Pitching wing

Kavithasan Patkunam, Wenyuan Yu

KTH Royal Institute of Technology, 100 44 Stockholm, Sweden

The case considered is a NACA0012 airfoil section at Re = 1.35 × 105 for

which detailed experimental measurements are available for both static as
well as dynamic cases. The main aim is to simulate the Static and Dynamic
conditions of the 2D airfoil and compare the results with the Experimental
values and then predict the Transition point.

I. Introduction

RANS[1] models are a vital tool for fluid flow simulations in engineering applications. These
models however can suffer from a number of drawbacks depending on the type of flow case that
is under investigation. Stall of retreating helicopter blades is one important industrial example of
unsteady flow phenomenon, one which has an important consequence of limiting the maximum
speed of helicopters. Vertical axis wind turbine rotors experience a constantly changing angle of
attack during their rotation cycle which affects their power generation. Flapping of wings in
insects and birds exhibit similar cycles of rapidly changing angle of attack (with added
complexities of deforming lifting surfaces). Even in stationary wings, incoming wind gusts or
structural vibrations can lead to a change in effective angle of attack and give rise to unsteady
phenomenon.

II. Method

A. Theorectical Calculations
While building the mesh, the mesh of first layer should be particularly refined based on
selected RANS model and wall function. In our case, the reference length of airfoil section
(0.15m) and Reynolds number (135000).
𝑢𝑢∞ 𝑙𝑙
Re =
𝜈𝜈
𝑢𝑢∞ ≈ 14𝑚𝑚/𝑠𝑠

According to empirical formula , the boundary layer thickness δ is :

δ = 0.37L𝑅𝑅𝑅𝑅 −1/5 = 5.2267 ∗ 10−3

The friction coefficient 𝐶𝐶𝑓𝑓 is:

 𝐶𝐶𝑓𝑓
= 0.0296𝑅𝑅𝑅𝑅 −1/5 = 2.788 ∗ 10−3
2

Wall friction velocity 𝑢𝑢𝑡𝑡 is:

𝐶𝐶𝑓𝑓
𝑢𝑢𝑡𝑡 = 𝑢𝑢∞ � = 0.7392
2
Viscous wall scale l*:
𝜈𝜈
l ∗= = 2 ∗ 10−5
𝑢𝑢𝑡𝑡

Dimensionless distance y+:

𝑦𝑦
y+=
𝑙𝑙 ∗

As y should be smaller than 1/10 of boundary layer thickness,

ny+≤ 0.1δ
n ≤26
As y+ should either smaller than 5 (viscous sublayer) or larger than 30 (log law region), in
order to avoid buffer layer, we could only put first layer in the viscous sublayer.

B. Turbulence Model

The k-ω model is a widely used turbulence model. It is a 2 equation model and first
transported variable is turbulent kinetic energy, K while the second transported variable in this
case is the specific dissipation, ω. The SST k-ω[4] model is used in this case. This model has an
advantage over the Standard k-ω model because it can be used as a Low-Turbulence model
without any problems. The SST (Shear Stress Transport) formulation has the tendency to switch
its behaviour to the k-ε[5] model and therefore avoids the problem of the model to be too
sensitive to the inlet free stream turbulence.

In Fluent, the wall boundary conditions for the k equation in the k-ω models are treated in the
same way as the k equation is treated when enhanced wall treatments are used with the k- epsilon
models. That means that the wall y+ should be equal to 1 but it not a problem to have a value
more than 1 but well inside the Viscous sublayer, y+ < 5 is acceptable. Since the y+ in our case
is less than 3, we chose to use the SST k-ω model.

Besides k-ω, in dynamic case, transition SST and k-epsilon models are also used. Trainsition
point is predicted with three different models, the results are different and will be shown in the
result section. The k-epsilon model is a 2 equation model like the k-ω but the second transported
variable will be Turbulent dissipiation. The Langtry-Menter 4-equation Transitional SST
Model[3] is also sometimes known as the "gamma-Retheta-SST" model, because it makes use of
equations for "gamma" and "Retheta" in addition to SST's "k" and "omega" equations.

2
C. ANSYS Setup
a. Geometry

Fluent is used as our choice of CFD tool. The airfoil of interest, NACA 0012 is imported to
the Design Modeler interface in ANSYS from a CAD software like CATIA V5. The trailing
edge is made a little rounded so that inflation layers can be used on the profile without any
problems. After importing to the Design Modeler, the domain is constructed in such a way that it
is large enough to not to interact with the flow. The length of the domain is chosen to be 25 times
the chord length and the Heigth to be 15 times the Chord length. But several other configurations
of the size where considered (Totally 4 Meshes) with different domain size and different
meshing techniques to prove that the Values we obtained where mesh independent and Domain
independent. To build the mesh in the domain. We chose C- topological structure domain to
ensure orthogonality, as airfoil section is a highly curved line.

Figure 1 : Domain
b. Mesh

The mesh type we select is unstructured quad mesh. We give up the choice of structured mesh
because we introduce inflation layer. Besides, compared with quad mesh, tri mesh is less
accurate. In total, there are around 130000 cells in the domain. Around the airfoil, two aspects
need to be considered while building inflation layer. First, more grids should be put around
leading edge and trailing edge of airfoil. Second, to ensure the quality if mesh, the sharp trailing
edge is transformed into round. In total, there are around 130000 cells, the average minimum
quality of mesh is around 0.62. The meshes below in the picture, were among the 4 mesh types
we build to check that the solution is mesh independent.

Figure 2 : Different meshes

3
 The 4 meshes and domain where checked for the similar coefficient of lift values at a
particular angle of attack (5°), time taken to complete the simulation, the order of convergence
and the wall y+ value[6]. The final mesh chosen mesh is shown below with a wall y+ less than 3
which is ideal for using the SST k-ω model as discussed before. The Skewness and the
orthogonal qualities where also checked and considered. After careful consideration, the
following mesh was considered as our final mesh to use it further in our analysis.

Figure 3 : Chosen refined mesh

Within inflation layer, when y+ is smaller than 5, there should be at least 60 layers. At the
place adjacent to far field boundary, the number of cells decreases in order reduce computation.

Figure 4 : Orthogonal quality of the mesh

Fluent can plot the contour of mesh quality. Color red and yellow indicates the mesh quality is
good and the value should be close to unity.

4
 Figure 5 : Chosen refined mesh

The graph below shows the Wall y+ plot with respect to the chord length. It can be noted that
the wall y+ is less than 3 which indicates that the mesh we build is reliable.

Figure 6 : Wall y+ plot

c. Fluent

In fluent, the Pressure based solver and SST k-ω model is chosen for both the Static and the
Dynamic parts. The velocity was set to 14m/s and the reference Area to 0.15m (chord length =
0.15m). FMG Initialisation was used only for the static case and Hybrid initialization for the
Dynamic case. The SIMPLE (Semi-Implicit Method for Pressure-Linked Equations) algorithm
was considered. For the Linear part of the Cl vs Alpha curve, second order upwind was
considered to give a more accurate results. But close to the stall, convergence could not be
obtained and therefore first order was considered for the stall angle which is 13°. For the
dynamic analysis, we considered the same and used first order upwind to enable a faster
simulation and convergence. The results will be discussed in the next Section.

The dynamic analysis was more challenging and we had to write our own UDF[7] (User
Defined Function) in C language to simulate the motion of the pitching wing. A UDF, is a
function is a program that can be dynamically loaded with the ANSYS Fluent solver to enhance

5
the standard features of the code. Source files containing UDFs can be either interpreted or
compiled in Fluent. For interpreted UDFs, source files are interpreted and loaded directly at
runtime, in a single-step process. For compiled UDFs, the process involves two separate steps. A
shared object code library is first built and then it is loaded into Fluent.

The macro cg_motion is used to specify the motion of a particular dynamic zone in Fluent by
providing it with the angular velocities at every time step. It uses these velocities to update the
node positions on the dynamic zone based on solid-body motion. The following equation was
given to simulate the motion.
θ = 7.5 sin 𝜔𝜔𝜔𝜔

θ is the pitching angle while 𝜔𝜔 is frequency angular velocity. we need the derivative of θ.

θ̇ = 7.5 𝜔𝜔sin 𝜔𝜔𝜔𝜔

𝜔𝜔 is 9.33 rad/s.

The UDF code is shown below:

#include "udf.h"
DEFINE_CG_MOTION(aero, dt, vel, omega, time, dtime)
{
real a, w, p, pi;
pi = 3.14159265;
/* define motion variables */
a = 7.5*pi/180;
w = 0.05*28/0.15;
p = pi/2;

omega[2] = a*w*sin(w*time+p); /* z-component of angular velocity */

}

The macro cg_motion can be used in fluent only when the code is compiled, not interpreted.
This added more complications while not all 64bit Windows systems had C language compilers
available straight away and it was required to install the compilers manually which added more
complications.

There are six arguments to present in this macro and they are name, dt, vel, omega, time, and
dtime. You supply name, the name of the UDF hich in this case is ‘aero’. dt, vel, omega, time,
and dtime are variables that are passed by the Fluent solver to your UDF. The angular velocities
are returned to Fluent by overwriting the arrays vel and omega, respectively.

6
 Figure 7 : Dynamic mesh zone window

In Fluent, instead of building dynamic mesh, we set the whole domain, including inlet, outlet,
mesh and wall as rigid bodies which will rotate around gravity center of airfoil( 1/4 chord
location). The inlet velocity is set a fixed direction so that the pitching angle varies
corresponding to time. Without using smooth or rebuild mesh function in dynamic mesh, the
quality of mesh is guaranteed. This was better ofcouse since smoothing, remeshing or even
layering methods available in dynamic meshing in Fluent will give Negative volumes and other
problems when attempted. The results we have obtained is convincing and close to the
experimental solution.

Figure 8 : Velocity inlet window

7
 The dynamic zones are picked and the compiled code is implemented on them as shown in the
previous image. All the zones are considered to be a rigid body. The Centre of Gravity or the
point about which the rotation should take place should be mentioned in the window. In our case,
the rotation is about 25% of the chord length from the leading edge and since our origin is
located at the trailing edge, we use a negative value (-0.1125m). It should be noted that the
velocity inlet in the boundary conditions is set to be flowing in the direction of the positive x-
axis even if the mesh itself is moving. This is can seen from the window above.

Figure 9 : Solution animation window

The animation is setup to record 1 cycle of the oscillations to save hard disk space. We had a
limitation on the disk space so we were not able to record all the output parameters in a graph or
contour. The time taken to complete 1 cylce was calculated to be 0.67 seconds and the
Simulation was recorded for about 0.7 seconds. The above picture is the window showing the
sequence that has to be recorded every iteration/timestep. The sequence needed should be
specified by clicking on ‘Define’ and selecting the needed plot.

III. Results

A. Static case
The static case analysis provided good results in fluent which is equivalent to the
experimental values of the coefficient of lift. But the results still vary to some level from the
experimental values although the trend of the curve remains the same. The variation in these
values can be because of the the change of the Second order upwind to the first order due to lack
of convergence near the Stall. The stall region was refined with more points to capture the slope.
This can be seen for the plot below.

8
 Figure 10(a) : Cl Vs angle of attack plot Figure 10(b) : Cl Vs angle of attack plot
(obtained from CFD) (obtained from experimental studies)

Figure 11 : Residuals

The above pictures show the residuals convergence. The image on the left shows a
convergence for an angle of attack at stall and the one in right shows the convergence for angles
more than the stall (20°). It can be seen that the convergence faced some trouble and it needs
more than 5000 iterations to get a good convergence. The stall angle (13°) took 16000 iterations
to reach convergence of the order 10-4 as seen in the image below.

For all of attack of angles, we use second order upwind scheme to ensure the accuracy of the
results. The results from first order scheme is a little smaller than second order upwind scheme,
but it will take much less iteration to let residuals to converge. Among the curves we monitored,
the continue equation is more difficult to be convergent than others, so we modified a little on
relaxation factor in some cases.

The difficulty in converging of big attack angle computation is reasonable, cause separation
of the flow occurs and it induces a large reverse flow area, where you can see in vector contour
fig14. This phenomenon causes the unsteady of the flow, which make monitored curves difficult
to converge.

9
 Figure 12 : Residuals for α> Stall angle

Figure 13 : Velocity vector for 5° AoA

The velocity vector representation in the above picture is for a small angle of attack. It shows
separation point is close to trailing edge. The image below represents the velocity vectors for an
angle of attack of more than the Stall angle (represents 17.5°) and it clearly shows that there is a
flow separation causing the flow to circulate.

10
 Figure 14 : Velcoity vector for 17.5° AoA

The following pictures show the Velocity contours of 5° and 20° angle of attack. Just like in
the velocity vectors, the wake region for the angle of attack more than stall can be seen clearly in
Figure 16.

Figure 15 : Velocity contour at 5° AoA

For all of attack of angles, we use second order upwind scheme to ensure the accuracy of the
results. The results from first order scheme is a little smaller than second order upwind scheme,
but it will take much less iteration to let residuals to converge. Among the curves we monitored,
the continue equation is more difficult to be convergent than others, so we modified a little on
relaxation factor in some cases.

The difficulty in converging of big attack angle computation is reasonable, cause separation
of the flow occurs and it induces a large reverse flow area, where you can see in vector contour
fig14. This phenomenon causes the unsteady of the flow, which make monitored curves difficult
to converge.

11
 Figure 16 : Velcoity contour at 17.5° AoA

B. Dynamic case
The dynamic case was simulated using the UDF code mentioned above in the precious
sections. The results obtained were ploted along with the static case as shown in the upcoming
image. In the plot, the Blue line denotes the dynamic case and the green line denotes the static
case. The dynamic case was limited to a maximum of 7.5° in the UDF code. By comparison, it
can be observed that the linear part of the static case overlaps with the dynamic part. This assures
the accuracy of the dynamic results.

Figure 17 : Comparison of Static and dynamic case from CFD

(Cl)

12
 Figure 18 : Comparison of Static and dynamic case
(Cl)

The above plot shows the experimental results obtained for the static and the dynamic parts
which was mentioned in the reference paper. The trend looks similar to the plot we obtained
from the CFD analysis.

As mentioned before, we used 3 turbulence model to find the transition point. The video
attachments will have the dynamic plots. By looking at the plots for Turbulent viscosity vs Chord
distribution along the x-axis and the plots for the Skin friction coefficient vs the Chord
distribution along x-axis, the following transition points have been obtained. It should be noted
that, in the animations, the trailing edge is set to zero and not the leading edge.

Figure 20 : Transition point prediction

13
 Figure 20 shows the results from three different RANS models, and the results are different.
As in turbulent boundary, there is a fluctuation in y direction, the friction is higher than that in
laminar layer. We can use this phenomenon to distinguish hybrid boundry layer with full
turbulent layer.

First, we notice the friction coefficients or wall shear stress computed from k-w and k-epsilon
models are much higher than that computed from Transitional SST model. Second, the peak of
friction coefficients is very close to the leading edge of airfoil in first two models, which
indicates there is no transition point, the whole boundary layer is regarded as full turbulent
boundary layers by k-w and k-epsilon models. Third, in the plot obtained with Transitional SST
model, the friction coeffcients are much lower and there is a reattachment at the location where
20% of chord length. This loction will vary corresponding to the attack angle.

Method k-ω (SST) k-ε (EWT) Transition SST

Transition point (x/c) 0 0 0.2

Figure 19 : Cd vs AoA obtained from CFD

Figure 21 : Cm vs AoA obtained from CFD

14
 The above two plots represent the Coefficient of Drag and the Coefficeint of moment with
respect to the angle of attack. It should be noted that the coefficient of moment is taken about
trailing edge in our case.

Figure 22 : Coefficents Vs Flow time

This above plot shows the coefficients (Cl,Cd,Cm) against the Flow time in seconds. It is
observed that the Coefficents curve with respect to the flow time are a sine curve.

IV. Conclusion

The main objective of this work has been to predict the transition location by the use of a
RANS solver. We use SST k-ω RANS model to analyse both static and dynamic performance of
NACA 0012 airfoil section and used k-ε and Transition SST models for comparing the transition
points for the 2D airfoil besides SST k-ω. The mesh quality is considered to be good since the
orthogonal quality are close to unity and the Skewness level are close to zero. Results from Static
case are trustable, while there are some uncertainities in Dynamic part, especially Cd and
transition point prediction (shear stress distributon).

V. Acknowlegement

We would like to thank Mr. Walter Fornari, Mr. Prabal Negi and Prof. Stefan Wallin for their
valuable contribution in getting this project done by helping us to get pass the problems we
faced.

VI. References

[1] T. Lee AND P. Gerontakos, “Investigation of flow over an oscillating airfoil”

15
[2] Prediction of Transition on Wings in a RANS Approach,

[3] Menter, F.R., Langtry, R., Volker, S. and Huang, P.G., “Transition Modelling for General
Purpose CFD”, C. de Nicola , L. Graziosi , R. S. Donelli

[4] http://www.cfd-online.com/Wiki/K-omega_models

[5] http://www.cfd-online.com/Wiki/SST_k-omega_model

[6] FLUENT 6.3 User's Guide

[7] ANSYS Fluent 12.0 UDF Manual

16