You are on page 1of 10

Index of Resolution Quality

for Large Eddy Simulations


In the light of rapidly increasing applications of large-eddy simulations (LES), it is
deemed necessary to impose some quality assessment measures for such studies. The
I. B. Celik verification of LES calculations is difficult because of the fact that both the subgrid scale
e-mail: ismail.celik@mail.wvu.edu
(SGS) model contribution and numerical discretization errors are functions of the grid
resolution. In this study, various indexes of quality measures, hereafter referred to as
Z. N. Cehreli LESគIQ, are proposed. The recommended LESគIQ is based on the Richardson extrapola-
tion concept. This method has been applied to various cases and the calculated LESគIQ
I. Yavuz results are compared with the relative total experimental and direct numerical simulation
(DNS) error, defined as IQគEX and IQគDNS, respectively. It is postulated that in practical
Mechanical and Aerospace Engineering, applications of LES, numerical dissipation will always be a significant part of the overall
West Virginia University, P.O. Box 6106, dissipation, and it must be accounted for in any assessment of the quality of LES. It is
Morgantown, WV, 25506-6106 further suggested that LESគIQ of 75% to 85% can be considered adequate for most
engineering applications that typically occur at high Reynolds numbers; the proposed
index is an indicator of good resolution (i.e., verification), but not necessarily a good or
accurate model (i.e., validation). 关DOI: 10.1115/1.1990201兴

Keywords: Numerical Uncertainty, Large Eddy Simulation (LES), Quality Index, Com-
putational Fluid Dynamics (CFD)

1 Introduction tests whereby model predictions are compared to the SGS stresses
obtained from DNS 共or experiments兲 by filtering; 共ii兲 a posteriori
It is anticipated that the use of LES 共large-eddy simulation兲 in
testing whereby actual LES results are compared with results from
engineering applications will increase significantly in the coming DNS or experiments. Vreman et al. 关9兴 dismiss the a priori testing
years. It is already becoming a user option in commercial CFD procedure, arguing that they usually are too pessimistic, since low
共computational fluid dynamics兲 packages such as FLUENT. It is correlations between stresses and predictions do not necessarily
then imperative that there should be some quality control on the lead to poor LES results. Pope 关11兴 also mentions that models that
simulations claimed to be LES. Uncertainty/quality assessment in perform poorly on a priori tests may perform well on a posteriori
RANS 共Reynolds-averaged Navier–Stokes兲 simulations has been tests. The second approach has the well-known difficulty of dis-
investigated at length 关1–4兴 and there are even some guidelines criminating between the numerical or discretization errors and the
关5–7兴 for assessment and reporting of numerical uncertainty in modeling errors 关12兴 in order to determine the magnitude of tur-
such simulations. The assessment of uncertainty in LES is not so bulent energy dissipation rate introduced by the SGS model, i.e.,
trivial, as both the numerical discretization error and the subgrid ␧SGS.
scale contributions are proportional to grid size. As it was rightly The total error 共hence the uncertainty兲 in LES, ␧LES, consists of
pointed out by Speziale 关8兴, a good LES is that which tends to two major components, namely numerical discretization error,
DNS 共direct numerical simulations兲 as the grid resolution tends to ␧num, and modeling 共i.e., subgrid scale, model兲 error, ␧SGS. Al-
the smallest scales, i.e., the Kolmogorov scales. Therefore, there is though these components are usually correlated, for brevity one
no such thing as grid-independent LES in theory, because a grid- can write
independent LES is essentially DNS, and the philosophy of LES
loses its meaning if it is grid independent; the advantage of LES ␧LES = ␧num + ␧SGS , 共1a兲
over DNS being that LES is much more economical while it only
requires the resolution of the most energetic eddies that determine ␧LES = ␶ijhS̄ij + ␶ijrS̄ij 共1b兲
the essential flow quantities and their consequences. If the main where is the numerical stress and ␶rij is the residual 共or subgrid兲
␶hij
purpose of an LES is to capture only the mean flow dynamics, the stress. In a good LES study, for a given filter width ⌬, the grid
above argument may not hold, but that would be a very limited
size h should be chosen such that ␶hij Ⰶ ␶rij. It is important to note
and costly purpose for LES. It is well known that physical phe-
that in some cases ␧num and ␧SGS may have different signs, thus
nomena such as mixing and combustion are strongly dependent on
canceling each other, which would result in a small total error.
intensity of turbulent fluctuations and the convection by these
In Eq. 共1a兲 ␧num must be minimized, which inevitably will re-
fluctuations that eventually exhibit themselves as turbulent diffu-
quire fine-grid LES. This is problematic not only because of the
sion. Hence, in most applications of LES the prediction of turbu- computational cost and time, but also it has the disadvantage that
lence statistics is at least as important as the prediction of mean
in most practical applications of LES as ␧num decreases so does
flow quantities. With such a premise, it is necessary that some
␧SGS. Such calculations seem to constitute the bulk of LES studies
quality assessment measures be formulated for LES geared to-
关13,14兴. Only in rare cases where ␧num goes to zero much more
wards engineering applications. This is the subject of the present
rapidly than ␧SGS would grid independent solutions be feasible.
study.
Consider for example the most widely used Smagorinsky model,
According to Vreman et al. 关9,10兴, there are two approaches for
where the subgrid eddy viscosity vSGS is proportional to ⌬4/3 共Ref.
testing the performance of SGS 共subgrid scale兲 models: 共i兲 a priori
关11兴, p. 588兲, ⌬ being the filter length which is usually taken
proportional to the grid size, h. If the numerical scheme is second
Contributed by the Fluids Engineering Division for publication in the JOURNAL OF order and the order of accuracy degenerates slightly for various
FLUIDS ENGINEERING. reasons such as maintaining stability, then the above procedure

Journal of Fluids Engineering Copyright © 2005 by ASME SEPTEMBER 2005, Vol. 127 / 949

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 09/01/2013 Terms of Use: http://asme.org/terms


will not work. It should be pointed out that LES using second- ⳵ūi
order schemes has many advantages 关15兴, and they are becoming ␧␮ = vS̄ij 共4a兲
⳵x j
more and more popular. Even with higher order schemes, Ref. 关9兴
observed that ␧num and ␧SGS are of comparable order and may
partially cancel each other. Thus, grid refinement may not neces- ⳵ūi
sarily lead to smaller ␧LES. ␧t = − ␶ij 共4b兲
⳵x j
A further concern is that the two approaches mentioned above
that use DNS or experiments as a benchmark will be limited in
most cases to simple flow problems, since DNS will either be too ␶ij = − 共Cs⌬兲2具S̄典S̄ij 共4c兲
costly, and/or impossible for many problems of engineering inter-
est. Moreover, for every application of LES, if DNS or experi-
ments are required for validation and uncertainty assessment, then 具vt典 = 共Cs⌬兲2具S̄典 共4d兲
the predictive purpose of these studies will be diminished and they It follows then
will not be viewed as tools for engineering analysis.
Therefore, a quality assessment procedure is necessary that can 具␧t典v
be applied independent of experimental or DNS data, such that it 具 ⑀ ␮典 = 共4e兲
does not require LES simulations at grid resolutions that are pro- 具 v t典
hibitively costly.
The present study is undertaken to formulate such an assess- Here, we have used the approximation 具S̄典 ⬵ 具S̄2典1/2 along the
ment procedure, which we call “LES Index of Resolution Quality, same lines of thinking as Lilly 关20兴. In Eqs. 共3兲 and 共4兲 Cs is the
in short, LESគIQ.” The procedure is based on the concept of Ri- Smagorinsky constant; vt is the SGS viscosity 共analogous to the
chardson extrapolation 关16,17,3兴. turbulent eddy viscosity兲; v is the molecular viscosity, and S̄ij is
The methods proposed in this paper are aimed at LES studies the mean rate of strain tensor.
where the filter length, ⌬, is implicitly related to the grid cell size Substituting Eq. 共4兲 into Eq. 共2兲 yields
h, and those with at least second-order accuracy in time and space,
but we have made an attempt to generalize them to include other 具 v t典 1
cases. s⬵ = 共5兲
具vt典 + v 共1 + v/具vt典兲
The errors in LES are primarily dependent on the filter length,
⌬, and the grid size, h. Geurts and Froehlich 关18兴 have developed Since in most LES applications it is anticipated that vt Ⰷ v, Eq.
error estimates for LES the philosophy of which are substantially 共5兲 indicates that s ⬵ 1.0, and it is not sensitive to grid resolution.
different than the current paper. Geurts and Froehlich 关19兴 present Thus, it will be difficult to use it as an assessment parameter.
their results in terms of a grid activity parameter 关Eq. 共2兲兴 and Geurts and Froehlich 关19兴 have suggested that the subgrid reso-
subgrid activity, s, and SGS resolution parameter, r = h / ⌬. The lution parameter, r = h / ⌬, and s be regarded as independent indi-
implications of using these parameters are elucidated in the next cators to classify and compare LES solutions. However, as we
section. The rest of the paper deals with testing and assessment of mentioned earlier, in most cases the filter width, ⌬ ⬇ h, which
the newly proposed LESគIQ, with examples to demonstrate the makes r ⬵ 1. As observed by Geurts and Froehlich 关19兴, with in-
feasibility of such an approach. creasing r value, the numerical error increases. The evaluation of
It must be mentioned that it is not the objective of this paper to the activity parameter “s” is difficult; it requires calculation of the
assess the performance or accuracy of SGS models; this falls in volume-averaged turbulent dissipation rate, which inherently in-
the realm of validation. Our aim is more of a verification rather cludes both the modeled dissipation and the numerical dissipation;
than validation 共see Ref. 关3兴, for appropriate definitions of verifi- segregation of the two is necessary but not easy. On the other hand
cation and validation兲. the modified activity parameter s* given by Eq. 共6兲 below can be
used with relative ease, but its derivation involves many assump-
tions and is based on the Smagorinsky model.
2 Subgrid Activity Parameter The s values were calculated from Eq. 共5兲, for the LES of the
Geurts and Froehlich 关19兴 suggested a subgrid activity param- wake behind an Arleigh–Burke class destroyer 共DDG51; see Sec.
eter, s, to be used as an independent parameter in assessment of 4兲 cruising on a circular track on a relatively coarse grid resolu-
errors in LES. It is stated that by definition s = 1 corresponds to tion of 300 K and 1000 K nodes. The results 共Fig. 1兲 indeed show
LES, whereas s = 0 corresponds DNS at infinite Reynolds number, s ⬵ 1.0. This is because in this case, as expected to be in many
Re. This parameter is defined as other cases, 具vt典 Ⰷ v. To bring 具vt典 down to the level of v would
具␧t典 require huge computational resources. This again demonstrates
s= 共2兲 that in a wide class of LES applications “s” is not sufficiently
具␧t典 + 具␧␮典 sensitive to grid resolution 共in that it is fixed approximately at s
Here, 具 典 denotes an averaged 共or filtered兲 quantity, ␧t is the tur- ⬵ 1.0兲. Hence, some other measure needs to be found to assess the
bulent dissipation, and ␧␮ is the molecular dissipation; ␧t can be quality of LES results. For example, recognizing the importance
expressed in terms of the turbulent viscosity. For example, for of numerical dissipation in most LES applications, one could de-
Smagorinsky-type models we have fine a modified activity parameter as

具␧t典 = 共Cs⌬兲2共S̄3兲 ⬵ 共Cs⌬兲2共S̄2兲3/2 共3a兲 具vt典 + 具vnum典


s* = 共6兲
具vt典 + 具vnum典 + v
具vt典 = 共Cs⌬兲 共S̄兲
2
共3b兲
where vnum is the numerical viscosity, and s* is always less than 1.
具 v t典 3
具␧t典 = 共3c兲
共Cs⌬兲4
where S̄ = 共2S̄ijSij兲1/2. The approximation 具S̄3典 ⬵ 具S̄2典3/2 is adopted 3 Proposed LESᠪIQ
from Pope 共Ref. 关11兴, p. 588兲 and it is due originally to Lilly 关20兴. Following a similar line of thinking that is inherent in Eq. 共6兲,
The molecular dissipation can also be related to turbulent eddy one can formulate an LES index of quality based on grid resolu-
viscosity as follows: tion relative to the Kolmogorov length scale, ␩␬, e.g.,

950 / Vol. 127, SEPTEMBER 2005 Transactions of the ASME

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 09/01/2013 Terms of Use: http://asme.org/terms


Fig. 1 The variation of subgrid activity parameter along the
Fig. 2 Index of quality versus h / ␩k or Švt,eff‹ / v
streamwise direction for LES of the wake of a turning ship „Ce-
hreli, 2004 †31‡….
end, Ret = 1200 and one at the low end, Ret = 300. These criteria
demand that for sufficient LES resolution h / ␩␬ and 具vt,eff典 / v both
1 be less than approximately 25. As shown in Appendix A, the

冉 冊
LES គ IQ␩ = m 共7兲 model constants should actually be made functions of the turbu-
h
1 + ␣␩ lence Reynolds number and ⌬ / h, but as a first indicator, we be-
␩␬ lieve that these values would be sufficient. We recommend these
where ␩k = 共v3 / ␧兲1/4; if the dissipation rate ␧ is not known from relations as an approximate quality measure everywhere for wall-
DNS or experiments ␩k should be calculated using its apparent resolved LES and away from the wall when LES is used with wall
value ␧eff = ␧res + ␧num, i.e., the numerical dissipation should not be functions.
ignored. Equation 共A3兲 共Appendix A兲 also indicates that for a good LES
An alternative index can be based on the ratio 具vt + vnum典 / v the grid size should be around h ⬵ ␭ / 8, which fortifies the judg-
= 具vt,eff典 / v, e.g., ment that there should be at least 8 cells to resolve an eddy of size
␭. It follows that, for example, in a channel flow the centerline
1 flow is easy to resolve; however, near the wall ␭ → 0 ⇒ h → 0,

冉 冊
LES គ IQv = 共8a兲
具vt,eff典 n which makes it difficult to resolve wall-bounded flows appropri-
1 + ␣v ately. Note also that ␭ = f共Ret兲; usually as Ret decreases both the
v
integral length scale ␭ and turbulent time scale ␶ increase, which
or in turn causes a moderate decrease in the velocity scale, u = ␭ / ␶
1 and an increase in the Kolmogorov length scale, ␩k. It then fol-

冉 冊
LES គ IQv = 共8b兲 lows that, at low Ret, there is less restriction on the grid size h to
s* n
1 + ␣v obtain a good LES; i.e., with the same grid size, h, one can obtain
共1 − s*兲 a higher LESគIQ, as Ret decreases. Similar arguments should ap-
The parameters ␣␩, ␣v, m, and n in Eqs. 共7兲 and 共8兲 can be deter- ply to Eq. 共8a兲 as well, noting that ␩k ⬃ v and ␭ ⬃ vt.
mined using the rationale explained in Appendix A, that h ⬵ ␩k for In both Eqs. 共7兲 and 共8兲 the molecular viscosity is used 关note
DNS, h ⬵ 25␩k for LES; the corresponding values are 具vt,eff典 / v that ␩k = 共v3 / ␧eff兲1/4兴 with the anticipation that LES are meaning-
⬵ 1 for DNS and 具vt,eff典 / v ⬵ 20 共see Appendix A兲 for LES. In ful only when it is applied to viscous laminar or turbulent flows
where the viscosity plays an important role in the generation of
DNS simulations, the residual viscosity vt is typically about twice
the molecular viscosity 共Ref. 关11兴, p. 603兲. If we consider that an shear and turbulence. Cases where v ⬅ 0 would be meaningless
index of quality greater than 80% is a good LES, and that 95% for LES; hence, these equations should not be applied.
In both of the above expressions the inherent difficulty is the
and above is considered as DNS, this yields m ⬵ 0.5, ␣␩ = 0.05,
calculation of ␩k = 共v3 / ␧eff兲1/4 and vt,eff both including the effect of
n ⬵ 0.53, ␣v = 0.05. Concerning resolution of DNS and LES, Pope
关11兴 suggests that 80% of the energy be resolved everywhere for numerical dissipation. It seems more practical to make use of the
LES with near-wall resolution. Moreover, in three-dimensional resolved turbulent kinetic energy kres versus the total ktot = kres
flow using the Kolmogorov spectrum for high Reynolds number + kSGS + knum, the latter two terms being the contribution from the
homogeneous isotropic turbulence, it can be shown 共Ref. 关11兴, p. SGS model and the numerical dissipation, respectively. If one can
577兲 that 80% of the energy is resolved when ⌬ ⬵ ␭, where ␭ is make the assumption that keffគSGS = 共ktot − kres兲 scales with the mesh
the integral length scale. To give an idea of the relative magnitude size h, then an index of quality similar to Eqs. 共7兲 and 共8兲 can be
of subgrid viscosity vt versus vnum, we refer to Froehlich and Rodi defined as 关see also Eq. 共18兲兴
关14兴. They reported that when the QUICK scheme was used, kres kres
vt / v ⬵ 1.0 and vnum / v ⬵ 180 were typical values near the center- LES គ IQk = = 共9兲
ktot kres + akh p
line of a channel flow simulated by employing a Smagorinsky
model. The functions 共7兲 and 共8兲 are plotted in Fig. 2 for two Here, kres is the resolved kinetic energy, ktot is the total kinetic
values of the turbulence Reynolds number Ret, one at the high energy, h is the grid size, ak is a coefficient to be determined, and

Journal of Fluids Engineering SEPTEMBER 2005, Vol. 127 / 951

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 09/01/2013 Terms of Use: http://asme.org/terms


p is the order of accuracy of the numerical scheme. context of “incomplete resolution” in LES, meaning that velocity
In the spirit of Richardson’s extrapolation, the error in any con- fluctuations that correspond to small scales are filtered out; hence,
tinuous dependent variable ␾ can be expressed as a polynomial in the turbulence field is not fully resolved. In other words, in LES it
h, i.e., ␾o − ␾h = a1h + a2h2 + ¯, here ␾h is the numerical solution is expected that, in general, ⌬kLES ⬎ 0 and monotonic conver-
on grid size h, and ␾o is the extrapolated solution as h → 0 共see gence to “exact” solution is obtained, as far as the turbulent ki-
Roache, 1998 关3兴, for more details兲. The effective SGS kinetic netic energy is considered. Unfortunately, this is not always the
energy can be approximated as case 共see for example Sreedhar and Stern’s results 关22兴兲, where
the coarse-grid simulations are in some regions closer to the DNS
ktot − kres effគSGS
1 = k1 = akh1p 共10兲 and experiments than the fine-grid simulations. In some cases,
convergence seems to be oscillatory and DNS and experiments do
ktot − kres effគSGS
2 = k2 = akh2p 共11兲 not agree well. There are also such cases, for example channel
where subscripts 共1兲 and 共2兲 denote quantities obtained on mesh flows, where it so happens that the coarse-grid kinetic energy
共1兲 and mesh 共2兲, respectively. In the present application, h values are higher near the wall compared to the fine-grid values or
= 共⌬x⌬y⌬z兲1/3, where ⌬x , ⌬y , ⌬z are the grid cell lengths in the higher than the DNS or experiments 共see Refs. 关23,14,24,25兴兲. It
x-, y-, z-directions, respectively. This relation can be used for is not clear why such trends arise in LES. A discussion on this
calculating a local mesh size if the details of the grid are known. issue is presented in Appendix B. Nevertheless, in those situations
However, in cases where the local grid information is not avail- Eq. 共9兲 should be modified as
able, a global grid index parameter can be defined as 兩ktot − kres兩

冋 册
LES គ IQk = 1 − 共18兲
1/3 ktot

1
heq = ⌬∀i 共12兲
N and ktot − kres = keffគSGS = akh p be approximated in the same way as
where ⌬∀i is the volume of a computational cell, N is the total before.
number of cells, and the summation is over all the cells. Equation In the above formulation it is assumed that the mesh size h and
共12兲 has been used in the literature 共see Refs. 关3,2,1兴兲 successfully the filter width ⌬ are approximately equal to each other. Hence,
to characterize both structured and unstructured grids, but it is the contributions from subgrid scale modeling and the discretiza-
more suitable for structured and geometrically similar grids. In tion errors can be lumped together as in Eqs. 共10兲 and 共11兲. How-
writing Eqs. 共10兲 and 共11兲, it is assumed that the leading term in ever, when the filter length is different than the grid size 共see for
the truncation error series is dominant; in other words, the grid example Ref. 关18兴兲, then the resolution index should be modified
resolution is in the asymptotic range 共see Ref. 关3兴兲. We should also to account for these two major components separately
mention that Richardson extrapolation works well for monotonic ktot − kres = keffគSGS = ah p + b⌬q 共19兲
grid convergence; if the convergence is oscillatory it should be
used with caution 共see Ref. 关21兴兲. where p denotes the order of accuracy of the numerical scheme
Equations 共10兲 and 共11兲 can be solved for ak to yield and q denotes the order of the modeling error, and a and b are the

冋 册
constants. Hence, the expression for LESគ IQk may be rewritten as
1 kres res
2 − k1
ak = 共13兲 kres
h2 ␣ − 1
p p LES គ IQk = res 共20兲
k + ah p + b⌬q
where ␣ = h1 / h2 ⬎ 1 is the grid refinement parameter.
If the order, p, of the discretization error is known and it is
Hence, the expression for LESគ IQk becomes
assumed that usually the SGS modeling introduces a second-order
kres dissipation term 共i.e., q = 2兲, Eq. 共20兲 can be evaluated with three

冉冊
LES គ IQk = 共14兲 sets of grids. If q = p Eq. 共20兲 reduces to Eq. 共9兲 with a modified
共kres
2 − k1 兲
res
h p
k res
+ constant a*k = 共ar p + b兲 共note: r = h / ⌬兲; then, two sets of calcula-
␣ −1
p
h2 tions on sufficiently refined grids would suffice. It should be noted
Rearranging Eq. 共14兲 for both grid sizes yields that the first term on the right-hand side of Eq. 共19兲 represents the
numerical/discretization error, and the last term represents essen-
Fine-grid index: tially the modeling error. However, due to the convolution of these
1 errors during an actual simulation, it is usually not possible to

冉 冊
LES គ IQkf = 共15兲 segregate the two from each other unless one is made dominant
kres
1 over the other by design.
1+ 1− 共␣ p − 1兲−1
kres
2
Coarse-grid index: 4 Application and Results
1 The proposed quality index concept has been applied to six

冉 冊
LES គ IQck = 共16兲 different cases, namely the LES of flow around a square cylinder
kres
2 关26兴, case 1; the turbulent mixing layer 关27兴, case 2; channel flow
1+ − 1 ␣ p共␣ p − 1兲−1
kres
1 关28兴, case 3; surface piercing flat-plate boundary layer 关22兴, case
4; ship wake flow 关29–31兴, case 5; and vertical buoyant jet flow
Furthermore, a relative index based on experimental or DNS 关32兴; case 6. The formal order of the schemes used in all six of
results can be defined as these cases was second order both in space and time discretiza-
tion. We also utilized the mixing layer simulations by Geurts and
兩kexp − kres兩 Froehlich 关18兴 to study the effect of filter size versus grid size. In
IQ គ EX = 1 − 共17a兲
kexp what follows we only briefly describe these cases taken from al-
ready published literature. The reader is referred to the original
兩kDNS − kres兩 reference for more details.
IQ គ DNS = 1 − 共17b兲 For the first case, two calculations were used with 185⫻ 105
kDNS
⫻ 25 共case 1.1兲 and 265⫻ 161⫻ 25 共case 1.2兲 computational cells,
The absolute value in Eq. 共17兲 is necessary for cases where where the grid distribution was uniform with a constant cell size,
⌬kexp = 共kexp − kres兲 or ⌬kDNS = 共kDNS − kres兲 is less than zero. Theo- h. The resolved turbulent kinetic energy values 关26兴 are presented
retically, both of these quantities should be larger than zero in the in Fig. 3共a兲 for two different sets of grids. We applied Eqs. 共15兲

952 / Vol. 127, SEPTEMBER 2005 Transactions of the ASME

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 09/01/2013 Terms of Use: http://asme.org/terms


Fig. 3 „a… The resolved kinetic energy. „b… Large-eddy simula-
tion index of quality „LESᠪIQ… †Eqs. „15… and „16…‡ for flow
around a square cylinder „Sohankar et al., 1999 †26‡…. „The Fig. 4 „a… The resolved turbulence kinetic energy. „b… LESᠪIQ
numbers in parentheses in the legend indicate grid nodes in †Eq. „18…‡ and IQᠪEX †Eq. „17…‡ for the mixing layer simulation
the x- , y- , z directions, respectively…. „Badeau, 2003 †27‡…. „The numbers in parentheses in the legend
indicate grid nodes in the x- , y- , z directions, respectively….

and 共16兲; the calculated quality index for case 1.1 and case 1.2 is periments include the periodic fluctuations due to the coherent
shown in Fig. 3共b兲. It is seen that for both grid resolutions, Kelvin–Helmholtz waves. This complicates the analysis further. It
LESគIQ is approximately 95%. Sohankar et al. 关26兴 have stated seems that Eq. 共18兲 is more appropriate for this case. The calcu-
that their results with these two grid resolutions were in good lated LESគIQ from Eq. 共18兲 is depicted in Fig. 4共b兲 along with the
agreement with experiments, reflecting the high LESគIQ we true index IQគEX calculated from Eq. 共17兲. The proposed LESគIQ
found. for both coarse- and fine-grid resolution follows closely the
The kinetic energy values for mixing layer flow simulations IQគEX values based on experimental values of k. The observed
关27兴 with grids of 200⫻ 80⫻ 40 共case 2.1兲 and 300⫻ 120⫻ 60 LESគIQ values indicate the quality of LES is good in the middle
共case 2.2兲 along with approximated experimental results 关33兴 are section of the shear layer but it deteriorates considerably near the
shown in Fig. 4共a兲. Here, kexp was calculated by assuming vrms outer boundaries; the quality is particularly low 共50%兲 for the
= wrms; Rightley 关33兴 reported only urms and vrms values. The di- coarse-grid resolution near the boundaries of the calculation do-
mensions of the calculation domain used by Badeau were 0.55 main.
⫻ 0.2⫻ 0.2 m, and uniform grid was used in all three directions. For the channel flow 共case 3兲, two calculations were carried out
Using Eqs. 共15兲–共17兲, LESគIQ and IQគEX are obtained for this with 343 共case 3.1兲 and 663 共case 3.2兲 grids with a nonuniform
case shown in Fig. 4共a兲, where ⌬kres attains both positive and grid distribution in the vertical direction 关28兴. The Reynolds num-
negative values. When the original formulation 关Eq. 共9兲兴 sug- ber, based on the half-channel height, is 33 000. These calcula-
gested by Celik et al. 关34兴 was used, the LESគIQ values in some tions were performed with an imposed pressure gradient corre-
locations were higher than 1.0, which is unrealistic. This can be sponding to Re= 33 000; only the fluctuating part of the pressure
due to the measurement errors or a high level of backscatter in the was calculated. This practice, combined with the periodicity of the
LES simulations 共see also the discussion in Appendix A兲. It flow in axial and spanwise directions, makes the LES of channel
should be pointed out that in this case both LES results and ex- flow easier. Hence, acceptable results can be obtained away from

Journal of Fluids Engineering SEPTEMBER 2005, Vol. 127 / 953

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 09/01/2013 Terms of Use: http://asme.org/terms


Fig. 6 „a… The resolved turbulence kinetic energy. „b… LESᠪIQ
†Eqs. „15… and „16…‡ for surface piercing flat-plate boundary
Fig. 5 „a… The resolved turbulence kinetic energy. „b… LESᠪIQ layer simulation „Sreedhar and Stern, 1998 †22‡… for the uniform
†Eq. „18…‡ and IQᠪEX †Eq. „17a…‡ for the channel flow simulation grid and 3% stretching in the transverse direction „The num-
„Shi, 2001 †28‡…: Curve fit to experiments was done by the bers in parentheses in the legend indicate grid nodes in the
present authors. „The numbers in parentheses in the legend x- , y- , z directions, respectively….
indicate grid nodes in the x- , y- , z directions, respectively….

the walls with relatively coarse grid resolutions. The computed space. Here, the details of the grid distribution are not known. We
turbulent kinetic energy and curve fit to the measurements of Hus- calculated the global grid index parameter from Eq. 共12兲 and
sain and Reynolds 关35兴 are shown in Fig. 5共a兲. Equation 共18兲 was LESគIQ from Eqs. 共15兲 and 共16兲. It is seen from Fig. 6共b兲 that the
used to compute LESគIQ; the results are shown in Fig. 5共b兲. calculated index of quality gives values in the range from approxi-
Shown also in this figure is the quality index IQ-EX based on mately 80%–100%. This range is consistent with the conclusions
experimental results 关Eq. 共17a兲兴. Again, the suggested quality in- drawn by Sreedhar and Stern 关22兴. When we used the 3% grid
dex is in good agreement with those obtained using the experi- stretching in the transverse direction, the LES index increased by
ments. It is noteworthy to observe that the coarse-grid indices near about 10% from 0.8 to 0.9. This shows that whenever possible,
the wall indicate less than 50%resolution, which is not surprising local grid refinement parameter should be used in calculation of
for wall boundary layers. LESគIQ.
For case 4, two calculations with 65⫻ 49⫻ 109 grids 共case 4.1兲, The application of the quality index, LESគIQ, is also done for
and with 97⫻ 65⫻ 129 grids 共case 4.2兲 are studied and in all these two more not as well resolved LES cases, case 5, namely the LES
simulations, the mesh size is taken to be uniform in all three of the flow in the wake of a ship cruising on a straight track and
directions; however, the actual grid was nonuniform in the trans- that of a ship cruising on a circular track 关29,30,21兴. For brevity,
verse direction 关22兴. The dimensions of the calculation box are we show here only the results from a ship cruising on a straight
共2␲ / ␣兲␦ in the streamwise direction, 40␦ in the transverse wall track. These studies examined the dynamics of turbulent flow in
normal direction, and 20␦ in the spanwise direction, where ␦ is the the wake of a surface ship using the LES technique with a three-
reference length and ␣ is the wave number of the linear stability dimensional incompressible Navier–Stokes solver, using nonor-
wave explained briefly in Sreedhar and Stern 关22兴. The turbulent thogonal curvilinear coordinates. The LES technique has been ap-
kinetic energy profiles are shown in Fig. 6共a兲. This case is used to plied throughout these studies in conjunction with a new random
illustrate the sensitivity of LESគIQ on variation of grid size, h, in flow generation 共RFG兲 technique originally developed at West

954 / Vol. 127, SEPTEMBER 2005 Transactions of the ASME

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 09/01/2013 Terms of Use: http://asme.org/terms


Fig. 7 „a… Resolved turbulence kinetic energy normalized with
respect to its inlet value, kinlet = 3.3E − 0.3 „ship cruising on a
straight track…. „b… LESᠪIQ †Eqs. „15… and „16…‡ along centerline
of the wake. „The numbers in parentheses in the legend indi-
cate grid nodes in the x- , y- , z directions, respectively….

Fig. 8 Vertical buoyant jet. „a… Comparison of resolved TKE


from DREAM® simulations „300Ã 238Ã 238; 150Ã 107Ã 107;
Virginia University to provide unsteady inflow boundary condi- and 75Ã 56Ã 56… to experiments „Anwar, 1969 †40‡… at x / Din
tions 关36,37兴. The formal order of the scheme is second order both = 20 „normalized with respect to the square of the inlet veloc-
in space and time discretization. ity…, „b…. Corresponding LESᠪIQ and IQEX of the above data
*
Shi 关28兴 and Shi et al. 关29兴 applied LES to the wake flow of the „wrms for the experimental data was estimated by the authors
as wrms · vrms….
ship model DTMB 5512 关38,39兴. A 3.048 m long unpropelled
model of a modern U.S. Navy fleet ship, Arleigh-Burke class de-
stroyer, DDG51, with a Reynolds number of 4.65⫻ 106 cruising
on a straight track, has been investigated. The computational do- ent nonuniform grids were used, being 75⫻ 57⫻ 57, 150⫻ 107
main was 1.5⫻ 0.3⫻ 0.6 共given in nondimensional units in ship ⫻ 107, and 300⫻ 238⫻ 238. A uniform grid is used in the axial
length兲 in the x-, y-, z-directions, respectively, with a grid size of direction for all simulations, with a nonuniformity being imposed
162⫻ 50⫻ 66 共⬃500 K兲 and 322⫻ 50⫻ 66 共⬃1000 K兲. Nonuni- in the spanwise and transverse directions. The first grid size for
form grid spacing with stretching less than 3% was used in both the fine grid, away from x = 0 and z = 0, equaled 0.000 38 m; and
the x- and y-directions. the coarse grid equaled 0.000 68 m. The experiments by Anwar
The resolved turbulent kinetic energy profile for the above- 关40兴 ensured that the flow was turbulent after x / D = 5, where x is
mentioned ship wake simulation is shown in Fig. 7共a兲. We applied the axial distance and D is the inlet diameter of the jet. The inlet
Eqs. 共15兲 and 共16兲 to the results of ship wake calculations. The mass flux of the primary water jet equaled 20.37 kg/ s / m, which
calculated quality index with two different grid resolutions for the yields an inlet Reynolds number of 1892 for the primary stream.
ship wake cruising on a straight track is shown in Fig. 7共b兲. For The inlet densimetric Froude number was 2.8.
the fine grid, LESគIQ in the near-wake region is ca. 55% com- Figure 8共a兲 shows the resolved turbulent kinetic energy, as
pared to a ca. 30% for the coarse grid. The quality deteriorates compared to the experimental data. The simulation data have been
downstream where the grid resolution is coarser 关see Fig. 7共b兲兴. time averaged over two flow-through times, as well as spatially
The low LESគIQ values indicate that for this case more grid re- averaged over eight nearest-neighbor cells on the centerlines. It is
finement is necessary. seen that the 75⫻ 57⫻ 57 grid resolution clearly underpredicts the
The quality indices are also applied to a study conducted by magnitude as compared to the experiments by as much as 50%.
Badeau and Celik 关32兴, on Vertical buoyant jet plumes, case 6, However, at the finer resolutions, agreement with experimental
that was investigated experimentally by Anwar 关40兴. Three differ- data is good. The LESគIQ and IQគEX are presented in Fig. 8共b兲

Journal of Fluids Engineering SEPTEMBER 2005, Vol. 127 / 955

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 09/01/2013 Terms of Use: http://asme.org/terms


order contribution, dependent on the filter width. Although the
spatial discretization scheme used by Geurts and Froehlich 关18兴
for convective terms was fourth order, the current procedure
yielded approximately second order, i.e., p = 2, that is consistent
with the order of the scheme they used for the viscous terms. The
LESគIQ 关Eq. 共20兲兴 and IQគDNS 关Eq. 共17b兲兴 have been calculated
using the grids 643 共coarse兲 and 963 共fine兲 where ⌬ = 2h, and the
results are shown in Fig. 9共b兲. The LESគIQ values fall in the range
85%–100%, and they follow the same trend as the IQគDNS re-
sults; however, the coarse- and fine-grid indices are lower than the
corresponding IQគDNS results, which is desirable to assure some
degree of conservatism in the calculated LESគIQ values. Calcula-
tions were also carried out for the case ⌬ = 4h. The results 共not
shown here兲 were similar to the case with ⌬ = 2h.
It may be thought that the suggested practice of p = 2 and q
= 2 might lead to significant uncertainty if the actual p were to be
much larger than the p used. To assess this, we did calculations
with p = 4. This resulted in ca. 10% improvement in the LESគIQ
which can be absorbed into the conservatism of the proposed in-
dex.

5 Conclusions
Various indices to measure the quality of resolution measures
for large-eddy simulations have been formulated and the underly-
ing rationale is elucidated. The formulation given by Eq. 共9兲 关or
Eq. 共18兲兴 targets specifically those applications where the filter
length and the mesh size are directly proportional to each other,
and the numerical dissipation and subgrid scale contribution to
dissipation of energy are both functions of mesh size. In such
applications, Richardson extrapolation can be used to estimate the
effective subgrid turbulent kinetic energy, including the contribu-
tions from both the modeling and discretization, hence leading to
an estimate of the apparent total kinetic energy. The proposed
index is a measure of the percentage of the resolved turbulent
kinetic energy to the total. Equation 共9兲 关or Eq. 共18兲兴 is recom-
mended for those cases where h ⬵ ⌬ and the order of accuracy of
the numerical scheme is approximately the same as that order
implied of SGS modeling. On the other hand, for ⌬ ⬎ h, LESគIQ
Fig. 9 „a… Resolved turbulence kinetic energy decay versus
time for ⌬ / h = 2; „b… Calculated LESᠪIQ †Eq. „20…‡ and IQᠪDNS
formulation should be modified such that the effective subgrid
†Eq. „17…‡. A curve fit to the turbulent kinetic energy data from turbulent kinetic energy is a function of both the filter width and
Geurts and Froehlich „2001… †18‡ was used in our analysis. the mesh size. Then, Eq. 共20兲 should be used assuming that p
= min共p , q兲. In situations where the values of p or q are not
known, it is recommended that p = q = 2 be used. This recommen-
for this case. Although the index indicates a good LES when dation is based on the assumption that, on one hand, the numerical
fine-grid distributions are used, it appears that the order of the schemes used for LES must at least be second order both in time
monotonic scheme used by Badeau and Celik 关32兴 reduces in the and space discretization, and on the other hand the quality or
case of a coarse grid due to the use of the power-law scheme. uncertainty measures should have some degree of conservatism. If
When this is taken into account by letting p = 1.0 the LESគIQ and more conservatism is desired the coarse-grid index rather than the
IQគEX agree quite well. fine-grid index can be used, or the fine-grid index can be used
with p = 1.0. We believe that the suggested quality index can be
Influence of Filter Size versus Grid Size. Geurts and Froe- applied everywhere in the calculation domain, preferably using a
hlich 关18兴 simulated turbulent flow in a mixing layer. DNS results local mesh size for LES with wall resolution; for LES with wall
with uniform grids that consisted of 1923 nodes and LES results functions it should be applied away from the walls.
with grids that consisted of 323, 483, 643, 963 nodes were re- Applications of the proposed quality index to a range of case
ported. The formal accuracy is second order in time, fourth order studies from the literature and comparison with DNS and experi-
in space for convective terms, and second order for viscous and ments show that newly proposed LESគIQ is a good indicator of
subgrid fluxes. The original study was completed for four cases the resolution quality in LES. When possible, the proposed index
with ⌬ / h = 1 , 2 , 3 , 4. However, grid convergence is not monotonic of quality should be supplemented by providing additional infor-
for the case ⌬ / h = 1. We calculated LESគIQ values for two cases mation on relative mesh size 关e.g., Eq. 共7兲兴, and relative magni-
with ⌬ / h = 2 , 4. The resolved kinetic energy profiles for these two tude of the subgrid viscosity 关e.g., Eq. 共8兲兴 along with spectral
cases are shown in Fig. 9共a兲 for a range of ⌬ / h = r values along information on resolved turbulent kinetic energy in frequency and
with the DNS results. wave number domains.
The four sets of kinetic energy values from cases with ⌬ / h The reader must be warned against a misconception that “good
= 2 , 4 are used to calculate the values of p, q, a, and b described in grid resolution indicates a good model or a good accuracy!” The
Sec. 3. The order of the SGS contribution, q, is calculated to be proposed LESគIQ is an indication of sufficient grid resolution so
approximately 2. This seems to be consistent with the general that the calculations can qualify as LES, and that the SGS model
observation that the Smagorinsky model introduces a second- reveals its characteristics. In this regard LESគIQ should be con-

956 / Vol. 127, SEPTEMBER 2005 Transactions of the ASME

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 09/01/2013 Terms of Use: http://asme.org/terms


sidered as a verification index rather than a validation index; the vt + vnum 共kres兲1/2L
latter necessarily requires comparison with DNS or physical ex- = 0.014 共A8兲
v v
periments, and involves assessment of accuracy.
Using the definition ␩k = 共v3 / ␧兲1/4 and Eq. 共A2兲 with ␭ = 0.55L,
Eq. 共A8兲 yields
Acknowledgments vt + vnum
This work has been partially supported by the Office of Naval = 0.017 Ret 共A9兲
v
Research 共ONR兲 as part of a DOD EPSCOR project. The techni-
cal program monitor is Dr. Patrick Purtel. We also acknowledge In deriving Eq. 共A9兲, the assumed 50% contribution to the unre-
the partial funding kindly provided by the WVU Research Coop- solved scales by numerical dissipation may be considered high,
eration, National Research Center for Coal and Energy 共NRCCE兲, but most probably this is not an exaggeration when practical en-
and the College of Engineering and Mineral Resources 共CEMR兲. gineering applications of LES are considered. For a high turbu-
We thank Lars Davidson of Chalmers University of Technology, lence Reynolds number of Ret = 1200 and low Ret = 300, Eq. 共A9兲
Steve Jordan of Naval Undersea Warfare Center, and Ugo Pi- gives veff / v to be approximately 20 and 5.
omelli of University of Maryland for their valuable discussions.
We also thank B.J. Geurts and J. Froehlich for their kind help in
providing references and data when needed. Appendix B: On The Consistence of LES With DNS
B.1 Celik’s question. While reviewing literature on predic-
tion of turbulent kinetic energy using large-eddy simulation, I
Appendix A: Rationale for Eqs. (7) and (8) in the Main have noticed that in some cases the calculated resolved turbulent
Text kinetic energy is larger than DNS and/or experiments. This is
It can be shown 共Pope, 2000, 关11兴兲 that 80% of the turbulent particularly common in wall-bounded flows. Intuitively, I thought
kinetic energy can be resolved when L / ⌬ ⬵ 12 for the sharp cutoff the resolved TKE and LES should always be smaller than 共i.e., a
filter and L / ⌬ ⬵ 17 for the Gaussian filter in case of high- fraction of兲 DNS results. Because there is usually higher energy
Reynolds number homogeneous isotropic turbulence with the cut- dissipation in simulations of LES, I cannot explain the reverse
off wave number ␬c = ␲ / ⌬. Here, L = k3/2 / ␧ is a length scale of trend. Could you provide your opinion on this?
turbulence and ⌬ is the filter width, k is the turbulent kinetic B.2 Reply by L. Davidson. Often poor resolution in wall-
energy, and ␧ is the dissipation rate of turbulence. Noting that ␭ bounded flows gives rise to too high resolved TKE. I think the
⬵ u3 / ␧ and u ⬵ 冑2 / 3k, we can relate the integral length scale ␭ to reason is that the resolved strain S̄ij gets too small and hence the
L by ␭ = cL, where c ⬵ 0.55. Taking an average value of 共12
resolved dissipation ␧SGS = 2vSGSS̄ijS̄ij gets too small. You’re right
+ 17兲 / 2 = 14.5 yields when you say that one could expect the resolved dissipation to be
␭/⌬ ⬵ 8 共A1兲 larger in LES than in DNS 共because vSGS ⬎ v兲, but then one for-
gets that the resolved dissipation is very dependent on the re-
The Kolmogorov length scale is given by solved strain rate tensor.
␩k = ␭ Ret−3/4 共A2兲
B.3 Reply by S. Jordan. This problem is a common occur-
where Ret = u␭ / v is the turbulence Reynolds number. rence when insufficient resolution is supplied in the spanwise di-
Equations 共A1兲 and 共A2兲 can be used to rewrite 共A1兲 as rection. The peak turbulent intensities can be twice the correct
values in the wake of bluff bodies if given poor spanwise resolu-
h Ret3/4 tion. This result can be easily seen in the coupled equation be-
= 共A3兲
␩k 8共⌬/h兲 tween pressure and velocity 共dot product of momentum兲.
For a typically high Ret ⬵ 1200 and ⌬ ⬵ h, Eq. 共A3兲 yields h / ␩k B.4 Discussion by I. Celik. Usually, the 具vv典 and 具ww典 are
⬵ 25 共Ret ⬵ 300⇒ h / ␩k ⬵ 9兲. Moreover, in the context of the lower than DNS data, while 具uu典 is higher; the coarser the calcu-
Smagorinsky SGS model, from the scaling of quantities in the
lation the higher 具uu典 共see, e.g., Froehlich and Rodi, 2002 关14兴兲.
inertial subrange the SGS residual viscosity is given by 共Pope,
2000, 关11兴兲 Froehlich and Rodi also observed that the discrepancy between
LES and DNS diminished rapidly by improving the resolution in
vt ⬵ Cs2⌬4/3␧1/3 共A4兲 the streamwise and the spanwise directions. But, still the resolved
turbulent kinetic energy in LES may attain higher values than that
A reasonable model for vnum is
seen in DNS. If we assume that in the LES of channel flow the
vnum ⬵ Cvhknum
1/2
共A5兲 production 具uv典 dU / dy is predicted more or less correct, then the
where Cv is a constant, and knum is the residual kinetic energy slope of the mean velocity profile should be more or less correct
caused by numerical dissipation. near the wall. The apparent dissipation rate ␧ ⬵ u3 / ␭ should bal-
Using the results mentioned above, i.e., L / ⌬ ⬵ 14.5, when 80% ance this production term. It appears that the apparent length scale
of the turbulent kinetic energy is resolved, and kres ⬵ 0.8k, L can in LES is usually too small near the wall, hence leading to large
be related to kres as rms values in order to balance the correctly predicted production.
On the other hand, if the mean flow is not right, it is under-
共kres/0.8兲3/2 1.4共kres兲3/2 standable that LES with a dissipative SGS model might give
L= = 共A6兲 larger fluctuations. What might be the reason for not predicting
␧ ␧
the mean flow? The reason probably is that the grid resolution is
One can then show that not sufficient to capture the slope of the mean velocity profile

vt + vnum = 1.12Cs2 冉冊 ⌬
L
4/3
L共kres兲1/2 + ch共knum兲1/2 共A7兲
correctly. This in effect will lead to inaccurate mean pressure gra-
dient in the streamwise direction, which in turn, may lead to
wrong strain rates that are used in SGS model. Then, the dissipa-
Also, Cs = 0.17, h ⬵ ⌬, c ⬵ 0.55, L / ⌬ ⬵ 14.5, and knum / kres tion rate could be underpredicted, thus polluting the overall re-
⬵ 0.125 共i.e., numerical dissipation contributes 50% of the unre- sults. Also, when the grid is insufficient to resolve the near-wall
solved energy budget; this is an intuitive approximation for a rea- energy carrying eddies, streaks, hairpins etc., an overshoot in tur-
sonably good quality LES兲; it follows that bulent kinetic energy may occur. As noted by Pope 共2000 关11兴兲,

Journal of Fluids Engineering SEPTEMBER 2005, Vol. 127 / 957

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 09/01/2013 Terms of Use: http://asme.org/terms


for low Reynolds number flows there is considerable overlap be- Application to the Stresses in a Masonary Dam,” Philos. Trans. R. Soc. Lon-
don, Ser. A, 210, pp. 307–357.
tween energy-containing eddies and dissipative ranges of scales. 关17兴 Richardson, L. F., and Gaunt, J. A., 1927, “The Deferred Approach to the
In LES of low Reynolds number flows the filter is likely to be Limit,” Philos. Trans. R. Soc. London, Ser. A, 226, pp. 299–361.
placed in this overlapping region. This argument is also applicable 关18兴 Geurts, B. J., and Froehlich, J., 2001, Numerical Effects Contaminating LES; A
to inherently low Reynolds number flows near a wall where the Mixed Story. Modern Strategies for Turbulent Flow Simulation, B. J. Geurts,
segregation of “large” and “small” scales is not possible. ed., Edwards Publishing, pp. 317–347.
关19兴 Geurts, B. J., and Froehlich, J., 2002, “A Framework for Predicting Accuracy
Recent observations also indicate that bad SGS models cannot Limitations in Large Eddy Simulation,” Phys. Fluids, 14, pp. 41–44.
predict transition to turbulence at the proper Reynolds number and 关20兴 Lilly, D. K., 1967, “The Representation of Small-Scale Turbulence in Numeri-
at the proper location. When this happens there is a phase shift in cal Simulation Experiments,” Proceedings of the IBM Scientific Computing
the profiles of turbulence quantities. For example, in LES of free- Symposium on Environmental Sciences, IBM-Form No. 320-1951, Yorktown
Heights, NY.
shear layers, if the location of transition in the axial direction is 关21兴 Celik, I. B., Li, J., Hu, G., and Shaffer, C., 2004, “Limitations of Richardson
not predicted right, the peak rms values are not predicted at the Extrapolation and Possible Remedies for Estimation of Discretization Error,”
right location; hence, at some points the LES results would show HT-FE2004-56035, Proceedings of the ASME Heat Transfer/Fluids Engineer-
a higher turbulent kinetic energy than the corresponding DNS ing Summer Conference, 11–15 July, Charlotte, NC.
关22兴 Sreedhar, M., and Stern, F., 1998, “Large Eddy Simulation of Temporally
results. Similar phenomenon will occur in wall layers, where there Developing Juncture Flows,” Int. J. Numer. Methods Fluids, 28, pp. 47–72.
is always transition from laminar 共viscous sublayer兲 to turbulent 关23兴 Temmerman, L., Leschziner, M. A., and Hanjalic, K., 2003, “A Combined
flow. As long as the integral kinetic energy agrees well with DNS RANS-LES Strategy With Arbitrary Interface Location for Near-Wall Flows,”
results, this is not necessarily a shortcoming of LES, other than 3rd International Symposium on Turbulence and Shear Flow Phenomena, Sen-
dai, Japan.
the fact that the location of transition is not predicted accurately, 关24兴 Stolz, S., Adams, N. A., and Kleiser, L., 2001, “An Approximate Deconvolu-
which is in general a very difficult task. tion Model for Large-Eddy Simulation with Application to Incompressible
Wall-Bounded Flows,” Phys. Fluids, 13, pp. 997–1015.
关25兴 Piomelli, U., 1993, “High Reynolds Number Calculations Using the Dynamic
References Subgrid-Scale Stress Model,” Phys. Fluids A, 5, pp. 1484–1490.
关1兴 “Quantification of Uncertainty in Computational Fluid Dynamics,” 1993, Ce- 关26兴 Sohankar, A., Davidson, L., and Norberg, C., 1999, “A Dynamic One-
lik, I., Chen, C. J., Roache, P. J., and Scheurer, G., eds. ASME Publ. No. Equation Subgrid Model for Simulation of Flow Around a Square Cylinder,”
FED-Vol. 158, ASME Fluids Engineering Division Summer Meeting, Wash- in Engineering Turbulence Modelling and Experiments 4, W. Rodi and D.
ington, DC, 20–24 June. Laurence, eds., pp. 227–236.
关2兴 Celik, I., and Karatekin, O., 1997, “Numerical Experiments on Application of 关27兴 Badeau, A. E., 2003, “Simulation of Buoyant Flows Using Implicit Turbulence
Richardson Extrapolation With Nonuniform Grids,” ASME J. Fluids Eng., Modeling,” Ph.D. dissertation, West Virginia University, Morgantown, WV.
119, pp. 584–590. 关28兴 Shi, S., 2001, “Towards Large Eddy Simulation of Ship Wakes,” Ph.D. disser-
关3兴 Roache, P. J., 1998, Verification and Validation in Computational Science and tation, West Virginia University, Morgantown, WV.
Engineering, Hermosa Publishers, Albuquerque. 关29兴 Shi, S., Celik, I., and Smirnov, A., 2006, “LES of a Spatially Developing
关4兴 Stern, F., Wilson, R. V., Coleman, H. W., and Paterson, E. G., 2001, “Com- Turbulent Wake Flows,” J. Ship Res., 共to appear兲
prehensive Approach to Verification and Validation of CFD Simulations-Part 1: 关30兴 Yavuz, I., Cehreli, Z. N., and Celik, I. B., 2002, “Large Eddy Simulation of the
Methodology and Procedures,” ASME J. Fluids Eng., 123, pp. 793–802. Wake Behind a Turning Ship,” 2002 Fluids Engineering Division Summer
关5兴 Roache, P. J., Ghia, K. N., and White, F. M., 1986, “Editorial Policy Statement Meeting, 14–18 July 2002, Montreal, Quebec, Canada.
on Control of Numerical Accuracy,” ASME J. Fluids Eng., 108, Mach Issue. 关31兴 Cehreli, Z. N., 2004, “Investigation of Ship Wakes Using LES With Various
关6兴 Freitas, C. J., 1993, “Journal of Fluids Engineering Editorial Policy Statement SGS Models,” Ph.D. dissertation, West Virginia University, Morgantown, WV.
on the Control of Numerical Accuracy,” ASME J. Fluids Eng., 115, pp. 339– 关32兴 Badeau, A. E., Jr., and Celik, I. B., 2004, “Simulation of Vertical Buoyant
340. Flow Events Using ITM,” J. Turbul. 共submitted兲.
关7兴 Freitas, C. J., Ghia, U., Celik, I., Roache, P., and Raad, P., 2003, “ASME’s 关33兴 Rightley, P., 1995, “Bubble Dispersion and Interphase Coupling in a Free
Quest to Quantify Numerical Uncertainty,” AIAA, 41st Aerospace Sciences Shear Flow,” Ph.D. dissertation, University of California, San Diego, CA.
Meeting& Exhibit, 6–9 January 2003, Reno, NV. 关34兴 Celik, I. B., Cehreli, Z., and Yavuz, I., 2003, “Index Of Quality For Large
关8兴 Speziale, C. G., 1998, “Turbulence Modeling for Time-Dependent RANS and Eddy Simulations,” 2003 ASME Fluids Engineering Division Summer Meet-
VLES: A Review,” AIAA J., 36. ing, 6–10 July 2003, Honolulu, Hawaii.
关9兴 Vreman, B., Geurts, B., and Kuerten, H., 1996, “Comparison of Numerical 关35兴 Hussain, A., and Reynolds, W., 1975, “Measurements in Fully Developed
Schemes in Large-Eddy Simulation of the Temporal Mixing Layer,” Int. J. Turbulent Channel Flow,” ASME J. Fluids Eng., 97, pp. 568.
Numer. Methods Fluids, 22, pp. 297–311. 关36兴 Smirnov, A., Shi, S., and Celik, I., 2000, “Random Flow Simulations with a
关10兴 Vreman, B., Geurts, B., and Kuerten, H., 1997, “Large-Eddy Simulation of the Bubble Dynamics Model,” ASME Fluids Engineering Division Summer Meet-
Turbulent Mixing Layer,” J. Fluid Mech., 339, pp. 357–390. ing, No. 11215, FEDSM2000, Boston, MA.
关11兴 Pope, S. B., 2000, Turbulent Flows, Cambridge University Press, Cambridge. 关37兴 Shi, S., and Celik, I., 2000, “Random Flow Generation Technique in Large
关12兴 Meneveau, C., 1994, “Statistics of Turbulence Sub-Grid-Scale Stresses: Nec- Eddy Simulations and Particle Dynamics Modeling,” ASME J. Fluids Eng.,
essary Conditions and Experimental Tests,” Phys. Fluids, 6, pp. 815. 123, pp. 359–371.
关13兴 Rodi, W., Ferziger, J. H., Breuer, M., and Pourquie, M., 1997, “Status of Large 关38兴 Longo, J., Stern, F., and Toda, Y., 1993, “Mean-Flow Measurements in the
Eddy Simulation: Results of a Workshop,” ASME J. Fluids Eng., 119, pp. Boundary Layer and Wake and Wave Field of a Series 60 CB= 0.6 Ship
241–254. Model-Part 2: Scale Effects on Near-Field Wave Patterns and Comparisons
关14兴 Froehlich, J., and Rodi, W., 2002, “Introduction of LES of Turbulent Flows,” With Inviscid Theory,” J. Ship Res., 37, pp. 16–24.
in Closure Strategies for Turbulent and Transitional Flows, B. Launder and N. 关39兴 Gui, L., Longo, L., and Stern, F., 1999, “Towing Tank PIV Measurements
Sandham, eds., Cambridge University Press, Cambridge, pp. 267–298. System and Data and Uncertainty Assessment for DTMB Model 5512,” 3rd
关15兴 Choi, H., and Moin, P., 1994, “Effects of Computational Time Step on Nu- International Workshop on PIV, Santa Barbara, CA, 16–18 September 1999.
merical Solution of Turbulent Flows,” J. Comput. Phys., 113, pp. 1–3. 关40兴 Anwar, H., 1969, “Experiment on an Effluent Discharging From a Slot Into
关16兴 Richardson, L. F., 1910, “The Approximate Arithmetical Solution by Finite Stationary or Slow Moving Fluid of Greater Density,” J. Hydraul. Res., 7, pp.
Differences of Physical Problems Involving Differential Equations, with an 411–430.

958 / Vol. 127, SEPTEMBER 2005 Transactions of the ASME

Downloaded From: http://fluidsengineering.asmedigitalcollection.asme.org/ on 09/01/2013 Terms of Use: http://asme.org/terms

You might also like