Experimental study of the turbulent flow

in a plane asymmetric diffuser

by

Olle Törnblom

January 2003
Technical Reports from
Royal Institute of Technology
Department of Mechanics
SE-100 44 Stockholm, Sweden
 Typsatt i LATEX med mekaniks avhandlingsstil.

Akademisk avhandling som med tillstånd av Kungliga Tekniska Högskolan i

Stockholm framlägges till offentlig granskning för avläggande av teknologie
licentiatexamen onsdagen den 15:e januari 2003 kl 13.00 i Seminarierum S40,
Teknikringen 8, Kungliga Tekniska Högskolan, Stockholm.

­
c Olle Törnblom 2002
Universitetsservice US AB, Stockholm 2002
Olle Törnblom 2003 Experimental study of the turbulent flow in a plane
asymmetric diffuser
Department of Mechanics, Royal Institute of Technology
SE-100 44 Stockholm, Sweden
Abstract
The separating turbulent flow in a plane asymmetric diffuser is investigated
experimentally. The considered flow case is suitable for fundamental studies
of separation, separation control and turbulence modelling. The flow case has
been studied in a specially designed wind-tunnel under well controlled condi-
tions. The velocity field has been mapped out with PIV and LDV techniques
in order to determine all three velocity components. Knowledge of all velocity
components allows the formation of several quantities of interest in turbulence
modelling such as the turbulence kinetic energy, the turbulence anisotropy
tensor and the turbulence production rate tensor. Pressures are measured thr-
ough the diffuser. The measured data will form a reference database which can
be used for evaluation of turbulence models and other computational investi-
gations. A comparative study is made where the measured turbulence data
are used to evaluate an explicit algebraic Reynolds stress turbulence model
(EARSM). A discussion regarding the underlying reasons for the discrepancies
found between the experimental and the model results is made. A simple form
of separation control using vortex generators is tested with positive result.

Descriptors: Fluid mechanics, wind-tunnel, turbulence, turbulence modelling,

EARSM, asymmetric diffuser, boundary layer, PIV, LDV, vortex generator,
control.
 Preface

This thesis is an experimental study of the turbulent flow in a plane asymmet-

ric diffuser and it is based on the following papers.

Paper 1. Lindgren, B. Törnblom O. & Johansson, A. V. 2002 Measurements

in a plane asymmetric diffuser with 8.5◦ opening angle. Part I: General flow
characteristics. To be submitted.

Paper 2. Törnblom O., Lindgren, B. & Johansson, A. V. 2002 Measurements

in a plane asymmetric diffuser with 8.5◦ opening angle. Part II: Comparison
with model predictions for turbulence characteristics. To be submitted.

Paper 3. Törnblom O., Lindgren, B. & Johansson, A. V. 2002 Design and

calibration of a plane asymmetric diffuser wind-tunnel.

Division of work between paper authors

The papers included in this thesis have been written in collaboration with other
researchers. Below follows a description of the contribution the respondent
made to the different papers. Arne V. Johansson acted as supervisor and
project leader in all investigations. The respondent made a major part of the
writing of the papers, if not otherwise stated below.
Paper 1. This work was made in collaboration on equal terms with Björn
Lindgren, who wrote the major part of the paper text. Parts of the work was
presented at the 9th European Turbulence Conference, Southampton, UK, 2002
by the respondent and at the 11th International symposium on application of
laser techniques to fluid mechanics, Lisbon, Portugal, 2002 by Björn Lindgren,
with written contributions to the conference proceedings.
Paper 2. The experimental work was made in collaboration on equal terms
with Björn Lindgren. The simulations were performed by Johan Gullman-
Strand. Parts of the work was presented at the 9th European Turbulence
Conference, Southampton, UK, 2002 by the respondent and at the 11th In-
ternational symposium on application of laser techniques to fluid mechanics,
v
vi PREFACE

Lisbon, Portugal, 2002 by Björn Lindgren, with written contributions to the

conference proceedings.
Paper 3. The work was made in collaboration on equal terms with Björn
Lindgren. Minor parts of the work has been presented at the 9th European
Turbulence Conference, Southampton, UK, 2002 by the respondent and at the
11th International symposium on application of laser techniques to fluid me-
chanics, Lisbon, Portugal, 2002 by Björn Lindgren, with written contributions
to the conference proceedings.
Till minne av Morfar
 Contents

Preface v
Division of work between paper authors v

Chapter 1. Introduction 1

Chapter 2. Basic concepts 4

2.1. Governing equations 4
2.2. Turbulent boundary layer separation 5
2.3. The plane asymmetric diffuser 7

Chapter 3. Experiments 8
3.1. Experimental facility 8
3.2. Measurement techniques 9
3.3. Results 11
3.4. Control 12

Chapter 4. Concluding remarks and outlook 14

Acknowledgment 15

Bibliography 16
Papers

Paper 1: Measurements in a plane asymmetric diffuser with 8.5◦

opening angle. Part I: General flow characteristics 21

Paper 2: Measurements in a plane asymmetric diffuser with 8.5◦

opening angle. Part II: Comparison with model
predictions for turbulence characteristics. 49

Paper 3: Design and calibration of a plane asymmetric diffuser

wind-tunnel 67

ix
x
 CHAPTER 1

Introduction

Fluid mechanics and fluid flow turbulence might (at a first glance) seem like
narrow subjects, slightly detached from the world we live in, but I would say
that it’s quite the opposite. Liquid and gas flows are everywhere in our daily
lives, as we are all living in the thin layer of gas covering the earth called the
atmosphere. Whether we walk, cycle, go by car or take an airplane, we have to
move through this layer of air, or rather, the air has to move around us. Even
if we don’t move at very high speed, it is likely that the flow around us will
become turbulent (even when walking quickly one can hear the turbulent flow
around the ears). Many other flows that we can find in our near environment
for instance rivers, fires, chimney plumes and the wind blowing around a house
are also turbulent, and many of the flows that we don’t encounter as often such
as inter-stellar gas clouds or the corona of the sun are turbulent too.
Turbulence is characterized by chaotic and three-dimensional motion of the
fluid on a wide range of scales both in space and time. The chaotic and non-
deterministic nature of turbulence makes it natural to analyze turbulence with
statistical methods. Trying to exactly describe the velocity in every location at
all times doesn’t increase the understanding of turbulent flows. Instead, it is
customary to study averages, standard deviations or higher order moments of
the probability density functions for different flow quantities such as velocity,
correlations between velocities in different directions or correlations between
velocity and pressure. These correlations can be taken both in a single point
or as a spatial correlation. A reason for studying these statistical quantities is
that they all can be found in the governing equations (after some manipulation)
and thus are important also in the mathematical modelling of turbulent flows.
A flow becomes turbulent when the so called Reynolds number (Re) becomes
large enough. The Reynolds number is a nondimensional measure of how large
the inertia forces of the flow are compared to the viscous forces. Simply put, a
high Reynolds number means that the viscosity of the flow is too small to be
able to damp out flow disturbances. In turbulent flows, the Reynolds number
determines the range of scales in space and time over which the kinetic energy
will be spread out.
This thesis deals with incompressible flows, which means that the density
of the fluid is assumed to be constant and not dependent on the pressure. The

1
2 1. INTRODUCTION

incompressibility assumption is valid for water and air flows at velocities that
are small compared to the speed of sound in the respective medium. Air and
water have another property in common, they are so called Newtonian fluids,
which means that the relation between stress and strain-rate in the fluid is
linear i.e. that the viscosity is constant with respect to the strain-rate. Under
these conditions, air and water behave very much in the same way and their
motions can be described by the same equations.
A fluid in relative motion to a solid object forms boundary layers in which
the relative velocity between the fluid and the solid is gradually decreased
until it is zero at the surface of the solid. This is called the no-slip boundary
condition and it is valid for all flows considered in the present context. The
velocity gradient in a boundary layer causes shearing of the fluid and shearing
is a necessary ingredient for disturbance growth and for turbulence to occur
and be sustained. Boundary layers are therefore one of the most important
issues in fluid mechanics, without the no-slip condition this subject would be
a lot less challenging.
A separating flow is the subject of this thesis. Separation occurs when the
flow next to a solid surface is slowed down and eventually reversed, forming
a region of circulating flow between the surface and the outer flow. Examples
of situations where separation might occur and be of importance are; stall of
wings at high angles of attack, in the rear end of bluff vehicles leading to in-
creased air resistance and in bent channels in hydro power plants or ventilation
systems leading to decreased performance. It is obvious that there are many
benefits to earn from better knowledge of separation prediction and separation
control. Separation can be divided into two groups depending on what caused
the separation; geometry and pressure gradient induced separation. Geometry
induced separation occurs in flows over objects with for example sharp corners
where the acceleration of the flow would have to be infinite for it to follow
the surface. Accordingly the separation point (the point where the separated
region starts) is fixed at one position i.e. on the corner. In pressure gradient
induced separation the deceleration of the flow next to the surface is a gradual
process and the location of the separation point is determined by a balance
between the inertia in the boundary layer and the adverse pressure gradient
(the rate at which the pressure increases in the streamwise direction). This
means that the prediction of the separation point is a much more complicated
matter than in the case with geometry induced separation.
Predicting a turbulent flow means solving the governing equations for the
flow, but this quickly becomes an impossible task as the Reynolds number, and
thus the range of scales, is increased. For practically all flow cases, the govern-
ing equations have to be solved numerically, but not even the most powerful
supercomputers of today have a chance of exactly solving for instance the flow
around a car or an airplane. This is because the number of variables needed to
resolve the smallest turbulent scales increases with Reynolds number as Re9/4 .
 1. INTRODUCTION 3

The alternative, if solving the problem with a computer isn’t an option, is to

make an experiment, this has been the traditional way of testing fluid dynamic
design. Many experiments are costly though, and some things such as auto-
matic optimization are easier to do with a computer than in an experiment.
Therefore, turbulence models are used in order to circumvent the problem of in-
creasing computational effort with increasing Reynolds number. In large-eddy
simulations (LES), only the largest scales of the flow are computed directly, and
the effects of smaller scales are modelled using the large scales as input to the
model. In engineering applications, the dominating approach is still to consider
only time- (or ensemble-) averaged momentum equations, complemented with
evolution equations for some set of turbulence quantities, in order to close the
whole set of equations (see chapter 2).
Accurate prediction of separating flows is one of the major challenges for
turbulence models, and the measurements presented in this thesis were made
both to give a well defined test case and to add detailed knowledge about where
the models need to be improved. A future task is to elucidate possible routes
towards separation control methods.
 CHAPTER 2

Basic concepts

2.1. Governing equations

The flow considered in this thesis and many other flows at moderate speed can
be described by the incompressible Navier-Stokes equations,
∂ ũi ∂ ũi 1 ∂ p̃ ∂ 2 ũi
+ ũj =− +ν (2.1)
∂t ∂xj ρ ∂xi ∂xj ∂xj
∂ ũi
= 0. (2.2)
∂xi
Cartesian tensor notation has been used when the above equations were formu-
lated and Einstein’s summation convention should be applied when the tenso-
rial expression is expanded. The velocity tensor ũi is of first rank and contains
the velocities corresponding to the three spatial directions xi , the scalar p̃ is
the pressure and ρ and ν are constants for the density and kinematic viscosity,
respectively. When analyzing turbulent flows, it can be useful to decompose
the velocities and the pressure into a mean part and a fluctuating part,
ũi = Ui + ui , (2.3)
p̃ = P + p, (2.4)
this is often called Reynolds decomposition. Here capital letter quantities are
ensemble averages and lower-case letters represent fluctuations. Taking the
ensemble average of equations 2.1 and 2.2 yields the following equations for the
mean flow,
∂Ui ∂Ui 1 ∂P ∂ 2 Ui ∂ui uj
+ Uj =− +ν − (2.5)
∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj
∂Ui
= 0. (2.6)
∂xi
These equations are called the Reynolds equations or Reynolds averaged Navier-
Stokes (RANS) equations and are the equations that are most often solved in
computations using turbulence models. Note the last term in equation 2.5,
which is a derivative of the Reynolds stress tensor −ρui uj . The overline is
used to indicate ensemble averages. The Reynolds stress appears in the equa-
tions for the mean velocity as an unknown extra stress caused by the turbulent
4
 2.2. TURBULENT BOUNDARY LAYER SEPARATION 5

4 0.7 0.5 0.3

0.1

0
0 5 10 15 20 25 30 35 40

Figure 2.1. Backflow coefficient χ for the plane asymmetric

diffuser flow.

fluctuations. The objective of single-point turbulence models is to model this

term as accurately as possible. Modelling is needed since equations 2.5 and
2.6 only constitute four equations but the number of unknowns are ten (six
unknowns in the Reynolds stress tensor due to index symmetry). This is called
the closure problem of turbulence modelling.

2.2. Turbulent boundary layer separation

Boundary layer separation on a smooth surface occurs if the boundary layer
is subjected to a strong enough adverse pressure gradient (APG) and if the
duration of the APG is long enough. APG boundary layers grow more rapidly
than for instance zero pressure gradient (ZPG) boundary layers since an APG
is associated with a deceleration of the external flow and we get from the
continuity equation (equation 2.6) that the wall normal velocity component U2
will depend on the deceleration of the streamwise velocity as
Z x2
∂U1
U2 (x1 , x2 ) = − dx2 , (2.7)
0 ∂x1
in a two-dimensional flow where U3 = 0. The boundary layer will hence grow
not only due to momentum diffusion, as in the ZPG case. If the APG is
sustained, the velocity in the near-wall region will continue to decelerate and
eventually reverse and form a separated region. Turbulent boundary layers
can sustain adverse pressure gradients longer than laminar ones due to the
increased momentum diffusion caused by the turbulence that more efficiently
brings down high-velocity fluid towards the wall.
Turbulent boundary layer separation is, in contrary to most laminar sep-
arations, a nonstationary phenomenon and the location of the point on the
wall where backflow first occurs will fluctuate. The most common definition of
the separation point is that it is the point where the mean wall shear-stress is
zero but this measure does not tell us anything about where the instantaneous
separation points are located. However, the backflow coefficient χ, defined as
the fraction of time that the flow is reversed does tell us something about the
separation point movement. Figure 2.1 shows the backflow coefficient mea-
sured in the plane asymmetric diffuser, and it can be seen that there is no
6 2. BASIC CONCEPTS

0
0 5 10 15 20 25 30 35

Figure 2.2. Comparison between different definitions of the

separation bubble; dividing streamline (- - -), zero streamwise
mean velocity (—) and backflow 50% of the time (− · −).

(or only a very small) region where there is backflow at all times. This is be-
cause the velocity fluctuations are of the same order or larger than the mean
velocity. In Simpson et al. (1981) it is suggested that the backflow velocity
is supplied intermittently by large-scale structures as they pass through the
separated flow. Hence, the mean streamlines of a turbulent separated region
may give the somewhat wrong impression that the backflow region is one large
coherent recirculation zone.
The flow case investigated in this thesis cannot be directly compared to
separating boundary layers with a free-stream as it is an interior flow in which
the separated region is highly affected by the presence of the straight wall.
I have after studying many articles and several textbooks and questioning
my colleagues, found that the term separation bubble have different meanings
depending on who you ask. I have encountered three different definitions or
interpretations for the limits of a separation bubble: (i ) the separation bubble
is the mean recirculating region within the dividing streamline (also called
separation streamline) reaching between the stagnation points on the wall at
the separation and reattachment points, (ii ) the separation bubble is the region
with mean backflow i.e. below the curve of zero mean velocity and (iii ) the
separation bubble is the region with backflow more than 50% of the time. The
differences between these three definitions are illustrated in figure 2.2 which is
based on the measurements presented in this thesis. Personally, I prefer the
first definition which also seems to be the most established one in the literature.
In a recent paper Haller (2002) presents a kinematic theory for an unsteady
laminar separation bubble, and it is shown that the instantaneous location of
the point of zero wall shear-stress does not always coincide with the separation
point, defined as the intersection between the wall and the separation profile
(a material line that collects and ejects fluid particles from the vicinity of the
wall). The separation profile is basically what one would see in a experimental
visualization of the flow.
 2.3. THE PLANE ASYMMETRIC DIFFUSER 7
24.8 H

o
y
8.5

4.7 H
Recirculation zone x
Turbulent channel inflow
H

Figure 2.3. The plane asymmetric diffuser.

2.3. The plane asymmetric diffuser

The plane asymmetric diffuser used in the present experiment (see figure 2.3)
is a generalized geometry for studies of two-dimensional duct flow separation,
separation control and turbulence model testing. The inlet flow condition is
fully developed turbulent channel flow, this makes it easy to match the ex-
perimental conditions in computations. The plane asymmetric diffuser, has
prior to this study, been investigated in two independent experiments by Obi
et al. and by Buice & Eaton. Results from these experiments have been pub-
lished at several occasions in Obi et al. (1993a), Obi et al. (1993b), Obi et al.
(1997), Buice & Eaton (1997) and Buice & Eaton (2000). Both studies con-
cern diffusers with an opening angle of 10◦ . The experimental data from these
experiments have been used for turbulence model evaluation in several studies,
e.g. Obi et al. (1993a), Apsley & Leschziner (1999), Hellsten & Rautaheimo
(1999), Gullman-Strand et al. (2002). These investigations did not concern so
much the physics of the flow but were more concentrated on turbulence mod-
elling aspects. Kaltenbach et al. (1999) made a computational investigation of
the flow using large-eddy simulation which besides computational aspects also
discusses the flow dynamics.
Unlike these previous investigations, this work concerns the flow in a dif-
fuser with an opening angle of 8.5◦ . We chose to have a smaller angle in order to
have a smaller separation and possibly have a better two-dimensionality of the
flow. The smaller separation will also be more sensitive to the control methods
which are to be tested in a later stage of this investigation. It is also more
common to find flows on the verge of separation in practical applications.
 CHAPTER 3

Experiments

The experimental approach is, at present, the only alternative if one wants
to investigate the turbulent plane asymmetric diffuser flow at reasonably high
Reynolds numbers. We are still some years from being able to do a fully resolved
direct numerical simulation (DNS) of this flow at the same Reynolds number
as in this study.

3.1. Experimental facility

The experiments were performed in the specially designed closed loop wind-
tunnel depicted in figure 3.1. A detailed description of the wind-tunnel design
can be found in paper 3 in this thesis.
The approximately three meter long inlet channel has the purpose of pro-
ducing the fully developed turbulent channel flow which is the inlet condition
for the diffuser. The channel height being 30 mm means that the inlet channel
is 100 heights long, which should be more than sufficient for a fully developed
flow. Furthermore, the width of the inlet channel is more than 50 heights in or-
der to reduce end-wall influences on the flow. Large parts of the measurement
section are made in transparent materials to enable the use of optical measure-
ment techniques (see sections 3.2.1 and 3.2.2). The flow after the diffuser is
allowed to relax in a two meter long channel in order to reduce the influences
from the downstream conditions on the diffuser flow. A circular pipe with 400
mm diameter leads the flow back to the fan inlet, this allows smoke to be used
for flow seeding.
Fan Inlet channel Measurement section Relaxation channel Contraction Return channel

3.0 m 1.5 m 2.0 m

Figure 3.1. The wind-tunnel used for the experiments.

8
 3.2. MEASUREMENT TECHNIQUES 9

image pair interrogation areas

cross correlation
function

∆y
∆x
y

Figure 3.2. The PIV analysis procedure.

3.2. Measurement techniques

Besides particle image and laser doppler velocimetry, measurements of pressure
differences have been performed in this study. A digital differential pressure
transducer was used to measure the static pressure distribution through the
diffuser and to monitor the wall shear-stress in the inlet channel using a Preston
tube.

3.2.1. Particle image velocimetry

A major part of the measurements presented in the papers in this thesis were
made using digital particle image velocimetry (PIV). The principle for PIV is
rather straightforward, it is based on two photographic images of an illuminated
plane in the fluid. If the fluid has been mixed with light reflecting particles (so
called seeding particles) these will appear in the images as bright dots. The
two images are taken within a time interval that is short enough to be able to
see the same particles in both images, knowing the time between the images
the velocity of the particles can be calculated from their displacement between
the images. As light source for the illuminated plane it is common to use a
pulsed laser, synchronized with the camera, which can deliver light at much
higher intensities than other light sources, allowing short exposure times for
sharp images. For example, the laser used for the measurements in this thesis
has a maximum power of approximately 40 MW in each light pulse, although
the duration of the pulse is only about 10 ns.
In digital PIV the images are taken with a digital camera. The digital
images are divided into interrogation areas of 2n × 2n pixels, the correspond-
ing interrogation areas from the the image pair are cross-correlated in order to
find the most probable displacement of the particles in each interrogation area.
10 3. EXPERIMENTS

Figure 3.2 gives a schematic view of this procedure. By choosing the interroga-
tion areas to be 2n × 2n pixels the two-dimensional fast Fourier transformation
(FFT) algorithm can be utilized for finding the autocorrelation function, this
decreases the computational effort needed for the image evaluation. The out-
put from PIV is an instantaneous field of velocity vectors which gives direct
information on the spatial structure of the flow. A disadvantage of the PIV
technique is the often poor temporal resolution and the large amount of data
that needs to be stored if many samples are taken in order to get reliable sta-
tistical averages. See section 2.2.2 in Paper 1 on page 27 for details on the PIV
measurements in this particular study.

3.2.2. Laser doppler velocimetry

Laser doppler velocimetry (LDV) was used in this investigation to measure the
velocity component perpendicular to the plane where the PIV measurements
were taken. LDV is primarily based on two physical principles, the Doppler
effect and the heterodyne principle. The Doppler effect is the phenomenon
that an object moving towards a wave source will receive waves with another
frequency than a stationary object would. For instance, if an object is moving
towards the wave source (sending with the frequency f ) with a velocity v, the
object will receive waves with the frequency fobj = (c + v)/λ, where c is the
propagation velocity of the wave and λ is wavelength of the wave. The object
will reflect waves with the shifted frequency fobj , which will have a wavelength
λs = (c − v)/fobj and frequency fs = c/λs when they are received by the
sender. The difference in frequency between the waves sent out and received
by the sender will be ∆f = fs − f = 2cv/(λ(c − v)) ≈ 2v/λ if the velocity of
the object is small relative to the propagation velocity of the wave.
The heterodyne principle says that if two waves are superimposed, the
amplitude of the resultant wave will vary at a frequency equal to the difference
in frequency between the two superimposed waves. This is utilized in LDV by
mixing the reflected wave with the wave that was originally sent out on the
detector surface of a receiver. The receiver will hence measure a signal with an

laser beam splitter Bragg cell focusing lens

detector

fluid flow

Figure 3.3. Sketch of a directional sensitive LDV setup in

backscatter mode.
 3.3. RESULTS 11

0
0 5 10 15 20 25 30 35 40

−3
0.0 3.2 6.4 9.6 12.8 16 10

Figure 3.4. The turbulence kinetic energy K/Ub2 .

amplitude varying with the frequency ∆f . Theoretically, it is not necessary to

make use of the heterodyne principle in order to determine the velocity of the
object but in practice it is very difficult to directly measure the frequency of
the reflected light.
Figure 3.3 shows the principal setup for a directional sensitive LDV-system
in backscatter mode, such as the one used in this study. Here, one has two laser
beams coming in from two different directions, the scattered light from the
upper beam in figure 3.3 will have a higher frequency than the incoming light
and the light from the lower beam will be scattered with a lower frequency.
The detector registers the difference in frequency between the two scattered
beams as a varying light intensity each time a particle passes through the
point of intersection between the two beams. The Bragg cell visible in figure
3.3 changes the frequency of one of the laser beams with a known constant
shift (often with 40 MHz) to allow determination of the flow direction. If no
Bragg cell is used, it is impossible to determine whether the flow is coming
from the top of the figure or from below since the frequency difference between
the reflected light from the upper and lower beams would be the same. See
section 2.2.1 in Paper 1 on page 26 for details on the LDV measurements in
this particular study.

3.3. Results
The measurements were made at a Reynolds number Reτ = uτ H/ν = 2000,
where uτ is the friction velocity in the inlet channel and H is the inlet channel
height. All PIV and LDV measurements where taken in a plane near the
centerline of the diffuser. The separation and reattachment points where found
to be located at a distance of 9 and 31 inlet channel heights downstream of the
diffuser inlet.
Figure 3.4 shows the turbulent kinetic energy K = 0.5u2i of the diffuser
flow, this quantity is composed of velocity measurements in all three spatial
directions. Notice especially the high levels of turbulence kinetic energy that
appear in the shear layer on top of the separated region and the tongue of low
12 3. EXPERIMENTS

0
0 5 10 15 20 25 30 35 40

Figure 3.5. The production rate of turbulence kinetic energy

PH/Ub3 + x/H. Comparison between measurements (◦) and
EARSM computation (—).

turbulence energy emerging from the inlet channel. The measured dataset has
been used to evaluate a RANS computation using an explicit algebraic Rey-
nolds stress turbulence model (EARSM). The computation was performed by
Johan Gullman-Strand using the same code as in the computations presented
in Gullman-Strand et al. (2002). The computational code is based on the finite
element method and the governing equations are implemented through a sym-
bolic computations interface using the femLego toolbox described in Amberg
et al. (1999). The EARSM used is the one described in Wallin & Johansson
(2000). In figure 3.5 the measured production rate of turbulence kinetic energy
is compared to the results of the EARSM computation. It can be seen that
the computation predicts higher production rates in the beginning of the dif-
fuser and this overproduction is believed to result in an overestimation of the
momentum diffusion towards the inclined wall leading to a delayed separation
in the computations. In figure 3.6 the static pressure coefficient on the plane
diffuser wall is shown and the smaller separation in the computation is percep-
tible as a higher pressure recovery and a less pronounced plateau region over
the separation.

3.4. Control
One of the final goals of this study, which will continue for some more years
after this thesis, is to develop techniques for separation control. So far only a
preliminary test using vortex generators has been done. However, the vortex
generator results were very positive. Figure 3.6 shows the pressure distribution
through the diffuser for the case when a spanwise row of 20 mm high vortex
generators were fitted on the inclined wall slightly above the location of mean
flow separation. For comparison the pressure distribution for the case without
vortex generators is included in the figure and the increase in pressure recovery
when using vortex generators is quite substantial bearing in mind also that the
vortex generators themselves cause flow losses.
 3.4. CONTROL 13

0.9
0.8
0.7
0.6
0.5
Cp
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30 35 40 45
x/H
Figure 3.6. The pressure distribution through the diffuser
for the uncontrolled flow (◦), the flow with vortex generators
(3) and for a computation with an EARSM turbulence model
(—).
 CHAPTER 4

Concluding remarks and outlook

The flow in a plane asymmetric diffuser with an 8.5◦ opening angle has been
carefully investigated and a database which can be used for evaluation and
improvements of turbulence models has been compiled. The flow case has been
found to include several important and challenging phenomena, for instance
streamline curvature effects around the diffuser’s upstream corner, rapidly dis-
torted turbulence in the beginning of the diffuser and pressure induced bound-
ary layer separation and reattachment. A special wind-tunnel has been built for
the experiment and tested with special attention paid to the two-dimensionality
of the flow and to the long time mean velocity stability.
In the future the dataset will be complemented with near-wall measure-
ments of the flow especially near the inlet corner of the diffuser where many
phenomena important for the flow development are believed to occur. The
streamline curvature correction described in Wallin & Johansson (2002) will
be implemented and tested in new RANS computations. Another interesting
thing to try, is to see if a differential Reynolds stress model (DRSM) satisfying
strong realizability, such as the one of Sjögren & Johansson (2000), will be
able to predict the near-wall asymptotics better than the DRSM:s tested in
Apsley & Leschziner (1999). Rapidly distorted turbulence which is not in weak
equilibrium is probably also handled better by a DRSM.
The control part of this project has yet merely been touched upon and more
thorough work remains to be done here. The vortex generator test turned out
to be a success and in the future similar approaches will be tested i.e. we believe
that it is more efficient to control the separation by breaking up the spanwise
homogeneity of the flow than only manipulating it in a two-dimensional fashion.
Possible actuators that will be tested are inclined blowing jets, synthetic jets
and plasma actuators. Computational studies of separation control might also
be done using the forth order accurate finite difference DNS code described in
Brüger (2002), and the large eddy simulation technique.

14
 Acknowledgment

First of all I would like to express my gratitude to my supervisor Professor Arne

Johansson for giving me the possibility to be his student and for introducing
me to the challenging field of turbulence research.
I also would like to thank my co-worker and office-mate Björn Lindgren for
introducing me to the field of fluid mechanics experiments. Björn has besides
being a good teacher also been a valuable friend.
I am very grateful to everyone at the department for contributing to a stim-
ulating, enjoyable, friendly and relaxed atmosphere. I particularly appreciate
the (sometimes lengthy) coffee breaks, not only for the coffee but also for the
often interesting discussions. I have learnt a lot about all sorts of things during
these breaks. Special thanks to all friends I have made at the department,
with whom I have sailed, skied, travelled, played indoor bandy and football,
barbecued, partied and much more.
The Swedish Energy Agency are gratefully acknowledged for their financial
support.

15
 Bibliography

Amberg, G., Törnhardt, R. & Winkler, C. 1999 Finite elemet simulations using
symbolic computing. Mathematics and Computers in Simulation 44, 275–274.
Apsley, D. D. & Leschziner, M. A. 1999 Advanced turbulence modelling of sep-
arated flow in a diffuser. Flow, Turbulence and Combustion 63, 81—112.
Brüger, A. 2002 Higher order methods suitable for direct numerical simulation of
flows in complex geometries. Tech. Rep. 2002:07. Lic. thesis, Dept. of Mechanics,
KTH, TRITA-MEK.
Buice, C. U. & Eaton, J. K. 1997 Experimental investigation of flow through an
asymmetric plane diffuser. Tech. Rep.. Department of mechanical engineering,
Stanford university.
Buice, C. U. & Eaton, J. K. 2000 Experimental investigation of flow through an
asymmetric plane diffuser. J. of Fluids Eng. 122, 433–435.
Gullman-Strand, J., Amberg, G. & Johansson, A. V. 2002 Study of separated
flow in an asymmetric diffuser. In Advances in Turbulence IX , pp. 643–646.
Southampton, U.K.
Haller, G. 2002 Kinematic theory of unsteady separation for two-dimensional flows.
Under consideration for publication in J. Fluid Mech.
Hellsten, A. & Rautaheimo, P., ed. 1999 Workshop on refined turbulence mod-
elling. ERCOFTAC/IAHR/COST.
Kaltenbach, H.-J., Fatica, M., Mittal, R., Lund, T. S. & Moin, P. 1999
Study of flow in a planar asymmetric diffuser using large-eddy simulation. J.
Fluid Mech. 390, 151–185.
Obi, S., Aoki, K. & Masuda, S. 1993a Experimental and computational study of
turbulent separating flow in an asymmetric plane diffuser. In Ninth Symp. on
Turbulent Shear Flows, p. 305. Kyoto, Japan.
Obi, S., Ishibashi, N. & Masuda, S. 1997 The mechanism of momentum transfer
enhancement in periodically perturbed turbulent separated flow. In 2nd Int.
Symp. on Turbulence, Heat and Mass Transfer, Delft, The Netherlands, pp.
835–844.
Obi, S., Ohizumi, K., Aoki, K. & Masuda, S. 1993b Turbulent separation control
in a plane asymmetric diffuser by periodic perturbation, pp. 633–642. Elsevier
Science Publishers B.V.

16
 BIBLIOGRAPHY 17

Simpson, R. L., Chew, Y.-T. & Shivaprasad, B. G. 1981 The structure of a

separating turbulent boundary layer. Part 2. Higher order turbulence results. J.
Fluid Mech. 113, 53–73.
Sjögren, T. & Johansson, A. V. 2000 Development and calibration of algebraic
nonlinear models for terms in the Reynolds stress transport equations. Phys.
Fluids 12 (6), 1554–1572.
Wallin, S. & Johansson, A. V. 2000 An explicit algebraic Reynolds stress modell
for incompressible and compressible turbulent flows. J. Fluid Mech. 403, 89–132.
Wallin, S. & Johansson, A. V. 2002 Modeling streamline curvature effects in
explicit algebraic Reynolds stress turbulence models. Int. J. of Heat and Fluid
Flow. 23, 721–730.
 P1
Paper 1
 Measurements in a plane asymmetric diffuser
with 8.5◦ opening angle. Part I: General flow
characteristics
By Björn Lindgren, Olle Törnblom and Arne V. Johansson
Dept. of Mechanics, KTH, SE-100 44 Stockholm, Sweden

To be submitted

The flow in an asymmetric plane diffuser with opening angle 8.5◦ has been
studied in detail experimentally. The inlet condition was fully developed tur-
bulent channel flow at a Reynolds number based on the inlet channel height
and the friction velocity of Reτ = 2000. The mean and fluctuating velocities
have been measured in all directions (streamwise, spanwise and wall-normal).
A separated region is found on the inclined wall with a mean separation point
at 9 and a mean reattachment point at 31 inlet channel heights downstream
the diffuser inlet. Separation never occurs upstream of 5 inlet channel heights
and reattachment never occurs downstream of 35 inlet channel heights. Two
regions of special interest have been observed in the flow. A strong shear-layer
between the separated region and the ”outer” flow fluctuates substantially in
wall-normal direction, interacting with the large scale vortices in the separated
region. It is also influenced by the large adverse pressure-gradient and the
rapid change in flow direction at the diffuser inlet corner on the inclined wall.
The static wall pressure through the diffuser was measured and compared to a
case where vortex generators are used to suppress the separation. The pressure
coefficient was found to be almost 10% higher at the diffuser outlet for the case
with vortex generators.

1. Introduction
Separation and the need for controlling the separation are essential in many
applications of fluid flows. Particularly diffuser flows, i.e. duct flows subjected
to a positive pressure gradient in the streamwise direction, are very common
in many industrial applications. The aim of this study is to determine the
characteristics of the separated plane asymmetric diffuser flow, to provide a
reliable data-base for the turbulence modeling community and to identify flow
mechanisms that can be used in future schemes for efficient separation control.
Many studies have been performed on geometry induced separated flow
and adverse pressure gradient flows with separation. The backward facing step
21
22 B. Lindgren, O. Törnblom and A. V. Johansson

and the blunt plate have been investigated by e.g. Cherry et al. (1984), Eaton
& Johnston (1981) and Kiya & Sasaki (1983). Ruderich & Fernholz (1975)
performed an investigation on a normal bluff plate with a splitter plate using
pulsed hot-wire anemometry. Some investigations closer to ours on turbulent
boundary layer separated flow are reported by e.g. Perry & Fairlie (1975)
Dianat & Castro (1991) and Angele (2002). There is also a review on the topic
by Simpson (1989). Here, however we will concentrate on the plane asymmetric
diffuser flow with fully developed turbulent channel flow as inlet condition.
The problem of computing the flow in this choice of geometry or to realize
it experimentally are both very challenging tasks. The range of scales, the high
turbulence intensities and the high strains encountered in this flow together
with a large separated region makes it a demanding, and thereby well suited,
problem for testing and developing turbulence models. The fairly simple geom-
etry, for which a numerical conformal mapping for an orthogonal curvilinear
grid can be derived, see e.g. Brüger (2002), simplifies the possibility to perform
LES and DNS calculations in the future.
In the first part of our study, presented in this paper, we will concen-
trate on the mean flow characteristics of the flow in the diffuser and its nearby
surroundings. The main reason is to provide good data of mean flow prop-
erties, such as mean and rms velocities, in the three spatial directions along
the streamwise centerline of the diffuser where the flow is expected to be most
two-dimensional. We will also investigate the extension of the separated region
and look at measures such as the back-flow coefficient determining the features
of the flow in the separated region.
In the second part of the study titled, Measurements in a plane asymmetric
diffuser with an opening angle of 8.5◦ opening angle. Part II: Turbulence char-
acteristics, we present properties especially interesting for turbulence modeling
comparisons. We will look at characteristics of existing models and compare
them to the behavior of the flow in the diffuser.
Flow in this geometry has previously been studied experimentally by e.g.
Obi et al. (1993a), Obi et al. (1993b), Obi et al. (1997), Buice & Eaton (1997)
and Buice & Eaton (2000). The opening angle of the diffuser was in all these
studies slightly larger, 10◦ , than that (8.5◦ ) used in the present work. The
reason for choosing a smaller diffuser opening angle was here to reduce the
size of the separated region, and thereby, in combination with a high aspect
ratio of the diffuser, achieve a high degree of two-dimensionality of the flow
and moderate unsteadiness of the separation and reattachment points. From
a control study point of view a smaller separation bubble is desired to keep
actuator amplitudes moderate. It is also more common to find flows at the
verge of separation in practical applications, e.g. draft tubes, wind-tunnel
diffusers, ventilation ducts etc.
 Measurements in a plane asymmetric diffuser. Part I 23

Furthermore the intention of this study was to measure all necessary ve-
locity components to be able to calculate the turbulent kinetic energy and to
form the anisotropy tensor throughout the diffuser. These are very useful mea-
sures for the turbulence modeling community when new improved models are
developed and evaluated. These data, mainly obtained from PIV images with
a rather coarse grid, (each interrogation area is 4 × 4 mm2 ), are to be comple-
mented with measurements of velocity profiles with more data points to resolve
the flow near the boundaries.
Among the above mentioned studies, Obi et al. (1993a) studied the plane
asymmetric diffuser flow using a single component LDV. The measurements
were then compared to calculations with two different turbulence models, (stan-
dard k − ε and a basic version of SMC (second moment closure)). The agree-
ment between the experimental data and the calculations were not very good
although the SMC was better than the k − ε model. They concluded that it is
essential to capture the redistribution of energy between the different compo-
nents, in this highly anisotropic flow, in the turbulence models.
Later Obi et al. (1993b) continued to study the same flow experimentally by
inserting a periodic perturbation to the flow. The perturbation was generated
by periodic blowing and suction through a slit in the spanwise direction. They
investigated the influence of the perturbation frequency on the size of the sepa-
ration bubble and found an optimum non-dimensional frequency, based on the
inlet channel height and the inlet channel centerline velocity, around St = 0.03.
They also concluded that at this perturbation frequency the enhancement of
transport of momentum across the diffuser was maximized.
In Obi et al. (1997) they further investigated the effect of the perturbation
on the production of the turbulent Reynolds shear stress separating the con-
tributions from the perturbation and the mean flow. They found that at the
optimum frequency the production of Reynolds shear stress and the interaction
between the mean flow and the perturbation were both enhanced.
Buice & Eaton (1997), Buice & Eaton (2000) also made an experimental
investigation in the same geometry (with 10◦ ) increasing the diffuser aspect
ratio and used primarily pulsed hot-wires to determine the flow in the separated
region. Thermal tufts were used to find the separation and reattachment points.
They also measured the wall shear-stress using pulsed wall-wire probes. Their
results were used for comparison in a Large Eddy Simulation by Kaltenbach
et al. (1999).
Brunet et al. (1997) made an experimental investigation on pressure effects
on turbulent flow in a plane asymmetric diffuser with a smooth (10◦ ) inclined
wall. They compared their results to calculations using the k − ε model and
found that the model was particularly deficient in capturing the behavior of
the Reynolds shear-stress.
24 B. Lindgren, O. Törnblom and A. V. Johansson

Fan Inlet channel Measurement section Relaxation channel Contraction Return channel

3.0 m 1.5 m 2.0 m

Figure 1. The wind-tunnel used in the experiments. The

flow is circulating counter clockwise.

In the introduction to the second part of this paper there is a review of the
numerous numerical and modeling efforts so far made using the plane asym-
metric diffuser geometry. Some of these efforts have used the Obi et al. (1993a)
and foremost the Buice & Eaton (1997) experimental data for comparison.
In future studies in the present set-up, control of the flow separation in
the diffuser will also be investigated in co-operation with developers of control
schemes. There have been a number of investigations involving control in var-
ious diffuser flow configurations, e.g. Coller et al. (2000); Obi et al. (1993b,
1997) but there is still much work to be done in this field.

2. Experimental setup
2.1. Description of the wind-tunnel
The experiments were performed in a closed loop wind-tunnel built specifically
for this investigation. The wind-tunnel can be seen in figure 1. It consists of a
blowing centrifugal fan delivering 11 kW of power followed by a section trans-
forming the rectangular cross section shape of the blower outlet to another one
with much higher aspect ratio. In this section splitter plates and screens ensure
an even distribution of the fluid over the cross section area. The transformer is
followed by a straight duct (settling chamber) which contains two screens at its
upstream end to further even out mean flow variations. The settling chamber
is followed by a two-dimensional contraction. The contraction further evens
out mean flow variations and decreases the cross section height to 30 mm.
The contraction is followed by a turbulence generating grid and a 3.2 m
long channel with a cross section area (width to height) of 1525 × 30 mm2 .
The length to height ratio of the channel is thus larger than 100 ensuring
fully developed channel flow at the outlet (see Comte-Bellot (1965)). This is
important in order to obtain a well defined inlet condition to the diffuser which
can also be easily produced in numerical calculations. At the downstream end
of the inlet channel, the end walls (limiting the spanwise width of the channel)
 Measurements in a plane asymmetric diffuser. Part I 25

24.8 H

o
8.5 y

4.7 H
Recirculation zone x
Turbulent channel inflow
H

Figure 2. The measurement section of the wind-tunnel. A

part of the inlet channel is seen to the left and part of the
outlet channel is seen to the right. The view is from the top
as referred to figure 1.

are perforated with 2 mm holes. This section is 100 mm long and the end
wall boundary layers are removed through these holes by means of suction to
prevent end wall boundary layer separation in the diffuser. Such a separation
would destroy the two-dimensionality of the mean flow.
The inlet channel is followed by the diffuser. The diffuser has an inclined
wall on one side and a straight wall on the other. The end walls are straight
and equipped with three rows of vortex generators to further minimize the
risk for boundary layer separation. The inclined wall has an angle of 8.5◦ ,
see figure 2. At the upstream corner of the inclined wall there is a radius of
100 mm to prevent separation at this corner. The inclined wall is made of an
aluminum sandwich plate. Pressure taps are located along the centerline in
the downstream direction at every 100 mm. These pressure taps were used to
confirm that the pressure on both the straight and inclined walls were similar
at all downstream measurement positions. The straight wall and the end walls
are made of Plexiglas to allow the use of optical measurement techniques such
as LDV and PIV. The straight wall is also equipped with pressure taps along
the centerline in the downstream direction at an interval of 25 mm. There
are also pressure taps in the spanwise direction 100 mm upstream the diffuser
inlet on each side at 100 mm interval. These pressure taps are used to check
the two-dimensionality of the incoming flow. The high aspect ratio, 50 at the
diffuser inlet, is crucial to achieve a high degree of spanwise uniformity.
The diffuser is followed by an outlet channel which is 141 mm high (4.7
inlet channel heights (4.7H)) and 2.5 m long. The purpose of this channel is to
avoid upstream influence on the flow from devises located further downstream.
This channel is partly made of Plexiglas to facilitate measurements and partly
of Plywood. Here, there are also two hatches giving access to the inside of the
tunnel. Along the extent where it is made of Plexiglas there are also pressure
taps along the spanwise centerline with the same interval as in the diffuser.
26 B. Lindgren, O. Törnblom and A. V. Johansson

Diffuser
Laser beams

x2 LDV-Optics
Traversing table
x1

Figure 3. A sketch of the LDV measurement setup. The

scattered light from the particles are received through the same
optics as the emitted light.

The outlet channel is followed by a three-dimensional transformer changing

the aspect ratio of the cross section area towards unity. The contraction is
followed by a heat exchanger which is necessary in a closed return wind-tunnel
to keep the temperature steady. A temperature sensor is located in the outlet
channel and a computer logged the temperature which was found to be steady
within ±0.5 ◦ C. This is sufficient since the experiments performed here are not
very temperature sensitive due to the fact that optical measurement techniques
(LDV and PIV) are used instead of hot-wire anemometry.
Following the heat exchanger the cross section is converted from a rectan-
gular to a circular shape. A pipe with 400 mm diameter then leads the flow
back to the fan. At the end of the pipe the seeding particles (in this case smoke,
see section 2.2.3) are injected into the flow. Just in front of the fan there is a
slit ensuring atmospheric pressure at the fan inlet. This is important since it
provides a constant reference pressure.

2.2. Measurement techniques

The measurement techniques used in this investigation were primarily Particle
Image Velocimetry, PIV and Laser Doppler Velocimetry, LDV. Pressures were
measured using a Furness Control FCO 510 differential pressure transducer
with an accuracy of 0.25% of full scale (2000 Pa).

2.2.1. LDV measurements

The LDV is a one component FlowLite system from Dantec. It has a He-Ne
Laser of 10 mW emitting light with a wave length of 632.8 nm. The light beams
have a Gaussian intensity distribution and the beam diameter was 3.23 mm
between the beam expander and the focusing lens. The measurement volume
 Measurements in a plane asymmetric diffuser. Part I 27
Camera

Flow

Laser sheet

Figure 4. A sketch of the PIV measurement setup. The laser

sheet is orientated horizontally, i.e. perpendicularly to the
paper.

diameter using a lens with a focal length of 310 mm and a beam separation of
73 mm was 77 µm. The measurement volume length was 0.66 mm. In figure 3
a sketch of the LDV setup is shown. Note that the scattered frequency shifted
light from the particles are received through the same optics as the emitted
light, i.e. backscatter receiver technique.
The LDV was calibrated against a rotating wheel and used to measure flow
components in the spanwise direction, i.e. W and wrms . The measurements of
the spanwise component were made along the spanwise centerline of the diffuser
at intervals of 50 mm. A total of 29 profiles were measured. In the wall normal
direction, the interval varied with proximity to a wall and the downstream
position, from 0.1 mm to 5 mm. These data were later also interpolated to
give values on a common measurement grid. This grid have a spacing between
points of 10 mm in the downstream direction and 2 mm in the wall normal
direction.
Between 25000 and 100000 samples were taken at each measurement point
for the ensemble average with the sampling rate varying between 200 and 1000
Hz. The time for collecting data was set to a minimum of 120 s at each point.

2.2.2. PIV measurements

In the streamwise and wall normal directions the velocity components were
measured using a PIV system from Dantec. In figure 4 a sketch of the PIV
setup is shown. The laser sheet is orientated horizontally and perpendicular to
the paper in the sketch, and the camera takes images from above.
28 B. Lindgren, O. Törnblom and A. V. Johansson

The system consists of a Spectra Physics 400 mJ double-pulse Nd-Yag

laser, a megaplus ES1.0 digital double frame camera from Kodak and hardware
processing and software from Dantec.
The infrared light emitted by the laser is doubled in frequency to obtain
a visible light of 532 nm. CCD chips are most sensitive to light around this
wave length. The duration of the pulses emitted by the laser is on the order of
a few nano seconds. The light beam is guided from the laser through a Dantec
flexible arm allowing changes in the light sheet position and orientation. A
cylindrical lens at the end of the flexible arm converts the circular Gaussian
beam into a light sheet with a thickness on the order of 1 mm. The intensity
distribution of the light in the sheet is approximately Gaussian.
The CCD camera has a light sensitive chip of 1018 × 1008 pixels and an
integrated memory chip allowing two images to be taken within 1 µs. The
gray-scale resolution of the CCD chip is 256 levels. A f=60 mm lens was used
together with a large aperture (f2.8) giving a very light sensitive image over an
area of about the size of the outlet channel height in both directions.
The hardware, including all of the image processing, was supplied by Dan-
tec. Through the hardware the camera and laser are synchronized. The image
processing allowed for 32 double frame samples to be taken into each burst at
a sampling rate of approximately 7.5 Hz depending on the maximum velocity
in the image. Extra time for image processing was then needed to empty the
memory containing the images before another burst could be recorded.
The software used to evaluate the images and calculate the velocity vectors
was FlowManager 2.12 from Dantec. The software used sub-pixel interpolation
to increase the velocity resolution assuming a Gaussian distribution. The data-
sets exported from the FlowManager software were, the velocity components,
the streamwise and wall-normal position in the laser sheet plane, the peak ratio,
i.e. the ratio between the two largest peaks in the cross correlation function
between the two images, and an indicator, (0 or 1), indicating if the velocity
vector was rejected according to the rejection criteria set in the software. A new
evaluation with extended rejection criteria was later performed with Matlab
based software.
The image size of the PIV was calibrated against a ruler that was inserted
into the laser sheet before each measurement. Data was collected from 2048
images to form reliable mean and rms values. The size of the interrogation areas
were 32×32 pixels. In physical space this is equivalent to approximately 4.5×4.5
mm2 . A 50% overlap between the interrogation areas were used throughout
the measurements. In the inlet channel just upstream of the diffuser a more
highly resolved image was used with another lens, (180 mm), to obtain more
velocity vectors in this narrow channel. With the 1018 × 1008 pixel CCD chip,
an interrogation area of 32 × 32 pixels and 50% overlap, the images contain
62×62 velocity vectors. Adding the fact that each interrogation area is 4.5×4.5
 Measurements in a plane asymmetric diffuser. Part I 29

mm2 the physical size of the images are 145 × 145 mm2 which just covers the
outlet channel height of 141 mm.
The time between frames was optimized for each set of pictures and varied
with the downstream position. In general each measurement of 2048 frames
took about 30 min to collect. To cover the entire measurement section PIV
images were obtained at 11 downstream positions. A GUI Matlab script was
used to match the streamwise and wall normal positions from the 11 sets of
data.
These data were then, together with the LDV wall normal profiles of the
spanwise velocity component interpolated on a common grid with a mesh size
of 10 mm by 2 mm in the streamwise and wall-normal directions respectively.
This grid is used whenever contour plots of the flow in the entire measurement
section are shown in the section 3 of this paper.

2.2.3. Seeding particles

In both the PIV and LDV measurements the air was seeded with smoke from
a smoke-generator ZR12-AL from Jem using a mixture of glycerol and water.
By varying the amount of water the density of the smoke can be optimized to
give good results. The smoke was inserted through the reference pressure slit
just upstream of the fan and was sufficiently spread out in the measurement
section due to the mixing in the fan and in the inlet channel. No attempt
to measure the physical size of the seeding particles has been made but from
previous investigation using the same type of smoke we can conclude that they
typically are a few µm in diameter.

2.2.4. Error sources in LDV and PIV measurements

It is very important to study the errors in the velocity measurements using the
PIV and LDV techniques. Kristian Angele has investigated the error sources
found in digital PIV (which is used here). His results will appear in a pa-
per currently under preparation and the references here to his investigation is
through private communication, Angele found that even carefully made mea-
surements can have large errors but most of the errors can be eliminated by
using appropriate validation criteria. PIV random error sources can e.g. be
noise which can come from background disturbances, poor image contrast or
resolution. Using a digital FFT to process the data will introduce small random
uncertainties.
Another error source is the so called loss of pairs which means that a parti-
cle is only present in one of the images in an image pair. This is evidently more
likely to happen for fast particles which gives a bias towards lower velocities.
If the particle density is too low this problem will increase since the number of
particles that form the correlation pair are fewer. This problem can be relieved
30 B. Lindgren, O. Törnblom and A. V. Johansson

by applying a weight function to the correlation plane, see e.g. Raffel et al.
(1997).
A large velocity gradient over an interrogation area will not only cause
the problem of integration but it will also decrease the signal to noise ratio.
Following a criterion based on the velocity gradient, the time between images
and interrogation area size in physical units and pixels introduced by Adrian
& Keane (1992) makes sure this error source is kept small.
Finally peak-locking errors, i.e. the tendency of instantaneous particle dis-
placements towards integer pixel values, is thoroughly investigated by Angele.
An interrogation area is usually too small to give more than 8-16 possible dis-
crete velocities. Therefore a curve-fit is applied in the correlation plane to get
sub-pixel accuracy. This does not always eliminate the error which can be seen
in the probability density distributions. Angele found that the limiting factor
for peak-locking is that the ratio between the discretization velocity and the
rms-velocity, i.e. the number of velocity peaks distributed over the probability
density distribution. A symptom of peak-locking error is a shaky urms -profile.
There are some symptoms of peak-locking problems in our measurements, see
figure 7 as the mean velocity variation in our case is rather high.
LDV error sources may be reflections close to the wall decreasing the signal
to noise ratio substantially and making it impossible to measure. Very close to
a wall vibrations of the wall causes erroneous velocity peaks in the probability
density distribution also largely affecting the measurements.
When measuring with an LDV system it is important to collect statistically
independent samples i.e a single burst shall only be sampled once. It is also
important to correct errors from high velocity bias, i.e. high velocity particles
are more often detected than low velocity particles. The correction can be made
by weighting the sample with its residence time, i.e. the time the particle are
within the measurement volume. Arrival time averaging can be used to correct
for low velocity bias at low data rates, see e.g. Fischer et al. (2001).
Averaging effects over the measurement volume while measuring in velocity
gradients is an error source that can be quite substantial but it can be corrected
for. Durst et al. (1998) derived an equations for correcting the mean and
fluctuating velocity assuming elliptical shape of the measurement volume and
that the scattering particles are mono-disperse.
Measuring at high mean velocity with limited fluctuating velocity can be
difficult if the mixing and filter frequencies are limited in the LDV system. This
is a result of the low resolution of the velocity probability density distribution
which causes large errors in the measured fluctuating velocity. Our LDV system
have these limitations and therefore we only used the LDV for the spanwise
velocity component which has zero mean velocity. Many of the difficulties
described above such as velocity gradient effects are therefore avoided.
 Measurements in a plane asymmetric diffuser. Part I 31
y
H

0
0 5 10 15 20 25 30 35 40
x
H
° ±
Figure 5. Streamwise mean velocity, Ufig = 10 UUb + H x
, at
8 streamwise positions (x/H = 5, 10, 15, 20, 25, 30, 35 & 40).
Dashed lines indicate zero level for each streamwise velocity
profile.

3. Results
We here focus on the mean velocity components and fluctuation intensities.
Results are also presented for the turbulence kinetic energy as well as for the
stream function and back-flow coefficient, which both characterize the mean
properties of the separated region at the inclined wall. Furthermore, the static
wall-pressure along the spanwise centerline downstream through the diffuser is
presented from which the pressure coefficient can be calculated.

3.1. Velocity components

In this study we measured all three velocity components and their fluctua-
tions using PIV for the streamwise and wall-normal components and LDV for
the spanwise component. Here, results for all these quantities are presented
with the exception of the mean spanwise component which is zero within the
measurement accuracy. The coordinate system is here defined as x in the
”streamwise” direction, y in the ”wall-normal” direction and z in the spanwise
direction. The corresponding average velocities are denoted respectively by U ,
V and W . The y-component is of course only normal to the lower, straight
wall but because of the relatively small opening angle, we prefer to here to
refer to the x and y velocity components as streamwise and wall-normal. In
all figures the spatial coordinates are normalized with the inlet channel height,
H = 0.03 m, and the velocities are normalized with the inlet channel bulk
velocity, Ub = 20.0 m/s, defined as

Z H
1
Ub = U dy. (8)
H 0
32 B. Lindgren, O. Törnblom and A. V. Johansson

3.1.1. Streamwise mean and fluctuating velocities

The diffuser can be divided into a region below the maximum velocity peak,
i.e. close to the straight wall, and a region, above the velocity peak, including
the separated flow region closer to the inclined wall, see figure 5. On the side
close to the inclined wall an inflection point is created almost immediately
downstream the inlet in the streamwise velocity profile. The inflectional profile
is enhanced downstream in the diffuser until the inclined part of the wall ends.
Thereafter it relaxes towards the same profile as on the straight wall although
at the maximum downstream position in this study the complete streamwise
velocity profile is far from symmetric. The peak in streamwise velocity diverges
slightly towards the centerline of the diffuser as it decreases in strength, but this
process is very slow and still at our most downstream measurement position
the streamwise velocity maximum is much closer to the straight wall than the
inclined wall.
The inflection streamwise velocity profile creates a strong shear-layer which
is important as it distributes momentum from the outer flow towards the sepa-
rated region. The non-zero components in the Reynolds stress tensor, (−u2rms ,
2 2
−vrms , −wrms and −uv) all have their peaks within this shear-layer.
The very high fluctuation level of this flow with instantaneous velocity
profiles very far from the mean value is illustrated in figure 6 through a sequence
of instantaneous velocity fields in the streamwise, wall-normal plane. We can
observe very large velocity variations which are of large scale in this region
which includes the first (upstream) part of the separated region. High speed
fluid (shown in white) emerging from the inlet channel is deflected towards the
inclined wall directly downstream the inlet, but occasionally, an almost jet like
character (figure 6q) can be observed with a resulting large separation region.
This motion can, with a larger opening angle, lead to instantaneous separation
along the straight wall. As will be shown later this is not the case here and we
can conclude that the 8.5◦ opening angle is small enough to also avoid straight
wall separation.
The motion of this high velocity region is of course intimately linked to
the shedding process in the separated region. Unfortunately the sampling fre-
quency of the measurement system is too low to be able to detect the shedding
process. A sampling frequency 10 times higher than the 7.5 Hz available with
our PIV system should be enough to resolve this process that is expected
(loosely from Obi et al. (1993b) and Kaltenbach et al. (1999)) to have a mean
frequency of about 20 Hz although a large variation in frequency is expected
from these earlier studies and also judged from our instantaneous velocity im-
ages.
The fluctuating streamwise velocity is in general very large in this flow
as was shown above in the instantaneous velocity fields. The fully developed
turbulent inlet flow generates a urms distribution which is symmetric and with
 Measurements in a plane asymmetric diffuser. Part I 33
y
H
3 (a) (b) (c)
2.5
2
1.5
1
0.5
0
3 (d) (e) (f)
2.5
2
1.5
1
0.5
0
3 (g) (h) (i)
2.5
2
1.5
1
0.5
0
3 (j) (k) (l)
2.5
2
1.5
1
0.5
0
3 (m) (n) (o)
2.5
2
1.5
1
0.5
0
3 (p) (q) (r)
2.5
2
1.5
1
0.5
0
10 11 12 13 14 10 11 12 13 14 10 11 12 13 14
x
H

Figure 6. Time sequence of instantaneous velocity fields (U-

2
V plane). ∆t = 15 s between images. White: High positive
velocity fluid (17 m/s). Black: Negative velocity fluid (-1 m/s).
34 B. Lindgren, O. Törnblom and A. V. Johansson
y
H

0
0 5 10 15 20 25 30 35 40
x
H

Figure° 7. Root mean square ±of streamwise fluctuating ve-

locity, urms,fig = 40 uUrms
b
+H x
, at 8 streamwise positions
(x/H = 5, 10, 15, 20, 25, 30, 35 & 40).

peaks close to each wall and a local minimum at the centerline where the
streamwise velocity gradient is zero. The development of these two peaks are
rather different as the flow propagates downstream through the diffuser. The
peak close to the straight wall is first slightly reduced downstream and eventu-
ally it is almost overtaken by the growing local maximum emerging from the
upper part of the inlet channel. This peak grows and reaches its maximum
at about x/H = 20 − 25, see figure 7. The location of this peak detaches
from the inclined wall and follows the strong shear-layer outside the separated
region. Eventually the fluctuating streamwise velocity reaches a symmetric
distribution across the outlet channel with small variations along the profile.
Thereby a distribution with two peaks at the inlet develops towards a distri-
bution with one peak at the centerline. As the flow in the outlet channel is
developing, the turbulent channel flow distribution will eventually be recovered
far downstream.

3.1.2. Wall-normal mean and fluctuating velocities

The wall-normal velocity, V (i.e. normal to the straight wall, and almost
normal to the inclined wall) is very small in most parts of the diffuser. Close
to the corner at the diffuser inlet where there is a rather abrupt change in the
direction of the wall causes the flow to turn leading to locally high values of
the wall-normal velocity (Not shown in figure 8). Further studies of this region
is required with a better spatial resolution.
The wall-normal velocity decreases downstream (until about x/H = 15 −
20) where its peak slowly starts to grow again due to the decreasing height of
the separation bubble. The location of this peak follows the strong shear-layer
outside the separated region, see figure 8. The maximum value is found close
to the centerline of the diffuser and the magnitude of the maximum increases
up to approximately x/H = 30. Further downstream the peak follows the
centerline of the outflow channel while decreasing in magnitude. The presence
 Measurements in a plane asymmetric diffuser. Part I 35
y
H

0
0 5 10 15 20 25 30 35 40
x
H
° ±
Figure 8. Wall-normal mean velocity, Vfig = 100 UVb + H x
,
at 8 streamwise positions (x/H = 5, 10, 15, 20, 25, 30, 35 & 40).
Dashed lines indicate zero velocity for each streamwise velocity
profile.
y
H

0
0 5 10 15 20 25 30 35 40
x
H

Figure°9. Root mean square ±of wall-normal fluctuating ve-

locity, vrms,fig = 40 vrms x
Ub + H , at 8 streamwise positions
(x/H = 5, 10, 15, 20, 25, 30, 35 & 40).

of a vertical velocity in the core region of the outflow channel is of course due
the asymmetry of the streamwise velocity component.
The fluctuating wall-normal, vrms , velocity behaves much like the stream-
wise fluctuating velocity with peaks on either side of the centerline in the inlet
channel and with a growth of the peak closest to the inclined wall and a reduc-
tion of the other peak. The variation of the wall-normal fluctuating velocity
is however substantially smaller than for its streamwise counterpart and it
reaches a symmetric profile at quite an early stage with a maximum at the
outlet channel centerline. The maximum is reached at a downstream position
approximately equal to the point where the inclined part of the wall ends as is
also the case for the wall-normal mean velocity.

3.1.3. Spanwise fluctuations, Reynolds stress and turbulence kinetic energy

The development of the fluctuating spanwise velocity, wrms , downstream thr-
ough the diffuser resembles very much its streamwise counterpart. The main
36 B. Lindgren, O. Törnblom and A. V. Johansson
y
H

0
0 5 10 15 20 25 30 35 40
x
H

Figure° 10. Root mean square± of spanwise fluctuating ve-

locity, wrms,fig = 40 wUrms
b
+H x
, at 8 streamwise positions
(x/H = 5, 10, 15, 20, 25, 30, 35 & 40).
y
H

0
0 5 10 15 20 25 30 35 40
x
H

Figure °11. Reynolds stress ±in the streamwise wall-normal

plane, uv fig = 1000 Uuv2 + H x
, at 8 streamwise positions
b
(x/H = 5, 10, 15, 20, 25, 30, 35 & 40).

difference lies in the smaller magnitude of the spanwise component. The wall-
normal location of the maximum is also slightly different, with the maximum
for the spanwise fluctuating velocity closer to the inclined wall, but still within
the shear-layer outside the separated region. The maximum in streamwise
direction is found at x/H = 20 − 25 which is similar to that for urms .
If the spanwise component cannot be measured in an experiment, it seems
possible to obtain a first order estimate by rescaling of the streamwise fluctu-
ating velocity. The main reason for making such an estimate would be to be
able to estimate the turbulence kinetic energy without having to measure all
three velocity components. This has been done in other studies, however, it is
of course more appropriate to directly measure all components to get a correct
value for the turbulence kinetic energy that can be used in comparison with
results from e.g. turbulence model predictions.
The Reynolds shear-stress is a very important quantity since it redistributes
momentum from high velocity regions to low velocity regions. A high level of
Reynolds shear-stress is thus instrumental in suppressing separation, see e.g.
 Measurements in a plane asymmetric diffuser. Part I 37
y
H

0
0 5 10 15 20 25 30 35 40
x
H
° ±
Figure 12. Turbulence kinetic energy, Kfig = 400 UK2 + H x
,
b
at 8 streamwise positions (x/H = 5, 10, 15, 20, 25, 30, 35 & 40).

Obi et al. (1993b). The amount of Reynolds shear-stress scales with turbulence
kinetic energy which makes a flow like this, with fully developed turbulent
channel flow as inlet condition, quite capable of handling a strong adverse
pressure gradient and thereby reduce the size of the separated region.
The Reynolds shear-stress in the streamwise wall-normal plane, uv, is an-
tisymmetric in the inlet channel. With this coordinate system the negative
peak is near the straight wall and the positive peak closer to the inclined wall.
As the flow develops downstream in the diffuser the positive peak grows while
the negative peak is suppressed, see figure 11. The maximum Reynolds shear-
stress is found at x/H = 20 − 25 and thereafter it decreases in magnitude.
The location of the positive peak follows the location of maximum shear in
the shear-layer outside the separated region. At the end of the measurement
region the Reynolds shear-stress attains a symmetric profile with the maximum
located near the outflow channel centerline.
The turbulence kinetic energy, K, is a primary quantity for comparisons
with turbulence model predictions since most models rely on transport equa-
tions for this quantity where one or more terms have to be modeled. Com-
parisons with numerical results can be improved if direct and accurate mea-
surements of the turbulence kinetic energy are made. New insights that can
assist in the development of new turbulence models can thereby be obtained.
The dissipation rate, ε, can here however not be measured directly although
some attempts using the PIV technique have been made e.g. by Baldi et al.
(2002). The dissipation rate is needed for a complete comparison with turbu-
lence models where it is included, directly or in some other form (ω, τ ), through
a transport equation.
In figure 12 the turbulence kinetic energy at eight streamwise positions
is plotted. It is for natural reasons, similar in character to the individual
intensities shown in previous figures. It is defined as
1 2 2 2
¡
K= u + vrms + wrms . (9)
2 rms
38 B. Lindgren, O. Törnblom and A. V. Johansson

The maximum value is located at about x/H = 20 − 25 in the shear-layer

outside the separated region as for the individual fluctuating components. The
kinetic energy profile does not quite reach a truly uniform shape at our most
downstream position.
It should be noticed that our measurements are not resolving the very near
wall regions. Therefore our conclusions are excluding phenomena occurring in
the immediate vicinity of the walls. In the future these measurements will be
complemented by new experiments covering the near wall regions as well as a
more detailed investigation of the upstream corner on the inclined wall where
the expanding part of the diffuser begins. There we expect to find locally high
levels of turbulence intensity and anisotropy as well as a strong streamline
curvature. Some clues to these statements can be seen if the whole velocity
field is studied rather than the isolated velocity profiles presented here.

3.2. Characterization of the separated region

To further study the separated region, extending the information gained by
looking at the velocity profiles the stream-function, generating streamlines and
the back-flow coefficient are useful tools. The back-flow is also studied through
a sequence of instantaneous images.

3.2.1. Stream-function
The stream-function is here defined as
Z y
1
Ψ(x, y) = 1 − U (x, y)dy. (10)
HUb 0
In figure 13, constant values of the stream function (streamlines) are plotted
throughout the measurement region. This definition of the stream function
gives a value of Ψ = 0 at the dividing streamline. The dividing streamline sep-
arates the (averaged) recirculation zone from the outer flow. The two positions
where the dividing streamline reaches the ”upper” wall are the mean separa-
tion and reattachment points, respectively. From figure 13, the mean separation
point is found to be located at 9 channel heights downstream the diffuser inlet
(x/H ≈ 9) and the mean reattachment point is located at x/H ≈ 31.
In figure 13 the
√ gray-scale levels represent a measure of the speed in the
diffuser, i.e. here U 2 + V 2 , with an increment of 2 m/s. The figure shows how
the flow with maximum velocity first is deflected at the inlet corner towards the
inclined wall but when it approaches the separation point it is deflected back
towards the straight wall. Thereafter, there is a very slow relocation of the
maximum towards the center of the outlet channel. It can also be noted that
the decrease in speed is slower through the diffuser compared to an attached
flow case. This follows from the constriction caused by the separated region,
which decreases the adverse pressure gradient and slows down the retardation
of the flow speed and spreads it out over a larger downstream length.
 Measurements in a plane asymmetric diffuser. Part I 39
y
H

4 0
0.2
0.4
2
0.6
0.8
0
0 5 10 15 20 25 30 35 40
x
H

Figure 13. Streamlines shown as white curves. The stream

function is integrated from the inclined wall. Gray-scale back-
ground with separating black curves shows the speed, with a
contour increment of 2 m/s.

3.2.2. Back-flow coefficient

A very interesting quantity is the back-flow coefficient which relates the number
of samples with negative velocity (along the inclined wall) to the samples with
positive velocity. A back-flow coefficient of 1 means that all samples have
negative velocity and a back-flow coefficient of zero thus means that all samples
have positive velocity. The back-flow coefficient then reads
N
1 X 1 − sgn(uk (x, y) cos(α) + vk (x, y) sin(α))
χ(x, y) = , (11)
N 2
k=1

where N is the total number of samples, sgn is the sign function, uk the stream-
wise velocity in the k th sample, vk the wall-normal velocity in the k th sample
and α = 8.5◦ the diffuser opening angle.
First, we notice in figure 14 that along the straight wall there seems to be
no samples with back-flow although the rather poor resolution might allow for
a very thin separation bubble very close to the wall. It is important that there
is no separation on the straight wall since this may destroy the ”stability” of
the separation bubble on the inclined wall with a separation altering side from
the inclined wall to the straight wall back and forth. Our choice of opening
angle, 8.5◦ is thus small enough to avoid this kind of flow state.
Focusing our attention on the separation bubble on the inclined wall the
back-flow coefficient gives us the streamwise locations on the inclined wall for
the most upstream instantaneous separation point that in figure 14 is shown
to be x/H ≈ 5. The flow is separated 80% of the time downstream x/H ≈ 14.
In real time the separation point moves back and forth along the inclined wall
and at some occasions the flow is completely attached. A back-flow coefficient
of 0.5 gives the mean separation point, already shown to be x/H ≈ 9 from
the dividing streamline, and the mean reattachment point at x/H ≈ 31. The
40 B. Lindgren, O. Törnblom and A. V. Johansson
y
H

4 0.7 0.5 0.3

0.1

0
0 5 10 15 20 25 30 35 40
x
H

Figure 14. back-flow coefficient, χ.

reattachment point never moves downstream of x/H ≈ 35 where the flow thus
always is attached.
To study the separation more in detail in time, figure 15 shows the in-
stantaneous back-flow in a sequence of images. Black is large negative velocity
along the inclined wall and white is positive velocity. It is clear from this se-
quence that sometimes the flow is almost fully attached (e.g. 15i and 15m) and
sometimes there are large regions of flow with negative velocity. The formation
of spanwise vortices can be seen in some of the frames in the time sequence of
figure 15.

3.3. Static wall-pressure distribution

The static wall-pressure has been measured along the spanwise centerline in
the downstream direction. Measurements on both the inclined and straight
walls were performed but with a much better spatial resolution on the straight
wall. The measurements on the inclined wall were made to confirm that the
static pressure is constant in the wall-normal direction (within our measure-
ment accuracy) through the diffuser. This is true for most of the downstream
measurement positions except at the upstream corner where the rapid change
in wall direction leads to large curvature of the streamlines and consequently
pressure differences between the two sides.
Shown in figure 16 is the pressure coefficient, Cp , defined as

pw (x) − pw (x = 0)
Cp (x) = 1 2
, (12)
2 ρUb

where pw is the static wall-pressure and ρ the density of the air is shown.
The diamonds in figure 16 represent the pressure coefficient with undisturbed
flow. The start of the separation can be seen where the diamonds deviate from
the circles. The following rather flat region is caused by the constriction to
the outer flow produced by the separated region that decreases the pressure
gradient.
 Measurements in a plane asymmetric diffuser. Part I 41
y
H
3 (a) (b) (c)
2.5
2
1.5
1
0.5
0
3 (d) (e) (f)
2.5
2
1.5
1
0.5
0
3 (g) (h) (i)
2.5
2
1.5
1
0.5
0
3 (j) (k) (l)
2.5
2
1.5
1
0.5
0
3
(m) (n) (o)
2.5
2
1.5
1
0.5
0
3
(p) (q) (r)
2.5
2
1.5
1
0.5
0
10 11 12 13 14 10 11 12 13 14 10 11 12 13 14
x
H

Figure 15. Time sequence of images of instantaneous nega-

2
tive streamwise velocity (x-y plane). ∆t = 15 s between im-
ages. White: Zero or positive streamwise velocity. Black:
Strong streamwise negative velocity.
42 B. Lindgren, O. Törnblom and A. V. Johansson

0.9

0.8

0.7

0.6
Cp 0.5

0.4

0.3

0.2

0.1

0 5 10 15 20 25 30 35 40
x
H

Figure 16. Static wall-pressure coefficient, Cp . Diamonds

represent flow without vortex generators. Circles represent
flow with vortex generators.

The circles represent a reference case where vortex generators have been
installed to suppress the separated region. These vortex generators are posi-
tioned in a spanwise row at a downstream location of about x/H = 7. Their
height is 20 mm and the angle of the vertical plates to the mean flow is 20◦ .
The spanwise spacing between the vortex generators is 100 mm which gives
a total of 15 generators covering the entire span of the diffuser. As can be
seen in figure 16 the pressure coefficient continues to rise where the flat re-
gion, indicating separation in the uncontrolled case (diamonds), starts. The
final pressure recovery increases with vortex generators from approximately
0.73 to 0.80 despite the extra pressure loss generated by the vortex generators
themselves through the streamwise vortices they create.

4. Concluding remarks
The flow in an asymmetric plane diffuser has been studied with extra attention
paid to ensuring good two-dimensionality of the mean flow. The focus in this
paper is on the overall flow characteristics, i.e. mean velocities, turbulence
intensities and Reynolds stresses, together with an investigation of the sepa-
rated region using instantaneous velocity fields, streamlines and the back flow
coefficient.
The streamwise mean velocity, U , gives us indications of where in the dif-
fuser flow the most interesting flow features can be found. These are apart from
 Measurements in a plane asymmetric diffuser. Part I 43

the separated region at the inclined wall the strong shear-layer outside the sep-
arated region and the region just downstream of the diffuser inlet. Studies of
instantaneous velocity fields reveal that a region of high velocity fluid emerges
from the inlet channel, that gives almost a ”jet-like” structure to the flow with
large amplitude and large scale coherence of the fluctuating field, interacting
with the large scale structures in the separated region. There is a certain risk
for a small separation on the straight wall when the flow is attached along the
inclined ditto but the rather moderate opening angle used here (8.5◦ ) does not
produce large enough gradients for this to happen. The inflectional stream-
wise velocity profile found around the edge of the separated region creates a
strong shear-layer that is important as it distributes energy from the outer
flow towards the separated region. The non-zero components in the Reynolds
stress tensor, (−u2rms , −vrms
2 2
, −wrms and −uv) all have their peaks in this
shear-layer.
The wall-normal velocity, V , is very small in most parts of the diffuser,
but close to the upstream corner at the diffuser inlet where the change in the
direction of the wall is rather abrupt, locally high values of the wall-normal
velocity and high streamline curvature are found. This region is not studied
in detail in this paper since a better spatial resolution in the measurements is
required.
The streamwise fluctuating velocity, urms , increases in magnitude down-
stream in the diffuser and peaks between x/H = 20 − 25. The typical, urms ,
two peak profile in the inlet channel flow is redistributed towards a one peak
profile with the peak following the shear-layer outside the separated region and
eventually its location in the outflow channel is close to the centerline. The
other two components of the fluctuating velocity, vrms and wrms , are subjected
to the same transformation as the streamwise component, although the mag-
nitude is smaller with vrms (about half the size of urms ) and the magnitude of
wrms lies in between vrms and urms . The location of the peaks in wall-normal
direction is also somewhat different from the streamwise case.
The Reynolds shear-stress, −uv, in the streamwise wall-normal plane is
large in the shear-layer outside the separated region. Since the Reynolds shear-
stress contributes to a redistribution of momentum toward regions with lower
mean velocity it helps to delay the separation in the diffuser. The turbulent
inlet channel flow contains a fair amount of turbulent kinetic energy which
makes this diffuser flow more resistant to separation than if a laminar flow
would have been chosen as the inlet condition.
The flow detaches on the inclined wall at approximately 9 inlet channel
heights downstream from the diffuser inlet. The separation point moves, in
time, up and down along the inclined wall but never above x/H = 5 as shown
by the back flow coefficient figure. This movement is coupled to the shedding
process in the separating region. The reattachment point is of course also
44 B. Lindgren, O. Törnblom and A. V. Johansson

moving in its location with the furthest downstream location of 35 inlet channel
heights. The mean reattachment point was found to be at x/H = 31. By
studying the instantaneous images of back flow in the separated region we get
a fairly good view of the vortices created and swept downstream through the
diffuser although the sampling rate of our PIV system is too low to be able to
determine a mean ”shedding” frequency. Using other measurement techniques
or combining the PIV with a detection/trigging signal it may be possible to
determine this frequency. One should of course keep in mind that one cannot
expect a well-defined single frequency in a phenomenon like this, but rather a
fairly broad-band peak.
The static wall pressure was measured at the spanwise centerline in the
streamwise direction along the straight wall. An array of vortex generators
was applied as a reference case of diffuser flow without separation. Comparing
the two curves from flow with and without vortex generators clearly shows the
separated region as a plateau present only in the case without vortex generators
caused by the constriction of the outer flow by the separated region. The
pressure coefficient at the measurement region outlet, achieved with vortex
generators (0.80) is almost 10% higher than for the original case (0.72).

5. Acknowledgment
The authors would like to thank Ulf Landén for aiding in the manufacturing of
the flow device and Kristian Angele for introducing us into the world of PIV.
The Swedish Research Council and The Swedish Energy Agency are gratefully
acknowledged for their financial support.

References

Adrian, R. & Keane, R. 1992 Theory of cross-correlation in PIV. Applied scientific

research 49, 191–215.
Angele, K. P. 2002 Pressure-based scaling in a separating turbulent APG boundary
layer. In Advances in turbulence IX (ed. I. Castro, P. E. Hancock & T. G.
Thomas), pp. 639–642. CIMNE, Barcelona, Spain.
Baldi, S., Hann, D. & Yanneskis, M. 2002 On the measurement of turbulence
energy dissipation in stirred vessels with PIV techniques. In 11th Int. Symp.
on Applications of Laser Techniques to Fluid Mech. Instituto superior técnico,
Center for innovation, technology, and policy research, Lisbon, Portugal.
Brüger, A. 2002 Higher order methods suitable for direct numerical simulation of
flows in complex geometries. Tech. Rep. 2002:07. Lic. thesis, Dept. of Mechanics,
KTH, TRITA-MEK.
Brunet, L., Calzalbou, J. B., Chassaing, P. & Jervase, L. 1997 Experimental
and computational study of pressure effects on turbulent flow in an asymmetric
plane diffuser. In 11th Symp. on Turbulent Shear Flows, Grenoble, France, pp.
114–119.
 Measurements in a plane asymmetric diffuser. Part I 45

Buice, C. U. & Eaton, J. K. 1997 Experimental investigation of flow through an

asymmetric plane diffuser. Tech. Rep.. Department of mechanical engineering,
Stanford university.
Buice, C. U. & Eaton, J. K. 2000 Experimental investigation of flow through an
asymmetric plane diffuser. J. of Fluids Eng. 122, 433–435.
Cherry, N. J., Hillier, R. & Latour, M. E. M. P. 1984 Unsteady measurements
in a separated and reattaching flow. J. Fluid Mech. 144, 13–46.
Coller, B. D., Noack, B. R., Narayanan, S., Banaszuk, A. & Khibnik, A. I.
2000 Reduced-basis model for active separation control in a planar diffuser flow.
In AIAA Fluids 2000, 19-22 June 2000, Denver, CO, USA, AIAA-2000-2562 .
Comte-Bellot, G. 1965 Écoulement turbulent entre deux paroix parallèles. Publi-
cations scientifiques et techniques 419. Ministère de l’air, 2, Avenue de la Porte-
d’Issy, Paris.
Dianat, M. & Castro, I. P. 1991 Turbulence in a separated boundary layer. J.
Fluid Mech. 226, 91–123.
Durst, F., Fischer, M., Jovanović, J. & Kikura, H. 1998 Methods to set-up
and investigate low Reynolds number, fully developed turbulent plane channel
flows. ASME J. Fluids Eng. 120, 496.
Eaton, J. K. & Johnston, J. P. 1981 A review of research on subsonic turbulent
flow reattachment. AIAA J. 19, 1093–1100.
Fischer, M., Jovanović, J. & Durst, F. 2001 Reynolds number effects in near-wall
region of turbulent channel flows. Phys. Fluids 13 (6), 1755–1767.
Kaltenbach, H.-J., Fatica, M., Mittal, R., Lund, T. S. & Moin, P. 1999
Study of flow in a planar asymmetric diffuser using large-eddy simulation. J.
Fluid Mech. 390, 151–185.
Kiya, M. & Sasaki, K. 1983 Structure of a turbulent separation bubble. J. Fluid
Mech. 137, 83–113.
Obi, S., Aoki, K. & Masuda, S. 1993a Experimental and computational study of
turbulent separating flow in an asymmetric plane diffuser. In Ninth Symp. on
Turbulent Shear Flows, p. 305. Kyoto, Japan.
Obi, S., Ishibashi, N. & Masuda, S. 1997 The mechanism of momentum transfer
enhancement in periodically perturbed turbulent separated flow. In 2nd Int.
Symp. on Turbulence, Heat and Mass Transfer, Delft, The Netherlands, pp.
835–844.
Obi, S., Ohizumi, K., Aoki, K. & Masuda, S. 1993b Turbulent separation control
in a plane asymmetric diffuser by periodic perturbation, pp. 633–642. Elsevier
Science Publishers B.V.
Perry, A. E. & Fairlie, B. D. 1975 A study of turbulent boundary-layer separation
and reattachment. J. Fluid Mech. 69, 657–672.
Raffel, M., Willert, C. & Kompenhans, J. 1997 Particle Image Velocimetry, A
practical guide. Springer Verlag.
Ruderich, R. & Fernholz, H. H. 1975 An experimental investigation of a turbulent
shear flow with separation, reverse flow, and reattachment. J. Fluid Mech. 163,
283–322.
46 B. Lindgren, O. Törnblom and A. V. Johansson

Simpson, R. L. 1989 Turbulent boundary-layer separation. Ann. Rev. Fluid Mech.

21, 205–234.
Paper 2
P2
 Measurements in a plane asymmetric diffuser
with 8.5◦ opening angle.
Part II: Comparison with model predictions for
turbulence characteristics
By Olle Törnblom, Björn Lindgren, Johan Gullman-Strand
and Arne V. Johansson
Dept. of Mechanics, KTH, SE-100 44 Stockholm, Sweden

To be submitted

The separating turbulent flow in a plane asymmetric diffuser is studied exper-

imentally. Careful measurements of the velocities have been made in all three
spatial directions. The Reynolds number based on the inlet channel height and
friction velocity was Reτ =2000. Emphasis is put on the study of the turbu-
lence in the diffuser. The Reynolds-stress anisotropies and the corresponding
anisotropy invariants are calculated and studied. The production terms in the
transport equations for the Reynolds stresses are investigated. Comparisons
with RANS-solutions using the Wallin & Johansson (2000) explicit algebraic
Reynolds stress model (EARSM) are made. It is observed that the EARSM
under-predicts the size of the separation bubble. The size of the predicted
bubble is approximately 60% of the measured one and the main reason for this
is the large level of wall-normal turbulence intensity predicted by the EARSM
at the diffuser inlet.

1. Introduction
This is the second part of a study of the flow in an asymmetric plane diffuser
with an opening angle of 8.5◦ . Here, we will concentrate our work on evaluating
measurement data for quantities that are particularly interesting in turbulence
modelling. The aim is to provide new insight into the dynamics that are essen-
tial in the modelling of this flow, and thereby help in testing existing models
or in the development of new models. We will compare the measurement data
with an Explicit algebraic Reynolds stress model (EARSM) calculation, based
on the code presented in Gullman-Strand (2002).
The first part of this study entitled Measurements in a plane asymmet-
ric diffuser with an 8.5◦ opening angle. Part I: General flow characteristics
(Lindgren et al. (2002)) concentrated on the general flow characteristics of this
49
50 O. Törnblom, B. Lindgren, J. Gullman-Strand and A. V. Johansson

24.8 H

8.5 y

4.7 H
Recirculation zone x
Turbulent channel inflow
H

Figure 1. The measurement section of the wind-tunnel. A

part of the inlet channel is seen to the left and part of the
outlet channel is seen to the right.

diffuser flow such as mean velocity and turbulence intensity in the three spatial
directions. It also includes measures of the extension of the separation bubble,
its character and strength and the average pressure distribution. For informa-
tion hereabout we refer to that paper. In this paper we focus on turbulence
statistics which are compared with predictions based on RANS computations
carried out by the third author.
The diffuser used in both these studies has one inclined wall with an open-
ing angle of 8.5◦ . The opposite wall is straight, see figure 1. The diffuser is
preceded by a 3.2 m channel with a height of 30 mm and a width of 1525 mm.
This ensures fully developed turbulent channel flow as inlet condition into the
diffuser (see e.g. Comte-Bellot (1965)). It is important when computations and
experiments are compared, to have a well defined inlet condition. The diffuser
is followed by a 2.5 m long outlet channel which is 141 mm high and 1520 mm
wide. The purpose is here to minimize upstream influence from disturbances
from other wind-tunnel parts located further downstream. The primary mea-
surement techniques used are Particle Image Velocimetry in the streamwise
wall-normal plane and Laser Doppler Velocimetry in the spanwise direction.
An advantage of these measurement techniques is that they can measure both
direction and absolute value of a fluid particle which is necessary when there
is backflow involved. The Reynolds number based on the inlet channel height
and the friction velocity is 2000.
The flow in the diffuser is characterized by a long thin separation bubble
located at the inclined wall. The separation point is located about 9 channel
heights downstream the diffuser inlet and the reattachment point is located 31
inlet channel heights downstream the inlet channel (see Lindgren et al. (2002)).
The maximum thickness of the separation bubble is approximately 1.6 inlet
channel heights occurring at the end of the diffuser.
Some numerical studies on this geometry have been performed previously.
However, all these studies used a geometry with a larger opening angle (10◦ )
 Measurements in a plane asymmetric diffuser. Part II 51

than that in the present study. An extensive numerical study of the plane
asymmetric diffuser flow was made by Kaltenbach et al. (1999), who performed
a large eddy simulation at a Reynolds number of 1000 based on the inlet channel
height and the inlet friction velocity. Their data showed good agreement with
the Buice & Eaton (1997) experimental data for velocity profiles. The point of
separation also agreed well but some discrepancy was found in the location of
the reattachment point. A possible reason for this can be the relatively small
spanwise width of the computational domain which may tend to artificially
enhance spanwise coherence of large scale structures. Kaltenbach et al. (1999)
found that the sub-grid scale model plays an essential role to calculate the flow
correctly, since sub-grid stresses are a major contribution to the wall-shear
stress. The sub-grid scale model must also adapt to the increase in turbulence
level in the downstream part of the diffuser.
Other numerical studies involving more or less advanced closures based
on eddy-viscosity models, (differential) Reynolds stress models (RSM) and
EARSM models have been performed by a number of research groups. For
instance, Apsley & Leschziner (1999) tested a number of linear and non-linear
eddy viscosity models as well as differential stress-transport models. They
found that strain dependent coefficients and anisotropy resolving closures are
needed. However no models tested were capable to resolve all flow features in
the diffuser. Apsley & Leschziner (1999) also points out the possibility to en-
counter problems related to the ”flapping” motion of the unsteady separation.
In an Ercoftac workshop (Hellsten & Rautaheimo (1999)), different numer-
ical approaches with varying turbulence models were used and compared to the
Buice & Eaton (1997) data-base. Models used comprised k − ε, k − ω, RSM
and LES. The agreement was, for the more simple models, in general fairly
poor due to the complex flow in the diffuser.
The plane asymmetric diffuser has also been used as a test case for commer-
cial codes. The investigation performed by Iaccarino (2000) aimed at finding
the limits of the versatile commercial codes in this complex geometry. The
codes tested were CFX, Fluent and Star-CD. Two turbulence models were
tested, (k − ¯ and v 2 − f ) for the three codes. The results were compared to the
Obi et al. (1993) and Buice & Eaton (1997) data-bases. The k − ¯ model was
unable to capture the recirculation zone but the v 2 − f model did so with an
accuracy in separation length of 6%. The agreement for the friction coefficient
was also fairly good.
As is seen above, an ample amount of numerical tests of closures in plane
asymmetric diffusers exists already today. Some of the major challenges in
turbulence modelling are related to near-wall turbulence and pressure-gradient
induced separation, phenomena which are represented in an ideally generic
manner in the plane asymmetric diffuser flow. The present choice of opening
angle ensures a separation-free flow near the straight wall, and the aspect ratio
52 O. Törnblom, B. Lindgren, J. Gullman-Strand and A. V. Johansson

together with end-wall boundary layer control measures ensure a high degree
of spanwise uniformity. This makes the present case ideal for detailed tests
of turbulence modelling aspects that may require a high degree of accuracy in
the turbulence statistics to evaluate differences in modelling approaches among
high-level single-point closures, and sub-grid scale models in the LES-approach.
The presently created data-base, containing information on all velocity compo-
nents and related second order statistics, would, for instance, be well suited for
tests of modern nonlinear RSM:s and newly developed approaches, including
curvature corrections etc, within the concept of EARSM.

2. Turbulence models
2.1. The RANS equations
The by far most common approach to compute turbulent flows is to decompose
the velocity and pressure field into a mean and a fluctuating part, and form
equations for the mean velocity and single-point turbulence statistics. This
approach is called the Reynolds decomposition and a detailed explanation of
this standard technique can be found in e.g. Johansson & Burden (1999). The
(incompressible) RANS equations,
∂Ui ∂Ui 1 ∂P ∂ ° ±
+ Uj =− + 2νSij − u0i u0j (1)
∂t ∂xj ρ ∂xi ∂xj
∂Ui
= 0, (2)
∂xi
which govern the ensemble averaged velocities (Ui ) and the average pressure (P )
are identical to the corresponding Navier-Stokes equations for the instantaneous
velocities and pressure with one important exception, the Reynolds stress term.
Since the flow considered is incompressible, we will consider the kinematic
Reynolds stress tensor −Rij ≡ −u0i u0j , where u0i are the velocity fluctuations.
The aim of single point closures is to construct a closed set of equations for
this quantity, to be inserted into equations 1. An ensemble average is herein
denoted with a bar over the respective quantity.

2.2. Eddy-viscosity based two-equation models

In CFD for engineering applications, the by far most commonly used method
of closing the RANS set of equations is the eddy-viscosity based two-equation
model approach. The eddy-viscosity hypothesis, introduced by the french
physicist and mathematician V.J. Boussinesq in 1877, approximates the in-
creased diffusivity due to turbulence by using a turbulent viscosity (νT ), anal-
ogous to the well known kinematic viscosity for a Newtonian fluid. Using the
eddy-viscosity hypothesis the Reynolds stresses are approximated by
2
−u0i u0j = 2νT Sij − Kδij . (3)
3
 Measurements in a plane asymmetric diffuser. Part II 53

Where K ≡ u0i u0i /2 is the turbulence kinetic energy and Sij ≡ (Ui,j + Uj,i )/2
is the mean rate of strain tensor. By simple dimensional analysis it is easy to
see that in order to estimate the turbulent viscosity one needs to know at least
one time (or velocity) scale and one length scale of the turbulence. In so called
two-equation models these scales are determined by solving transport equations
for two turbulence quantities. Usually one of these quantities is the turbulence
kinetic energy K and the other can be e.g. the dissipation rate of turbulence
kinetic energy ε, or the inverse timescale of the most energetic eddies ω.
Such standard two-equation models can be reasonably accurate for predict-
ing attached flows without large influences from mean flow streamline curvature
or system rotation and the inclusion of two transport equations for turbulence
quantities means that some history effects of the turbulence can be captured.
In the Boussinesq hypothesis there is no dependence of the stresses on the
rotation rate tensor (Ωij ≡ (Ui,j − Uj,i )/2), a deficiency which can influence
the ability to predict e.g. separating flows. A more elaborate description of
eddy-viscosity based two-equation models can be found in e.g. Wilcox (1993)
and Johansson & Burden (1999).

2.3. Differential Reynolds stress models (DRSM)

A straightforward way to generalize the modelling approach as compared to
the two-equation model approach, is to introduce the transport equations for
the Reynolds stresses
DRij
= Pij − εij + Πij + Dij . (4)
Dt
Such a model is referred to as a differential Reynolds stress model (DRSM)
and pioneering work on this kind of model was done by Launder et al. (1975).
Due to the symmetric nature of Rij , six equations plus one for a turbulence
‘length’ scale, has to be solved in 3D flow problems. The production term Pij
is explicit in the Reynolds stresses and the mean velocity gradient tensor and
does not have to be modelled within this modelling context.
The trace of the dissipation rate tensor εij is usually determined by an
extra transport equation for ε which is also used to determine the turbulent
‘length’ scale, and the anisotropy of the dissipation rate is often assumed to be
negligible or modelled through use of the Reynolds stress anisotropies.
The pressure strain rate tensor Πij is a correlation between the fluctuating
pressure and the fluctuating strain rate. This term represents the intercompo-
nent redistribution of the Reynolds stresses and its modelling is a key element
in this type of closure. The models are based on solutions of the Poisson equa-
tion for the fluctuating pressure. The diffusion term Dij is normally modelled
with a gradient diffusion formulation.
In a DRSM much more of the turbulence physics are of course captured as
compared to a two-equation eddy-viscosity model, but at the prize of five more
54 O. Törnblom, B. Lindgren, J. Gullman-Strand and A. V. Johansson

equations to solve. As an alternative, the performance of a two-equation model

can be improved if the eddy-viscosity approach is abandoned and the trans-
port equations for the Reynolds stresses are approximated with an algebraic
equation for the Reynolds stress anisotropy tensor.

2.4. Explicit algebraic Reynolds stress models (EARSM)

From equation 4 an equation for the Reynolds stress anisotropy tensor can be
derived. The Reynolds stress anisotropy tensor is defined as
u0i u0j 2
aij ≡ − δij . (5)
K 3
Assuming that the flow is in so called ’weak equilibrium’ one can discard the ad-
vection and diffusion terms in the transport equation for the anisotropy tensor.
Weak equilibrium means that the timescale on which the anisotropy relaxes to
some quasi-equilibrium state prescribed by the mean flow and the turbulent
scales, is small. If the pressure-strain rate and dissipation rate anisotropy are
modelled in terms of Sij , Ωij , aij and the turbulence velocity and length-scales
(e.g. K and ε), the weak equilibrium assumption implies that the Reynolds
stress anisotropy is completely determined by the local values of the mean
strain and rotation rate tensors (normalized by the turbulence time scale), i.e.
fij (a, S∗ , Ω∗ ) = 0. (6)
Equation 6 represents an implicit relation between the anisotropy tensor (a)
and the normalized strain (S∗ ) and rotation rate (Ω∗ ) tensors. The weak
equilibrium assumption does not hold in slowly distorted turbulence where
P/ε ¼ 1, e.g. in the outer part of a boundary layer or in the center of a jet or
a channel-flow.
Using the weak equilibrium assumption together with isotropic assumption
for the dissipation rate tensor and linear model for the pressure strain rate
tensor one can derive an explicit algebraic equation for the anisotropy (see
e.g. Gatski & Speziale (1993) and Wallin & Johansson (2000)), i.e.
aij = aij (S∗ , Ω∗ ). (7)
This equation depends exclusively on the mean flow, or more precisely on the
mean strain-rate and rotation-rate tensors. Such an expression requires very
little effort to be evaluated, so the increase in computational effort for an
EARSM compared to a standard two-equation model is almost negligible. The
explicit expression is also a good way of ensuring robustness of the computa-
tional scheme.
The EARSM has several advantages over an ordinary eddy-viscosity based
two-equation model, where
∗
aij = 2Cµ Sij . (8)
 Measurements in a plane asymmetric diffuser. Part II 55

The EARSM has been shown to improve near-wall behavior and thereby re-
duce the need for near-wall damping, as compared to eddy-viscosity models.
Effects of system rotation and streamline curvature can also be captured in the
EARSM:s through extensions analyzed by e.g. Girimaji (1997) and Wallin &
Johansson (2002). Curvature corrections are not included in the model used
for comparisons herein.

3. Results
In this section results from the analysis of the turbulence data are presented
and comparisons with computations are made. For a detailed description of the
experiment, the mean flow and the Reynolds stresses, consult Lindgren et al.
(2002) and for details on the computational aspects, consult Gullman-Strand
(2002).
The numerical results have been obtained using a finite element code,
created by an automated code generation procedure, first described by Am-
berg et al. (1999). The system of equations were the RANS equations with
the EARSM by Wallin & Johansson (2000) combined with Wilcox low-Re
K − ω. The RANS equations were solved in a time-dependent fractional step
scheme described by Guermond & Quartapelle (1997) and a decomposition of
ω = ω̃ + ωw decreased the demand of grid resolution close to the walls. A
more detailed description of the code generation procedure, formulation of the
equations and solution strategy can be found in Gullman-Strand (2002).
The geometry used in the computations were identical to the experiments
with respect to diffuser angle and height ratio but with inlet and outlet lengths
of 10 x/H and 60 x/H respectively. The mesh was a structured triangular
grid with 318 nodes in the streamwise direction, of which 100 were located in
the diffuser and 75 nodes stretched in the wall-normal direction. The short
inlet channel was possible since the turbulent channel inlet conditions were
calculated by the same code but in a separate channel geometry, consistent
with the guidelines of Hellsten & Rautaheimo (1999) for the 10◦ case.
To facilitate interpretation most of the data presented in this section have
been plotted in the diffuser geometry, in all these plots the axis scales are x/H
and y/H in the horizontal and vertical directions respectively, where H is the
inlet channel height. Velocities are normalized with Ub , being the inlet channel
bulk velocity.

3.1. Mean velocities

A comparison of profiles of the streamwise average velocity can be seen in
figure 2, the most noticeable difference between the computation and the mea-
surements is that the separation bubble in the computation is much thinner.
The maximum thickness of the recirculation zone is only 60% of that of the
56 O. Törnblom, B. Lindgren, J. Gullman-Strand and A. V. Johansson

0
0 5 10 15 20 25 30 35 40

Figure 2. The mean velocity in the streamwise direction for

experiment (◦) and EARSM (—); 10 UUb + Hx
.

0
0 5 10 15 20 25 30 35 40

Figure 3. The mean velocity in the wall-normal direction for

experiment (◦) and EARSM (—); 100 UVb + H x
.

0
0 5 10 15 20 25 30 35 40

Figure 4. The turbulence kinetic energy for experiment (◦)

and EARSM (—); 400 UK2 + H
x
.
b

measured separation. This discrepancy has a large influence on the whole flow-
field. Since the smaller separation bubble constitutes a smaller displacement
the whole flow-field is shifted towards the inclined wall as compared to the
measured field. The experimentally determined separation point is located at
x/H = 9 while in the computation it is at x/H = 11. The height of the com-
puted bubble also grows much slower in the downstream direction than the
measured. The reattachment points are located at x/H = 31 and x/H = 27
for the experiment and the computation respectively.
The wall-normal mean velocity in figure 3 is highly affected by the smaller
separation zone in the simulation since the flow follows the inclined wall more
closely. Accordingly it is larger than in the experiments in the diverging part
of the diffuser and smaller in the exit channel where the flow is attached.
 Measurements in a plane asymmetric diffuser. Part II 57

4 (a)
2

0
0 5 10 15 20 25 30 35 40

4 (b)
2

0
0 5 10 15 20 25 30 35 40

4 (c)
2

0
0 5 10 15 20 25 30 35 40

4 (d)
2

0
0 5 10 15 20 25 30 35 40

Figure 5. Measured (◦) and computed (—) variance of tur-

bulent fluctuations and Reynolds shear stress; (a) u0 2 , (b) v 0 2 ,
u0i u0j
(c) w0 2 and (d) u0 v 0 , displayed as 500 Ub2
+ x
H.

3.2. Turbulence intensities and Reynolds shear stress

There is good agreement for the turbulence kinetic energy K (see Figure 4) at
the first station, at x/H = 5, when neither the simulated nor the experimental
flow has separated. The agreement then deteriorates further downstream. In
general the turbulent kinetic energy is smaller in the simulation, something
which might be attributed to the lower levels of shear in the region outside the
separation bubble. The mean velocity profiles at x/H = 35 in figure 2 are very
similar but the simulated K-levels at the same position are much smaller than
the measured. This indicates that the supply of turbulent kinetic energy to
this position from upstream positions is underestimated.
Looking closer into the components of K, shown in figure 5a–d, one can
see that the computed v 0 2 is larger than the measured at the beginning of the
58 O. Törnblom, B. Lindgren, J. Gullman-Strand and A. V. Johansson

0
0 5 10 15 20 25 30 35 40

Figure 6. Measured Reynolds stress production rates; P11

P H x
(—), P22 (- - -) and P12 (· · · ), displayed as 1000 Uij3 + H .
b

0
0 5 10 15 20 25 30 35 40

Figure 7. Calculated Reynolds stress production rates; P11

P H x
(—), P22 (- - -) and P12 (· · · ), displayed as 1000 Uij3 + H .
b

diffuser, at the same location u0 2 is slightly under-estimated in the computa-

tions while the peak heights of w0 2 agree well. The experimental data show
broader and higher variance peaks in the shear-layer over the separated region
at the end of the diverging section than the computations, which is natural due
to the larger separation bubble, while the peak near the plane wall is more ac-
centuated in the computations. This is extra obvious in the profiles of w0 2 .The
agreement of u0 v 0 between the measurements and the computations is good ex-
cept at the first two positions, where it is over-estimated in the computations
near the inclined wall.

3.3. Reynolds stress production

The turbulent stresses in a flow produce new turbulence by interaction with
the mean rates of strain and rotation
∂Uj ∂Ui
Pij = −u0i u0k − u0j u0k . (9)
∂xk ∂xk
In the plane asymmetric diffuser flow, there is homogeneity in the spanwise
direction, so only P11 , P22 and P12 = P21 are non-zero. Hence, all turbulent
energy in the spanwise component (w0 w0 ), which is fairly high (see figure 5)
has been redistributed from the two other components via the pressure strain
rate term in equation 4, illustrating the importance of adequate modelling of
Πij .
Figure 6 shows the three non-zero components of the production rate ten-
sor. The highest production rates are found in the 11-component in the strong
shear-layer above the separation bubble. This is due to the high level of mean
 Measurements in a plane asymmetric diffuser. Part II 59

0
0 5 10 15 20 25 30 35 40

Figure 8. Turbulence kinetic energy production rate, P; ex-

periment (◦), EARSM (—), displayed as 1500 PH
U3
+H x
.
b

shear (∂U/∂y) in combination with the peak in u0 v 0 at this location cf. figure
5. Almost no production of the 11-component takes place at the location of the
maximum peak for the streamwise velocity, which is consistent with the small
mean shear in that region. For the P12 -component the maximum peak is lo-
cated at the same position as for the P11 -component, also due to the high mean
shear, but its strength is only about 60% of the 11-peak. The P12 -component
cross the zero level at the y-position where the streamwise mean velocity peaks
and has a minimum in the ’boundary layer’ on the straight wall. Compared to
the other two components the P22 -component is very small. At some locations
this component attains negative values due to the deceleration of the flow along
the plane wall. At the most downstream profile all production components are
very small due to the very flat mean velocity profile here.
Figure 7 shows the computed components of the production rate tensor.
All three components are significantly overestimated at the two most upstream
positions, while the agreement is better further downstream.
In figure 8, a comparison is made for the production rates of turbulence
kinetic energy (P). In the first profiles at x/H = 5 and 10 the EARSM sub-
stantially overestimates the production rate. The peak levels are almost twice
those of the experiment at the most upstream position. However, since the K
levels in this region agree well, the dissipation rate must also be overestimated.
The shift of the EARSM profiles towards the inclined wall can, of course, be ob-
served in P too. Further downstream the predicted production rate decreases
in amplitude faster than the measured and this can explain why the produced
quantity, K, is underestimated in the downstream part of the diffuser and in
the exit channel.

3.4. Reynolds stress anisotropy tensor

The aij -tensor has two independent scalar measures which are invariant to the
choice of coordinate system. These are defined as
IIa ≡ aij aji and IIIa ≡ aij ajk aki , (10)
and are referred to as the second and third anisotropy invariants respectively.
The experimental determination of these scalar measures is quite sensitive since
60 O. Törnblom, B. Lindgren, J. Gullman-Strand and A. V. Johansson

0
0 5 10 15 20 25 30 35 40

Figure 9. The second invariant of the turbulence anisotropy

x
tensor for experiment (◦) and EARSM (—); 5IIa + H .

an error in the rms of the fluctuating velocity is taken to the fourth and sixth
power in IIa and IIIa , respectively.
Figure 9 shows the variation of IIa over the cross section at different
streamwise positions in the diffuser. Throughout the whole diffuser IIa has
maxima near the walls due to the damping of the wall-normal turbulence in-
tensity. A minimum is also found in all IIa -profiles at the same y-position as
the maximum U -velocity. The low levels of shear here allows the turbulence
to relax and become more isotropic. The position of the maximum in IIa lies
directly above the minimum in the positive y-direction. This is in between the
locations of the maximum shear and the maximum velocity.
Figure 10 shows the four non-zero Reynolds stress anisotropy components.
The streamwise component a11 is generally the largest, the only exception to
this is found far downstream close to the wall were the spanwise component is
larger. The wall-normal component a22 is negative at all positions, signifying
a v 2 content smaller than 2/3 of the kinetic energy. The perhaps most obvious
discrepancy of the computed data in figure 11 as compared to the measurements
is that the computed a33 is identically zero making a22 a mirror image of a11 ,
this approximation in the EARSM is not altogether valid for this flow according
to the measurements.
The second invariant of the Reynolds stress anisotropy tensor IIa is com-
pared in figure 9. At the first station, where the computed profile of the
turbulence kinetic energy agrees well with the experimental values, the sec-
ond invariant is severely underestimated. The large difference between the
streamwise and the wall-normal turbulence components (c.f. figure 10) is not
captured correctly. The position and value of the minimum near the center is
well predicted but the high levels of anisotropy on either sides of the minimum
are underestimated. This underestimation of the anisotropy comes from an
overestimation of the wall-normal velocity fluctuations (vrms ) by the EARSM
in the region near the upstream corner of the inclined wall. The fact that the
EARSM overestimates the vrms , and also u0 v 0 slightly, in the beginning of the
diffuser is probably the reason for the delayed separation. This overestimation
can in turn be a consequence of the omission of streamline curvature effects
in this particular model. An addition of a curvature correction would reduce
 Measurements in a plane asymmetric diffuser. Part II 61

0
0 5 10 15 20 25 30 35 40

Figure 10. Measured Reynolds stress anisotropies; a11 (—),

x
a22 (- - -), a33 (- · -) and a12 (· · · ), displayed as 5aij + H .
4

0
0 5 10 15 20 25 30 35 40

Figure 11. Computed Reynolds stress anisotropies; a11 (—),

x
a22 (- - -), a33 (- · -) and a12 (· · · ), displayed as 5aij + H .

the turbulence levels near the upstream corner of the inclined wall. Further
downstream at the stations x/H =10–25 the maximum values of the experi-
mental anisotropy is gradually decreased and the quantitative agreement with
the computations is increased, but on the other hand the discrepancy in the
location of the minimum peak increases due to the smaller separation bubble
in the computation. The best agreement is found at the most downstream
stations in the exit channel, but the behavior close to the walls is, as in all
profiles, quite different. The measured second invariant increases more slowly
with increasing distance from the walls as compared to the computed invariant.
The pressure distributions on the plane wall are plotted in 12, normalized
with a dynamic pressure based on the inlet channel bulk velocity. They clearly
reflect the difference in strength of the separation. The EARSM gives a slightly
higher pressure recovery than the experiment and the effects of displacement
by the separation bubble are much smaller.

4. Concluding remarks
The plane asymmetric diffuser with 8.5◦ opening angle has shown to be a
challenging flow case for turbulence model testing with several important phe-
nomena, e.g. flow separation and reattachment, highly anisotropic turbulence,
high levels of turbulence and some streamline curvature effects. These complex
flow phenomena all occur within a relatively simple two-dimensional geome-
try. The smaller angle of the diffuser compared to that (10◦ ) in the earlier
experiments by Obi et al. (1993) and Buice & Eaton (2000) seems to give a
separation which is even more difficult to predict and hence constitutes a truly
challenging test case for turbulence closures.
62 O. Törnblom, B. Lindgren, J. Gullman-Strand and A. V. Johansson

0.9
0.8
0.7
0.6
0.5
Cp
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30 35 40 45
x/H
Figure 12. Comparison between the measured pressure dis-
tribution (◦) and the EARSM (—).

The data-base of measurements has proven to be useful for comparison

with computations, but more data from the near-wall regions are desirable.
Especially in the region near the upstream corner of the inclined wall where
streamline curvature is suspected to influence the turbulence. Furthermore,
Kaltenbach et al. (1999) noted in an LES-calculation that a very small sepa-
rated region formed in the viscous layer in this region. Accurate modelling and
fine resolution of the flow in this region is crucial in order to obtain the correct
behavior of the separating boundary layer on the inclined wall.
The EARSM manages to predict a separation bubble, but the size of it
is not in close agreement with the measurements. The main reason for this
is believed to be the overestimation of vrms and u0 v 0 in the beginning of the
diffuser. This can be related to the above-mentioned complicated flow in this
region. The curvature correction described in Wallin & Johansson (2002) has
not been implemented in the EARSM with which the comparisons are made.
Adding curvature corrections would reduce the overestimated Reynolds-stress
components at the inclined wall. History effects related to advection of the
anisotropy along streamlines may also be of influence here. To account for
these a DRSM modelling approach would be the normal choice.
However, Apsley & Leschziner (1999) tested four different DRSM:s on the
10◦ case, two of the models used wall functions and two were low Reynolds
number models, without getting satisfying results. It was argued that the
 Measurements in a plane asymmetric diffuser. Part II 63

difficulties related to the modelling of the wall-asymptotic behavior could be a

reason. This would indicate that a DRSM with nonlinear models for Πij etc.
would be interesting for this case. The DRSM of Sjögren & Johansson (2000)
is such a model that has been proven to satisfy near-wall asymptotic behaviors
through satisfaction of strong realizability. Apsley & Leschziner (1999) also
mention the possible problem of periodic shear-layer instabilities provoked by
the upstream corner of the diffuser as a reason for the general failure of all
models to resolve the initial development of the boundary layer on the inclined
wall.

5. Acknowledgement
The authors would like to thank Ulf Landén for aiding in the manufacturing of
the wind-tunnel and measurement equipment. The Swedish Research Council,
The Swedish Energy Agency and the Integral Vehicle Structures research school
are gratefully acknowledged for their financial support.

References

Amberg, G., Törnhardt, R. & Winkler, C. 1999 Finite elemet simulations using
symbolic computing. Mathematics and Computers in Simulation 44, 275–274.
Apsley, D. D. & Leschziner, M. A. 1999 Advanced turbulence modelling of sep-
arated flow in a diffuser. Flow, Turbulence and Combustion 63, 81—112.
Buice, C. U. & Eaton, J. K. 1997 Experimental investigation of flow through an
asymmetric plane diffuser. Tech. Rep.. Department of mechanical engineering,
Stanford university.
Buice, C. U. & Eaton, J. K. 2000 Experimental investigation of flow through an
asymmetric plane diffuser. J. of Fluids Eng. 122, 433–435.
Comte-Bellot, G. 1965 Écoulement turbulent entre deux paroix parallèles. Publi-
cations scientifiques et techniques 419. Ministère de l’air, 2, Avenue de la Porte-
d’Issy, Paris.
Gatski, T. B. & Speziale, C. G. 1993 On explicit algebraic stress models for
complex turbulent flows. J. Fluid Mech. 254, 59–78.
Girimaji, S. S. 1997 A Galilean invariant explicit algebraic Reynolds stress model
for turbulent curved flows. Phys. Fluids 9, 1067–1077.
Guermond, J. L. & Quartapelle, L. 1997 Calculation of incompressible vicous
flow by an unconditionally stable projection FEM. Journal of Computational
Physics 132 (CP965587), 12–33.
Gullman-Strand, J. 2002 Turbulence modeling using automated code generation
applied to asymmetric diffuser flow. Dept. of mechanics, KTH, Stockholm, Swe-
den, licentiate thesis.
Hellsten, A. & Rautaheimo, P., ed. 1999 Workshop on refined turbulence mod-
elling. ERCOFTAC/IAHR/COST.
64 O. Törnblom, B. Lindgren, J. Gullman-Strand and A. V. Johansson

Iaccarino, G. 2000 Prediction of the turbulent flow in a diffuser with commersial

CFD codes. Tech. Rep.. Center for Turbulence Research.
Johansson, A. V. & Burden, A. D. 1999 An introduction to turbulence mod-
elling. In Transition, Turbulence and modelling (ed. A. Hanifi, P. Alfredsson,
A. Johansson & D. Henningsson), chap. 4. Kluwer academic publishers.
Kaltenbach, H.-J., Fatica, M., Mittal, R., Lund, T. S. & Moin, P. 1999
Study of flow in a planar asymmetric diffuser using large-eddy simulation. J.
Fluid Mech. 390, 151–185.
Launder, B., Reece, G. & Rodi, W. 1975 Progress in the development of a
Reynolds-stress turbulence closure. J. Fluid Mech. 68, 537–566.
Lindgren, B., Törnblom, O. & Johansson, A. V. 2002 Measurements in a plane
asymmetric diffuser with 8.5◦ opening angle. part I: General flow characteristics.
To be submitted .
Obi, S., Aoki, K. & Masuda, S. 1993 Experimental and computational study of
turbulent separating flow in an asymmetric plane diffuser. In Ninth Symp. on
Turbulent Shear Flows, p. 305. Kyoto, Japan.
Sjögren, T. & Johansson, A. V. 2000 Development and calibration of algebraic
nonlinear models for terms in the Reynolds stress transport equations. Phys.
Fluids 12 (6), 1554–1572.
Wallin, S. & Johansson, A. V. 2000 An explicit algebraic Reynolds stress modell
for incompressible and compressible turbulent flows. J. Fluid Mech. 403, 89–132.
Wallin, S. & Johansson, A. V. 2002 Modeling streamline curvature effects in
explicit algebraic Reynolds stress turbulence models. Int. J. of Heat and Fluid
Flow. 23, 721–730.
Wilcox, D. C. 1993 Turbulence modeling for CFD. DCW Industries Inc., USA.
Paper 3

P3
 Design and calibration of a plane asymmetric
diffuser wind-tunnel
By Olle Törnblom, Björn Lindgren and Arne V. Johansson
Dept. of Mechanics, KTH, SE-100 44 Stockholm, Sweden

The design process of an experimental setup for studying a plane asymmetric

diffuser flow is described. Prior to the diffuser the flow passes through a 3 m
long and 30 mm high channel to generate a turbulent channel flow as inlet
condition to the measurement section, which is a plane asymmetric diffuser.
The opening angle of the inclined diffuser wall can be adjusted between 8-10◦
in order to be able to set the size of the separated region, within a limited
range. The wind-tunnel is of closed circuit type in order to allow for flow
seeding in an indoor environment. The wind-tunnel is driven by an 11 kW
radial fan, capable of generating inlet channel flow Reynolds numbers of up
to HUcl /ν =60000, H being the channel height and Ucl the channel centerline
velocity. The calibration of the wind-tunnel is described and special attention
is paid to the two-dimensionality of the flow and the long time variations of
the wind-tunnel velocity.

1. Introduction
Study of diffuser flows, i.e. duct flows subjected to an adverse pressure gradient,
is motivated by the large number of applications in which these flows appear
and the energy savings that may result from an increased knowledge about
proper diffuser designs and separation control. The experimental facility for
which the design is described herein was built to study a generic flow with
separation, both in order to get a database for turbulence model comparisons
as well as to test separation control techniques.
If one desires to study turbulent separating flow, separation control and
provide reference data for computations, in the same geometry, it has to be
chosen with care. The most generic way to study separation is to do a flat plate
experiment where an adverse pressure gradient is produced by altering the free-
stream velocity with a variable geometry of the opposing wind-tunnel wall (see
e.g. Perry & Fairlie (1975) and Angele (2002)). Dianat & Castro (1991) used
a porous cylinder with a small flap at the trailing edge mounted above a flat
plate to generate the pressure gradient. Another category of separated flows
is that associated with sharp edges or bluff body geometries. Examples are
67
68 O. Törnblom, B. Lindgren and A. V. Johansson
24.8 H

o
y
8.5

4.7 H
Recirculation zone x
Turbulent channel inflow
H

Figure 1. The plane asymmetric diffuser.

flow over flat plates or cylinders with blunt leading edges (e.g. Cherry et al.
(1984) and Ruderich & Fernholz (1975)) or flow over a backward facing step
(e.g. Etheridge & Kemp (1978) and Yoshioka et al. (2001)). Symmetric diffuser
flows have been investigated by e.g. Sovran & Klomp (1967) and Chithambaran
et al. (1984). All of these flow cases have been used to study different aspects of
separation, but some are more suited for the study of generic mechanisms and
comparison between experiments and numerical simulations. For instance the
varying free-stream velocity in a separating flow on a flat plate can be difficult
to mimic accurately in a CFD simulation. In blunt edge flows and flows over
backward facing steps the challenge of predicting the separation point vanishes,
as the separation inevitably will occur at the corner. In symmetric diffusers the
separation can fluctuate between the two diffuser walls in a random manner and
at long time-scales making ensemble averaging a very tedious task. In a CFD
calculation, this unstable flow situation may also cause convergence problems.

The plane asymmetric diffuser with turbulent channel inflow, shown in

figure 1, had the properties we were looking for. The fully developed turbulent
channel flow used as inflow condition is a well known flow, simplifying the task
of setting appropriate boundary conditions, both in CFD computations and
in direct numerical simulations. The asymmetric geometry favors separation
on the inclined wall and eliminates the problems associated with symmetric
diffusers. If the upstream corner of the inclined wall is sufficiently smooth
with a finite radius of curvature, the separation point will be located at the
inclined wall, giving a challenging flow to predict in simulations and a location
of the separation point sensitive to control. Furthermore, the relatively simple
geometry with two dimensional flow and mostly straight walls facilitate CFD
grid generation and the determination of positions inside the measurement
section.

2. Design
2.1. Pre-design considerations
At the time when the first design steps where taken, there existed two (to the
authors known) previous experiments in a very similar geometry i.e. those by
 Design and calibration of a plane asymmetric diffuser wind-tunnel 69

Obi et al. (1993) and Buice & Eaton (1997). In particular the latter one, which
contains a detailed description of the experimental setup, was studied in or-
der to learn about the difficulties related to the design of a plane asymmetric
diffuser experiment. The major difficulty of this flow case is the inherent ten-
dency, of a separating flow, to be three-dimensional. Since we required the flow
to be two-dimensional (in an average sense), the three-dimensional behavior
must be controlled. Since three-dimensional effects enter the flow via the end
walls (the walls which limit the spanwise extent of the diffuser), the ideal, but
unrealistic, solution would be to have an infinitely wide diffuser. However, a
relatively large spanwise width compared to the height of the diffuser (i.e. a
large aspect ratio) is a necessary requirement to have a two dimensional flow
field.
A preliminary idea was that the newly built low speed wind-tunnel at the
department (cf. Lindgren & Johansson (2002)) could be used for the experi-
ment. But that alternative was abandoned due to the large pressure losses that
the inlet channel would cause and because of the limited aspect ratio it would
allow.
Another decision which needs to be taken before the actual design-work
start is how large the diffuser should be. The size of the facility partly deter-
mines what Reynolds number the flow will have and also what fan power is
needed to drive the wind-tunnel at the desired flow speeds. In a small facility
the flow will be more influenced by small imperfections in the design e.g. small
edges in joints between different wall sections. A too small test-section also
limits the accessibility, due to the physical size of our hands, arms etc. and it
also requires better precision in the positioning of the measurement probe than
in a larger test-section, for the same relative error.
With these arguments in mind we decided to make our diffuser larger than
the ones used in the experiments by Obi et al. (1993) and Buice & Eaton
(1997). We also wanted a larger aspect ratio than the previous experiments.
Obi’s and Buice’s inlet channel heights were 20 mm and 15 mm and their
spanwise widths were 700 mm and 450 mm respectively. We chose to build a
30 mm high channel and an inlet aspect ratio of 50, giving a channel width of
1500 mm (due to the standard width of plywood sheets the final channel was
built with a width of 1525 mm). In turbulent channel flow an aspect ratio of five
is the lowest possible in order to have (a reasonably) two-dimensional flow in an
about two channel heights wide region in the center of the channel (Johansson
& Alfredsson (1986)). If the aspect ratio is lower than this, secondary flow
effects caused by end walls and corners may influence the flow. In our facility
this requirement is well fulfilled also in the exit channel where the aspect ratio
is 10.8. However, in flows with adverse pressure gradients these rules of thumb
can only be considered valid if separation is avoided on the end walls.
70 O. Törnblom, B. Lindgren and A. V. Johansson

In order to produce a fully developed turbulent channel flow it was decided

that the length of the inlet channel should be at least 100 channel heights.
According to the experiments by Comte-Bellot (1965) the higher order statis-
tical moments are fully developed at 60 channel heights from the inlet. Similar
observations were made by Johansson & Alfredsson (1982).
Accordingly, the final choice of channel geometry was a 30 mm high, 3000
mm long and 1500 mm wide channel. It was concluded that, in order to min-
imize the floor area occupied by the wind-tunnel, the channel should be built
so that the spanwise direction was oriented vertically, which is also favorable
in the sense of minimizing wall deflections caused by gravity.
One of the obvious difficulties when studying separating flows experimen-
tally is the occurrence of velocities with changing sign, due to the region of back-
flow. This difficulty makes it impossible to measure with hot-wire anemometry,
an otherwise well established and accurate technique. If one wants to deter-
mine both the value of the velocity and its direction one has a few alternatives;
pulsed hot-wires, laser doppler velocimetry (LDV) or particle image velocime-
try (PIV). The natural choice for us was to use the two latter techniques, as
these techniques are well established at our department and complement each
other well, LDV gives rather high data rates for accurate point measurements
and PIV gives whole field measurements but at a slow rate. LDV and PIV are
both optical methods, and thereby require optical access to the flow. Hence,
the measurement section of the wind-tunnel has to be made in a transparent
material. If one wants to make a large transparent wall, as in a wind-tunnel
test section, at a reasonable cost one has practically two choices, window glass
or polymer glass. Window glass typically has a more even and precise thickness
than polymer glass. An even thickness is of high importance when measuring
with LDV, since a distortion of the laser-beams (or rather of one of the beams)
can make them misaligned in the measurement volume, resulting in bad, or
no, doppler signals. Furthermore, window glass has a higher modulus of elas-
ticity as compared to polymer glass which allows thinner sheets to be used
for a certain desired stiffness. Thinner sheets will also contribute to lowering
the optical distortion. If one desires to use polymer glass a popular choice is
Plexiglas which is made from polymethyl methacrylate (PMMA). This mate-
rial is usually cheaper than window glass and not as brittle, furthermore it is
easier to machine than window glass. Another material property which could
influence the choice is the refractive index, but for PMMA and window glass
this can be disregarded since it is approximately 1.5 for both materials. The
need to be able to drill pressure taps, the cheaper prize and the possibility to
join Plexiglas sheets together with screws made us choose this material for
large parts of the measurement section.
 Design and calibration of a plane asymmetric diffuser wind-tunnel 71

11

5 5 7 9
2 3 4 6 8 10
1

3.0 m 1.5 m 2.0 m

Figure 2. A sketch of the complete wind-tunnel. Refer to

table 1 for a description of the numbered parts.

Part Description
1 Fan, 11 kW centrifugal
2 Transformer from 322×229 mm2 to 1525×100 mm2
3 Settling chamber, 1525×100×500 mm3
4 Contraction, from 1525×100 mm2 to 1525×30 mm2
5 Inlet channel, 1525×30×3000 mm3
6 Diffuser/measurement section, from 1525×30 mm2 to 1525×141 mm2
7 Exit channel, 1525×141×2000 mm3
8 Contraction, from 1525×141 mm2 to 400×300 mm2
9 Heat exchanger
10 Transformer, from 400×300 mm2 to 400 mm diameter
11 Return pipe, 400 mm diameter with five 90◦ bends
Table 1. Description of the numbered parts in figure 2.

2.2. Fan and motor

Having decided the spanwise width and height of the inlet channel and the
desired Reτ in the inlet channel the required flow rate can be calculated using
the logarithmic friction law and Dean’s relation,
² ³
Ucl 1 uτ H
= ln +C (1)
uτ κ 2ν
² ³−0.0116
Ucl Ub H
= 1.28 (2)
Ub 2ν

with the additive constant C = 6.0 and κ = 0.4. Where Ucl is the channel cen-
terline velocity, uτ the friction velocity at the wall, Ub the channel bulk velocity
and H is the channel height. A desired Reτ = uτ H/ν = 4000 gives, according
to these relations and the chosen channel area, a volume flux of approximately
1.8 m3 /s. In order to choose a fan one has to know also at what pressure the
72 O. Törnblom, B. Lindgren and A. V. Johansson

required flow should be delivered. This has to be estimated. A very rough es-
timate yielded that the wind-tunnel would cause pressure losses corresponding
to approximately three times the dynamic pressure of the inlet channel bulk
velocity. Later this estimate was proven to be reasonably accurate for the first
design, which did not have a return circuit and heat exchanger. When choosing
a fan it is advisable to examine not only the maximum performance, but also
the performance over a region of the fan characteristics curve. It is desirable
to have a characteristics curve in the volume flow region where the fan will be
used, that gives as small a flux variation as possible for a given variation in the
pressure, for instance due to a nonstationary separation bubble.
Based on these estimates a centrifugal fan capable of delivering 2.1 m3 /s
at 3000 Pa was ordered from Ventur tekniska AB. To control the speed of the
11 kW induction motor, the fan was complemented with a frequency converter
and a filter, from ABB Motors AB.

2.3. Transformer and settling chamber upstream the channel

The fan has an outlet section of 322×229 mm2 that has to be transformed into
the large aspect ratio cross-section (1525×30 mm2 ) of the inlet channel. This
transformation had to take place over a limited length, putting high demand on
the design in order to minimize losses and ensure an even flow over the whole
cross-section.
A common way to achieve a spanwise homogeneous flow is to have a settling
or stagnation chamber, with large cross-section area where the velocity is very
low, followed by a contraction into the channel. Here, we chose to have a
rather small settling chamber with a cross-section area of 1525 × 100 mm2 . By
doing so, the transformer between the fan and the settling chamber will have
a shape which diverges in the spanwise direction and contracts in the other
direction, see figure 3. The total expansion ratio, i.e. the ratio between the
cross-section area of the settling chamber and the fan outlet area, is thus only
about two. Two splitter plates are mounted immediately after the fan outlet
to avoid large scale flow separation here and to direct the flow outwards in the
spanwise direction.
Three screens, shown in figure 3, were placed in the transformer and settling
chamber in order to even out mean flow variations. The screens used have a
wire diameter d of 0.71 mm and a (wire center to wire center) mesh width M
of 3.2 mm. These values give a porosity β = (1 − d/M )2 = 0.6. Using screens
will not only even out mean flow variations but also reduce angle deviations
relative to the screen according to the relation

φ = αθ, (3)

where θ is the angle of the incoming flow and φ the angle of the out-going flow.
The constant α is related to the local pressure-loss coefficient K0 through the
 Design and calibration of a plane asymmetric diffuser wind-tunnel 73

Figure 3. Four view drawing of the transformer and settling

chamber upstream the channel.

empirical relation (see Laws & Livesey (1978))

1.1
α= √ . (4)
1 + K0
The local pressure-loss coefficient is highly Reynolds number dependent for
subcritical Reynolds numbers but can for higher Reynolds numbers be approx-
imated with
1 − β2
K0 = 0.5 , (5)
β2
if 0.6 < β < 0.7. With a volume flow of 2.1 m3 /s the wire-diameter Reynolds
number will be well above the critical value of Red = 100. Using equations 4
and 5 it can be seen that the local pressure-loss coefficient will be K0 = 0.9 and
that an error in the incoming flow angle relative to the screen will be reduced by
a factor 0.8 after the screen. Streamwise mean velocity variations are reduced
74 O. Törnblom, B. Lindgren and A. V. Johansson

when the flow passes through a screen according to the expression

∆u2 1 + α − αK0
= , (6)
∆u1 1 + α + K0

derived by Taylor & Batchelor (1949), where ∆u1 and ∆u2 are the variations
of the streamwise mean velocity upstream and downstream the screen respec-
tively. For the screens of concern herein this ratio will be 0.4.
In the diffusing part, a bent screen was used in order to both even out
mean velocity variations and to direct the flow towards the desired direction
according to equation 3. Further downstream, in the first part of the settling
chamber, two screens are placed to further reduce possible mean flow variations.
The screens were mounted onto frames made from 15×15 mm stainless steel
tubes and were manufactured by AB Derma in Gråbo, Sweden.
Both the transformer and the settling chamber were manufactured in ply-
wood sheets joined together with L-shaped aluminum profiles and screws and
sealed with tightening material. All the wind-tunnel parts are mounted in a
framework (as can be seen in figure 2) of 60×40×3 mm3 and 60×40×4 mm3
steel tubes and steel L-profiles respectively. The steel framework is bolted to
the concrete floor in the laboratory.

2.4. Contraction and inlet channel

The settling chamber is followed by a two-dimensional contraction, changing
the channel height from 100 mm to 30 mm. The shape of the contraction,
shown in figure 4, is described by two sinus hyperbolic functions as described
in Lindgren & Johansson (2002). This shape is an optimal shape regarding
the pressure gradient along the walls and it was derived originally for the MTL
wind-tunnel by Henrik Alfredsson and Alexander Sahlin at the department,
using inviscid/boundary layer calculations. The contraction was manufactured
in polystyrene foam which was cut to the appropriate shape using a hot-wire
saw. The polystyrene foam was covered with a glassfiber-epoxy laminate in
order to have a smooth and hard surface.
The inlet channel consists of two 3000×1525×21 mm3 plywood sheets with
plywood end walls joined together with L-shaped aluminum profiles. Four L-
shaped aluminum profiles oriented in the spanwise direction are screwed to
the large plywood sheets in order to prevent the walls from bulging due to
the larger pressure inside the channel. Furthermore, the large plywood sheets
are supported along their centerlines with adjustable screws in order to ensure
a constant (30 mm) channel height at all locations. The channel height was
controlled with a movable 30 mm plastic cube (mounted on a long stick to be
able to reach all points in the channel) and adjusted with the screws. The
horizontal bar with the adjustment screws can be seen in figure 2.
 Design and calibration of a plane asymmetric diffuser wind-tunnel 75

Glass-fiber reinforced epoxy

Polystyrene foam

Figure 4. The design for the contraction before the inlet channel

2.5. Measurement section

The measurement section (part number 6 in figure 2) is connected to the inlet
channel and the exit channel with high precision joints milled in aluminum
with steel guide taps. The joints are held together with three DE-STA-CO
clasps at each side. This design allows the measurement section to be removed
in order to facilitate modification of the diffuser or measurement installations.
As mentioned in section 2.1, optical access to the measurement section is
necessary and Plexiglas was believed to be the most appropriate transparent
material. The plane wall of the measurement section was made from a single
1500×1525×25 mm3 Plexiglas sheet and the end walls are two 350×1500×25
mm3 sheets. It was discovered that the delivered Plexiglas varied in thickness
as much as ±1 mm, and later on a 4 mm thick float-glass window was inserted to
replace the Plexiglas in the upper end wall. The 1250×155 mm2 piece of float-
glass was taken from the mid portion of a large glass plate where the thickness
homogeneity are best according to a representative for Pilkington floatglas AB
in Halmstad, Sweden. The plane wall and the end walls are joined with screws
that go through the end walls and are screwed into threaded holes in the plane
wall.
The inclined wall of the measurement section is made as an aluminum
sandwich construction. Square tubes (20×20 mm2 ) are screwed and glued
with epoxy to a 4 mm sheet of aluminum. A 1 mm thick sheet of aluminium
is then glued onto the 4 mm sheet to have a smooth surface. The smooth
transitions between the inlet channel and the inclined wall and between the
inclined wall and the straight wall in the exit channel are achieved by omitting
the 4 mm sheet in these areas and letting the 1 mm sheet bend to form a
smooth curve. This design allows the angle of the inclined wall to be changed
76 O. Törnblom, B. Lindgren and A. V. Johansson

Figure 5. Cross section of the measurement section showing

the design of the inclined wall and between which angles the
diffuser can be altered. The holes for end wall boundary layer
removal can also be seen.

between approximately 8◦ and 10◦ , see figure 5. The inclined wall is fastened
to a framework through adjustable screws which are used to alter the angle of
the inclined wall. The inclined wall can be adjusted to the desired angle with
high accuracy by inserting precision-milled blocks with correct angle into the
measurement section and adjusting the screws. The flat wall in the measure-
ment section downstream the inclined wall is made of 21 mm thick plywood
and has two hatches that allow easy access to the measurement section.
When the turbulent channel flow develops in the 3 m long inlet channel
boundary layers will form on the channel end walls. These boundary layers
will decrease the level of momentum in a region near the end walls, making
the end wall flow sensitive to adverse pressure gradients. By reducing the
thickness of the end wall boundary layers through suction before the diffuser
where the adverse pressure gradient starts, end wall separation can be avoided.
The suction is applied through a 100 mm long perforated section of the end
wall ending 50 mm before the diffuser starts. The suction is produced by a
separate fan connected to the perforated parts through hoses. The air that is
sucked out is brought back to the wind-tunnel circuit by leading the exit of
the fan to a connection on the return channel. After the perforated section a
new boundary layer will start to grow and to prevent this boundary layer from
separating small vortex generators are put on the end walls in the diffuser.
A row of 57 pressure taps were drilled along the centerline of the plane
wall, the distance between the pressure taps is 25 mm and the holes which are
 Design and calibration of a plane asymmetric diffuser wind-tunnel 77

drilled directly in the Plexiglas has a diameter of 0.4 mm. Similarly a courser
row of pressure holes were drilled on the opposite side of the diffuser, along
the inclined wall. Pressure taps were also drilled in spanwise rows on both
walls 100 mm before the upper corner of the diffuser in order to measure the
spanwise homogeneity of the incoming flow.

2.6. Exit channel and transformer before the heat exchanger

The exit channel is built in plywood in a similar way as the inlet channel
and has the purpose of reducing the effects of the downstream wind-tunnel
parts on the flow in the measurement section. The size of the exit channel
is 2000×1525×141 mm3 and stiffening aluminum L-profiles and adjustment
screws along the centerline are used, in the same fashion as for the inlet channel,
to ensure that these dimensions are kept.
The exit channel is followed by a transformer, changing the cross section
to 400×300 mm2 in order fit the dimensions of the following heat exchanger.
The shape of the contraction was determined using the same formula as for the
contraction upstream the inlet channel (see section 2.4). Since the transformer
is contracting in one direction and expanding in the other the total contraction
ratio is only 1.8, smaller than what one can conceive from figure 2. The sides of
the transformer were sawed into the proper shape from a 25 mm thick plywood
sheet and the (bent) upper and lower walls were made by steel tinplates nailed
to the plywood.

2.7. Heat exchanger and return channel

When running the wind-tunnel, the fan is constantly adding kinetic energy to
the ’closed’ system. Energy which through viscous dissipation eventually will
become heat. This heat has to be removed at the same rate as it is added
if a constant flow temperature is desired in the experiment. In order to do
so a heat exchanger, consisting of a car radiator (VW Golf II), was built into
the wind-tunnel circuit. In the heat exchanger the cooling medium is tap
water. No automatic temperature control system has been implemented as
the temperature variations proved to be small enough for optical measurement
techniques without automatic control.
After the heat exchanger the cross section shape is gradually changed from
rectangular to circular with 400 mm diameter. This is done in a 650 mm long
transformer section manufactured in glass fiber reinforced epoxy molded on a
plug which was formed from a block of styrofoam using a hot-wire saw.
The return channel is made of standard 400 mm diameter ventilation pipes
and standard 90◦ bends. The total length of the return channel is approxi-
mately 12 m and five 90◦ bends were used. The return channel is supported by
consoles bolted to the ceiling. The return channel ends approximately 50 mm
before the inlet of the fan in order to have a point in the wind-tunnel circuit
78 O. Törnblom, B. Lindgren and A. V. Johansson

with a constant reference pressure (the pressure in the laboratory) and thus
avoid pulsating pressures from appearing. The last bend of the return channel
has two connectors for hoses one for introducing seeding particles and one for
returning the air from the boundary layer suction (see section 2.5).

3. Calibration
The wind-tunnel was first built without the return channel and preliminary
tests using pitot-tubes and LDV were made in this configuration mainly with
the aim to evaluate the inflow conditions and the two-dimensionality of the
flow.

3.1. Pitot-tube measurements of the inflow

Pitot tube measurements of inlet velocity profiles were performed at six differ-
ent spanwise positions 100 mm upstream of the first corner of the diffuser in
order to evaluate the turbulent channel inflow and the spanwise homogeneity
of the inflow. In order to reduce the perturbation of the flow, the pitot tube
was positioned at the end of a long carbon fiber reinforced sting (see figure 6).
The pitot tube was traversed from the wall towards the channel center with
a DC servo controlled motor operated from a computer. In order to measure
from the opposite wall the sting had to be rotated 180◦ around its axis. The
pitot tube senses the total pressure, the static pressure was measured in a hole
on the plane wall and the pressure difference was measured with a Furness
Control FCO 510 differential pressure transducer with an accuracy of 0.25%
of full scale (2000 Pa). Although the sting had been stiffened by carbon fibers
the pitot-tube vibrated during the measurements with an estimated amplitude
of 1 mm, this in combination with the rather large ’probe volume’ of a pitot
tube introduce some uncertainties in the measured data. Small differences in
the shapes of the static pressure holes is another error source.
Figure 7 shows the six measured inlet profiles compared to DNS-data by
Kim et al. (1987) at Reτ = 590. The DNS-data have been scaled so that the
centerline velocity is 22 m/s. Table 2 shows the maximum velocities measured
at the six positions. The spanwise positions are measured from the center-
line of the diffuser and the positive direction is upwards in the experimental
setup. The y-coordinates of the half profiles measured from each side has been
matched so that their maximum velocities are at the channel center line. The
qualitative shape of the profiles agrees well with the DNS-data. The centerline
velocities are varying with a few percent over the whole spanwise width. It
should be noted that these measurements where taken before the velocity con-
trol, described in section 3.3, was implemented and that the different profiles
where measured at different times.
 Design and calibration of a plane asymmetric diffuser wind-tunnel 79

Inclined diffuser wall

H
Y
H
H Plane diffuser wall

ptot pstat

Figure 6. Sketch of the setup used for the pitot-tube measurements.

Spanwise position [mm] 700 400 100 -100 -400 -700

Centerline velocity [m/s] 22.1 22.3 22.0 22.0 22.4 22.0
Table 2. Centerline velocities at the inlet.

3.2. Two-dimensionality
The previous section concerned the spanwise homogeneity of the inflow but
the largest difficulty of this experiment is to get a two-dimensional flow in the
adverse pressure gradient region in the diffuser. The methods used to prevent
end wall separation are described in 2.5. In order to see if the end wall flow
really is attached tufts taped to the end walls were used. With the boundary
layer suction turned on, the only backflow that could be observed on the end
walls was in the regions where the ’desired’ separation is located i.e. near the
downstream half of the inclined wall and on a limited area downstream of that.
Notable is that if the boundary layer suction was turned of, very large scale
vortices, which were visualized with smoke, formed. These large vortices could
extend 1/3 of the spanwise width from the end walls towards the center line and
would probably give rise to a significant flow rate increase in the downstream
parts of the diffuser if measurements were taken along the diffuser centerline.
80 O. Törnblom, B. Lindgren and A. V. Johansson

z=700 mm z=400 mm
30 30

20 20

10 10

0 0
0 5 10 15 20 25 0 5 10 15 20 25
z=100 mm z=-100 mm
30 30

20 20

10 10

0 0
0 5 10 15 20 25 0 5 10 15 20 25
z=-400 mm z=-700 mm
30 30

20 20

10 10

0 0
0 5 10 15 20 25 0 5 10 15 20 25

Figure 7. Inlet profiles from six spanwise positions compared

to DNS-data at Reτ = 590 by Kim et al. (1987). The horizon-
tal scale is velocity in m/s and the vertical scale is wall normal
distance in mm.

The flow-rate is a good measure of the two-dimensionality of a flow and

for this flow a suitable definition of the flow rate is
Z
1
U (y)dy. (7)
HUb
Where U is the velocity in the x-direction, H is the inlet Rchannel height and Ub
is the bulk velocity of the inlet channel flow (i.e. Ub = Uinlet dy/H). Figure
8 shows the flow rate measured with PIV (see Lindgren et al. (2002)) in the
center region of the measurement section. For comparison the flow rates of
Obi et al. (1993) and Buice & Eaton (1997) (the data have been taken from
Kaltenbach et al. (1999)) have been included in the figure. The bulk velocity
used for normalization was calculated using equations 1 and 2 and the friction
 Design and calibration of a plane asymmetric diffuser wind-tunnel 81

1.15

1.1
U dy/HUb

1.05
R

0.95
-10 0 10 20 30 40 50 60 70

x/H

Figure 8. Flow rate measured with PIV (◦) compared to

measurements by Obi et al. (1993) (3) and Buice & Eaton
(1997) (2).

velocity at the inlet was measured with a Preston tube. The PIV-data show a
peak at x/H around 5 with 5 % higher flow rate as compared to the inlet bulk
flow, and after that a decrease to around 3 % lower flow rates.

3.3. Long time velocity variations

It was noticed when the first measurements with LDV were taken that the
mean values of the measured velocity converged very slowly indicating some
very long timescale variation, or drift, of the velocity. In order to handle
this problem a LabView program (see figure 9) was built allowing the fan to
be controlled from the computer. The LabView program measures the air
temperature in the wind-tunnel with a PT-100 temperature sensor, the friction
velocity at the inlet with a Preston tube connected to a Furness Control FCO
510 differential pressure transducer and the ambient pressure with an absolute
pressure transducer connected to the FCO 510. From the measured data the
density and viscosity of the air is calculated. These values are used to calculate
Reτ at the inlet and a proportional controller adjusts the fan speed to keep
Reτ at a constant level. Figure 10 illustrates the effect of the controller, the
control is turned off after 4800 s and turned on again after 70800 s.

4. Concluding remarks
An experimental setup for studying the flow in an asymmetric diffuser has been
built and verified. The two-dimensionality and the long time velocity variation
was given special attention in the wind-tunnel design and evaluation. The
82 O. Törnblom, B. Lindgren and A. V. Johansson

TEMPERATURE FROM PT-100 SENSOR THROUGH HP34401A WITH 4-WIRE COMPENSATION

Plot 0
TEMP GRAPH
Type of loop
22.00
For While
21.50
No of loops loo 21.00

 20.50
0 2184
20.00
19.50
No of sets set
19.00
12 12
18.50
date 18.00
021118 17.50
17.00
Yes New file?
16.50
No 16.00
0 10000 20000 30000 40000 50000 60000 70000 80000 90000

device position
Plot 0
1 5 Retau GRAPH

tube diameter Patm 1010.00

Retau set
0.0016 1013 1007.50

Retau Retau set 1005.00

998.45 1000.00
1002.50
Port K
1000.00
0 0.003
997.50
sampling Offset (Pa)
995.00
0.00
Sampler data
992.50
Sets 30 990.00
Scans 1024 0 10000 20000 30000 40000 50000 60000 70000 80000 90000
Scanrate 500.00

Buffersize 4096 R of loop R of set T of loop T of set delay time time t1 fan speed fan status

channels 1 106.4715 106.4701 16.5990 16.5955 0 86774.02 3916809.73 5.62

Figure 9. Screen dump from the wind-tunnel control program.

diffuser was designed in such a way that the angle of the inclined wall can be
altered between 8◦ and 10◦ .
The design of a compact-length transformer between the fan outlet and the
inlet channel was successful. Screens, including one bent screen, and splitter
plates where used in the transformer in order to expand the flow over a short
distance. The inlet channel was designed to be long enough and wide enough
to have a fully developed two-dimensional channel flow. The channel flow mean
velocity profiles are verified by comparing measurements with DNS data and
the spanwise homogeneity is checked by taking measurements at six different
spanwise positions.
The two-dimensionality is evaluated by measuring the flow rate in a two-
dimensional plane using PIV. The flow rate is found to be constant through
the measurement section within ±5%.
The wind-tunnel is controlled from a computer program, logging the at-
mospheric pressure, the temperature and the friction velocity at the inlet. A
controller adjusts the fan speed to keep the inlet Reynolds number constant, a
necessary condition for experiment repeatability.
 Design and calibration of a plane asymmetric diffuser wind-tunnel 83
22
temp [◦ C]

20
18
16
4
0 1 2 3 4 5 6 7 8 9 x 10
1010
Reτ

1000
velocity control off
990 4
0 1 2 3 4 5 6 7 8 9 x 10
1020 x 10
Patm [mbar]

1015

1010 4
0 1 2 3 4 5 6 7 8 9 x 10
time [s]

Figure 10. Output information from the velocity control pro-

gram illustrating the effect of velocity control.

5. Acknowledgement
The authors would like to thank Ulf Landén for aiding in the manufacturing of
the wind-tunnel and measurement equipment. The Swedish Research Council
and The Swedish Energy Agency are gratefully acknowledged for their financial
support.

References

Angele, K. P. 2002 Pressure-based scaling in a separating turbulent APG boundary

layer. In Advances in turbulence IX (ed. I. Castro, P. E. Hancock & T. G.
Thomas), pp. 639–642. CIMNE, Barcelona, Spain.
Buice, C. U. & Eaton, J. K. 1997 Experimental investigation of flow through an
asymmetric plane diffuser. Tech. Rep.. Department of mechanical engineering,
Stanford university.
Cherry, N. J., Hillier, R. & Latour, M. E. M. P. 1984 Unsteady measurements
in a separated and reattaching flow. J. Fluid Mech. 144, 13–46.
Chithambaran, V. K., Aswatha, P. A. & Chandrasekhara Swamy, N. V. 1984
Mean velocity characteristics of plane diffuser flows with inlet velocity distortion.
J. Indian Inst, Sci. 65(A), 79–93.
84 O. Törnblom, B. Lindgren and A. V. Johansson

Comte-Bellot, G. 1965 Écoulement turbulent entre deux paroix parallèles. Publi-

cations scientifiques et techniques 419. Ministère de l’air, 2, Avenue de la Porte-
d’Issy, Paris.
Dianat, M. & Castro, I. P. 1991 Turbulence in a separated boundary layer. J.
Fluid Mech. 226, 91–123.
Etheridge, D. W. & Kemp, P. H. 1978 Measurements of turbulent flow downstream
of a rearward-facing step. J. Fluid Mech. 86, 545–566.
Johansson, A. V. & Alfredsson, P. H. 1982 On the structure of turbulent channel
flow. J. Fluid Mech. 122, 295–314.
Johansson, A. V. & Alfredsson, P. H. 1986 Structure of tubulent channel flows.
In Encyclopedia of Fluid Mechanics, chap. 25, pp. 825–869. Gulf publishing
company.
Kaltenbach, H.-J., Fatica, M., Mittal, R., Lund, T. S. & Moin, P. 1999
Study of flow in a planar asymmetric diffuser using large-eddy simulation. J.
Fluid Mech. 390, 151–185.
Kim, J., Moin, P. & Moser, R. 1987 Turbulence statistics in fully developed channel
flow at low Reynolds number. J. Fluid Mech. 177, 133–166.
Laws, E. & Livesey, J. 1978 Flow through screens. Ann. Rev. Fluid Mech. 10,
247–266.
Lindgren, B. & Johansson, A. V. 2002 Design and evaluation of a low-speed
wind-tunnel with expanding corners. Technical report ISRNKTH/MEK/TR–
02/14-SE. Dept. of mechanics, KTH.
Lindgren, B., Törnblom, O. & Johansson, A. V. 2002 Measurements in a plane
asymmetric diffuser with 8.5◦ opening angle. part I: General flow characteristics.
To be submitted .
Obi, S., Aoki, K. & Masuda, S. 1993 Experimental and computational study of
turbulent separating flow in an asymmetric plane diffuser. In Ninth Symp. on
Turbulent Shear Flows, p. 305. Kyoto, Japan.
Perry, A. E. & Fairlie, B. D. 1975 A study of turbulent boundary-layer separation
and reattachment. J. Fluid Mech. 69, 657–672.
Ruderich, R. & Fernholz, H. H. 1975 An experimental investigation of a turbulent
shear flow with separation, reverse flow, and reattachment. J. Fluid Mech. 163,
283–322.
Sovran, G. & Klomp, E. D. 1967 Experimentally determined optimum geometries
for rectelinear diffusers with rectangular, conical, or annular cross-section. In
Fluid mechanics of internal flow (ed. G. Sovran), pp. 270–319. Elsevier publish-
ing company.
Taylor, G. & Batchelor, G. K. 1949 The effect of wire gauze on small disturbances
in a uniform stream. Quart. J. Mech. and App. Math. 2, part 1,1.
Yoshioka, S., Obi, S. & Masuda, S. 2001 Turbulence statistics of periodically
perturbed separated flow over backward facing step. Int. J. of Heat and Fluid
Flow 22, 391–401.