Sei sulla pagina 1di 7

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/222132491

On the Sensitization of Turbulence Models to


Rotation and Curvature
Article in Aerospace Science and Technology July 1997
DOI: 10.1016/S1270-9638(97)90051-1

CITATIONS

READS

171

347

2 authors:
Philippe Spalart

Michael L. Shur

The Boeing Company

St. Petersburg State Polytechnical University

204 PUBLICATIONS 13,076 CITATIONS

88 PUBLICATIONS 2,499 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate,


letting you access and read them immediately.

SEE PROFILE

Available from: Michael L. Shur


Retrieved on: 01 August 2016

Aerospace Science and Technology,

1997,

no 5, 297-302

On the Sensitization of Turbulence Models


to Rotation and Curvature
P. R. Spalart cl), M. Shur c2)
(I) Boeing Commercial Airplane
P.O. Box 3707, Seattle, WA 98124-2207,

Group
U.S.A.

(1 Federal Scientijk Center Applied Chemistry


St. Petersburg 197198, Russia.
Manuscript

Spalart P. R., Shur M., Aerospace


Abstract

Scierzce and Technology,

Turbulence

1997, no 4, 297-302.

models - Rotation - Curvature - Shear flow - Boundary

OF SYMBOLS

c.f

skin-friction
coefficient
R
radius of curvature
of the sreamlines
Szj = (X$/axj
+ aUj/axi)/2
strain tensor
streamwise
and
cross-stream
mean velocity
u, v

components
streamwise and cross-stream fluctuating
velocity components
azimuthal velocity
oe
y-derivative of U
UY
streamwise and cross-stream coordinates
5; Y
6
boundary-layer thickness (99.5% velocity)
~ijk = (i - j) (j - k)(k - i)/2 alternating tensor
r = 2 7rrUo circulation in azimuthal flow
spanwise vorticity
WZ
u; v

Aerospace

August 29, 1995; accepted June 14, 1996.

Empirical alterations of eddy-viscosity turbulence models to account for system rotation and streamline
curvature are discussed. Except in a narrow class of flows, the streamline curvature itself is an inadequate
entry into a model, because it is not Galilean-invariant.
We propose a measure of the extra influence on
the turbulence which is invariant, fully defined in three dimensions, and unifies rotation and curvature
effects. This is at the expense of involving higher derivatives than the traditional
(non-invariant)
terms
do. It is closely related to an idea of Knight & Saffman [l]. Its interpretation is straightforward,
at least
in two dimensions, and coincides with the classical measures in a few building-block
flows such as
rotating shear flow or a curved boundary layer. It can be used in one-equation, two-equation, and similar
models in conjunction with empirical functions which depend on the model; an example is given.
Keywords:

LIST

received

Science and Tecltnology,

0034.1223,

97/05/$

7.00/O

Gauthier-Villars

layer.

R
system rotation rate
average value of U~
?P
wi Ai = CiE1,2 wi Xi summation convention

I - INTRODUCTION
The explicit appearance of rotation and curvature
terms in the turbulence equations is cited as
a fundamental
advantage of full-Reynolds-stress
turbulence models over the simpler eddy-viscosity
models. An effective alteration to the simple models
would nevertheless fill a widespread need [2], because
for a large class of computational
activities Reynoldsstress models are not yet affordable, or have not
been made to converge reliably and demonstrate
their superior accuracy [3].
In addition,
exact
influences appear in the Reynolds-stress models only

298

as production-like terms. There is little consensus


about more arcane influences, for instance through
the pressure terms, and we leave these issues to the
experts.
Our knowledge of rotation and curvature effects is
fair for two extreme types of flows. In thin shear
flows with weak rotation (compared with the shear
rate) or weak curvature (compared with the inverse
of the shear-layer thickness), we observe a sizeable,
even surprising, effect on the level of the turbulent
shear stress [4]. The other extreme is homogeneous
rotating shear flow; there, in particular, strong rotation
extinguishes the turbulent shear stress [Sj. Another
case of strong rotation is the core of a free vortex,
a flow which has not been as well explored but
also shows dramatic effects [4], and poses serious
difficulties for unprepared turbulence models and even
scaling theories [6]. The measure we propose below is
based on intuitive arguments developed in thin shear
flows; it is then shown to be useable under strong
rotation, although not without further empirical input.
For convenience we often refer to a steady thin
shear flow with velocity primarily in the 17:direction,
and all quantities varying primarily in the y direction.
Let U (y) be the velocity profile and assume that
U, > 0, so that w, < 0. In the eddy-viscosity
context, we naturally focus on the Reynolds shear
stress --U/U which is positive. Rotation at a rate
R contributes
a term 2 R (u~ - v~) to its budget.
-Because U~ > v~, the shear stress is enhanced when
R is positive, and vice versa. Thus the production-like
term arises from a fairly subtle feature of the Reynoldsstress tensor, which is not rendered by the common
eddy-viscosity models (they are just as likely to return
vf2 > u12). With curvature, a similar term appears
if the transport equation is written with respect to
curvilinear coordinates aligned with the flow direction,
leading to an effective rotation rate equal to U/R
(here R is taken as positive if the streamlines are
concave). Now U/R can also be written as V,. In
our view a difficulty in past discussions, which label
V, as an extra rate of strain [4], has been to explain
why values of V, much smaller than that of U, have
a significant effect, when the two are equal partners
in the vorticity w,.
We further note that V,, although a velocity
derivative, is not a Galilean invariant, because it is
referred to axes aligned with the velocity. Therefore
its use outside shear flows that are either steady or
constrained to follow a fixed part (such as boundary
layers) can hardly be defended.
Measures based
on the transverse pressure gradient [7] fall in the
same category, the more so because the spatiallyuniform acceleration of an incompressible flow alters
the pressure gradient, and not the turbulence. Although
this fact is rigorous only for solid-wall boundary
condition, we observe by analogy that a body of
turbulent fluid in an unsteady Ilow will experience
accelerations, and that the attendant pressure gradients

P. R. Spalart,

M. Shur

must not be mistaken as curvature by a turbulence


model.
We are not implying that the community has
made widespread use of the streamline curvature in
obviously inappropriate situations. We have however
identified a few marginal uses, and the need for
a more general measure is clear. We recognize
a risk of elevating invariance requirements to a
level inconsistent with the drastic approximations
inherent in any closure of turbulence, and we are not
contending that a non-invariant model cannot be of any
value. However we feel that any addition proposed
for a model should be as resistant to coordinate
transformations as the existing model is. We are
targeting a class of models, typically with one or
two transport equations, which are Galilean-invariant
and are now routinely used in three-dimensional (3D)
or time-dependent flows. A high enough level of
invariance is needed for the calibration secured in
simple flows to, possibly, carry over to more complex
flows.
A Galilean-invariant measure can be derived by
comparing the vorticity magnitude w E 1w, 1 and the
strain rate s E dm
in two-dimensional flows.
Having noted that with unstable rotation or curvature
s exceeds w, and vice versa, we could design an
empirical correction based on s/w, for instance. It
would unify weak rotation and weak curvature in the
traditional way, that is, R and U/R play the same
role [4]. However the difference s - w is slight and a
large coefficient would be needed [4]. In addition we
anticipate spurious effects when the difference results
from other features of the flow, especially in 3D. Thus
we regard the trend of s/w as too restricted and seek
a more trustworthy, albeit empirical, connection. It
turns out that s/w is used in conjunction with the
new measure in our example (3 IV), but this is not
necessary for other models.
In 1971 Knight & Saffman [l] presented a term
called the gyroscopic stability effect, an attempt
to model the effects of streamline curvature. The
term also addressed rotation and was successfully
applied to 2D curved flows, in the context of the
Saffman e-w model. However, it was overlooked
by later two-equation modelers [2], and apparently
not promoted even by its authors. In $ III it will
become clear that we essentially re-invented that
term, but we deem that our formulation is easier to
apply especially in 3D. In any case the idea deserves
renewed attention because of its soundness, because
the original discussion was very brief, and because of
the current demand in Computational Fluid Dynamics.
II - TWO-DIMENSIONAL
INCOMPRESSIBLE
CASE
The role played by the inequality u12 > vj2 in
a thin U (y) shear flow suggests the following
Aerospace

Science and Technology

On the Sensitization

of Turbulence

Models to Rotation

approach. This inequality is equivalent to stating that


the principal axes of the Reynolds-stress tensor are
not aligned with those of the strain tensor (which
are at 45 to the (5, y) axes), but rotated counterclockwise.
Thus the stress axes lead or lag the
strain axes, as the system- or flow-aligned axes
rotate. depending on the sign of 0. If they are
leading, the rotation brings the strain axes closer
to the (more slowly responding) stress axes which,
plausibly, benefits turbulence production. We make
the following hypothesis: under weak rotation or
curvature, turbulence is enhanced if the Reynoldsstress principal axes are leading the strain axes, and
vice versa. In a one-equation turbulence model,
turbulence enhancement is translated by eddyviscosity increase and vice-versa.
Two-equation
models offer additional choices, such as whether to
act on the energy variable, on the dissipation variable,
or on the eddy-viscosity directly (for instance, the cP
quantity of the k-E model).
This hypothesis unifies rotation and curvature effets,
giving 0 and U/R the same role. Although this
unification is by no means rigorous, it is consistent
with observation [4], it simply settles an auxiliary
question in the approach, and it allows for a crossvalidation between rotation and curvature cases. In
similar work on the k-w model, Wilcox & Chambers
[3] offered a nonempirical,
although physically
based, argument that gives the same role to -R and
to U/(2 R). We differ with their emphasis on @, as
we contend that U u has more leverage. In addition
their own analysis, reflected by the values of the p
constants, shows that the use of R/2 in place of R
would vastly improve the agreement with experiments.
The system rotation R is well defined and
understood [5]. It remains to dispose of U/R, or
equivalently V, , because of its lack of Galilean
invariance. This is the central difficulty. Instead of the
streamline direction we propose to track the direction
of the principal axes of the strain tensor, as did Knight
& Saffman [l]. In a weakly curved thin shear flow,
the flow direction, the direction of the strain principal
axes, and that of the Reynolds-stress principal axes
all evolve at the same rate, U/R. The strain axes are
invariant and useable in a simple turbulence model.
This leads us to the fundamental quantity
DCX
Dt

(1)

where the angle Q gives the direction of the straintensor principal axes with respect to an inertial
reference frame. Being the Lagrangian derivative of
a quantity which is defined with respect to an inertial
frame, Da/Dt
is Galilean-invariant.
Which of the
two principal axes we consider is unimportant, since
they are orthogonal. Situations in which the axes are
undefined, because the tensor is isotropic, will require
some care.
In a homogeneous rotating flow with timeindependent deformation,
Da/Dt
= R. In an
1997. no 5

299

and Curvature

inhomogeneous flow its calculation is more involved,


but we can avoid the troublesome step of computing
and numerically differentiating the principal directions,
thanks to the following identity. This question was not
addressed by Knight & Saffman [l], who treated only
flows in which U/R was adequate. If we consider
an incompressible flow, calculate the strain-tensor
eigensystem,
differentiate its direction analytically,
and simplify, we arrive at
1
D&2
D&I
S -s12 ~
2 6%

S,22)

l1

Dt

Dt

Here the strain tensor S;j and the Lagrangian


derivative in the right-hand-side of (2) are defined
with respect to the reference frame of the calculation,
which is rotating at a rate 0. Szl and 522 do
not appear because we have used the symmetry and
trace-free properties of the tensor. In a compressible
flow, we take the deviator of the tensor (i.e., we
remove the velocity divergence) to recover a tracefree tensor before applying (2). Taking the deviator
does not affect the principal directions. As expected
this formula diverges if the deviator is identically zero,
or equivalently if the tensor is isotropic.
The sign of Da/Dt
is relevant only as compared
with that of the vorticity, because the stress-strain
misalignment is in the direction of the vorticity.
Therefore the quantity e z (Da/Dt)
sign (wZ + 2 fi)
is a prime candidate, see 8 IV. We use the nondimensional combination rX E e/w,. Small positive
values of ? suppress turbulence activity; small negative
values enhance it; values of order fl suppress it. Such
values are outside the range of our hypothesis above;
however, a useful model can result, in particular if the
function of ? is even for large 1r I.
Depending on the model, e could be used in raw
form, but we think nonlinear functions of rXare more
likely. For instance in a typical boundary layer the
ratio of Knight & Saffmans curvature contribution
(x term in their eq. 2) to the production (cv term)
is roughly O.EXJ+[y/R], where U+ is the velocity in
wall units, y is the distance to the wall, and R the
radius of curvature. Since U+ takes values of the
order of 20 in the outer region of the boundary layer,
it is apparent that the factor multiplying y/R is quite
large, so that for strongly curved flows (with y/R of
the order of 0.1) the linear term as used by Knight &
Saffman is not likely to be valid.
The new measure e reduces to the old ones when
appropriate. In solid-body rotation, e = w/2, so
that 7 = l/2 (stabilizing).
In a pure azimuthal
flow, represented in polar coordinates by .UB(T), we
verify that e = (UQ/T) sign (d [r Ue] /dr). Therefore,
we recover the streamline curvature per se, Ue/r.
In addition e has the sign of the classical quantity
d 1l? I/&, known since Rayleigh for its influence on
the stability of azimuthal flows. On the other hand,

P. R. Spalart, M. Shur

300

this equivalence of e and U/R is not a general


property, even for thin shear flows (see discussion
of Fig. 2). We have failed to find even a fair
approximation of e that can be generally computed
from first velocity derivatives such as Vi. The second
derivatives of velocities (in oSij/ot)
add complexity
when programming the model, as does the time
derivative in unsteady calculations.
Unlike the case of V,, it is not a surprise that
e can significantly alter the turbulence even if it is
much smaller than U,. Our interpretation is that
e modulates the mechanism by which U, sustains
turbulence.
III - THREE-DIMENSIONAL

CASE

Here we face two additional difficulties. The first is


that simple algebraic formulas are not available for the
principal directions, so that obtaining a closed form
such as (2) hardly seems possible. The second is that
the rotation of the principal axes is now represented
by a rotation vector (since the axes are orthogonal,
there exists a solid-body rotation that follows their
direction). Exploiting this vector is not just a matter
of multiplying it by sign (w,), as it was in two
dimensions. Knight & Saffman use (with a change
of notation) the dot product w; X; where Xi is the
angular velocity of the principal axes of the mean
rate-of-strain tensor following a fluid particle. They
do not discuss analytical difficulties in calculating Xi,
and only present results in simple geometries. The
spirit of our approach is the same, but we produce an
explicit formula.
We seek a measure of whether the rotation of the
principal axes (following a particle), and the vorticity,
are in the same general direction; this is similar to the
wi Xi dot product. For this we invoke the dot product
between two tensors. The first is the Lagrangian
derivative DS;j/Dt;
the second is the tensor SSij/St,
defined as the result of a rotation of the strain tensor
Sij itself at a rate given by the vorticity vector. The
two-dimensional equivalent of this procedure is simply
to rewrite sign (w,) as wZ/l w, / in e. We find through
algebra that SSijIGt = (Xli/axk
- XJk/~xi) Sjk +
(dU~/ds~ - aUk/&:j) S;,. The dot product of the
tensors simplifies a little as a result of symmetry,
leading to the triple sum

that the 2D and 3D approaches are consistent.It points


to the following normalization, as well as the addition
of system rotation :

e=

1
as,,s,,

aUi
~8%

8Uk
z

Tj--&

DSij

+2Elki

01

1(
sjk

+ 00[%Jp
sj, + Ejop
Sip]
>
This is our proposal for a scalar measure of rotation
and curvature effects in three dimensions. It has
the dimension of an inverse time. Small positive
values are stabilizing, and vice versa. The rest of the
implementation is model-dependent.
In the S,, S,, factor of the denominator of
(4) it would be consistent to use the deviator of
the strain tensor; this factor is then proportional
to the se aration of the eigenvalues (defined as
(Xi - X2)+ (X1 - X3)2 + (X, - Xs)2, with obvious
notation, but not that of Knight & Saffman).
On the other hand, using the full tensor instead
of the deviator will produce fewer indeterminate
values (it makes S,, S,, more positive) and may
benefit the stability of the calculations. When the
three eigenvalues coincide, in general the numerator
vanishes linearly, and the denominator quadratically,
so that the ratio does take large values. We need
more experience before making a recommendation
regarding the deviator, and of course there is no effect
in incompressible flow. Similarly, e is likely to be
used is a non-dimensional argument r E e/S, with S
a measure of the deformation tensor. For S a choice
will be made between w, S and others, depending on
the model [8]. With a two-equation model, the internal
time scale of the model would also be available, so
that some of the singularities that interfere with the
application to one-equation models may be removed.
Users engaged in checking their programming of
(4), which is a fairly complex expression, can use the
following diagnostics. In a region of pure rotation,
? = l/2. Near a flat surface, I? 1 << 1. In curved
thin shear flows, a separate estimate of U/R can be
compared with the computed values. e and especially
? tend to behave erratically outside the regions with
shear. This has little impact on the solution, provided
the turbulence model is passive there, as shown in the
next section.
IV - APPLICATION
TO A
ONE-EQUATION
MODEL

Also as a result of symmetries, the quantity (3)


depends only on the deviators of Sij and of DSijlDt;
the divergence of the velocity field has no influence.
This can be verified by substituting S,, for Sjk, or Sij
for DSij/Dt
in (3).
Equation (3) reduces to 4w, [S11 (DSlz/Dt)
SrZ (DS,,/D)]
in 2D incompressible flow, showing

The application of the present ideas to the SpalartAllmaras one-equation model [8], although it is
preliminary, is useful in showing the capability of
the approach in a simple flow and its stability in a
complex one. Note an error on p. 21 [S]: the correct
definition of C,, is on p. 10.
Aerospace

Science

and Technology

On the Sensitization

of Turbulence

Models to Rotation

The production term cblwG of the model


multiplied by the rotation function h-l>
fT1

CT*;

and Curvature

301

is

+)

= (1+ Cd) G

(a)
(1 - erg tan-l

(~2,~?)) - crl. (5)

Here, ? = S/w, T* = S/w and the constants are


CT1= 1, ($2 = 12, CT3 = 1. These tentative values are
based on the wingtip calculations of Dacles-Mariani et
al. [9], in the case of crl, and on curved and rotating
boundary layers, in the case of c,2 and cT3.
The form of frl ensures that when rX is singular
as a result of the strain tensor vanishing, the function
becomes insensitive to ? because of the T* factor.
The tan- dependence allows a large slope for small r
without large values for large rX.Finally, the constraint
frl (1, 0) = 1 is satisfied, for thin shear flows without
curvature.
Figzwe 1 shows the effect on the skin-friction
coefficient C, in a curved boundary layer. The
boundary-layer thickness is S and the radius of
curvature is R. Each boundary layer was allowed
to grow to a momentum-thickness Reynolds number
RO = 2000 with a constant R. The local S and C,
are then used for the figure. The standard model has
a very weak effect of rotation, and of the wrong sign
(a de-stabilized boundary layer has a lower vorticity,
leading to lower production in the model). Other
models which base their production on the strain rate
may have no effect at all. In contrast the model
modified with I,1 reproduces the stabilization and
de-stabilization in fair agreement with experiments, at
least for small 6/R. The alteration of the flow saturates
for [6/R] near 0.06 (this depends on c,~). Strong
convex curvature nearly causes relaminarization.

0.2 1

-0.08

-0.04

0.00

0.04

0.08

SIR
Fig. 1. - Skin-friction
variation
layer at Rg = 2000.

due to wall curvature

in a boundary

Figure 2 uses the backward-facing


step as an
application in a more complex geometry. The value of
c,3 was 0.6 instead of 1, as we are still experimenting
with the function frl.
Converged solutions were
obtained without special difficulties. They are not very
1997. no 5

(4
2
Y

1
0

2
Y

04

2
Y

(4
0
0

10

Fig. 2. - Flow over a backward-facing


step. (a) streamlines;
(b) eddy
viscosity,
standard S - A model; (c) production
term, standard model;
(d) eddy viscosity,
modified
S - A model;
(e) production
term,
modified
model.

different from the standard model; for instance the


peak eddy viscosity is lowered by 7% but the velocity
field (including the reattachment length) changes by at
most 1%. The production term is noticeably different;
its peak is 13% lower and it takes negative values
around (2: 9) = (0.5, 1.5) and (0.7, 0.4). However,
the integrated effect (on the eddy viscosity, and then
on the velocity) is weak. Although the quantity ? has
large excursions above the vertical region and in the
region near (z, y) = (0.4, 0.4) (because w, crosses
zero) the production term is still well-behaved.
In Figure 2 the effects of the jr1 function are
not correlated with the streamline curvature (compare
Figure 2 c and 2 e, and use 2 a). Therefore, this flow
is complex enough to create a conflict between our
proposed criterion and conventional thinking. We do
not know of a conclusive physical argument in either
direction. In fact, no rigorous argument is likely to
indicate a definitive trend for a simplistic quantity
such as the eddy viscosity. The validation of model
modifications outside the domain of thin shear flows
will be a lengthy process.
In summary, although extensive tests are yet to
be made and the implementation requires additional
decisions for each model, our goal of an empirical

302

scalar quantity that unifies rotation and curvature,


is Galilean-invariant,
fully defined in 3D, and of
manageable complexity, seems to be attained with (4).

REFERENCES
[l] Knight D. D., Saffman P. G. - Turbulence Model
Predictions for Flows with Significant Mean Stramline
Curvature, AZAA-78-258.
[2] Wilcox D. C., Chambers T. L. - Sreamline Curvature
Effects on Turbulent Boundary Layers, AZAA J., 1977,
Vol. 15, No. 4, 574-580.
[3] Lien F. S., Leschziner M. A. - Modeling 2D
Separation From a High Lift Aerofoil with a NonLinear Eddy-Viscosity Model and Second-Moment
Closure, Aeronautical J., April 1995. 125144.
[4] Bradshaw P. - Effects of Streamline Curvature on
Turbulent Flow, AGARD-AG-169, 1973.
[5] Speziale C. G. - Analytical Methods for the Development of Reynolds-Stress Closures in Turbulence, Ann.
Rev. Fluid Mech. Vol. 23, 1991, 107-157.

P. R. Spalart,

M. Shur

[6] Govindaraju S. P., Saffman P. G. - Flow in a Turbulent


Trailing Vortex, Phys. Fluids, 1971, 14, 10, 2070-2080.
[7] Childs R. E., Caruso S. C. - Assessment of Modeling
and Discretization Accuracy for High-Speed Afterbody
Flows, J. Propulsion, 1991, 7, 4, 607-616.
[S] Spalart P. R., Allmaras S. R. - A One-Equation
Turbulence Model for Aerodynamic Flows, La
Recherche Aerospatiale, 1994, 1, 5-21.
[9] Dacles-Mariani, J., Zilliac, G. C., Chow J. S., Bradshaw
P. - Numerical/Experimental Study of a Wingtip Vortex
in the near Field, AMA J., 1995, 33, 9, 1561-1568.

Acknowledgements. - This work is dedicated to Professor


Ha Minh. We greatly benefited from numerous discussions
with Dr. M. Strelets; with Dr. A. N. Secundov, who
drew our attention to Knight & Saffmans paper and to a
simplification in Eq. 3; and with Dr. F. Menter. Dr. S. R.
Allmaras reviewed the manuscript.

Aerospace

Science

and Technology

Potrebbero piacerti anche