Wesleyan University Physics Department The Effects of Non-Universal Large Scales on Conditional Statistics in Turbulence by Daniel Brian Blum A dissertation submitted to the faculty of Wesleyan University in partial fulfillment of the requirements for the Degree of Doctor of Philosophy Middletown, Connecticut April, 2011
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Wesleyan University Physics Department
The Effects of Non-Universal Large Scales onConditional Statistics in Turbulence
by
Daniel Brian Blum
A dissertation submitted to the
faculty of Wesleyan University
in partial fulfillment of the requirements for the
Degree of Doctor of Philosophy
Middletown, Connecticut April, 2011
Dedication
For Karen Blum, who packed me lunch every day of school for the first 12 years, and called
every week for the remaining 10.
Acknowledgements
As I near the end of my time at Wesleyan I feel many things, but two emotions that stand our
are relief, and more strongly, thankful. This has been a great undertaking, (great, both in terms
of magnitude, and in positive impact) and it absolutely would not have been possible without
many people. First and foremost I would like to thank my advisor, Professor Greg Voth. Having
a knowledgeable advisor is common, but having such a patient, helpful and enthusiastic advisor
seems to be particular to Greg, and for this I am very grateful. I would also like to express my
gratitude to the other members of my thesis committee, Professors Fred Ellis and Tsampikos
Kottos for their support over the years. The Wesleyan physics department as a whole has been
very supportive, and I owe them particular gratitude for helping to fund my class at Yale. For
their collaboration, which makes up chapter 5, I would like to thank Eberhard Bodenschatz,
Mathieu Gibert, Armann Gylfason, Laurent Mydlarski, Haitao Xu, and P.K. Yeung. In addition,
I’d like to express my gratitude to Zellman Warhaft, Mark Nelkin, and Nick Ouellette who have
contributed their expertise to this work. My fellow Voth lab researchers have helped a great
deal over the years, Jim Johnson helped construct the tank, Dominic Stitch, Dennis Chan, and
Susantha Wijesinghe created the image compression circuits, Shima Parsa Moghaddam, Tom
Glomann, Emmalee Reigler, Rachel Brown, Nick Rotile, Reuben Son, and Surendra Kunwar
have all helped both with technical projects and camaraderie. This research was financially
supported by NSF Grant No. DMR-0547712 and the Alfred P. Sloan Foundation. It is common
to thank ‘the machine shop guys’ for their work on the apparatus, but I can truly say that
Dave Boule, and Bruce and Dave Strickland are the finest machine shop guys I have worked
with. I can also safely say that I would not have passed the qualifying exams without studying
ii
intensely with Joshua Bodyfelt. I also appreciate the camaraderie of my classmates over the
years, particularly Luis Fernando Vargas. I would also like to thank those at the beginning
of my physics career, Professors Sean Washburn, Lloyd Carrol, and Daniel Lathrop, as well as
Josh LaRocque, Adam Brooks, Rachel Rosen, Dan Zimmerman, Greg Bewley, and Santiago
Traiana.
Graduate school can be difficult, but it would have been impossible for me without the support
of my friends and family. For his years of dinner ‘collaborations,’ rides to New York, and dear
friendship I am very grateful to Joe Fera. There have been times when my best buddy, Joanna
Tice has carried me on her broad swimmer’s shoulders, and brought great joy, companionship and
strength into my life, and for this I thank her sincerely. I have shared in a condensed American
dream with the Bravos, Daniel, Felipe, Valentina, Laura, and Claudia Antonio, starting with
little and getting so much. It has also been a lot of fun, and I’d like to thank everyone for being
a part of it, Anna Haensch, Nathan Boon, Annie Rorem, Hiram Navarrete, Eric and Shannon
Paul, Charlie Mcintosh, James Ricci, Eliz Cox, Nicole Bobitski Olcese, Eran and Marmo Bugge,
and Evan Jones. As well as the MVC and the eternal email chain, Mike Shea, Mike Martin, Dave
Kanner, Seamus Scott, Beth Weil, Justin Hillman, Kevin Leahy, and Danny Shekhtman.
And of course my family, the bedrock of my support, Shirlie, Theodore, Karen, Michael, Andrew,
Natalie, and David Blum, Samuel and Estelle Ginsberg, and Diane, Eddie, Jordan and Rebecca
data for Reλ > 104, some of the largest Reynolds numbers ever measured. They find that the
structure functions conditioned on the large scale velocity show a strong dependence, and they
also show that direct numerical simulation in a periodic box (DNS) and passive grid turbulence
measurements show almost no dependence. They attribute the dependence in the atmospheric
boundary layer to large scale shear, which is not present in the DNS and passive grid turbulence.
They also show how to remove the large scale dependence through experimentally found fit
parameters in order to improve power law scaling. Kholmyansky and Tsinober [10] also measured
conditional structure functions in a high Reynolds number atmospheric boundary layer and
found a dependence that is weaker than Sreenivasan and Dhruva found. They attribute the
strong dependence they observe to a direct and unavoidable coupling between the large and
small scales in turbulence [11].
A natural initial objection when discussing how the small scales depend on the large scales is that
the Reynolds number is not sufficiently large enough. One recalls how important a sufficiently
high Reynolds number is to Kolmogorov’s hypotheses. However, the atmospheric boundary layer
data is measured at some of the highest Reynolds numbers ever recorded. We cannot discount
the possibility that any infinite Reynolds number system has small scale statistics that are always
independent from the large scales. However, this is an asymptotic limit, whose value is doubtful
in the real world. Our aim is to better understand turbulent flows as they actually are, and to
some extent move away from analyzing the idealized constituent parts. Upon accepting that
real world flows almost always have small scales which depend on the state of the large scales
our work tries to quantify what properties of the large scales are responsible for the observed
dependence.
One challenge in discussing interactions between large scales and small scales is the very non-
Chapter 1 - Introduction 9
universal nature of the large scales. Each flow has a unique set of large scales, which may
depend on time, geometry, or driving parameters. So it has been difficult to isolate the aspects
of the large scale flow that are affecting the small scales. Anisotropy is the aspect that is best
understood. Extensive work has identified persistent anisotropy at small scales even at very
high Reynolds numbers [9, 12], and analysis using spherical tensor decomposition has placed
this problem on solid footing [13, 14]. However, this is not the only effect of the large scales.
Here we have found two additional aspects of the large scales that are particularly important.
Inhomogeneity is the spatial variation of statistics, and will be discussed in chapter 3. Large
scale intermittency is temporal fluctuations on time scales longer than the eddy turnover time,
L/U , and will be discussed in chapter 4. Both inhomogeneity and large scale intermittency
often occur together in real flows, but are distinct properties since flows can be conceived that
have each without the other. For example, a homogeneous turbulent flow in DNS can have large
scale intermittency by having the energy injection varied in time.
It is my hope that this work can be helpful to better understand the dependence of small scale
statistics in turbulence on non-universal large scales which will help in the identification of
universal statistics, and the comparison of different flows.
Chapter 2Apparatus and Experiment
This work1 is based on optically tracking passive tracer particles seeded in a turbulent flow. Es-
sentially, 3D positions are determined by the 4 cameras in a similar fashion to 2 eyes determining
the 3D position of an object in nature. Tracking each particle is accomplished by comparing
adjacent frames and identifying a particle by its possible trajectory as it moves through the
frames. Once a particle has been identified over several frames its velocity is then determined
by calculating the distance traveled over time. After we have acquired a significant amount of
velocity data we can begin our analysis which this thesis is based on. Our system has the distinct
advantage of real time image compression; with compression factors of 100-1000 data could be
acquired continuously and nearly endlessly. This allowed for greatly increasing the amount of
data that could easily be acquired using a relatively simple apparatus. In later chapters we will
see how the key contributions of conditional statistics require very large amounts of data.
This chapter will describe the experimental apparatus and the protocols used in the experiments.
I will describe the turbulence tank in section 2.1, the detection of the particles in section 2.2,
the calibration method in section 2.3, 3D particle finding in section 2.4, and the method for
tracking and velocity determination in section 2.5.
1Note this chapter is largely adapted from work that has been published as Blum et al. “Effects of nonuniversal
large scales on conditional structure functions in turbulence” Physics of Fluids 22, 015107 (2010)
10
Chapter 2 - Apparatus and Experiment 11
1 m
1.5 m
Figure 2.1: Experimental apparatus diagram. Two oscillating grids were held 56.2 cm apart in an 1,100
l octagonal prism Plexiglas tank. Four high speed cameras were used to stereoscopically image an illuminated
volume in order to record 3D particle positions. Illumination was provided by a Nd:YAG laser with 50 W average
power.
2.1 The turbulence tank
The turbulence tank is a 1 × 1 × 1.5 m3 Plexiglas prism and is filled with approximately 1,100
l (300 gallons) of filtered, degassed water. Two grids generate the turbulence, the grids have 8
cm mesh size, 36% solidity, and are evenly spaced from the top and bottom of the tank with
a 56.2 cm spacing between the grids, and a 1 cm gap between the grids and the tank walls as
shown in figure 2.1. The stroke was 12 cm peak to peak, powered by an 11 kW motor. A typical
grid frequency was 3 Hz, but could be raised up to 5 Hz safely. Water cooling maintains the
temperature at ± 0.1◦ C during each run. Grid position is determined by a simple photogate
placed underneath one of the rods such that the oscillation of the grids breaks the photogate’s
beam. The grids were at their bottom position when the photogate beam has been broken for
half of the total time it was broken. This position was assigned the phase angle 0=2π.
The water used for the experiment was degassed overnight and filtered through 2 filters, the first
had a 50µm pore size and the second 0.2 µm. Keeping the water and the tank free from debris
and air bubbles was a challenge. To address debris in the tank (often composed of biological
Chapter 2 - Apparatus and Experiment 12
material) we found degassing did an adequate job inhibiting biological growth, replacing the need
for adding antibacterial chemicals. However, some debris and biological material still found its
way into the tank via the hoses running into the tank and times when a port is open in the tank.
Another possible source of debris was corrosion in the tank, although only anodized aluminum
and plastic come in contact with the water to try to minimize corrosion, it still persisted. One
possible reason for the corrosion was the a chemical reaction between the different aluminum
alloys used in the grids and the top and bottom surfaces of the tank. This could be minimized by
resurfacing or replacing the grids. To remove as much debris as possible the water was filtered
continuously while the motor occasionally stirred the water. This was moderately successful,
however this introduced air bubbles in to the tank which should be avoided so as not to be
confused with tracer particles when recording. An additional source of air bubbles in the tank
was air being drawn in to the water past the seals. To minimize this source the seal housing was
modified so that both sides of the seal have a reservoir of water. This effectively eliminated air
bubbles entering the water while the motor was running. Overall, degassing overnight had the
benefit of minimizing the effect of any air bubbles that were introduced in to the tank. Although
no completely satisfactory solution was found to eliminate debris and air bubbles in the tank,
these contaminants were estimated to represent less than 1% of the particles found, and usually
not tracked.
Neutrally buoyant 136 µm diameter polystyrene tracer particles were added to the flow. While
2 cameras were in place the seeding density could be up to 50 particles per frame without
significant stereomatching errors. After 4 cameras were in place the seeding density could be
raised to over 180 particles per frame without significant stereomatching errors. This has not
been limit tested due to the expense of particles.
One difficulty inherent in the oscillating grid experiment is the vibrations introduced from the
oscillatory driving. The benefits of imaging particles with high spatial resolution could be lost
if the imagers themselves are vibrating significantly. To minimize the source of vibrations the
flywheel was milled to precise tolerances, and lightweight Aluminum was used in the construction
of the grids. To minimize vibrations coupling to the cameras, a lightweight custom built camera
support was used to hold the cameras rigidly with respect to each other. In addition, the camera
support is mounted on an optical table that is not connected to the tank, and rests on rubber
vibration reducing pads.
Chapter 2 - Apparatus and Experiment 13
It should be noted the Plexglas tank did rupture once, before I began working on the project.
The tank was filled with cold water, then the all-thread rods were tightened. The tank was then
drained and refilled with warmer water which caused the Plexiglas to thermally expand under
compression. The Plexiglas joints failed, spilling 300 gallons of water onto the floor. The tank
was repaired and the all thread rods were buffeted with springs, so the Plexiglas could thermally
expand more easily, which has proven successful to the present date.
2.2 Detection
These data were acquired using three dimensional particle tracking velocitmetry measurements.
The data in chapter 3 were acquired using 2 Bassler A504K video cameras capable of 1280 ×1024 pixel resolution at 480 frames per second (a data rate of approximately 625 Mbyte per
second per camera). Later, 2 newer cameras were added, model Mikrotron MC1362 which have
similar pixel resolution and data rates, but have the advantage of greater sensitivity. The noise
(frame to frame deviations from the mean) depend more on brightness level than camera. The
key difference in camera model is the sensitivity; given the same light on each pixel the Mikrotron
cameras can give a much higher pixel value, thus allowing them a greater dynamic range.
Recording such high data rates with 4 cameras is a significant technological hurdle. A well
equipped desktop system could store data in 4 Gbyte of video RAM, so that one run could last
just 7 s before waiting approximately 7 min for the data to download to the hard disk. We
have developed an image compression circuit to threshold images in real time so that only pixels
above a user defined brightness limit are regarded as particle data and retained, while the dark
background pixels are discarded [15]. This technique produces a dynamic data compression
factor of 100-1000, which enables continuous data collection and storage to hard disk.
Our first implementation of the image compression circuit faced two major challenges. First, the
simple thresholding compression reduces particle center accuracy. However, particle finding is
typically degraded by only 0.1 pixel, which is typically less than the uncertainty in the particle
finding from unthresholded images. To help this, plans are underway to implement a nearest
neighbor algorithm which will record pixels adjacent to the bright pixels. Second, because
frame number information was created and recorded separately on each computer, any operating
system delay can lead to frames lost and timing mismatch between the cameras. Some frame
Chapter 2 - Apparatus and Experiment 14
number errors were corrected in postprocessing. Updated versions of the image compression
circuit solved this problem by including camera frame number in the data stream the computers
record.
Particles are illuminated using a 532 nm pulsed Nd:YAG laser with 50 W average power. The
beam was expanded to create an illumination volume approximately 7 × 4 × 5 cm3. The Condor
model laser manufactured by Quantronix corporation has had some operational issues that
required a return shipment to the manufacturer. The issues were: a faulty transient absorber on
the power control circuit, loose wires, and biological growth in the deionized cooling water line.
These issues were addressed, and the maintenance procedure now includes powering down the
laser from the circuit breaker, and running the laser for at least 1 hour each month to inhibit
biological growth.
2.3 Calibration
The goal of the calibration is to set up a real space coordinate system with a known origin and
camera positions. This will be used to find the 3D particle positions in real space. The calibration
procedure involved placing a calibration mask into the tank at the desired observation volume,
filling the tank with water, and recording calibration images of the well illuminated calibration
mask. The calibration mask (manufactured by Applied Image Inc.) is simply a 50 × 100 mm
sheet of glass with an array of 0.2 mm dots placed 2 ± 0.001 mm apart. The perspective each
camera has of the known calibration mask distorts the regular array of dots slightly, the columns
farthest away seem contracted. This distortion contains enough information for the cameras 3D
position to be calculated. A detailed description of the procedure and algorithm used for the
measurements in chapter C can be found in Ref. [16] and [17]. A similar procedure, although
different algorithm was used for the measurements in chapter 4, greater detail can be found in
refs. [18],
This traditional calibration method can obtain a particle finding accuracy of approximately 2 pix-
els, or 100 µm. Particle finding accuracy can be greatly increased by using known stereomatched
pairs from the cameras and running a nonlinear optimization to minimize the stereomatching
error and find optimal camera position parameters. This can increase particle finding accuracy
to less than half of a pixel.
Chapter 2 - Apparatus and Experiment 15
2.4 3D Particle Finding
The procedure for finding particles in 3D is essentially creating a 3D ray which starts at the
known camera position and goes through the particle which that camera sees. With multi-
ple cameras all seeing the same particle the rays should intersect at one point in 3D space,
which is the position of the particle. This technique, referred to as stereomatching, has 3 chal-
lenges.
First, in real world experiments the rays never exactly intersect. There is always some mea-
surement error, both in the 2D images and in tracing the real space ray pointing to the particle
(having a straight ray is an approximation of how the light travels to the cameras). Fortu-
nately, these effects are small, and the stereomatching intersection error can be less than half a
pixel.
Second, there are pathologies which can yield false particles. One can imagine two rays that
intersect, but do not have a particle where they intersect. Instead, there are two particles each
out of the frame of the other camera, but they are positioned in such a way that the rays
intersect in the observation volume, and report a false particle. Fortunately this pathology is
rare, and becomes exceedingly rare when more than two cameras are used.
The third, and most challenging problem, is that of particle density. Imagine a particle seen
on one camera is a ray seen from another camera. If the particle density is low, and each
camera sees only one particle in its frame there is no ambiguity as to which which ray on one
camera is pointing to which particle seen on a different camera. If each camera sees a handful
of particles calculating the position of a particle would require taking one ray on one camera
and calculating which rays on the other cameras it comes closest to intersecting. The correct set
of rays will be much closer to intersecting than any other combination of ray sets. This works
fine for low particle densities. The problem occurs when an experiment calls for a high particle
density, the brute force method of trying every ray set combination quickly breaks down. For
example if there are 600 particles in a frame, the brute force method requires one ray from
one camera to be matched to every ray from each other camera. With 4 cameras this requires
6004 or 1.3×1011 combinations, even at 1 millisecond computation time per combination to
evaluate the 3D matching, (which is a generously fast estimate) each frame would take over 4
years of computation. This is for only 1 frame, our measurements require millions of frames.
Chapter 2 - Apparatus and Experiment 16
Camera 1’s view
Camera 2
[x1,y
1] Camera coordinates of
particle seen in camera 1
[xc,y
c] Coordinates of camera 2
if it could be seen by camera 1
[x2,y
2]
+
θ1
θ2
θ1=tan-1( )
yc-y
1
xc-x
1
θ2=tan-1( )
yc-y
2
xc-x
2
Figure 2.2: Diagram of angle discrimination definitions. The red + is the particle seen on camera 1. The red
o can be placed anywhere along the ray, it does not effect the angle θ2.
We have developed a solution to this problem called angle discrimination which, although more
conceptually difficult, greatly reduces the computation time.
The key to solving this problem is to limit the number of ray set combinations to try. Consider
two cameras on a horizontal plane pointed at the same observation volume. If one camera sees
a particle at the very top of its frame we can safely choose only rays originating near the top of
the frame of the second camera to try for an intersection. Even if we eliminate half of the rays,
the number of combinations to try has greatly decreased. Taking this concept further, and using
some simple geometry, we can do better still. We can accurately predict which particles seen on
a 2d frame are viable candidates to be matched to particles seen on other cameras, using only
their position and the position of the cameras.
Figure 2.2 shows a diagram for a method of determining which rays are good candidates for
3D matching. Camera 1’s field of view is shown, it sees only one particle denoted by a red +.
Camera 2 cannot be seen by camera 1, but since we know it’s 3D coordinates from the calibration
we can project it on to camera 1 coordinate space. In other words, if camera 2 is a point particle
camera 1 sees it at [xc, yc]. In this diagram camera 2 is directly above camera 1. The ray which
travels through the red circle is the most important part of this diagram. Camera 2 can point
Chapter 2 - Apparatus and Experiment 17
a 3D ray in real space to the particle it sees, here the particle it sees is drawn as a red circle.
The ray shown in the diagram is the 2D projection of that ray on to camera 1 coordinate space.
Note the red circle can be placed anywhere along the ray without affecting θ2. The angles θ1
and θ2 can now be compared. Only rays that point to the same particle will have similar angles,
typically within small fractions of a degree. Each frame needs to be sorted by angle only once,
and the angle can quickly be used to discriminate against unlikely rays. This comparison can
be calculated very quickly and can greatly reduce the number of combinations of rays needed
to find the ray set that intersects at a real particle position. The increase in speed depends on
the tolerances used for the angle discrimination and the particle density, but an average frame
can now be processed in under 1 second, far better than 4 years.
2.5 Particle Tracking and Velocity
After the 3D positions of the particles have been determined a particle tracking program (writ-
ten by John C. Crocker and Erik Weeks 2) is run which identifies particles through multiple
frames. This works by recording the particle positions on one frame, then considering all pos-
sible new positions on the next frame. It then calculates all possible identifications of the old
positions with the new positions and minimizes the total displacement. Parameters such as
maximum displacement between frames and gaps is a track are considered. This program has
proven adequate for the experiment described here. An improved version (which was started)
could include the probable particle trajectories and splice particle tracks together that became
separated by a particle absent over several frames.
After particles tracks were identified, particle velocities were then calculated. The straightfor-
ward derivative calculation ~v = ∆(~x)/∆(t) is too noisy in this experiment. Instead, a particle
trajectory was fit to a second order polynomial function. The derivative of this function was
then found, and the velocity was determined. This method reduces the noise in the velocity
measurement but can artificially smooth over velocity features if the fit is applied over too many
frames. Details of finding the best duration to fit over can be found in Voth et al. [19].
2http://www.physics.emory.edu/ weeks/idl/
Chapter 3Effects of Non-Universal Large Scales on
Conditional Structure Functions
This chapter1 reports measurements of conditional Eulerian and Lagrangian structure functions
in order to assess the effects of nonuniversal properties of the large scales. As stated in the
introduction, a foundational concept in the study of turbulence is that of the energy cascade.
A process where kinetic energy enters the system at a large scale, and cascades down to smaller
and smaller scales through the chaotic break up of eddies. It is supposed that in this chaotic
process large scale information is lost by the time the kinetic energy is at the small, universal
scales and should be independent of the large scales. The purpose of these experiments is to
explore this notion of the small scale independence from the large scales.
We begin in section 3.1 by characterizing the flow. Section 3.5 discusses conditioning structure
functions in order to assess large scale influence. In section 3.7 specific properties of the large
scales are considered.
1Note this chapter is largely adapted from work that has been published as Blum et al. “Effects of nonuniversal
large scales on conditional structure functions in turbulence” Physics of Fluids 22, 015107 (2010) [20]
18
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 19
Mea
n V
ertic
al V
eloc
ity (
cm/s
) V
ertic
al V
eloc
ity V
aria
nce
(cm
2 /s2 )
-6
-4
-2
0
2
4
6
-20 -15 -10 -5 0
300
250
200
150
100
50
0
Distance From Center (cm)
a
b
Center
Region
Near
Grid
Region
Figure 3.1: Mean and variance of the vertical velocity along the central vertical axis of the tank. Grid
frequency is 3 Hz and grid separation distance is 56.2 cm. The dot-dash line represents the grid height at
maximum amplitude. We will focus on measurements in the two regions designated by the vertical dashed lines:
one at the center of the tank and one near the grid.
3.1 Characterizing the Flow
With the large scales being such an important part of this discussion it is important to adequately
characterize them in this particular flow. Through a good deal of effort a flow profile was created
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 20
-28.0
-16.8
-5.6
5.6
16.8
28.0D
ista
nce
From
Cen
ter
(cm
)
C
NG
Top Grid
Bottom Grid
. .
Figure 3.2: Scale diagram of 56 cm × 100 cm area between grids showing the mean circulation torii which
are nearly rotationally symmetric about the central vertical axis. Center (C) and near grid (NG) overreaction
volumes are drawn in dashed lines, which shows the relative size and position of the observation volume in Fig.
3.1. Horizontal dot-dashed lines represent the range of motion of the top and bottom grids.
that spans from the middle of the tank to bottom grid as shown in figure 3.1.
We define a characteristic velocity by U = (〈uiui〉/3)1/2 and a characteristic length scale by
L = U3/ε where ε is the energy dissipation rate per unit mass defined in section 3.3. For the
center region U = 6.0 cm/s, L = 9.0 cm, and for the near grid region U = 8.3 cm/s, L = 4.5 cm.
The Taylor Reynolds number, Reλ = (15UL/ν)1/2, (where ν is the kinematic viscosity) ranges
from 285 for 3 Hz grid frequency to 380 for 5 Hz grid frequency in the center. Near the grid
at 3Hz Reλ = 230. The Kolmogorov length and time scales are η = 140µm, τη = 20 ms in the
center region, and η = 94 µm, τη = 8.8 ms in the near grid region.
Figure 3.1a shows the mean vertical velocity as a function of the vertical position along the
central axis of the tank. The top and bottom grids are separated by 56.2 cm, approximately 7L.
In Fig. 3.1 the dot-dashed line indicates the maximum amplitude of the bottom grid, 22.1 cm
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 21
below the center of the tank. Data were collected at five separate heights in order to measure the
complete flow profile from the center to the bottom grid. Mapping the bottom half of the talk is
sufficient because the geometrical symmetry produces a mirror image above the midplane. The
two volumes which we will focus on are bounded by the dashed lines and will be referred to as
the center and near the grid observation volumes. At this grid separation distance the mean flow
traces four torii, two above and two belove the center plane of the tank, as shown in the sketch
in Fig. 3.2 (drawn to scale). In the large central region, the effect of the mean flow is to pump
highly energetic fluid from the region near the grid toward the center of the tank. In Fig. 3.1(a),
there are two points where the mean vertical velocity approaches zero: one near the center; the
other 18 cm from the center just below the near grid observation volume. The existence of this
second stagnation point and reverse circulation region depicted in Fig. 3.2 is a common feature
in mean flows generated by oscillations [21]. In all of the following measurements, the mean
velocity field has been subtracted so that we study the fluctuating velocity.
Figure 3.1(b) shows the vertical velocity variance along the central axis as a function of the
vertical position. The velocity variance is large near the grid and quickly falls off toward the
center where it is nearly homogeneous. The center and near grid observation volumes were chosen
to provide a contrast between the large homogeneous region in the center and the much more
inhomogeneous region near the grid. In the center, the variance of the velocity is homogeneous
for several L in either direction. The velocity variance ranges moderately in the near grid
observation volume and enormously with one L below this region. In Fig. 3.1 deviations from a
smooth curve are not due to statistical uncertainty, but are a result of patching five calibrated
regions together with the majority of error coming from measuring absolute position in the
tank.
It is interesting to note that we made measurements in a flow with smaller grid separation of
35 cm and found the Reynolds number in the center was lower. The characteristic velocity in
the center did increase due to the closer proximity of the grids, but L was reduced by a larger
amount resulting in approximately 8% decrease in Reλ. The reason for the unexpected decrease
in Reynolds number is a reversal of the mean velocity compared with larger grid separations.
For larger grid separation distances, energetic fluid from near the grids is carried to the center
by the mean flow. However, at 35 cm grid separation the mean velocity reverses which results
in a lower Reynolds number in the center.
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 22
10 1000.1
1.0
10.0
< ∆ur >2
(εr)2/3
r/η
r/η10 100
0.5
1.0
1.5
2.0<
∆u r
>2
(cm
/s)2
Figure 3.3: Eulerian second order longitudinal velocity structure function shown as a function of pair separation
r normalized by the Kolmogorov length η. The inset shows this data compensated by Eq.3.1 for p=2.
3.2 Structure Functions
We focus on the second order structure function because it is a easily measurable quantity that
has been studied thoroughly in the literature, is scale dependent, is clearly predicted in turbu-
lence theory, and requires an amount of statistics that is accessible to our measurements. To
measure the Eulerian structure functions we first find the instantaneous longitudinal velocity
difference between two particles a distance r apart ∆ur = [u(x)-u(x+r)]L, where the L sub-
script denotes the longitudinal component, found by projecting the 3D velocity difference vector
onto the vector connecting the two particles. The longitudinal structure functions are defined as
Dp = 〈(∆ur)p〉 where p represents the order of the structure function and the brackets represent
the ensemble average. In the inertial range, Kolmogorov (1941) gives
〈(∆ur)p〉 = C(E)
p (εr)p/3 (3.1)
where C(E)p are the Eulerian Kolmogorov constants and ε is the energy dissipation rate.
Figures 3.3 and 3.4 show measured second and third order longitudinal velocity structure func-
tions with the straight thin lines representing Kolmogorov’s prediction from Eq.(3.1). The insets
show the structure functions compensated by Eq.(3.1). At Reλ = 285, any scaling range is very
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 23
10 1000.01
0.10
1.00
10.00
r/η
< ∆ur >3
(εr)
r/η10 1000.0
0.2
0.4
0.6
0.8
1.0<
∆u r
>3(c
m/s
)3
Figure 3.4: Eulerian third order longitudinal velocity structure function. The inset shows this data compen-
sated by Eq.3.1 for p=3.
limited, but the plateaus can be used to estimate the inertial range.
Lagrangian structure functions were measured from temporal velocity differences along a particle
trajectory. The velocity difference now becomes ∆uτ = u(t) − u(t + τ), where τ is the time
interval between measurements. We use the vertical velocity for Lagrangian velocity differences
throughout this thesis, although results for the other components are similar. For Lagrangian
structure functions Kolmogorov (1941) predicts
〈(∆uτ )p〉 = C(L)
p (ετ)p/2. (3.2)
3.3 Energy Dissipation Rate Measurement
The energy dissipation rate ε is an important value throughout this analysis, let us discuss
how it is determined. Limitations in particle density preclude direct measurement via the
definition
ε = 2ν〈sijsij〉
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 24
with
sij =1
2
(
∂ui
∂xj+
∂uj
∂xi
)
.
Instead we utilize Kolmogorov’s 4/5 law: Eq. 3.1 with p = 3 where the coefficient C(E)3 = −4/5
can be derived from the Navier-Stokes equations. We identify the inertial range with the plateau
in the compensated third order structure function (Fig. 3.4 inset). The inertial range is chosen
to be 25 to 91 r/η (.35 to 1.3 cm). If the same inertial range is used in the second order
structure function, the energy dissipation rate determined from it (using the empirical coefficient
C(E)2 = 2.0) [2] is within 3% of the value calculated from the third order.
3.4 Phase Dependence
A simple energy cascade has constant energy input at the largest length scales. An obvious
departure from constant energy input is the oscillating grid driving mechanism. The sinusoidal
motion of the grid directly corresponds to energy with periodic time dependence. It seems likely
that such a strongly periodic energy input would have a signature throughout the whole energy
cascade.
The method we employ throughout this work to detect signatures of the large scales is to
condition structure functions on some measurement of the state of the large scales. This utilizes
the structure functions described in 3.2 and determines how they depend on the large scales.
It is using this method we can comment on how the large scales effect all scales in a turbulent
flow, from near the characteristic length scale L, down to near the dissipation scales η. In this
case, we condition of the phase of the grid motion, φ. Conditioning instantaneous single particle
statistics such as the mean and variance of the velocity shows some sinusoidal dependence on
grid phase. For example, in the center the conditional variance, 〈(u − 〈u〉)2|φ〉, varies by 1%
over the cycle of the gird. In the near grid region, the conditional variance varies by 10%. The
mean vertical velocity in the center 〈u|φ〉, varies by 0.8 cm/s over a cycle of the grid which is
10% of the standard deviation. Near the grid, the conditional mean velocity varies by 2 cm/s
which is 20% of the standard deviation at that location.
Figure 3.5 shows the second order longitudinal structure functions conditioned on phase, 〈∆u2r|φ〉.
In the center of the flow (fig.3.5(a)) the structure functions have essentially no change with
phase. Near the grid (3.5(b)) there is a slight dependence on grid phase. To emphasize the
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 25
10 1001.0
1.2
1.4
1.6
1.8
2.0
< ∆ur | φ >2
(ε3r)2/3
r/η3
10 100r/η
3
1.0
1.2
1.4
1.6
1.8
2.0
< ∆ur | φ >2
(ε3r)2/3
a
b
Figure 3.5: Second order compensated velocity structure functions conditioned on grid phase. The collapse
shows the very weak phase dependence: (a) center of the tank, (b) near the grid. Zero and 2π phase represents
differences between structure functions at different phases, we compensated the structure func-
tions by a single energy dissipation rate in each figure, ε3 = 2.46 × 103 mm2/s3 in the center
and ε3 = 1.31 × 104 mm2/s3 near the grid. These values were determined when the grid is
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 26
in mid-amplitude (the third bin). The good collapse of the structure functions at all phases
across the entire range of r shows the minimal dependence of the small scales on the large scale
periodicity of the flow created by the oscillating grids. One possible source of dependence of
small statistics on the large scales has been shown to be minimal.
3.5 Dependence on Large Scale Velocity
3.5.1 Eulerian structure functions conditioned on the large scale ve-
locity: Center region
A more revealing dependence on the large scales of the flow is found by conditioning the velocity
structure functions directly on the large scale velocity. A convenient measurable quantity that
reflects the local instantaneous state of the large scales is the average velocity of the particle
pair used for the structure function, defined as Σu = (u(x)+u(x+ r))/2. This velocity sum can
also be decomposed into longitudinal (Σu‖) transverse Σu⊥) or directional Σui) components.
Alternatively, conditioning on the average velocity of many particles, not just one pair, was
studied and found to have similar results, but we choose to focus on Σu because it can be more
easily measured and does not depend on the observation volume and seeding density. Additional
conditioning quantities will be discussed in Sec.3.7.
Figure 3.6(a) shows the second order Eulerian velocity structure function conditioned on Σuz.
The smallest values of the structure function correspond to pair velocities near zero, represented
by ⋄, while large |Σuz| results in larger values of the structure function. For the bins we chose, the
structure function conditioned on large values of Σuz, is nearly twice the value when conditioned
on Σuz near zero.
Figure 3.6(b) shows the data in Fig. 3.6(a) compensated by Kolmogorov inertial range scaling.
The functional forms are quite similar, confirming the impression from Fig. 3.6(a) that all length
scales are affected similarly by the instantaneous state of the large scales. In Fig. 3.6(b) we
used a different energy dissipation rate εuzto compensate each of the five individual large scale
vertical velocity bins. This insures all conditions plateau at approximately the same value and
allows for direct comparison of the functional forms of the conditional structure functions.
The strong dependence of the conditional structure functions on the large scale velocity at all
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 27
10
100
10 1000.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
r/η
a
b
< ∆
u r |
Σu z
>2
(εΣ
u r)
2/3
z
< ∆
u r |
Σu z
>2
(cm
/s)2
Figure 3.6: Second order velocity structure function conditioned on particle pair velocity (vertical component)
in the center of the tank. (a) Uncompensated structure function. (b) Individually compensated by the energy
dissipation rate for each conditional data set. Symbols represent the following dimensionless vertical velocities,
Σuz/√
〈u2z〉: + = 4.2 to 2.5, ∗ = 2.5 to 0.84, ⋄ = 0.84 to -0.84, △ = -0.84 to -2.5, � = -2.5 to -4.2.
scales reveals that the small scales are not statistically independent of the large scales in this
flow. There is no detectable trend toward the smaller scales becoming less dependent on the
large scale velocity than somewhat larger scales.
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 28
10-1
100
101
102
1
2
3
4
5
< ∆
u τ | Σ
u z >2 (εΣ
u τ
)z
10-1 100 101τ/τ
η102
a
b
< ∆
u τ | Σ
u z >
2(c
m/s
)2
Figure 3.7: Second order Lagrangian velocity structure function conditioned on instantaneous velocity (vertical
component) in the center of the tank. (a) Uncompensated structure function. (b) Individually compensated to
have the peak values match. Symbols represent the following dimensionless vertical velocities, Σuz/√
〈u2z〉: + =
3.1 to 1.9, ∗ = 1.9 to 0.62, ⋄ = 0.62 to -0.62, △ = -0.62 to -1.9 , � = -1.9 to -3.1.
3.5.2 Lagrangian Structure Functions Conditioned on the Large Scale
Velocity: Center Region
In much the same way we can evaluate the conditional Lagrangian structure functions. Figure
3.7(a) shows the second order Lagrangian structure function conditioned on the vertical compo-
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 29
nent of the large scale velocity, Σuz. Here we find Σuz by averaging the velocity of the particle
at the two times used to determine ∆uτ . The conditional structure functions for different large
scale velocity are different by a factor of about 2.5 and they remain nearly parallel throughout
the entire time range.
Figure 3.7(b) shows the second order conditional Lagrangian structure function compensated by
Eq. 3.2, where ε is individually chosen so the maxima of all the conditioned structure functions
coincide. This aids comparison of the functional forms of the conditioned structure functions.
Again, the functional form is nearly identical for different large scale velocities, indicating that
the large scales affect all time scales in the same way. There may be a small trend toward larger
values of the compensated Lagrangian structure functions at small times when the magnitude
of the large scale velocity is large.
It should be noted that there is a bias present in Lagrangian measurements that is not present
in Eulerian measurements. A sample of measured Lagrangian trajectories is biased toward low
velocity particles since the high velocity particles are more likely to have left the measurement
volume. This bias becomes larger for larger τ . Berg et al. [22] studied this bias and find
that it can be quite large for typical experimental conditions. We quantified this bias in our
data by measuring the Lagrangian structure functions using trajectories that remained inside
artificially restricted measurement volumes. From a simple extrapolation of the dependence
on the size of the artificial detection volume, we estimate that our experimental Lagrangian
structure functions underestimate the true value by 17% for τ = 8τη. This is roughly consistent
with the size of the error we expect based on the critical time lag defined in Ref. [22]. Note that
we have not performed the compensation they recommend and we are roughly translating their
uncompensated results. Because of this bias, we will focus attention on τ < 10τη. As we will
discuss in Sec.3.6.3, the dependence of the conditioned Lagrangian structure functions on the
large scale velocity does not seem to be significantly influenced by this bias.
3.5.3 Eulerian Structure Functions Conditioned on the Large Scale
Velocity: Near Grid Region
By comparing separate regions of the tank we are able to explore the effects of inhomogeneity
on this conditional dependence. Figure 3.8(a) shows the Eulerian structure functions, similar
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 30
to Fig. 3.6(b), but with data collected near the inhomogeneous region near the bottom grid.
The separation between the Eulerian structure function conditions doubled to approximately a
factor of 4. Note the different ordering of the structure functions. The up-down symmetry is
now broken since these data were not taken in the center of the tank. Fluid traveling upwards (∗symbols) has a large structure function, while fluid traveling downward with the same magnitude
vertical velocity (△ symbols) has the lowest value of the structure function. We interpret this
as highly energetic fluid originating near the bottom grid and being turbulently advected into
the observation volume. Similarly, fluid carried down from the more quiescent region above the
detection volume has low energy and a smaller structure function.
Figure 3.8(b) shows the compensated Eulerian structure functions, similar to Fig. 3.6(b), but
reveals a novel insight. Stepping through the vertical velocity bins is equivalent to stepping
through the energy cascade. Fluid coming directly upward from the bottom grid (symbol +)
carries energy that was recently injected into the large scales. As a result, the compensated
structure function for upward moving fluid is biased toward the large scales. Fluid that has
downward vertical velocity (symbol △) comes from the center region far away from the grid. It
has had more time to mature, and in this process the energy is transported to smaller length
scales. Conditional structure functions appear to be effective tool to evaluate whether or not a
turbulent flow is fully developed and has established a stable cascade.
3.5.4 Third order Eulerian Structure Functions Conditioned on the
Large Scale Velocity: Center Region
Figure 3.9 shows the third order structure function individually compensated and conditioned
on Σuz in the center of the tank. Convergence of third order statistics was more difficult, so
elimination of the two extreme conditions was required. The third order structure function
proves to be similar to the second order in separation, symmetry and collapse to a single func-
tional form. The energy dissipation rates found for the three conditions are ε∗ = 25.2cm2/s3,
ε⋄ = 21.7cm2/s3,ε△ = 28.7cm2/s3.
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 31
3.6 A Powerful Method for Plotting Conditional Struc-
ture Functions
3.6.1 Eulerian Structure Functions Conditioned on the Large Scale
Velocity: Center Region
An alternative, and in many ways more powerful, method of visualizing the data is presented
in Fig. 3.10. Here we show the second order Eulerian structure function conditioned on the
1001.2
1.4
1.6
1.8
2.0
2.2
2.4
10
100
a
b
r/η
< ∆
u r |
Σu z
>2
(εΣ
u r)
2/3
z
< ∆
u r |
Σu z
>2
(cm
/s)2
Figure 3.8: Second order velocity structure function conditioned on particle pair vertical velocity (z direction)
in the region near the bottom grid. The condition with the largest downward velocity has been eliminated due
to lack of statistical convergence. Symbols represent the following particle pair vertical velocities Σuz/√
〈u2z〉:
+ = 3.8 to 2.3, ∗ = 2.3 to 0.75, ⋄ = 0.75 to -0.75, △ = -0.75 to -2.3, a) Uncompensated structure functions b)
Individually compensated by the energy dissipation rate for each conditional data set.
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 32
10 100
0.2
0.4
0.6
0.8
r/η
< ∆
u r |
Σu z
>3
(εΣ
ur)
z
Figure 3.9: Third order velocity structure function plots conditioned on particle pair vertical velocity and
individually compensated for each conditional data set. Data are taken in the center region of the tank, and the
extreme vertical velocity plots have been eliminated due to lack of statistical convergence. Symbols represent the
following vertical velocities Σuz/√
〈u2z〉: ∗ = 2.5 to 0.84, ⋄ = 0.84 to to -0.84, △ = -0.84 to -2.5.
vertical component of the large scale velocity (the same data as Fig. 3.6). However, the scaled
vertical pair velocity is plotted on the horizontal axis with the conditioned structure functions
on the vertical axis. When the structure functions are scaled by their value at Σuz = 0, we
find very good collapse of the data. The fact that these curves for different r/η collapse so well
is a striking demonstration that the large scales affect all length scales in the same way. The
fact that the conditional structure functions vary by a factor of 2.5 demonstrates the strong
dependence on the large scales. Note that for Gaussian random fields, Fig.5.5 is flat (this will
be discussed further in section 3.9, and again in 5.3). A nearly flat result is also observed in
DNS and grid turbulence [1] and reproduced in figures 5.2a and 5.2b. Note there is a significant
difference in the structure functions when the large scale velocity is larger than two standard
deviations away from the mean between these two simulations, this will be discussed in section
5.3. We will also discuss the measurement error in this type of plot in appendix A.
In Fig. 3.10 we see the dependence is a steep parabola. This plotting method makes clear the
extent to which all the smaller scales are affected by the large scale velocity; in fact, all length
scales collapse nearly perfectly onto one parabola. The large scale velocity affects all length
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 33
2
<Δur|Σu
z>
<Δur|Σu
z>
Σu = 0
2
Z
Σuz
√<uz>2
-4 -2 0 2 40.5
1.0
1.5
2.0
2.5
3.0
Figure 3.10: Eulerian second order conditional structure function versus large scale velocity. Data taken in
the center region. Each curve represent the following separation distances r/η: + = 0 to 40, ∗ = 40 to 70, ⋄ =
70 to 110, △ = 110 to 140, � = 300 to 370, × = 370 to 440.
scales in nearly the exact same way, all the way down to the dissipative range.
3.6.2 Eulerian Structure Functions Conditioned on the Large Scale
Velocity: Higher Reynolds Number
Figure 3.11 shows the effect of increasing Reynolds number. These data are at the center of the
tank with the grids oscillating at 5 Hz instead of 3 Hz which increases Reλ from 285 to 380. The
collapse of the structure function remains. The curvature in the this figure is not significantly
different from the lower Reynolds number data in Fig. 3.10 indicating that if there is a Reynolds
number dependence it is weak.
Figure 3.12 shows a comparison of our data with data taken in the atmospheric boundary
layer [1] with Reλ > 104. Atmospheric boundary layer turbulence shows a similar collapse
of conditional structure functions at all length scales. The curvature is also similar in both
data sets, indicating that the dependence on the large scales is similar even at these very large
Reynolds numbers.
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 34
2
<Δur|Σu
z>
<Δur|Σu
z>
Σu = 0
2
Z
Σuz
√<uz>2
-4 -2 0 2 40.5
1.0
1.5
2.0
2.5
3.0
Figure 3.11: Eulerian second order conditional structure function versus large scale velocity. Data taken in the
center region at higher grid frequency, 5Hz, resulting in higher Taylor Reynolds number 380. Symbols represent
the following separation distances r/η: + = 0 to 50, ∗ = 50 to 100, ⋄ = 100 to 150, △ = 150 to 200, � = 310 to
420, × = 420 to 520.
3.6.3 Lagrangian Structure Functions Conditioned on the Large Scale
Velocity: Center Region
Figure 3.13 shows the Lagrangian structure functions plotted versus the large scale velocity,
comparable to the Eulerian data shown in Fig. 3.10. The parabolic shape remains, but the
curvature is greater for all Lagrangian time scales that it is in the Eulerian data. All time scales
are affected by the large scale velocity. To determine the effect of measurement volume bias, we
have done this analysis for artificially restricted measurement volumes. By decreasing the volume
by a factor of 2, we observe the large τ curves shift by approximately the deviations between
the curves. We conclude that the bias does not have a significant effect on the conditional
dependence shown in Fig.3.13 for the time differences presented (τ ≤ 10τη).
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 35
2
<Δur|Σu
z>
<Δur|Σu
z>
Σu = 0
2
Z
Σuz
√<uz>2
-4 -2 0 2 40.5
1.0
1.5
2.0
2.5
3.0
Figure 3.12: Eulerian second order conditional structure function versus large scale velocity. The thin plots
are from atmospheric boundary layer data [1] r/η: ∗ ∼ 100, △ ∼ 400, � ∼ 1000, × ∼ 1250. The thick line is
from fig. 3.10, which has been overlaid for comparison, r/η: ⋄ = 70 to 110.
-4 -3 -2 -1 0 1 2 3 4
1
2
3
4
5
Σuz
√<uz>2
2
<Δuτ|Σu
z>
<Δuτ|Σu
z>
Σu = 0
2
Z
Figure 3.13: Lagrangian second order conditional structure function versus large scale vertical velocity. Data
taken in the center region. Symbols represent the following τ/τη : + = 0.42 , ∗ = 1.3, ⋄ = 3.5, △ = 10.
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 36
3.6.4 Eulerian Structure Functions Conditioned on the Large Scale
Velocity: Near Grid Region
Figure 3.14 shows a conditional Eulerian structure function similar to Fig. 3.10, but measured
in the inhomogeneous region near the grid. The structure function here is strikingly different
than in the center. The minimum is shifted by more than one standard deviation to the left.
The inhomogeneity breaks the up-down symmetry so that fluid coming directly up from the
bottom grid is markedly different than fluid coming down from the more quiescent region above
(analogous to the ∗ and △ separation in Fig.3.8). It follows that fluid with an upward velocity
has higher energy than fluid with the same velocity magnitude in the downward direction.
The atmospheric boundary layer data in Fig.3.12 also show this effect with a minimum at
Σuz/√
〈u2z〉 = −0.5, presumably as a result of weaker inhomogeneity. Also notable is that the
collapse of plots for various r values is not as complete as in the central region. This is consistent
with Fig. 3.8(b) which shows that the conditional structure functions have somewhat different
r dependence.
3.6.5 Lagrangian Structure Functions Conditioned on the Large Scale
Velocity: Near Grid Region
Figure 3.15 shows a Lagrangian structure function measured in the near grid region, similar to
the Eulerian data in Fig. 3.14. The minimum is shifted to the left here also as a result of the
inhomogeneity in this region of the flow. The conditional dependence on the large scale velocity
is again somewhat larger than in the Eulerian case, and the collapse at different time scales is
not as complete.
3.6.6 Third Order Eulerian Structure Functions Conditioned on the
Large Scale Velocity: Center Region
The third order Eulerian velocity structure function plotted versus the large scale vertical veloc-
ity is shown in Fig.3.16 using data from the center of the tank. Statistical convergence is weaker
than the second order which limits the large scale velocity range available for analysis. The
collapse seems similar to the second order case shown in Fig.3.10, although the measurement
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 37
-4 -2 0 2 40.5
1.0
1.5
2.0
2.5
3.0
2
<Δur|Σu
z>
<Δur|Σu
z>
Σu = 0
2
Z
Σuz
√<uz>2
Figure 3.14: Eulerian second order conditional structure function versus large scale velocity. Data taken in the
near grid region of the tank. The structure function is heavily influenced by the bottom grid which has skewed
the symmetry of the plot minima in the negative direction. Symbols represent the following non-dimensional
separation distances r/η: + = 0 to 50, ∗ = 50 to 110, ⋄ = 110 to 160, △ = 270 to 320, � = 330 to 450, × = 450
to 560.
uncertainties are larger here. The curvature seems to be slightly larger for the third order than
for the second order case.
3.6.7 Second Order Eulerian Structure Functions Conditioned on the
Velocity Magnitude: Center Region
The second order Eulerian structure function plotted versus the magnitude of the pair velocity
is shown in Fig.3.17 using data from the center observation volume. The magnitude of the pair
velocity is also a useful indicator of large scale activity. It has no preferred direction and it is
a more direct indicator of the instantaneous local energy. A similar dependence remains as in
Fig.3.10, the collapse seems similar, and the curvature is significantly larger.
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 38
-4 -3 -2 -1 0 1 2 3 4
2
3
4
5
6
Σuz
√<uz>2
2
<Δuτ|Σu
z>
<Δuτ|Σu
z>
Σu = 0
2
Z
Figure 3.15: Lagrangian second order conditional structure function versus large scale vertical velocity. Data
taken in the near grid region of the tank. Symbols represent the following τ/τη : + = 0.94, ∗ = 2.8, ⋄ = 8.0.
-2 -1 0 1 20.5
1.0
1.5
2.0
2.5
3.0
Σuz
√<uz>2
3
<Δur|Σu
z>
<Δur|Σu
z>
Σu = 0
3
Z
Figure 3.16: Eulerian third order conditional structure function versus large scale vertical velocity in the center
region. Symbols represent the following non-dimensional separation distances r/η: + = 0 to 40, ∗ = 40 to 70, ⋄
= 70 to 110, △ = 110 to 140, � = 220 to 300, × = 300 to 370.
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 39
0.0 1.0 2.0 3.00.5
1.0
1.5
2.0
2.5
3.0
2
<Δur|Σ|u|>
<Δur|Σ|u|>
Σ|u| = 0
2
Σ|u|2
√<|u| >
Figure 3.17: Eulerian second order conditional structure function versus magnitude of the velocity pair in the
center region. Symbols represent the following non-dimensional separation distances r/η: + = 0 to 40, ∗ = 40
to 70, ⋄ = 70 to 110, △ = 110 to 140, � = 300 to 370, × = 370 to 440.
3.7 Properties of the Large Scales and Their Effects
The measurements provided in this chapter show the dependence of all scales including the
inertial range and small scales on the current state of the large scales. In this section the
properties of the large scales that might be responsible for this dependence on the large scales
will be discussed. It is important to note that this list is not exhaustive, and involves some level
of speculation. However, once this dependence has been shown and quantified the next step of
identifying possible contributors to the large scale dependence is valuable. We begin to address
the effects of the properties of the large scales here, and continue by comparing different flows
with unique large scales in chapter5.
3.7.1 Reynolds Number
An immediate concern when discussing large scale effects is if the oscillating grid flow has
a Reynolds number insufficient for adequate scale separation, and it is this which leads to
contamination of the small scale statistics by the large scales. Evidence points to large scale
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 40
dependence not being caused by limited Reynolds number. The comparison in Fig. 3.12 shows
that atmospheric boundary layer data [1] with very large Reynolds number (Reλ > 104) have
nearly the same dependence on the large scales as our flow. Increasing the Reynolds number
in our flow makes very little difference. Additionally, all length scales collapse to nearly the
same functional form indicating that limed separation of scales is not the primary factor. Taken
together, these lead us to the conclusion that merely high Reynolds number alone is not enough
to create small scales that are statistically independent of the large scales.
3.7.2 Anisotropy
The flow between oscillating grids is somewhat anisotropic. The ratio of vertical to horizontal
velocity standard deviations is 1.5:1 in the center. The effects of large scale anisotropy on the
small scales have been studied extensively, [13] but we suspect it is not a major factor in the
conditional dependence studied here. We analyzed the data by averaging over particle pairs
with all orientations, so when the structure functions are conditioned on a quantity with no pre-
ferred direction like the velocity magnitude (Fig.3.17) there should be very little contribution
from anisotropy. In fact, we find that the conditional dependence on velocity magnitude is even
stronger than the dependence on the vertical velocity component. We also observe the condi-
tional dependence remains when conditioned on other quantities without preferred directions
like Σu‖ and Σu⊥. We believe anisotropy is not a significant cause of the conditional dependence
we observe, but anisotropic effects is an intriguing topic currently being investigated by another
graduate student in the lab.
3.7.3 Shear
Sreenivasan and Dhruva [1] attribute the strong conditional dependence of the Eulerian structure
functions on the large scale velocity to shear in the atmospheric boundary layer. In comparing
conditional structure functions from both the atmospheric boundary layer and the flow between
oscillating grids a similar dependence is found despite significantly different shear. The mean
velocity gradient normalized with the eddy turnover time is 1.2 in the center of the oscillating
grid flow, and is estimated to be in the range of 5 or greater for their atmospheric boundary layer
data. There must be some other properties that exist in the shear flow, but are also important
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 41
in our flow with small shear.
3.7.4 Inhomogeneity
In figure 3.1(b) the homogeneity in the center and the near grid observation volumes is similar,
the velocity r.m.s. varies only slightly within the defined regions. What varies greatly for the
two regions is the inhomogeneity in velocity r.m.s. directly adjacent to the observation areas.
The center observation volume is surrounded by homogeneous turbulence for several L on along
the central axis. Whereas the velocity r.m.s. varies greatly within one L. We observe that
the Eulerian and Lagrangian data near the grid in Figs. 3.14and 3.15 show that the structure
functions depend greatly on the origin of the fluid being swept into the observation volume.
Fluid coming from energetic regions of the tank has larger structure functions than fluid coming
from more quiescent regions. Inhomogeneity is directly responsible for the shift of the minima
in Figs. 3.14and 3.15 away from zero vertical velocity. In the center of the tank (Fig. 3.10), the
inhomogeneity is much smaller, but it could be responsible for some part of the curvature since
both fluid coming downward and fluid coming upward would be coming from more energetic
regions symmetrically.
3.7.5 Large Scale Intermittency
Inhomogeneity alone does not account for all of the large scale dependence observed. There is
also a significant contribution from large scale intermittency. Fernando and De Silva [23] show
large scale intermittency can exist in an oscillating grid flow depending on boundary conditions.
Although we use their recommended boundary conditions we observe clear signatures of large
scale intermittency. The velocity distribution in the center of the flow is bimodal, indicative of
switching between two flow states. This effect is more prominent in preliminary data we took
for grid spacings of 66 and 100 cm that it is in the present data taken at 56.2 cm.
Our measurements show a dependence of the conditional structure functions on the large scale
velocity that cannot be fully attributed to inhomogeneity, and large scale intermittency is the
most likely cause. Clear evidence for this comes from conditioning the structure functions
on the horizontal components of the large scale velocity, Σux and Σuy instead of the vertical
component, Σuz. The horizontal midplane (x and y directions) is much more homogeneous
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 42
than the vertical axis (Z direction). Yet, the structure functions conditioned on Σux or Σuy
show a large scale dependence that is only moderately smaller than the Σuz condition (72% and
85% of the dependence seen in Σuz). If the inhomogeneous direction shows similar conditional
dependence on the large scales as the homogeneous directions show, then it seems that a large
part of the conditional dependence must come from fluctuations of the large scales and not
directly from inhomogeneity. Praskovsky et al. [24] attribute large scale intermittency as a
crucial component of the large scale dependence they observe. To investigate this effect further
we devised an experiment described in chapter 4.
3.8 Kinematic Correlations
A reasonable suspicion might be that the observed dependence is a kinematic correlation, mean-
ing that particle pairs with large velocity may also have a large velocity difference simply because
the same measurements are used in both cases and they are intertwined in a mathematical sense.
Hosokawa [25] identified that Kolmogorov’s 4/5ths law requires that velocity sums and differ-
ences be correlated so that
〈u2+∆u−〉 =
ǫr
30(3.3)
where u− is half the longitudinal velocity difference and u+ is half the sum. (For comparison, we
have used ∆ur = 2u− and Σu‖ = u+.) Kholmyansky and Tsinober [10] provide an experimental
confirmation of this and in a more recent paper [26] present a list of kinematic relations. In
addition, several lines of evidence indicate that kinematic correlation does not account for the
majority of the dependence we observed.
First, note that two independent random samples with identical Gaussian distributions have a
difference that is uncorrelated with the sum, so that the conditional dependence seen in Fig.
3.10 would be flat. This remains true for velocity differences and sums from Gaussian random
fields. Both of these results can be obtained by considering the joint pdf of the two samples and
the the rotating 45◦ to the coordinate system of sums and differences. Because the samples are
interchangeable, the sum and difference axes have to be principle axes of the joint Gaussian pdf,
and the conditional variance of the difference is independent of the sum.
Of course, turbulent velocities are not joint Gaussian. However, from kinematic relations in the
literature we have not been able to derive predictions for the conditional structure functions
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 43
that we consider or for the correlation 〈(Σu)2(∆u)2〉 that would capture the main conditional
dependence we see.
An analysis of conditioning in Gaussian random fields shows that conditioning on the velocity
at either point alone produces a kinematic correlation that is avoided by our conditioning on
the velocity sum.
To make an experimental estimate of the effect of kinematic correlations, we conditioned the
velocity differences on several different quantities. For each particle pair, we calculated the longi-
tudinal and transverse components of the average velocity of the particles. We then conditioned
the longitudinal structure functions on the longitudinal and transverse pair velocities. The idea
here is that while conditioning on the longitudinal component (Σu‖) could have a kinematic cor-
relation, conditioning on the transverse component (Σu⊥) should have no kinematic correlation.
We found that conditioning on (Σu‖) had a roughly 30% larger effect on the structure functions
that conditioning on (Σu⊥). Conditioning on Σuz should have less kinematic correlation than
conditioning on Σu‖. So more than 70% of the effect remains unexplained by kinematic corre-
lation. We conclude that while kinematic correlation may possibly make a contribution to the
conditional dependence, the majority of the effect comes from the large scales.
3.9 Discussion
An immediate concern when discussing large scale effects is if the oscillating grid flow has
a Reynolds number insufficient for adequate scale separation, and it is this which leads to
contamination of the small scale statics by the large scales. Evidence points to large scale
dependence not being caused by limited Reynolds number. The comparison in Fig. 3.12 shows
that atmospheric boundary layer data [1] with very large Reynolds number (Reλ > 104) has
nearly the same dependence on the large scales as our flow. Increasing the Reynolds number in
our flow makes very little difference. Additionally, all length scales collapse to nearly the same
functional form indicating that limited separation of scales is not the primary factor. Taken
together, these lead us to the conclusion that merely high Reynolds number alone is not enough
to create small scales that are statistically independent of the large scales.
Our flow is somewhat anisotropic. The ratio of vertical to horizontal velocity standard deviations
is 1.5:1 in the center. The effects of large scale anisotropy on the small scales has been studied
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 44
extensively [14], but it is not a major factor in conditional dependence studied here. We have
analyzed our data by averaging over particle pairs with all orientations, so when the structure
functions are conditioned on a quantity with no preferred direction like the velocity magnitude
(Figure 3.17) there should be very little contribution from anisotropy. In fact, we find that
the conditional dependence on velocity magnitude is even stronger than the dependence on
the vertical velocity component. We also observe the conditional dependence remains when
conditioned on other quantities without preferred directions like Σu‖, and Σu⊥. We conclude
that anisotropy of the large scales is not a significant cause of the conditional dependence we
observe.
Sreenivasan and Dhruva [1] attribute the strong conditional dependence of the Eulerian structure
functions on the large scale velocity to shear in the atmospheric boundary layer. In making
this argument, they show an important piece of information in their figure 6 which shows
conditioned structure functions in homogeneous turbulence from both DNS and wind tunnel
grid turbulence. The conditional statistics in these homogeneous and isotropic flows show no
apparent dependence on the large scale velocity (we will investigate these systems independently
in chapter 5). However, we conclude that shear is not the fundamental property responsible in
our flow since the oscillating grid flow has a much lower shear but produces much the same
dependence on the large scale velocity. The mean velocity gradient normalized with the eddy
turnover time is 1.2 in the center of our flow and we estimate it is in the range of 5 or greater
for their atmospheric boundary layer data. There must be some other properties that exist in
the shear flow, but also are important in our flow with small shear.
Our data clearly show the role that inhomogeneity plays in the observed large scale depen-
dence. Our Eulerian and Lagrangian data near the grid in Figs. 3.10 and 3.15 show that the
structure functions depend greatly on the origin of the fluid being swept into the observation
volume. Fluid coming from energetic regions of the tank have larger structure functions than
fluid coming from more quiescent regions. Inhomogeneity is directly responsible for the shift of
the minimum in Fig. 3.14 away from zero vertical velocity. In the center of the tank (Fig. 3.10),
the inhomogeneity is much smaller, but it could be responsible for part of the curvature since
both fluid coming downward and fluid coming upward would be coming from more energetic
regions symmetrically.
However, inhomogeneity alone does not account for all of the large scale dependence observed.
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 45
There is also a significant contribution from large scale intermittency, and it is possible that
this is the dominant contribution in the center of the tank. Large scale intermittency has been
difficult to quantify. It can be defined as any temporal fluctuations in the large scales that occur
on timescales longer than the eddy turnover time, L/U . We will discuss it briefly here, and it is
the focus for the next chapter.
Fernando and DeSilva [23] show large scale intermittency can exist in an oscillating grid flow de-
pending on boundary conditions. We have observed clear signatures of large scale intermittency
in our flow. Although we use their recommended boundary conditions, the velocity distribution
in the center of the flow is slightly bimodal indicative of switching between two flow states. This
effect is more prominent in preliminary data we took for grid spacings of 66 cm and 100 cm
than it is in the data for 56.2cm presented in this thesis.
Our measurements show a dependence of the conditional structure functions on the large scale
velocity that can not be fully attributed to inhomogeneity, and large scale intermittency appears
to be the most likely cause. The clearest evidence for this comes from conditioning the structure
functions on the horizontal components of the large scale velocity, Σux and Σuy instead of
on the vertical component, Σuz. The horizontal midplane (x and y directions) is much more
homogeneous than the vertical axis (z direction). Yet, the structure functions conditioned on
Σuy or Σux show a large scale dependence that is only moderately smaller than for the Σuz
condition (85% and 72% of the dependence seen in Σuz). If the inhomogeneous direction shows
similar conditional dependence on the large scales as the homogeneous directions show, then
it seems that a large part of the conditional dependence must come from fluctuations in the
large scales, and not directly from inhomogeneity. Praskovsky et al. [24] attribute large scale
intermittency as a crucial component of the large scale dependence they observe. This is not
enough to draw any conclusions about the role intermittency plays, so the next chapter is
dedicated to further investigation.
We have largely ignored considerations of power law scaling which has been a focus of much
of the previous work on this subject. Because of the relatively low Reynolds number of our
experiment, we can not make sensitive tests of scaling. However, our data provides a plausible
picture about how the large scales should affect power law scaling. If the data in Fig. 3.10
collapses to a single curve, then the dependence of the conditional structure functions on r and
uz are separable and the large scale dependence will have no effect on the scaling exponents of
Chapter 3 - Effects of Non-Universal Large Scales on Conditional Structure Functions 46
unconditional structure functions. When this type of plot does not collapse as in Figs. 3.14 and
3.15, then the power law scaling will be affected by the large scales.
Chapter 4Effects of Large Scale Intermittency on
Conditional Structure Functions
The previous chapter gave experimental evidence that the large scale velocity affects all scales
in a turbulent flow in our oscillating grid experiment. This is an intriguing finding because it
contradicts a strict interpretation of Kolmogorov, that the small scales are completely indepen-
dent of the large scales. This result appears robust, as we will discuss in the next chapter, the
dependence is not an exception, but actually typical for all but the most controlled flows. Once
we have accepted the dependence exists a natural next step is to gain some understanding as
to what role various properties of the large scales play in this dependence. In section 3.7 we
discussed several properties of the large scales and what affects they may have on the observed
dependence. Two particular properties of the large scales that are likely candidates for influ-
encing the dependence small scales have on the large scales are inhomogeneity and large scale
intermittency. Whereas the experiments in the previous chapter focused on inhomogeneity by
measuring separate regions of the tank, this chapter will focus on experiments done to quantify
the effects of large scale intermittency.
As defined before, large scale intermittency refers to temporal fluctuations on time scales longer
than the eddy turnover time, τL = L/U , where U is the characteristic velocity U = (〈uiui〉/3)1/2
and L is the characteristic length scale L = U3/ε where ε is the energy dissipation rate per unit
47
Chapter 4 - Effects of Large Scale Intermittency on Conditional Structure Functions 48
mass. We can easily estimate a time scale associated with large scale intermittency in our flow.
For continuous driving at 3Hz the characteristic length scale L is 7.9 cm, and the characteristic
velocity U is 5.7 cm/s, making the eddy turnover time τL = 1.4 seconds. For an upper estimate
of the time scale of large scale intermittency we can approximate the length scale to be half the
width of the tank. This is an appropriate approximation because large scale intermittency can
include any fluctuations on the mean flow. As we saw in diagram 3.2 the mean flow consist of
large torii, where the size of one torus is approximately half the width of the tank; this is the
essentially the largest coherent structure allowed by the system dimensions. To calculate the
velocity scale we use (〈〈ui〉〈ui〉〉/3)1/2 where 〈ui〉 is the mean velocity which we had previously
subtracted off the velocity measurements to attain the fluctuating velocity. We arrive at a time
scale of roughly 49 seconds. In other words, we can estimate large scale intermittency to be
time scales from τL to 35τL.
Directly measuring length scales that are half the size of the tank is restricted by the maximum
size attainable of the observation volume. However, our apparatus has the advantage of being
able to measure very long durations, and can easily measure time scales over the duration of
interest. In fact, inspection of a velocity vs. time scatter plot for data taken in the center of
the tank reveals some long duration coherent structures. Due to the intermittent nature of this
phenomenon these structures are not immediately apparent, and elude simple quantification by
Fourier transform for example. However, close inspection reveals the velocity does show some
switching behavior, transitioning, for example from, predominately upward to predominately
downward velocity on the order of 30 seconds, which is in the range of interest. This can
also be seen in probability distribution function of velocity (particularly the vertical velocity
component), where a clearly bimodal distribution is not uncommon for data sets of several
hundred τL in duration.
Quantification of large scale intermittency is a challenge. The time scales may be long with
a broad range (as our example just showed), and there is no obvious way to control it. That
is to say, there is no obvious parameter than can be adjusted that can dial up or down the
large scale intermittency, and there is no accepted way to measure it. De Silva and Fernando
suggested the perimeter of the oscillating grid plays a role in the large scale intermittency in the
flow [23]. We have followed their suggestion to minimize large scale intermittency, yet as we will
see it remains to some extent. There does not seem to be a straightforward way to lessen the
Chapter 4 - Effects of Large Scale Intermittency on Conditional Structure Functions 49
large scale intermittency further. Faced with this, we instead measured the effect of large scale
intermittency by attempting to control the large scale energy injection. In essence, large scale
intermittency varies the amount of energy in the observation volume on time scales longer than
the eddy turnover time. Since we could not lessen the large scale intermittency, our strategy to
measure its effects was to augment the large scale intermittency in a typical continuous driving
case and measure the results. To accomplish this we varied the energy input into the system over
a range of time scales. For example, the energy can be injected into the system by running the
grids at a certain frequency, then periodically halting the grids suddenly, thus briefly stopping
any energy injection, and lowering the energy measured in the observation volume. This can
effectively augment the large scale intermittency which occurs with constant driving. There are
3 parameters that we varied in order to control the addition of large scale intermittency: the
amplitude of the energy injection, the duty cycle over which we varied the energy injection, and
the duration of the energy injection.
There has been some previous work on attributing the dependence of the small scales on specific
properties of the large scales. The seminal work of Praskovsky et al. was done in two high
Reynolds number wind tunnel flows, a return channel and a mixing layer. They observed
strong correlations between the large scales and the velocity structure functions. All length
scales were largely affected in the return channel; at two standard deviations greater than
the mean large scale velocity the conditioned structure function was 50% greater than that of
the unconditioned structure function. The mixing layer was only slightly less affected by the
conditioning on the large scales, and had a large difference in the spanwise and streamwise flow.
Praskovsky et al. concluded Kolmogorov theory was not contradicted, the small scales could
still be independent of the large scales. The observed correlation between the large scales and
the velocity structure functions ”... reflects the spatial and temporal variability of the energy
flux through a cascade” [24]. Although the terminology was different this equates to attributing
the observed dependence of the small scales on the large scales on the fluctuations of the energy
into or out of the large scales. These fluctuations occur on time scales longer than the eddy
turnover time, a characteristic of large scale intermittency.
In a slightly different manner turbulent shear flows also contain large scale intermittency. S. B.
Pope shows through an analysis of direct numerical simulations that a large amount of quiescent
ambient fluid is entrained in the turbulent fluid produced by round jets [2]. He gives the example
Chapter 4 - Effects of Large Scale Intermittency on Conditional Structure Functions 50
of a turbulent round jet, where at x/d = 100, where x is the position downstream, and d is the
jet diameter, 80% of the fluid in turbulent motion has been entrained between x/d = 20 and
x/d = 100. In such a manner the fluid within the jet simultaneously contains fluid elements
with very different origins, and different turbulent properties. In addition to fluctuations of the
energy flux which Praskovsky observes, entrainment of fluid with different turbulent properties
could possibly lead to fluctuations of energy longer than the eddy turnover time, a characteristic
of the large scale intermittency.
What we wish to accomplish in this chapter is to begin to quantify the effects the large scale
intermittency. Specifically, how it effects the dependence of the small scales on the large scales
which we have previously observed. We hypothesize the large scale intermittency can affect the
measured small scales in a mixture of the two ways exemplified by Praskovsky et al. and Pope:
fluctuations in the energy flux into and out of the largest scales, and the entrainment of fluid
with different turbulent properties present in the observation volume. Future studies may be
able to measure the relative weight of these to mechanisms of large scale intermittency, but in
this chapter, regardless of it’s cause, we will see large scale intermittency can have a significant
effect on small scale statistics.
4.1 Data
In this section we quantify experimentally how conditional structure functions depend on large
scale intermittency. Throughout this chapter we will use continuous driving of the oscillating
grids at 3 Hz as a base case. As we showed in figure 3.10 the structure function at all scales
(including the small scales) is affected by the large scale velocity. We attributed this dependence
in part to both inhomogeneity and large scale intermittency. To isolate the role that large scale
intermittency plays we attempt to augment it by varying the time dependence, and amplitude
of the energy injection.
In order to add large scale intermittency in a controlled way the motor driving frequency was
modulated, which created two distinct energy inputs. In an ideal experiment the motor could be
changed instantaneously from one motor frequency to another. However, in the real world ex-
periment the motor was attached to the grids and rods in the apparatus which have considerable
rotational inertia. This was most pronounced when halting the motor completely going from 3
Chapter 4 - Effects of Large Scale Intermittency on Conditional Structure Functions 51
to 0 Hz. Efforts were made to minimize this effect by replacing the flywheel with a much smaller
one, with less inertia, and studying how best to ramp down the motor current. Ultimately,
the time it took to fully start or stop the motor was less than the time of one revolution and
accounted for a small fraction of the data. For example, the single revolution before and after
the grid was halted amounted to 2% of the total data that could have possibly been affected
for the common case of alternating between 3 and 0 Hz at 30 second periods. A greater concern
in the analysis of motor phase dependence was the instrumentation used to record the motor
frequency. As described in section 2.1 a photogate was used to determine the frequency of the
motor. This added complexity to data acquisition; the motor would sometimes stop when the
beam was broken, sometimes stop when the motor was not broken, and sometimes stop when
the beam was just on the edge of being broken. In addition, the duration of a period could not
be precisely controlled, it may vary by a few tenths of a second. This caused difficulty in when
that variation in period meant the period would end with the beam broken, or not broken. Two
methods were employed to solve this problem. First, a modeling program was implemented to
determine the cycle phase at any time based on the recorded beam breaks. This worked well
for runs when the motor was not fully stopped such as frequency modulations between 3 - 2
Hz. For runs when the motor was fully stopped the non-consistent periods made this method
difficult. As a result a second method was implemented where no modeling took place, only
straightforward rules assigning long and short periods between beam breaks to different cycle
phases. Some accuracy was lost in this method, but the effect was very modest, and it primarily
affected runs involving very short periods, and only data near the motor on/off transition.
4.1.1 Driving Frequency Modulation
Varying the Energy Input
We modulated the frequency of the driving in order to artificially increase the large scale in-
termittency, so that in addition to any effects that large scale intermittent energy fluctuations
play on the continuous 3 Hz case there would be the additional effects of the modulated driving
frequency. For example, the driving frequency was modulated from 3 Hz to 0 Hz (the grids were
halted) at 15 second intervals. The grids would run for 15 seconds then be halted for 15 seconds,
and would repeat in this fashion for over an hour, until enough data was acquired. One could
Chapter 4 - Effects of Large Scale Intermittency on Conditional Structure Functions 52
rightly point out that this energy fluctuation is not intermittent, as it would be in the base case
and most experimental flows. We briefly considered randomly halting the grids, but concluded
it would make phase averaging difficult, and periodic halting is preferable because it allows us
to better control the energy input at no real loss of generality.
Of course modulating the driving frequency between 3 and 0 Hz is just one data point in
parameter space, and an extreme one at that. In addition to modulating between 3-0 Hz, we
also modulated between intermediate frequencies, namely 3-1 Hz and 3-2 Hz all with the same
cycle duration and duty cycles. These additional data points allow us to see a trend in the large
scale dependence as a function of driving modulation frequency.
Figure 4.1 shows the Eulerian second order conditional structure functions versus large scale
velocity. Note Fig. 4.1 is very similar to Fig. 3.10 where the component of the large scale
velocity and the normalization has changed. Here we condition the structure function on the
velocity component that is transverse to r denoted Σu⊥. We normalize the vertical axis by
the unconditioned structure function, and the horizontal axis by the characteristic velocity
(〈u2〉)1/2 = ((u2x + u2
y + u2z)/3)
1/2. We have made these changes so the conditional structure
functions can be more easily compared, this should not have any major affect on the appearance
or interpretation of the plots.
As before, we note the curvature of the conditional structure function and its collapse on to one
functional form. In Fig. 4.1(a) the curvature is significant, but relatively small; the structure
function does depend somewhat on the state of the large scales. The collapse of the structure
function at all length scales on to a single curve is not absolute, there is some r/η dependence,
but it is slight. Figure 4.1(b) shows the dependence of the conditional structure function in the
presence of grid frequency modulation. We see alternating the grid frequency between 3 and 2 Hz
every 15 seconds has a large effect on the curvature. The curvature has increased significantly,
and the collapse remains approximately the same. Comparing these two plots alone is enough
to derive the conclusions of this chapter. By artificially increasing the large scale intermittency
we see a striking increase in the dependence of the structure function at all length scales on
the instantaneous state of the large scales. We can continue increasing the level of large scale
intermittency by increasing the disparity in the two grid frequencies. Figure 4.1(c) shows the
conditioned structure function when the grid oscillation frequency modulates between 3 and
1 Hz every 15 seconds. We see the curvature increases further from 4.1(b), and the collapse
Chapter 4 - Effects of Large Scale Intermittency on Conditional Structure Functions 53
-2 -1 0 1 2 3
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
-3
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
Σu┴
√<u2>
<(Δur)2|Σu
┴
<(Δur)2>
>
<(Δur)2|Σu
┴
<(Δur)2>
>
<(Δur)2|Σu
┴
<(Δur)2>
>
<(Δur)2|Σu
┴
<(Δur)2>
>
a
c
b
d
Figure 4.1: Eulerian second order conditional structure function versus large scale velocity, plotted in a similar
manner to figure 3.10. The grid oscillation frequency was modulated such that the frequency would switch from
high to low at 15 second intervals. The frequencies modulated were: a) 3 Hz continuous, b) 3-2 Hz, c) 3-1 Hz, d)
3-0 Hz. Each curve represents the following separation distances r/η: + = 2.67 to 5.33, ◦ = 5.33 to 10.67, ∗ =
10.67 to 21.33, × = 21.33 to 42.67, � = 42.67 to 85.33, ⋄ = 85.33 to 170.67, △ = 170.67 to 341.33, ▽ = 341.33
to 682.67.
Chapter 4 - Effects of Large Scale Intermittency on Conditional Structure Functions 54
-3 -2 -1 0 1 2 3
0.8
1
1.2
1.4
1.6
1.8
2
2.2
<(Δur)2|Σu
┴
<(Δur)2>
>
Σu┴
√<u2>
Figure 4.2: Second order conditional structure functions vs. large scale velocity (the same data as figure 4.1).
All curves represent one separation distance, r/η = 10.67 to 21.33 measured from the different driving frequency
modulations. Each curve is measured from the following driving frequency modulations: + = 3 Hz continuous,
◦ = 3-2 Hz, ∗ = 3-1 Hz, × = 3-0 Hz
on to one functional form remains approximately the same. In figure 4.1(d) we modulate the
grid between oscillating at 3 Hz and being completely stopped at 0 Hz, again alternating every
15 seconds. The curvature increases still further, but the jump in curvature is no longer as
dramatic as comparing (a) and (b), or (b) and (c). Again, the collapse on to one functional
form remains.
Since all length scales in figure 4.1(a, b, c, and d) essentially collapse onto one functional
form the 4 plots can easily be condensed into one. Figure 4.2 extracts only the 3rd r/η bin
where r/η = 10.67 to 21.33 from figure 4.1(a, b, c, and d), and plots them on one graph for
better comparison. The minima all fall on the mean large scale velocity, but at different values
vertically. The symmetry around the mean large scale velocity is due to the inherent isotropy
in the measurement of the transverse component of the large scale velocity. Since the particles
Chapter 4 - Effects of Large Scale Intermittency on Conditional Structure Functions 55
101
1020
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
r/η
b
Figure 4.3: Curves in figure 4.1 were fit to au4 + bu2 + c. The coefficient b is a measure of the dependence
of the conditional structure function on the large scales. The coefficient b is shown here versus the separation
distance r/η. The drive frequency period was 30 seconds (15 seconds with the motor on, and 15 seconds with
the motor off). Each line represents the following driving frequency modulations: + = 3 Hz continuous, ◦ = 3-2
Hz, ∗ = 3-1 Hz, × = 3-0 Hz
are randomly distributed r is randomly oriented, and therefore the component of Σu that is
transverse to r is isotropic. The result is Σu⊥ is isotropic, and therefore all dependence on it
must be symmetric. As discussed previously the curvature in +, ◦, ∗, and × is dictated by the
amount of large scale intermittency. The relation of curvature and the value at the minima
is due to the normalization. The low curvature lines are necessarily closer to unity. This is
because the less dependent the structure function is on the large scales the closer it is to the
unconditioned structure function, thus producing a ratio closer to unity.
Another convenient way to represent the data is shown in figure 4.3. The curves in figure 4.1
were all fit to the functional form au4 + bu2 + c, and the coefficient b is shown here versus
the separation distance r/η. The coefficient b gives the second derivative at the origin, a useful
Chapter 4 - Effects of Large Scale Intermittency on Conditional Structure Functions 56
measure of the dependence of the conditional structure function on the large scales. For example
when b is zero, there is no dependence on the large scales and the curve is flat, and when b is large
there is a steep parabola and a large dependence on the large scales. The coefficient a was not
used because it is dominated by data at the extreme large scale velocities, in addition measuring
the coefficient of the second order term is in keeping with previous studies [1]. As figure 4.3
shows there is a major increase in the dependence of the conditional structure function on the
large scales as the large scale intermittency increases. The dependence increases by a factor of
more than 5 when the intermittency is increased from 3 Hz continuously driven to alternating
between 3 and 0 Hz. Even at the most modest driving frequency modulation between 3 and 2
Hz the dependence increased by more than a factor of 2 over the 3 Hz continuously driven case.
The collapse of all length scales on to a single functional form can also be evaluated by figure 4.3.
The curvature is plotted as a function of separation distance r/η, so a flat line (no separation
distance dependence) would mean all of the length scales are affected by the large scales in the
same way, and they have all collapsed on to the same functional form. Figure 4.3 shows for the
low intermittency case of 3 Hz continuously driven there is only a modest dependence on the
separation distance. This dependence increases with increasing intermittency, where the most
intermittent case of 3-0 Hz has a curvature that ranges from about .35 to .45.
Constant Energy Input
It is worthwhile to compare the driving frequency modulation (figures 4.2-4.3) to the case where
the driving is not modulated. Figure 4.4 shows conditional structure functions versus large scale
velocity for constant grid frequencies from 1 to 5 Hz. In these cases no additional intermittency
has been added by modulating the driving frequency. It is important to note how the curvatures
in the various constant driving cases are all somewhat similar, and relatively low. We can draw
from this that the large scale effects, including large scale intermittency, are all somewhat similar
regardless of driving frequency and therefore Reynolds number. We can see clearly in figure 4.5
how low the curvature is compared to figure 4.3, all of the constant driving cases have curvature
between .05 and .14, which is lower than any of the cases in which the driving frequency was
modulated and the large scale intermittency was increased. This helps to illustrate the extent
to which large scale intermittency can affect the dependence on the large scales. Changing the
Reynolds number by a factor of approximately 4 results in a smaller difference in the dependence
Chapter 4 - Effects of Large Scale Intermittency on Conditional Structure Functions 57
-3 -2 -1 0 1 2 30.8
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
Σu⊥√
〈u2〉
〈(∆ur)2|Σu⊥〉
〈(∆ur)2〉
Figure 4.4: Conditional structure function vs. large scale velocity showing one separation distance, r/η =
10.67 to 21.33 (similar to figure 4.2). The grid oscillation frequency was kept constant. Each curve represents
Figure E.1: A data processing flowchart, rounded edges represent programs, rectangles represent data files,
and shaded areas occur on the computer cluster
bright pixel the program should search for other bright pixels. Histogram threshold of 30, this
sets the minimum pixel intensity to be considered as part of a particle. I keep this at the image
compression circuit threshold. Minimum total intensity of 100, the total intensity is the sum
of all the pixel intensities in a particle. Minimum center intensity of 35, one pixel must have
this brightness intensity in order to be considered a particle. Minimum number of pixels of 3,
just as the name implies. Choosing fewer than 3 pixels will lead to the loss of particle finding
accuracy.
The algorithm for calculating particle centers from 2D images was also written in a C++ pro-
gram and uploaded to Wesleyan’s Swallowtail computer cluster where it can be used on many
processors in parallel. Although it is possible to calculate to 2D positions locally in CPVplayer
I recommend processing the cpv files into featdat.gdf files on the computer cluster. Below I will
describe the procedure for doing so, it is not as straightforward as using the CPVplayer, but is
orders of magnitude faster.
First one must get access to the cluster by contacting the administrator (currently Henk Meij).
The particle recognition programmust be set up on their home directory on the cluster. The files
required are located here lab2-C:\DocumentsandSettings\dbblum\MyDocuments\ImportantComputerPrograms\
ParticleRecognitiononCluster. Set up may not be straightforward, and may require assistance,
be needs only to be done once.
To help facilitate the processing on the computer cluster CPVplayer creates a script for use on
the cluster that will instruct the program to run on many nodes in parallel. Create a cluster
script by entering the number of frames to be calculated in a given node (typical is 5,000 frames)
and click the ’create cluster script’ button. This will produce a text document called runjobs.sh
in the local directory of the cpv file. The objective now is to put the cpv files and cluster
scripts on the cluster and run the particle recognition algorithm there. First one will note that,
unfortunately, the current version of CPVplayer does not have a user defined input path, so
the script files need to be altered manually. This means a typical local cpv path would read
C:/dbblum/2010-03-29/run3_45.cpv, but this must be altered to the location of the cpv file on
the cluster, something like /cpv2gdf_project_new_8-2007/home/dbblum/2010-03-29/run3_45.cpv
the exact location will depend on where the cpv file is located on the cluster. Renaming the
path can be done manually with the replace function in Notepad, or I wrote an even more
automated version in Matlab ext_hd1-G:\Lab1backup\Matlab_local_lib\dbblum\utilities\db_
runjobs_path_renamer.m.
Operating on the cluster will be unfamiliar for those who do not have Unix experience. I will
give a brief overview of the steps involved, but I recommend asking someone with experience
for help. To upload the cpv files to the cluster I recommend the SSH Secure File Transfer program
video6-C:\ProgramFiles\SSHCommunicationsSecurity\SSHSecureShell\SshClient.exe for both the
transfer of files and for the terminal to execute commands on the cluster. The cluster’s address
is ’swallowtail.wesleyan.edu’ (users will have to get a password from the administrator). Put the
cpv files in a folder on the home directory and put the runjobs.sh files in the same directory as
the cpv2gdf executable file. The runjobs.sh file can technically be executed, but does so in a way
which is unfair to other cluster users. I therefore strongly recommend a further step of calling on
the runjobs.sh scripts from another script called C:\DocumentsandSettings\dbblum\MyDocuments\
ImportantComputerPrograms\ParticleRecognitiononCluster\simple_edit.txt. This script enters
the processing program on to the cluster nodes properly using queuing. The only parts of the
simple_edit.txt file that need to be altered are the exact runjobs.sh file name and the intended
queue. The options for which queue to choose are ’imw’, ’elw’, ’emw’, ’ehw’,or ’ehwfd’. Choose
one based on the most open nodes as reported here http://petaltail.wesleyan.edu/cgi-bin/
bqueues_web.cgi, ’imw’ or ’elw’ are typically used. Save the simple_edit.txt file and execute
it. It can be executed by typing the command ./simple_edit.txt while the terminal is in the
appropriate directory. It should take a few moments to begin, and when it does the number
of nodes used in the previously mentioned queue status website should increase with your new
jobs.
A few additional notes, do not leave data files on the cluster. It has limited disk space for
everyone to share, bring the files down to lab computers as soon as it is convenient. The
command to check the data used in a folder (including subfolders) is ’du-k’. To check to see
what nodes are being used enter ’bjobs’. The command to stop a program that is running is
’bkill’. Also, the eccentricity calculation does not currently work, and gives meaningless results,
but this is not used in the analysis.
While the featdat.gdf files are being created on the cluster the calibration can be preformed as
described in appendix C. The PTVsetupPrep calibration program will require the 4 averaged
calibration images as input and the output will be a camParaCalib.mat file which will have 4
structures containing the calibration information for each camera.
E.1 Stereomatching
An initial round of stereomatching should now be performed. The stereomatching program
db_Stereomatch_RUN_multiframe.m is rather complex and is outlined in section 2.4. Here I will
summarize the main functions as shown in figure E.2.
db Stereomatching RUN multiframe
Input: files with 2D particle centers for each camera, parameters for each subprogram
This program should be the only one to actually be executed by the user, it should contain all
the parameters for each subprogram. The algorithm is designed to run by iterating through
the featdat.gdf files with the 2D particle centers listed in ascending frame number. Each file
should be a reasonable length, 5,000 frames is typical. This limit is meant to keep array sizes
manageable, the main array in subsequent programs will be the length of all the particles in all
the frames in a file (approximately 300 particles per frame * 5,000 frames = 3,000,000 rows).
One of the parameters is ’moreinfoflag’, this changes the number of columns and 2d particle
information in the output file. The successful execution of this program will be files which
contain accurate particle velocity data.
Output: vel3d2d.gdf, 3D particle information and 2D particle information from each camera,
with up to 28 columns. The columns are described in figure E.3
db file finder3
Input: path for files with 2D particle center data, and number of cameras.
This small function creates a list of where all the files are, and assures they are in the proper
order
db cam2d loader2
Input: list of files with 2D particle center data
This small function opens the 2D particle center files in a robust way. It is designed to account
for missing frames.
db_Stereomatch_RUN_multiframe
db_file_finder3
db_cam2d_loader2
read_gdf
db_Frame_Number3
db_Angle_Determination5
gv_pixel2unitvector_fast
calibProj_Tsai
db_Candidate_Sort9
db_Candidate_Match6
db_RayIntersection_fast
db_Position3d_Determination_dist_method3
db_Tracking_and_velocity_function
db_track_weeks
dbsp_velocity
Figure E.2: A flowchart for the functions used in stereomatching.
vel3d2d columns: 1 2 3 3D position with respect to the origin (mm): x y z 4 5 ray intersection error(mm): best match, 2nd best match 6 frame number 7 8 9 10 11 12 13 14 2D position(pix):cam1 x y cam2 x y cam3 x y cam4 x y 15 16 17 18 total particle intensity: cam1 cam2 cam3 cam4 19 20 21 22 radius of particle(pix): cam1 cam2 cam3 cam4
23 particle id number
24 25 26
velocity (mm/s): x y z 27 (optional) motor phase 28 (optional) grid phase
Figure E.3: The meaning of each column in the vel3d2d files.
db Frame Number3
Input: number of cameras, 2D data
This function finds the indices of the frames within the 2D particle data files for each camera.
Determining this once in the beginning will save time for many of the upcoming functions which
will use this information.
db Angle Determination5
Input: maxsearchdist, camParaCalib, files for 2d particle centers for each camera
This is the key program, besides this the remainder of the code is essentially bookkeeping. The
idea as described in section 2.4, is to determine what particles seen in 2D are likely candidates
for 3D matches. Doing so greatly reduces the processing time. If a researcher needs to alter
this program please do so with caution, the bookkeeping may seem opaque, and will require a
good deal of effort to understand. It should also be noted that I often refer to the angles of a
camera pointing to a 2D particle. I usually convert this angle to a distance so it can be more
easily compared.
There are several key subfunctions within this function, briefly described here.
gv_pixel2unitvector_fast: Input: 2D pixel coordinates, camera calibration parameters
Output: a 3D unit vector pointing from the camera towards the 2D pixel in the 2D image that
it sees
calibProj_Tsai: Input: camera calibration parameters, 3D particle position coordinate
Output: 2D pixel coordinates of where that 3D point would be on that camera’s image plane.
It is important to note the 3D point can be anywhere, not just in a small box in 3D space in
front of the camera, this function will project the 3D point onto what the camera would see if
it had an infinite image plane.
db Candidate Sort9
Input: maxsearchdist, maxclosematches, angles associated with each 2D particle
The function compares the angles and records rays that are probably pointing to the same
particle, so later these candidates can be submitted to the ray intersection code.
For the inputs, maxclosematches is the maximum number of possible matches for rays pointing
to a particle. If this is a large number there may be combinatoric problems. A typical number is
9 close matches. The maxsearchdist parameter is the distance in pixels that the rays vary when
pointing to the same pixel. If this parameter is too high computation slows down, if it is too
low good matches will be list. An ideal system would have zero error, however a typical number
is .09 pixels, but this can be raised for the initial run, and lowered when dynamic calibration
has been established.
db Candidate Match6
Input: maxintersectionerror, mincamonpart
This function takes all of the combinations of candidate rays generated by db_Candidate_Sort9
and finds their intersection. An important subfunction is db_RayIntersection_fast, which is
the algorithm for determining the 3D point in space where the rays come closest to intersect-
ing.
For the inputs, maxintersectionerror is an important parameter that limits the distance the rays
can intersect and still be considered a single particle. The intersection error is recorded with the
3D particle information, I recommend looking at this statistic carefully and becoming familiar
with it (see how it varies in space, calibration, etc). If it is too small good matches will be lost,
if it is too large incorrect matches will be processed as particles. A typical number is 0.2 mm for
the initial run, and lowered to 0.06 mm after dynamic calibration has been established. However,
the mean for intersection error after dynamic calibration can be as low as 0.025 mm.
The parameter mincamonpart is the minimum number of cameras that need to see a particle in
order for it to be considered. Of course this is between 2 and 4, typically 2. It may be helpful
to set this at 4 for the initial run so that all cameras are properly dynamically calibrated and
the particles found are surely good matches.
db Position3d Determination dist method3
Input: all 3d points found by all camera sets, mingrouponpart, maxgrouperror
The current algorithm in db_Stereomatching_RUN_multiframe works by having the angles of a
main camera compared to the angles determined for each other camera (where ’angles’ refers to
the distance between a reference ray and the ray pointing from the camera to the 2D particle
on the image). The algorithm is somewhat redundant in that it cycles through each camera as
the main camera, but this is useful when one camera does not see a particle. The result is that
3D particles are found for different camera sets, up to 4 camera sets. As an example one camera
set would be cameras 1,2,3,4 where camera 1 is the main camera. A second camera set would
be 2,3,4,1 where 2 is the main camera. Both of these camera sets could find a 3D particle in a
nearly identical place in space. This function was made so that this 3D point is not recorded
twice (once for each camera set). Instead it goes through and finds particles that are nearly
identical and picks one randomly to be recorded.
For the inputs, mingrouponpart is the minimum number of camera sets that a 3D particle needs
to have in order to be considered valid. A typical number is 2 camera sets. The maxgrouperror,
is maximum distance between particles found from different camera sets. It is expected that this
distance is quite small, so a typical number is 0.001 mm, much smaller than the maxintersec-
tionerror. Be sure not to make this number too small, or valid particles will be lost. If it is too
large particles that just happen to be near each other will be recorded as one particle.
E.2 After Stereomatching
Stereomatching is now complete. A list of 3D particles is produced and the processing is nearly
complete. The stereomatching program also includes tracking and velocity functions within
the function db_Tracking_and_velocity_function.m. The tracking program (db_track_weeks.m)
identifies which frames a single particle appears in. Tracking a single particle through frames is
essential for measuring its velocity. As for the inputs for tracking, mem is the number of frames
into the future the program should search for the current particle, a typical number is 4 frames.
The maxdisp parameter is the maximum displacement of the particle between frames, and is the
most important parameter. A typical number is 0.6 mm, but should be judged based on each
experiment because it limits the maximum velocity that can be recorded. The dim is simply
the number of dimensions the particle exists in, of course this is set to 3. The parameter good
is the minimum number of frames a particle needs to be tracked for in order to be recorded, a
typical number is 5 frames. The velocity program dbsp_velocity.m determines the velocity of
each particle. A simple derivative gives unnecessarily noisy results. Instead the algorithm fits
a polynomial to the positions as described in 2.5. The fitorder is the order of the polynomial
that should be used to fit the position data, a typical number is order 2. The window refers to
the points considered within a trajectory for the velocity measurement, the halfwindow is half
of this, and a typical number is 2 frames.
Now that the velocities have been found I recommend running the dynamic calibration as de-
scribed in section C.3. The results of dynamic calibration are more accurate calibration files
which in turn should be used to recreate the 3D particle data more accurately. Usually one
iteration is enough, but I have found it tempting to create more velocities with the dynamic
calibration and then repeat the process.
In addition to running the sterematching code on local machines it can be run on the computer
cluster. This can be very useful if there is a lot of data needs to be processed. Eight nodes
are available for use with a Matlab program, although, I recommend not using every node as
a courtesy to other users. In order to run the stereomatching program on the cluster it must
be executed as a function (see db_Stereomatch_function19.m and related subfunctions as an
example). After the proper files have been uploaded to the computer cluster they can be run by
executing the file MyJob.m. This is done by opening the terminal and accessing the folder with all
of the sterematching function and launching the Matlab program on the cluster (note only one
user can do this at Wesleyan at one time). The command to launch Matlab is simply ’matlab’
once opened simply type the name of the MyJob script to be run. Be sure to type ’exit’ after it
has run to free up the one instance of Matlab that can be running at a time.
This concludes the data processing. It will certainly take some time to become familiar with the
process and develop one’s own best practices. From here I can point future researchers to a few
additional programs I have developed. The program ext_hd1-G:\Lab1backup\local_lib\dbblum\
DataInspection\VelocityField\db_smoothed_velocity_vector_field_creation_ML.pro and ext_hd1-G: