Documente Academic
Documente Profesional
Documente Cultură
Numerical Recipes in FORTRAN: The Art of Scienti® c Computing, 2nd ed., came from the comparison of the computed solution with the ex-
Cambridge Univ. Press, Cambridge, England, UK, 1992, pp. 436±448. periments in the region downstream of the vortex formation region
4 Goldberg, D. E., Genetic Algorithms in Search, Optimization and Ma- (5.0 < x / D < 10.0) where the mesh was relatively coarse. It was
chine Learning, Addison±Wesley, Reading, MA, 1989, pp. 59±88. found that, in this region, numerical dissipation overwhelmed the
5 Quagliarella, D., and Cioppa, A. D., ª Genetic Algorithms Applied to the
contribution of the SGS model, and the three computed solutions
Aerodynamic Design of Transonic Airfoils,º Journal of Aircraft, Vol. 32, were virtually indistinguishablebeyond x / D > 7.0. The simulation
No. 4, 1995, pp. 889±891. with a seven-order scheme also showed that energy in a substantial
6 Doorly, D., ª Parallel Genetic Algorithms for Optimization in CFD,º
Genetic Algorithms in Engineering and Computer Science, edited by G. portion of the resolvable wave number range was damped due to
Winter, J. PÂ eriaux, M. GalÂan, and P. Cuesta, Wiley, Chichester, England, numerical dissipation, and it was concluded that these high-order
UK, 1995, pp. 251±270. upwind-biased schemes were unsuitable for use in LES.
7
Katz, J., and Plotkin, A., Low-Speed Aerodynamics: From Wing Theory Analysis of the truncation error5 indicates that higher-order
to Panel Methods, International ed., McGraw±Hill, New York, 1991, pp. upwind-biasedschemes provide good resolution in about two-thirds
301±377. of the wave number range, and the upper-half of the wave number
8 Vanderplaats, G. N., ªADSÐ A FORTRAN Program for Automated De-
range is affected by numerical dissipation. In contrast to upwind-
sign Synthesis, Version 3.00,º Engineering Design Optimization, Inc., Santa
Barbara, CA, March 1988. biased schemes, which control aliasing errors through numerical
dissipation, in central schemes aliasing must be controlled by en-
A. Plotkin forcing kinetic energy conservation. Such schemes do not exhibit
Downloaded by UNIVERSITY OF OKLAHOMA on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/2.253
mesh similar to the one used by Beaudan and Moin5 were to be used,
the ¯ ow in the region of the separated shear layer would experience
large streamwise grid stretching as it would go from being aligned
with one family of grid lines to the other, and the ¯ ow® eld would be
contaminated with 2-D waves. This problem does not arise when a
C-mesh is used. In the downstream region of the wake where de-
tailed comparisonsbetween the simulationswill be made, both grids
are nearly Cartesian and, therefore, the grid topology is not a factor.
The solution is advanced in time using a second-order accurate,
semi-implicit fractionalstep scheme where the convection-diffusion
terms are advancedfollowed by the pressure correctionstep.9 A ver-
sion of the dynamic model suitable for application in generalized
coordinates has been used. Details of the dynamic SGS modeling
procedurecan be foundin Refs. 1 and 2. Test ® lteringis performedin
the streamwise and spanwise directions, and the least-squares min-
imization approach10 is used for obtaining the model coef® cient.
Furthermore, the model coef® cient is obtained as a spanwise aver-
aged quantity, and the total viscosity is constrained to be positive
through clipping of large negative eddy viscosityvalues. The current Fig. 1 One-dimensional streamwise velocity spectra E11 along the
wake centerline: Ð Ð , Ong and Wallace6 ; ± ± ± , central difference; and
simulation has been carried out on a 401 120 48 (streamwise
£ £
£of wall normal spanwise) mesh with a time-step size (U D t / D)
£ 1 been ², upwind biased.5 Vertical lines indicate the grid cutoff: Ð Ð , central
difference and Ð Ð , upwind biased.5
¢
about 0.007. All statistics for the current simulation have
Downloaded by UNIVERSITY OF OKLAHOMA on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/2.253
Description of Results
It is found that key wall statistics (mean base pressure coef® cient,
mean drag coef® cient, Strouhal number, and mean separationangle)
obtained from the current central difference-basedsimulation are in
good agreement with experiments and show less than 2% deviation
from the results of the upwind-biased simulations.11 Furthermore,
mean velocity pro® les in the near wake (x / D < 3.0) also compare
reasonably well with the pro® les obtained from the upwind-biased
simulations.11 Inasmuch as the drag and base pressure coef® cient
depend strongly on accurate prediction of near-wake features such
as vortex rollup and formation of streamwise vortical structures, Fig. 2 Mean velocity and Reynolds stress pro® les at two downstream
locations: , Ong and Wallace6 ; Ð Ð , central difference; and ± ± ± ,
good prediction of these quantities suggests that the development 4
upwind biased.5
and evolution of the vortical structures in the near wake is being
simulated reasonably accurately.
Because the ¯ ow® elds in the near wake obtained from the two extending from about x / x st = 2±7. Figure 1 clearly shows that the
simulations are in reasonable agreement with each other, we ex- spectra from the current simulation match the experimental spectra
pect that difference in the downstream portion of the ¯ ow (5.0 < much better than the simulation of Beaudan and Moin.5 A closer
x / D < 10.0) will be primarily due to differences in the resolution. look at the spectra at the three locations obtained from the upwind-
Thus, comparison of the statistics in the downstream portion of the biased simulation5 shows that only the energy in the lower 20±25%
wake should allow us to compare the performance of the different of the resolved wave numbers matches with the experiment. On
schemes. Also, based on the estimate of the Kolmogorov length the other hand, in the current simulation the damping at the higher
scale at these downstream locations provided by Ong and Wallace,6 wave numbers is not as severe and spectra in the lower 40±50% of
Beaudan and Moin5 determinedthat the ¯ ow was about three to four the resolved wave number range match well with the experiment.
times better resolved in the vertical and spanwise directionsas com- It might be more appropriate to compare the spectra obtained from
pared to the streamwise direction in their simulations. In the current LES with a suitably ® ltered experimental spectra. However, there
simulation, too, the ¯ ow is better resolved in these directions, thus is no straightforward way of accurately determining the grid ® lter
allowing us to base the comparison between the two simulations function, and we have chosen instead to just indicate the grid cutoff
solely on the streamwise resolution. The streamwise grid spacing in on the spectra plot. Beaudan and Moin5 attributed the marginal per-
the current simulation is 20±30% smaller than Beaudan and Moin’s formance of the upwind-biased schemes in the downstream wake
between x / D = 5 and 7. However, at x / D = 10.0 both simula- region to the dominance of numerical dissipation. Thus, given the
tions have roughlythe same streamwisegrid spacing.This difference fact that the spectra for the current simulationshow better agreement
cannot be avoided because, as mentioned earlier, the grid cannot be with the experiment than Beaudan and Moin’s,5 it is reasonable to
stretched in the streamwise direction as fast in the central difference expectthat the turbulencestatistics obtainedfrom current simulation
simulation as was done in the upwind-biased simulations. will also be better predicted.
In Fig. 1 we have plotted the one-dimensional frequency spectra In Fig. 2 the mean streamwise velocity and normal stress pro® les
of the streamwise velocity at two locationsin the downstream region at these two locations are shown. We observe that the streamwise
of the wake. The current solver utilizes a variable time-step size and, velocity pro® les obtained from both simulations agree reasonably
therefore, the time series obtained is not evenly sampled. To obtain well with the experiment. Furthermore, at x / D = 7.0, the simu-
the spectra from these unevenly sampled data, we have used the lation of Beaudan and Moin5 underpredicts the peak streamwise
Lomb periodogram with an oversampling factor of four.12 Spectra normal stress, whereas the current simulation shows better agree-
from both simulationsand experiment6 are plotted together for com- ment in both the magnitude of the peak stress and shape of the stress
parison. The streamwise grid spacing limits the highest frequency pro® le. However, at x / D = 10.0, streamwise stress pro® les from
that can be locally resolved in the simulation,and this represents the both simulations match quite well, and both underpredict the ex-
implicit ® lter that is imposed by the grid on the ¯ ow® eld. The verti- perimental stress level. Because streamwise Reynolds stress at the
cal lines in the plots indicate the grid cutoffs for the two simulations. wake centerline is directly related to the area underneaththe spectra
The experimentalspectra shows about half a decade of inertial range shown in Fig. 1, it is somewhat surprisingthat the current simulation
AIAA JOURNAL, VOL. 35, NO. 8: TECHNICAL NOTES 1417
does not predict a streamwise stress level that is signi® cantly higher Circular Cylinders at Sub-Critical Reynolds Numbers,º Dept. of Mechanical
than the simulation of Beaudan and Moin.5 However, this can be Engineering, Rept. TF-62, Stanford Univ., Stanford, CA, Dec. 1994.
6 Ong, L., and Wallace, J., ª The Velocity Field of the Turbulent Very Near
explainedby noting that, for this ¯ ow, most of the contributionto the
Reynolds stress comes from ¯ uctuationsin a narrow frequencyband Wake of a Circular Cylinder,º Experiments in Fluids, Vol. 20, No. 6, 1996,
pp. 441±453.
extending from about 0.5x s t to 3.0x st , and in this frequency band, 7 Akselvoll, K., and Moin, P., ª Large-Eddy Simulation of Turbulent Con-
the energy in both simulations is comparable. Thus, even though ® ned Coannular Jets,º Journal of Fluid Mechanics, Vol. 315, May 1994, pp.
the simulations of Beaudan and Moin5 exhibit signi® cant damping 387±411.
8 Lund, T. S., and Moin, P., ª Large-Eddy Simulation of a Concave Wall
of the higher frequencies,this does not have a signi® cant impact on
the low-order turbulence statistics. Boundary Layer,º International Journal of Heat and Fluid Flow, Vol. 17,
By comparing the vertical stress pro® les at these locations, we No. 3, 1996, pp. 290±295.
9 Choi, H., Moin, P., and Kim, J., ª Direct Numerical Simulation of Flow
observe that prediction from the two simulations at x / D = 7.0 is Over Riblets,º Journal of Fluid Mechanics, Vol. 255, Oct. 1994,pp. 503±539.
quite similar. At x / D = 10.0, the two simulations predict roughly 10 Lilly, D. K., ª A Proposed Modi® cation of the Germano Subgrid-Scale
the same peak stress level; however, the shape of the experimental Closure Method,º Physics of Fluids A, Vol. 4, No. 3, 1992, pp. 633±635.
11 Mittal, R., ª Progress in LES of Flow Past a Circular Cylinder,º CTR
pro® le matches the pro® le of Beaudan and Moin5 better than it does
for the current simulation. Furthermore, we have found that vertical Annual Research Briefs, Center for Turbulence Research, Stanford Univ.,
velocity and shear stress pro® les (not shown here) from both the Stanford, CA, 1996, pp. 233±242.
12 Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P.,
simulations are also in reasonable agreement with experiments.11 Numerical Recipes in Fortran, Cambridge Univ. Press, New York, 1992.
Conclusions C. G. Speziale
Associate Editor
It is found that in the downstream portion of the wake, where the
grid is relatively coarse, the numerical dissipation inherent in the
Downloaded by UNIVERSITY OF OKLAHOMA on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/2.253