Flow Around A Surface Mounted Cubical Obstacle: Comparison of LES and RANS-results

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/281077975

Flow around a surface mounted cubical obstacle: Comparison of LES and RANS-
results

Article  in  Notes on Numerical Fluid Mechanics · January 1996


DOI: 10.1007/978-3-322-89838-8_4

CITATIONS READS
31 218

3 authors, including:

Michael Breuer Djamel Lakehal


Helmut Schmidt University / University of the Federal Armed Forces Hamburg AFRY Switzerland
229 PUBLICATIONS   5,126 CITATIONS    133 PUBLICATIONS   2,682 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Modeling and simulation of particle agglomeration, droplet coalescence and particle-wall adhesion in turbulent multiphase flows View project

Simulation methods for dynamic fluid-structure interactions of lightweight flexible structures View project

All content following this page was uploaded by Michael Breuer on 23 August 2015.

The user has requested enhancement of the downloaded file.


FLOW AROUND
OBSTACLE: A SURFACE
COMPARISON MOUNTED
OF LES CUBICAL
AND RANS{RESULTS
M. Breuer, D. Lakehal, W. Rodi
Institute for Hydromechanics, University of Karlsruhe, Germany

SUMMARY
The paper deals with a comparative study of LES and RANS (k " model) results for a typical blu {body
ow, namely the ow around a surface mounted cubical obstacle placed in a plane channel. For this test
case detailed experimental data (Re=40,000) are available [11]. Two slightly di erent numerical solution
procedures based on a 3{D nite{volume method are used in this investigation. The Reynolds{averaged
equations for incompressible ow are solved implicitly [10], whereas in the LES code [1, 2, 3, 4, 5] an
explicit second order Adams{Bashforth scheme is applied. Di erent formulations of the k " turbulence
model are used in the RANS simulations, the standard version with wall functions, a RNG version, a
modi ed version proposed by Kato and Launder [7], and a two{layer approach. For modeling the non{
resolvable subgrid{scale motion in the LES two di erent models are applied, namely the well known
Smagorinsky model [16] as well as the dynamic model originally proposed by Germano et al. [6]. The
capability of the di erent methods is demonstrated by comparison with the measurements.

1. INTRODUCTION
Turbulent ows of practical interest are in general very complex including phenomena such as separation,
reattachment and vortex shedding. An appropriate description by Reynolds{Averaged Navier{Stokes
(RANS) equations combined with statistical turbulence models is dicult to achieve. This is due to the
necessity of modeling the whole spectrum of turbulent scales. The method of direct numerical simulation
requires no model assumptions but will not be applicable to engineering ows in the foreseeable future
because of extremely high computing costs. The concept of large{eddy simulation (LES) seems to be a
promising way of solving such ow problems. In LES the large eddies that depend strongly on the special
ow con guration are resolved numerically whereas only the ne{scale turbulence has to be modeled by
a subgrid{scale model.
The goal of the work reported here is the development of a large{eddy simulation technique for practically
relevant ows and a comparative study of LES and RANS (k " model) results. The ow around
a surface mounted cubical obstacle inside a plane channel was chosen as a typical blu {body ow.
Detailed experimental data (Re = UB H= = 40; 000, UB = bulk velocity, H = obstacle height) have been
provided by Martinuzzi et al. [11]. In the rst section of the paper the governing equations, turbulence
models, methods of solution, and boundary conditions for both RANS and LES approach are explained.
The second part deals with a detailed comparison of the di erent computed results as well as with the
experimental data.

2. GOVERNING EQUATIONS
Three{dimensional unsteady ows are described by the Navier{Stokes equations. These are the starting{
point for both the RANS approach as well as the LES technique. The procedures to derive the governing
equations are also very similar. For the RANS approach the Reynolds averaging procedure is introduced
which separates all instantaneous quantities in a turbulent ow eld into a time or ensemble averaged
mean value and a uctuating part. Then the Reynolds averaged equations describe the motion of the time
or ensemble averaged mean ow. Due to the averaging of the non{linear convective terms the unknown
Reynolds stress tensor ij appears. In the LES approach a similar averaging procedure is applied to
the governing equations. However, in contrast to the RANS approach the averaging is now accomplished
with respect to space and not to time. In the context of LES this is called ltering rather than averaging.
The goal is to separate di erent length scales in the turbulent ow eld. All large scale structures which
can be resolved by the numerical method applied should be separated from the small scale structures
(subgrid{scales) which cannot be captured on a given grid. The governing equations for LES have the
same form as for the RANS approach. However, the meaning of some variables has changed. In LES the
equations describe the motion of the resolvable part (grid scale) of the ow eld. The new stress tensor
ij , resulting from the ltering of the convective terms, is now called Reynolds subgrid{scale stress
tensor. This expresses the main di erence between RANS and LES. In a RANS simulation the Reynolds
stress tensor has to be modeled, which describes the in uence of the whole spectrum of turbulent scales
on the time or ensemble averaged mean ow. In LES, however, a large part of the spectrum of turbulent
motions is directly computed by the numerical scheme and only the subgrid{scales have to be modeled.
It is well known that the smaller eddies in a turbulent ow are easier to model (more homogeneous and
isotropic) than the whole spectrum of turbulent motions, but the price which has to be paid in LES for
this advantage is the necessity to resolve the large scale structures. In both approaches the stress tensor
ij is modeled using the eddy viscosity concept [19]. In this, ij is linearly related to the deformation
tensor Sij = 1=2 (Ui ; j + Uj ; i) by ij = 2=3 ij k 2 t Sij . In the RANS approach the trace of the
stress tensor 1=3 ij kk is expressed by the turbulent kinetic energy k = 1=2 ui ui, whereas in LES this
part is normally added to the pressure.

3. RANS: TWO{EQUATIONS TURBULENCE MODELS


The eddy viscosity is determined according to the algebraic expression [14] t = C k2=". This relation
requires values of the turbulent kinetic energy k as well as the dissipation rate ".

3.1 Standard k " model [14]


In the standard k " model the turbulence parameters k and " are obtained by solving transport equations
for these quantities together with those describing the mean ow. In both transport equations the term
Pk = ij Ui ; j appears which represents the rate of production of turbulent kinetic energy resulting from
the interaction of the turbulent stresses and the mean ow. The empirical constants appearing in the
model are given the following standard values: C = 0:09; C1 = 1:44; C2 = 1:92; k = 1: and " = 1:3.

3.2 RNG k " model [20]


In [20], the renormalization group (RNG) theory was applied to derive the k " turbulence model. In
this theory, the small scale uctuations are removed successively from the governing equations leading
to averaged equations. The resulting k " model has an extra term  R" in the "{equation, for which a
model is introduced, which reads: R" = C 3 (1 =0)= 1 + 3 "2 =k where  = (2Sij Sij )1=2k=" and
0 = 4:38 , = 0:015 . The model constants following from the RNG theory are: C = 0:084; C1 = 1:42;
C2 = 1:68; and k = " = 0:7179.
3.3 Kato{Launder k " model [7]
One of the most important drawbacks of the standard k " model is that it leads to an excessive production
of k in stagnation regions, i.e. in impinging ows. This phenomenon is a consequence of the inability
of eddy viscosity models to simulate correctly the di erence in normal Reynolds stresses governing the
production of k in such regions. Hence, Kato and Launder [7] suggested as an adhoc measure to replace
the original production term Pk = C " S 2 by Pk = C " S
, where S = (1=2 Sij Sij )1=2 k=" is the strain
rate parameter and
= (1=2
ij
ij )1=2 k=" the rotation parameter (
ij = 1=2 (Ui ; j Uj ; i) ). In shear
ows
 S so that the original production term is recovered while in stagnation regions
 0 so that
the desired e ect of suppressing the k{production is achieved.

3.4 Two{Layer Model [15]


The assumptions on which the wall functions used in the standard k " model are based are not re-
ally valid. Hence, a two{layer model combining the k " model in the outer region (t= > 36) with
a one{equation model in the viscosity{a ected near{wall region (t= < 36) was proposed [15]. In the

2
one{equation model, the dissipation rate " needed for both eddy viscosity formulation and k trans-
port equation is not determined via a transport equation but from the following prescribed length{scale
distribution:
k 3=2  C 
"
" = L 1 + k1=2L with: L = CD  yn ; C" = 13:2; CD = 6:41;  = 0:41 (3.1)
L stands for the near{wall length{scale, and yn the normal distance from the wall. However, the turbulent
kinetic energy k is determined in the same way as in the k " model.The eddy viscosity t is obtained
from:
t = f C0 k1=2L with: f = 1 exp( 0:0198Ry); Ry = k1=2yn =; C0 = 0:084 (3.2)
It should be noted that Ry involves k1=2 as velocity scale and not U which changes sign in separated
ows.

4. LES: SUBGRID{SCALE MODELS


4.1 Smagorinsky Model [16]
In principle the LES concept leads to a similar closure problem as the RANS approach. Therefore a similar
classi cation of turbulence models starting with zero{equation models and ending up with Reynolds stress
models is possible. However, the non{resolvable small{scale turbulence in a LES is much less problem{
dependent than the large{scale turbulence so that the subgrid{scale turbulence can be represented by
relatively simple models, e.g. zero{equation eddy{viscosity models. The well known and mostly used
Smagorinsky model [16] is based on the Boussinesq's approach. The eddy viscosity t itself is a function
of the strain rate tensor Sij :
p
t = l2 j Sij j = (Cs )2 j Sij j with: j Sij j = 2 Sij Sij  = (x y z )1=3 (4.3)
Cs is called the Smagorinsky constant. Taking into account the reduction of the subgrid length l near solid
walls, the length scale is usually multiplied by a Van Driest damping function. Here the main advantages
of LES become clear. The dicult problem of RANS to determine a characteristic length scale of the
turbulent ow does not appear in LES, because it is xed by the size of the computational cell.

4.2 Dynamic Model [6]


One of the major drawbacks of the Smagorinsky model is that the Smagorinsky constant Cs was found
to depend on the ow problem considered. Secondly, in an inhomogeneous ow, the optimum choice
for Cs may be di erent for di erent points in the ow. Furthermore, the Smagorinsky model needs
some additional assumptions to describe ows undergoing transition or near solid walls. The dynamic
model, originally proposed by Germano et al. [6], eliminates some of these disadvantages by calculating a
`Smagorinsky constant' as a function of time and position from the smallest scales of the resolved motion.
Based on the local equilibrium approach (production = dissipation), the eddy viscosity is again evaluated
from eq.(4.3). However, in contrast to the Smagorinsky model Cs is no longer a constant but a local,
time{dependent variable. The dynamic model itself represents a method for determinating this unknown
variable Cs from the information already contained in the resolved velocity eld. A systematic procedure
for computing turbulent ows by LES without the necessity of prior experience to properly adjust the
Smagorinsky constant was derived by Germano et al. [6]. The basic formalism behind the method need
not be repeated here. Following a suggestion of Lilly [9] the original model was slightly improved by using
a least{squares approach to obtain values for Cs.
Tests with this formulation of the dynamic model have shown that in principle negative values of Cs are
possible. However, in practice these are not able to represent the backscatter e ect because negative eddy
viscosities strongly destabilize the numerical algorithm. Furthermore, the value Cs is an instantaneous and
local quantity varying very much in space and time. This also leads to numerical instabilities. Di erent
possibilities have been tested to remove this problem. Depending on the ow considered, di erent kinds
of averaging procedures can be applied. If the ow is homogeneous in a certain direction, averaging can
be applied over this direction. For fully inhomogeneous ows, like the ow around a surface mounted

3
cubical obstacle, only an averaging procedure in time is applicable. In order not to restrict the values
of Cs to a fully time{independent function and to allow variations with low frequencies, a special form
of time averaging (lowpass ltering) is chosen, which is well known as a recursive lowpass digital lter
[1]. With an appropriate value for the parameter of this lter function all high frequency oscillations are
damped out and only the low frequency variations remain. This seems to be better than fully freezing
Cs . In addition, negative eddy viscosities are clipped.

5. METHODS OF SOLUTION
Because the LES code (LESOCC [1, 2, 3, 4, 5]) is a descendant of the RANS program (FAST{3D
[10]) both have many features in common, which will be described rst. Both methods are based on a
nite{volume approach for solving the incompressible Navier{Stokes equations on general body{ tted,
curvilinear grids. A non{staggered, cell{centered grid arrangement is used. In order to avoid the well
known pressure eld checkerboard problem, the momentum interpolation technique due to Rhie and
Chow [13] is applied. The pressure{velocity coupling is achieved with the SIMPLE algorithm of [12]. In
both codes the viscous uxes are approximated by central di erences of second order accuracy. The linear
discretized system of equations is solved using the strongly implicit solution procedure of Stone [17] which
can be accelerated by a FAS multigrid technique.
The main di erences between both codes is given by the temporal discretization as well as the spatial
discretization of the convective uxes. Due to totally di erent goals, di erent time stepping schemes are
used. An implicit decoupled solution method is prefered for the RANS code, because it guarantees a fast
convergence to the desired steady state solution of the RANS equations. In a LES it is necessary to resolve
turbulent uctuations in time with at least an accuracy of second order. A LES further requires small
time steps which can be treated much more eciently by an explicit scheme. Therefore the temporal
discretization of the LES code consists of a predictor{corrector scheme, where the predictor step is an
explicit Adams{Bashforth scheme for the momentum equations (second order in time) and the corrector
step involves the implicit solution of the Poisson equation for the pressure correction.
Another important requirement for LES is a higher order approximation of the convective uxes where
in general the numerical dissipation produced by a scheme is a much better measure for its accuracy
than the order of the discretization itself. Central di erences of second order accuracy have been found
to be a reasonable discretization of the convective uxes for LES. In the RANS model the HLPA (hybrid
linear{parabolic approximation) second order low{di usive and oscillation{free scheme of [21] is applied
for the convective part.

6. BOUNDARY CONDITIONS
For the three RANS turbulence models, described in chapters 3.1 { 3.3, the boundary conditions at
impermeable walls involve the well known wall functions [14]. In order to obtain a fully developed chan-
nel ow solution as initial condition, the in ow pro les are approximated using a logarithmic pro le
U (y)=U =  1  ln(y=y0 ) for thepvelocity, and k(y) = 1:5(I (y)  U (y))2 for the k pro le respectively. y0
is the roughness length and I = u02 =UB the turbulence intensity. The on{coming ow is assumed to
have a dissipation rate according to fully developed channel ow, leading to "(y) = C3=4k3=2 =Lu. The
turbulence length scale Lu is set equal to 0:1H [8]. At the lateral planes, symmetry conditions are applied.
In the case of LES, the in ow is fully developed turbulent channel ow, generated by LES of plane channel
ow (same grid in the cross{sectional plane). For the lateral boundaries (x{y plane) at z=H = 3:5
periodic boundary conditions are chosen. At solid wall the wall function approach of Werner/Wengle [18]
is applied. A convective boundary condition is used at the out ow boundary.

7. GRIDS AND CALCULATION DOMAINS


For all RANS computations except the two{layer one, the same grid and computational domain is used.
A mesh consisting of 110  32  66 grid points (streamwise/normal to the channel walls/lateral) is
applied to form a computational domain with an upstream length of x1=H = 3:5, a downstream length

4
of x2=H = 10, and a width of b=H = 9. Here the grid covers the whole domain. The smallest cell volume
in the vicinity of the obstacle walls has a size of (0:01 H )3. However, for the two{layer model simulation
a ner grid with 142  84  64 grid points is used, which covers only one half of the calculation domain
taking into account the symmetry condition of the ow eld (smallest cell (0:001 H )3).
For all LES a computational domain with an upstream length x1=H = 3 and a downstream length of
x2 =H = 6 is used. The width is set to b=H = 7. The restriction to a smaller integration domain in the
LES case compared with the RANS case is necessary to achieve a sucient resolution of the ow eld in
the vicinity of the obstacle. Of course LES cannot take advantages of the symmetry of the time{averaged
ow eld as RANS can do. All LES computations are performed on a stretched grid with 165  65  97
grid points for the x; y and z directions. In the streamwise direction, 70 grid points are distributed in the
region in front of the obstacle. On the surface of the obstacle, 31 grid points are used in all directions.
The smallest cell volume in the vicinity of the solid walls has a size of (0:0125 H )3.

8. RESULTS
Fig. 1 shows a rst qualitative comparison of the results (Re = 40; 000). The streamlines in the plane of
symmetry and at a horizontal plane close to the channel oor are plotted for the experimental and numer-
ical results. In the LES the velocities are averaged over a long period of more than 100 dimensionless time
units (H=UB ) to achieve good statistics. It appears clearly that the stagnation point is well simulated by
the di erent numerical approaches, whereas the primary upstream separation location (labeled A in the
experimental oil ow pattern) caused by the strong adverse pressure gradient imposed by the obstacle,
is slightly shifted upstream vis a vis the experiment (x=H = 0:9). Moreover only LES and two{layer
calculations produce a correct separation bubble on the roof, with a somewhat better agreement with the
experiment for the LES (no reattachment of the time{averaged ow). The Kato{Launder modi cation of
the standard k " model seems to improve the results compared with the original version. However, the
calculated separation bubble on the roof of the obstacle is still too at compared with the experiment
and reattachment takes place. Indeed, the standard k " model as well as the RNG version simulation
show a very poor description of the ow in this region. Furthermore the extension of the large separa-
tion region (xr ) behind the obstacle is highly overpredicted using the RANS models (standard k ":
xr =H = 2:20, RNG: xr =H = 2:08)) compared with the experimental value xr =H = 1:62. The use of
both Kato{Launder model (xr =H = 2:73) and two{layer approach (xr =H = 2:68) shows an unexpected
overprediction of the size of the recirculating zone, resulting from an underpredicted turbulent viscosity
level. The agreement between the experiment and the time{averaged ow eld calculated by LES is much
better. The reattachment length behind the obstacle is only slightly overpredicted by the LES with the
Smagorinsky model (LES{S) (xr =H = 1:69) and underpredicted somewhat by the LES with the dynamic
model (LES{D) (xr =H = 1:43).
Fig. 1 displays also a comparison of the experimental versus the calculated time{averaged surface stream-
lines at the channel oor. The di erence between the locations of the calculated primary separation line
using the di erent calculation approaches is clear. Furthermore one can notice that amongst the di erent
RANS models used, only the two{layer model allows to capture the secondary recirculation at the front
base of the obstacle (C ), as well as LES. In comparison with the other RANS models the two{layer
approach reproduces more details of the ow structure near the walls. However, this is at least partly
the result of a ner resolution in the vicinity of the solid walls for the two{layer model in contrast to the
coarse grids used for the wall function approaches. The same gure shows that the horseshoe vortex gen-
erated between the primary and the secondary separation lines (B ) is fairly well predicted by the di erent
approaches. The ow patterns suggest also that the structure of the outer limit of the wake region formed
by the lateral arms of the horseshoe vortex (line D), varies between the di erent calculation approaches.
In the experiment, the width of this wake decreases up to approximately the reattachment point; then it
increases again. This behavior is well described only by the two{layer and LES models. Both RANS and
LES approaches seem to predict correctly the corner vortices (N12 ) generated downstream of the vertical
leading edges of the cube at the channel{body junction. The location of the simulated corner vortices
behind the obstacle (N14 ) shows clearly the di erences between RANS and LES results. Except for the
two{layer approach, the center of the vortices produced by the RANS methods is shifted downstream
compared with the experimental observation. The LES results agree fairly well with the experiment.

5
Fig. 2 displays the calculated versus the measured streamwise velocity pro les U pro les at six dif-
ferent locations in the symmetry plane. One obstacle height H in front of the cube (x=H = 1:0 ) all
streamwise velocity pro les agree fairly well the measurements. However, large di erences can be ob-
served for the next pro le in the middle of the roof (x=H = 0:5). Here the best result compared with
experiment is provided by the LES with the dynamic model (LES{D). The size of the separation bubble
and the magnitude of the reversed ow velocity is well reproduced. An attempt to classify the rest of the
simulations results in: LES{S, two{layer model, Kato{Launder model, standard k " model and nally
RNG model. The last two do not show any separation at this position on the roof of the obstacle at
all, while both the Kato{Launder modi cation and the two{layer model results show a better behavior.
This latter observation allows us to believe that combining the Kato{Launder modi cation with the
two{layer approach would provide a better description of the ow in this region. At the third location
(x=H = 1:), the same tendency can be detected. Moving further downstream, the e ect of the variations
in the computed length of the recirculation region is clearly visible. In the wake region (x=H = 1:5), the
computed velocity magnitudes are globally underestimated; here LES{S results are closest to experiment.
Far from the reattachment point at x=H = 4: , again both LES give a better representation of the ow
than all RANS models. The bad agreement between experiment and RANS computations concerning the
recirculation length behind the obstacle con rms the unsatisfactory results in the velocity pro les at this
position. This also demonstrates the low level of recovery of the ow eld. Therefore the RANS com-
putations would require a much longer distance to establish fully developed channel ow conditions again.
In Fig. 3 three pro les of the turbulent kinetic energy k are plotted in the symmetry plane. It should be
noted that for LES only the resolved part of the turbulent kinetic energy is included. At x=H = 0:5 all
simulations give similar peak values for k, however, the form of the pro les is di erent, e.g. the standard
k " model shows too large values above the separation bubble. In this gure the in uence of the
Kato{Launder modi cation can be clearly observed. The two{layer approach produces a k{pro le at this
location very similar to LES. At x=H = 1: the di erences in the peak values of k become larger. None
of the simulations gives the experimentally observed peak at the right position; in all computations, it is
located higher than in the measurements. Further downstream (x=H = 2: ) the scatter in the computed
pro les for k increases. Again, both LES are in closer agreement with the experimental values than all
RANS results, even if the two LES provide a slightly di erent behavior. Overall the level of turbulent
kinetic energy is much too small in the RANS computations which may cause the far too long recirculation
region behind the obstacle.

9. CONCLUSION
A typical blu {body ow, namely the three{dimensional turbulent ow around a surface{mounted cubical
obstacle placed in developed channel ow, has been investigated by four di erent 2{equation RANS
models as well as LES with two subgrid{scale models. This is a geometrically simple but physically very
complex ow with multiple, unsteady separation regions and vortices. Concerning all quantities considered
LES in general shows better results compared with the experiment than the RANS approaches. Some
qualitative features of the ow eld are not even captured by some of the RANS models, e.g. the large
separation region on top of the roof without reattachment of the time{averaged ow. Here the standard
k " model and the RNG modi cation do a rather poor job. Only for the two{layer approach the agreement
with the experiment is better in this region. The length of the recirculation region behind the obstacle
is highly overpredicted by all RANS models. Both LES show better agreement with the measurements.
Depending on the applied subgrid{scale model the recirculation length is slightly overpredicted (LES{S)
or even underpredicted (LES{D). Taking the surface streamlines at the bottom wall as well as the pro les
of mean velocity and turbulence kinetic energy as the basis of assessment, the tendency is always the
same. However, the price for better agreement with experiments is rather high and has to be mentioned
here. It is well known that LES is a very CPU{time consuming way of computing turbulent ows. This
is on the one hand due to the requirements concerning the spatial resolution of the ow eld. On the
other hand the most expensive part especially for fully inhomogeneous ows is the necessity to simulate
the instantaneous ow over a long period in time to achieve good statistical values. In our case the ratio
between the CPU{time requirements can be approximated by 1 : 25 : 200(400) where the three RANS
models (standard k " model, RNG, Kato{Launder) with nearly similar values are taken as the basis of
reference (approximately 15 CPU{min. on SNI S600/20). Switching to the two{layer approach already
increases the costs by a factor of about 25 due to the necessary resolution in the vicinity of solid walls

6
and lower rates of convergence. A factor of about 200 is present between standard RANS models and
LES for the mean quantities, where the value in brackets is an estimation for reasonable higher order
moments. However, if the instantaneous features of the ow eld are more interesting than the time{
averaged results, e.g. for uid{structure aerodynamic coupling problems, LES may become a reasonable
alternative to RANS models. Such simulations, however, still require powerful vector or parallel machines,
whereas 3{D RANS simulations (except the two{layer approach) can be performed on workstations.

ACKNOWLEDGMENTS
The work reported here was sponsored by the Deutsche Forschungsgemeinschaft and the Human Capital and Mo-
bility Programme of the European Union. The calculations were carried out on the SNI S600/20 vector computer
of the University of Karlsruhe (Computer Center).

REFERENCES
[1] Breuer, M., Rodi, W.: Large{Eddy Simulation of Turbulent Flow through a Straight Square Duct and a 180o
Bend, Fluid Mech. and its Appl., vol. 26, Direct & LES I, Sel. papers f. the First ERCOFTAC Workshop
on Direct & LES, Guildford, Surrey, U.K., 27{30 March 1994, ed. Voke, Kleiser & Chollet, Kluwer Acad.
pub., (1994).
[2] Breuer, M., Rodi, W.: Large{Eddy Simulation of Turbulent Flow through Straight and Curved Ducts, ER-
COFTAC Bulletin, vol. 22, Sept. (1994).
[3] Breuer, M., Pourquie, M., Rodi, W.: Large Eddy Simulation of Internal and External Flows, 3rd Interna-
tional Congress on Industrial and Applied Mathematics, Hamburg, 3{7 July, (1995), to be published in
ZAMM, (1996).
[4] Breuer, M., Pourquie, M.: First Experiences with LES of Flows past Blu Bodies, accepted for the 3rd
Intern. Symposium of Engineering Turbulence Modeling and Measurements, Crete, Greece, May 27{29,
(1996).
[5] Breuer, M., Rodi, W.: Large Eddy Simulation of Complex Turbulent Flows of Practical Interest, in prepara-
tion for the nal report of DFG Priority Programme: 'Flow Simulation with High{Performance Computers
II', Notes on Numerical Fluid Mechanics, Vieweg Verlag, Braunschweig, (1996).
[6] Germano, M.; Piomelli, U.; Moin, P.; Cabot, W. H. : A dynamic subgrid{scale eddy viscosity model, Phys.
Fluids A. 3 (7), pp. 1760{1765, (1991).
[7] Kato, M., Launder, B.E.: The Modeling of Turbulent Flow around Stationary and Vibrating Square Cylin-
ders, Proc. 9th Symp. Turb. Shear Flows, Kyoto, 10-4-1, (1993).
[8] Lakehal, D.: Simulation Numerique d'un Ecoulement Turbulent autour de Batiments de Formes Courbes,
Thesis at the University of Nantes, Ecole Centrale of Nantes, (1994).
[9] Lilly, D.K.: A proposed modi cation of the Germano subgrid{scale closure method, Phys. Fluids A 4 (3),
pp. 633{635, (1992).
[10] Majumdar, S., Rodi, W., Zhu, J. : Three{dimensional nite{volume method for incompressible ows with
complex boundaries, J. Fluid Eng., vol. 114, pp. 496{503, (1992).
[11] Martinuzzi, R. and Tropea, C.: The Flow around surface{mounted, prismatic obstacle placed in a Fully
Developed Channel Flow, J. of Fluids Engineering, vol. 115, (1993).
[12] Patankar, S.V., Spalding, D.B.: A calculation procedure for heat, mass and momentum transfer in three{
dimensional parabolic ows, Int. J. Heat & Mass Transfer, vol. 15, pp. 1778{1806, (1972).
[13] Rhie, C.M., Chow, W.L.: A numerical study of the turbulent ow past an isolated airfoil with trailing edge
separation, AIAA{J., Vol. 21, pp. 1225{1532, (1983).
[14] Rodi, W.: Turbulence Models and their Application in Hydraulics, International Association for Hydraulic
Research, Delft, The Netherlands, (1980).
[15] Rodi, W.: Experience with two{layer models conbining the k " model with a one{equation model near the
wall, AIAA paper, AIAA{91{0216, (1991).
[16] Smagorinsky, J.: General circulation experiments with the primitive equations, I, The basic experiment,
Mon. Weather Rev. 91, pp. 99{165, (1963).
[17] Stone, H.L.: Iterative solution of implicit approximations of multidimensional partial di erential equations,
SIAM J. on Num. Anal., vol. 5, pp. 530{558, (1968).
[18] Werner, H., Wengle, H.: Large{Eddy Simulation of Turbulent Flow over and around a Cube in a plate
Channel, 8th Symp. on Turb. Shear Flows, (Schumann et al., eds.), Springer Verlag, (1993).
[19] Wilcox, D. C.: Turbulence Modeling for CFD, DCW Ind., ING, Las Canada, California, USA, (1993).
[20] Yakhot, V., Orszag, S. A., Tangham, S., Gatski, T. B., Speziale, C. G.: Development of Turbulence Models
for Shear Flows by a Double Expansion Technique, Physics of Fluids A. 4, pp. 1510-1520, (1992).
[21] Zhu, J.: A low{di usive and oscillating{free convective scheme, Communications in Applied Numerical
Methods, vol. 7, pp. 225{232, (1991).

7
FIGURES
2.0

1.5

1.0 EXP
0.5

0.0
-2 -1 0 1 2 3 4
2.0 2

1.5 1

1.0
k{E 0

0.5 -1

-2
0.0
-2 -1 0 1 2 3 4 -2 -1 0 1 2 3 4

2.0 2

1.5
RNG
1

1.0 0

0.5 -1

-2
0.0
-2 -1 0 1 2 3 4 -2 -1 0 1 2 3 4

2.0 2

1.5
KATO{L 1

1.0 0

0.5 -1

-2
0.0
-2 -1 0 1 2 3 4 -2 -1 0 1 2 3 4

2.0 2

1.5 TWO{L 1

1.0 0

0.5 -1

-2
0.0
-2 0 2 4 -2 0 2 4

2.0 2

1.5 LES{S 1

1.0 0

0.5 -1

-2
0.0
-2 0 2 4 -2 0 2 4

2.0 2

1.5
LES{D 1

1.0 0

0.5 -1

-2
0.0
-2 0 2 4 -2 0 2 4

Fig. 1: Streamlines of the time{averaged ow in the symmetry plane of the 3{D obstacle
and surface streamlines in the bottom wall of the channel, Re = 40,000

8
Mean Velocity Profile U: x = -1.0 Mean Velocity Profile U: x = 0.5 Mean Velocity Profile U: x = 1.0
2 2 2
1.8 1.8 1.8
1.6 K-E 1.6 1.6
RNG
1.4 KATO-L 1.4 1.4
1.2 TWO-L 1.2 1.2
LES-S
LES-D
y

y
1 1 1
Exp.(u-v)
0.8 Exp.(u-w) 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 -0.5 0 0.5 1 1.5 0 0.5 1 1.5
U U U

Mean Velocity Profile U: x = 1.5 Mean Velocity Profile U: x = 2.5 Mean Velocity Profile U: x = 4.0
2 2 2
1.8 1.8 1.8
1.6 1.6 1.6
1.4 1.4 1.4
1.2 1.2 1.2
y

y
1 1 1
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
-0.5 0 0.5 1 -0.5 0 0.5 1 0 0.5 1
U U U

Fig. 2: Comparison of mean velocity pro les U of the time{averaged ow in the symmetry
plane of the 3{D obstacle, Re = 40,000

Turbulence Kinetic Energy: x = 0.5 Turbulence Kinetic Energy: x = 1.0 Turbulence Kinetic Energy: x = 2.0
2 2 2
1.8 1.8 1.8
1.6 1.6 1.6
1.4 1.4 1.4
1.2 1.2 1.2
y

1 1 1
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
k k k

Fig. 3: Comparison of turbulent kinetic energy pro les k of the time{averaged ow in the
symmetry plane of the 3{D obstacle, Re = 40,000, (same legend as in Fig. 2)

9
View publication stats

You might also like