Computers and Mathematics with Applications 70 (2015) 2238–2251

Contents lists available at ScienceDirect

Computers and Mathematics with Applications

journal homepage: www.elsevier.com/locate/camwa

An anisotropic shear stress transport (ASST)

model formulation
F. Vitillo a , C. Galati a , L. Cachon a,∗ , E. Laroche b , P. Millan b
a
DEN/CAD/DTN/STCP/LCIT, Bât.204 13108 Saint-Paul-lez-Durance, France
b
ONERA – The French Aerospace Lab, F- 31055 Toulouse, France

article info abstract

Article history: Computational fluid dynamics is nowadays a primary tool for a variety of engineering
Received 4 February 2015 applications. Aeronautic, chemical, nuclear, marine engineering use available models to
Received in revised form 13 May 2015 provide high technological products. Isotropic Reynolds-average Navier–Stokes (RANS)
Accepted 20 August 2015
turbulence models are widely used and their strengths and drawbacks are very well
Available online 26 September 2015
known. Nevertheless, since they are not supposed to provide accurate results for the most
complex flows, Reynolds stress transport (RST) models are proposed to take into account
Keywords:
Non-linear eddy viscosity model
for strong anisotropies. However its numerical behavior is often too expensive for industrial
Anisotropy applications. Alternative models as URANS, LES or DNS are far from an intensive industrial
RANS model application. Hence, a non-linear eddy viscosity (NLEV) model is proposed, which is shown
ASST to give very accurate results for a large variety of flow application with minor computation
effort needed, compared to traditional RANS models. In particular it consists of a quadratic
stress–strain relation with variable coefficients, which can capture secondary fluid motion,
coupled with the SST turbulence model proposed by Menter, which is supposed to correctly
represent the shear stress for adverse pressure gradient flows.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction

In the last decades Non-Linear Eddy Viscosity (NLEV) as well as Explicit Algebraic Reynolds Stress (EARS) models
received a great deal of attention due to their performance with regard to traditional Eddy Viscosity models and the minor
computation cost necessary to produce accurate results. Table 1 shows some of the better-known NLEV/EARS models
proposed.
All of them use a polynomial approximation to express the anisotropy tensor according to the Cayley–Hamilton theorem.
The main differences lie in the formulation of the closure coefficients (which are either constants or general functions of the
non-dimensional deformation or rotation rate) and in the non-linear approximation order. In particular, Craft et al. [3] and
Wallin and Johansson [5] showed that the third order terms are directly responsible for the correct description of rotating
system effects.
Given the success they obtained in modeling complex flows, an innovative Anisotropic Shear Stress Transport (ASST)
model is proposed, which will be shown to produce reliable results for a large variety of flows and applications. Since rotation
effects are often negligible for industrial applications (even though in turbo-machinery applications they are of primary
interest), the present model presents a 2nd order approximation of the anisotropy tensor, deferring the development of
third order terms (accounting for such rotation effects [3]) to future refinements of the model itself.

∗ Corresponding author. Tel.: +33 0 4 42 25 74 25.

E-mail address: lionel.cachon@cea.fr (L. Cachon).

http://dx.doi.org/10.1016/j.camwa.2015.08.023
0898-1221/© 2015 Elsevier Ltd. All rights reserved.
 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251 2239

Table 1
Some NLEV and EARS models proposed in literature.
Reference Non-linear approximation order Underlying model

Speziale [1] 2nd k–ε/k–kl

Shih, Zhou, Lumley [2] 2nd k–ε
Craft, Launder, Suga [3] 3rd k–ε̃
Hellsten [4] 3rd k–ε/k–ω BSL
Wallin, Johansson [5] 3rd k–ε/k–ω
Menter, Garbaruk, Egorov [6] 3rd k–ω BSL

2. SST model formulation

The Shear Stress Transport (SST) model has been proposed by Menter [7–9]. The basic idea is to take advantage of the
point of strength of both the k–ε and the k–ω models. In particular the first one is used in the flow bulk to use its stability
and proven capability to correctly reproduce bulk flow; the second one is used in the near wall region to use its capability to
correctly integrate transport equations all the way down to the wall with no need of specific near-wall treatments. To better
understand the formulation of the new model proposed hereafter, it is useful to remember the SST model formulation. The
chosen version of the SST model to be presented is the last one, i.e. that of Menter [7].
∂U
 
∂ Ui
The baseline of an isotropic model is the Boussinesq approximation, i.e. ρ u′i u′j = 2
ρ k − µt Sij , where Sij =
+ ∂ xij .
3 ∂ xj
Transport equations of the turbulent variables (the turbulence kinetic energy k and the specific dissipation rate ω) for an
incompressible flow are:
∂ ∂  ∂ ∂k µt
  
(ρ k) + ρ kuj = µ+
+ G˜k − ρβ ∗ kω

(1)
∂t ∂ xj ∂ xj ∂ xj σk
∂ ∂  ∂ µt ∂ω
  
(ρω) + ρωuj = µ+ + Gω − ρβω2 + Dω .

(2)
∂t ∂ xj ∂ xj σω ∂ xj
The turbulent Prandtl numbers for k and ω are expressed in the following way:
1
σk = F1 (1−F1 )
(3)
σk,1
+ σk,2

1
σω = F1 (1−F1 )
. (4)
σω,1
+ σω,2
The function F1 is computed as:
F1 = tanh φ14 ,
 
(5)
  √  
k4ρ k 500µ
φ1 = min max , , (6)
0.09ωy ρ y2 ω
σω,2 D+ωy
2

1 1 ∂ k ∂ω
 
D+
ω = max 2ρ , 10−10
. (7)
σω,2 ω ∂ xj ∂ xj
The turbulence kinetic energy production term in the SST models has a limiter, which provides a pseudo-realizability as
expressed by Durbin [10]. The limiter is defined as:

G˜k = min Gk , 10ρβ ∗ kω

 
(8)
where Gk is given by Eq. (30).
The specific dissipation rate production is given by:
ρα ˜
Gω = Gk , (9)
µt
α = F1 α∞,1 + (1 − F1 ) α∞,2 , (10)
βi,1 κ 2
α∞,1 = − √ , (11)
β ∗ σω,1 β ∗
βi,2 κ2
α∞,2 = − √ . (12)
β ∗ σω,2 β ∗
The coefficient β is expressed by:
β = F1 βi,1 + (1 − F1 ) βi,2 . (13)
2240 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251

The Cross-Diffusion term Dω is expressed as follows:

1 1 ∂ k ∂ω
Dω = 2 (1 − F1 ) ρ . (14)
σω,2 ω ∂ xj ∂ xj
Finally the model is closed by the definition of the eddy viscosity, which is limited within the boundary layer by the
Bradshaw’s assumptions [11], i.e.:
ρk 1
µt = (15)
ω max 1,
 
SF2
a1 ω

F2 = tanh φ22
 
(16)
 √ 
2 k 500µ
φ2 = max , (17)
0.09ωy ρ y2 ω
S is the modulus of the strain rate tensor, i.e.
 1/2
1
S= Sij Sij . (18)
2
Observe that the eddy viscosity formulation presents two terms: one is the standard High-Re eddy viscosity formulation,
ρk
i.e. µt = ω . The second one is based on Bradshaw’s assumption in boundary layer, i.e. τ = −ρ u′ v ′ = a1 ρ k. Since the
Reynolds stress in a boundary layer is: τ = −ρ u′ v ′ = µt S it results Eq. (15). Observe that the modulus of the strain rate
tensor has been chosen as the generalized velocity gradient. As the F1 function, the F2 function blends the High-Re and the
boundary layer formulation of the eddy viscosity. Finally, model constants are:
σk,1 = 1.176, σk,2 = 1.0, σω,1 = 2.0, σω,2 = 1.168, β ∗ = 0.09, a1 = 0.31,
βi,1 = 0.075, βi,2 = 0.0828.

3. New ASST model formulation

The basic idea of a Non-Linear Eddy Viscosity Model is to express the Reynolds stress tensor (and hence indirectly the
eddy viscosity) in a nonlinear polynomial form. In particular, following the approach of Shih et al. [2], the Reynolds stress
tensor can be expressed as (up to second order terms):
 
2 1
ρ u′i u′j = ρ k − µt Sij + ρ C1 kτ 2 Sik Skj − δij S kl Skl + ρ C2 kτ 2 Ωik Skj + Ωjk Ski
 
3 3
 
1
+ ρ C3 kτ Ωik Ωkj − δij Ω kl Ωkl
2
(19)
3
∂U
 
∂U
where Ωij = ∂ x i − ∂ x j and τ being the turbulent time scale. Usually, for an ω-based model it can be expressed as:
j i

1
τ= (20)
β ∗ω
with β = 0.09.
∗

The goal is to find the expression for the unknown coefficients C1 , C2 and C3 .
Refs. [3,2,5] affirm that these coefficients must be non-constant: indeed Wallin and Johansson [5] affirm that the
coefficient should be found from an implicit form of the anisotropy tensor based on well-known and well-established models
of the Reynolds stress tensor transport equations; instead [2] found the coefficient by application of Realizability on the
Reynolds stresses. We will retain the same approach as [2,12], applying the Realizability condition on the Reynolds Stresses,
obtaining:
CNL1
C1 = (21)
CNL4 + CNL5 · (τ S )3
CNL2
C2 = (22)
CNL4 + CNL5 · (τ S )3
CNL3
C3 = . (23)
CNL4 + CNL5 · (τ S )3
It is worth noting the values of the unknown coefficients, given in Table 2. Note that the major model improvement in [12]
is the increased value of the CNL2 coefficient to be able to represent the scale of the secondary motion.
 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251 2241

Table 2
NLEVM closure coefficient in [2,12] and in the present ASST
model.
Model coefficients [2] [12] Present model

CNL1 0.75 0.8 0.8

CNL2 3.75 11 11
CNL3 4.75 4.5 4.5
CNL4 1000 1000 1000
CNL5 1.0 1.0 1.0

For the present model we retain values of [12] given in Table 2.

In the original versions of the model (that were coupled with a standard k–ε model), to assure realizability condition a
non-constant definition of the Cµ coefficient in Eq. (23) was employed: for example in [12]:
2/3
Cµ = (24)
A1 + τ S
where coefficient A1 =3.9 is used to obtain Cµ = 0.09 for an equilibrium boundary layer where τ S = 3.5. For an ω-based
model the equilibrium value is rather equal to one: hence, to be consistent with the original model, the Cµ formulation must
be divided by β ∗ 1 to obtain:

7.407 7.4
Cµ = ≈ (25)
A1 + τ S A1 + τ S
with A1 = 3, 9 as in the original model.
The final High-Re eddy viscosity formulation results to be as follows:
ρk
µt = Cµ . (26)
ω
Anyway it is interesting to retain the Bradshaw’s assumption related limitations of the original SST model, since it is supposed
to provide very good results for adverse pressure gradient flows. Remember that the Bradshaw’s assumption states that in
boundary layers there is a linear relation between the principal turbulent shear stress and the turbulent kinetic energy,
i.e. τ = −ρ u′ v ′ = a1 ρ k for a flat plate channel flow. It is worth noting however that for a NLEVM there is a difference in the
formulation of the Reynolds Stress, i.e. it depends on the quadratic velocity gradient:

τ = −ρ u′ v ′ ∝ µt S − ρ C 1 kτ 2 S 2 − ρ C2 kτ 2 S 2 − ρ C 3 kτ 2 S 2 (27)
where the modulus of the strain rate tensor is chosen again as the generalized velocity gradient, as in Menter et al. [6] Hence,
a formally correct geometry independent eddy viscosity formulation should be:

ρ a1 k + C1 kτ 2 S 2 + C2 kτ 2 S 2 + C3 kτ 2 S 2 ρB
 
µt = = . (28)
S S
The introduced correction is supposed to correctly take into account the constant ratio of the principal turbulent Reynolds
stress to the turbulent kinetic energy in an important range of boundary layer flows [11], especially for complex geometries,
where non-zero values of the second order velocity gradients play an important role. Therefore the generalized formulation
given by Eq. (28) provides the right first and second velocity gradient scale thanks to the use of the modulus of the strain
rate tensor.
Given the previous considerations, the proposed eddy viscosity formulation taking into account the realizability condition
(i.e. strain-dependent eddy viscosity in the High-Re region) and the Bradshaw relation in boundary layers is:
ρB
µt =   (29)
Bω
max Cµ k
, F2 S
where the F2 function is the same as the SST model. The model is closed through the k and ω transport equations (1) and (2).
Finally, to be consistent with the NLEV formulation, the turbulence kinetic energy production term can be expressed in
the proper form:
∂ Ui
Gk = −ρ u′ı u′ȷ (30)
∂ xj
where uı ′ uȷ ′ is given by Eq. (19). Note that, due to the realizability condition given by the eddy viscosity modification and
the model coefficient expressions, no turbulence kinetic energy production limiter is needed.

1 Always remember that ω = ε and k = 1 hence there is always a factor equal to β ∗ between an ε -based and an ω-based formulation.
β∗ k ε β∗ ω
2242 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251

Finally the specific dissipation rate is expressed in the original form i.e.
ρα
Gω = Gk . (31)
µt
The other parameters not mentioned so far maintains the same definition as the original SST model formulation. Closure
constants are:
σk,1 = 1.176, σk,2 = 1.0, σω,1 = 2.0, σω,2 = 1.168,
β ∗ = 0.09, a1 = 0.31, βi,1 = 0.075, βi,2 = 0.0828.

4. ASST model analysis of the logarithmic layer

To justify the fact that no SST model closure coefficient has been changed in the ASST formulation and to motivate the
value of the A1 coefficient, we will show that the latter does not impact the model’s behavior in the logarithmic layer of the
law of the wall. Remember that the only coefficient changed by Menter from the original k–ω is the σk,1 coefficient in order
to recover the correct behavior for a flat plate boundary layer. The original k and ω equations are satisfied in the logarithmic
region by the following solutions (see Wilcox [13]):
uτ u y
τ
u= ln +C (32)
κ ν
u2
k = √τ (33)
β∗
uτ
ω = √ ∗  (34)
β κy
where C is a constant and κ is the von Karman constant.
It is worth studying the eddy viscosity solution first, since it has been modified by the new ASST formulation with respect
to the original SST formulation. In fact, as already mentioned, in the original SST model, the eddy viscosity solution in the
log-layer is (the hypothesis F2 = 1 has been made, since we are inside the boundary layer):
 
a1 k k
νt = min , . (35)
S∗ ω
Since S = √1 ∂∂Uy = √1 uκτy for simple shear flow log-layer, the following relation can be written:
2 2
√ ∗   √ 
u2 √ κ y u2 β κy

a1 2
νt = min a1 √ τ ∗ 2 , √ τ ∗ = min √ ∗ uτ κ y, uτ κ y . (36)
β uτ β uτ β
√
The value √1 β ∗ ≈ 1.46 lets us affirm that the High-Re formulation is used in a shear flow log-layer. For the ASST model the
a 2

Cµ modification has to be taken into account and the B coefficient has to be defined. It results that:
B = a1 k + C1 kτ 2 S 2 + C2 kτ 2 S 2 + C3 kτ 2 S 2
B = a1 + C1 τ 2 S 2 + C2 τ 2 S 2 + C3 τ 2 S 2 · k.
 
(37)
Remembering Eqs. (20)–(23) and Eq. (34), it can be demonstrated that:
√ 2
1 1 β ∗κ y uτ 2 1
τ 2S2 = S2 = = ≈ 5.6 (38)
β ∗ω β ∗2 uτ 2
2 (κ y)2 2β ∗
CNL1 0.8
C1 τ 2 S 2 = 3/2 τ S =
2 2
· 5.6 ≈ 0.0044 (39)
CNL4 + CNL5 · τ 2 S 2 1000 + 5.61.5


CNL2 11
C2 τ 2 S 2 = 3/2 τ S =
2 2
· 5.6 ≈ 0.061 (40)
CNL4 + CNL5 · τ S 1000 + 5.61.5

2 2

CNL3 4.5
C3 τ 2 S 2 = 3/2 τ S =
2 2
· 5.6 ≈ 0.025. (41)
CNL4 + CNL5 · τ S 1000 + 5.61.5

2 2

Hence the B coefficient results to be equal to:

B = a1 + C 1 τ 2 S 2 + C 2 τ 2 S 2 + C 3 τ 2 S 2 · k
 

= (0.31 + 0.0044 + 0.061 + 0.025) · k

≈ 0.4 · k. (42)
 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251 2243

Following the analysis previously done and with the hypothesis of F2 = 1, it results from Eq. (29):
7.4
   
B k B k
νt = min , Cµ = min , . (43)
S ω S A1 + τ S ω
Therefore, applying the already discussed conditions, the eddy viscosity is:
√
 
uτ 2 √ κ y 7.4 uτ 2 β ∗κ y 
νt = min 0.4 √ ∗ 2 , √ ∗ √
β uτ A1 + 1∗ β κ y √1 uτ β∗ uτ
β uτ 2 κy
 √ 
0.4 2 7.4
= min  √ ∗ uτ κ y, uτ κ y ≈ min (1, 88 uτ κ y, 1, 18uτ κ y) . (44)
β A1 + √1
2β ∗

Therefore we can conclude that the anisotropic formulation does not differ from the SST model behavior in the log-layer: in
fact the High-Re formulation is used again in this region, as already found in Eq. (36) for the SST model.
In the logarithmic layer of a simple shear flow the momentum, k and ω equations reduce respectively to:
∂ ∂U
 
0= νt (45)
∂y ∂y
∂U 1 ∂ ∂k
 2  
0 = νt − β ∗ ωk + νt (46)
∂y σk,2 ∂ y ∂y
∂U 1 ∂ ∂ω
 2  
0=α − βω + 2
νt (47)
∂y σω,2 ∂ y ∂y
which are the same formulations as the standard SST model and where the solutions given by (32)–(34) are still valid.
To demonstrate that these equations are the log-layer equations for the ASST model as well, note that the eddy viscosity
solutions does not change, as already demonstrated. Moreover, only the turbulence kinetic energy production term (and
hence the specific eddy dissipation production term) have been modified in the k and ω transport equations with regard to
those of the SST model. Hence, if there was a difference, it would be in such two terms. Indeed the second order terms never
appear, since in the simple shear flow only one principal velocity component U is present and only gradients along the y
direction (normal to velocity) are allowed. To demonstrate this, observe that the turbulent kinetic energy production term
in the logarithmic layer is:
∂ Ui ∂U ∂U

Gk
= −uv + β1 Sxx Sxy + Sxy Syy + Sxz Szy
 
= −uı uȷ = − −νt
ρ ∂ xj ∂y ∂y
+ β2 Ωxx Sxy + Ωxy Syy + Ωxz Szy + Ωxy Sxx + Ωyy Syx + Ωyz Szx
 

 ∂U ∂U
  2
+ β3 Ωxx Ωyx + Ωxy Ωyy + Ωxz Ωyz = νt .

(48)
∂y ∂y
Second order terms do not appear in the final formulation, since all other terms but Sxy = Ωxy = ∂∂Uy are equal to zero.
Once demonstrated that the Gk term does not change in the ASST model formulation, it results from Eq. (47) that the Gω
does not change too, since the two are related by multiplying coefficient. Hence it is verified that the ASST acts exactly like
the SST model in the logarithmic boundary layer. This is consistent with the fact that the non-linear terms are not supposed
to influence the flow main characteristics for simple shear flow cases, but only to act on flow anisotropy.

5. Model verification and validation

The ASST is now validated against four test cases of practical industrial interest: a fully developed channel flow, a 90°
bend developing flow, the squared cross-section straight tube and a tight-lattice rod bundle subchannel. Experimental as
well as DNS data will be used for validation. Both ASST and SST computational results are shown, to underline the differences
between the two models and the eventual gain obtained by the ASST model.
The solver is always Pressure-based and the pressure–velocity coupling is done with Semi-Implicit Method for Pressure-
Linked Equations (SIMPLE) algorithm. Gradients are evaluated through the Least Square Cell Based method. Finally the
Second Order Upwind Scheme is used for the spatial discretization of momentum, turbulent kinetic energy and turbulent
dissipation rate. Convergence criteria are based on a 10−5 value of absolute residuals of transported quantities.

5.1. Channel flow

The test case of fully developed channel flow has been studied extensively to increase the understanding of the mechanics
of wall-bounded turbulent flows. Its geometric simplicity is attractive for theoretical investigations of complex turbulence
interactions near a wall.
2244 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251

Fig. 1. Channel flow mesh.

Table 3
Channel flow grid convergence evaluation.
Mesh A Mesh B Mesh C
+
Y 2.242 1.487 0.743
τw 5.864e−04 Pa 5.801e−04 Pa 5.790e−04 Pa

In order to obtain a fully developed flow the height H and body length L of the plate is chosen respectively H = 1 m and
L = 80 H. Flat plate’s geometry is symmetric about the middle plane. The inlet boundary condition is taken as a velocity inlet
and the outlet boundary condition as a pressure outlet. The inlet velocity is set based on the desired flow Reynolds number.
In particular several Reynolds numbers, corresponding to different database, have been used to demonstrate the model’s
applicability on different flow conditions and turbulence levels. Experimental database of Laufer [14] (Re = 14 000) and
DNS database of Kim et al. [15] (Re = 5600) are used for validation.
Fig. 1 provides an example of the channel flow mesh.
To establish mesh convergence, the average wall shear stress τw is evaluated: Table 3 shows the results.
The comparison shows that there is almost no difference between the two finer grids (difference of 0.2% between the fine
and refined grids). Therefore we will adopt the finest calculation grid to capture the finest turbulence phenomena occurring.
Results are shown in Fig. 2.
The results shown in Fig. 2 have been first compared with the theoretical law of the wall. Results are obtained considering
velocity at x = 75 H location, where the flow can be considered as fully developed.
Fig. 2 shows that both the SST and the ASST models produce very similar good predictions. The same behavior is observed
for the computation of non-dimensional turbulent kinetic energy k+ = k2 , where both models fail to capture the turbulence
uτ
kinetic energy peak due to the form of the k-omega model used in the near-wall region, particularly due to the absence of
low-Reynolds formulations in the viscous sublayer. Anyway the two models correctly predict the trend in the logarithmic
region as well as in the outer region. Computational results based on the previously described turbulence models have been
compared with the DNS data available (Kim et al. [15]).
u′ v ′
Fig. 2 shows also that Shear Reynolds Stress is very well predicted by both the turbulence models. This Reynolds
u2τ
stress is of primary importance in channel flow since it is often the most important in terms of magnitude, justifying
the macroscopic phenomena occurring in such a flow. Therefore it is not surprising to see such results, since turbulence
models are usually tuned to be able to reproduce the turbulent shear stress. In Fig. 2 the normal Reynolds stress difference
τyy –τxx = v ′ 2 −u′ 2 , predicted by the ASST and by the SST model is compared as well with the experimental data of Laufer [14].
The standard SST model predicts that τyy –τxx = 0 since no effect of the only velocity gradient ∂ U /∂ y is present on the
principal Reynolds stresses (remember that only the Sxy = ∂∂Uy strain component is present in simple shear flow).
The difference between the SST and ASST predictions is entirely related to the effects of anisotropy. To verify this
statement, observe from Eq. (19) how the non-zero normal Reynolds stresses are calculated by the ASST model:
   
2 k 1 k  k 1
u′ 2 = 2
+ C2 2 2 ·Ω xy Syx + C3 2 ·Ω 2xy .

k + C1 · Sxy (49)
3 ω2 3 ω ω 3
This behavior is of primary importance for the prediction of more complex flows, since this parameter is considered to
be responsible to the prediction of the reattachment length downstream a backward facing step as well as the secondary
motion of the second type in curved ducts of circular cross-section [16].
In the end, the analysis of the channel flow shows that the formulated ASST model correctly represents the basic turbulent
characteristics of a turbulent flow. Finally, it has been demonstrated that the anisotropic formulation can reproduce flow
anisotropies even in a simple shear flow. This is very promising for the application to complex flows, where anisotropy is
supposed to play a major role.

5.2. Turbulent developing flow in 90° bend

Fluid flow in a bend is a baseline flow studied for several applications in applied sciences. Configurations of interest are,
for instance, centrifugal pumps, heat exchanger tube-side and industrial pipelines. It is of particular interest for the present
 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251 2245

25

SST
u+ = 2.5 ln(y+)+5.5
20

LOG-law of the
15
wall

u+
10 LINEAR law of the
wall
5
u+ = y+ ASST

0
1 10 100 1000
y+
5

DNS (Kim et al.)

3
k+

SST
2
ASST

0
0 0.1 0.2 0.3 0.4 0.5

y/h
1

DNS (Kim et al.)

0.8

SST
u'v'/u2τ

0.6

0.4 ASST
0.2

0
0 0.1 0.2 0.3 0.4 0.5

y/H
5
Exp.Data [Laufer]
v'2-u'2/(U0 x 10-3)

4 SST
3
ASST

-1
0 0.2 0.4 0.6 0.8 1

y/(H/2)

Fig. 2. Channel flow validation.

application since the in-bend flow coming from mixing zones cannot be considered as fully developed flow: therefore it is
interesting to verify the model capabilities to correctly describe this flow configuration. A characteristic which distinguishes
such flows from those in straight ducts is the generation of streamwise vorticity, or secondary motion, within the duct,
resulting in a pressure loss, the spatial redistribution of streamwise velocity and increased heat transfer at the duct wall.
Hence, the aim of this work is to numerically investigate the turbulent flow inside a 90 ° bend of squared cross-section,
and determine the appropriate turbulence modeling level required to describe this flow.
The present work will study the bend geometry proposed by Taylor et al. [17] (Re = 40 000): an entry duct enters the
90° bend, whose mean curvature radius is 92 mm. The squared cross-section side is 40 mm (Fig. 3). A 1 m/s velocity inlet
and a gauge pressure equal to 0 Pa pressure outlet (with regard to the atmospheric operating pressure) boundary conditions
are used. The working fluid is water at atmospheric pressure and temperature.
The geometry proposed in [17] (entry length of 7.5 D) did not give satisfactory numerical results due to the inability of
the uniform velocity inlet boundary condition to describe the experimental configuration before the entry length. Hence
a slightly longer inlet length of 11.25 D has been used to obtain a closer main velocity profile in the inlet section without
changing the paper base idea of developing flow into a bend. The original 50 D exit length has been retained.
2246 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251

Fig. 3. 90° bend geometry [17].

Table 4
90° bend used grids.
Mesh configuration Y+ τw (Pa)
A 0.27 2.61
B 0.15 2.70
C 0.12 2.72

Fig. 4. 90° bend studied profiles.

Using the described geometry, three different meshing configurations have been preliminarily studied in order to
evaluate the influence of mesh size on the solution. The three structured mesh configurations (hereafter named as A, B,
C) differ by a factor of around 2 on the mesh size in the first layer thickness and in the global size of a single mesh cell
(Table 4):
The Y + average values are 0.27 and 0.12 and 0.12 for configurations A, B and C whereas the maximum Y + are 0.78, 0.40
and 0.21 respectively. Meshing convergence has been evaluated by comparing the average wall shear stress. The comparison
shows a good agreement for the B and C configurations (difference in the wall shear stress less than 1%). The final choice is
to use Configuration B, which gives a good balance between calculation results and total number of mesh nodes.
The comparison has been done on the values of the non-dimensional principal velocity U /Vc , the non-dimensional radial
velocity V /Vc and non-dimensional the Reynolds stress—u′ v ′ /Vc 2 . These values will be shown in two different lines of a
cross-section: the line AA, defined as the line with r ∗ = 0.5 and 0 < Z ∗ < 1 and the line BB defined as 0 < r ∗ < 1 and
Z ∗ = 0. See Fig. 4 for explanation. See that the BB line covers only half cross-section due to symmetry considerations.
Results of the mixed Reynolds stress u′ w ′ are for the AA and BB lines in the inlet and outlet cross-sections are shown in
Figs. 5–8.
Observe in Fig. 5 that there is no major difference in the description of u′ w ′ Reynolds stress at line AA inlet: indeed this is
not surprising, since the mixed Reynolds stress is well predicted by the three models as a result of the changed inlet length.
Regarding the AA line in the outlet section, the mixed Reynolds stress is not well predicted by the two models: in
particular the SST model predicts a spurious local minimum at Z ∗ ≈ 0.5, whereas the ASST generally provides better results.
Even if in the inlet AA line there is no major difference in the description of the mixed Reynolds stress, this is not the case
in line BB. In fact, the mixed Reynolds stress is quite well predicted by ASST model, whereas the SST fails in providing the
correct value in the fluid bulk. A possible explanation can be found looking at the Reynolds stress formulation for the SST
 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251 2247

u'v' *1000 [m2/s2]

2
Exp. Data (Taylor et al.)
1

0 ASST
-1

-2 SST
-3
0 0.2 0.4 0.6 0.8 1

r* [-]

Fig. 5. 90° bend AA line inlet cross-section u′ w ′ profile.

Fig. 6. 90° bend AA line outlet cross-section u′ w ′ profile.

Fig. 7. 90° bend BB line inlet cross-section u′ w ′ profile.

Fig. 8. 90° bend BB line outlet cross-section u′ w ′ profile.

model, i.e.:
∂u ∂v
 
u′ v ′ = −ντ + . (50)
∂r∗ ∂x
The ∂∂ru∗ gradient can be detected in Fig. 9 in the bulk region, where both experimental and numerical results show a non-
zero slope in the vicinity of r ∗ = 0, 5. Remember that, strictly speaking, the r ∗ = 0.5 point of Fig. 9 and the Z ∗ = 0 of Figs. 7
and 8 are actually coincident.
Hence, it is likely that the non-zero velocity gradient in Fig. 9 is responsible for the non-zero Reynolds stress in Fig. 7.
Obviously, the ASST model can prevent this issue by the anisotropic formulation given by the positive squared term tensor
Ar ∗ x :
∂u ∂v
 
u′ v ′ = −ντ + + Ar ∗ x . (51)
∂r ∗ ∂x
Stated the previous considerations, there is no surprise in noting that the mixed Reynolds stress on outlet BB line is poorly
predicted by SST model: on the other hand see again that the ASST model provides correct trends. In the end we can conclude
that the illustrated 90° bend represents a very hard task to be accomplished for turbulence models. The fact that the flow
entering the channel is not fully developed makes it harder to describe turbulence phenomena which are dominated both
2248 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251

Fig. 9. 90° bend AA line inlet cross-section velocity profile.

Fig. 10. Straight duct of squared cross-section configuration.

Table 5
Straight duct of squared cross-
section grid independence evalu-
ation.
Mesh ID Y+ τw
A 0.81 4.82 mPa
B 0.35 4.73 mPa
C 0.13 4.71 mPa

by the curvature adverse pressure gradient and bend-induced anisotropy. By the way the results of the ASST model are very
valuable, providing a dramatic improvement with respect to the SST model thanks to its anisotropic formulation.

5.3. Turbulence-driven secondary motions in squared cross-section straight duct

Straight channel flow has been already investigated in Section 5.1. However the straight duct of squared cross-section
is a test-case of primary importance due to the onset of corner vortices that are particularly hard to predict. The corner
vortices consist in eight vortices at the four corners of the square cross-section: they are known to be secondary motion of
the second type (turbulence driven), since no pressure gradient imbalance can justify their onset as for the Dean Vortices.
Since isotropic eddy viscosity models are not supposed to predict any corner vortex [16], the interest of the present
investigation is to determine the behavior of the ASST model (which should be able to provide some anisotropy). The results
will be very interesting since the poor behavior of isotropic eddy viscosity models can be easily extended to more complex
flows, where the applicability of such models should be carefully evaluated. Test configuration is shown in Fig. 10: only a
quarter of the total cross-section has been simulated: hence symmetry conditions have been imposed on the two sides to
take into account the rest of the cross-section. Periodic boundary conditions have been imposed on the streamwise faces in
order to consider a fully developed flow. The working fluid is air at atmospheric pressure and temperature.
Three meshes differing by around two in the Y + value have been used to test grid convergence; results of this study are
shown in Table 5.
Note that there is practically no difference between meshes B and C. Therefore mesh C is chosen, since such a finest mesh
did not result in an excessive computational effort.
Results of the analysis are shown in Fig. 11, where the comparison between the vertical velocities V (with respect to the
mean velocity U0 ) is plotted against experimental data of Huser and Biringen [18] (Reτ = 600). The difference between the
SST and the ASST models is dramatic: as expected, the isotropic SST model cannot account for flow anisotropies resulting
in the corner vortices, giving a vertical velocity identically equal to zero. On the other hand, the anisotropic formulation of
the ASST model provides results that approach the experimental data. In particular, ASST model provides a lower maximum
value than experimental data, but still it seems to reproduce the global trend correctly.
 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251 2249

Fig. 11. Vertical secondary velocity distribution in a straight channel of squared cross section.

Fig. 12. Tight lattice rod bundle test configuration [12].

Table 6
Rod bundle grid convergence evaluation.
Mesh ID Y+ τw
A 1.53 0.845 Pa
B 0.92 0.823 Pa
C 0.56 0.819 Pa

5.4. Secondary motion in tight-lattice rod bundles

Fluid flow inside rod bundles is of interest in many engineering applications such as shell side of a shell and tube heat
exchanger or nuclear fuel assemblies. Secondary motion in tight lattice rod bundle is one of the phenomena occurring and
it is responsible for the high transversal mixing and turbulence. The model capability to accurately predict this secondary
motion is of primary importance to apply the model to more complex flows, where secondary motion is of the same order
of magnitude as the primary flow.
The geometry of triangular lattice is shown in Fig. 12. This geometry is composed of six elementary modules (in grey),
which are connected by symmetry conditions. Hence the chosen geometry is the one in dark grey in Fig. 12 to minimize the
geometry extension and maximize the meshing size. The geometry is that of measurements of Mantlik et al. [19] for a pitch
over diameter P/D ratio equal to 1.17 and Re = 109 400.
Three different grids have been used to verify the mesh convergence. They are shown in Table 6 and differ by a factor of
around two in the Y + value.
Again, since no major difference is detected between meshes B and C the final mesh is configuration C, since no greatly
increased computational time is needed to calculate the finest mesh. Mesh C is shown in Fig. 13.
The fact that secondary motion plays a primary role in this flow configuration is clearly visible when looking at the wall
shear stress distribution shown in Fig. 14. In fact, the isotropic SST model predicts a monotonically increase of the wall shear
stress with the angular position, which is in contrast with the experimental measurements. On the other hand the anisotropic
formulation dramatically improves the prediction for the ASST model that correctly reproduces the trend both qualitatively
and quantitatively. Baglietto et al. [12] affirm that the improvement is due to the CNL2 coefficient (see Table 2), whose value
has been increased with regard to the original one proposed by Shih et al. [2]. In this sense, it is worth showing that the shown
wall shear stress is equal to τw = µ ∂∂Wy , where W is the velocity component tangential to the wall and y is the direction
normal to the wall. Therefore it is clear that the anisotropic formulation better captures the secondary flow field, resulting
in the proper velocity scale all along the wall. It has also to be noted that the isotropic formulation of the SST model results
2250 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251

Fig. 13. Tight lattice rod bundle final mesh.

Fig. 14. Tight lattice rod bundle subchannel non-dimensional wall shear stress distribution.

in a quasi-symmetric wall shear stress distribution, since the model cannot capture the correct scale of secondary motion.
In this sense, it has already shown by Baglietto et al. that the original value of the CNL2 coefficient did not result in a major
improvement. Moreover, it could not represent the non-monotonic shape of the wall shear stress distribution, which seems
to be one of the hardest result to obtain. In this sense, the new CNL2 coefficient equal to eleven does produce better results.
Note that the present test case is the principal case used by Baglietto et al. to motivate the increased CNL2 coefficient value.

6. Conclusions and perspectives

The aim of the present work was to present a new eddy-viscosity model, namely the Anisotropic Shear Stress Transport
Model. Its formulation has been presented and motivated based on realizability conditions and correct behavior in the log-
sublayer of a shear flow. The proposed model has been investigated together with its baseline model, i.e. the SST model.
The two models have identical performances for straight channel flow (basically having been developed for such a flow),
differing only in the ability to reproduce turbulence driven anisotropies. This macroscopically results in a good behavior of
the ASST model to reproduce corner vortices in a straight duct with squared cross-section, turbulence characteristics in a
90° bend and the wall shear stress in a rod-bundle flow. In this sense, it seems that the ASST model can provide a valuable
alternative, giving global results that are very promising for engineering application not involving hard system rotation.
Moreover, it keeps a very short computational time, almost identical to that of the SST model.
The new proposed ASST turbulence model couples the linear SST model with a 2nd order non-linear Reynolds stress
formulation. However, as it has been mentioned, a 3rd order formulation would better account for flow swirling and system
rotation. For example, Baglietto et al. [20] proposed a 3rd order formulation, which could be a starting point for ASST model
improving.

Acknowledgment

This work has been done under the CEA CFR founding for the Ph.D of the principal author.
 F. Vitillo et al. / Computers and Mathematics with Applications 70 (2015) 2238–2251 2251

List of acronyms
NLEVM: Non-Linear Eddy Viscosity Model
EARSM: Explicit Algebraic Reynolds Stress Model
SST: Shear Stress Transport model
Ui : Mean velocity along the Xi axis
ASST: Anisotropic Shear Stress Transport model
DNS: Direct Numerical Simulation
RSM: Reynolds Stress Model
Re = Reynolds number based on average velocity
Reτ = Reynolds number based on the friction velocity.

References

[1] C.G. Speziale, On nonlinear k- and k–ε models of turbulence, J. Fluid Mech. 178 (1987) 459–475.
[2] T.H. Shih, J. Zhu, J.L. Lumley, A realizable Reynolds stress algebraic equation model, NASA TM 105993.
[3] T.J. Craft, B.E. Launder, K. Suga, Prediction of turbulent transitional phenomena with a nonlinear eddy-viscosity model, Int. J. Heat Fluid Flow 18 (1997)
15–28.
[4] A. Hellsten, S. Laine, Explicit algebraic Reynolds-stress modelling in decelerating and separating flows, in: AIAA Paper, 2000, pp. 2000–2313.
[5] S. Wallin, A.V. Johansson, An explicit algebraic Reynolds stress model for incompressible and compressible turbulent flows, J. Fluid Mech. 43 (2000)
89–132.
[6] F.R. Menter, A.V. Gabaruk, Y. Egorov, Explicit algebraic Reynolds stress models for anisotropic wall-bounded flows, in: EUCASS—3rd European
Conference for Aero-Space Sciences, Versailles, France, July 6–9 2009.
[7] F.R. Menter, M. Kuntz, R. Langtry, in: K. Hanjalic, Y. Nagano, M. Tummers (Eds.), Turbulence, Heat and Mass Transfer, vol. 4, Begell House, Inc., 2003.
[8] F.R. Menter, Zonal two equation k–ω turbulence models for aerodynamic flows, in: AIAA Paper 93-2906, 1993, pp. 1–21.
[9] F.R. Menter, Two equation eddy-viscosity turbulence models for engineering applications, AIAA J. 32 (8) (1994) 1598–1605.
[10] P.A. Durbin, Limiters and wall treatments in applied turbulence modeling, Fluid Dyn. Res. 41 (2009).
[11] P. Bradshaw, D.H. Ferris, N.P. Atwell, Calculation of boundary layer development using the turbulent energy equation, J. Fluid Mech. 28 (1967) 593–616.
[12] E. Baglietto, H. Ninokata, T. Misawa, CFD and DNS methodologies development for fuel bundle simulations, Nucl. Eng. Des. 236 (2006) 1503–1510.
[13] D.C. Wilcox, Turbulence Modeling for CFD, DCW Industries, Inc., La Canada, 1998.
[14] J. Laufer, The structure of turbulence in fully developed pipe flow, NACA Report 1174.
[15] J. Kim, P. Moin, R. Moser, Turbulence statistics in fully developed channel flow at low Reynolds number, J. Fluid Mech. 56 (1987) 337–351.
[16] Y.G. Lai, R.M.C. So, H.S. Zhang, Turbulence-driven secondary flows in a curved pipe, Theor. Comput. Fluid Dyn. 3 (1991) 163–180.
[17] A.M.K.P. Taylor, J.H. Whitelaw, M. Yianneskis, Measurements of laminar and turbulent flow in a curved duct with thin inlet boundary layer, NASA
Contractor Report 3367, 1981.
[18] A. Huser, S. Biringen, Direct numerical simulation of turbulent flow in a square duct, J. Fluid Mech. 257 (1993) 65–95.
[19] F. Mantlik, J. Heina, J. Chervenka, Results of local measurements of hydraulic characteristic in triangular pin bundle, UJV-3778-R, Rzez, Czech Republic,
1976.
[20] E. Baglietto, H. Ninokata, Anisotropic eddy viscosity modeling for application to industrial engineering internal flows, Int. J. Transp. Phenom. 8 (2006)
109–125.