1988 Wilcox PDF
1988 Wilcox PDF
1988 Wilcox PDF
A comprehensive and critical review of closure approximations for two-equation turbulence models has been
made. Particular attention has focused on the scale-determining equation in an attempt to find the optimum
choice of dependent variable and closure approximations. Using a combination of singular perturbation
methods and numerical computations, this paper demonstrates that: 1) conventional A:-e and A>w2 formulations
generally are inaccurate for boundary layers in adverse pressure gradient; 2) using "wall functions'' tends to
mask the shortcomings of such models; and 3) a more suitable choice of dependent variables exists that is much
Downloaded by IMPERIAL COLLEGE LONDON on March 10, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.10041
more accurate for adverse pressure gradient. Based on the analysis, a two-equation turbulence model is
postulated that is shown to be quite accurate for attached boundary layers in adverse pressure gradient,
compressible boundary layers, and free shear flows. With no viscous damping of the model's closure coefficients
and without the aid of wall functions, the model equations can be integrated through the viscous sublayer.
Surface boundary conditions are presented that permit accurate predictions for flow over rough surfaces and for
flows with surface mass addition.
includes results of attached boundary-layer computations for where, by definition, the mean strain-rate tensor Su is
boundary layers subjected to adverse pressure gradient,
surface mass injection and compressibility, and free shear u^ dUj
(8)
flows. Xj dX{
II. Equations of Motion We invoke the Boussinesq approximation that the Reynolds
This section states the postulated equations of motion, stress tensor is proportional to the mean strain-rate tensor,
including established values of all closure coefficients. Physi- that is,
cal interpretations of turbulence field properties are given and,
additionally, arguments are presented that have been used in (9)
setting values of several of the closure coefficients.
A. Postulated Equations Finally, the heat flux vector QJ is approximated as
For general compressible turbulent fluid flows, the turbu-
lence model equations are written in terms of Favre13 mass- Qj =
( V ^T\ dh (10)
~~ ~ \PrL+ PrT) dXj
averaged quantities as follows.
where PrL and PrT are the laminar and turbulent Prandtl
Mass conservation:
numbers, respectively.
Downloaded by IMPERIAL COLLEGE LONDON on March 10, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.10041
B. Establishing Closure Coefficient Values In summary, the arguments presented in this subsection are
In this subsection, we present straightforward arguments sufficient to uniquely set the values of 7*, ft*, and ft. Also,
from which values of the four closure coefficients /3, 0*, 7, Eq. (17) determines 7 in-terms' of the as-yet undetermined
and 7* can be established. A review of the arguments generally value of a. As a by-product of analysis in the next section, the
presented by turbulence model researchers indicates that the values of a and a* will be established.
following are as physically sound as possible within the
context of two-equation turbulence models. III. Defect-Layer Analysis
Considering first the coefficient 7*, we rewrite Eqs.(1-10) In this section, we use singular perturbation methods to
in terms of the quantity o>/7*. Inspection of the resulting analyze model-predicted structure of the classical defect layer.
equations shows that this rescaling of u is equivalent to setting The analysis presented is a generalization of that done by
7* = 1. Hence, with no loss of generality, we conclude that the Wilcox and Traci.9 In contrast to the Wilcox and Traci
value of 7* is indeed unity. analysis, effects of pressure gradient have been included.
Next, we turn to the ratio of ft to ft*. For decaying Additionally, the analysis has been done for three turbulence
homogeneous, isotropic turbulence, Eqs. (4) and (5) simplify models: the model postulated in Eqs. (1-10), the Wilcox-
to Rubesin10 model, and the Jones-Launder5 model. First, we
review details of the perturbation solution procedure. Next,
ct) we compare solutions for the three models in the absence of
and — = -ftu2 (13) pressure' gradient. Then, effects of pressure gradient are
at
studied for the three models. Finally, we justify the values
Downloaded by IMPERIAL COLLEGE LONDON on March 10, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.10041
from which the asymptotic solution for k is readily found to chosen for a and a*.
be
A. Perturbation Solution
k~t-0*'* (14) In the past, the only detailed analyses of the defect layer for
6/5 any turbulence model have been those of Bush and Fendell16
Experimental observations indicate that k ~ t ~ for decay- (for the mixing-lenth model) and Wilcox and Traci (for a k-u>2
ing, homogeneous, isotropic turbulence, which implies that model). In neither case were effects of pressure gradient
delineated. In this section, we extend the Wilcox-Traci
Values for the coefficients 7 and ft* can be established by analysis to include presssure gradient.
examining the so-called "wall layer." The wall layer is defined To study the defect layer, we confine our analysis to
as the portion of the boundary layer far enough from the incompressible flow and we seek a perturbation solution. The
surface to render molecular viscosity negligible relative to expansion proceeds in terms of the ratio of friction velocity to
eddy viscosity, yet close enough for convective effects to be the boundary-layer-edge velocity uT/Ue and the dimensionless
negligible relative to the rate at which the turbulence is being vertical coordinate 77, defined by
created and destroyed. In the limiting case of an incompress-
ible constant-pressure boundary layer, defining VT - IJLT/P,
Eqs. (1-10) simplify to (18)
mental data presented are those at the highest Reynolds strengths closest to values inferred from data over the
number for which data are reported. This is consistent with complete range considered. Consistent with predictions of
the defect-layer solution that is formally valid for very large Chambers and Wilcox,17 the Jones-Launder model exhibits
Reynolds number. Numerical results are shown for three the largest differences, with predicted wake strength 50-100%
models: the new (Ar-co) model, the Wilcox-Rubesin (k-u2) lower than inferred values when /3r is as small as 2!
model, and the Jones-Launder (k-e) model. Figure Ib compares computed velocity profiles with experi-
As shown, all three models predict velocity profiles that mental data of Clauser2 for PT = 8.7. As with the constant-
differ from measured values by no more than about 3% of pressure case, computed and measured skin friction are
scale. Interestingly, the new model shows the smallest included in the insert. Consistent with the wake-strength
differences from the Wieghardt data. Additionally, skin predictions, the new model yields a velocity profile and skin
friction C/ can be inferred from the defect-layer solution (see friction closest to measurements, whereas the Jones-Launder
AppeMix). Corresponding computed and measured vahies are model shows the greatest differences . The Wilcox-Rubesin
summarized in the insert on Fig. la; the largest difference is profile and skin friction lie about midway between those of the
less than 3%. Thus, based on analysis of the constant-pressure other two models.
defect layer, there is little difference among the three models. The explanation of the Jones-Launder model's poor perfor-
mance for adverse pressure gradient can be developed from
C. Effects of Pressure Gradient inspection of the asymptotic behavior of solutions as r/^0.
Turning now to the effect of pressure gradient, we have For the three models tested, the velocity behaves as
computed defect-layer solutions for the equilibrium parameter
Downloaded by IMPERIAL COLLEGE LONDON on March 10, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.10041
0r, ranging from - 0.5 to +9.0, where positive &T corre- (Ue - u)/uT~ - +A - as iy— 0 (23)
sponds to an adverse pressure gradient. The choice of this
range of /3T has been dictated by the requirement of the where Table 1 summarizes the constants A and C. Note that
perturbation somtion that j8r be constant. This is as wide a while the coefficient A is determined as part of the solution
range as we have been ables'tp find for which experimental data (from the integral constraint that mass be conserved), the
have been taken with 0r more or less constant. coefficient C follows directly from the limiting form of the
Figure Ic compares computed wake strength f with values solution as 77 ->0. As seen from Table 1, C is largest for the
inferred by Coles and Hirst2 from experimental data. For the Jones-Launder model and smallest for the new model. The
sake of clarity, note that the wake strength appears in Coles' presence of the rj&iTj term gives rise to the inflection in the
composite law-of-the-wall/wake profile, velocity profile as ??-*0 that is most pronounced for the
Jones-Launder model. In terms of turbulence properties, the
turbulence length scale, £= &1/2/w, behaves according to
(22)
'
as r/- (24)
Inspection of. Fig. Ic reveals provocative differences among
the three models. Most notably, the new model yields wake Table 1 also includes the coefficient L for each model. Again,
we see that the' 'y/ka\^ term is largest for the Jones-Launder
model and smallest for the new model. Thus, in the presence
of adverse pressure gradient, the Jones-Launder turbulence
length scale tends to be too large in the near-wall region. Note,
of course, that this shortcoming is not evident in the
constant-pressure case, which has .$T = 0.
The manner in which the new model achieves smaller values
of £ than the Jones-Launder model can be seen by changing
dependent variables. That is, starting with the k-co formula-
tion, defining e = (3*uk and vT = k/u, we can deduce the
following incompressible equation for e implied by the new
model:
de
de_ '.- dk
a) Velocity profiles, B r = b) Velocity profiles, B r = 8.?• vT- (25)
**y
All terms except the last one on the right-hand side of Eq. (25)
are identical in form to those of the Jones-Launder model (see
• COLES AND HIRST Appendix). The last term is negligibly small as r?^0 for
constant-pressure boundary layers because k-» constant as
r?—0. However, dk/dy is nonvanishing when (3T7±Q and
d(€/k)/dy generally is quite large as 77-»0. The net effect of this
additional term is to suppress the rate of increase of I close to
the surface.
(26)
[du d^
(29) As with the defect layer, we solve the sublayer equations by
dy 1) adding unsteady terms to the left-hand sides of Eqs. (28)
Five boundary conditions are needed for this fifth-order and (29), 2) making an initial guess at the solution, and
system, two of which follow from matching to the law of the 3) using an implicit, time-marching, second-order-accurate
program to generate the long-time solution in which the
wall as y + -*<»:. unsteady terms tend to zero. The velocity is computed at each
+
time step using the fourth-order Runge-Kutta method. The
arid co —> - as y -~ c (30) program used is an improved version of that developed in the
Vis* study by Wilcox and Traci.9
where y+ =uTy/v. Two more boundary conditions follow Using this program, we find that Eqs. (27-32) predict the
from "no slip" at the surface, which implies that u and k smooth-wall value of B as
vanish at y = 0. Thus,
,6 ==. lirn I "u + - - = 5.1 (34)
y + —oo j K
u =k = (31)
1304 D. C. WILCOX AIAA JOURNAL
That this value is well within the scatter of measured values of Figure 3 shows the computed value of B for a wide range of
B strongly suggests that no further viscous modifications are values of SR. As shown, in the limit SR-~oo, B tends to 5.1. In
needed. the limit S/?—0, an excellent correlation of the numerical
Figure 2a compares computed and measured2'18'19 sublayer predictions is given by
velocity profiles. As shown, computed velocities generally fall
within experimental data scatter. In Fig. 2b, we compare
- - as (37)
computed and measured18 turbulence production and dissipa- 100
tion terms. Again, predictions fall well within experimental
error bounds. By experimental means, as summarized by Schlichting,21
Perhaps the only deficiency of predicted smooth-surface Nikuradse found that for flow over very rough surfaces,
sublayer structure is that, very near the surface, the model
predicts (38)
as y^ (35)
where kR is the average height of sand-grain roughness ele-
By contrast, the Wilcox-Rubesin model predicts that k ~y4, ments. (Note that, in our computations, we use K = 0.41,
which suggests that k ~ (v'2}\ a point this researcher has whereas Nikuradse used /c = 0.40.) Thus, if we make the corre-
made9 as a more plausible interpretation than saying that k is lation
the kinetic energy of the turbulence. By letting the closure 100
Downloaded by IMPERIAL COLLEGE LONDON on March 10, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.10041
Fig. 3 Computed variation of the constant in the law of the wall, B-9 Fig. 4 Comparison of computed and measured sublayer velocity
with the surface value of the specific dissipation rate. profiles for "completely rough'* surfaces.
NOVEMBER 1988 EQUATION FOR ADVANCED TURBULENCE MODELS 1305
u =k = 0 at y=0 (47)
where H is given by
For compressible flows, we specify either surface temperature
(44) Tw or surface heat flux qw, so that the enthalpy boundary
condition for a calorically perfect gas becomes
The variation of £ predicted in Eqs. (43) and (44) is consistent
with the data of Andersen e t a l . Including appropriate T ._
or at y = 0 (48)
convective terms in Eqs. (27-29), we have performed sublayer
computations for the cases experimentally documented by
Andersen et al. In each case, the surface value of co has been where Cp is specific heat at constant pressure and subscript w
given by denotes surface (y = 0) value. For the compressible flow
at y=0 (45)
.005
and the value of SB has been varied until optimum agreement
between measured and computed velocity profiles is achieved.
The final correlation between SB and v + is given in analytical
form as
20
(46)
SYMBOL . v+
O o . '. -
• .0303
• .0.683 ANDERSEN
D .1819
A -7393
————— COMPUTED
0 0 . 5 1.0 1.5 2 . 0 2 . 5
x(m)
computations of Sec, VB, we assume that our fluid is air so compressible constant-pressure (flat-plate) boundary layer.
that PrL = 0.72, and the value used for PrT is 8/9. Finally, for First, we apply the model to the incompressible flat-plate case.
perfectly smooth surfaces with no surface mass injection, we The next two applications are to boundary layers in an adverse
require that pressure gradient. The final incompressible application is for
a boundary layer with surface mass injection. We conclude
v=Q and o>^ —-^ as y—0 (no mass injection) (49) with application of the model to the compressible flat-plate
• - ;
' ' boundary layer with and without heat transfer for Mach
numbers 0-5. Numerical details of the computations are given
For the computation with surface mass injection (Sec. VB), by Wilcox.26
the surface conditions satisfied by tr and ox are given by
Incompressible Flat-Plate Boundary Layer
v :•='. v# and o> = — SB at y = 0 (with mass injection) (50) The first application is for the constant-pressure incom-
pressible boundary layer. Although this application does not
provide a severe test of the new model, it is nevertheless
where the dimensionless coefficient SB is defined in terms of necessary to be sure the boundary-layer program has been
the blowing rate parameter v+ = vw/ur by Eq. (46). coded properly. Also, the new model would be of little use as
The manner in which the surface boundary condition on QJ a predictive tool if it were inaccurate for the simplest of all
is implemented when mass addition is present is straightfor- boundary layers.
ward. For the perfectly smooth surface, the method used is The computation begins at a plate-length Reynolds number
less obvious and requires special care. As indicated in Eq. (49),
Downloaded by IMPERIAL COLLEGE LONDON on March 10, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.10041
friction differ by less than 4% of scale whereas computed and VI. Summary and Conclusions
measured velocity profiles are within 3% of each other.
Although this flow has zero pressure gradient, corresponding The primary objectives of this study have been accom-
skin friction predicted11 by the k-e model is as much as 50% plished: we have made a critical review of closure approxima-
higher than measured, probably because the wall functions tions used in two-equation turbulence models and determined
used by Rodi in this computation are inadequate for flows what appears to be an optimum choice of dependent variables.
with mass injection. As a result, we have developed a new two-equation model that
promises more accuracy for boundary layers in an adverse
pressure gradient than any similar model.
Compressible Flat-Plate Boundary Layers As in our prior turbulence modeling efforts, we have made
In order to test the k-u model for effects of compressibility, extensive use of perturbation methods (Sees. Ill and IV). In
we consider flows 8101 and 8201 of the second Stanford contrast to prior studies, our analysis of the defect layer
conference.11 We first consider the adiabatic-wall case with includes pressure gradient. As discussed in Sec. Ill, limiting
freestream Mach number 0-5. For each Mach number, com- the defect-layer analysis to the constant-pressure case displays
putation begins at anRex of 1.0-106 and continues to the point little difference among the various two-equation models in
where momentum-thickness Reynolds number Re9 is 10,000. general usage. However, as soon as an adverse pressure
Figure 7a compares the computed ratio of skin friction C/to gradient is included, the models exhibit large differences. As a
the incompressible value Cf09 with the correlation developed general observation, the second Stanford conference demon-
by Rubesin. As shown, differences between computed ratios strated that modern turbulence models are not much more
accurate than those in use in 1968 if the flow of concern is a
Downloaded by IMPERIAL COLLEGE LONDON on March 10, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.10041
V C fo V C fo
1.0 1.0
0 RUBESIN O RUBESIN
—— COMPUTED —— COMPUTED
0 J______t 0
0 1 2 3 4 5 0 .2 .4 .6 .8 1.0
a) Mach number ^oo b) Surface cooling ^^aw
Fig. 7 Comparison of computed and measured effects of freestream Mach number and surface cooling on flat-plate boundary-layer skin friction.
1308 D. C. WILCOX AIAA JOURNAL
successful. Now that this step has been accomplished, In the Jones-Launder model, we compute dissipation e
complicating effects such as anisotropic shear, streamline directly, so that additional equations are
curvature, unsteadiness, and separation can be addressed by
building on the solid foundation offered by the k-u model (A5c)
introduced in this paper.
(A9)
In analyzing the defect layer, we focus on three turbulence
models: the new model postulated in Sec. II, the Wilcox-
Rubesin10 model, and the Jones-Launder5 model. For all three (A10)
of the models, we must solve the equations for mean mass and *A
momentum conservation, an equation for turbulent energy, 3
and an equation for a turbulence dissipation scale. For all
(All)
three models, the first three equations assume the following
form:
where
du dv
(Al)
:Z A = -^ (A12)
du dUe a
(A2) Inserting Eqs. (A8-A12) into Eqs. (A1-A7), neglecting
higher-order terms, letting AToO?) denote dimensionless eddy
viscosity, and defining
dk du d_
(A3)
'dy
(A13)
where d/dt = ud/dx + vd/dy. Note that Eqs. (A 1-A3) do not
include molecular viscosity. This is a valid approximation in
the defect layer as the eddy viscosity is proportional to £46*, we obtain the following equations:
where Ue is the boundary-layer-edge velocity and 6*
is displacement thickness. Hence, the ratio of molecular to at/!
eddy viscosity varies inversely with displacement-thickness
Reynolds number and is thus very small. The difference
among the three models is in the way the dissipation e and the = 2oTx -^
i
(A1.4)
kinematic eddy viscosity VT. are computed. ox
For the new model, in addition to Eqs. (A1-A3), we have
€ = /3*a>/: (A4a)
(7/VQ I _n
—— + (<xT ~ 2pr - 2wr)
_ a^0
(A5a) (A15)
(A6a) where
Wilcox-Rubesin model:
(A17b)
dx
Jones-Launder model:
dE =0
« + .- 20r - 2o,T)n ^
dr,
(A23b)
Jones-Launder model:
Downloaded by IMPERIAL COLLEGE LONDON on March 10, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.10041
(A17c)
(A23c)
where the parameters a y , j37, a r , and w r are defined in terms
of 8\ ur, and skin friction, Cf = 2(uT/Uef, i.e.,
Boundary Conditions
At the outer edge of the defect layer, we require that the
velocity equal the freestream velocity. Additionally, we let the
5* dp turbulent energy assume a small value and insist that the
turbulent length scale have zero slope at the boundary-layer
edge. In their defect-layer analysis, Wilcox and Traci used
these boundary conditions, as well as explicitly prescribing
both KQ and WQ. Thus,
as (A19)
as (A20)
Thus, self-similar solutions exist, provided only that 0T is The coefficients «i, k\, Wi, and e\ are as follows:
independent of x .
In summary, the defect-layer equations for the leading-or- All models:
der terms in the perturbation expansions become
(A2I)
New model:
(A27a)
(A22) (A28a)
1310 D. C. WILCOX AIAA JOURNAL
2
Wilcox-Rubesin model: Coles, D. E. and Hirst, E. A., Computation of Turbulent
Boundary Layers—1968 AFOSR-IFP-Stanford Conference, Vol. II,
. Stanford Univ. Press, Stanford, CA, 1969.
(A27b) 3
Bradshaw, P., Ferriss, D, H., and Atwell, N. P., "Calculation of
Boundary Layer Development Using the Turbulent Energy Equa-
tion," Journal of Fluid Mechanics, Vol. 28, Pt. 3, 1967, pp. 593-616.
4
(A28b) Donaldson, C. duP., "Calculation of Turbulent Shear Flows
4- for Atmospheric and Vortex Motions," AIAA Journal, Vol. 10,
Jan. 1972, pp. 4-12.
5
Jones-Launder model: Jones, W. P. and Launder, B. E., V'The Prediction of Laminariza-
tion with a Two-Equation Model of Turbulence," International
Journal of Heat and Mass Transfer, Vol. 15, 1972, pp. 301-314.
(A27c) 6
Rodi, W., "Progress in Turbulence Modeling for Incompressible
Flows/'AIAA Paper 81-0045, Jan. 1981.
7
(1 + a* Wilcox, D. C. and Alber, I. E., "A Turbulence Model for High
(A28c) Speed Flows," Proceedings of the 1972 Heat Transfer and Fluid
- C2) Mechanics Institute, Stanford Univ. Press, Stanford, CA, 1972,
pp. 231-252.
Additionally, the coefficient w0 is determined from the integral 8
Saffman, P.O. and Wilcox, D.C., "Turbulence-Model Predic-
constraint for mass conservation, that is,
Downloaded by IMPERIAL COLLEGE LONDON on March 10, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.10041