performed for all Rayleigh numbers considered. The turbulence model used for Rayleigh
numbers greater than 10” is a (k N E)two-equation model of turbulence, that includes gravity z density gradient
interactions. The results are presented in tabular and graphical form, and as correlations of the Nusselt
and Rayleigh numbers. Furthermore, the results for Rayleigh numbers up to lo6 are compared with the
benchmark numerical solution of de Vahl Davis.

NOMENCLATURE by Ostrach [S] and discussed more fully in refs. [6, 71.
The purpose of this work is to describe a computa-
D cavity width [m]
tional procedure for solving the non-linear, coupled
g gravitational acceleration [m s-‘1
differential elliptic equations over the entire flow
!i kinetic energy of turbulence per unit mass
domain, with no assumptions concerning the core
CJhi-‘1 configuration or any other ad hoc simplification for
NU Nusselt number
Rayleigh numbers up to 106, and with the speculative
Pr Prandtl number
use of a two-equation turbulence model for higher
Ra Rayleigh number, p2gD3/lAT PrJp’
Rayleigh numbers ; and to demonstrate that this can be
S source term
accomplished without excessive demands on computer
T temperature [K]
time or storage. The presented results are restricted to
V velocity component in the y-direction
rectangular cavities of aspect ratio 1, fluids of Prandtl
[m s-r]
number 0.71, and Rayleigh numbers ranging from 10”
W velocity component in the z-direction
to 10t6. However, the procedure is general and can be
[m s-‘1
easily applied for practical computations in cavities of
Y,Z spatial coordinates [m]
different aspect ratios, fluids of different Prandtl
Y,z normalized coordinates.
numbers, three-dimensional (3-D) enclosures or tilted
enclosures such as. for example, those studied by
Catton [4].
1. INTRODUCTION

THE DETERMINATIONof buoyancy-driven flow in an

enclosed cavity provides a suitable comparison prob- 2. THE PROBLEM CONSIDERED
lem for evaluating the performance of numerical
The problem considered is depicted schematically in
methods dealing with viscous flow calculations [l-3].
Fig. 1, and refers to the two-dimensional (2-D) flow in a
Furthermore, the above process has many practical
square cavity, of side D. The cavity is assumed to be
applications of which the most widely known is that of
of infinite depth along the x-axis and is heated
double glazing. Other applications include nuclear-
differentially along the South (low-y) and North
reactor insulation, ventilation of rooms, solar-energy
(high-y) walls. The other two (horizontal) walls are in-
collection and crystal growth in liquids. There is an ever
sulated.
increasing amount of research on confined natural
The no-slip condition is applied on the velocity at all
convection and refs. [4-l 0] are typical examples ofsuch
four walls, and friction is calculated by invoking ‘wall-
work. A recent review of the existing literature is given
functions’ [ 121, see Appendix. The hot and cold walls
by Ostrach [ 1 I]. Despite all the recent research activity,
are considered to be isothermal, and the other two
a central problem that has remained unsolved is the
adiabatic. Heat transfer through the walls causes
coupling between boundary layers and core flows. The
density changes to the fluid in the cavity, and leads to
latter depend on the boundary layer, which, in turn, is
buoyancy-driven recirculation. The resulting flow is
influenced by the core. This problem was first identified
treated as steady and, depending on the Rayleigh
number, laminar (Ra < 106) or turbulent (Ra > 106).
* Present address: Faculty of Science and Mathematics,
This criterion for switching over to turbulence cal-
School of Mathematics, Statistics and Computing, Thames culations is based on experimental observations [25,
Polytechnic, London SE18 6PF, U.K. 261 and has been used widely in the literature [l--3,20,

755
756 N. C. MARKATOSand K. A. PERICLEOUS

dT/dr =0

-D-

(a) (b)
FIG. 1. (a) The process considered. (b) The solution grid used for Ra < 106.

221. Reference fluid properties were calculated at the (k - F) turbulence model was used. Then the above
ambient temperature T, = 293 K. equations are time-averaged equations and p and rb
are replaced by their ‘effective’ values pLerfand rerfr as
3. MATHEMATICAL FORMULATION given by the turbulence model [12,13].
AND METHOD OF SOLUTION The generation term in the k-equation includes the
buoyancy production
3.1. The dzferential equations
For steady flow, the equations
velocity components and temperature
for continuity,
take the form :
G, = -Bs$$-
continuity
In stable stratification, G, becomes a sink term so that
G4 I Gw) _ o,
the turbulent mixing is reduced. In unstable
(1)
ay az ’ stratification, the buoyancy will enhance turbulence
since Ga is positive. The buoyancy term appearing
y-direction momentum tentatively in the c-equation, in other k - Ecalculations
for buoyant flows [ 14,241, has been omitted. There is
;(Puw)+ =; (II;)+~(f$)+s.:
fpuu) no obvious physical reason for including such a term
and other related work indicated that it is completely
(2)
insignificant [ 151. The turbulence model contains five
z-direction momentum constants which were assigned the following values
WI
&ww)+ (PZ)+;(lrg)+s,:
~(Puwl=; C1 = 1.44; C, = 1.92; C, = 0.09;
(3) ok = 1.0; aE = 1.314.
general transported fluid scalar, 4 (e.g. T, k, E)
More details may be found in refs. [14, 153.

=f (r,m&
;bw9)+&4) ““>
+
(r,,y““1
+s,; a
ay

(4)
Assuming that p is proportional
are obtained from the temperature
invoking the Boussinesq-type
to l/T the densities
field, without
approximations.

where r+ is the exchange coefficient for the transport of 3.2. The solution procedure
property I$. The source terms in the momentum The above equations were solved on a square mesh
equations are by the finite-domain method outlined in refs. [16-18-J.
Finite-domain equations are derived by integration of
(5) the differential equations over an elementary control
volume or cell surrounding a grid node. Upwind
differencing is used in the convective terms and the
integrated source term is linearized. Both these
practices are widely used to enhance numerical
ST = 0,
stability. The upwind scheme has come under much
where 0 is the temperature rise above ambient TO. criticism recently, but it is only grid-refinement that can
For Rayleigh numbers above lo6 a two-equation detect the ‘false diffusion’ associated with the various
 Laminar and turbulent natural convection in an enclosed cavity 151

schemes. Therefore, care was taken to obtain grid- formula gave results very close to the ones calculated by
independent results. Pressures are obtained from a equation( 10) for Ra up to lOa, but it became erroneous,
pressure-correction equation which yields the pressure as expected, for the high Ra numbers.
change needed to procure velocity changes to satisfy
mass continuity. The ‘SIMPLEST’ practice [17] is
4. RESULTS AND DISCUSSION
followed for the momentum equations. The most
significant difference between ‘SIMPLEST’ and the The main results are presented in graphical and
well-established ‘SIMPLE’ algorithm [ 193 is that in the tabular form, and as Nu vs Ra correlations. The
former the finite-domain coefficients for momenta graphical results are obtained by the post-processor
contain only diffusion contributions, the convection GRAFFIC [21].
terms being added to the linearized source term of the Solutions were computed for Pr = 0.71 (air) and Ra
equations. This implies that, in the absence of diffusion, between lo3 and 1016. The solutions for Ra up to lo6
the momentum equations are solved by a Jacobi point- were obtained by using normalized variables, in such a
by-point procedure as opposed to the more popular way that dimensionless velocities were of the order of
‘simultaneous’ line-by-line procedures. This mixed unity, in order to improve the numerical accuracy of the
practice (the use of Jacobi for convection terms, the use results. Indeed it was found that this procedure led to a
of line-by-line for diffusion terms) derives from 3% improvement in the important parameters over the
experience and intuition. For example, use of Jacobi for dimensional solution, due to the limited accuracy of the
diffusion in a pipe would take an extremely long time to mini-computer used.
spread the viscous effects of the wall, while a The validity of the relationship O(y,z) = /3(1--y,
simultaneous solution would do that in virtually a 1 -z), where 0 is the dependent variable, was found
single iteration. This is not so for the convection terms, to hold for all the solutions obtained, with very close
however, because of their special links with the approximation, everywhere in the flow field. An ex-
pressure-correction equation. The above mixed ception to this was in the cavity core for Ra > lo’*,
practice was found to accelerate convergence probably due to the round-off error in the very small
significantly. velocities encountered in this region, as compared with
The equations are solved by a line-by-line procedure the large ones at the wall boundary layers; but even
which is similar to Stone’s Strongly Implicit Method then the departure from the above centro-symmetry
but free from parameters requiring case-to-case adjust- was within a few percent.
ment and so less complex and slower. The pressure- The presented results are practically grid in-
correction equation is solved in a ‘whole-field’ manner, dependent (see Section 6 on grid dependence).
2-D simultaneous. Further details may be found The velocity components at the domain centre-lines
in refs. [16, 171. are shown in non-dimensional form (using K/D as a
scale factor, where ti is the thermal diffusivity) in Figs. 2-
3.3. 7he Nusselt number 4, for Rayleigh numbers Ra = 103, lo5 and 106.
From the engineering viewpoint, the most important It can be seen from these graphs that as the Rayleigh
characteristic of the flow is the rate of heat transfer number increases, the velocity maximum moves closer
across the cavity. The Nusselt number on the hot wall at to the wall and its amplitude increases. At the same
y = 0 is given by time, the velocity between the two maxima becomes
progressively smaller and at Ra > lo6 flow reversal is
observed immediately outside the boundary layers.
The reason for the above behaviour will be explained in
conjunction with the thermal distribution inside the
section.
Figure 5 shows streamlines for Ra = 103-1016.
.,I
Corresponding maps of temperature are shown in
Nu(z)=+$ (10) Fig. 6. The presented contours, in all the figures are
H c
labelled according to a well-known convention. For ex-
where the heat flux at the wall, $Lall, is calculated by the ample, in Fig. 6 the contours are labelled via the
program, from wall-functions [12,19]. statement 283(2)303 K. This means that temperature
The average Nusselt number is given by contours are presented every 2 K with the first (at the
extreme right) corresponding to 283 K and the last to
fi = (11) 303 K.
At Ra = 103, streamlines are those of a single vortex,
The mid-plane value Nu,,, is also computed as it is with its centre in the centre of the domain. Corre-
probably [3] a more accurate quantity in describing the sponding isotherms are parallel to the heated walls,
heat flow across a cavity with adiabatic ‘end’ walls. The indicating that most of the heat transfer is by heat
three- and five-point formulae commonly used by other conduction. The effect of convection is seen as the
investigators [20] for resolving the wall temperature departure of the isotherms from the vertical. The vortex
gradient in equation (8) were also used. The five-point is generated by the horizontal temperature gradient
758 N. C. MAKKATOS and K. A. PERICLEWS

FIG. 2. (a) The vertical velocity component : Ra = 103. (b) The horizontal velocity component: Ra = 103.

IO-
(a)

08-

06-
a
$ Z/D=0 5
04-

wD/k

FIG. 3. (a) The vertical velocity component: Ra = 105. (b) The horizontal velocity component: Ra = 10’.
 Laminar and turbulent natural convection in an enclosed cavity 759

06
Q
3 t

wD/k

I I I I I
-100
0 02 04 06 08 IO

Z/D
FIG. 4. (a) The vertical velocity component: Ra = 106. (b) The horizontal velocity component: Ra = 106.

across the section. This gradient, aTjay, is nega- vortices are formed. Viscous diffusion appears to
tive everywhere in this case, giving rise to positive prohibit the development of these vortices for Ra
(clockwise rotation) vorticity. < 105. The vortices at Ra = lo5 are sufficiently strong
As the Rayleigh number increases (Ra = 104) the to convect the temperature fields to the extent that the
central streamline is distorted into an elliptic shape and isotherms are nearly horizontal in the centre,
the effect of convection is more pronounced in the preventing any vertical motion there (see w-velocity
isotherms. Temperature gradients are now more severe plots in Fig. 3).
near the vertical walls, but diminish in the centre. Increasing Ra to 106, causes the secondary vortices
This behaviour continues to Ra = 105; the central to move closer towards the walls and are convected
streamline is further elongated and two secondary further downstream. A third vortex appears in the
vortices appear inside it. Its long axis is now tilted in the centre of the section, again rotating clockwise. This is
direction of the flow, as the secondary vortices are surprising, as one would expect this last vortex to rotate
convected by the flow in the periphery. Heat transfer counterclockwise, to reduce the shear between the
by convection in the viscous boundary layers alters other two vortices. Mallinson and de Vahl Davis [22]
the temperature distribution to such an extent that attribute this to the presence of a small positive
temperature gradients in the centre are close to zero, or temperature gradient in the centre. Viscous diffusion
change sign, thus promoting negative vorticity. This between the secondary vortices dissipates any
causes the development of secondary vortices in the counterclockwise vortices that might appear. Heat
core. transfer is now mostly by convection in the rapidly
As discussed in ref. [22], the secondary vortices in the moving fluid near the walls. The boundary layers
square cavity do not result from an instability of the adjacent to the vertical walls have become thin and fast.
base flow but are a direct consequence ofthe convective In the central region the vertical stratification in the
distortion of the temperature field. As Ra increases, the temperature distribution, shown in Fig. 6 with
development of thermal boundary layers intensifies increasing values from the bottom to the top of the
aTlay in the vicinity of the walls, and the convection cavity, prevents any vertical motion as confirmed by the
within each layer leads to negative aT/ay in the centre. w-plots in Fig. 4. In general, as Ra is increased w tends
A vorticity sink thus separates the regions cf to become comparatively small outside the vertical
concentrated vorticity generation and two secondary boundary layers, and is virtually zero over the central
760 N. C. MARKATCIS and K. A. PERICLEOUS

(b)

(e)

FIG. 5. Streamlines :(a)Ra = lo3 ;(b) Ra = lO’;(e)Ra = 10S;(d)Ra = 106;(e)Rn = 108;(f)Ra = 10’ c.
(9) Ra = 10”; (h) Ra = 10L6.
 Laminar and turbulent natural convection in an enclosed cavity 761

(a) (b)

(cl (d)

(g) (h)
-9 0
-8 7
A 7

6
6
J I

5
A 5
V f

F 4
r
4
V
I 3
X ,Y 2

FIG. 6. Isotherms: (a) Ra = 103;(b) Ra = 104;(c) Ra = 10’; (d) Ra = 106;(e) Ra = 10’; (f) Ra = 10”;
(g) Ra = 10” ; Ra = 1016. Contours at 283(2)303 K, except(f) which is at 288(1)298 K.
762 N. C. MARKATOS and K. A. PERICLEOUS

60”/;,ofthecavitywidthat Ra = 106.Themaximumand
minimum w values for each profile are strongly affected
bythe~anumberasshowninTable l.In this table(and
also in Table 2) the velocities are normalized using D as
a scale factor for length and D’/K as a scale factor for
time. The results obtained for Ra = lo6 were identical
whether the (k - E) model was used or not, due to
virtually zero k generation.
As the Ralyeigh number increases from lo6 to 10’.
the secondary vortices generated in the central core,
are convected further upstream and closer to the
differentially heated walls. The central vortex has now
diffused into the other two vortices which become a
dominant feature of the flow. The boundary layers on
the heated walls are now very thin. At the upstream
corner of each boundary layer, the low momentum
outer layer is absorbed by the adjacent vortex, while
the rest of the boundary layer follows the adiabatic
wall under the action of the secondary vortices.
For Ra = 10” the central vortex reappears [Fig.
5(f)] and the other two vortices are shifted closer to the
walls and get thinner and elongated. As Ra increases
further, the vortex system becomes progressively
weaker and eventually (Ra = 1016) disappears com-
pletely. High velocities now only occur within the thin
boundary layers, and the Row in the central core is
stratified. The streamline patterns of Fig. 5(h) show this
feature very clearly. Recirculation now only exists
within the isothermal layers of fluid and the sense of
rotation alternates between adjacent layers.
The accompanying temperature maps, for Ra = lo*
and LO’* show the temperature range in the core
dimillishing, from between 289 and 297 K to 291 and
295 K. A steep temperature gradient accompanies the
location of the secondary vortices. The horizontal
extent and magnitude of this gradient determines the 2 “0
extent and strength of the vortices described earlier. -z
c
The negative temperature gradient in the central vortex
is observed. For high Ra, the T-profiles have a very
steep slope within the thermal boundary layers; and
this becomes steeper as Ra increases. Outside the
boundary layers the T-profiles are almost horizontal “0
and temperature increases with z.
Surprisingly, as the Rayleigh number increases even
further, the temperature range in the core increases
againto between288and29~K(Ra = 1012)andfinally
to 287 and 299 K at Ra = 1016. This is due to weak
interchange of heat and momentum between adjacent
fluid layers as the flow becomes increasingly stratified.
At the same time, the temperature gradient in the
horizontal direction diminishes to zero. This is to be
expected, since now the walls are ‘too far away’ to have
any influence in the core (Ra CCD3, while 6 cc Dm3j4).
Figures 7 and 8 show horizontal and vertical velocity
contours at various Ra numbers. For Rayleigh
numbers up to lo6 these contours are normalized, but
not for the higher ones. The main observations
discussed so far can also be inferred from these figures,
i.e. the location and thickness of the boundary layers
and the location of reverse flow regions accompanying
 Laminar and turbulent natural convection in an enclosed cavity 163

Table 2. Comparison of the two solutions

Ra
lo3 lo4 lo5 lo6
I 2 1 2 1 2 1 2

“,,X 3.544 3.649 16.18 16.178 35.73 34.73 68.81 64.63

z 0.832 0.813 0.832 0.823 0.857 0.855 0.872 0.850
W,,X 3.593 3.691 19.44 19.617 69.08 68.59 221.8 217.36
4 0.168 0.178 0.113 0.119 0.067 0.066 0.0375 0.0379
NU 1.108 1.118 2.201 2.243 4.430 4.519 8.754 8.799
Nu,,, 1.496 1.505 3.482 3.528 7.626 7.717 17.872 17.925
0.092 0.1425 0.143 0.0825 0.081 0.0375 0.0378
‘Numi,, 0.0825
0.720 0.692 0.643 0.586 0.824 0.729 1.232 0.989
F 0.9925 1.0 0.9925 1.0 0.9925 1.0 0.9925 1.0

Note: (1) Present solution; (2) de Vahl Davis’ solution 131.

the secondary vortices. The vertical velocity maximum logarithmic in nature (neglecting the near-wall points
moves closer to the hot wall as the Rayleigh number where a logarithmic profile was imposed by the use of
increases, and it is close to the centre of the hot wall for the ‘wall functions’). This is in agreement with the
Rayleigh numbers up to lo6 (see also Table 1). The results of George and Capp [27].
maximum horizontal component also moves closer to Figure 12 presents w-velocity profiles within the
the hot wall as the Rayleigh number increases, and is narrow boundary layer near to the hot wall, for Ra
shifted upwards. = 10” with two different scaling factors. Three profiles
Figure 9 shows the resulting velocity fields. The are presented at three horizontal stations, z/D = 0.25,
boundary layer profiles and the diminishing velocity 0.50 and 0.75. Figure 13 presents the same information
field in the centre are clearly seen. An enlarged view of for Ra = 10”.
the central portion of Fig. 9(d) (indicated by the dotted The important quantities for the problem con-
line) shows clearly the presence and direction of the sidered, e.g. the Nusselt numbers and the maximum
three secondary vortices (Fig. 10). velocities are summarized in Table 1. The table shows
Figure 11 presents boundary layer velocity profiles at the calculated Nusselt numbers at the ‘hot’ wall.
the ‘hot’ wall for Ra = lo”, at various Y-locations. The The maximum and minimum values are given in the
boundary layer gets progressively thicker as we move table, together with the Nusselt number at j = y/D
from the bottom (IZ = 15, Z = 0.0075) up to z = 0.5. = 0.5. The same table, contains values of the maximum
Further up, Z = 0.795, negative flow is observed at the vertical velocity component on the horizontal mid-line
edge of the boundary layer, which is now thinner. The and its location, the maximum horizontal velocity on
reason for this was explained earlier in connection with the vertical mid-line and its location and the maximum
the presence of the secondary vortices. The boundary horizontal and vertical velocities over the whole
layer at Z = 0.9925 has almost disappeared as it domain and their location.
approaches separation at the top corner. Inspection of The heat transfer coefficient is seen to increase with
the detailed results indicates that both velocity Rayleigh number, as convection becomes dominant,
components are indeed negative at that corner. The but not as fast as the flow. Thus, passing from Ra = lo3
velocity profiles within the boundary layers are not to lo6 leads to an increase of w,,, from 3.59 to 221.8 ;

Table 3. Comparison of present and experimental correlations (arrows indicate range of validity of
equations used)

Nu (laminar) Nu (turbulent)
log Ra Equation (12) Equation (20) A% Equation (14) Equation (21) A%

3 1.128 1.102 + 2.4

4 2.246 1.960
. . . ...? + 14.6 1.698 1.546 +9.8
5 4.470 3.485 28.3 .~
+ 3.621 3.329 +8.8

6 1.723 7.167 + 7.2

8 .” .~. . . . . . . . . . . . ..t .., .., __
35.14 33.40 + 5.2

10 159.9 155.0 +3.1

12 727.5 720.0 + 1.0
 N. C. MARKATOS and K. A. PERICLFDUS

(a)

(e)

F1~.7.Horizontalvelocity(v)contours:(a)Ra = 103,contoursat -3.55(0.711)3.55;(b)Ra = 104,contoursat

- 16(3.2) 16;(c)Ra = 106,contoursat -43@62)43;(d)Ra = 106,contoursat - 117.88(23.58)117.88;(e)Ra
= 108,contoursat ~0.005,0.01,0.02,0.03,0.04,0.05;(f)Ra = lO”‘,contoursat +0.01,0.025,0.05,0.075.0.1.

but the maximum Nusselt number increases only from comparison is provided by the Nusselt numbers and the
1.496 to 17.87, despite the passage from diffusion- maximum velocity values in Table 2. Agreement is
dominated to convection-dominated transfer, as generally good. Differences exist in the minimum
revealed by the configuration of the isotherms. Nusselt number, particularly at Ra = 106. Although
The maximum Nusselt number occurs at the bottom the agreement of the maximum and centre-line Nusselt
of the cavity and the minimum at the top. numbers is better than 1.5% over the whole Ra range,
The results for Rayleigh numbers up to lo6 are the present predictions indicate a higher minimum
compared with the benchmark solution provided inref. Nusselt number than the benchmark solution.
[3] which can be considered as accurate. The Numerical errors and the first-order differencing
qualitative agreement of the presented plots of stream scheme used in this work do not account for the
function, temperature and velocity maps with those of observed differences, since the results are practically
the benchmark solution is very good. Quantitative grid independent. However, there are two other
 Laminar and turbulent natural convection in an enclosed cavity 765

(cl)
Y

(d)

i) :
‘.‘* ,.!
,/
,.-’
;“I
‘:’

12

(Q)
‘i A 7

0 *!
6

FIG. 8. Vertical velocity(w) contours: (a) Ra = lo’, contours at - 3.555(0.71I)3.5.55;(b) Ra = 104,contours at

- 19.24(3.848)19.24; (c) Ra = 105, contours at -67.55(13.51)67.55; (d) Ra = 106, contours at -217.38
(43.47)217.38;(e)Rn = lOs,contoursat ~0.0075,0.0125,0.02,0.03,0.04,0.05,0.06;(f)Rn = lo”,- 20,
---- <O,contoursat +0.01,0.025,0.05,0.075,0.1,0.15;(g) Ra = 10IZ,contours at +0.01,0.025,0.05,0.1,0.2.
 , I s . ‘

, * . . 1 ,

, I 6 f

, . . . a ,

I ‘ . . I ‘

I 1 . . I , , ,

t--r,,

/,- \

I --\\\----‘\ 1
“, \ ,“‘I/!
 Laminar and turbulent natural convection in an enclosed cavity 767

0 Oi 02 03 04

w (m s-‘1

FIG. 1 I. Boundary layer velocity profiles at Ra = 10xz,

‘hot’ wall.

differences between the procedures, namely the example that at Ru = lo6 the solution predicts a near-
elimination of the Boussinesq assumption in the wall maximum temperature of 302.81 K instead of the
present work and the implied use of the linear velocity correct value of, say, 302.84 K. This represents a very
profile between wall and near-wall points, in the shear- accurate temperature prediction with an error of less
stress calculation. The latter should not account for any than 0.01% (well within the convergence criterion used,
difference either, provided that the benchmark solution see Section 5). For a wall temperature of 303 K, the
has properly computed the wall shear-stresses. application of a one-sided formula for calculating a
The computations were repeated by using the temperature gradient at the wall, would lead to a NQ,,
Boussinesq approximation. Minor differences were which for the present grid spacing would be 1.267
observed for Ra = 103and 101,with a maximum of 1%. instead of 1.067, e.g. in error by 19%. The same is not
For Ra = lo5 and lo6 the differences were up to 3.7%; true in the case of Nu,,,. Thus the same accuracy in
the minimum Nusselt number being 0.804 at Ru = 10’ temperature (say 300.43 K instead of 300.4 K) leads to
and 1.188 at Ru = 106. The large errors in the Nu,,, = 17.13insteadof 17.33,e.g.inerror by only 1%.
calculation of the minimum Nusselt number appear to Therefore, despite the large errors at Ra = lo6 in the
be due to the sensitivity of its calculation. Suppose for derived quantities, it is suggested that the present

w (m 5’)

0 02 03 0 4

FIG. 12. w-Velocity profiles within boundary layer, Ra = 10” FIG. 13. w-Velocity profiles within boundary Iayer, Ra = lOi
(two scaling factors). Three horizontal stations at: (I) z/D (two scaling factors). Three horizontal stations at: (1) z/D
= 0.25; (2) z/D = 0.50; (3) z/D = 0.75. = 0.25; (2) zJD= 0.50;
(3) z/D = 0.75.
768 N. C. MARKATOS and K. A. PERICLEOUS

solution may still be about 1% accurate in terms of the It can be seen that extrapolation of the experimental
computed primitive variables. turbulence correlation up to Rayleigh numbers of 10”
Direct comparisons are not possible for the is in good agreement with the present predictions,
turbulence runs, due to the scarcity of experimental despite the fact that the former is applicable to cavities
evidence. However, the use of the k - E model in other of aspect ratios between 3 and 42. The aspect-ratio
related work (e.g. buoyant turbulent flow in buildings) etfect explains the relatively large differences observed
has led to favourable comparisons with experiment for the lower Ra numbers, and the closer agreement
Cl5]. obtained for higher Ra numbers; the effect of aspect
Furthermore, the presented solutions certainly show ratio diminishes as Ra increases.
some of the features that are observed experimentally The switch in the exponent of the Rayleigh number
(position of vortices and their shifting with increasing for the correlations for Ra > 10” was introduced
Ra number, thermal stratification, etc.) and indicate, simply by requiring a zero error from the least-square
qualitatively at least, the correct change in flow linear regression. Experimental evidence, however
structure as Ra increases. limited, indicates an exponent of l/3. If the exponent of
Finally, the correlations of the present results given Ra is assumed to be l/3, then the best correlation of the
below are in good agreement with well-known present results over the whole of the turbulent range
correlations derived from a large number of gives
experiments for Ra up to 1O8 [23], and appear to extend -
the validity of those correlations up to Ra = 10”. For Nu = 0.060Ra”3. (22)
higher Ra a set of new correlations is proposed. The above correlation gives higher values than those
obtainedbyMacGregorandEmery(Nu = 0.046Ra1’3)
4.1. Derived Nu - Ra correlations -
The following correlations for maximum, minimum [28] and by Cowan et al. (Nu = 0.043 Ra”3) [26].
and average Nusselt numbers were derived from the However, the former [28] refers to cavities of aspect
present predictions by least-square linear regression. ratio lo-40 with constant heat flux boundary
Laminar (lo3 < Ra < 106) conditions at the hot face, and isothermal cold face and
the latter [26] refers to water and to Ra numbers up to 2
NM1,z = Nu = 0.143Ra0.299; (12) x 10”.

Nu,,, = 0.130Ra0~356. (13)

The minimum Nu is not given as the log Nu vs log Ra
curve is not .a straight line. 5. CONVERGENCE AND COMPUTER TIME
Turbulent (lo6 < Ra ,< 10”)
A converged solution was defined as one that met the
Nu = 0.082Ra0.329 ; following criterion for all dependent variables
(14)
Nu,,, = 0.057Ra0.37g; max I@+’ -@I < 10-4,
(15)
Nu,,,~, = 0.016Ra0.3’5. between sweeps n and n+ 1. At this stage mass
(16)
continuity errors per slab were of the order of lo- ’ and
Turbulent (10” < Ra < 1016) therefore insignificant. Further sweeps of the solution
domain confirmed no changes and were actually
Nu = 1.325Ra0.245; (17)
pointless since the accuracy limit of the computer was
Num,, = 0.34Ra0,31 6 ; (18) approached. The sum of the absolute volumetric error
over the whole field was again insignificant, at 10m6.
Nu In,” = 0.137Ra0.233. (19)
Convergence was found to be affected by the
Rayleigh number. Hence, the high Rayleigh number
4.2. Comparison with experimental correlations
cases required more sweeps than the low ones. To
The above correlations are compared with the
improve convergence, a false time-step relaxation was
following experimental ones in Table 3 [23] :
used. This was reduced by an order of magnitude, from
laminar 0.1 to 0.01 for both velocities, for Ra > lo’, and to 0.001
- 119 for Ra > 10”. This is to be expected since this
Nu = 0.196Ra’i4 k relaxation factor is proportional to the fluid residence
0 time in a typical cell; and it is therefore appropriate
(2.8 x lo3 < Ra < 2.8 x 104); (20) to reduce it as the velocity in the boundary layers in-
creases with Rayleigh number.
turbulent
To economize on run-times each Rayleigh number
case was restarted from the previous one (starting with
Ra = 103) and 200 sweeps were found adequate for
convergence. The domain dimension D was the
(2.8 x lo4 < Ra < 1.55 x 107); (21)
parameter adjusted to get the required Rayleigh
where L/D is the cavity aspect ratio. number value.
 Laminar and turbulent natural convection in an enclosed cavity

04-

‘x-
01
0 02xw 0 4x 10-I 0.6 x 10-l

w (m s-l)

FIG. 14. Velocity comparison (Ra = 106)for 40 x 40 and 80 x 80 solutions.

A 100~sweep run required 11 min CPU time on the The grid was refined again to 80 x 80 by halving all
Perkin-Elmer 3220 mini-computer, for the 30 cells and the Ra = lo6 case was run for another 300
x 30 grid, and 15 min for a 40 x 40grid. The above mini- sweeps. Very minor changes were observed, as shown in
computer is several times slower than mainframe Fig. 14, for the velocity profile inside the hot wall
machines. In general, the time per finite-domain cell, boundary layer. The maximum velocity increased by
per sweep, per variable was 1.5 x 10m3 s CPU on the 0.4’/” in the Ra = lo6 case. The same observations
Perkin-Elmer machine. A full run at Ra = 1014 apply also to the temperature-field changes with grid
using a 60 x 60 grid would require 300 sweeps and take refinement. It is concluded that the results up to Ra
2.5 h on the mini-computer for solving the six equations = lo6 are practically grid-independent.
(0, W,P, k, E, ‘0. For the turbulent cases care was taken to place l&l 5
It is not an easy matter to compare the above grid cells inside the wall boundary layers. The thickness
reported CPU times with those of other methods in the of these layers is given by 6/D = 4.86Ra-‘I4 for Pr
literature, because of different computers used and = 0.71. Care was also taken to concentrate several of
different practices in reporting these times. these 10-15 points between the velocity peaks and the
However, it appears to the authors after their survey wall, and to place the first grid point very close to the
that the present computer-time requirements are wall (y/D = 1 x 10m6 for the high Ra numbers, y+
modest, and possibly an order of magnitude less than varying between 1 and 12 along the hot wall). The above
those reported in the literature for the same cases. If considerations are very important for treating the
true, this is certainly due to thespeed ofthe SIMPLEST boundary layers in accordance with the essential
algorithm and the associated whole-field pressure physics, since in the wall region temperature gradients
solver, coupled with the thorough optimization of are most severe and hence provide an important source
FORTRAN arithmetic and the very orderly bookkeep- of vorticity. If the first grid point is too far from the wall
ing in the software, embodying the solution procedure. then the results will be grossly distorted by this effect.
Three grids were used,(y, z) = 40 x 80,60 x 120 and 100
x 160. Practically grid-independent results seemed to
6. GRID DEPENDENCE AND COMPUTER STORAGE
be obtained using a grid of (y, z) = 60 x 120 for all
Initial investigations were performed on a uniform Ra > 10s; the 100 x 160 grid leading to a maximum
30 x 30 grid. This was found to be adequate for Ra further change of 1.2% for the maximum velocities and
= 103, but not for higher values. An improved 30 x 30 1.5% for temperature. For Ra > 1014 the resulting grid
grid was then used, with closer spacing near the walls cells in the core were thought to be too ill-conditioned
(Fig. l), to increase boundary layer resolution. (aspect ratio of up to 10: 1) for sufficient accuracy.
This grid was used for all Rayleigh number cases up However, the flow in the core for these high Ra numbers
to Ra = lo6 and the results were stored. is nearly stagnant and the cell aspect-ratio effect
To check grid dependency, the grid was further appears not to be important, as was indicated by
refined to 40 x 40 by adding intermediate cells in the repeating the runs with half the above ratios.
central region. All cases (up to Ra = 106) were re-run The program required 90 K-words ; of these, 35 K-
restarting from interpolation of the 30 x 30 results, for words were required for data storage (30 x 30 grid).
another 100 sweeps. In their bulk, the results showed
little change, although the secondary vortices at Ra
7. CONCLUSIONS
= lo5 and lo6 were, as expected, better resolved. The
maximum velocity increased by 5% in the Ra = 10J The problem considered represents a 2-D approxi-
case, and these are the results presented in this paper. mation to a series of practical problems. The study
770 N. C. MARKATOS and K. A. PERICLEOUS

demonstrated that numerical solutions can be obtained 4. 1. Catton. Natural convection in enclosures. in Heat
quickly and economically for such problems, where Transfer 1978, Vol. 6. National Research Council of
Canada (1978).
buoyancy effects are dominant.
5 S. Ostrach, Completely confined natural convection.
Accurate results were obtained at both ends of the Developments in Mechanics, Proc. Tenth Midwestern
Rayleigh number scale, at least for the laminar range, Mechanics Co@, Vol. 4. pp. 53-81. Johnson, Chicago.
indicating that both diffusion-only or convection-only Illinois (1967).
heat transfer problems can be tackled with ease. The 6. S. Ostrach, Natural convection in enclosures. in Advances
in Heat Transfer (edited by J. P. Hartnett and T. F. Irvine.
results were presented in graphical and tabular form,
Jr.), Vol. 8, pp. 161-227. Academic Press, New York
and as correlations between the Nusselt and Rayleigh (1972).
numbers. 7. S. Ostrach and R. G. Hantman, Natural convection inside
The main points of the present method can be a horizontal cylinder, Chem. Enyny Commun. 9,213-243
(1981).
summarized as follows: (a) the model consists of the
8. S. Ostrach, Laminar flows with body forces. in Theory of
coupled differential elliptic equations which are solved Laminar F/ows(edited by F. K. Moore). Vol. 4. High Speed
over the entire flow domain, with no assumption Aerodynamics and Jet Propulsion, Princeton University
concerning the core configuration ; (b) the procedure is Press, Princeton, New Jersey (1964).
strongly convergent and results were easily obtained 9. S. Ostrach, Low-gravity fluid flows, Annual Review of
Fluid Mechanics, Vol. 14 (1982).
even at Rayleigh numbers as high as 10’ 6 ;(c) practically 10. B. E. Boyack and D. W. Kearny, Heat transfer by laminar
grid independent results were obtained with only natural convection for low aspect ratio cavities, ASME
modest computer storage and CPU time requirements. Paper No. 72-HT-52 (1972).
Indeed, a survey revealed that the present method may 11. S. Ostrach, Natural convection heat transfer in cavities
and cells, Proc. Int. Heat Transfer Conf. pp. 3655379.
be up to an order of magnitude faster than other
Hemisphere, Washington. DC (1982).
available procedures, enabling very fine grid runs to be 12. B. E. Launder and D. B. Spalding. The numerical
performed within practicable resources ; (d) the computation of turbulent flows, Comp. Methods Appl.
procedure and associated computer program are Me&. Engng 3,269 (1974).
general and can be used immediately for 3-D natural 13. N. C. Markatos. Transient flow and heat transfer ofhauid
sodium coolant in the outlet plenum of fast nuclear
convection problems in cavities of any aspect ratio and
reactors, Int. J. Heat Mass Transfer 21, 1565 (1978).
orientation, and fluids of any Prandtl number. It is 14. W. Rodi, Turbulent models and their application in
therefore immediately applicable to all related practical hydraulics-a state of the art review, SFB 80/T/127,
problems ; (e) the speculative use of the (k - E)model in University of Karlsruhe (1978).
15. N. C. Markatos, M. R. Malin and G. Cox, Mathematical
this work has indicated that, despite its well-known
modelling of buoyancy induced smoke flow in enclosures,
deficiencies in terms ofphysical realism, it may still lead Int. J. Heat Mass Transfer 25, 63-75 (1982).
to a reasonable prediction of the overall flow structure 16. D. B. Spalding. A general purpose computer program for
of the problems considered; (f) apart from the multi-dimensional one- and two-phase flow. in Mathe-
uncertainty connected with the (k - E) model itself matics and Computers in Simulation, Vol. XXIII, pp. 267-
276. North Holland, Amsterdam (1981).
another source of uncertainty is provided by the ‘wall- 17. D. B. Spalding, Mathematical modelhng of fluid-
functions’. More work is required to establish more mechanics, heat-transfer and chemical-reaction pro-
realistic ‘wall-functions’ for buoyancy-dominated cesses, A Lecture Course, CFDU Report. HTS/80/1,
flows. Once established, it is a very simple matter to Imperial College, London (1980).
18. N. C. Markatos, P. J. Phelps and B. Purslow, Computer
incorporate in the present model.
simulation of the thermal-hydraulic behaviour of fast-
Although only time-averaged steady-state results reactor pools. Ann. Nucl. Energy 9, 179-193 (1982).
have been presented, the procedure can also be used in 19. S. V. Patankar and D. B. Spalding, A calculation
its in-built transient mode. Also radiation and variable procedure for heat, mass and momentum transfer in three-
property effects, that have been neglected in the present dimensional parabolic flows, Int. J. Heat Mass Transfer
15, 1787 (1972).
study, must be included to investigate their im- 20. I. P. Jones and C. P. Thompson, Numerical solutions for a
portance at high Rayleigh numbers. Finally, much comparison problem on natural convection in an
more experimental work is required for high Ra enclosed cavity, Report No. AERE-R 9955, UKAEA
numbers to provide data for improving and validating Harwell(1981).
21. G. D. Mallinson and K. A. Pericleous. GRAFFIC: the
the computer models.
graphical representation and analysis of fluid flow by
interactive computation, CHAM Technical Report,
REFERENCES TR/71. June (1981).
22. G. D. Mallinson and G. de Vahl Davis. Three-
1. I. P. Jones. A
comparison problem for numerical methods dimensional natural convection in a box: a numerical
in fluid dynamics: the ‘double-glazing’ problem, in studv. J. Fluid Mech. 83. l-31 (1977).
Numerical Methods in Thermal Problems(edited by R. W. 23. M. Jacob, Heat Transfer, Vol. 1. Wiley, New York (1949).
Lewis and K.-t, ~33%348. Pineridge Press, 24. M. S. Hossain and W. Rodi, Influence of buoyancy on the
Swansea. UK. (1979). turbulence intensities in horizontal and vertical jets, in
2. G. de Vahl Davis and I. P. Jones, Natural convection in a Heat Transfer and Turbulent Buoyant Convection (edited
square cavity-a comparison exercise, Int. J. Num. by D. B. Spalding and N. Afgan), Vol. 2, pp. 39-51.
Methods Fluids 3, 2277248 (1983). Hemisphere, Washington, DC (1976).
3. G. de Vahl Davis, Natural convection of air in a square 25. J. W. Elder, Turbulent free convection in a vertical slot. J.
cavity: a benchmark numerical solution, Int. J. Num. Fluid Mech. 23,99-l 11(1965).
Methods Fluids 3, 249-264 (1983). 26. G. H. Cowan, P. C. Lovegrove and G. L. Quarini.
 Laminar and turbulent natural convection in an enclosed cavity 111

Turbulent natural convection heat transfer in vertical functions intersect) the above laminar wall-function was also
single water-filled cavities, Proc. 7th Int. Heat Transfer used. If the Reynolds number was greater than 132.25, the
Conf, Vol. 2, pp. 195-201 (1982). presumed velocity variation was logarithmic, and the
21 W. K. George and S. P. Capp, A theory for natural corresponding shear-stress coefficient was evaluated. Both k
convection turbulent boundary layers next to heated and Ewere fixed at the near-wall grid nodes at the values which
vertical surfaces, Int. J. Heat Mass Transji~ 22,813-826 would prevail there ifindeed the universal logarithmic velocity
(1979). profile prevailed. The wall heat transfer rate was evaluated
28 R. K. MacGregor and A. F. Emery, Free convection from the Chilton-Colburn form of the Reynolds analogy, in
through vertical plane layers-moderate and high which the Stanton number (St) is related to the friction
Prandtl number fluids, Am. Sot. Mech. Enyrs, Series C, J. coefficient (C,) as follows :
Heat Transfer 91, 391403 (1969).
St = C, Prmzi3, (Al)
29 A. S. Monin and A. M. Obukov, Basic laws of turbulent
mixing in the ground layer of the atmosphere, Trudy.
where C, is related to the wall shear stress (7,) and the parallel-
Geqfiz. Inst. An-SSR 151(24), 163-187 (1954).
to-wall velocity /WIas follows:
APPENDIX
C, = r& I#). (A2)
THE WALL BOUNDARY CONDITIONS
The heat transfer rate per unit area at the wall (4;) is then
The treatment of wall boundary conditions for turbulent deduced from :
buoyant flows presents a source of uncertainty in the present
work. It is therefore worthwhile describing briefly what was 4; = St plWC,(T,- T.), (A3)
done. ‘Wall functions’ were used for both the laminar and
turbulent calculations. In the laminar cases this simply meant where T, is the temperature at thegrid node in question, and T,
that the wall shear stress was evaluated from the presumption the temperature at the wall.
of a linear velocity variation between the calculated near-wall It is realized that the wall functions may be influenced
value and the zero wall value. considerably by temperature variations, and indeed that the
The shear stress calculated in this way was then included as a logarithmic form used is probably not appropriate since
source term for the velocity component parallel to the wall. buoyancy effects are ignored in the log layer. A better wall
For the turbulent calculations, the Reynolds number for the function could be based on either the Monin-Obukov log-
near-wall point was first evaluated. If this was less than 132.25 linear profile 1291, or on the cube root profiles of George and
(y’ = 11.5, the value at which the laminar and turbulent wall- Capp ~271.

CONVECTION LAMINAIRE ET TURBULENTE DANS UNE CAVITE FERMEE

Resume---Onpresente unemethodenumeriquepour obtenirdes solutionsd’ecoulement nature1 et de transfer1

de chaleur dans une cavite carree avec des parois laterales chauffees differemment. On Btudie un domaine de
nombres de Rayleigh entre lo3 et 106. On utilise une differentiation de cellules donatrices et des etudes de
maillage sont developpees pour tousles nombres de Rayleigh consider&. Le modele de turbulence utilist pour
les nombres de Rayleigh superieurs a lo6 est un modele (k--E) i deux equations qui inclut les interactions
gravite-gradient de masse volumique. Les resultats sont present& sous forme de tables et de graphiques et de
formules de nombres de Nusselt et de Rayleigh. En outre, les resultats de nombres de Rayleighjusqu’a lo6 sont
compares avec la solution numerique de Vahl Davis.

LAMINARE UND TURBULENTE FREIE KONVEKTION IN EINEM GESCHLOSSENEN

HOHLRAUM

Zusammenfassung-Es wird eine Berechnungsmethode beschrieben, die dazu dient, Losungen fur die
laminare und turbulente, von Auftriebskraften bestimmte Strbmung und den Warmelbergang in einem
Hohlraum mit quadratischem Querschnitt und unterschiedlich beheizten Seitenwinden zu erhalten. Die
Rayleigh-Zahl wurde im Bereich von 10’ bis lo6 variiert. Es werden “donor-cell”-Differenzen verwendet.
Einfliisse der Gitterverfeinerung wurden bei allen betrachteten Rayleigh-Zahlen untersucht. Als
Turbulenzmodell fiir Rayleigh-Zahlen griiBer lo6 wurde ein (k N &)-Model1 verwendet, welches
Wechselwirkungen zwischen Schwerkraft und Dichtegradienten beriicksichtigt. Die Ergebnisse werden in
tabellarischer und grafischer Form und als Korrelationen von Nusselt- und Rayleigh-Zahlen dargestellt. Die
Ergebnisse fiir Rayleigh-Zahlen bis lo6 werden mit den Referenz-Losungen von de Vahl Davis verglichen.
112 N. C. MARKATOS and K. A. PERICLEOUS

.JlAMMHAPHAIl M TYP6YJIEHTHAR EmECTBEHHAII KOHBEKUMII

B 3AMKHYTOR IlOJlOCll4

AHHOTa,@,n--OIIBCbIBaeTCB WCneHHbIfi MeTon pe”IeHHn 3ana’I CBO60JlHOKOHBeKTBBHOrO naMNHapHOr0

A Typ6yneHTHoro TeqeHm H TennonepeHoca B KBaflpaTH0i-I IIOnOCTB C 60KOBbIMH CTeHKaMB,
HarpeTbv.48 no pa3HoB TemepaTypbI. MccnenoeaHm npoeonsnecb B mianasose 3HaveHaB mcna
Penes OT lo3 a.0 1oL6. npH nOCTpOeHHH pa3HOCTHOfi CXeMbI liCnOnb30BaHbI ,.,OHOpCKIte WIeiiKW.
,Qna BCCX paccMaTpwBaeMblx 3HaneHwii wicna Penea npoeonenmb m22nenoBaHm Emimm A3MenbqeHm
ce~w Ha pemeaue. npa 3HaveHmx wicda Penea, npeBbImamUix 106, wnonb3oBanacb neyxnapa-
MeTpmecKar (k s E) MOnenb Typ6yneHTHOCW yWiTbIBafOIUaK BsaaMoneficrme Meany canok
TIlTeCTR w rpanrteHToM NIOTHOCTW. Pe3ynbTaTbI npencra9neHbI B mine Ta6nw A rpa@iKoB, a
TaKme B Bltfle 0606IqeHHbIX 3aBHWMOCTei-4 Mexny ‘iHCnaMI( HyCCWIbTa U PeneSI. KpoMe TOTO,
npoaeneH0 CpaBHemie pe3ynbTaTos. nonyveHHbIx npe wicnax Penen, MeHbuwix 106. c WCJIeHHbIM
pemeH&ieM ne Baanb flaBma.