You are on page 1of 16

Journal of Wind Engineering

and Industrial Aerodynamics 90 (2002) 11911206

Large eddy simulation of the ow past a


circular cylinder at ReD=3900
J. Franke*, W. Frank
Institute of Fluid- and Thermodynamics, University of Siegen, Paul-Bonatz-Strasse 9-11,
D-57076 Siegen, Germany

Abstract

A large eddy simulation (LES) is performed for the turbulent ow around a circular cylinder
at ReD 3900 with a cell-centered nite volume code that solves the compressible Navier
Stokes equations. The results are compared with the direct numerical simulation of Ma et al.
(J. Fluid Mech. 410 (2000) 29) and the experiments of Ong and Wallace (Exp. Fluids 20 (1996)
441). It is shown that short averaging times, which have been used by several previous LES do
not lead to converged mean values. For the largest averaging time used in this study, the
results are in good agreement with the aforementioned data. In addition, the realizability of
the modeled subgrid scale stresses and the computed Reynolds stresses is analyzed. r 2002
Elsevier Science Ltd. All rights reserved.

Keywords: Large eddy simulation; Turbulent ow; Circular cylinder; Realizability

1. Introduction

The interest of wind engineers in the turbulent ow around bluff bodies is due to
several questions. The pressure distribution on the bodies is required in the design
process with regard to failure under wind loading, while knowledge of the velocities
around the bodies is necessary for the assessment of, e.g., pedestrian comfort and
pollutant dispersion. Research on this uid mechanical problem is conducted by
means of experiments and, increasingly, by numerical simulation. The latter is
mainly based on the NavierStokes equations as mathematical model of the physics
of turbulent uid ow. In wind engineering the two predominant approaches for
their numerical solution are the Reynolds averaged simulation (RAS) and the large

*Corresponding author.
E-mail address: franke@ift.mb.uni-siegen.de (J. Franke).

0167-6105/02/$ - see front matter r 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 1 6 7 - 6 1 0 5 ( 0 2 ) 0 0 2 3 2 - 5
1192 J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206

eddy simulation (LES). Both methods solve the averaged NavierStokes equations,
where the average is dened as a statistical or temporal average in RAS and as a
spatial average over a small volume in LES. Due to the non-linearity of the Navier
Stokes equations their averaged forms contain additional terms for which models are
needed. They are termed turbulence models in RAS and subgrid or sublter scale
models in LES. Independent of their respective formulation and the underlying
simplications of the physics, these models should lead to uctuations that do not
violate the physics of turbulence.
Whether a computed ow eld is a physical ow eld can be conveniently
analyzed by evaluating in-equalities that the modeled terms have to fulll. These
realizability conditions have been rst formulated by Du Vachat [1] and Schumann
[2] for the Reynolds stress tensor, which has to be modeled in an RAS. They have
been used in the derivation and analysis of turbulence models and it has been shown
that models complying with the realizability conditions do not necessarily lead to
better quantitative results [3,4]. However, violation of the realizability conditions
clearly indicates that the predictions are in error. Therefore, we are of the opinion
that the realizability of the Reynolds stress tensor should be always checked for an
RAS to identify results that violate the fundamental physics of turbulence,
especially, if no experiments are available for comparison.
These statements apply also to the time-averaged results of LES. As this
computational approach solves the unsteady ow eld together with a model for the
subgrid stresses, realizability also has to be fullled by the instantaneous subgrid
stress tensor, as has been shown by Vreman et al. [5] and Fureby and Tabor [6]. The
latter examined the realizability of the LES of forced homogeneous isotropic
turbulence and a turbulent channel ow. In both cases they found a violation of the
realizability condition in a small part of their computational domain. However, their
time-averaged results did not suffer from the non-realizability of the instantaneous
ow elds. For LES around bluff bodies no examination of realizability has been
reported.
Therefore we will present among other results, which are detailed in Section 1.2, a
methodology for the extensive analysis of realizability in LES for the ow past a
circular cylinder. This idealized bluff body is of practical relevance for wind
engineering with regards to the ow around towers, chimneys, cables or bridge piers,
to name a few. In addition, it has been investigated extensively by experiments and
numerical simulations for the Reynolds number used in the present LES, as is
summarized in the following section.

1.1. Previous computational results

The ow around a circular cylinder at a subcritical Reynolds number, based on


cylinder diameter D and freestream velocity U0 ; has become a benchmark for LES
on the way towards their application for complex ows of technical relevance [7].
While the geometry of the ow is still simple, the ow phenomena are complex,
including laminar separation with no xed separation point, transition to turbulence
in the thin shear layers, that are separating, and shedding of large scale vortices [8].
J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206 1193

Due to the availability of measurements of the mean oweld in the very near wake
(x=Dp3) by Lourenco and Shih [9] and in the near wake (3Xx=Dp10) by Ong and
Wallace [10], the ow at ReD 3900 has been studied numerically, beginning with
the work of Beaudan and Moin [11]. They solved the compressible NavierStokes
equations on an O-grid with a high order upwind scheme, which was shown to be
very dissipative when the grid coarsens in the wake. Therefore, Mittal and Moin [12]
performed the LES with the incompressible equations on a C-mesh with a central
difference scheme of second order. While their results for the mean oweld did not
differ much from the ones of Beaudan and Moin, their power spectra in the near
wake were in better agreement with the experiments [10]. Even better agreement for
the spectra was obtained by Kravchenko and Moin [13], who solved the
incompressible equations with a high order method based on B-splines on an
O-grid with zonal renement. They also showed that inadequate grid resolution of
the separating thin shear layers leads to a shorter length of the mean separation
region, dened by zero mean velocity on the wake centerline. Their high-resolution
LES resulted in a larger recirculation length than the one measured by Lourenco and
Shih, but in agreement with the results of the previously mentioned LES. In contrast
to those ndings, the LES of Breuer [14] and Frohlich . et al. [15] predict a
recirculation length, which is shorter than the one measured by Lourenco and Shih.
They used a nite volume method with central differences to solve the incompressible
equations on an O-grid. In agreement with the resolution studies of Kravchenko and
Moin, Breuer also found that doubling the cylinder span from pD to 2pD and
keeping the spanwise resolution does not change the results much.
To clarify the discrepancies between the LES results and the experimental data of
Lourenco and Shih, Ma et al. [16] performed direct numerical simulations (DNS)
with a spectral nite element method, solving the incompressible equations in a box-
shaped domain. They showed that two converged states of the oweld in the very
near wake exist, that are related to shear layer transition and depend on the spanwise
extent of the computational domain. For a domain size of 2pD; a short recirculation
length is obtained, while a domain size of pD; which has been used in all
aforementioned LES, leads to a long recirculation length, when the same spanwise
resolution is retained. They also conrmed the previously cited inuence of
the resolution on the mean oweld and in addition showed with their LES that the
recirculation length increases with increasing the constant in the xed constant
Smagorinsky subgrid scale model, i.e. with increasing subgrid viscosity.
Another DNS was performed by Tremblay et al. [17]. They also used a box-shaped
computational domain with spanwise extent of pD; in which they solved the
incompressible equations with a nite volume method and central differences on a
cartesian grid. They also obtained a larger recirculation length than in the
experiment of Lourenco and Shih, but nearly 20% shorter than the one of Ma
et al. for the corresponding spanwise size. So from the numerical simulations also
one nds that the ow around a circular cylinder at ReD 3900 is very sensitive to
the boundary conditions, the cylinder span and small disturbances, caused by
insufcient resolution and/or increased viscosity. This is in accordance with
experimental results, which are known to be very sensitive to aspect ratio, blockage
1194 J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206

effect, end conditions, turbulence level, etc. Therefore, we will compare our results
mainly with the results of Ma et al., for reasons explained in Section 3.1.

1.2. Objectives

The main objective of the current study is the validation of our nite volume code,
which solves the compressible NavierStokes equations with a high-resolution
central scheme with explicitly added articial viscosity and employs the subgrid
scale model of Yoshizawa [18], by a comparison with the DNS of Ma et al. and
the experiment of Ong and Wallace. Thereby we investigate the inuence of the
averaging time on the mean oweld, as the LES of Moin and co-workers used much
.
shorter averaging times than the differing results of Breuer and Frohlich et al. In
addition, we examine whether the instantaneous owelds represent physically
possible realizations, as only those are allowed to be used in ensemble or time
averaging, in principle. We use the realizability conditions as a criterion to judge
whether the oweld is physical or not. The aforementioned formulation of Du
Vachat [1] and Schumann [2] was extended to the subgrid stress tensor of an
incompressible uid by Vreman et al. [5]. Fureby and Tabor [6] then derived the
corresponding inequalities for the subgrid stress tensor in the general compressible
case. We will use this formulation in the present analysis, which we consider as
essential with regard to the aim of performing LES for complex industrial
applications, which was mentioned before. The computed owelds being physical
possible owelds is the main prerequisite, if the results of LES shall be used as data
eventually. Finally, the realizability of the Reynolds stress tensor of the time-
averaged results is checked.

2. Numerical method

2.1. Governing equations

As has been already stated in Section 1.2, the equations solved are the
compressible NavierStokes equations. In ltered form the continuity, momentum
and energy equation are:
qr%
% v 0;
r  r* 1
qt
%v
qr*
% vv*  r  T# r  Tsgs 0;
r  r* 2
qt
qr% E* #  v* r  q# r  q 0:
*  r  T
% vE
r  r* sgs 3
qt
Here the tilde denotes Favre-ltered, i.e. density weighted variables, e.g. v* rv=r: % E*
* *
is the ltered total energy of an ideal gas, E cV T 0:5*v  v* ksgs ; where ksgs is the
kinetic energy of the subgrid scales and T* the ltered temperature. As Lesieur and
J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206 1195

Comte [19] have recently shown, ksgs is small for air and can be neglected. But as the
evaluation of ksgs is necessary for the analysis of realizability, it is kept in the present
formulation. Therefore, the ltered pressure p% is obtained from r% and T* via the ideal
* where R is the specic gas constant of air. p% is part of the stress
gas law, p% Rr% T;
tensor of a Newtonian uid:
T# pI
% m *vr *vr  23r  v* ;
T
4
|{z}

S#

where 2S# is the deviator of the deformation tensor. The formula of Sutherland is
used to determine the dynamic viscosity of air as a function of T: * The heat ux vector
in Eq. (3) follows from Fouriers law, q# mTc * p =Pr rT;
* where the conductivity is
expressed in terms of mT * and the Prandtl number Pr; for which the value of 0.71 is
used.
In Eqs. (2) and (3) only the subgrid terms arising from ltering the convective
uxes have been retained. In principle more subgrid terms arise when ltering the
compressible NavierStokes equations, due to the nonlinearity of the viscous uxes
and the presence of the energy equation [20]. For ow at low Mach and high
Reynolds numbers these terms are normally neglected [21,19]. Therefore, only
models for the subgrid heat ux qsgs rE pv  r% E* p*% v and the subgrid stress
tensor Tsgs rvv  r*% vv* are necessary to close the system of equations. For Tsgs the
eddy viscosity model of Yoshizawa [18] is used:
q
Tsgs p 2 # 2 # 2
% 3ksgs  vsgs S; ksgs Ck D I2 S; vsgs CS D I2 S # ; 5
# is the quadratic invariant of S# (cf. Eq. (7)). The lter width D of the
where I2 S
implicit box-lter is obtained from the volume V of the computational cells,
multiplied by the damping function of Van Driest to account for the damping of the
subgrid stresses near walls, D V 1=3 1  expy =253 1=2 : For the constant
coefcients in Eq. (5) the values of CS 0:1 and Ck 0:0886 are used.
For this model to be realizable, it should have the same mathematical and physical
properties as the exact subgrid scale stresses. The latter form a positive semi-denite
tensor, as has been shown by Vreman et al. [5] for the incompressible case, and by
Fureby and Tabor [6] for the compressible case. Therefore, the three invariants of
Tsgs have to obey
I1 Tsgs trTsgs X0; 6

I2 Tsgs 12 I12 Tsgs  I1 T2sgs X0; 7

I3 Tsgs detTsgs X0: 8


From Eqs. (7) and (8) two inequalities for ksgs and the subgrid viscosity vsgs can be
derived, leading to a lower bound for ksgs and an upper bound for vsgs [6]. The former
leads to an inequality for the model constants Ck and CS [5], which is fullled by the
present choice, although CS has been reduced from 0.16, originally proposed by
Yoshizawa [18], to 0.1, which is the value that has been used in the LES of Breuer
1196 J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206

.
[14] and Frohlich et al. [15]. Therefore, Eqs. (6) and (7) are fullled by the model for
Tsgs on this analytical level.
For qsgs nally, an eddy diffusivity model is used, qsgs rv * with a
% sgs cp =Prsgs rT;
subgrid Prandtl number Prsgs 0:7 [6]. Also for models for the subgrid heat ux
realizability conditions exist, namely the Schwarz inequality for the components of
qsgs [22,6]. But these three inequalities have not been analyzed in the present work.
Just Eqs. (6)(8) are examined at positions, which are specied in the following.

2.2. Numerical approximations

Eqs. (1)(5) are solved by a cell-centered nite volume method [23]. All surface
integrals for the ux terms are approximated by the midpoint rule, which is of second
order. For the convective terms the high-resolution scheme of Jameson et al. [24] has
been applied in the formulation of Martinelli [25]. This method is a central scheme of
second order supplemented with an articial viscosity term, consisting of second-
and fourth-order viscosity. While the second-order term is only activated, when
shocks are detected, which is not the case for the ow studied here, the fourth-order
viscosity serves as a background dissipation to suppress oddeven decoupling. The
1
magnitude of this term is controlled by a factor, which is normally set to 64 for the
computation of steady ows [26]. Here we reduced the factor to 7.8125 105 to
keep the dissipative character of this scheme as small as possible. To quantify the
inuence of the articial viscosity term, the ratio of the net articial viscosity uxes
to the net subgrid uxes has been evaluated in each control volume for several time
steps. The ratio does not exceed 3%, conrming the small amount of articial
viscosity added by this approach.
The viscous uxes are also obtained from numerical integration using the
midpoint rule. The values of the cartesian components of the stress tensor and
the heat ux vector in the centers of the control volume surfaces are obtained in the
following way. First, the gradients of the velocity components and the temperature
are calculated in the four vertices of the face. This is done with the aid of an auxiliary
control volume, constructed around the vertices, and the Gauss theorem [34, p. 218].
Then the cartesian components of the stress tensor and the heat ux vector
are calculated in the vertices with these gradients. The dynamic and subgrid scale
viscosity needed at the vertices are linearly interpolated from the cell centers, where
they are calculated. The components of the stress tensor and heat ux vector in
the center point of the surface are nally obtained by arithmetically averaging the
four values of the vertices.
While the dynamic viscosity is directly computed from the temperature in the
control volume, the subgrid scale viscosity depends on the deviatoric part of the
deformation tensor via Eq. (5). In the present implementation this is computed in
the control volumes center by approximating the velocity gradients by central
differences. Therefore two approximations for the subgrid stresses must be taken
into account when analysing their realizability: the subgrid stresses in the control
volumes center and the subgrid scale stresses in the center points of the surfaces,
leading to the subgrid scale uxes.
J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206 1197

For the time integration of the resulting semi-discrete system of conservation laws
an explicit three stage RungeKutta scheme of third order accuracy [27] has been
used with a CFL number of 1.0.

2.3. Description of the test case

The computational domain, the coordinate sytem and the denition of the angle y;
which is used to measure the seperation points, are depicted in Fig. 1 on the left. The
size of the domain in normal direction has been chosen in accordance with the DNS
of Ma et al. [16] and of Tremblay et al. [17], to compare the present results with those
DNS. Therefore also the extent in spanwise direction has been set to pD; which is the
value that has been used in all previously mentioned LES.
In that computational domain a block structured H-grid has been used with an
O-grid with radius 5D around the cylinder. The height of the rst cell at the cylinder
surface is 3.5 103 D; which is approximately 30% larger than the value used by
.
Breuer [14] and Frohlich et al. [15], but still small enough for y+ being of the order of
1. The grid is then stretched in the normal direction of the cylinder wall with a
stretching factor of 1.032 inside of the O-grid. Part of this grid is shown in Fig. 1 on
the right. Outside of the O-grid, the stretching factor is increased to a maximum of
1.05.
The number of grid points along the wake centerline is 185. In the normal
direction at y 7901; 153 points are used and on the circumference of the cylinder
193. The grid therefore has a higher resolution in circumferential direction than all
other LES while the resolution on the wake centerline is approximately the same. In
spanwise direction 33 grid points are used which is the same as that used by Breuer
.
and Frohlich et al., and less than that used by Beaudan and Moin [11], Mittal and
Moin [12] and Kravchenko and Moin [13], who used 48 points. The total amount of
hexahedral control volumes of the present grid is 1138688.
At the inow and at the boundaries in normal direction laminar far eld boundary
conditions are prescribed. For those the velocity U0 ; normal to the inow boundary,

Fig. 1. Computational domain (left) and details of the grid close to the cylinder (right).
1198 J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206

of the parallel free stream and its temperature T0 are prescribed to yield the
Reynolds number ReD r0 U0 D=m0 3900 and the Mach number M0 U0 =c0
0:2; where c0 is the sound speed in the free stream. The temperature T0 is also used
for the isothermal boundary condition at the cylinder wall. For the velocities the no-
slip condition is applied there. At the outow the far eld condition with the static
pressure p0 of the free stream is prescribed. This boundary condition is known to
be reective [28]. When an inhomogeneous pressure distribution like the one in the
large Ka! rma! n vortices is convected across the outow boundary, small pressure
disturbances are reected back into the computational domain, travelling towards
the inow boundary where they are again reected back into the computational
domain. As instantaneous distributions of the pressure and of the solution variables
do not show any visible reections, these have been considered as small. But it
should be kept in mind, that this boundary condition imposes disturbances on the
approaching ow, which are purely numerical.

3. Results and discussion

The code used for the simulation has full multigrid capability. Therefore the
computation was started from uniform ow on the coarsest grid level and run there
until periodic vortex shedding occured. This solution was then transferred to the next
ner grid level and the simulation continued. On the nest grid, which is the one
described in the previous section, the simulation was run for ten shedding cycles [13],
before time averaging was begun. The three dimensional solution variables were then
sampled at every ten time steps, corresponding to a dimensionless time interval of
Dt DtU0 =DE2:64 103 : In addition squares and products of the solution
variables were sampled, which are necessary to compute the components of the
Reynolds stress tensor, which arises in time averaging. As the compressible Navier
Stokes equations are solved, the time averages are also dened as Favre averages.
The temporal mean of the streamwise velocity component is therefore dened as
fUg
/r% uS=/
* %
rS; where /?S denotes the usual Reynolds time average of a
variable. For the turbulent normal stress, which is computed from the resolved
values only, one has analogously fU 00 U 00 g
/r% u* 00 u* 00 S=/rS
% fu* 2 g  fug
* 2 ; where
dashed quantities are the uctuations around the Favre mean. The other mean
values and components of the Reynolds stress tensor are dened accordingly. These
time averages have additionally been averaged over the spanwise direction to yield
the mean values that will be presented next.
The inuence of compressibility on the results can be quantied by the relative
deviation of the calculated mean density from the reference density, D/rS=r0
/rS  r0 =r0 ; and the intensity of the uctuations, rrms =r0 : While D/rS=r0 is in
the range ] 0.035, 0.030[, with the maximum at the stagnation point at y 01 and
the minimum near the end of the recirculation region, rrms =r0 does not exceed 1.3%.
Therefore the inuence of the density variations on the results can be considered as
negligible. Whether the variation of the mean density has a great inuence on the
J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206 1199

results will become apparent, when the present results are compared to the results of
Ma et al. [1], whose DNS is based on the assumption of constant density.

3.1. Mean integral quantities

First the integral results of the simulation are shown. Those are the mean drag
/cd S; the mean back pressure coefcient /cpb S; the mean seperation angle /ys S;
the mean recirculation length Lr and the Strouhal number St. These quantities are
listed in Table 1 for ve dimensionless time averages T TU0 =D together with
experimental and other numerical results from DNS and LES.
The present results indicate that the mean integral values depend strongly on the
averaging time. For short integration times the results coincide with the data of
the LES of Kravchenko and Moin [13] and Mittal [32], who used about the same
integration times. The good agreement between the present results and those, which
were obtained with a dynamic Smagorinsky model, can be attributed to the ner grid
of the present computation.
With increasing averaging time, the results approach the data of the DNS and
LES of Ma et al. [16] with a substantially longer recirculation region, which is
accompanied with an increase in the base pressure and a decreasing mean drag
coefcient.
The results also show, that the maximum averaging time of T 200 is not
enough to reach a statistically converged mean oweld. While the changes from
T 150 to T 200 in /cd S; /cpb S and /ys S are getting smaller, compared to

Table 1
Computed integral mean quantities for different averaging cycles Nc and times T TU0 =D; respectively,
in comparison to experimental and other numerical results

Data from Nc ; T* /cd S /cpb S ys (1) Lr =D St

Experiment , 0.987 0.907 85.07 1.337 0.2157


, 0.05 [29] 0.05 [29] 2 [30] 0.2 [31] 0.005 [31]
DNS [16]: Case II 131, 624.89 0.84 1.59 0.219
DNS [17] 60, 300 1.03 0.92 85.7 1.30 0.220

LES [16]: Case V 131, 624.89 0.765 1.76 0.208


LES [13] 7, 35 1.04 0.94 88.0 1.35 0.210
LES [14]: C2 56, 1.099 1.049 87.9 1.115
LES [15]: LRUN2 86, 1.08 1.06 88 1.09 0.210
LES [32] 12, 1.00 0.93 86.9 1.40 0.207
LES [11] 6, 31.22 0.92 0.81 84.8 1.74 0.209

Present LES 10, 50 1.005 0.94 89.0 1.34


16, 75 0.999 0.90 88.7 1.44
21, 100 0.994 0.88 88.5 1.51 0.209
31, 150 0.985 0.86 88.3 1.57
42, 200 0.978 0.85 88.2 1.64
1200 J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206

the respective changes between T 100 and T 150; the change in Lr is about the
same. This indicates that the converged Lr will be close to the LES of Beaudan and
Moin [11] and of Ma et al. The long recirculation length of the former can be
attributed to the dissipative numerical scheme when regarding the averaging time
and the value for the Smagorinsky constant (CS 0:065), because Ma et al. showed
that a higher value for the Smagorinsky constant leads to a longer recirculation zone.
In their LES (Case V), cited in Table 1, they used CS 0:196; which is only twice the
value that is used in the present simulation. As the results of their LES correspond
to their DNS (Case II), the results of the present simulation will be mainly compared
to this DNS and not the one of Tremblay et al. [17], because it differs considerably
from the one of Ma et al.
The large difference between the DNS of Ma et al. and the DNS of Tremblay et al.
can be explained with the different boundary conditions in normal direction. Ma
et al. used zero gradient conditions at y=D 79; which corresponds to the ow
between two mirrored cylinders, where seperation occurs at the sides of the cylinders
which are facing each other, due to the symmetry. In contrast to that, the ow of
Tremblay et al. is the one between two cylinders, where one cylinder has been merely
translated in normal direction, as they used periodic boundary conditions in the
normal direction at y=D 710: Therefore seperation occurs at the same side of the
cylinders. Thus the ow is accelerated at the boundary closer to the seperation and
decelerated at the opposite boundary.
The fareld boundary conditions applied in the present simulation x the
tangential velocity component at y=D 710 to the freestream value U0 : This
condition is closer to the boundary conditions of Ma et al. than to the ones of
Tremblay et al. Therefore the present results are in better agreement with the DNS
and LES of the former. But as Mittal [32] has stated, application of the fareld
boundary condition at y=D 710 leads to an acceleration of the ow at the edge of
the wake region. This can be also seen in the present mean velocity proles, which
are presented in the next section. While Mittal decided to enlarge his domain in
normal direction to y=D 725; the present domain size has been chosen to render
the results comparable with the DNS of Ma et al.

3.2. Mean flow and turbulence statistics

The mean oweld and the turbulence quantities in the wake are compared with
the corresponding data of the DNS (Case II) of Ma et al. [16], as their simulation is
closest to the present LES, concerning the boundary conditions in normal direction
and the spanwise extent of the computational domain. In Fig. 2 the mean velocity
components are plotted at three locations in the very near wake (x=Dp3). The
different lines correspond to different averaging times, namely T 50; 100 and 200.
At x=D 1:06 the prole of the mean streamwise velocity develops from a V-
shape prole to a U-shape prole with increasing T : At this location and at x=D
1:54; a pronounced peak of the mean streamwise velocity can be observed at the edge
of the wake region. As has been stated in the last section, this is due to the fareld
boundary condition at y=D 710: Nevertheless, the present results justify this
J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206 1201

Fig. 2. Mean velocity proles at x=D 1:06; 1.54, 2.02 (from top to bottom, respectively) for different
averaging times. T 50 (dotted line), T 100 (dashed line), and T 200 (solid line). Squares denote
DNS results of Ma et al. [16]: (a) streamwise velocity and (b) crossow velocity.

Fig. 3. Normal components of the Reynolds stress tensor at x=D 1:06; 1.54, and 2.02 (from top to
bottom, respectively) for different averaging times. Symbols as in Fig. 2: (a) streamwise stress and
(b) crossow stress.

choice, when intending to compare the results with the DNS of Ma et al. Further
downstream the velocity at y=D 0 reduces with increasing T in accordance with
the increasing length of the recirculation region.
For the mean crossow velocities the agreement with the DNS results is also very
good, as can also be seen in Fig. 2. While the differences between the values for the
three averaging times are small at most locations, this is not the case for the normal
components of the Reynolds stress tensor, which are plotted in Fig. 3.
The streamwise stress reduces with increasing T at x=D 1:06: The lower
uctuations correspond to the more at U-shaped prole of the mean streamwise
velocity, with less ow mixing. Contrary to that, the uctuations grow at x=D
2:02; which is due to the fact that the mean recirculation length is closer to that point
1202 J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206

with increasing T : For the crossow stress a permanent decrease can be observed,
with a larger deviation from the DNS results for T 200: In addition, the
asymmetry of the normal stresses for short averaging times indicates that these
averaging times are too short.
The mean velocities and Reynolds stresses in the near wake (3px=Dp10) are
compared with the measurements of Ong and Wallace [10] in the following, as
complete data for comparison are not available from Ma et al. and their available
results match the experiments except for the mean streamwise velocity at x=D 4:
Furthermore, only the present results for T 200 are shown.
In Fig. 4 the mean streamwise velocity and the streamwise normal Reynolds stress
are plotted at three locations.
The lower velocity at x=D 4 on the wake centerline indicates that the
recirculation length in the experiment has been smaller than the computed one.
Again the result is in good agreement with the DNS of Ma et al. Accordingly the
calculated streamwise Reynolds stress is higher than the experimental stress, but
the results of Ma et al., not shown here, are closer to the experiment. Further
downstream the agreement of the mean velocities is good, whereas the Reynolds
stresses of the LES are too small in the wake, also smaller than the DNS, which
is very close to the experiments. As Kravchenko and Moin [13] have pointed out, this
is due to the truncation error of central schemes on coarse grids.
For the crossow normal stresses and the Reynolds shear stresses, plotted in
Fig. 5, there is also an overprediction at x=D 4: As the crossow stress is close to
the experimental values, the large deviations can be attributed to the streamwise
normal stress, which is predicted to be too large. Further downstream the crossow
stresses show that the calculated wake is wider than the measured one.
The comparison between the present simulation and the experimental results of
Ong and Wallace in the near wake, and the DNS results of Ma et al. in the very near
wake displays the programs capability of performing LES, at least for the ow

Fig. 4. LES results at x=D 4; 7; 10 (from top to bottom, respectively) for T 200: Circles are data of
Ong and Wallace [10], squares are DNS of Ma et al. [16]: (a) mean streamwise velocity and (b) streamwise
normal Reynolds stress.
J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206 1203

Fig. 5. LES Reynolds stresses at x=D 4; 7; 10 (from top to bottom, respectively) for T 200: Circles
are data of Ong and Wallace [10]: (a) crossow normal Reynolds stress and (b) Reynolds shear stress.

around a circular cylinder at ReD 3900: From the good agreement of the present
results with the aforementioned incompressible DNS, especially in the very near
wake, it can be concluded that the solution of the compressible equations is feasible
for this ow.
Whether the presented mean values have been obtained from physical, i.e.
realizable instantaneous ow elds, is revealed in the next section, where the analysis
of the subgrid scale stresses realizability is presented.

3.3. Realizability

As has been stated in the introduction, realizability is a fundamental property of


turbulence, which should also be fullled by the computational results. In LES this
requires the analysis of the inequalities given in Eqs. (6)(8) for the Reynolds stress
tensor of the time-averaged solution and of the subgrid stress tensor for the
instantaneous ow elds. For the Reynolds stress tensor these inequalities are
fullled with all the averaging times, for which the integral quantities are presented
in Section 3.1. The averaging time therefore has no inuence on the realizability of
the solution.
The realizability conditions (6)(8) of the subgrid stress tensor are analyzed at
several positions in each control volume. Unlike Fureby and Tabor [6], who
analyzed a time average of the invariants in forced homogeneous isotropic
turbulence and a turbulent channel ow, the current data are only analyzed on a
momentary basis. Thus, no correlations between non-realizability and other
quantities of the ow are examined in the present work.
The deviator of the deformation tensor in the cell centers is calculated to yield the
subgrid scale viscosity vsgs and the turbulent kinetic energy of the subgrid scales ksgs
via Eq. (5). As ksgs is dened to fulll the inequalities for the rst and the second
invariant of the subgrid stress tensor, calculated with S# in the cell centers, only the
third invariant has to be monitored during the calculation in the entire domain. Its
1204 J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206

value is always positive. Therefore, the subgrid stress tensor in the cell centers is
always realizable.
The subgrid stress tensor on each face of the control volumes is analyzed only at
14 points of time. The invariants are calculated from the subgrid stress tensor in the
center of each face. The procedure to calculate this tensor has been described in
Section 2.2. The invariant I1 ; which equals twice ksgs ; has to be always positive, as it
is obtained from the positive values in cell centers by linear interpolation. Therefore
only I2 > 0 and I3 > 0 have to be checked. It is found that these conditions are also
satised in the entire domain at all points of time. Therefore, all instantaneous
owelds that have been used in calculating the mean values represent physical
possible owelds.
The most probable reason for our results being always realizable in contrast to the
ndings of Fureby and Tabor [6] is the ne grid that is used in the present simulation.
Based on the experimentally determined Kolmogorov scale in the near wake of the
cylinder [10], Jordan [33] designed an O-grid for a DNS of the ow under
consideration. For the resolution of the cylinder wake, which had a length of 12D; he
used 241 grid points, and for the circumference and for the span of 321 and 64
points, respectively. Therefore, our grid is approximately less than twice as coarse in
each coordinate direction as a grid required for a DNS, with a corresponding small
amount of subgrid scale energy to be modeled. Contrary to this, Fureby and Tabor
performed their LES of forced isotropic turbulence with a grid that consisted only of
one-fourth of the grid points that were used in the corresponding DNS, 323 vs. 1283.
Also, for the aforementioned channel ow their grid was relatively coarser than
our grid. However, they used an implicit method for time integration, which might
as well have an inuence on the realizability, because of the convergence error
of iterative schemes. Also, no reference was made to the use of damping of the
subgrid stresses near the channel wall. This may also have an inuence on the
realizability, as Fureby and Tabor observed most of the unrealizable regions near
the channel walls.

4. Summary

In this study, we performed a large eddy simulation of the ow over a circular


cylinder for ReD 3900; to validate our nite volume code, which solves the
compressible NavierStokes equations together with the xed coefcient subgrid
scale model of Yoshizawa [18] and the high-resolution switch scheme of Jameson
et al. [24] for the numerical approximation of the convective uxes. For the
dimensionless averaging time of T 200 the present results in the very near wake
(x=Dp3) agree very well with the DNS of Ma et al. [16], who used the same cylinder
span and comparable boundary conditions in normal direction. The agreement with
the DNS and the experimental results of Ong and Wallace [10] in the near wake
(3Xx=Dp10) is fair. Therefore, we conclude that the solution of the compressible
NavierStokes equations is a feasible approach for LES of bluff body ows, when
the unsteady forces and the mean and uctuating velocities shall be predicted.
J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206 1205

For shorter averaging times we found a fair agreement of the mean recirculation
length with the large eddy simulations of Beaudan and Moin [11], Mittal [32] and
Kravchenko and Moin [13], who reported shorter recirculation lengths than Ma et al.
From this and the fact that the present results do not yet show a statistically
converged solution of the mean oweld, it is concluded that averaging times even
larger than T 200 are necessary for the case investigated.
As in principle only physical possible owelds are allowed to be used in
calculating ensemble or time averages, we additionally analyzed the realizability of
the subgrid scale stresses at the cell centers, where the subgrid viscosity and the
subgrid kinetic energy are calculated, and at the control volumes surfaces, where the
subgrid scale uxes that enter the balance equations are nally evaluated. At both
places no violation of realizability has been observed. Therefore, we conclude that
our time averages have been obtained from merely physical instantaneous owelds.
The according analysis of the Reynolds stress tensor of these time averages showed
also satisfaction of the realizability condition. The results consequently are in
accordance with the physics of turbulence. Furthermore, the method presented for
determining whether a solution of a LES is a physical ow or not can be easily
included in the analysis of computational results, of course not only in wind
engineering applications.

References

[1] R. Du Vachat, Realizability inequalities in turbulent ows, Phys. Fluids 20 (1977) 551556.
[2] U. Schumann, Realizability of Reynolds-Stress Turbulence Models, Phys. Fluids 20 (5) (1977)
721725.
[3] T.-H. Shih, W.W. Liou, A. Shabbir, Z. Yang, J. Zhu, A new ke Eddy Viscosity Model for high
Reynolds number turbulent ows, Comput. Fluids 24 (3) (1995) 227238.
[4] T. Rung, F. Thiele, S. Fu, On the realizability of nonlinear stressstrain relationships for Reynolds
stress closures, Flow Turbulence Combust. 60 (1999) 333359.
[5] B. Vreman, B. Geurts, H. Kuerten, Realizability conditions for the turbulent stress tensor in large-
eddy simulation, J. Fluid Mech. 278 (1994) 351362.
[6] C. Fureby, G. Tabor, Mathematical and physical constraints on large-eddy simulations, Theoret.
Comput. Fluid Dynamics 9 (2) (1997) 85102.
[7] M. Breuer, Towards technical application of large eddy simulation, Z. Angew. Math. Mech. 81
(Suppl. 3) (2001) S461S462.
[8] M.M. Zdravkovich, Flow around circular cylinders, Vol 1: fundamentals, Oxford University Press,
Oxford, New York, 1997.
[9] L.M. Lourenco, C. Shih, Characteristics of the plane turbulent near wake of a cylinder, A particle
image velocimetry study, Unpublished, 1993 (data taken from Beaudan and Moin [12]).
[10] L. Ong, J. Wallace, The velocity eld of the turbulent very near wake of a circular cylinder,
Exp. Fluids 20 (1996) 441453.
[11] P. Beaudan, P. Moin, Numerical experiments on the ow past a circular cylinder at a sub-critical
Reynolds number, Report No. TF-62, Thermosciences Division, Department of Mechanical
Engineering, Stanford University, USA, 1994.
[12] R. Mittal, P. Moin, Suitability of upwind-biased nite difference schemes for large-eddy simulation of
turbulent ow, AIAA J. 35 (8) (1997) 14151417.
[13] A.G. Kravchenko, P. Moin, Numerical studies of ow over a circular cylinder at ReD=3900,
Phys. Fluids 12 (2) (2000) 403417.
1206 J. Franke, W. Frank / J. Wind Eng. Ind. Aerodyn. 90 (2002) 11911206

[14] M. Breuer, Large eddy simulation of the sub-critical ow past a circular cylinder: numerical and
modeling aspects, Int. J. Numerical Methods Fluids 28 (1998) 12811302.
.
[15] J. Frohlich, W. Rodi, Ph. Kessler, S. Parpais, J.P. Bertoglio, D. Laurence, Large eddy simulation of
ow around circular cylinders on structured and unstructured grids. In: E.H. Hirschel (Ed.),
Numerical ow simulation: CNRS DFG collaborative research programme, Vol. 66, Vieweg,
Braunschweig, 1998, pp. 319338.
[16] X. Ma, G.-S. Karamanos, G.E. Karniadakis, Dynamics and low-dimensionality of a turbulent wake,
J. Fluid Mech. 410 (2000) 2965.
[17] F. Tremblay, M. Manhart, R. Friedrich, DNS of ow around a circular cylinder at a subcritical
Reynolds number with cartesian grids. In: Proceedings of the Eighth European Turbulence
Conference, Barcelona, Spain, EUROMECH, CIMNE, 2730 June, 2000, pp. 659662.
[18] A. Yoshizawa, Statistical theory for compressible turbulent shear ows, with the application to
subgrid modeling, Phys. Fluids 29 (1986) 21522164.
[19] M. Lesieur, P. Comte, Favre ltering and macro-temperature in large-eddy simulations of
compressible turbulence, C. R. Acad. Sci. Paris, S!erie II B 329 (2001) 363368.
[20] A.W. Vreman, B.J. Geurts, J.G.M. Kuerten, Subgrid-modelling in LES of compressible ow, in:
P.R. Voke, L. Kleiser, J.P. Chollet (Eds.), Direct and Large-Eddy Simulation I, Kluwer, Dordrecht
Boston, London, 1994, pp. 133144.
[21] B. Vreman, B. Geurts, H. Kuerten, A priori tests of large eddy simulation of the compressible plane
mixing layer, J. Eng. Math. 29 (1995) 299327.
[22] U. Schumann, Large-eddy simulation of turbulent diffusion with chemical reactions in the convective
boundary layer, Atmos. Environ. 23 (1989) 17131727.
[23] F. Magagnato, KAPPAKarlsruhe parallel program for aerodynamics, TASK Q. 2 (2) (1998)
215270.
[24] A. Jameson, W. Schmidt, E. Turkel, Numerical solutions of the Euler equations by nite volume
methods using RungeKutta time-stepping schemes, AIAA 81-1259, 1981.
[25] L. Martinelli, Calculations of viscous ows with a multigrid method, Ph.D. Thesis, Department of
Mechanical and Aerospace Engineering, Princeton University, 1987.
[26] R.C. Swanson, E. Turkel, Multistage central differencing schemes for the Euler and NavierStokes
equations. Lecture Series 1996-06, von K!arm!an Institute for Fluid Dynamics, 1996.
[27] C.-W. Shu, S. Osher, Efcient implementation of essentially non-oscillatory shock-capturing schemes,
J. Comput. Phys. 77 (1988) 439471.
[28] T. Poinsot, S.K. Lele, Boundary conditions for direct simulations of compressible viscous ows,
J. Comput. Phys. 101 (1992) 104129.
[29] C. Norberg, Effects of Reynolds number and low-intensity free-stream turbulence on the ow around
a circular cylinder, Publ. No. 87/2, Department of Applied Thermoscience and Fluid Mechanics,
Chalmers University of Technology, Sweden, 1987.
[30] J. Son, T.J. Hanratty, Velocity gradients at the wall for ow around a cylinder at Reynolds numbers
from 5 103 to 105, J. Fluid Mech. (1969) 353368.
[31] G.S. Cardell, Flow past a circular cylinder with a permeable splitter plate, Ph.D. Thesis, Graduate
Aeronautical Laboratory, California Institute of Technology, 1993.
[32] R. Mittal, Progress on LES of ow past a circular cylinder, CTR Annual Research Briefs, Center for
Turbulence Research, Stanford, CA 94305, 1996, pp. 233241. http://ctr.stanford.edu/ARB96.html
[33] S.A. Jordan, Dynamic subgrid-scale modeling for large-eddy simulations in complex topologies,
J. Fluids Eng. 123 (2001) 619627.
[34] J.H. Ferziger, M. P!eric, Computational methods for uid dynamics, Springer, Berlin, Heidelberg,
New York, 1996.

You might also like