You are on page 1of 37

Flow Turbulence Combust

DOI 10.1007/s10494-015-9622-4

A One-Equation Local Correlation-Based Transition


Model

Florian R. Menter1 Pavel E. Smirnov1 Tao Liu2


Ravikanth Avancha3

Received: 19 September 2014 / Accepted: 22 May 2015


Springer Science+Business Media Dordrecht 2015

Abstract A model for the prediction of laminar-turbulent transition processes was formu-
lated. It is based on the LCTM (Local Correlation-based Transition Modelling) concept,
where experimental correlations are being integrated into standard convection-diffusion
transport equations using local variables. The starting point for the model was the -Re
model already widely used in aerodynamics and turbomachinery CFD applications. Some
of the deficiencies of the -Re model, like the lack of Galilean invariance were removed.
Furthermore, the Re equation was avoided and the correlations for transition onset predic-
tion have been significantly simplified. The model has been calibrated against a wide range
of Falkner-Skan flows and has been applied to a variety of test cases.

Keywords Laminar-turbulent transition Correlation Local variables

1 Introduction

Modelling of laminar-turbulent boundary layer transition has long been elusive at least in
general-industrial CFD simulations. As a result, most industrial CFD computations have
been carried out fully-turbulent neglecting the effect of the transitional process on the
details of the flow field as well as on the overall performance. There are several reasons

 Pavel E. Smirnov
pavel.smirnov@ansys.com
Florian R. Menter
florian.menter@ansys.com

1 ANSYS Germany GmbH, Staudenfeldweg 20, 83624 Otterfing, Germany


2 General Electric Company, Global Research Center, One Research Circle,
Niskayuna, NY 12309, USA
3 GE Aviation, GE India Technology Centre, Bangalore, 560066, India
Flow Turbulence Combust

for that history. From a modelling standpoint, one of the most difficult aspects is that tran-
sition can take multiple paths (natural, bypass, separation induced, crossflow, ...). It was
therefore unlikely that all these mechanisms (or just several of them) could be modelled in
a physics-based equation framework, especially, when considering the varying importance
and interplay of linear and non-linear effects.
In order of providing a unified concept, which could eventually handle all the differ-
ent mechanisms, the current group proposed a decade ago the so-called Local-Correlation
based Transition Modelling (LCTM) concept ([14]). Within LCTM, no attempt was made
of actually modelling the different transition processes, but to formulate a set of CFD-
compatible transport equations which allow combining experimental correlations in a local
fashion with the underlying turbulence model. The correlations enter the equations mainly
through triggering functions initiating the transition process. In the LCTM formulation
given in Refs. [24] two additional transport equations have been formulated one for the
turbulence intermittency and another one for the transition onset correlation. This specific
realization of the LCTM concept was termed -Re model (or sometimes - model). As
the turbulence intensity, Tu, enters into the experimental correlations, the -Re model is
not Galilean invariant. This means that the model is only applicable to simulations where
the transitional walls are stationary relative to the coordinate system (or additional logic is
required in the code in case of multiple moving walls in a single domain). In most practi-
cal applications, this poses no severe limitation, however, in general-purpose CFD codes,
Galilean invariance is desirable.
Originally, only the conceptual framework of the -Re model was published, without a
full disclosure of the underlying correlations. For that reason, several groups engaged in the
re-construction or alternative formulation of these correlations (e.g. [59]). Eventually, the
original -Re model, including all correlations, was fully disclosed by the original authors
[4] and has since become widely used in numerous industrial applications, ranging from
external aerodynamics to turbomachinery flows, to wind turbine flows all the way to general
industrial applications.
The -Re model has also been modified and extended by various groups. The Re -
equation is required in the original model in order to transport the information of the
freestream turbulence level and the non-dimensional pressure gradient into the boundary
layer, where it is needed to trigger the transitional correlation. In order of avoiding the
need for this auxiliary equation, Coder and Maughmer [10] introduced a shape function-like
parameter, which allowed them to include the effect of the pressure gradient locally inside
the boundary layer. This is an attractive simplification, which significantly reduces the
model complexity. The downside is that, like the original -Re model, the shape-parameter
uses the local velocity relative to the wall and is therefore also not Galilean invariant. In
addition, the effect of the turbulence intensity is not captured automatically by this approach,
restricting the models applicability to external aerodynamic flows with known and essen-
tially constant freestream turbulence levels. This same parameter was recently used in the
development of a transition model by the same authors [11], which approximately tracks
the growth of Tollmien-Schlichting waves in a local transport equation based environment.
Extensions for crossflow instability for the -Re model were proposed by Seyfert and
Krumbein [12], Grabe and Krumbein [13], Medida and Baeder [14] and also in [15]. A fur-
ther extension for the -Re model is the inclusion of wall roughness effects by Dassler et
al. [16]. These extensions show the versatility of the model for inclusion of essentially any
transitional effect for which correlations can be formulated.
Flow Turbulence Combust

Another interesting transition model based on a single intermittency transport equation


using only local variables was recently proposed by Durbin [17] and Ge et al. [18]. The
basic framework of this model is similar to the one introduced by Menter et al. [1], but
more general and with a significantly wider range of calibration. The model of Durbin
[17] and Ge et al. [18] is geared on bypass transition for which the authors observed that
it could be modelled by relying essentially on the diffusion processes of the equations
for turbulence quantities and intermittency, the mechanism being that lower freestream
turbulence levels require longer running lengths than high freestream values to penetrate
into the boundary layer. This mechanism seems to work surprisingly well for numerous
bypass cases. The model does not involve external data correlations but empirical cali-
bration input in the source and sink terms in the intermittency equation, where the sink
term ensures a laminar boundary layer upstream of transition. Reliance on the diffusion
process is likely not to be suitable for natural transition and transition through cross-flow
instabilities, but a modification was introduced to predict transition in separated boundary
layers.
An alternative to LCTM are physics-based RANS model formulations. The most
prominent of these models being the Laminar Kinetic Energy model (k-kl model) by Wal-
ters and Leylek [19] and Walters and Cokljat [20]. These approaches use a model equation
for the laminar fluctuation energy which is then linked to the underlying turbulence model
through source and sink terms. By this mechanism, the laminar kinetic energy can be
transferred into the turbulent kinetic energy equation and initiate the transitional process.
Considering that the pre-transitional phase is very different for different transition scenar-
ios, it is at first surprising that such a formulation would be able to handle in one formulation
e.g. natural and bypass transition. However, the k-kl model uses numerous threshold func-
tions to trigger the transition onset which are similar in nature to the triggering functions
in the -Re model. For example, for the natural transition mechanism the k-kl threshold
function reads: NAT = max (Re CNAT /fNAT , 0), where Re is the vorticity Reynolds
number, CNAT is a constant and fNAT a calibration function. The triggering function in the
-Re model essentially reads Fonset = max (Re /(Re crit 2.2) 1, 0) where Re crit is
an experimental correlation. Therefore, CNAT /fNAT can be interpreted as an equivalent to
the natural transition correlation part in the -Re model. In other words, similar triggering
concepts have been employed independently in different modelling frameworks, making
both model families close cousins. The advantage of the k-kl model over the -Re model
so far has been that it requires only one additional transport equation and that the formula-
tion is Galilean invariant. On the downside, the transition k-kl model is currently combined
with the standard Wilcox model [21] and inherits that models freestream dependency in the
fully turbulent regime [22]. Due to the relatively high model complexity, it is also not eas-
ily fine-tuned and adjusted to other underlying turbulence models (like e.g. one-equation
models), or additional physics.
The current group is convinced that the LCTM concept is still the most flexible method
for including transitional effects into industrial CFD simulations. Due to its generic char-
acter, it allows the inclusion of essentially any transitional effect for which sufficient
experimental information is available. In addition, the intermittency concept is generic with
respect to the underlying turbulence model and can easily be combined with existing and
new model formulations. Finally, the model can be fine-tuned by improved and optimized
transition correlations, without a need for understanding all details of the underlying model
formulation.
Flow Turbulence Combust

The current article is devoted to the formulation of a simplified version of the -Re
model. The goals are to avoid the transport equation for Re by evaluating the correla-
tions inside the boundary layer and not outside. This also eliminates a current deficiency of
the Re model, which is not Galilean invariant due to the use of the freestream turbulence
intensity. The model formulation is therefore only valid for walls which are stationary with
respect to the coordinate system of the simulation (or would require additional software
infrastructure for moving walls). Furthermore, the -Re model features fairly complex
formulations for the experimental correlations, which will be simplified in the current for-
mulation. The goal of the current work is therefore to modernize the -Re model in order
to:

Maintain the LCTM concept, including the ability to model essentially all transitional
processes for which correlations can be formulated.
Reduce the formulation to only one additional equation ( -equation).
Achieve a Galilean invariant formulation.
Simplify the model drastically compared to the original -Re model.
Provide meaningful coefficients which can easily be fine-tuned to match specific
application areas.

Concerning the last bullet point, it needs to be stressed that the current calibration of the
model is covering a wide range of flows and that the model constants should be of sufficient
generality for most applications. Only if consistent deviations of results for a given type
of applications are observed, should the constant and or correlations be fine-tuned. All test
cases in the current work were computed with a fixed set of constants.
The new model is based on a transport equation for the turbulence intermittency, , and
maintains the LCTM concept of explicit correlations for triggering the transition onset. The
resulting model will be termed model. Like the -Re model, it uses transition onset cri-
teria in terms of Re , but computes the latter algebraically using local variables. At that,
it does not require the velocity U and maintains Galilean invariance. The model has
currently been combined with the SST model [23] (as well as other -equation based mod-
els). It can also be used in combination with other models. The main requirement for any
underlying turbulence model is however, that such models require a suitable viscous sub-
layer formulation, which will not interfere with the laminar and transitional flow behavior
(models need to predict an earlier transition location than the transition model).
During the model formulation and calibration, a large number of test cases have been
computed. It was not the goal to capture all cases in perfect agreement with the experimen-
tal data, as this would have lead again to a highly complex set of correlations, similar to
those of the -Re model. The strategy was therefore to calibrate the model carefully for
self-similar flows (Falkner-Skan family) and to extend this calibration with limited com-
plexity into non-equilibrium flows, especially flows with separation. Some differences to
experimental data were accepted as being part of the modelling approach. This seems sensi-
ble, as experimental data are not fully consistent and at times in contradiction with expected
behavior (correlations). The current article discusses the basic calibration and validation test
cases.
The new model was implemented in both ANSYS CFD solvers, ANSYS CFX 15.0 and
ANSYS Fluent 15.0. The simulations in this work have been computed mostly with ANSYS
CFX, but several cross-comparisons with an in-house boundary layer code and with ANSYS
Fluent have been performed to ensure implementation consistency.
Flow Turbulence Combust

2 Model Formulation

2.1 Transport equation for intermittency

The form of the transport equation for the intermittency has formally not been changed
from the one used in the -Re model and reads as:
  
( ) (Uj ) t
+ = P E + + (1)
t xj xj xj
The transition source term has been slightly simplified and is now defined as:
P = Flength S (1 ) Fonset (2)
where S is the strain rate magnitude. As previously, this term is designed to be equal to
zero (due to the Fonset function) in the laminar boundary layer upstream of transition and
is active once the local transition onset criteria is met. The magnitude of this source term
is controlled by the transition length function, Flength , which used to be a correlation, but
has been changed to a constant. Another modification is the absence of the square root from
the Fonset combination. The density and strain rate are present in order to achieve the
correct units for the source term and because the strain rate is the driving force behind the
transitional process.
The destruction/relaminarization source is identical to the one used in the -Re model
and is defined as follows:

E = ca2  Fturb (ce2 1) (3)


where  is the magnitude of the absolute vorticity rate.
The formulation of the function Fonset , which is used to trigger the intermittency produc-
tion (i.e. activate source term (2)) is similar to the one used in the -Re model. It contains
the ratio of the local vorticity Reynolds number ReV (in the current formulation the strain
rate is actually used inside ReV which is equivalent for boundary layers), to the critical
Reynolds number Rec . However, Rec is not computed from a transport equation but alge-
braically, using k and other local variables. As a result, the transition onset is controlled by
the following functions:

Rev
Fonset1 = , Fonset2 = min (Fonset1 , 2.0)
2.2 Rec
   
RT 3
Fonset3 = max 1 , 0 , Fonset = max (Fonset2 Fonset3 , 0) (4)
3.5
 4

RT
k dw2 S
Fturb = e 2
, RT = , Rev = , Rec = f (T uL , L ) (5)

The model constants are:
Flength = 100, ce2 = 50, ca2 = 0.06, = 1.0. (6)
In these equations, dw is the wall distance and is the turbulence frequency (obtained
from the -equation). The boundary condition for at a wall is zero normal flux while
at an inlet the value of is equal to 1. Note that the current equation does not enforce
= 1 through the entire turbulent boundary layer, but has a limit of = 1/ce2 at the wall,
similar to the -Re model. This region is limited to the viscous sub-layer where it has no
Flow Turbulence Combust

noticeable impact on the underlying SST model, as turbulence production in the sub-layer
is essentially zero. This could be alleviated, albeit at the cost of higher complexity.

2.2 Local formulations for T uL and L

The main change from the -Re model is that the arguments T uL and L , entering the
correlation Rec in Eq. 5 are now approximated locally. This removes the need for a second
transport equation for Re and renders the model formulation Galilean invariant. These two
arguments are designed so that Rec has meaningful values in a region between the middle
and the edge of boundary layer, where triggering of the transition model starts. A local
formulation which provides TuL levels in the middle of the boundary layer similar to the
freestream turbulence intensity, Tu, is expressed as:
 
2k/3
T uL = min 100 , 100 (7)
dw
where dw is the wall distance and the combination dw provides a velocity scale inside the
boundary layer, replacing the freestream velocity U used in the -Re model:
U Sdw dw (8)
This formulation of T uL ensures that in the middle of the boundary layer, T uL defined
by Eq. 7 approximately equals the freestream T u values. In turbulence models where is
not available, could suitably be replaced by S/0.3.
The pressure gradient parameter typically used in empirical correlations is defined as
follows:

2 1 dP 2 dU
= = (9)
U ds ds
where dU/ds is the acceleration/deceleration in the streamwise direction at the edge of the
boundary layer. For a flat plate, one can write:

2 dU 2 dV
= = (10)
dx dy
where V and y are the wall-normal velocity and coordinate in the freestream, respectively.
For a general geometry case, dV
dy can be computed as follows:

dV
( n V ) n (11)
dy
Strictly speaking, the relation in Eqs. 9, 10 is only correct outside the boundary layer.
However, the quantity dV /dy can also be used inside the boundary layer as an indicator of
pressure gradients imposed from the freestream. In order of provding a local approximation
of , the momentum thickness is replaced with the wall distance, dw . Note that the
transition model is activated near the center of the boundary layer. At that location one has:
dw = /2 ( being the boundary layer thickness) meaning that dw provides the proper
scaling when replacing . Scaling coefficients for L are selected such that in the middle
of the boundary layer it is approximately equal to for Falkner-Skan profiles. It should
be noted that the -model typically transitions form the middle of the boundary layer (see
Fig. 3). The final formula for L reads as:

dV dw2
L = 7.57 103 + 0.0128 (12)
dy
Flow Turbulence Combust

For numerical robustness L was bounded as follows:

L = min(max( L , 1.0), 1.0) (13)


The constant of 0.0128 accounts for the fact that in the middle of the boundary layer the
velocity gradient dV /dy is not zero even for zero pressure gradient flows, due to the growth
of the boundary layer.
In [18] the point is being made that turbulence models do typically not depend on pres-
sure gradients, which is correct. As the current form of L given in Eq. 12 is based on
the velocity field and not the pressure field it is actually closer to a shape-function than an
explicit pressure gradient.
It should be noted that the current parameter is based on 2-dimensinonal considerations;
however this is true for most parameters used in RANS based transition models. The main
three-dimensional effect in transition models is the inclusion of crossflow instabilities (see
e.g. Menter and Smirnov 2015).

2.3 Correlation

Like in the -Re model, any experimental correlation can be used in combination with
the new model. As it was felt that the correlations in the -Re model [4] were overly
complex, a simpler formulation was developed for the current model:

Rec (T uL , L ) = CT U 1 + CT U 2 exp [CT U 3 T uL FP G ( L )] (14)

CT U 1 = 100.0, CT U 2 = 1000.0, CT U 3 = 1.0 (15)


For zero pressure gradient flows, the form of the correlation (14) for Rec is similar to
the one proposed by Abu-Ghannam and Shaw (AGS) [24]. The constants (15) are model
parameters and can be changed by the user. The CT U 1 constant defines the minimal value of
the critical Rec number (for very high T uL levels the exponential approaches zero) which
can be utilized by the model. The sum of CT U 1 + CT U 2 defines its maximal value (for
very low T uL levels the exponential approaches one), while CT U 3 controls how fast Rec
decreases as the turbulence intensity T u increases.
The function FP G ( L ) is introduced to sensitize the transition onset to the streamwise
pressure gradient. This function is purely empirical and is calibrated using the Falkner-
Skan series of profiles as described in Section 3.3. From Eq. 14 it becomes clear that Rec
is independent on L if T uL is zero. This is different from the AGS [24] and also the
Langtry [4] correlations where Rec depends on for low and zero T u conditions and
where this dependency increases as the T u decreases. However, for the new -transition
model the expression (14) is not purely the correlation that fits the experimental data. It also
accounts for the fact that the transition triggering function is based on ReV rather than on
the momentum thickness Reynolds number Re computed from the velocity profile. The
relative difference between Re and ReV as a function of shape factor H is shown in Fig. 1.
The figure shows also the correction coefficient to the 2.2 factor as the function in the
corresponding range of . It is seen that the ratio can change by as much as a factor 2. This
strong effect of the pressure gradient can be accounted for either directly as a correction
function for the 2.2 factor in Eq. 4 or implicitly through correlations like equation (14). In
the transition model, it was decided in favor of the latter approach.
Figure 2 shows that the model with the constant Rec for zero T uL reasonably repro-
duces the data from AGS correlations in the region of adverse pressure gradients up to the
Flow Turbulence Combust

Fig. 1 Relation between the maximum value of the ratio of the vorticity Reynolds number ReV and the
momentum thickness Reynolds number Re and (left) the boundary layer shape factor H and (right) the
pressure gradient parameter

separation point ( = 0.0681), where transition occurs in most of cases. In other words,
the correlations specified are not formulated with respect to Re , but with respect to ReV .
The influence of the pressure gradient is indirect through the change of the shape of the lam-
inar velocity profile. This effect is most likely better represented by the local value ReV max
than by the integral value of Re , resulting in a simpler correlation when using the former.
For non-zero T uL , corrections to Rec using L appeared to be necessary, as will be
shown in Section 3.3. As the result, the formula for FP G entering into the correlation of
Eq. 14 reads as:



min 1 + CP G1 L , CPlimG1 , L 0
FP G ( L ) =
min 1 + CP G2 L + CP G3 min[ L + 0.0681, 0], CP G2 , L < 0
lim
(16)

CP G1 = 14.68, CP G2 = 7.34, CP G3 = 0.0 (17)


CPlimG1 = 1.5, CPlimG2 = 3.0 (18)

Fig. 2 Values of Re at
transition point as predicted by
Abu-Ghannam and Shaw
correlations and the transition
model for zero turbulence
intensity case and adverse
pressure gradients
Flow Turbulence Combust

A limiter is then applied to FP G in order to avoid negative values:

FP G = max (FP G , 0) (19)


The constants in Eq. 17 form another set of model parameters accessible by the user. The
CP G1 constant controls the value of the critical Rec number in areas with favorable pressure
gradient, CP G2 with adverse pressure gradient and CP G3 becomes active in regions with
separation, allowing correcting the Rec value there if necessary. The constants in Eq. 18
define the limits of the FP G function and should typically not be adjusted.

2.4 Coupling with the SST turbulence model

The coupling between the transition model and the SST turbulence model is the same as
for the -Re model and is accomplished by modifying the equations of the original SST
model as follows:


t (k) + xj (uj k) = Pk + Pk Dk + xj ( + k t ) x
lim k
 j (20)
Pk

t () +
xj (u j ) = t D + Cd +
xj ( + t )
xj

Pk = Pk (21)

D k = max( , 0.1) Dk (22)

a1 k
t = (23)
max (a1 , F2 S)
where Pk and Dk are the production and destruction terms from the turbulent kinetic energy
equation in the original SST turbulence model (for details and constants see [23]). ThePk
term is computed using Kato-Launder formulation [25]:

Pk = t S  (24)
where  denotes the magnitude of the absolute vorticity rate. This formulation helps avoid-
ing excessive levels of the turbulence intensity in stagnation regions. It is known that the
Kato-Launder formulation affects the fully turbulent model behavior (e.g. for swirling flows
etc.). An alternative to Kato-Launder would be a tight production, or realizability limiter
(note that the production limiter [23] is de-activated when the Kato-Launder formulation is
used).
An additional production term Pklim has been introduced into the k-equation to ensure
proper generation of k at transition points for arbitrary low (down to zero) Tu levels. The
need for such a term is based on the observation that for very low values of freestream turbu-
lence, the underlying SST model itself requires a relatively long running length to produce
turbulence inside the boundary layer, even if triggered properly by the transition model. The
additional term is designed to turn itself off when the transition process is completed and the
boundary layer has reached the fully turbulent state. The expression for the Pklim reads as:

Pklim = 5Ck max( 0.2, 0) (1 ) Fon


lim
max(3CSEP t , 0)S  (25)

ReV
lim
Fon = min(max( 1, 0), 3) (26)
2.2 Rec
lim

lim
Rec = 1100 (27)
Flow Turbulence Combust

Ck = 1.0, CSEP = 1.0 (28)


The max( 0.2,0) part ensures that the term is only activated once the tran-
sition model is triggered and has increased beyond a value of 0.2. The part
max( 3CSEP t , 0) switches the additional source term off in fully turbulent regions
where t >3CSEP . The relatively high limiting value for Rec lim = 1100 ensures that

the term is only activated for high Reynolds numbers and/or separating flows, where ReV
becomes larger than this limit.
This additional source term helps to make the transition process more reliable when it
is developing under low Tu conditions and/or when it starts in a laminar separation bubble.
This is illustrated in Fig. 3 and Fig. 4, respectively. If the freestream Tu is low and the addi-
tional source term is not used, then the locations where intermittency and eddy-to-molecular
viscosity ratio (RT ) start to grow are quite distant from each other see the left column
of Fig. 3. This disconnect significantly complicates the calibration of the model. The use
of the source term makes the transition process more compact (right column of Fig. 3) and
thus more in line with the underlying correlation. A similar effect takes place in separation
bubbles at low freestream Tu levels. Figure 4 shows that without the source term, the eddy
viscosity is not generated at the upper part of the bubble where intermittency is close to
unity (non-zero RT values closer to the wall are convected from a downstream region by the
backflow). Therefore, the solution is insensitive to the location where intermittency starts to
grow, since the eddy viscosity felt by the momentum equations is not produced accordingly.
Activation of the source term forces the RT generation at the expected area, which sensitize

Fig. 3 Effect of an additional production term Plim


k on the Schubauer-Klebanoff test case, T u = 0.03%
Flow Turbulence Combust

Fig. 4 Effect of additional production term Plimk on separation bubble developing on a flat plate under
lowpressure turbine airfoil (Pak-B) conditions, Re = 50 000, T u = 0.2%

the solution to the transition onset. It also allows controlling the bubble size by changing the
constant CSEP in front of the molecular viscosity. The default value of the constant is set to
1.0, a higher value will result in a shorter bubble. This mechanism is different from the one
used for the -Re model, where the effective intermittency eff was allowed to exceed 1.0
whenever the laminar boundary layer separates. Nevertheless, the result is similar to using
eff and allows improved predictions for separation induced transition as well as for flows
at low Tu levels, as will be shown later.
The modification of the blending function F1 responsible in the SST model for switching
between the k and k models is inherited from the Re model. The F1 is formulated
such that it will always be equal to 1.0 in a laminar boundary layer, which ensures that
k model is consistently active in all near-wall regions, including laminar, transitional
and turbulent boundary layers. The modified blending function is defined as follows:


Ry 8

y k
Ry = , F3 = e 120
, F1 = max(F1orig , F3 ) (29)

where F1orig is the original blending function from the SST turbulence model.

3 Model Calibration

3.1 Calibration technique

Empirical correlations like the ones shown in Fig. 5 served as a basis for the initial model
calibration. The idea is to tune the new model such that for each pair of Tu and it transi-
tions at a momentum thickness Reynolds number Re close to that from the correlations. In
particular, in case of zero steamwise pressure gradients the model should be able to match
the Re dependency versus Tu shown on the right chart of Fig. 5.
Flow Turbulence Combust

Fig. 5 Correlations for transition momentum thickness Reynolds number: (left) the correlation of Abu-
Ghannam and Shaw; (right) Mayle correlation, reproduced from [26]

For calibration, a very large number of cases have been run with different Tu and
parameters using an in-house boundary layer code. The applied technique resulted in the
model constants as given in Section 2. Due to the low cost of the boundary layer simulations,
all cases were run on excessively fine meshes ensuring grid independence. In addition,
consistency checks were performed between the boundary layer and the full Navier-Stokes
codes.

3.2 Results for zero pressure gradient

To calibrate the model for cases with zero streamwise pressure gradient ( = 0), 18 com-
binations of inlet turbulence intensity and eddy-to-molecular viscosity ratio were selected
for the test matrix. Note that the inlet values specified for eddy-to-molecular viscosity ratio
mainly determine the rate at which the turbulence kinetic energy decays, whereas the val-
ues for Tu determine the transition location. For the final set of the model constants, results
are shown in Fig. 6. In this figure, data of the -transition model for each parameter combi-
nation (black symbols) are compared with three different versions of empirical correlation,

Fig. 6 Results of model


calibration on boundary layer
flow developing on flat plate and
under zero streamwise pressure
gradient
Flow Turbulence Combust

Table 1 Inlet condition for the flat plate test cases

Case U [m/s] Tu (%) t /

T3A 5.18 3.3 12


T3B 9.4 6.2 90
T3A- 19.8 0.9 8
Schubauer-Klebanof 50.1 0.03 1
T3C2 5.29 3.0 8
T3C3 4.0 3.0 5
T3C4 1.37 3.0 2
T3C5 9.0 3.0 15

namely, AGS [24], Mayle [26] and the correlation from Langtry and Menter [4]. Transition
for the model location was defined based on minimal wall shear stress.
One can see that the model fits the correlations quite well. Three of the selected combi-
nations of inlet parameters correspond to T3 cases with zero pressure gradient, namely to
T3A-, T3A and T3B as given by Table 1. Results for these three cases have corresponding
labels in Fig. 6.

3.3 Results for non-zero pressure gradient

When the distribution of the external boundary layer edge velocity Ue (x) obeys a certain
power law of x, solutions of the laminar boundary layer equations become self-similar with
a parameter corresponding to a constant . The solution of these self-similar equations is
known as Falkner-Skan series of profiles (see, for example, [27]). A profile from this series
is defined by the parameter which enters the formula for Ue (x) in the following way:

Ue (x) = U0 (x x0 ) 2 (30)
Figure 7 shows the dependency of the shape of the Falkner-Skan profile on the param-
eters , as well as the connection between and . For zero , is also zero and this
case corresponds to the Blasius profile. Positive correspond to favorable pressure gradi-
ents characterized by positive , while negative stand for adverse pressure gradients. The
value of = 0.19884 corresponds to = 0.0681 at separation onset.

Fig. 7 Falkner-Skan series of profiles, dependency on parameter: (left) velocity; (right) pressure gradient
parameter
Flow Turbulence Combust

Fig. 8 Critical momentum thickness Reynolds number predicted by the current model for equilibrium
boundary layer flows developing on a flat plate under various adverse and favorable streamwise pressure
gradient conditions (Falkner-Skan equilibrium conditions). Comparison with established correlations

Figure 8 shows the results of the transition model for a series of values which cor-
responds to values as indicated on each plot. These simulations have been used as a
basis for the calibration of the coefficients CP G1 and CP G2 in equation (17). In general, the
selected set of constants provides a good agreement with the empirical correlations.
For strong favorable pressure gradients ( = 0.0614), the model follows the lower cor-
relation bound given by the Mayle criteria, but this was calibrated deliberately for the sake
of a compromise for the T3C5 test case (Section 4). Matching the experimental data of this
test case requires somewhat earlier triggering of transition in areas with favorable pressure
gradient.

3.4 Relaminarization under strong favourable pressure gradient

The last test case in this chapter is designed to investigate the ability of the transition model
to predict re-laminarization. Re-laminarization should occur in a strong favorable pressure
gradient when the acceleration parameter

dU
K= (31)
U 2 dx

exceeds 3106 (see e.g. [26]). Conversely, transition should never occur until the accelera-
tion parameter drops below 3106 . A hypothetical test case has been proposed in order to
test the ability of the new transition model to predict relaminarization. The test case is iden-
tical to T3A up to the streamwise coordinate x = 0.7 [m], at which point a strong favorable
pressure gradient is imposed in the boundary layer code. Figure 9 shows the distribution of
the parameter K using U (x) = tanh(7x 9) + 2 as boundary layer edge velocity.
The skin friction calculated by the boundary layer code for the laminar, transitional and
turbulent cases are shown also in Fig. 9. It is seen that the model does in fact predict re-
laminarization once the acceleration parameter exceeds 3106 . It also predicts re-transition
downstream once the strong favorable pressure gradient has subsided.
Flow Turbulence Combust

Fig. 9 Relaminarization test case: (left) acceleration parameter K and (right) skin friction coefficient

4 Validation for Flat Plate Test Cases

4.1 Introduction

The -transition model detailed in the previous section has been applied to predict the
ERCOFTAC (European Reseach Community on Flow, Turbulence and Combustion) T3
series of experimental flat plate test cases. The development of boundary layers on flat plates
with zero and non-zero pressure gradients has been studied in much detail by researchers
at Rolls Royce in the early 1990s [28]. The wind tunnel geometry consisted of a flat plate
with a very small rounded leading edge mounted in a test section. The radius of the rounded
leading edge was 0.75 mm and the flat plate had a length of 1.5 m. The T3 series of test
cases (T3A, T3B, T3A-, T3C2, T3C3, T3C4 and T3C5) have often been used to evaluate the
ability of transition models to predict the onset and length of transition. All the T3 series of
test cases had freestream turbulence intensities of about 1% or larger, and as a result bypass
transition was the dominant transition mode. In order to test the new transition models abil-
ity to predict natural transition, the Schubauer and Klebanoff [29] flat plate experiment has
also been computed.
The T3 measurements were performed on a flat plate subjected to different levels of
freestream turbulence and pressure gradients. Test cases T3A, T3B and T3A- had a zero
streamwise pressure gradient with freestream turbulence intensities of 3%, 6% and 0.9%
respectively. The different freestream turbulence levels where imposed by inserting different
turbulence grids into the wind tunnel test section upstream of the flat plate. The Schubauer
and Klebanoff zero pressure gradient flat plate experiment was performed in a relatively
quiet wind tunnel and had a freestream turbulence intensity of only about 0.03%.
The T3C test cases had a streamwise pressure gradient that is representative of an aft-
loaded turbine blade and freestream turbulence levels of 3.0%. The wind tunnel geometry
was identical to the T3A, T3B and T3A- cases except the tunnel wall opposite of the flat
plate was contoured in order to impose a streamwise pressure gradient on the flat plate and
thus alter the development of the flat plate boundary layer. The main variation for the var-
ious T3C test cases was that the wind tunnel velocity U0 (and hence the tunnel Reynolds
number) was progressively lowered so that the transition onset location moved from the
favorable pressure gradient region at the leading edge to the adverse pressure gradient at
the end of the plate. The T3C test cases have become very popular for testing the combined
effect of pressure gradient and boundary layer history on the predicted onset and length of
Flow Turbulence Combust

transition. However, simulations done in the present work revealed some unusual character-
istics of these test cases which are not entirely understood. Among them is the deviation of
the skin friction coefficient measured upstream of transition from the laminar solution, as
well as some inconsistencies of the measured transition location with established empirical
correlations, e.g. of AGS. The latter aspect is discussed in Section 4.3.
In many industrial transition predictions, it was found that transition is triggered by a
laminar separation bubble, followed by transition in the bubble and turbulent reattachment.
For this reason, a family of flat plate test cases with laminar separation bubbles (Hultgren
and Volino[30]) was included in the current study of basic cases.

4.2 Numerical setup T3 series

All the results shown in the present chapter have been computed with the ANSYS CFX
15.0 solver. A mesh refinement study was employed to all cases to ensure mesh indepen-
dence. For details on meshing requirements for the model, see also Appendix A. For
the T3A, T3B and T3A- cases, the solutions shown are for a mesh with 231148 nodes
in x and y directions, respectively (mesh refinement studies not shown here proved mesh-
independence). The distance to the first near-wall node was 2105 [m], which provided
average y + values below 1 for all the three cases. The expansion ratio for the grid lines in
the wall normal-direction was 1.05. Also, the grid was refined towards the inlet section for
better resolution of the region where the laminar boundary layer starts to develop; the first
grid step in x-direction was 103 [m] and the expansion ratio was again 1.05. The compu-
tational domain for the Schubauer-Klebanoff test case had the same dimensions, but due to
a higher Reynolds number, the first near-wall node was placed 1105 [m] away from the
wall in order to maintain y + values below 1.0, and the number of nodes was increased up to
245162 to keep the same grid expansion ratio and also the grid aspect ratio near the inlet
plane as for the T3 cases.
For all cases with pressure gradient, the upper boundary contour of the computational
domain was computed from the velocity profiles measured at streamwise stations distributed
over the plate length. Figure 10 shows the computed streamline points for each case from
the T3C series. The polynomial fits of discrete points were built in order to provide a smooth
curve for the upper boundary of the computational domain. These curves are also shown in
Fig. 10. For all boundary layer simulations, the upper wall was specified as a free slip wall.
For all four pressure gradient cases: T3C2, T3C3, T3C4 and T3C5 the computational
grids had the same dimensions with 257129 nodes x and y directions, respectively. The
distance to the first near-wall node was identical for all meshes and equal to 2105 [m],
which provided average y + values below 1 for all the cases. Expansion ratios and refinement
near the inlet section were the same as for the zero pressure gradient cases.

Fig. 10 Streamline coordinates


used as upper boundaries of
computational domains for T3C
cases. Symbols points
computed at measurement
stations, curves polynomial fits
Flow Turbulence Combust

The working fluid in the experiment was air at atmospheric pressure and room tempera-
ture. Although the measured air properties were somewhat different between measurement
stations and the test cases, these variations were totally less than 4%. Hence, con-
stant air properties were assumed for all CFD simulations with the following values:
= 1.2 [kg/m3 ], = 1.8 105 [kg/m/s]. At the inlet, constant values of streamwise velocity
and turbulence characteristics were specified in accordance with Fig. 11. The inlet turbulent
quantities were specified in order to match the measured decay of freestream turbulence.

4.3 Results T3 series

Figure 12 shows the predicted decay of freestream turbulence in comparison with the exper-
iment. It is seen that the values of inlet turbulence characteristics given in Fig. 11 provide
the correct distribution of Tu.
Figure 13 presents a comparison of the measured and computed skin friction coefficients
for all T3 cases and the Schubauer-Klebanoff case. The results of all test cases with zero
pressure gradient (T3B, T3A, T3A- and the Schubauer-Klebanoff) are summarized in the
upper row of Fig. 13. In general, the agreement of the new model with the experimentally
measured skin friction and transition locations is good. The model was able to predict tran-
sition from a Reynolds number (based on plate length and inlet velocity) of 50 000 all the
way out to 3106 as the freestream turbulence was lowered. The results of the new model
are compared against those of the -Re model. The new model gives a steeper transition,
which is especially noticeable for the T3A- case.
Results for the cases with streamwise pressure gradient are summarized in the lower row
of Fig. 13. For the test case with favorable pressure gradient (T3C5), the new model results
in good predictions. The -Re model gives too late transition, which is apparently due
to the fact that the model was tuned on higher inlet turbulence intensity for this test case.
Namely, T u = 4% was used for the T3C5 cases, while the experimental value was 3% as
for all other T3C cases [38]. Naturally, the model gives later transition when Tu is reduced

Fig. 11 Freestream turbulence intensity versus distance from the leading edge of the flat plate for the T3 test
cases
Flow Turbulence Combust

Fig. 12 Skin friction coefficient for flat plate test cases: (upper row) zero pressure gradient cases; (lower
low) cases with streamwise pressure gradient

back to 3%. In the present work, this decay was matched with the original Tu value from
the experiment equal to 3% (see Fig. 12).
For the case with transition near the suction peak (T3C2) the new model delays the
transitional process and in the adverse pressure gradient zone (T3C3) it results in lami-
nar separation, which is not the case in the experiment. The reason for this behavior was
investigated by comparing the exact momentum thickness Reynolds number from the lami-
nar boundary layer code with the critical Reynolds number from the AGS correlation (note
that the transition model reproduces the AGS correlation fairly closely in the Falkner-
Skan simulations). Fig. 14 shows the AGS correlation for the T3C2 and the T3C3 test cases
against the actual momentum thickness of the boundary layer. For the T3C2 case, transi-
tion starts in the experiment at around x = 0.8-0.9. At that location, the AGS correlation is
much higher (Rec 600) than the actual Re of the boundary layer (Re 390). It is not
clear, why the experiment starts to transition at such a low Re for this case. The intersection
between the AGS correlation and the boundary layer Re occurs at around x = 1.05. The
model does actually start to produce intermittency at that location, but then the remaining
running-length to separation is not sufficient to complete the transitional process, so that the
model transitions only after separating.
The situation is similar for the T3C3 case, where the AGS correlation and the actual Re
do not or only barely intersecting at x = 1.2 and then diverge again from each other.
The -Re model overcomes this problem by using modified correlations specially tuned
for the T3C cases together with additional empirical calibration of the Re equation aimed
at reducing the original freestream Re . This is one of the reasons for the complexity of
the -Re model formulation and the large number of constants involved. The goal of the
current formulation is to avoid this level of complexity in favor of a simpler formulation
which can be more easily calibrated by the user. The quality of predictions of the T3C2 and
T3C3 cases could be improved simply by decreasing the constant CP G2 responsible for the
sensitivity to adverse pressure gradient in equation (17). However, such a change breaks the
calibration for Falkner-Skan profiles. Application of the transition model to 2D blade test
cases revealed that the calibration based on Falkner-Skan profiles is preferable over the one
Flow Turbulence Combust

Fig. 13 Distributions of skin friction coefficient, momentum Reynolds number and critical Reynolds number
along flat plate for (left) T3C2 and (right) T3C3 test cases

tuned specially for T3C. Therefore, the Falkner-Skan based calibration is used for all further
studies.
It should also be noted for completeness that there are significant differences between
the simulations and the experiments in the wall shear stress for the T3C cases already in
the laminar region. It was not possible to trace the source of this discrepancy. Even fully
laminar simulations without any turbulence maintained a significantly higher laminar wall
shear stress than observed in the experiments. It is a discrepancy which puts a question mark
on the data. It was attempted to modify the Re number in order to match the laminar wall

Fig. 14 Computational mesh for Hultgren and Volino [30] test case, Re=50 000, Tu =0.2%
Flow Turbulence Combust

shear stress. The required Re number increase for the T3C2 and T3C3 cases was about 30%.
This lead to a significantly improved transition location, but the change is clearly outside
any conceivable measurement uncertainty. Similar discrepancies between model predictions
and data have been observed by Ge at al. [18].

4.4 Numerical setup Hultgren and Volino series

In the experiments of Hultgren and Volino [30], a flat plate boundary layer was subjected to
a streamwise pressure gradient. The superimposed pressure gradients were produced such
that laminar separation of the boundary layer occurred for most of the studied cases, with
subsequent turbulent flow reattachment. The resulting pressure distribution represents that
on the suction side of a highly loaded turbine blade. The Reynolds number varied between
50 000 and 300 000, and the freestream turbulence intensity was either 0.2% or 7%.
The present simulations have been carried out for two different Reynolds numbers, Re
= 50,000 and Re=300,000 (based on nominal exit velocity, Uref , and the wetted plate
length, L = 0.208 [m]) and two freestream turbulence intensities, Tu =0.2% and Tu =7%.
The inlet section of the computational domain was created at the same location as the first
measurement station in the experiment and the measured velocity profile was specified
as inlet boundary condition for the simulations. As stated in the experiments, freestream
Tu levels of 0.2% and 7% correspond to about 0.2% and 2.5%, respectively, in the test
section, so the latter pair of values was used to compute turbulence characteristics at the
computational domain inlet. At the flat plate, a no-slip wall condition was assumed. The
computational hexahedral meshes provided y + values smaller than 1 and had 31270 nodes
for Re = 50 000 and 31289 for Re = 300 000. The mesh for Re = 50 000 and Tu=0.2%
is illustrated in Fig. 14.
The upper boundary of the computational domain was a prescribed streamline taken
from the work of Suzen et al [31]. The authors obtained the external streamline shapes
by integrating the measured velocity profiles for each combination of Re and Tu. In the
present work, streamline coordinates were digitized from a plot in Suzen et al [31] and a
free-slip boundary condition was applied at the upper surface. It should be noted that the
integration of the outer free slip boundary from the experimental velocity profiles suffers
from the experimental uncertainty that no negative velocity components could be measured
in the experiment. For the cases with large separation bubbles, this introduces a significant
uncertainty into the CFD set-up. Still, the case is attractive as it offers a variation of both,
Tu and Re number.

4.5 Results Hultgren and Volino series

Figure 15 shows streamlines for all computed cases as predicted by the transition model.
The size of the separation bubble is decreasing as the Reynolds number or turbulence inten-
sity increase. For Re = 300 000 and Tu = 7% no bubble is predicted and transition is
triggered via the bypass mechanism at x/Lof about 0.65. Distributions of the skin friction
coefficient computed from the and -Re models are given in Fig. 16 in comparison with
the experimental data. In addition, the estimated reattachment point from Suzen et al [31]
is added to the reference data for Re = 50 000 and Tu = 0.2%, since for this case the last
measurement station in the experiments was still located in the recirculation zone.
Overall, the simulations capture the effect of bubble size variations under changing flow
conditions reasonably well. For Re = 50 000 and Tu = 7% the model delays transition rel-
ative to the experiment similar to the -Re model. In contrast, the model gives somewhat
Flow Turbulence Combust

Fig. 15 Effects of Reynolds number and freestream turbulence intensity on streamlines computed for
Hultgren and Volino [30] test case

earlier reattachment for Re = 300 000 and Tu = 0.2% which is closer to the experiment.
Due to the uncertainty in the specification of the external streamline, no model optimiza-
tion was performed based on the current data, also in light of fairly consistent predictions
of bubble sizes for real turbine blade geometries shown in the next sections.

Fig. 16 Distributions of skin friction coefficient for Hultgren and Volino [30] test case
Flow Turbulence Combust

Fig. 17 Computational mesh for


NACA 0021 airfoil

5 Application Test Cases

5.1 NACA 0021 airfoil

The NACA 0021 airfoil is a 21% thick symmetrical airfoil representative for horizontal-axis
wind turbine (HAWT) applications [32]. The experimental results were obtained in the wind
tunnel at the Monash University. The airfoil profile can be seen from the computational
mesh used for the present calculations (Fig. 17). The grid was of a C-type with the total
number of hexahedral elements 118 000.
In this study, the experimental results are compared to the numerical results obtained
using the transition model, the -Re model and the SST (fully turbulent) model. All
calculations were performed assuming incompressible flow and a Reynolds number of
2.7105 . The inlet turbulence level was selected so that the freestream turbulence level
around the airfoil was approximately 0.6% which corresponds to the turbulence level in the
experiments. The simulations were carried out for a range of angles of attack from 0up to
28with the increasing step of 1. This range of angles covers the so-called deep stall regime
when the flow separates from the leading edge of the airfoil. The predicted CL distribution
as a function of angle of attack is shown in Fig. 18.
For angles lower than the critical one (deep stall angle, 18in the experiments), transi-
tional simulations predict a separation bubble close to the leading edge which alters the
pressure distribution in this region and gives higher CL values compared to the fully tur-
bulent simulations. The positive result of capturing this effect is especially visible for the
medium angles between 5and 15. Airfoils stall occurs when the trailing edge turbulent

Fig. 18 Lift coefficient for


NACA 0021 airfoil
Flow Turbulence Combust

Fig. 19 Blade loading for the Pak-B Low-Pressure turbine at various freestream turbulence intensities (T u)
and Reynolds numbers (Re)

separation meets the back-end of the laminar separation bubble. This critical angle is overes-
timated by the -Re model, while the result of the model is quite close to the experiment,
due to a slightly larger laminar bubble predicted by the model.

5.2 Pratt and Whitney Pak-B low pressure turbine cascade

The Pratt and Whitney Pak-B low pressure turbine blade is a particularly interesting airfoil
because it has a loading profile similar to the rotors found in many modern aircraft engines

Fig. 20 Computational domain and mesh fragment for the T106 cascade
Flow Turbulence Combust

Fig. 21 Contours of turbulent kinetic energy at various wake passing periods

[33]. The low-pressure rotors on modern aircraft engines possess extremely challenging
flow fields. This is because in many cases the transition occurs in the free shear layer of
a separation bubble on the suction side [26]. The onset of transition in the free shear layer
determines whether or not the separation bubble will reattach as a turbulent boundary layer.
Huang et al. [34] conducted experiments on the Pak-B blade cascade for a range of
Reynolds numbers and turbulence intensities. The experiments were performed at the design
incidence angle for Reynolds numbers of 50 000, 75 000, and 100 000 based on inlet veloc-
ity and axial chord length, with turbulence intensities of 0.08%, 2.35% and 6.0% (which
corresponded to values of 0.08%, 1.6%, and 2.85% at the leading edge of the blade). In the
experiments of Lake et al. [35, 36] the Pak-B cascade was investigated for Reynolds num-
bers of 43 000, 86 000, and 172 000 and freestream turbulence intensities of 1.5% and 10%
(1.0% and 4.0% at the leading edge). Some of these cases were selected for the simulations
in the present work. The computational grid consisted of 143 336 hexahedral elements.
Figure 19 shows the results of the simulations. In this figure, the computed pressure
coefficient distributions for various Reynolds numbers and freestream turbulence intensi-
ties are compared to experimental data. The most important feature of this test case is
the extent of the separation bubble on the suction side, characterized by the plateau in
the pressure distribution. The size of the separation bubble is actually a complex function
of the Reynolds number and the freestream turbulence value. As the Reynolds number or
freestream turbulence decrease, the size of the separation and hence the pressure plateau
increases.
The computations with the new transition model compare quite well with the experi-
mental data for all of the cases considered, illustrating the ability of the model to capture
the effects of Reynolds number and turbulence intensity variations on the size of a lami-
nar separation bubble and the subsequent turbulent reattachment. It should be stressed, that
the fully turbulent simulations completely miss this phenomena since the boundary layer
remains attached over the entire length of the suction surface.

5.3 Unsteady wake/blade interaction on T106 blade cascade

The turbine blade cascade T106 has often been used for experimental studies of the effects
of periodically passing wakes on laminar-to-turbulent transition and separation in low-
pressure turbines (see, for example, [37]). The cascade was also studied numerically by
Wu and Durbin [38] by means of a DNS model where the wakes were generated by imag-
inary circular cylinders moving upstream of the blade. Conditions of the DNS were used
Flow Turbulence Combust

Fig. 22 Distributions of (left) skin friction coefficient and (right) intermittency for the T106 cascade
obtained using unsteady simulations with the wake from cylinder

in the present work to set-up the simulations and results from DNS were adopted for the
cross-comparisons.
Incompressible flow through a turbine stator passage was computed for a Reynolds num-
ber of Re = 1.48105 based on the blade axial chord and upstream flow velocity. The
computational domain and mesh are shown in Fig. 20. With the blade axial chord Lax =
1.0 [m], the blade pitch is 0.93 [m] and true chord length is 1.1639 [m], so the ratio is 0.799
the design value for T106. Inlet and outlet planes are at x = 0.33 [m] and x = 2.0
[m], respectively. The mesh comprises 195 023 hexahedral elements and has fully match-
ing nodes on periodic boundaries. The element size in the freestream was 0.003 [m], which
allows several cells to resolve the wake from the cylinder being 0.01177 [m] in diameter.
For modelling the cylinder wake, steady inlet profiles were moved across the inlet plane.
The profiles for velocities and turbulent characteristics were extracted from a precursor
RANS simulation of flow past the circular cylinder. In order to obtain a steady solution,
symmetry conditions were assumed and only one half of the cylinder was computed and
then mirrored. Extracted velocity profiles were aligned with the main flow direction being
inclined to the x-axis with an angle of = 37.7and the wake was moved along the inlet
plane at Ucyl = 1.2048 [m/s] to generate an unsteady wake at the blade. For the given pitch
and Ucyl , passing period T = 0.772 [s]. The bar to cascade pitch ratio is 1:1, so inlet profile

Fig. 23 Pressure coefficient obtained using (left) DNS [38] and (right) present unsteady simulations for the
T106 blade cascade with the wake from a cylinder
Flow Turbulence Combust

Fig. 24 Computational mesh for


V103 compressor blade

transformation is not needed. The background turbulence level was about 1% at the inlet
plane.
The computed turbulent kinetic energy contours at various time levels are shown in
Fig. 21. The incoming wake interacts with the trailing-edge vortex shedding, resulting in
distributions of skin friction and intermittency as shown in Fig. 22. In the aft part of the
suction side, evolution of the skin friction clearly shows periodic regions of earlier transi-
tion caused by the passing wake. The distribution of the intermittency responds dynamically
to the changing flow topology, showing the ability of the model to re-laminarize and to
re-transition within one wake period. Figure 23 shows the cross-comparisons of the Cp
distributions from the present simulations with the DNS of Wu and Durbin [38].
Distributions of the pressure coefficient near the separation bubble are somewhat
different between the DNS and the present unsteady RANS simulations with the new one-
equation transition model, but the trend is the same: the unsteady interaction of the blade
with the wake from the cylinder suppresses the separation bubble on the suction side. This
is the most important outcome of the study, because capturing the separation bubble sup-
pression plays a significant role in the accurate prediction of losses for low pressure turbine
blades.

5.4 V103 compressor cascade

The flow field around this compressor blade is simulated at four different free-stream tur-
bulence intensities varying from 0 to 8% and the results are compared with the DNS data of
cases T0-T3 in [39]. The coordinates of the blade are based on NACA 65 profile [39]. The
simulation Reynolds number is Re =138 500 and the inflow angle is 42. The grid used for
the CFD computations consisted of 58 166 hexahedral elements and is shown in Fig. 24.
The inflow plane is located at x/C = 0.4, the blade leading edge is at x/C = 0.

Fig. 25 Decay of freestream turbulence intensity for V103 test case plotted at mid-pitch
Flow Turbulence Combust

Fig. 26 Pressure coefficient for V103 compressor cascade

For the cases with non-zero Tu, inlet values of eddy-to-molecular viscosity ratios are
selected in order to fit the DNS data on the decay of freestream turbulence at mid-pitch
as close as possible. Figure 25 shows a reasonable agreement of the downstream Tu decay
from the present RANS simulations with DNS cases T1-T3 in [39].
Pressure coefficient and skin friction on the surfaces of the blade are plotted in Fig. 26
and in Fig. 27 for the and -Re transition models in comparison with the DNS data.
On the suction side, both models are in close agreement with the DNS data and are able to
capture the overall effect of the change in Tu. For the case with zero turbulence intensity
the wiggles in the DNS Cf distribution indicate unsteadiness due to the open separation
bubble all the way past the trailing edge. Overall, the model gives a slightly later transition
onset and is somewhat closer to the DNS data than the -Re model. For all simulations,
the levels of the skin friction coefficient after transition are lower than for the DNS data,
especially on the suction side. This issue is however not related to transition modeling and
is rather a difference between the turbulence model (SST) and the DNS. A similar problem
was found in [18] when computing this test case using a transition model coupled with k-
turbulence model. The possible reason could be an impact from the Kelvin-Helmholtz (KH)
vortex shedding inside the laminar bubble in the DNS on the downstream boundary layer
development. This could overlay the turbulence field and lead to increased mixing near the
trailing edge region. This would not explain the shift in Cf downstream of transition for the
Tu=8% case, where no laminar bubble exists (but where the difference is also smaller).

Fig. 27 Skin friction coefficient for V103 compressor cascade plotted for (upper row) suction side and
(lower row) pressure side
Flow Turbulence Combust

Fig. 28 General view of the test rig for four stage high-speed axial compressor. Courtesy of TFD Hannover

Table 2 General data for 4AV four stage high-speed axial compressor

Compressor component Number of blades Re Tip clearance, mm

IGV 26 234 000 -


Rotor 1 23 420 000 0.4
Stator 1 30 340 000 -
Rotor 2 27 520 000 0.4
Stator 2 32 390 000 -
Rotor 3 29 560 000 0.5
Stator 3 34 420 000 -
Rotor 4 31 550 000 0.4
Stator 4 36 400 000 -

Fig. 29 Details of structural analysis for rotor blades


Flow Turbulence Combust

Fig. 30 Computational domain for CFD of 4AV four stage axial compressor

The model also captures the separation bubble on the pressure side for zero Tu, but
the reattachment point is located considerably further downstream relative to the DNS data.
It should be noted however, that the DNS for this case showed a two-dimensional instanta-
neous flow characteristics, meaning essentially a laminar unsteady flow. This is of course
different from what the combination of the transition model and the SST model would
predict. For other Tu levels, both transition models delay the transition on the pressure side
relative to DNS, with the -Re somewhat closer to the DNS.

5.5 3D test case: 4AV four stage high-speed axial compressor

The four stage high-speed axial compressor 4AV is a research compressor studied at the
University of Hannover [40] (Courtesy of Prof. Seume, TFD Hannover). A general view of
the experimental test rig is illustrated in Fig. 28. The test rig is driven by a 1300 kW DC
motor. The compressor itself is a 4-stage high speed machine fully equipped with controlled

Fig. 31 Fragment of computational mesh for CFD simulations


Flow Turbulence Combust

Fig. 32 Total pressure ratio and isentropic efficiency at design rotational speed

diffusion airfoil (CDA) blading. The maximum speed is 18000 rpm, the design speed of
the tests selected for the present simulations is 17100 rpm. Table 2 shows the number of
blades in each component of the compressor, the Reynolds numbers at the design speed and
sizes of rotor tip clearances measured at zero speed. To account for the displacements due
to rotational velocity of the rotor blades and the corresponding decrease of tip clearances,
the mechanical deformation of the rotors at the design speed has been simulated using the
ANSYS Mechanical solver. The computed deformations have been used to modify the CFD
grids for rotors built originally for zero speed conditions. The grid used for the structural
analysis of Rotor 1 is shown in Fig. 29 together with the predicted total deformations.
The CFD computations have been limited to steady state and employed the mixing plane
approach (Stage) to connect rotating and stationary components of the compressor. Only
one passage of every blade row has been modeled. The computational domain is shown in
Fig. 30. Every single row is resolved with about 900 000 hexahedral cells, including 14
radial cell layers for rotor tip clearance. The total compressor is meshed with 7.6 million
cell volumes. A part of the mesh is shown in Fig. 31.
Air as ideal gas was activated in the solver to model the flow within the compressor, and
the total energy heat transfer model was selected to predict the temperature throughout the
flow. The turbulence models were the SST for fully turbulent simulations and the new
transition model for the simulations accounting for the transitional effects. For all rotors,
solutions were computed in the rotating frame of reference using the relative velocity for-
mulation. Stationary rotor shrouds were therefore modeled as counter-rotating walls. The
no-slip condition was employed on all walls treated as adiabatic when solving the energy
equation. Constant values of total pressure and total temperature were assigned at the inlet
plane of the computational domain. Magnitudes of these quantities were equal to 0.6 bar
and 288.15 K, respectively. Values of turbulence characteristics at the inlet corresponded to
a freestream turbulence intensity of 5%, and the eddy-to-molecular viscosity ratio equal to
10. A mass flow boundary condition was specified at the outlet. At this condition, the exit
pressure value is adjusted in order to obtain a target mass flow rate corresponding to the
desired operating point. The target mass through the machine is obtained at convergence of
steady-state simulations.
Flow Turbulence Combust

Figure 32 shows the measured and computed overall total pressure ratio characteristic
and the corresponding isentropic efficiencies of the compressor at design speed. The left-
most points of the computations are the last operating points where a converged solution
could be obtained. If the mass flow rate at the outlet boundary is decreased further, the
computations become unstable. The inclusion of transition through the model leads to a
significant improvement with the experiments relative to the fully turbulent SST solution.
There is still a shift by about 0.8% to the experimental choke point in the computations using
the transition model. However, there are a number of issues not fully consistent between
the CFD and the experimental set-up which could explain this shift, especially omission of
leakage flows in the simulation.

Fig. 33 Distribution of wall shear stresses and surface streamlines computed at design point (mass flow rate
7.82 [kg/s])
Flow Turbulence Combust

Figure 33 illustrates the effect observed in the global characteristic curves. Shown are the
distribution of wall shear stresses and surface streamlines computed at the mass flow rate
corresponding to the design point (mass flow rate 7.82 [kg/s]). When the transition model
is turned on, significant flow patches on the suction sides of all blades are predicted laminar
up to a laminar separation bubble, from where the flow reattaches as turbulent. A higher
momentum of the boundary layers gained through transition delays the onset of turbulent
separation and results in higher pressure ratios and increased efficiency.

6 Conclusions

A new transition model with one equation for the intermittency has been developed and
tested. The transition model is part of the LCTM model family and a further development
of the -Re model. It solves only one transport equation for the turbulence intermittency,
, and avoids the need for the second Re equation of the -Re model. The new model
has several advantages over the -Re transition model. Firstly, it reduced the computa-
tional effort by solving only one transport equation instead of two. In addition, it avoids the
dependency of Rec equation on the velocity U . This makes the transition model Galilean
invariant. It can therefore be applied to surfaces which move relative to the coordinate sys-
tem for which the velocity field is computed. Finally, the model formulation is simple and
can be fine-tuned based on a small number of user parameters. Like the -Re model, the
model is based strictly on local variables. The model preserves the main characteristics
of LCTM formulations of combining conventional transport equations with experimental
correlations.
The transition model has been calibrated against a wide range of generic as well as
turbomachinery and external aeronautical test cases. In general, good agreement with the
available experimental data was observed.
Some of the model coefficients are made adjustable to allow for fine tuning of the
model. However, this is not encouraged unless there is a clear indication of systematic dis-
crepancies for a certain application and should only be done based on detailed experimental
data. The most likely area in need of fine-tuning is for external aerodynamic flows at low
Tu and under favorable pressure gradients, where no test cases have been computed so far
with the new model.
The present formulation of the transition model can serve as the basis for the inclusion
of additional effects and flow regimes such as roughness, freestream turbulent length scale,
streamline curvature, cross-flow transition and Mach number effects. A first formulation of
a locally formulated crossflow transition criteria has already been combined with the model
and will be presented separately. Aided with these features, the new model should offer an
attractive and simple framework for including different types of transition phenomena into
industrial CFD simulations.

Acknowledgments The current work has been jointly sponsored by GE Global Research and ANSYS Inc.

Appendix A: Meshing Guidelines

The application of transition models requires somewhat finer grids than are typically used
for routine design purposes. The main reason being that one needs to resolve the thin laminar
Flow Turbulence Combust

boundary layer accurately. In addition, a sufficient number of grid points in the streamwise
as well as in wall-normal direction are needed to resolve the transitional region. It has been
observed that at least 30 nodes are necessary across the boundary layer in order to get an
accurate and fully grid-independent solution. The effect of the boundary layer resolution
for the T3A test case is shown in Figure A1. All meshes have a y + value of about 0.5.
Two different types of near wall meshes have been created for demonstration purposes. The
first mesh law has a constant wall-normal expansion ratio (ER) over the whole thickness
of the boundary layer, while the second one reduces the ER as the wall distance increases
and eventually ER becomes unity. The two mesh laws are available in ANSYS ICEM CFD
mesh generator and are named there as Geometric and BiGeometric, respectively. The
geometric mesh law results in a fast growth of elements height and for the ER = 1.3 the
boundary layer is covered with only 15 nodes (Fig. A2) which gives a solution that is notice-
ably different from those on finer meshes (left plot in Figure A1). If the number of nodes
stays above 30 which is the case for the bigeometric mesh law even for the ER = 1.3,
then the solution is virtually insensitive to the mesh. It should be noted however, that the
ER = 1.1 provides an acceptable solution for both mesh laws, therefore the recommended
best practice mesh guidelines developed previously for the -Re model, namely a max y +
of 1 and a ER = 1.1, are also applicable for the new transition model.
The effect of y + variation is shown in Figure A3. For all meshes, the bi-geometric law
with the ER=1.05 was applied. The solutions for y + values between 0.05 and 1.0 are very
close to each other, while at higher values the results start to deviate. This is however not
surprising, because there are too few nodes to resolve the boundary layer in wall-normal
direction. For the case with y + = 6 there are less than 10 nodes, as illustrated in Figure A4
in comparison with the mesh that has a y + of 0.5. This example clearly demonstrates that
wall-function grids cannot properly resolve the laminar boundary layer and therefore cannot
be used for transition modeling.
The requirements for the grid resolution in the streamwise direction depend on the appli-
cation. For an attached flat plate boundary layer, 50 nodes from the start of the laminar
boundary layer to the point where the transition is completed are found sufficient to provide
the grid-independent solution. This requirement is normally satisfied when there is at least
100 streamwise grid points from the leading edge to the trailing edge (i.e. one chord length)

Fig. 34 Figure A1. The effect of the wall-normal expansion ratio (ER) for the flat plate T3A test case: (left)
Geometric mesh law; (right) BiGeometric mesh law
Flow Turbulence Combust

Fig. 35 Figure A2. Boundary layer resolution for the flat plate T3A test case: (left) Geometric mesh law;
(right) BiGeometric mesh law

on each side of a blade. However, a better streamwise resolution will most likely be neces-
sary if separation induced transition takes place. In this case, the number of nodes has to be
sufficient to resolve the separation bubble. A general rule is that at least 20 nodes should
cover the bubble length. For low Reynolds numbers (<105 ) the bubble size is comparable
to the blade chord length and this requirement is easy to satisfy. Even the recommended of
100 grid points from the leading to the trailing edge are typically sufficient. However, the
bubble decreases as the Reynolds number increases, therefore the grid has to be refined in
the regions of the expected separation. An example on how the mesh density influences the
resolution of separation bubble near the leading edge of a wind turbine airfoil is illustrated
in Figure A5. The two meshes shown differ by the number of nodes in streamwise direc-
tion in the region covered by the bubble. On the Mesh 1, the bubble is resolved by just a
few nodes and appears to be much smaller compared to that for the finer Mesh 2. Clearly,
if there are not enough streamwise nodes, the model cannot resolve the rapid separation-
induced transition and the boundary layer on the suction side stays laminar, which might
have a large influence on the lift coefficient and stall angle predictions.
The best practice mesh guidelines can be summarized as follows. The requirements for
the wall-normal resolution are universal and the proper mesh should have a maximal y + of
1 and the wall normal expansion ratio of below 1.1. The number of nodes in streamwise

Fig. 36 Figure A3. The effect of


the y+ for the flat plate T3A test
case
Flow Turbulence Combust

Fig. 37 Figure A4. Boundary layer resolution for the flat plate T3A test case: (left) y+ = 6.0; (right)
y+ = 0.5

Fig. 38 Figure A5. Effect of streamwise grid density for resolving the separation-induced transition due to
a leading edge separation bubble for a wind turbine airfoil

direction is case dependent, but there must be always at least 100 grid points from the
leading edge to the trailing edge on each side of a blade. If separation induced transition is
present, the mesh has to provide at least 20 points streamwise for each separation bubble. A
local mesh refinement might be necessary to satisfy this requirement.

References

1. Menter, F.R., Esch, T., Kubacki, S.: Transition Modelling Based on Local Variables. In: Proc. 5th Int.
Sym. on Engineering Turbulence Modelling and Measurements, Mallorca, Spain (2002)
2. Menter, F.R., Langtry, R.B., Likki, S.R., Suzen, Y.B., Huang, P.G., Volker, S.: A Correlation based
Transition Model using Local Variables Part 1 - Model Formulation (2004). ASME Paper No. GT-2004-
53452
3. Menter, F.R., Langtry, R.B., Volker, S.: Transition Modelling for General Purpose CFD Codes. Journal
Flow Turbulence and Combustion 77, 277-303 (2006)
4. Langtry, R.B., Menter, F.R.: Correlation-Based Transition Modeling for Unstructured Parallelized
Computational Fluid Dynamics Codes. AIAA J. 47(12), 2984-2906 (2009)
5. Content, C., Houdeville, R.: Application of the -Re Laminar-Turbulent Transition Model in Navier-
Stoles Computations AIAA Paper 2010-4445 (2010)
6. Suluksna, K., Juntasaro, V., Juntasaro, E.: Capability Assessment of Intermittency Transport Equations
for Modeling Flow Transition. In: Proc. 19 Conference of Mechanical Engineering Network of Thailand,
Phuket, Thailand (2005)
Flow Turbulence Combust

7. Malan, P., Suluksna, K., Juntasaro, E.: Calibrating the -Re Transition Model for Commercial CFD.
AIAA Paper 2009-1142 (2009)
8. Misaka, T., Obayashi, S.: Application of Local CorrelationBased Transition Model to Flows around
Wings. AIAA Paper 2006-918 (2006)
9. Piotrowski, W., Elsner, W., Drobniak, S.: Transition Prediction on Turbine Blade Profile with Intermit-
tency Transport Equation. ASME J. of Turbomach 132(1) (2009)
10. Coder, J.M., Maughmer M. D.: One-Equation Transition Closure for Eddy-Viscosity Turbulence Models
in CFD. AIAA Paper 2012-0672 (2012)
11. Coder, J.M., Maughmer M. D.: Computational Fluid Dynamics Compatible Transition Modelling using
an Amplification Factor Transport Equation. AIAA J. (2014). doi:10.2514/1.J052905. to be published in
AIAA Journal
12. Seyfert, C., Krumbein, A.: Correlation-Based Transition Transport Modeling for Three-dimensional
Aerodynamic Configurations. AIAA Paper 2012-0448 (2012)
13. Grabe, C., Krumbein, A.: Extension of the -Re Model for Prediction of Crossflow Transition. AIAA
Paper 2014-1269 (2014)
14. Medida, S., Baeder, J.: A New Crossflow Transition Onset Criterion for RANS Turbulence Models.
AIAA Paper 2013-3081 (2013)
15. ANSYS CFX, Release 17.0, Help System, Theory Guide, ANSYS, Inc.
16. Dassler, P., Kozulovic, D., Fiala, A.: Transport Equation for Roughness Effects on Laminar-Turbulent
Transition (2012)
17. Durbin, P.A.: An intermittency model for bypass transition. Int. J. Heat Fluid Flow 36, 16 (2012)
18. Ge, X., Arolla, S., Durbin, P.: A Bypass Transition Model Based on the Intermittency Function. J. Flow
Turbulence and Combustion (2014). doi:10.1007/s10494-014-9533-9
19. Walters, D.K., Leylek, J.H.: A New Model for Boundary-Layer Transition Using a Single-Point Rans
Approach. ASME J. of Turbomach. 126(1), 193202 (2004)
20. Walters, D.K., Cokljat, D.: A Three-Equation Eddy-Viscosity Model for Reynolds-Averaged Navier-
Stokes Simulations of Transitional Flows. J. of Fluids Eng. 130, 2008
21. Wilcox, D.C.: Turbulence Modeling for CFD. DCW Industries, Inc., La Canada, CA (1993)
22. Menter, F.R.: Influence of freestream values on k- turbulence model predictions. AIAA J. 30(6), 1657-
1659 (1992)
23. Menter, F.R.: Two-equation eddy-viscosity turbulence models for engineering applications. AIAA J.
32(8), 269-289 (1994)
24. Abu-Ghannam, B.J., Shaw, R.: Natural Transition of Boundary Layers - The Effects of Turbulence,
Pressure Gradient, and Flow History. J. of Mech. Eng. Sci. 22(5), 213228 (1980)
25. Kato, M., Launder, B.E.: The Modeling of Turbulent Flow Around Stationary and Vibrating Square
Cylinders. In: Proc. 9th Sym. on Turbulent Shear Flows, Kyoto, Japan (1993)
26. Mayle, R.E.: The Role of Laminar-Turbulent Transition in Gas Turbine Engines. ASME J. of Turbomach.
113(4), 509537 (1991)
27. Loitsyanskii L.G.: Mechanics of Liquids and Gases. Begell House, New York (1995)
28. Savill, A.M.: Some recent progress in the turbulence modelling of by-pass transition. In: So, R.M.C.,
Speziale C.G., Launder, B.E. (eds.) Near-Wall Turbulent Flows, Elsevier, p. 829 (1993)
29. Schubauer, G.B., Klebanoff, P.S.: Contribution on the Mechanics of Boundary Layer Transition. NACA
TN 3489 (1955)
30. Volino, R.J., Hultgren, L.S.: Measurements in Separated and Transitional Boundary Layers Under Low-
Pressure Turbine Airfoil Conditions. ASME Paper No. 2000-GT-0260 (2000)
31. Suzen, Y.B., Huang, P.G., Hultgren, L.S., Ashpis, D.E.: Predictions of Separated and Transitional Bound-
ary Layers Under Low-Pressure Turbine Airfoil Conditions Using an Intermittency Transport Equation.
ASME J. of Turbomach. 125(3), 455-464 (2003)
32. Swalwell, K.E.: The effect of turbulence on stall of horizontal axis wind turbines. PhD thesis, Monash
University, Australia (2005)
33. Dorney, D.J., Lake, J.P., King, P.L., Ashpis, D.E.: Experimental and Numerical Investigation of Losses
in Low-Pressure Turbine Blade Rows. AIAA Paper AIAA-2000-0737 (2000)
34. Huang, J., Corke, T.C., Thomas, F.O.: Plasma Actuators for Separation Control of Low Pressure Turbine
Blades. AIAA Paper AIAA-2003-1027 (2003)
35. Lake, J.P., King, P.I., Rivir, R.B.: Low Reynolds Number Loss Reduction on Turbine Blades With
Dimples and V-Grooves. AIAA Paper AIAA-00-0738 (2000)
36. Lake, J.P., King, P.I., Rivir, R.B.: Reduction of Separation Losses on a Turbine Blade With Low Reynolds
Number. AIAA Paper AIAA-99-0242 (1999)
Flow Turbulence Combust

37. Stieger, R., Hollis, D., Hodson, H.: Unsteady Surface Pressures due to Wake Induced Transition in a
Laminar Separation Bubble on a LP Turbine Cascade. ASME Paper GT-2003-38303 (2003)
38. Wu, X., Durbin P.A.: Evidence of longitudinal vortices evolved from distorted wakes in a turbine passage,
vol. 446, pp. 199-228 (2001)
39. Zaki, T.A., Wissink, J.G., Rodi, W., Durbin, P.A.: Direct numerical simulations of transition in a
compressor cascade: the influence of free-stream turbulence, vol. 665, pp. 5798 (2010)
40. Fischer, A., Riess, W., Seume, J.R.: Performance of strongly bowed stators in a 4-stage high speed
compressor. ASME Paper No. GT2003-38392 (2003)

You might also like