Sei sulla pagina 1di 17

Aerospace Science and Technology 7 (2003) 493–509

www.elsevier.com/locate/aescte

An evaluation of RANS turbulence modelling


for aerodynamic applications
Pietro Catalano ∗ , Marcello Amato
C.I.R.A., Italian Aerospace Research Center, 81043 Capua (CE), Italy
Received 25 January 2002; received in revised form 10 March 2003; accepted 23 June 2003

Abstract
In a numerical simulation the choice of a turbulence model must be a compromise between physical modelling and computational cost.
The CIRA RANS flow solver has been applied, by employing a large set of turbulence models, to typical aerodynamic applications for which
certified experimental data are available in literature. The transonic flows over an airfoil and a wing placed in a wind tunnel, both characterized
by a strong shock-boundary layer interaction with an induced separation, and the high lift flow around a multi component airfoil are taken into
consideration. An evaluation, in terms of accuracy and numerical behaviour, of some common turbulence models ranging from one-equation
to high order, using the same code and numerics, is presented.
Satisfactory and consistent results have been achieved; the more sophisticated the turbulence model the more accurate the simulation is.
The SST Menter κ–ω turbulence model has shown, for the applications investigated, the best compromise between the physical capabilities
and the numerical stiffness.
 2003 Éditions scientifiques et médicales Elsevier SAS. All rights reserved.
Riassunto gnas
In una simulazione numerica la scelta di un modello di turbolenza deve essere un compromesso tra la modellistica fisica ed il costo
computazionale. II solutore fluidodinamico del CIRA ZEN é stato impiegato, utilizzando un’ ampia gamma di modelli di turbolenza, per
tipiche applicazioni di aerodinamica per le quali esistono in letteratura dati sperimental certificati. Sono stati presi in considerazione i flussi
transonici intorno ad un profilo ed ad un’ ala in galleria del vento, entrambi caratterizzati da una forte interazione urto-starto limite, ed
il flusso in condizione di alta portanza intorno ad un profilo multi-componente. É presentata una valutazione, in termini di accuratezza e
comportamento numerico di alcuni diffusi modelli di turbolenza, da modelli ad un’ equazione a modelli di ordine superiore, utilizzando lo
stesso codice e la stessa numerica.
Risultati sodisfacenti e consistenti sono stati ottenuti; piú sofisticato é il modello di turbolenza piú accurata é la simulazione. II modello
κ–ω SST di Menter ha mostrato, per le applicazioni prese in considerazione, il miglior compromesso tra potenzialitá fisiche e rigidezza
numerica.
 2003 Éditions scientifiques et médicales Elsevier SAS. All rights reserved.

Keywords: RANS; Turbulence modelling; Transonic; High lift

Parole chiave: RANS; Modellistica della turbolenza; Transonico; Alta portanza

1. Introduction The Reynolds Averaged Navier–Stokes (RANS) equa-


tions, obtained by time averaging the NS equations, are
The accuracy that can be achieved in a Navier–Stokes usually employed to simulate turbulent industrial flows.
simulation strongly depends on the prediction of the char- These equations need a closure to compute an unknown, the
acteristics of the turbulent flow field. Important physical Reynolds tensor constituted by the double correlation of the
phenomena, such as boundary layer separation and shock- turbulent fluctuations, that derives from the convective term
boundary layer interaction, can be predicted with sufficient of the NS equations. The task of a turbulence model is to
accuracy only by a correct simulation of the turbulent flows. close the RANS equations by computing the components of
the Reynolds stress tensor.
* Corresponding author. In this paper an evaluation, using the same code and nu-
E-mail address: p.catalano@cira.it (P. Catalano). merics, of some among the most used turbulence models for
1270-9638/$ – see front matter  2003 Éditions scientifiques et médicales Elsevier SAS. All rights reserved.
doi:10.1016/S1270-9638(03)00061-0
494 P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509

Nomenclature

α Incidence of the flow and generic exponent µt Eddy viscosity


b Wing span Re Reynolds number
c Airfoil chord Ret Turbulent Reynolds number
z Airfoil coordinates ρ Density
Cζ1 , Cζ2 , Cζ3 Closure coefficients of the turbulence S Mean strain rate tensor
model σκ , σζ Closure coefficients of the turbulence model
Cµ Closure coefficient of the turbulence model τ Reynolds tensor
δij Kronecker delta
u Mean velocity
ε Turbulent dissipation rate
u Fluctuating velocity
κ Turbulent kinetic energy
κa von Kármán constant x Airfoil coordinates
fµ , fζ1 , fζ2 Viscous damping functions y Distance
l Turbulent length scale y+ Distance in wall units
M Mach number ω = ε/κ Turbulent specific dissipation
µ Molecular viscosity Ω Mean vorticity tensor

typical aerodynamic applications is presented. The transonic 2.1. Linear models


flows over an airfoil and a wing placed in a wind tunnel,
both characterized by a strong shock-boundary layer interac- The Reynolds stress tensor, by applying the Boussinesq
tion with an induced separation, and the high lift flow around hypothesis, results linearly related to the mean flow strain
a multi component airfoil are taken into consideration. The tensor through the eddy viscosity :
CIRA flow solver ZEN has been applied, by employing the  
∂ui ∂uj 2 ∂uk 2
Spalart–Allmaras, the Myong–Kasagi κ–ε, the Wilcox κ–ω, τij = µt + − δij − ρκI , (2)
∂xj ∂xi 3 ∂xk 3
the Kok TNT κ–ω, the Menter SST κ–ω, and a non linear
κ–ε model, to simulate specific flow conditions around the where κ is the turbulent kinetic energy. The eddy viscosity
RAE 2822 airfoil, the wing RAE M2155, and a wing section depends on the velocity and the length scale of the turbulent
of the A310 aircraft. Certified experimental data are avail- eddies
able in literature for these test cases.
µt ∝ κ 1/2 l α (3)
In order to highlight the turbulence models giving the best
compromise between physical modelling and computational and plays the same role as the molecular viscosity in the
cost, both accuracy and numerical behaviour are discussed. molecular diffusion.
A short overview of the models, and their implementation in Several turbulence models, ranging from algebraic to the
the CIRA code ZEN is also given. Reynolds stress models have been developed and can be
found in literature. Each model is calibrated on a specific
range of applicability and requires constants specification.
The simplest turbulence models are the algebraic models
2. RANS turbulence modelling where the eddy viscosity is completely determined in terms
of local mean flow variables. These models are cheap and
robust, but are not able to take into account important
A set of transport equations to directly compute the
effects of the flow history. The well-known and widely used
components of the unknown Reynolds stress tensor
Baldwin and Lomax [4] model is adopted in ZEN.
In the one-equation turbulence models, only one or a
τij = −ρui uj (1) combination of the turbulent scales is computed by solving
a transport equation. The Spalart–Allmaras model [21],
can be directly derived by the Navier–Stokes equations. widely used in aeronautical applications, is implemented in
Currently, the difficult numerical handling and the high the CIRA code.
computational cost required to solve the transport equations The two-equations turbulence models are complete in the
system, makes this technique not suitable for the simulation sense that two transport equations for both the turbulent
of turbulent engineering flows. The standard approach is to scales are solved, and the Reynolds stress tensor can be
relate the unknown Reynolds stresses to the known mean completely determined from the local state of the mean flow
flow quantities through a turbulence model. and of the mean turbulent quantities. The velocity scale is
P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509 495

chosen to be the square root of the turbulent kinetic energy • κ–ω models
κ, while the length scale is usually determined from κ and κ
µt = fµ Cµ ρ . (8)
an auxiliary quantity. ω
One of the most popular two-equations turbulence mod-
els, is the κ–ε model that makes use, as second turbulent Some models (i.e., κ–ε) employ, for boundary layer flows in
scale, of the turbulent dissipation rate ε. This model, pro- the near wall regions, the so-called damping functions that
posed in the standard form by Jones and Launder [13], has multiply the coefficients of the ζ equation and Cµ in (7). The
been the subject of many researchs and efforts mainly de- constants and the damping functions of the κ–ζ turbulence
voted to the stiffness of the ε equation. The Myong and models coded in the CIRA code ZEN are summarized in
Kasagi κ–ε turbulence model [16] is coded in ZEN. Table 1.
The κ–ω turbulence models are becoming more and more
popular in the last years, because it has been shown that they 2.2. High order models
are less stiff, and are more accurate for boundary layers flows
subject to adverse pressure gradient than the κ–ε models The turbulence models that make use of the linear
[26]. constitutive relation (2) are not able to resolve the normal
The original κ–ω model developed by Wilcox [27] stress anisotropy. Furthermore, the linear models are based
shows, especially for free shear layers but also for boundary on the hypothesis of local equilibrium between stress and
layes, a free stream dependency; the results depend on the strain which does not allow to take into account important
free stream values of the turbulent variable and in particular stress relaxation effects.
of ω. Since the κ–ε model generally does not present Turbulence models that provide an anisotropic general-
this free stream dependency, Menter proposed to include a ization of the eddy viscosity concept avoiding the complex-
∂κ ∂ω ity of the Reynolds stress models have been developed. Ex-
cross diffusion term (∝ ∂x j ∂xj
) with blending functions that
allows to switch from κ–ω to κ–ε approaching the edge of a amples are the explicit algebraic Reynolds stress models
boundary layer [15]. Kok, extending to the κ–ω the analysis (EARSM) [10,25] that represent a simplified solution of the
originally applied to the κ–ε model [7], has proposed a Reynolds stress transport equations, and the nonlinear eddy
model (TNT) that seems to overcome the drawback of the viscosity models [22,23] where the constitutive relation (2)
free stream dependency [14]. In the TNT κ–ω model, the is considered as the leading term of a series expansion of
cross diffusion term is taken into account only if positive. functionals.
The standard Wilcox, the Menter SST and the Kok TNT Non linear κ–ε turbulence models have been developed
κ–ω turbulence models have been implemented in the CIRA by the authors, during the E.G. funded project AVTAC [11],
RANS flow solver [6]. coupling the constitutive relation by Shih [24] to the Myong
and Kasagi κ–ε model [2,3]. The Shih Reynolds stress
2.1.1. κ–ζ turbulence models tensor with quadratic and cubic terms reads as:
The transport equations of a generic κ–ζ turbulence  
∂ui ∂uj 2 ∂uk 2
model (where ζ can be assumed to be ε or ω) can be written τij = µt + − δij − ρκδij
as ∂xj ∂xi 3 ∂xk 3
3  
∂(ρκ) ∂(ρκuj ) A3 κ ∂uk ∂uk ∂ui ∂uj
+ + ρ −
∂t ∂xj 2 ε2 ∂xi ∂xj ∂xk ∂xk
   4 
∂ µt ∂κ κ ∂uk ∂uk ∂up ∂uk ∂uk ∂up
= Pκ − Dκ + µ+ , (4) + A5 ρ 3 +
∂xj σκ ∂xj ε ∂xi ∂xp ∂xj ∂xj ∂xp ∂xi
 
∂(ρζ ) ∂(ρζ uj ) ζ 2 1 ∂ul ∂ui ∂uk ∂uj ∂uk 2
+ = (fζ1 Cζ1 Pκ − fζ2 Cζ2 Dκ ) − Π2 δij − + − Π1 δij
∂t ∂xj κ 3 2 ∂xl ∂xk ∂xj ∂xk ∂xi 3
      
∂ µt ∂ζ ρ ∂κ ∂ζ 1 ∂ul ∂uk ∂uk ∂ui ∂uj 2
+ µ+ + Cζ3 , (5) − + − Π2 δij , (9)
∂xj σζ ∂xj ζ ∂xj ∂xj 2 ∂xl ∂xi ∂xj ∂xk ∂xk 3
where Pκ , and Dκ stand for the production and destruction where the invariants Πi are defined as:
terms of κ, and are evaluated as
∂ui ∂uj ∂ui ∂ui
∂ui Π1 = , Π2 = , (10)
Pκ = τij , Dκ = ρε. (6) ∂xj ∂xi ∂xj ∂xj
∂xj
∂ui ∂ui ∂up
The eddy viscosity is computed through the turbulent vari- Π3 = . (11)
∂xk ∂xp ∂xk
ables as
The functions A3 and A5 , in front of the quadratic and
• κ–ε models cubic term respectively, depend on the turbulent variables
κ2 and on the main strain and rotation tensor, and ensure the
µt = fµ Cµ ρ ; (7) realizability of the model.
ε
496 P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509

Table 1
Constants and damping functions of the implemented κ–ζ turbulence models

Turbulence model σκ σζ Cζ1 Cζ2 Cζ3 Cµ

κa2
Myong–Kasagi κ–ε 1.4 1.3 Cζ2 − √ 1.8 0 0.09
σζ Cµ
5 − κ2
Wilcox κ–ω 2 2 √a 0.075 0 1
6 σζ 0.09
5 − κ2
Kok TNT κ–ω 1.5 2 √a 0.075 0.5 1
6 σζ 0.09
1 0.3ζ
Menter SST κ–ω F1 σκω +(1−F1 )σκε F1 Cκω + (1 − F1 )Cκε 2(1 − F1 )σζ2 max(0.3ζ,ΩF2 )
Cζ2 σ κ2
√ζ a
κ–ω type 0.85 0.5 0.09 − 0.09 0.075
Cζ2 σ κ2
√ζ a
κ–ε type 1.0 0.856 0.09 − 0.09 0.0828

F1 = tanh(a14 )

κ 4ρσζκε κ
a1 = min[max( 0.09ζy , 500ν
2 ), ] F2 = tanh(a22 )
ζy CDκζ y 2

∂κ ∂ζ , 10−20 ]
CDκω = max[2ρσζκε ζ1 ∂x
2 κ
a2 = max[ 0.09ζy , 500ν
j ∂xj 2 ]
ζy

Turbulence model fζ1 fζ2 fµ

2 + +/70 )
Myong–Kasagi κ–ε 1 (1 − 29 e(Ret /6) )(1 − e(−y /5) ) (1 + √3.45 )(1 − e(−y )
Ret
Wilcox κ–ω 1 1 1

Kok TNT κ–ω 1 1 1

Menter SST κ–ω 1 1 1


1 − (9/2)Cµ2 (κS ∗ /ε)2 The use of a quadratic constitutive relation is effective when
A3 = , (12) the anisotropy of turbulence is important, while the cubic
0.5 + (3/2)(κ 2/ε2 )Ω ∗ S ∗
term is necessary only for swirling flows.
1.6µt
A5 = (13)
(ρκ 4 /ε3 )(7(S ∗ )2 + (Ω ∗ )2 )/4
3. Numerical method
with
 1
S∗ = Sij∗ Sij∗ , Sij∗ = Sij − Skk δij (14) The CIRA code ZEN (Zonal Euler Navier–Stokes) has
3 been used in the numerical simulations. ZEN is a multiblock
and solver for the Euler, TLNS and RANS equations based on
 

 1 ∂ui ∂uj a finite volume cell centered approach. A central differenc-
Ω = Ωij Ωij , Ωij = − . (15) ing with blended self adaptive second and fourth order artifi-
2 ∂xj ∂xi
cial dissipation is employed. The discretization inside blocks
The coefficient Cµ is not a constant but a function of κ, ε, and at block interfaces is done in a conservative and uniform
and of the main strain and rotation tensor: flow consistent way and allows for local block refinement on
1 a block by block basis. The solution procedure is based on
Cµ = , (16)
4 + As U ∗ (κ/ε) a time marching concept. The multigrid scheme is used to
where accelerate the convergence of the solution, and performs re-
√ √  laxations, by using the Runge–Kutta algorithm, on different
1
As = 6 cos φ, φ= arccos 6W ∗ (17) grid levels. Local time stepping, residual averaging and local
3 block relaxation [8] are the techniques employed. The code
with is parallelized by using different message passing softwares
Sij∗ Sj∗k Ski

like MPI, PVM, and OMP [17]. A preconditioning method
W∗ = (18) for low speed flows has been recently implemented [18].
(S ∗ )3
The turbulence equations are uncoupled by the RANS
and
 equations and are solved, inside a multigrid cycle, only on
U∗ = (S ∗ )2 + (Ω ∗ )2 . (19) the finest grid level.
P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509 497

3.1. Discretization of the turbulence equations focusing on the accuracy and on the numerical behaviour
of the adopted turbulence models.
The turbulence equations (4) and (5) can be written in
integral form, by applying the Gauss theorem, as (e.g., for 4.1. RAE 2822 airfoil
the (4))
The RAE 2822 airfoil is a transonic subcritical airfoil
d
ρκ dV = (Pκ − Dκ ) dV deeply investigated from both theoretical and experimental
dt point of view. It has been tested in the RAE wind tunnel in
V V
    11 different flow conditions in the range of Mach numbers
µt ∂κ
+ ρκui − µ + ni dS, (20) 0.676–0.750, and at several Reynolds numbers [1]. The
σκ ∂xi case 10, that presents a shock induced separation with a
∂V
re-attachment upstream of the trailing edge, has been the
where V is the computational domain and ∂V is its boundary subject of many numerical simulations, and has been chosen
surface. Eq. (20), by means of the cell centered finite volume as a test case to validate and assess turbulence models in
approach, reduce for each cell (i, j, k) to E.C. funded projects such as EUROVAL [12] and AVTAC
dUij k [11], and is part of the database of the E.U. thematic network
Vij k + Rijc k − Rijv k − Vij k Qij k = 0, (21) QNET-CFD [5]. A single-block mesh with a C topology and
dt
272 × 80 cells has been employed.
where Ui,j,k is the vector of the unknown ρκ and ρζ , Rijc k
and Rijv k are the total net fluxes (convective and dissipative 4.1.1. Case 10
respectively), and Qij k stands for the source term. The The simulation of the case 10 flow condition represents a
convective fluxes are evaluated by means of a first order severe test for the CFD codes. The prediction of the shock
upwind scheme, while a central difference scheme is used location, of the separation extension, and of the pressure
for the dissipative fluxes. recovery behind the shock and in the trailing edge zone are
real challenges for the turbulence models.
The geometry adopted in the EUROVAL project [12] has
4. Results and discussion been used in the numerical simulations. A positive camber
correction
The CIRA flow solver ZEN has been applied to simulate
the flow around the RAE 2822 airfoil, the wing RAE M2155 3(z/c) = 0.006x/c(1 − x/c) (22)
placed in a wind tunnel, and around a three component is applied to the measured airfoil contour [1], and the flow
airfoil in high lift conditions by using the same numerics specification, after the wind tunnel correction, is:
and a large set of turbulence models. The first two cases
are representative of a transonic flow with a strong shock- • Mach = 0.754,
boundary layer interaction with an induced separation. The • α = 2.57◦ ,
last case is indicative of a high lift flow with its complex • Reynolds = 6.2 × 106 .
phenomena such as confluence between wakes and boundary
layers, strong adverse pressure gradients and stream lines The pressure and friction coefficients are presented in
curvature effects. The results are presented and discussed Fig. 1. With respect to the experimental data all the model

Fig. 1. RAE 2822 case 10 – pressure and friction coefficient.


498 P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509

Fig. 2. RAE 2822 case 10 – velocity profiles.

predict a shock located more downstream. The Kok TNT κ– 4.2. RAE M2155 wing
ω, the Myong–Kasagi κ–ε, and the Spalart Allmaras give
practically the same result, while the Wilcox κ–ω provides The RAE M2155 is a low aspect ratio wing (3.27) deeply
the worst and the SST Menter κ–ω and the nonlinear κ–ε tested by D.R.A. in the RAE wind tunnel in the Mach
the best shock location. The pressure recovery behind the number range 0.6–0.87 [9]. The wing RAE M2155 has been
shock is under-predicted by all the models with the SST the subject of many numerical simulations, has been chosen
Menter κ–ω and the nonlinear κ–ε showing the best pressure to validate and assess turbulence models in the E.C. funded
recovery in the trailing-edge zone. All the turbulence models project AVTAC [11], and is part of the data base of the E.C.
predict a shock induced separation with the Wilcox κ–ω funded thematic networks FLOWNET and QNET-CFD [5].
providing the smallest flow separated zone. The κ–ε models The grid employed in the simulations has 36 blocks and
yield the lowest values of the friction coefficient upstream of about 1.2 × 106 points.
the shock.
The velocity profiles at three stations are presented 4.2.1. Case 2
in Fig. 2. At x/c = 0.40 the flow is attached and all Experimental data for several combinations of Mach
the numerical results are in excellent agreement with the number and incidences are available. The flow condition
experimental data. The station x/c = 0.65 is just behind the at M∞ = 0.806, Re∞ = 4.1 × 106 , α = 2.5◦ , named as
shock, and the flow is separated. The Wilcox κ–ω, due to case 2, seems to be the most severe and relevant to assess
the bad prediction of the shock location, yields the worst the capabilities of a NS solver.
velocity profiles, while the results provided by the SST κ– At these conditions (Fig. 3), the flow on the upper surface
ω and the nonlinear κ–ε are the closest to the experimental of the wing is characterized by a triple shock wave system
data. At x/c = 0.90 all the numerical results lie in a narrow from the root to about the 50% of the span, and by a single
band except very close to the wall where the Wilcox κ–ω shock wave from about the 50% to the tip. Inboard the 50%
model gives a wrong behaviour of the velocity profile. span, changes in the flow direction occur in the region of the
P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509 499

TNT κ–ω yield a similar result, while the SST Menter κ–ω
provides a solution similar to the nonlinear κ–ε. The Wilcox
κ–ω predicts the flow to separate more outboard than all
the other models and provides the smallest flow separated
zone. The flow three dimensionality, visible by the curva-
ture of the skin friction lines, are more pronounced, espe-
cially in the outboard part of the wing, for the nonlinear κ–ε
model.
The pressure coefficients at several stations along the
span are shown in Figs. 6 and 7. All the numerical results lie
in a narrow band, and the comparison with the experimental
data is generally good. In the sections where the shock-
boundary layer interaction is stronger, the nonlinear κ–ε and
the Menter SST κ–ω better predict the shock location and
the pressure recovery behind the shocks and in the trailing
edge zones (Fig. 7).
The friction coefficients are shown in Fig. 8. The highest
values of the friction coefficient are provided by the Wilcox
Fig. 3. Wing RAE M2155 case 2 – oil flow visualization. κ–ω and the lowest by the Myong–Kasagi and the nonlinear
κ–ε models. The considerations, already expressed, about
the flow separation could be repeated.
forward leg of the triple shock wave system and in trailing The velocity profiles at three stations along the wing
edge zone but without flow separation. The flow separation span are shown in Fig. 9. The first two plots (top part of
starts where the three shock waves join together and ends the figure) are relevant to stations (x/c = 0.85, y/b = 0.30,
at about 90% of the span. The separation extends for about x/c = 0.43, y/b = 0.35) where the flow is attached and all
10% of the local chord. the turbulence models give a satisfactory result. The last plot
The flow field on the upper surface of the wing, as pre- (bottom part of the figure) refer to a very critical station
dicted by the turbulence models employed in the numeri- (x/c = 0.40, y/b = 0.77) where the flow is separated or just
cal simulations, is shown in Figs. 4 and 5. The shock wave re-attached. The Wilcox and the Kok TNT κ–ω provide the
system is simulated in a satisfactory manner, and all the worst results, while the Menter SST model, as the nonlinear
models predict a flow separation where the shocks meet. κ–ε, yields a velocity profile in excellent agreement with the
The Spalart–Allmaras, the Myong–Kasagi κ–ε, and the Kok experimental data.

Fig. 4. Wing RAE M2155 case 2 – pressure contour map and skin friction lines (upper surface). Spalart–Allmaras (left), Myong–Kasagi κ–ε (right).
500 P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509

(a) (b)

(c) (d)

Fig. 5. Wing RAE M2155 case 2 – pressure contour map and skin friction lines (upper surface). (a) κ–ω Wilcox standard, (b) κ–ω TNT, (c) κ–ω Menter SST,
(d) MK κ–ε + Shih O(2).

4.3. A310 wing section flap settings corresponding to the landing configuration (slat
deflection = 24.4 ◦ , flap detection = 32.4◦ ). For this case,
The CIRA RANS flow solver ZEN has been applied to experimental data at Mach number of about 0.22, Reynolds
simulate the flow around a multi-component airfoil in high
number of about 4 million, and at several incidences have
lift condition. The set of turbulence models coded in ZEN
has been employed, and a verification of the models, look- been made available during the E.C. funded project EURO-
ing at local properties and at the aerodynamic coefficients, LIFT [20]. The computational mesh was originally devel-
has been performed [19]. The geometry chosen is the 59% oped by DLR, and has been modified in house; it has 76992
wing span section of the A310 aircraft, with the slat and cells with 30 blocks.
P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509 501

Fig. 6. Wing RAE M2155 case 2 – pressure coefficients.

Fig. 7. Wing RAE M2155 case 2 – pressure coefficients. Zoom in the shock (left) and trailing edge (right) regions.

Fig. 8. Wing RAE M2155 case 2 – friction coefficients.

4.3.1. Pressure coefficients results. At α = 12 ◦ (top left plot of Fig. 10), the pressure
Pressure coefficients at several angles of attack are distribution is rather accurate on the slat and on the main;
presented in Fig. 10. The turbulence models used in the on the latter an underprediction is visible in the upper
computations provide, before the stall occurs, quite similar trailing edge region. On the flap the experimental data
502 P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509

Fig. 9. Wing RAE M2155 case 2 – velocity profiles.

show a clear separation which is not present in any of 4.3.2. Velocity profiles
the numerical simulations performed during the EUROLIFT At the angle of attack of α = 12.24 ◦ and α = 21.40◦ ex-
project. The disagreement could be due to the different perimental velocity profiles at several locations (see Fig. 11)
condition and turbulence levels of the boundary layer and of are available. The comparison between the numerical and
the wake impinging the flap. Increasing the angle of attack experimental data is presented in the Figs. 12 and 13 and
the agreement is generally good. The experimental results
the pressure distribution on the rear part of the main and
are best reproduced over the main component (rakes 3 and
on the flap gets in better agreement with the corresponding 4). At the stations located on the main and on the flap, the
experimental values. The suction peak of the slat and of influence of the wake coming from the upstream elements
the main are generally underpredicted, the κ–ω Menter SST and the related momentum loss is clearly visible. The results
yielding the results closest to the experimental data. At coming from all the turbulence models lie in a rather narrow
α = 21◦ the nonlinear κ–ε turbulence model has shown band and show the same flow features.
poor convergence levels, while only the Kok TNT and the
Menter SST κ–ω have been applied at incidences greater 4.3.3. Aerodynamic coefficients
than 22◦ in order to compute the drag polar of the airfoil. Two turbulence models, the Kok TNT and the Menter
At α = 25◦ (bottom right plot of Fig. 10), the airfoil, from SST κ–ω, have been used to compute the aerodynamic
coefficients of the airfoil. The results are presented in Figs.
the experimental point of view, is stalled. The Kok TNT
14–19 in terms of lift, drag (from pressure integration), and
model does not predict the stall yet, thus the agreement with pitching moments, for the complete configuration and for
the experimental data is poor. The κ–ω Menter SST clearly each single component, as functions of the angle of attack.
detects the stall with a large decrease of the pressure levels; The lift and the drag coefficient are overestimated and the
however, the numerical-experimental comparison is not very pitching moment underestimated, with the TNT κ–ω model
good. providing results slightly closer to the experimental data.
P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509 503

Fig. 10. A310 59% wing span section – pressure coefficients.

Looking at the contribution of the different components,


it is evident how the differences between the numerical
and the experimental results are mostly due to the main
and to the flap elements, while the slat contribution is in
quite good agreement with the experimental data. The stall,
with respect to the experimental data, is anticipated by the
κ–ω Menter SST and postponed by the Kok TNT κ–ω.
The sudden characteristic of the stall, as presented by the
experimental data, is predicted somehow by the Menter SST
κ–ω model. The stall, for the Menter SST κ–ω model, starts
with the formation of a bubble on the upper surface of the
slat that causes a sudden lost of lift for the slat and a much
stronger wake that affects the flow field on the downstream
components. This can be noticed in Fig. 20, where the stream
lines at different angles of attack, in the forward region
of the airfoil, are shown. The re-attachment point in the
cove of the slat, that is located close to the trailing edge
at α = 12 ◦ , moves upstream when the incidence increases, Fig. 11. A310 59% wing span section – rakes positions.

with the separated zone becoming almost negligible. A clear


flow separated zone appears again in the cove of the slat at
the stall angle.
504 P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509

Fig. 12. A310 59% wing span section – velocity profiles α = 12.24◦ .
P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509 505

Fig. 13. A310 59% wing span section – velocity profiles α = 21.40◦ .
506 P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509

Table 2
RAE2822 case 10–CPU time comparison

Turbulence model CPU time

Spalart–Allmaras 1.00

Myong–Kasagi κ–ε 1.33

κ–ε MK + Shih (2) 1.45

Wilcox κ–ω 1.25

TNT κ–ω 1.28

Menter SST κ–ω 1.39

Fig. 16. A310 59% wing span section – pitching moment coefficient.

tions presented in the previous sections, and the results are


discussed only with reference to the RAE 2822 airfoil.
A comparison of the convergence histories of the ρ and κ
equations is reported in Fig. 21. The κ–ω and the Spalart
Allmaras show a similar behaviour, while the stiffness of
the Myong–Kasagi κ–ε and mainly of the nonlinear κ–ε is
clearly visible. It is worth noting that the results shown are
relevant to the same convergence level and not to the same
number of iterations.
An estimate of the cpu time required by the different
turbulence models to run a fixed number of iterations
is summarized in Table 2 where the Spalart Allmaras
turbulence model has been considered as a reference. The
Fig. 14. A310 59% wing span section – lift coefficient.
cpu time of the two-equation turbulence models are of the
same order of magnitude and, as expected, bigger than the
cpu time of the Spalart–Allmaras model.

5. Concluding remarks

An evaluation of some popular turbulence models for


2-D and 3-D typical aerodynamic applications has been
presented. The CIRA flow solver ZEN, by employing the
Spalart Allmaras, the Myong and Kasagi κ–ε, the Wilcox,
Kok TNT, and SST Menter κ–ω, and a nonlinear eddy
viscosity turbulence model, has been applied to simulate
specific flow conditions around the RAE 2822 airfoil,
the RAE M2155 wing placed in a wind tunnel, and a
wing section of the A310 aircraft. The first two cases are
representative of a transonic flow with a strong shock-
boundary layer interaction with an induced separation. The
Fig. 15. A310 59% wing span section – drag coefficient.
last case is indicative of an high lift flow with its complex
phenomena such as confluence between wakes and boundary
4.4. Numerical behaviour layers, strong adverse pressure gradients and stream lines
curvature effects.
An analysis of the numerical behaviour of the turbulence The Spalart–Allmaras, the Myong and Kasagi κ–ε, and
models employed in the simulations has been performed. the Kok TNT κ–ω turbulence models have provided a
The conclusions achieved are the same for all the applica- similar result more accurate than the Wilcox κ–ω.
P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509 507

(a) (b) (c)

Fig. 17. (a) A310 59% wing span section, (b) aerodynamic coefficients, (c) slat contribution.

(a) (b) (c)

Fig. 18. (a) A310 59% wing span section, (b) aerodynamic coefficients, (c) main component contribution.

(a) (b) (c)

Fig. 19. (a) A310 59% wing span section, (b) aerodynamic coefficients, (c) flap contribution.
508 P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509

(a) (b) (c)

Fig. 20. (a) A310 59% wing span section, (b) stream lines in the slat region, (c) κ–ω menter SST.

(a) (b)

Fig. 21. RAE2822 case 10 – convergence history.

The Kok TNT has shown to be quite flexible. The reproduced with reasonable accuracy by both the models,
characteristic to avoid the computation of distances from but only the SST Menter κ–ω has predicted the sudden
solid walls allows to this model to preserve one of the main characteristic of the stall presented in the experimental data.
advantages of the standard κ–ω turbulence model. In conclusion, the SST Menter κ–ω, for the transonic and
The nonlinear κ–ε model makes use of a quadratic stress- high-lift flows investigated, has shown the best compromise
strain relationship that has shown to be effective whereas between the physical capabilities and the numerical stiffness.
the flow presents a strong shock-boundary layer interaction.
The results have been improved in terms of shock location,
References
pressure recovery behind the shock and in the trailing edge
zone and also in terms of the extension of the separated flow [1] AGARD-AR-138, Experimental data base for computer program
regions. Nevertheless the constitutive relation adopted has assessment.
enhanced the stiffness of the model, and it has been not [2] M. Amato, P. Catalano, An evaluation of stress–strain relationships for
possible to achieve acceptable results for the high lift flow aeronautical applications, XV AIDAA National Congress, November
15–19, 1999, Turin, Italy.
at the highest incidences. [3] M. Amato, P. Catalano, Non linear κ ε turbulence modeling for
The SST Menter κ–ω has provided results similar to industrial applications, ICAS 2000 Congress, August 28–September 1
the nonlinear κ–ε, but being less stiff and computational 2000, Harrogate, UK.
expensive. The shear stress transport formulation has shown [4] B.S. Baldwin, H. Lomax, Thin layer approximation and algebraic
to be effective and important to improve the results. This model for separated turbulent flows, AIAA paper 78-257, 1978.
[5] P. Catalano, M. Amato, A numerical evaluation of 2-D and 3-D ap-
model and the Kok TNT have been applied to compute the plication challenges of transonic aerodynamics by applying different
drag polar of a three-component airfoil in high lift conditions turbulence models QNET-CFD, Network Newsletter 1 (2) (2001) 25–
up to the stall. The aerodynamic coefficients have been 28.
P. Catalano, M. Amato / Aerospace Science and Technology 7 (2003) 493–509 509

[6] P. Catalano, M. Amato, Assessment of κ–ω turbulence modeling [17] V. Puoti, E. Bucchignani, A. Matrone, Different domain decomposi-
in the CIRA flow solver ZEN, ECCOMAS 2001 CFD Conference, tion techniques to solve computational fluid dynamics problems, Par-
September 4–7, 2001, Swansea Wales, UK. allel CFD Conference, May 11–14, 1998, Hsinchu, Taiwan.
[7] J.B. Cazalbou, P.R. Spalart, P. Bradshaw, On the behaviour of two- [18] V. Puoti, A preconditioning method for low speed flows 15th CFD
equation models at the edge of a turbulent region, Phys. Fluids 6 (5) Conference, June 11–15, 2001, Anaheim, CA, USA. AIAA paper
(1994) 1797–1804. 2001-2555.
[8] C. De Nicola, R. Tognaccini, V. Puoti, Local block relaxation method [19] V. Puoti, P. Catalano, M. Amato, High lift flow modeling for multi-
for the solution of the equations of gasdynamics, AIAA J. 38 (8) component airfoils, XVI AIDAA National Congress, September 24–
(2000) 1377–1384. 28, 2001, Palermo, Italy.
[9] DRA TR 92016, Measurement of the flow over a low aspect-ratio wing [20] R. Rudnik, Towards CFD validation for 3D high lift flows – EURO-
in the Mach number range 0.6 to 0.87 for the purpose of validation of LIFT, ECCOMAS CFD 2001 Conference, September 4–7, 2001,
computational methods. Swansea Wales, UK.
[10] T.B. Gatski, C.G. Speziale, On explicit algebraic stress models for [21] P.R. Spalart, S.R. Allmaras, A one equation turbulence model for
complex turbulent flows, J. Fluid Mech. 254 (1993) 59–78. aerodynamics flows, 30th Aerospace Science Meeting and Exhibit,
[11] A. Gould, J.C. Courty, M. Sillen, E. Elsholz, A. Abbas, The AV- January 6–9, 1992, RENO, USA. AIAA paper 92-0439.
TAC project – a review of European aerospace cfd, ECCOMAS 2000 [22] T.H. Shih, J. Zhu, A realizable Reynolds stress algebraic equation
Congress, September 11–14, 2000, Barcelona, Spain. model, NASA TM 105993, ICOMP-92-27, CMOTT-92-14.
[12] W. Haase, F. Bradsma, E. Elsholz, M. Leschziner, D. Schwamborn, [23] T.H. Shih, J.L. Lumley, Remarks on turbulent constitutive relations,
EUROVAL – An European Initiative on Validation of CFD Codes, in: Math. Comput. Modeling 18, 9–16.
Notes on Numerical Fluid Mechanics, Vol. 42, Vieweg, 1992. [24] T.H. Shih, K.-H. Chen, N.-S. Liu, A nonlinear κ–ε model for
[13] W.P. Jones, B.E. Launder, The prediction of laminarization with a two turbulent shear flows, 34th AIAA/ASME/SAE/ASEE Joint Propulsion
equation model for turbulence, Int. J. Heat Mass Transfer 15, 301–314. Conference and Exhibit, July 13–15, Cleveland, OH, AIAA paper 98-
[14] J.C. Kok, Resolving the dependence on free-stream values for the κ–ω 3983.
turbulence model, AIAA J. 38 (7) (2000) 1292–1295. [25] S. Wallin, A.V. Johansson, An explicit algebraic reynolds stress model
[15] F.R. Menter, Two equation eddy viscosity turbulence models for for incompressible and compressible flows, J. Fluid Mech. 403 (2000).
engineering applications, AIAA J. 32 (8) (1994) 1598–1605. [26] D.C. Wilcox, Turbulence Modeling for CFD, DWC Industries, Inc. La
[16] H. Myong, N. Kasagi, A new approach to the improvement of the κ–ε Canãda, California.
turbulence model for wall bounded shear flow, JSME Intern. J. Ser. 2 [27] D.C. Wilcox, Reassessment of the scale-determining equation for
33 (1) 63–72. advanced turbulence models, AIAA J. 26 (11) (1988) 1299–1310.

Potrebbero piacerti anche