Professional Documents
Culture Documents
1
the technical challenges posed by such a model mean 1. No use of normal-to-wall distance. Most three
that an interim capability is required. dimensional calculations contain areas where two
or more boundary layers are interacting, such as
Two-equation models can be seen as an advance on wing-body junctions. It is dicult, therefore, to
algebraic models, since they are transport in nature and locate a suitable wall distance for cells in such a
therefore can take account of history eects. On the region. For a Multiblock scheme it is also possible
other hand, by making use of the Boussinesq approx- that the upper part of a boundary layer will be
imation there is no longer a need for a six-component separated from the solid surface by a block bound-
Reynolds-stress closure and the number of equations is ary. Locating the relevant surface in a generalised
reduced. The present study discusses how a robust and Multiblock mesh is a complicated and expensive
ecient two-equation solution method has been achieved process.
and demonstrates the capabilities of the system using re-
cent BAe AVRO research on a new regional jet. 2. Simple source terms. The turbulence model
source terms should be functions of readily avail-
able quantities. For example, models which make
Solution Method use of second derivatives are undesirable. Source
terms which become extremely large very close to
As its name suggests, a two-equation model requires the walls cause stiness problems.
solution of two transport equations in addition to the
Navier-Stokes equations. The turbulence equations are 3. Straightforward boundary conditions. The
dierent from the mean
ow equations in that they in- turbulence model variables should have boundary
clude important source terms which can cause the system conditions that can be readily applied by setting
to become numerically very sti. The approach adopted `halo'-cell values. The use of more complex con-
by BAe is to solve the turbulence equations explicitly ditions, such as forcing the dependent variables to
in the same manner as the mean
ow. The system is match a function over the near-wall region is un-
kept stable through the use of relaxation on the multi- desirable since this imposes severe constraints on
grid, limiters applied to certain quantities, and careful the grid generation process.
application of the articial viscosity. In RANSMB the
dissipation takes the form of blended second and fourth The degree to which a transport model can represent
dierences. For the mean
ow the blending uses a shock key physical phenomena is also very important. For at-
sensor to switch on the second dierence at shocks. For tached
ows, most traditional models perform very well.
the turbulence equations the second dierencing is in- However, these models often fail to predict more com-
voked at discontinuities via a non-linear second dier- plex
ows containing features such as boundary layer
ence sensor on the turbulence quantities, and has a co- separation, either shock induced or pressure gradient
ecient large enough to give a rst order TVD prop- induced5; 6 . Work is underway to analyse and imple-
erty. An alternative method of solving such equations ment some of the numerous corrections and modica-
in such an explicit scheme is given by Davidson4, who tions which have been proposed to overcome such limi-
proposes a semi-implicit method for the turbulence equa- tations. The present study is concerned with the choice
tions, which is coupled in some way to the explicit mean of the basic model on which to proceed with such work.
ow solution. This method has proved successful, but
due to its implicit nature becomes complex to imple- All two-equation models use the Boussinesq approx-
ment, especially in a Multiblock environment. imation to model the turbulent stresses:
Choice of Model ? ui uj = t ( @x + @x ) ? 3 @x ij ? 23 kij ; (1)
0 0 @ui @uj 2 @uk
j i k
In order to install a transport turbulence model into an where the eddy viscosity is given by:
industrial CFD code the developer faces a number of
issues which may determine whether or not a certain k2
model is viable. In addition to the numerical stiness, t = C : (2)
"
there are further features of the turbulence equations
which are unique to each new model. When assessing a The widely accepted transport equation for turbulent
particular turbulence model for use in RANSMB there kinetic energy, k, is modelled as:
are certain properties which are desirable in the context
@ (ui k)
of a structured, Multiblock method. These include: @xi
= @ ( + t ) @k + P ? "; (3)
@xi k @xi k
2
where Pk is the production of k, and " is the specic a one-equation model in the near-wall region, which is
dissipation. matched to the two-equation model at a certain distance
from the wall. The most commonly used near-wall model
To achieve closure of the system an expression for the is that of Wolfshtein10, which solves Eqn.(3) in conjunc-
dissipation, ", is required. A large number of closures tion with an algebraic expression for ".
have been proposed over the last 20-30 years. Most of
these give a transport equation for " itself, based on the This approach has been shown to be very eective6 .
`standard' form: A major drawback in the context of the present study
is the need to nd a suitable location for the transition
@ (ui ") @ t @" " "2
@xi
= @x ( + ) @x + C"1 k Pk ? C"2 k : (4) from the near-wall model to the two-equation model. In
i " i a structured grid system it is usual to select a particular
parallel-to-wall grid line which represents a y+ value of
An alternative closure method is provided by around 200. For three-dimensional
ows with several
Wilcox7. This is written in terms of ! which represents intersecting surfaces it becomes dicult to locate a grid
the ratio between " and k: surface which will be satisfactory.
"
!= :
k
(5) use The most appropriate method would appear to be the
of a single model which is modied through the addi-
tion of wall damping terms to allow its use in both high
The transport equation for ! is given as: and low Reynolds number regions. A large number of
@ (ui !) @
t @!
! such models exist in the literature. Many of these have
= (+ ) + Pk ? !2 : (6) been tested including Jones-Launder11 , Chien12, Lam-
@xi @xi ! @xi k
Bremhorst13, Lien-Leschziner14 and Speziale9 . In prac-
tice it has been found that all these models give very
Near-wall asymptotic analysis8; 9 shows that these similar solutions for most
ow problems. They do, how-
basic forms of k-" and k-! models give incorrect pre- ever, dier substantially with regard to the ease with
dictions of key turbulent quantities approaching a solid which they can be made to produce stable solutions.
surface. In the case of the k-" equations, it is necessary
to add damping terms or use wall functions to reproduce Implementation of k-"
experimentally observed behaviour. In the case of the k- One key issue encountered when applying low Reynolds
! model it has been found that despite the near-wall number k-" models is the boundary condition for ".
inconsistencies, good predictions of boundary layer pro-
les and skin friction can be obtained. The suitability of Whilst
not
k tends to zero as the wall is approached, " does
vanish. The low Reynolds number models fall into
a large number of both the k-" and k-! type of models two categories, those which apply a specic boundary
have been assessed for use within the BAe solver. condition for ", and those which replace " with new vari-
To test the basic implementations, and to illustrate goables, such as "~ or , which are dened such that they
the following discussions, a
at plate boundary layer test to zero at the surface.
case is used. This is calculated for a
ow at M =0.5, One of the most popular of this latter category, and
Re=7 106, using a grid with cell distributions similar one which has been in use at BAe for many years, is the
to those typically used for aerofoil or wing calculations. Chien k-~" model. It is worthwhile repeating the deni-
The cell nearest the wall is within y+ 1 and there are tion of Chien's model here, to facilitate the description
around 30 cells below y+ =1000. Transition is applied at of the implementation.
x=l = 0:1 and the proles are extracted at x=l = 0:9.
The new dissipation is related to " by:
Assessment of k-" "~ = " ?
2k ; (7)
y2
As already mentioned, wall-eects are not suciently where y is the normal-to-wall distance.
well modelled by the standard k-" equations, and these
equations are only valid away from a surface, in high The new dissipation equation is:
turbulent Reynolds number areas of the
ow. The low
@ (ui "~) @ "~
Reynolds number region can be bridged with wall func-
@x
= @
@x
(+ t)
@x
+ C"1 k"~ Pk
tions, which allow relatively coarse meshes to be used. i i " i
However, no signicant increase in capability over the "~2 2
? C"2 k f2 ? y2 e "~ ?C4 y+
: (8)
algebraic models is obtained. An alternative is to use
3
Here, f2 is a wall damping function, cell. This procedure is used for all blocks adjacent to
2 walls. This requires that the grid be generated in such
f2 = 1 ? 0:22e?(Re =6) : t
(9) a way that blocks adjacent to walls are suciently large
One additional damping function is used to factor the to contain the whole low Reynolds number region. For
eddy viscosity in Eqn.(2), situations such as wing-body junctions where more than
one surface is present, y is set to the smaller of the two
(10) wall distances.
? C y +
f = 1 ? e 3 :
Whilst the boundary conditions for "~ are now straight-
35 160
p
= C4 ky
5
1 20
y +
: (11) 0 0
0 1 2 3 4 5 0 1 2 3 4 5
log(y+) log(y+)
Comparing this model with the three criteria listed
above it is clear that, although the boundary conditions Figure 1: Log law comparison (left) and eddy viscos-
appear to be straightforward (k=0, "~=0), the normal-to- ity prole for the
at plate boundary layer using the
wall distance measure, y, is used. For mesh blocks which Chien(+Wolfshtein) model.
do not have a wall boundary the damping functions are
not applicable, and the model reverts to the `standard'
high Reynolds number " model, Eqn.(4). This is a valid Towards the outer edge of the boundary layer, a
assumption, except immediately downstream of a trail- new problem is introduced by the ratio of k2 =" which
ing edge, where although there is no solid surface, the is present in the eddy viscosity, Eqn.(2). As the bound-
boundary layer(s) from the wing surface create a
ow, ary layer edge is approached, both k and " fall to zero.
which for a short distance, approximates a wall bounded Any slight error in the evaluation of these quantities may
ow. Without a special treatment for wake regions the cause the ratio to become signicant. This eect can
sudden transition from "~ to " here can lead to conver- cause large values of eddy viscosity to be predicted, even
gence diculties. though the turbulent quantities themselves are small.
To control the production of `spurious' eddy viscosity
Further, it can be seen that Eqn.(8) contains source outside the boundary layer a minimum limiter for " is
terms which become very large in the near-wall region introduced 16
,
as y ! 0. Small imbalances in these terms can lead to k 32
"min = : (12)
large production or destruction of "~, with undesirable lmax
consequences for the convergence of the solution.
This is derived from dimensional considerations, and
To overcome the near-wall and wake problems a small the maximum length scale, lmax , is typically set to 0:1
layer of cells next to the surface is solved using the Wolf- aerofoil chord. This limiter improves stability and also
shtein model. This approach resembles the method men- helps to reduce the production of spurious eddy viscosity.
tioned above to bridge the gap between high Reynolds However, this eddy viscosity is not eliminated altogether.
number models and the wall. However, since the Chien
model is low Reynolds number compatible, the choice of The
at plate solution for the Chien-plus-Wolfshtein
grid line for separating the two regions is not critical. In model, together with the corresponding prole of eddy
RANSMB four to eight cells next to a wall are used for viscosity, t , is shown in Fig. 1. A satisfactory compar-
the one-equation model. ison with the log-law velocity prole is obtained. How-
ever, it can be seen that the eddy viscosity remains at a
The distance, y, is found by calculating the distance level which is still much higher than the laminar viscosity
along a normal grid line from the solid surface to a given outside the boundary layer.
4
Assessment of k-! 35
Eqn. (14)
1.2
The k-! model appears to meet two of the three crite- 30 use of R limiter
1.0
U+ 0.6
5
The source terms in Eqn.(14) are multiplied by the ratio approached, approaches zero as a function of y2 . Since
t =R and the transport equation for is now written as: the method used to calculated local gradients uses rst
2 order averages (as in most CFD methods), the quadratic
@ (ui ) @ @ nature of is not properly accounted for. The last term
= (+ t) ? R Pk
@xi @xi @xi in Eqn.(17) contains a product of the gradients of , and
@ @ it is the incorrect evaluation of this source term which
+ Rt ? + t 2Rk @x @x
: (17) leads to the poor modelling of the
ow very close to the
i i
wall.
The limiter, R, is chosen such that in regions of the
ow
where the eddy viscosity is larger than 1% of the laminar To eliminate this discretisation error a higher order
viscosity the source terms are unaected. A useful side- gradient calculation method could be employed near the
eect of such a limiter is that the primary turbulence surface. A simpler alternative is to modify the model to
variables are now only used in the denominator when eliminate the quadratic behaviour. A further transfor-
the local turbulence level is above the 1% cut-o. This mation is carried out by the rst two authors to re-cast
precludes the possibility of division by zero. the equation in terms of a new variable, g, which rep-
resents p . Using the same source term limiter, R, the
The solution for this modied model on the
at plate new model can be written as:
is also given in Fig. 2. Once the edge of the boundary
layer is reached, the eddy viscosity decreases rapidly, and @ (ui k) = @ ( + t ) @k + Pk ? 2 k2 ; (18)
the limiter is activated, causing to be reduced smoothly @xi @xi k @xi R
to a very low level in the freestream. Meanwhile, the
velocity prole is seen to remain largely unaected. @ (ui g) @
t @g
g3
@xi
= @x ( + ) @x ? 2R Pk
Care is required in selecting the appropriate value
kg
i 3g kgi @g @g
for the limiter. Very close the wall the eddy viscosity + 2R ? + t
2
: (19)
is again small. All the calculations presented here use g R @xi @xi
meshes where the minimum wall cell height has a y+
value between 0.5 and 2. For these meshes the limit of The eddy viscosity is:
0:01 is found to work well. For ner meshes, a lower t = kg2 ; (20)
limit would be required.
35
and the constants are:
30
g ! = 2:0; C = 0:09;
k- = 5=9; = 0:075
k-g
25
The new solution is shown in Fig. 3 and it can be seen
20 that the correct velocity prole is now obtained.
U+
15 A similar transformation has been carried out by
Gibson and Dafa'Allap18 to convert a k-" model to q-
10
, where q represents k. However, this transformation
adds new source terms to the k equation. Since the k
5
diusion term is the only term involving gradients of k in
the k-g model the discretisation errors at the wall have
a much reduced eect. Hence, a transformation of the k
0
-1 0 1 2 3 4 5
log(y+) equation does not appear to be necessary in the context
Figure 3: Flat plate velocity proles for the k- of the present study.
model, Eqn.(18) and the k-g model, Eqn.(19).
2D Validation
Concerning the velocity prole predicted by the k-
model, it is apparent in Fig. 2 that the model gives Both the k-~" and k-g implementations described above
an incorrect prediction of U + implying that there is an have been validated on a large number of external
ows
over-prediction of skin friction. This can been traced for both 2D and 3D geometries. To demonstrate the
to an error in the calculation of the gradient of very capabilities of both models a smaller number of test cases
close to the solid surface. It is known that as the wall is will be described. First, a transonic aerofoil, and a low
6
4.5
0.008
-1.0
0.006 4.0
-0.5
0.004
3.5
Lift Coefficient
Cp Cf 0.002
0.0
3.0
0.0
0.5
Experiment -0.002 2.5
k-
k-g
1.0 -0.004 Experiment
2.0
k-
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 k-g
X/c X/c 1.5
7
shortcoming of the Chien implementation. A certain
-14
amount of `spurious' eddy viscosity outside the boundary
-12
Experiment layers can be seen, which aect a large area of the
ow
k-
above the geometry. As observed in the
at plate test
-10 k-g
case, this value of eddy viscosity is controlled, but not
-8 eliminated, by the use of limiters. The presence of such
Cp eddy viscosity can trigger undesirable diusion or even
-6 production of k which in turn can aect the levels of tur-
bulence at the boundary layer edge further downstream.
Meanwhile, the k-g implementation behaves well at the
-4
6 Exp. 6
dled by these two models. Fig. 9 shows contours of nor- Flap, 50% Flap, t.e.
o
malised eddy viscosity for the 21:18 case. Immediately
7 7
5 5
8
number, or the addition of extra components such as
winglets and
ap track fairings. Also required is the
ability to reliably detect the buet boundary.
Primary areas of interest in Navier-Stokes results will
be associated with shock locations and degradation of
the boundary layer towards the trailing edge including
3D eects, separations, and trailing edge pressure rise.
Related design areas where CFD has a particular contri-
bution to make are in the details of
ow in wing-fuselage
and pylon-wing junctions, these being very sensitive to
Reynolds number dependent separations, often in the
presence of quite strong shocks in the wing-nacelle gap.
Thus validation of the code for representative cases is
vital.
Figs. 10 and 11 show results for a BAe AVRO re-
gional jet wing-body research conguration for which
tunnel results are available. Flow conditions are tran-
sonic cruise at M = 0:79, = 1:0o, transition at 15%
chord, CL 0:42, but at a tunnel Reynolds number
based on c of 2.7 million. The mesh has 1.2 million
cells in 230 blocks with a rst cell depth of y+ 2 and
around 25 cells in the boundary layer. Pressure con-
tours in Fig. 10 are obtained by use of the k-g turbulence
model. Fig. 11 shows surface streamlines using both k-g
and k-~" models. These streamlines can be compared to
wind tunnel oil-
ow visualisations. The k-~" results show
a separation at the wing root trailing edge which is not
present in the wind tunnel. This separation is not shown
by the k-g results, for which the streamlines match the
behaviour seen in the wind-tunnel. Note that the wing
is attached to a belly fairing which presents a vertical
wall at the wing root thus avoiding acute angles at the
wing-fuselage junction, and that the leading edge root
fairing further suppresses any tendency to separation.
Elsewhere the two sets of streamlines show very similar
Figure 9: Contours of eddy viscosity (t =) for the properties, notably the curvature and incipient separa-
Chien model (top) and the k-g model. Values range tion near the trailing edge crank which is also suggested
from 0 (Red) to 800 (Violet). in the tunnel oil
ow. Although not clear in these g-
ures, the surface
ow pattern around the leading edge
fairing is well predicted by both models.
3D Applications It is signicant that trailing edge root separations
Two applications are described from current BAe AVRO similar to that seen with k-~" are also produced for this
research studies on regional jet congurations. At case when Baldwin-Lomax and Johnson-King turbulence
present, the primary requirement for this
ow solver is models are used in RANSMB, the former model being
transonic cruise conditions with very accurate drag pre- noted for its resistance to separation. The one common
dictions for a wide spread of lift coecients around the feature in these three models that is not shared with k-g
design point extending well into the wave drag rise. The is the need for wall distance, which may be one source
wing designer would routinely compute CL/CD polars of error.
over a CL range of 0.1 to 0.7. Initial accuracy of 1% Fig. 12 shows pressure contours for a dierent
of total aircraft drag is sought together with drag incre- wing/body plus pylon/nacelle research conguration us-
ments to substantially less than one count. These in- ing the k-g model. Conditions are M = 0:8, = 1:0o ,
crements may be due to variation of shape or Reynolds
9
Figure 10: Contours of pressure coecient on a typical regional jet transport wing/body predicted by the
k-gmodel. M = 0:79, = 1:0o , Rec = 2:7 106 .
Figure 11: Skin friction lines on the same wing/body for k-"~ (left), and k-g.
10
Figure 12: Pressure coecient contours for a wing/body/pylon/nacelle research conguration calculated
using the k-g model. M = 0:8, = 1:0o , Re = 6:0 106 .
11
transition again at 15% chord, Rec = 6 106, and while both models give robust convergence on moder-
CL 0:41. The mesh has 2 million cells in 900 blocks ately complex geometries, the solutions given by the k-g
with a rst cell depth of y+ = 2. This is part of model are at least as good, and often better than those
an ongoing study into nacelle-pylon installation and is of the k-~" model. When applied to three-dimensional
shown here with a very closely coupled nacelle and over- design tests, stable solutions can be obtained with both
wing pylon. Flow disturbance can be seen at the wing- transport models. The k-g model, with its lack of de-
fuselage junction. A small horseshoe vortex appears at pendency on wall distance, appears particularly useful
the leading edge root. No belly or wing root fairings are for complex aircraft congurations. Initial results indi-
present in this simulation and the wing-fuselage junc- cate that accurate and reliable solutions can be obtained.
tion presents an acute angle at the trailing edge. This The increased condence in the solutions which are now
could be expected to lead to incipient separation in the being obtained is helping to increase the use of CFD by
wing-fuselage junction. Tunnel tests are planned. the design teams.
12
[9] Speziale C.G., Abid, R. & Anderson E.C. Criti-
cal Evaluation of two-equation models for near-wall
turbulence. AIAA Journal, 30 pp324-331, 1992.
[10] Wolfshtein M.W. The Velocity and Temperature
Distribution in One-Dimensional Flow with Turbu-
lence Augmentation and Pressure Gradient. Int.
Journal of Heat and Mass Transfer, 12 pp301-312,
1969.
[11] Jones W.P. & Launder B.E. The prediction of lami-
narisation with a two-equation model of turbulence.
Int. Journal of Heat and Mass Transfer, 15 pp301-
314, 1972.
[12] Chien K-Y. Predictions of Channel and Boundary-
Layer Flows with a Low-Reynolds-Number Turbu-
lence Model. AIAA Journal, 20 pp33-38, 1982.
[13] Lam C.K.G. & Bremhost K.A. Modied Form of
k- Model for Predicting Wall Turbulence. ASME
Journal of Fluids Engineering, 103 456-460, 1981.
[14] Lien F.S. & Leschziner M.A. Computational Mod-
elling of 3D Turbulent Flow in S-Diuser and Tran-
sition Ducts. Engineering Turbulence Modelling
and Measurements, 2 Elsevier, 1993.
[15] Kalitzin G. Validation and Development of Two-
Equation Turbulence Models. In: ECARP (Vali-
dation Area) Report, to be published in Notes on
Numerical Fluid Mechanics Series, by Vieweg, 1996.
[16] Benton J.J. Description of Methods Used by British
Aerospace. In: `EUROVAL, A European Initiative
on Validation of CFD Codes.' Haase, Brandsma,
Elsholz, Leschziner, Schwamborn (eds.) Notes on
Numerical Fluid Mechanics Vol 42, Vieweg, 1993.
[17] Menter F.R. In
uence of Freestream Values on k-!
Turbulence Model Predictions. AIAA Journal, 30
pp1657-1659, 1992.
[18] Gibson M.M. and Dafa'Alla A.A. A Two-Equation
Model for Turbulent Wall Flow. AIAA Journal, 33
pp1514-1518, 1995.
[19] Cook P.H., MacDonald M.A. & Firmin M.C.P.
Aerofoil 2822 - Pressure Distributions, Boundary
Layer and Wake Measurements. Appendix A6,
AGARD AR-138 , 1979.
[20] Moir I.R.M. Measurements on a Two-Dimensional
Aerofoil With High Lift Devices. Appendix A2,
AGARD AR-303 , 1994.
13