Thesis

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

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, Ármann 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
Steinberg.
iii
Contents
1 Introduction 1
1.1 Basic Turbulence Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Scaling Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Apparatus and Experiment 10

2.1 The turbulence tank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 3D Particle Finding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Particle Tracking and Velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Effects of Non-Universal Large Scales on Conditional Structure Functions 18

3.1 Characterizing the Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Structure Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Energy Dissipation Rate Measurement . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 Phase Dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.5 Dependence on Large Scale Velocity . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5.1 Eulerian structure functions conditioned on the large scale velocity: Cen-
ter region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
iv
CONTENTS CONTENTS
3.5.2 Lagrangian Structure Functions Conditioned on the Large Scale Velocity:

Center Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.5.3 Eulerian Structure Functions Conditioned on the Large Scale Velocity:
Near Grid Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.5.4 Third order Eulerian Structure Functions Conditioned on the Large Scale
Velocity: Center Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.6 A Powerful Method for Plotting Conditional Structure Functions . . . . . . . . . 31
Center Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Higher Reynolds Number . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Center Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Near Grid Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Near Grid Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.6.6 Third Order Eulerian Structure Functions Conditioned on the Large Scale
Velocity: Center Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.6.7 Second Order Eulerian Structure Functions Conditioned on the Velocity
Magnitude: Center Region . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.7 Properties of the Large Scales and Their Effects . . . . . . . . . . . . . . . . . . . 39
3.7.1 Reynolds Number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.7.2 Anisotropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.7.3 Shear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.7.4 Inhomogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.7.5 Large Scale Intermittency . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.8 Kinematic Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4 Effects of Large Scale Intermittency on Conditional Structure Functions 47

4.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.1.1 Driving Frequency Modulation . . . . . . . . . . . . . . . . . . . . . . . . 51
v
CONTENTS CONTENTS
4.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5 Signatures of Non-Universal Large Scales in Conditional Structure Functions

Compared in Various Turbulent Flows 73
5.1 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6 Conclusions 87
A Measurement Error Considerations 90
B Tank Preparation 93
C Calibration 96
C.1 Acquiring Calibration Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
C.2 Running the Calibration Programs . . . . . . . . . . . . . . . . . . . . . . . . . . 97
C.3 Dynamic Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
D Data Acquisition 101

D.1 Filling the tank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
D.2 Using the laser . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
D.3 Using the image compression circuits . . . . . . . . . . . . . . . . . . . . . . . . . 103
D.4 Using external trigger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
D.5 Running the motor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
D.6 Water cooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
D.7 Adding particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
D.8 Recording the grid position (optional) . . . . . . . . . . . . . . . . . . . . . . . . 108
D.9 Recording data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
E Data Processing 110

E.1 Stereomatching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
E.2 After Stereomatching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
vi
List of Figures
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. 11
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 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
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. . . . . . . . . . . . . . . . . . . . . 19
3.2 Scale diagram of 56 cm × 100 cm area between grids showing the mean circula-
tion 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. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
vii
LIST OF FIGURES LIST OF FIGURES
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. . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.4 Eulerian third order longitudinal velocity structure function. The inset shows
this data compensated by Eq.3.1 for p=3. . . . . . . . . . . . . . . . . . . . . . . 23
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 grid at lowest possible amplitude. φ:
+ = 0 - 2π/5, ∗ = 2π/5 - 4π/5, ⋄ = 4π/5 - 6π/5, △ = 6π/5 - 8π/5 , = 8π/5 -
2π. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
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 ve-
p
locities, Σuz / hu2z i: + = 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. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.7 Second order Lagrangian velocity structure function conditioned on instantaneous
velocity (vertical component) in the center of the tank. (a) Uncompensated struc-
ture function. (b) Individually compensated to have the peak values match. Sym-
p
bols represent the following dimensionless vertical velocities, Σuz / hu2z i: + =
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. 28
3.8 Second order velocity structure function conditioned on particle pair vertical ve-
locity (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
p
Σuz / hu2z i: + = 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. . . . . . . . . . . . . . . 31
viii
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
p
the following vertical velocities Σuz / hu2z i: ∗ = 2.5 to 0.84, ⋄ = 0.84 to to -0.84,
△ = -0.84 to -2.5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
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. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Data taken in the center region at higher grid frequency, 5Hz, resulting in higher
Taylor Reynolds number 380. Symbols represent the following separation dis-
tances r/η: + = 0 to 50, ∗ = 50 to 100, ⋄ = 100 to 150, △ = 150 to 200, =
310 to 420, × = 420 to 520. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
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. . . . . . . . . . . . . . . . . . . . . . 35
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. . . . . . . . . . . . . . . . . . . . . . . . . . 35
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 min-
ima 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. . . . . . . . . . . . . . . . . . . . . . . . 37
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. . . . . . . . . . . . . . . . . . . . . . 38
ix
3.16 Eulerian third order conditional structure function versus large scale vertical ve-
locity in the center region. Symbols represent the following non-dimensional sep-
aration distances r/η: + = 0 to 40, ∗ = 40 to 70, ⋄ = 70 to 110, △ = 110 to 140,
= 220 to 300, × = 300 to 370. . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.17 Eulerian second order conditional structure function versus magnitude of the ve-
locity 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. . . . . . . . . . . . . . . . . . . . . . . . . 39
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. 53
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 continu-
ous, ◦ = 3-2 Hz, ∗ = 3-1 Hz, × = 3-0 Hz . . . . . . . . . . . . . . . . . . . . . . . 54
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 . . . . . . . . . . . . . 55
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 fre-
quency was kept constant. Each curve represents the following grid oscillation
frequencies: + = 1 Hz, ◦ = 2 Hz, ∗ = 3 Hz, × = 4, Hz = 5 Hz . . . . . . . . . 57
x
4.5 The dependence of the conditional structure function on the large scales is quan-
tified. The coefficient b (where the polynomial au4 + bu2 + c was fit to the
conditional structure function) is shown here as a function of r/η. This is plotted
in the same way as figure 4.3. Each curve represents the fittings for the following
grid oscillation frequencies: + = 1 Hz, ◦ = 2 Hz, ∗ = 3 Hz, × = 4 Hz, = 5 Hz. 58
4.6 The number of samples shown as a function of large scale velocity for frequency
modulation 3-0 Hz, duty cycle 50%, and period 384 seconds. The curves represent
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. . . . . . . . . . . . . . . . . . . . . . . . 59
4.7 Second order conditional structure functions vs. large scale velocity. All curves
represent one separation distance, r/η = 10.67 to 21.33 measured from the dif-
ferent driving frequency modulations. Each curve is measured from the following
driving frequency modulations: + = 4 Hz continuous, ◦ = 4-2.67 Hz, ∗ = 4-1.33
Hz, × = 4-0 Hz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.8 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:
+ = 4 Hz continuous, ◦ = 4-2.67 Hz, ∗ = 4-1.33 Hz, × = 4-0 Hz . . . . . . . . . 62
4.9 Conditional structure function vs. large scale velocity showing one separation dis-
tance, r/η = 10.67 to 21.33 (similar to figure 4.2). The grid oscillation frequency
was modulated between 3 and 0 Hz, where the motor was in the high state for
different percentages of the period, also called duty cycles. Each curve represents
the following duty cycles: + = 100%, ◦ = 75% Hz, ∗ = 50%, × = 25% . . . . . . 64
4.10 The dependence of the conditional structure function on the large scales is quanti-
fied. The coefficient b (where the polynomial au4 +bu2 +c was fit to the conditional
structure function) is shown here as a function of r/η. This is plotted in the same
way as figure 4.3 Each curve represents the fittings for the following grid duty
cycles: + = 100%, ◦ = 75% Hz, ∗ = 50%, × = 25% . . . . . . . . . . . . . . . . . 65
xi
4.11 The phase average energy shown as a function of cycle phase. The motor was
halted at 2pi and turned on at pi. Each curve represents the following cycle
periods r/η + = 1.5/3”, ◦ = 3/6”, ∗ = 6/12”, × = 12/24” = 24/48” ⋄ =
192/384” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.12 Conditional structure function vs. large scale velocity showing one separation dis-
tance, r/η = 10.67 to 21.33 (similar to figure 4.2). The grid oscillation frequency
was modulated between 3 and 0 Hz and the period was varied. Each curve rep-
resents the following time the motor was on/total period (seconds): + = 1.5/3”,
◦ = 3/6”, ∗ = 6/12”, × = 12/24”, = 24/48”, ⋄ = 192/384” Note all have a
duty cycle of 50%. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.13 The dependence of the conditional structure function on the large scales is quanti-
fied. The coefficient b (where the polynomial au4 +bu2 +c was fit to the conditional
structure function) is shown here as a function of r/η. This is plotted in the same
way as figure 4.3 Each curve represents the fittings for the following cycle periods:
+ = 1.5/3”, ◦ = 3/6”, ∗ = 6/12”, × = 12/24” = 24/48” ⋄ = 192/384” Note
all have a duty cycle of 50% . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.1 Eulerian second order longitudinal velocity structure functions compensated by

(εr)2/3 . a) Direct Numerical Simulation b)Passive Grid Wind Tunnel c) Active
Grid (Synchronous Driving) d) Active Grid (Random Driving) e) Counter Rotat-
ing Disks f) Oscillating Grids g) Lagrangian Explorer Module (Constant Driving)
h) Lagrangian Exploration Module (Random Driving). . . . . . . . . . . . . . . . 77
5.2 (Color Online) The Eulerian second order longitudinal structure functions are
conditioned on the transverse velocity sum (Σu⊥ ), and plotted versus Σu⊥ . Sym-
bols represents the following separation distances: r/η:+ = 4, ◦ = 8, ∗ = 16,
× = 32, = 64, ♦ = 128, △ = 256, ▽ = 512 ⊲ = 1024. a) Direct Numerical
Simulation b)Passive Grid Wind Tunnel c) Active Grid (Synchronous Driving)
d) Active Grid (Random Driving) e) Counter Rotating Disks f) Oscillating Grids
g) Lagrangian Explorer Module (Constant Driving) h) Lagrangian Exploration
Module (Random Driving). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
xii
5.3 (Color Online) Structure functions conditioned on the vertical component of the
velocity sum, Σuz . Symbols represents the following separation distances r/η:+
= 4, ◦ = 8, ∗ = 16, × = 32, = 64, ♦ = 128, △ = 256, ▽ = 512 ⊲ = 1024. a)
Counter Rotating Disks b) Oscillating Grids . . . . . . . . . . . . . . . . . . . . . 79
5.4 (Color Online) The Eulerian second order longitudinal structure functions are
conditioned on the longitudinal velocity sum (Σuk ), and plotted versus Σuk .
Symbols represents the following separation distances r/η:+ = 4, ◦ = 8, ∗ =
16, × = 32, = 64, ♦ = 128, △ = 256, ▽ = 512 ⊲ = 1024. a) Direct Numerical
Simulation b)Passive Grid Wind Tunnel c) Active Grid (Synchronous Driving)
d) Active Grid (Random Driving) e) Counter Rotating Disks f) Oscillating Grids
g) Lagrangian Explorer Module (Constant Driving) h) Lagrangian Exploration
Module (Random Driving). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.5 Conditional structure functions for Gaussian random fields. The Eulerian second
order longitudinal structure function is conditioned on the longitudinal velocity
sum. Symbols represents the following separation distances r/η:+ = 4, ◦ = 8, ∗
= 16, × = 32, = 64, ♦ = 128, △ = 256, ▽ = 512 ⊲ = 1024. . . . . . . . . . . 81
A.1 Eulerian second order conditional structure function versus large scale velocity,
plotted in a similar manner to figure to 4.1a, with error bars representing the
statistical error. The following separation distances are shown in separate plots
for clarity: r/η: ◦ = 5.33 to 10.67, × = 21.33 to 42.67, ⋄ = 85.33 to 170.67, ▽ =
341.33 to 682.67. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
C.1 An example calibration image that is an average of 10 calibration images taken

from camera 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
C.2 An example of a calibration image that has been edited so all that remains is the
dots which are viewed by each camera . . . . . . . . . . . . . . . . . . . . . . . . 99
D.1 A flowchart for the path of the camera trigger. . . . . . . . . . . . . . . . . . . . 105
E.1 A data processing flowchart, rounded edges represent programs, rectangles rep-
resent data files, and shaded areas occur on the computer cluster . . . . . . . . . 111
E.2 A flowchart for the functions used in stereomatching. . . . . . . . . . . . . . . . . 115
E.3 The meaning of each column in the vel3d2d files. . . . . . . . . . . . . . . . . . 116
xiii
Chapter 1
Introduction
In a reductionist sense most of fluid dynamics was solved with the Navier-Stokes equations in the
early 1800’s. The fundamental equations that govern fluid dynamics are essentially conservation
laws for a continuous fluid, deterministic equations that can be written down rather simply.
Once the equations are solved with the proper boundary conditions the entire velocity field of a
system can be found. However, once one tries to describe a real world flow, it quickly becomes
apparent that the Navier-Stokes equations are analytically solvable for only the most basic,
hypothetical systems. Real world flows generally have complex boundary conditions, and move
quickly enough to have the complex non-linear terms become important, and the velocity field
quickly becomes unpredictable, and turbulent.
In a way this is what makes this field exciting: the fundamental equations have been known
for centuries, but attempts to predict individual trajectories are almost by definition futile in
most cases. Instead, we must invent different approaches to even hope to shed some light on
to the field. It quickly became clear that since predicting individual trajectories can not be
done, a statistical description of turbulence is needed to progress forward. There has been some
wonderful successes in searching for a statistical description of turbulence. Certain statistical
quantities have been shown to be very robust, appearing in a wide range of flows, which have
very real engineering benefits. The field has grown very wide encompassing many perspectives
and affecting other branches of research and engineering for instance: turbulent mixing, insect
flight, polymer drag reduction, to name just a few. The list of fields where turbulence research
1
Chapter 1 - Introduction 2
can help guide progress will certainly grow. It has become a model problem for digging into an
unpredictable, highly nonlinear system with a wide range of length and time scales- it is the
model problem for problems that are difficult to model.
1.1 Basic Turbulence Theory
After we have accepted the limitations of attempting to directly solve the Navier-Stokes equa-
tions we can begin to explore other insights in to turbulence. I will now give a basic background
to the turbulence theory needed to understand this thesis. A student is hard pressed to find
a better text than Turbulent Flows by S.B. Pope [2], which is the major reference for this
section.
In 1922 L.F. Richardson introduced the idea of an energy cascade in turbulence. The idea
begins with kinetic energy entering a fluid at a large scale through some stirring mechanism.
This stirring energy will create an eddy at a large scale. An eddy can be loosely defined as a
turbulent motion localized within a certain region with a definite size, that is at least moderately
coherent within that region. As an example, the stirring mechanisms in our experiment are the
two oscillating metal grids. In the notion of the energy cascade they are injecting energy at a
length scale of approximately the mesh size of the grids (the spacing between the bars of the grids
which is 8 cm). This begins the energy cascade with a large scale eddy that is approximately
8 cm. Large scale eddies that are created by the stirring mechanism are unstable, and break
up. Energy is thus transferred from the large eddies to smaller eddies. The smaller eddies are
themselves unstable and chaotically break up in to even smaller eddies. This process continues
until, at the smallest length scales the energy is dissipated into heat by the viscous action of the
fluid. Richardson even wrote a little poem about it [3]:
Big whirls have little whirls;

That feed on their velocity;
And little whirls have lesser whirls,
And so on to viscosity
(in the molecular sense).
Hopefully, the energy cascade makes some intuitive sense. We can progress further by seeing
how such a notion can be quantitative and what predictions can be made.
The Reynolds number is a key number in almost all of fluid dynamics and represents the ratio of
the inertia to the viscous forces, Re = U L/ν where U is a characteristic velocity (usually the rms
velocity), L is a characteristic length scale (usually can be approximated by the energy input
length scale, in our case 8 cm), and ν is the kinematic viscosity of the fluid (the resistance of
the fluid to motion scaled by the fluid density), note that the Reynolds number is dimensionless.
The Reynolds number is important because it can immediately describe the qualitative motion
of a fluid. If the Reynolds number is very small the viscous forces dominate, the fluid velocity is
relatively low, and the flow is laminar. If the Reynolds number is very large the inertial forces are
dominant, the viscosity plays a relatively small role in the flow, and the flow is unpredictable
and turbulent. To give a feel for its importance, the Navier-Stokes equations have only one
free parameter, and that is the Reynolds number. We will not go into great depth about the
Reynolds number here, but if it is helpful it can be thought of as a ’turbulence intensity’. As an
example when the oscillating grids in our experiment are moving at 5 Hz the Reynolds number
is 9,400, when they are moving at 3 Hz the Reynolds number is 5,400, when they are moving at
1 Hz the Reynolds number is 1,300, and as they move slower the Reynolds number will continue
to decrease until the flow is completely laminar.
One can assign a Reynolds number to the eddies in a turbulent flow. We can write Re0 = u0 l0 /ν,
where u0 is the characteristic velocity of an eddy, l0 is the length scale of an eddy, and these can
be related in the simple definition τ0 ≡ l0 /u0 where τ0 is the characteristic timescale associated
with the eddy. So at the largest eddies Re0 is approximately the same as Re. As we examine
smaller and smaller eddies l0 gets smaller as does u0 (more on why they both decrease later),
and ν grows in importance until it is completely effective at dissipating the kinetic energy.
This is important because one can hypothesize that at large enough Reynolds number the large
length scales would be unaffected by the viscosity, all of the energy of one large length scale
can be transferred directly into the somewhat smaller ones with negligible loss due to viscosity.
These eddies have energy of order u20 and a timescale τ0 . So the rate of energy transfer is just
u20 /τ0 = u30 /l0 . Since essentially no energy is lost in the energy cascade until l0 is small, once the
energy finally reaches the energy dissipation scales it has the same amount of energy as when
the energy started to be transferred. This means that at sufficient Reynolds number the energy
transfer rate is equal to the energy dissipation rate per unit mass, independent of ν, ε = U 3 /L
where ε is the energy dissipation rate. One can already see how knowing the rate at which
energy is dissipated in a flow by only knowing U and L could be very useful, especially from an
engineering perspective.
At this point the idea of the energy cascade may seem plausible, and possibly intriguing, but
still has many questions unanswered. We still need to answer why when l0 decreases u0 also
decreases, and try to better define ’small’ scales.
In 1941 Kolmogorov posited 3 hypotheses which further the energy cascade idea, and have been
very useful in turbulence theory. All assume a sufficiently large Reynolds number and are stated
here less formally than in Kolmogorov’s original work.
· Local Isotropy - Small scale turbulent motions are statistically isotropic.
Directional information of the large scales is lost as the energy is transferred down the cascade.
In fact, all information about the the large eddies, geometry, direction, boundary conditions,
etc., is lost as the eddies are chaotically broken apart into smaller and smaller scales. This is
a crucial point that this thesis hopes to address, to what extent is all information about the
large scales lost? If all large scale information is lost the small scale motions have some amount
of universality, they should be similar in every turbulent flow with sufficient Reynolds number.
This leads to the next hypothesis:
·The First Similarity Hypothesis - Small scale motions have a universal form that is uniquely
determined by ν and ε.
Since all large scale information is lost, all small scales -independent of flow and therefore
universally- depend on two parameters, ν and ε. One result of this hypothesis is that we can
now define unique time, length, and velocity scales, which are called Kolmogorov scales:
τη ≡ (ν/ε)1/2
η ≡ (ν 3 /ε)1/4 (1.1)
uη ≡ (νε)1/4
These are derived from the two parameters which characterize the smallest, dissipative eddies.
Derivation is a simple matter of dimensional analysis, arranging the two parameters so that
they have the proper units for time, length and velocity respectively. This creates a definition
for the smallest scales. A consistency check shows the Reynolds number for the Kolmogorov
scales shows ηuη /ν = 1, which agrees with the earlier notion that smaller eddies have smaller
Reynolds numbers.
To arrive at the next hypothesis, one can further suppose that given a sufficiently high Reynolds
number there is a special range of scales, which are called the inertial range, that are larger
than the dissipative scales (where ν and ε play a role). The inertial range includes length scales
that are larger than the dissipative range, and large enough so that ν will no longer play a
role, yet smaller than the largest length scales where the boundary conditions and other non-
universal attributes play a role. They exist (again given sufficiently high Reynolds number) in
a intermediate range. This leads to the third hypothesis:
·Kolmogorov’s Second Similarity Hypothesis - There exists a certain range of length scales such
that there is a universal form which is uniquely determined by only ε.
This means that if a flow has sufficient Reynolds number some range of length scales will
depend only on ε, and will do so in a way that is universal, independent of the unique large
scale boundary conditions, and the viscosity which dominates the small scales.
An important note about language, it is commonplace to categorize all scales in a turbulent

system in to “large” and “small.” Of course there is energy in the turbulent flow at length
scales which range continuously from near the dimensions of the system to below the Kolmogorov
lengths. In this work, the large scales refer to any scale which is size L or larger. Conversely, the
small scales refer to any scale which is smaller than L, either in the inertial or dissipative ranges,
where the large scale information is predicted to have been lost in the energy cascade.
Now, using this Second Similarity Hypothesis and some more dimensional arguments we can find
the functional form in the inertial range that only depends on ε and is predicted to be universal.
Testing the universality of these predictions is the basic motivation for this thesis.
1.1.1 Scaling Laws
Let’s begin by considering the easily measurable quantity, the velocity difference, ∆ur = [u(x) - u(x+r)]L
which is the instantaneous velocity difference between two points where x is the position, r is
the separation distance, and the subscript L denotes the longitudinal component of the veloc-
ity difference vector (the component of the velocity difference vector projected on the the ray
connecting the two points). The measurement of ∆ur is useful because as a velocity difference
between two points it can easily describe the dependence on length scale, r, while not being as
susceptible to the effects of non-universal large scales that would be inherent in a single point
statistic. We will commonly raise the velocity difference to a power and take the ensemble
average, this is called the structure function Dp = h(∆ur )p i.
We see that ∆ur depends on r, and if we assume homogeneity for a moment, the Second
Similarity Hypothesis implies that in the inertial subrange ∆ur = F (ε, r), the functional form
only depends on ε and r. Now some simple dimensional analysis, r has units of length, ε has
units of energy per unit mass per unit time, and ∆ur has units of velocity,
r → [m] ,
2
m
ε→ ,
s3
h m ip
h(∆ur )p i → .
s
We can use these to find the functional form of F (ε, r). We set the structure function propor-
tional to some combination of ε and r
h(∆ur )p i ∝ εα rα
We now use dimensional analysis to determine this relation

h m ip m2 α
α
∝ [m] .
s s3
This directly leads to α = p/3. So now, by using dimensional analysis we have
h(∆ur )p i = Cp (εr)p/3 (1.2)
where Cp is a universal constant which is generally found experimentally. This is a testable

result, and has been the beginning for much of the experimental turbulence research since its
inception in 1941. Experiments have examined all aspects of the relation, the universality of
Cp , the robustness of the p/3 exponent, and as we will test the degree to which the relation
is universal. An additional very nice check is that the relation for p = 3 can be been derived
directly from the Navier-Stokes equations. This gives great confidence in the result, as well
as gives C3 = −4/5. Since this can be shown analytically, the p = 3 case is the standard for
evaluating an experiment.
1.1.2 Motivation
The theories of Kolmogorov I have just sketched have been very successful. Granted, there have
been significant modifications to some of the specific predications, but I classify the theory as a
success because of its utility as a framework from which progress has been made. Not only has
the inertial range been shown to exist in countless systems, there are statistics in the inertial
range which adhere very closely to Kolmogorov’s predictions. It has been carefully shown that
there are statistics that are nearly identical in different flows, the non-universal properties of the
large scales which are unique in each flow do not affect them [4–6]. This work is by no means an
attempt to dispute these findings. Instead we hope to shed some light on small scale statistics
which are affected by the non-universal properties of the large scales within a flow, and perhaps
what differentiates them from the universal small scale statistics. Other work has shown that
the coefficients of scaling laws [7, 8], or the scalar derivative skewness [9] are not universal. We
will be primarily focused on structure functions with p = 2 or 3, as defined equation 1.2.
A traditional approach to determine if a small scale statistic is universal is to first categorize

flows such as: free shear flows (jets, mixing layers, etc.), wall bounded shear flows (boundary
layers, channel flows, etc.), or isotropic turbulence (wind tunnel grid turbulence, or numerical
simulations in a box with periodic boundary conditions, etc.). Then a careful empirical compar-
ison of small scale statistics between different flows can show which statistics are independent
of the large scales.
However, we will not be comparing different flows to determine if a statistic is universal. Instead,
we will use a more direct method, the small scale measurements in a flow will be conditioned1
on a measurement of the state of the large scales. Any dependence of a small scale statistic
on the non-universal large scales within the flow will be obvious, and can even be quantified.
Three previous studies have done just that, and shown strong dependence of the small scales of
Eulerian structure functions on the instantaneous velocity in the flow, which is dominated by
the large scales.
Praskovsky et al. conducted experiments in two high Reynolds number wind tunnel flows, a
1 Conditioned statistics will be used heavily throughout this work. The concept is simple, when a variable is
measured the condition of a different variable is also measured. The conditional statistic can then show how the
two measurements are related. For example, measure the distribution of marble sizes, and condition it on marble
color. You could see that the biggest marbles are red, for instance.
return channel and a mixing layer. The Eulerian structure functions were conditioned on the
large scale velocity and showed a strong correlation. All length scales were significantly affected,
including the small length scales in both flows. In their discussion Praskovsky et al. discounted
the possibility of having a Reynolds number that was too small, their apparatus achieved a
very large Reynolds number. They also discounted that Kolmogorov’s theory was incorrect.
Instead, they concluded their observations were consistent with Kolmogorov’s theory when it is
correctly applied to a flow with a fluctuating energy injection at large scales. Sreenivasan and
Dhruva [1] measured Eulerian velocity structure functions from atmospheric boundary layer
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-
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 2
Apparatus 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.
1 Note 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
1m
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
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.
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
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.
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.
Camera 2
[xc,yc] Coordinates of camera 2
if it could be seen by camera 1
( yx -y-x )
θ1=tan-1
c
c
1
y -y
θ =tan ( x -x )
c 2
-1
2
c 2
[x2,y2] θ2
θ1
+
[x1,y1] Camera coordinates of
particle seen in camera 1
Camera 1’s view
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
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].
2 http://www.physics.emory.edu/ weeks/idl/
Chapter 3
Effects 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.
1 Note 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
6
a
Mean Vertical Velocity (cm/s)

4
2 N
e
0
a
r C
-2
e
G n
-4
r t
i e
-6
d r
b
Vertical Velocity Variance (cm2/s2)
300 R R
e e
250 g g
i i
200 o o
n n
150
100
50
0
-20 -15 -10 -5 0
Distance From Center (cm)
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
Top Grid
28.0
Distance From Center (cm)
16.8
5.6
C
-5.6
NG
. .
-16.8
-28.0
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 = (hui ui i/3)1/2 and a characteristic length scale by

L = U 3 /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 = (15U L/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
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.
10.0
< ∆ur > (cm/s)2

2.0
1.5
2
< ∆ur >
2
1.0 (εr)2/3
1.0
0.5
10 100
r/η
0.1
10 r/η 100
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 = h(∆ur )p i where p represents the order of the structure function and the brackets represent
the ensemble average. In the inertial range, Kolmogorov (1941) gives
h(∆ur )p i = Cp(E) (εr)p/3 (3.1)
(E)
where Cp 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
10.00
< ∆ur > (cm/s)3

1.0
0.8
1.00
0.6
3
3
< ∆ur >
(εr)
0.4
0.10
0.2
0.0
10 r/η 100
0.01
10 100
r/η
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
h(∆uτ )p i = Cp(L) (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 = 2νhsij sij i
with
1 ∂ui ∂uj
sij = + .
2 ∂xj ∂xi
(E)
Instead we utilize Kolmogorov’s 4/5 law: Eq. 3.1 with p = 3 where the coefficient C3 = −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
(E)
C2 = 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, h(u − hui)2 |φi, 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 hu|φi, 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, h∆u2r |φi.
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
a
2.0
1.8
2
< ∆ur | φ >
2/3 1.6
(ε r)
3
1.4
1.2
1.0
10 100
r/η3
b
2.0
1.8
2
< ∆ur | φ >
2/3 1.6
(ε r)
3
1.4
1.2
1.0
10 100
r/η3
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
grid at lowest possible amplitude. φ: + = 0 - 2π/5, ∗ = 2π/5 - 4π/5, ⋄ = 4π/5 - 6π/5, △ = 6π/5 - 8π/5 , =
8π/5 - 2π.
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
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 (Σuk ) 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 εuz to 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
100
a
< ∆ur | Σuz > (cm/s)2

2
10
2.2 b
2.0
1.8
< ∆ur2 | Σuz >
(εΣu r)2/3
1.6
z
1.4
1.2
1.0
0.8
10 100
r/η
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,
p
Σuz / hu2z i: + = 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.
2
a
10
< ∆uτ | Σuz > (cm/s)2 10

1
0
10
2
-1
10
5 b
< ∆uτ2 | Σuz >
4
(εΣu τ)
z
10-1 100 τ/τη 101 102
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
p
have the peak values match. Symbols represent the following dimensionless vertical velocities, Σuz / hu2z i: + =
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-
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
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 .
3.6 A Powerful Method for Plotting Conditional Struc-

ture Functions

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
a
< ∆ur | Σuz > (cm/s)2
100
2
10
2.4 b
2.2
< ∆ur2 | Σuz >
(εΣu r)2/3
2.0
z
1.8
1.6
1.4
1.2
100
r/η
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
p
to lack of statistical convergence. Symbols represent the following particle pair vertical velocities Σuz / hu2z i:
+ = 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.
0.8
< ∆ur3 | Σuz >

0.6
(εΣur)
z
0.4
0.2
10 r/η 100
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
p
following vertical velocities Σuz / hu2z i: ∗ = 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
3.0
2.5
2
<Δur|Σuz> 2.0
2
<Δur|Σuz>Σu = 0
Z
1.5
1.0
0.5
-4 -2 0 2 4
Σuz
√<u2z>
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.

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.
3.0
2.5
2 2.0
<Δur|Σuz>
2
<Δur|Σuz>Σu = 0
Z
1.5
1.0
0.5
-4 -2 0 2 4
Σuz
√<u2z>
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.

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 ).
3.0
2.5
2 2.0
<Δur|Σuz>
2
<Δur|Σuz>Σu = 0
Z
1.5
1.0
0.5
-4 -2 0 2 4
Σuz
√<u2z>
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
2
<Δuτ|Σuz>
2
3
<Δuτ|Σuz>Σu = 0
Z
-4 -3 -2 -1 0 1 2 3 4
Σuz
√<u2z>
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.

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
p
Σuz / hu2z i = −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.

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
3.0
2.5
2 2.0
<Δur|Σuz>
2
<Δur|Σuz>Σu = 0
Z
1.5
1.0
0.5
-4 -2 0 2 4
Σuz
√<u2z>
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.
5
2
<Δuτ|Σuz>
2 4
<Δuτ|Σuz>Σu = 0
Z
-4 -3 -2 -1 0 1 2 3 4
Σuz
√<u2z>
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.
3.0
2.5
3
<Δur|Σuz> 2.0
3
<Δur|Σuz>Σu = 0
Z
1.5
1.0
0.5
-2 -1 0 1 2
Σuz
√<u2z>
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.
3.0
2.5
2
<Δur|Σ|u|> 2.0
2
<Δur|Σ|u|>Σ|u| = 0
1.5
1.0
0.5
0.0 1.0 2.0 3.0
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
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 Σuk 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
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
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
ǫr
hu2+ ∆u− i = (3.3)
30
where u− is half the longitudinal velocity difference and u+ is half the sum. (For comparison, we
have used ∆ur = 2u− and Σuk = 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
that we consider or for the correlation h(Σu)2 (∆u)2 i 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 (Σuk ) could have a kinematic cor-
relation, conditioning on the transverse component (Σu⊥ ) should have no kinematic correlation.
We found that conditioning on (Σuk ) 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 Σuk . 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
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 Σuk , 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.
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
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 4
Effects 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 = (hui ui i/3)1/2
and L is the characteristic length scale L = U 3 /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 (hhui ihui ii/3)1/2 where hui i 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
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
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
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
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
(hu2 i)1/2 = ((u2x + u2y + 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
2.2
2
1.8
1.6
<(Δur)2|Σu┴>
1.4
<(Δur)2>
1.2
1
a
0.8
0.6
2.2
2
1.8
1.6
<(Δur)2|Σu┴>
1.4
<(Δur)2>
1.2
1
0.8
0.6
2.2
b
2
1.8
1.6
<(Δur)2|Σu┴>
1.4
<(Δur)2>
1.2
1
c
0.8
0.6
2.2
2
1.8
1.6
<(Δur)2|Σu┴>
1.4
<(Δur)2>
1.2
1
0.8
0.6
-3
d -2 -1 0 1 2 3
Σu┴
√<u2>
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.
2.2
1.8
1.6
<(Δur)2|Σu┴>
<(Δur)2> 1.4
1.2
0.8
-3 -2 -1 0 1 2 3
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
0.5
0.45
0.4
0.35
0.3
b 0.25
0.2
0.15
0.1
0.05
0 1 2
10 10
r/η
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
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
1.8
1.7
1.6
1.5
1.4
h(∆ur )2 |Σu⊥ i
h(∆ur )2 i
1.3
1.2
1.1
0.9
0.8
-3 -2 -1 0 1 2 3
√Σu⊥2
hu i
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
the following grid oscillation frequencies: + = 1 Hz, ◦ = 2 Hz, ∗ = 3 Hz, × = 4, Hz = 5 Hz
0.15
0.1
0.05
0 1 2
10 10
r/η
Figure 4.5: The dependence of the conditional structure function on the large scales is quantified. The
coefficient b (where the polynomial au4 + bu2 + c was fit to the conditional structure function) is shown here as a
function of r/η. This is plotted in the same way as figure 4.3. Each curve represents the fittings for the following
grid oscillation frequencies: + = 1 Hz, ◦ = 2 Hz, ∗ = 3 Hz, × = 4 Hz, = 5 Hz.
9
10
8
10
7
10
Number of Samples
6
10
5
10
4
10
3
10
2
10
-3 -2 -1 0 1 2 3
Σu┴
√<u2>
Figure 4.6: The number of samples shown as a function of large scale velocity for frequency modulation 3-0
Hz, duty cycle 50%, and period 384 seconds. The curves represent 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.
than from increasing the large scale intermittency.
Although the curvatures in Fig. 4.3 are relatively similar, there is a trend for higher constant
grid frequencies to have higher curvatures. It is very interesting to note that this is contrary to a
simple attribution of the observed dependence on insufficient Reynolds number. If it were simply
insufficient Reynolds number responsible for the dependence one would expect the dependence
to decrease with increasing Reynolds number. In addition, the dependence on the length scale
for the constant grid frequencies is fairly small with one notable exception. At 2 Hz constant
grid frequency (symbol ◦) the curvature does not follow the trend of the other constant grid
frequencies, and more prominently, has a much larger dependence on length scale. Mechanisms
for this anomalous behavior are unknown at this time.
An Illustrative Example
In order to interpret the results shown in figure 4.1, and throughout this chapter it is fruitful
to think of an illustrative example. Imagine the two modes of the motor rigorously correspond
to two modes in the fluid. When the motor is the high state the fluid is very energetic with a
high mean energy, and therefore large standard deviation of velocity. In the other mode, the
motor is in the low state and the fluid has a low mean energy, and therefore a small standard
deviation of velocity. After many periods it would become apparent the low energy fluid would
dominate the statistics near zero large scale velocity and would have a low structure function.
Quite simply, the low energy fluid would not have very many particle pairs that make it out
to 3 standard deviations large scale velocity. Instead, the structure function at the extrema of
large scale velocity would be dominated by the high energy fluid, and would have a much higher
value for the structure function.
In fact, we took experimental data that very clearly shows this illustrative example. Figure
4.6 shows the number of samples as a function of large scale velocity for different separation
distances. The frequency modulation was 3-0 Hz, the duty cycle 50%, and the period was a
very long 384 seconds. This long period allowed the weight of the transitions to be very small
so the motor on time was dominated by very energetic fluid with a large standard deviation
of velocity, and the motor off time was dominated by low energy fluid with small standard
deviation of velocity. The peculiar bumps in the center of each curve are essentially all the
additional samples of the low energy fluid to the high energy fluid.
Simply connecting the conditional structure functions of the high energy fluid at the extrema to
the low energy fluid near the center in a smooth way will create the curvature seen in figure 4.1
(and later we will see it even more clearly in figure 4.12 symbol ⋄). The greater the disparity
in energy between the fluids the greater the disparity between the structure function value of
the lower point in the center and the higher points at the extrema, and thus the higher the
curvature.
Varying the Energy Input (Large Amplitude)
Figures 4.7-4.8 are similar to figures 4.2 and 4.3 but with different grid modulation frequencies,
namely 4Hz continuous, 4-2.66 Hz, 4-1.33 Hz, and 4-0 Hz. These frequencies were chosen to
2.5
2
h(∆ur )2 i
1.5
0.5
-3 -2 -1 0 1 2 3
√Σu⊥2
hu i
Figure 4.7: Second order conditional structure functions vs. large scale velocity. 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: + = 4 Hz continuous, ◦ = 4-2.67 Hz, ∗ =
4-1.33 Hz, × = 4-0 Hz
0.8
0.7
0.6
0.5
b
0.4
0.3
0.2
0.1
0 1 2
10 10
r/η
Figure 4.8: 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: + = 4 Hz continuous, ◦ = 4-2.67
Hz, ∗ = 4-1.33 Hz, × = 4-0 Hz
have the same ratios of high to low as those in figures 4.1-4.3, the duty cycle and period were
also chosen to be the same. Note the primary result is reinforced, increasing the large scale
intermittency increases the dependence all scales have on the large scales. Comparing the cases
3 and 3-1 Hz to 4 and 4-1.33 Hz shows the ratios of the curvature are not the same, as would
be expected, but have a moderate discrepancy in the ratio of curvature from 4.3 to 6.4. The
differences between Figs. 4.7 - 4.8 and Figs. 4.2 - 4.3 highlight the non-linear nature of this
discussion. Increasing the magnitude of the frequencies, even while maintaining the ratio of the
frequencies and all other parameters, we suspect, can have secondary effects such as increasing
the mean flow, anisotropy, or other large scale properties that can not easily be predicted, but
have a measurable affect on the observed large scale dependence. Unfortunately this must remain
as speculation at this point, we have made no direct measurements to confirm this suspicion
and can draw no hard conclusions.
Varying the Duty Cycle
Let us now examine the effects of the duty cycle parameter on the observed dependence. Here
the duty cycle is defined as the fraction of the period where the motor is in its high state.
This measurement sheds some light on the sensitivity of the dependence to fluctuations in the
energy input that may occur only seldomly. It is common in actively modulated turbulence to
have the driving mechanism change states some fraction of the time. In addition, in steadily
driven turbulence fluid that had recently been in a relatively quiescent region may sweep by the
observation volume occasionally. Our goal here is to measure the impact of the relative duration
of these fluctuations on the dependence of the small scales on the large scales.
Figures 4.9 and 4.10 are plotted similarly to figures 4.2 and 4.3 with the duty cycle varied. The
grid frequency was modulated from 3 to 0 Hz with a period of 30 seconds, but the duration
the motor was high was varied. The duration high was varied from 30, 22.5, 15, to 7.5 seconds
which corresponds to duty cycles of 100%, 75%, 50%, and 25%. These results are similar to
those shown in figures 4.2 and 4.3; an increase in large scale intermittency results in an increase
in the dependence of all scales on the large scales. As the duty cycle decreases the curvature
increases. This can be explained by the limiting cases given in the illustrative example earlier;
as the duty cycle is decreased the system tends towards the limit of a quiescent fluid with
low energy and small standard deviation of velocity being combined with a high energy, large
3.5
2.5
<(Δur)2|Σu┴>
<(Δur)2> 2
1.5
0.5
-3 -2 -1 0 1 2 3
Σu┴
√<u2>
Figure 4.9: 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 modulated between 3 and 0 Hz, where the
motor was in the high state for different percentages of the period, also called duty cycles. Each curve represents
the following duty cycles: + = 100%, ◦ = 75% Hz, ∗ = 50%, × = 25%
0.6
0.5
0.4
b 0.3
0.2
0.1
0 1 2
10 10
r/η
function of r/η. This is plotted in the same way as figure 4.3 Each curve represents the fittings for the following
grid duty cycles: + = 100%, ◦ = 75% Hz, ∗ = 50%, × = 25%
standard deviation of velocity fluid. The disparity between these two modes increases as the
duty cycle decreases. For small duty cycle cases the energy in the active mode has enough time
to dissipate thoroughly, and dictates the statistics near low large scale velocity. When the motor
is on, the standard deviation of velocity increases and the large magnitude large scale velocity
is represented and has a large structure function. This creates the higher curvature, and thus
higher dependence on the large scales we observe.
Varying the Period
As we will see here the simple picture of two discrete and separate states of high or low energy
that was described in the illustrative example earlier was a hypothetical limit of the system,
and here we will experimentally explore how the flow transitions between these two states. We
will achieve this by varying the period of the frequency modulation. The grid frequency was
modulated from 3 to 0 Hz with a a duty cycle of 50%, but the duration of the period was varied
from 3,6,12,24,48, and 384 seconds.
To illustrate the two regimes and the transition between them figure 4.11 shows the energy vs.
cycle phase for various periods. Energy is simply u2x + u2y + u2z . The motor was turned off at
2π and turned on at π. The long period cases (symbols and ⋄) show the energy saturating
regime. Their maximum is at about 95 (cm/s)2 and they reach a lower limit of less than 20
(cm/s)2 . The shortest period case (symbol +) does not come close to the energy saturation
regime. It does not saturate in the maximum nor the minimum energy limit. This is because
the turbulence takes a finite time to decay and rise, approximately on the time scale of the eddy
turnover time. The shortest period case modulates the driving at about the eddy turnover time
scale, thus not allowing for much decay or rise. Intermediate cases (symbols ◦, ∗,and ×) shows
the transition between the two regimes.
Predicting the results of altering the period on the observed dependence is not as straightforward
as the previous cases, it may not be obvious which case would have the highest intermittency.
The base case (with continuous driving) may no longer have the least influence from large scale
intermittency. Guidance on this topic was provided by von der Heydt et al. [27], which showed
the response to very fast modulations was minimal. This means that in the limit of very short
periods the frequency modulation was just like oscillating at a single, constant frequency which
was approximately an average between the two driving frequencies.
120
100
80
2
energy (cm/s)
60
40
20
0
0 1.57 3.14 4.71 6.28
motor phase
Figure 4.11: The phase average energy shown as a function of cycle phase. The motor was halted at 2pi and
turned on at pi. Each curve represents the following cycle periods r/η + = 1.5/3”, ◦ = 3/6”, ∗ = 6/12”, × =
12/24” = 24/48” ⋄ = 192/384”
2.5
h(∆ur )2 i
1.5
0.5
0
-3 -2 -1 0 1 2 3
√Σu⊥2
hu i
Figure 4.12: 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 modulated between 3 and 0 Hz and
the period was varied. Each curve represents the following time the motor was on/total period (seconds): + =
1.5/3”, ◦ = 3/6”, ∗ = 6/12”, × = 12/24”, = 24/48”, ⋄ = 192/384” Note all have a duty cycle of 50%.
Whereas, for longer periods the turbulence had time to fully reach equilibrium and the fluid
could reach two distinct states. When the motor is off the energy decays rapidly, and has reached
a low energy state. When the motor is on the energy increases rapidly and has time to plateau
at a high energy limit dictated by the 3 Hz motor driving.
Figures 4.12 and 4.13 are plotted in the same manner as figures 4.2 and 4.3, These results are
similar to those shown in figures 4.2 and 4.3, in that an increase in large scale intermittency
results in an increase in the dependence of all scales on the large scales.
Figure 4.13 shows how great the disparity in curvatures can be from varying the modulation
period. The two fastest modulation periods (symbols + and ◦) have very similar and relatively
low curvatures. this suggests we have approached the lower limit in cycle period where we are
no longer increasing the large scale intermittency even though the driving frequency is being
1.5
b
1
0.5
0 1 2
10 10
r/η
function of r/η. This is plotted in the same way as figure 4.3 Each curve represents the fittings for the following
cycle periods: + = 1.5/3”, ◦ = 3/6”, ∗ = 6/12”, × = 12/24” = 24/48” ⋄ = 192/384” Note all have a duty
cycle of 50%
modulated. This contrasts to the longer modulation periods, as the period grows the curvature
grows significantly greater. This is again, due to the disparity in energy in the fluid in the two
modes. The greater the disparity in energy between the fluid with the motor on and off, the
greater the disparity between the higher points at the large Σu⊥ and lower point in the center,
and thus higher curvature. This is particularly striking for symbol ⋄ (the longest period, 384
seconds) where the statistics within ± 1 standard deviation of Σu⊥ are clearly dominated by
the lower energy fluid. As mentioned earlier, for the longest period the disparity in the amount
of samples between the high and low modes is seen clearly in Fig. 4.6 at all length scales.
Table 4.1: Experimental parameters and resulting statistics for large scale intermittency tests
Frequency (Hz) Duty Cycle Duration On (s) U (cm/s) L (cm) τ (s) ε (cm2 /s3 ) Reλ
3 100% n.a. 5.46 7.69 1.41 21.2 250
3-2 50% 15 4.72 7.48 1.58 14.1 230
3-1 50% 15 4.23 7.13 1.69 11 213
3-0 50% n.a. 4.21 7.16 1.7 10.4 212
4 50% 15 7.14 6.87 0.96 52.9 271
4-2.66 50% 15 6.04 7.23 1.2 30.5 256
4-1.33 50% 15 5.41 7.22 1.33 21.9 242
4-0 50% 15 5.38 7.61 1.41 20.5 247
3-0 50% 1.5 4.44 8.28 1.86 10.6 235
3-0 50% 3 4.71 8.51 1.81 12.3 245
3-0 50% 6 4.54 8.44 1.86 11.1 240
3-0 50% 12 4.42 7.87 1.78 11 228
3-0 50% 24 4.07 6.48 1.59 10.4 198
3-0 50% 192 4.06 5.76 1.42 11.6 187
3-0 75% 22.5 4.92 7.58 1.54 15.7 236
3-0 25% 7.5 3.27 6.86 2.1 5.1 183
5 100% n.a. 7.71 5.95 0.77 77 262
2 100% n.a. 4.05 9.3 2.3 7.15 237
1 100% n.a. 1.96 6.59 3.36 1.15 139
Table 4.1 shows the experimental parameters and resulting statistics for the large scale inter-
mittency tests.
4.2 Discussion
Although rarely addressed, large scale intermittency is almost certainly present in a large fraction
of real world and experimental flows. This experiment tries to shed light on this prevalent, yet
poorly understood topic. It is experimentally very difficult to lower the large scale intermittency
in our flow; however, by modulating the energy injection in a methodical way we can effectively
add large scale intermittency to our flow and systematically observe the results. By doing so we
demonstrate that large scale intermittency affects the structure functions at all scales, including
the small scales, significantly.
It is a concern that the addition of large scale intermittency into our flow is not a realistic
comparison to the large scale intermittency that exists in real world and experimental flows.
Addressing this concern without an accurate and accepted method for measuring large scale
intermittency is difficult. However, there are important flows where the driving mechanism
is significantly modulated, such as the active grid wind tunnel [28] and LEM [29], and it is
reasonable to expect that these flows have considerable large scale intermittency. In addition
to these flows where the driving mechanism is actively modulated, there is reason to suspect
that most steadily driven flows, and almost any low mean velocity flow (usually designed for
Lagrangian experiments) has some large scale intermittency, which will be discussed in the
next chapter. This experiment can grant some experimental backing to concerns that actively
modulated, and low mean velocity flows bring up. As an example for comparison, consider the
most modest case we measured, where the motor frequency was modulated between 3 and 2 Hz,
two turbulent states that are distinct, yet less pronounced than the 40% variation in propeller
driving frequency seen in the LEM. Even in this case the dependence on large scales increased
by approximately a factor of 2 at all length scales. Because this effect is large it is reasonable to
suspect that the majority of real world and experimental flows would also be affected by such a
sensitive dependence on large scale intermittency.
As mentioned above, an accurate and accepted method for measuring large scale intermittency is
an open problem. Work has been done to quantify intermittency in the transition from laminar
to turbulent flow in simulations, however it is unclear if this work can be translated to measuring
transitions between turbulent fluid and turbulent fluid with a different energy level.
4.3 Conclusion
We have previously established that the small scales in turbulence are not completely indepen-
dent of the current state of the large scales. This chapter has provided a comprehensive set of
experiments that has quantified the affect of one of large scale property, large scale intermit-
tency, has on all scales. This was achieved by adding large scale intermittency in a controlled
way and measuring how the structure functions dependence on large scales changed. We found
that large scale intermittency can play a significant role in the dependence structure functions
show on the large scales. We therefore conclude large scale intermittency should be included as
a large factor when considering the interactions between large and small scales.
Chapter 5
Signatures of Non-Universal Large Scales
in Conditional Structure Functions
Compared in Various Turbulent Flows
This chapter1 contains a systematic comparison of conditional structure functions in eight turbu-
lent flows. The flows studied include direct numerical simulation of turbulence in a periodic box,
passive grid wind tunnel turbulence, active grid wind tunnel turbulence (in both synchronous
and random driving modes), the flow between counter-rotating disks, oscillating grid turbu-
lence, and the flow in a Lagrangian exploration module (in both constant and random driving
modes). We compare longitudinal Eulerian second-order structure functions conditioned on the
instantaneous large-scale velocity in each flow (for a description of this method see section 3.6).
This is used to assess the ways in which the large scales affect the small scales in a variety
of turbulent flows. Structure functions are shown to have larger values when the large-scale
velocity significantly deviates from the mean in all flows except the passive grid wind tunnel,
suggesting that dependence on the large scales is typical in many turbulent flows. The effects
of the large-scale velocity on the structure functions can be quite dramatic, with the structure
1 Note this chapter is largely adapted from work which I lead and organized that has been submitted to be
published in the Journal of Turbulence as Blum et al. “Signatures of non-universal large scales in conditional
structure functions from various turbulent flows”
73
Chapter 5 - Conditional Structure Functions Compared in Various Turbulent Flows 74
function varying by up to a factor of 2 when the large-scale velocity deviates from the mean by
± 2 standard deviations. In general, the effects of the large-scale velocity are similar at all the
length scales we measured in the conditional structure functions. This indicates the large-scale
effects are scale independent.
5.1 Experiment
Parameters for each of the eight flows are summarized in Table 5.1. We define a characteristic
velocity by U = (hui ui i/3)1/2 where U represents the fluctuating velocity, and a characteris-
tic length scale by L = U 3 /ε, where ε is the dissipation rate of turbulent kinetic energy per
unit mass. The Taylor Reynolds number reported is Reλ = (15uL/ν)1/2 , where ν is the kine-
matic viscosity. We briefly describe each system with further details available in the referenced
work.
The direct numerical simulation data set analyzed has a Reλ of approximately 650, with 20483
grid points and periodic boundary conditions. The turbulence is maintained stationary by
stochastic forcing at large scales [30]. The domain size is 2.1L. Statistics are calculated from
15 snapshots of the simulation. The pseudospectral algorithm of Rogallo was used. The time
stepping is second-order Runge-Kutta. Aliasing errors were controlled by a combination of
truncation and phase shifting techniques.
One wind tunnel was used for passive and active grid turbulence measurements. The wind
tunnel is 0.914 × 0.914 m2 and 9.1 m long, with a horizontal, open circuit design. In the passive
grid configuration it produced a flow with Reλ =66. The grid mesh was M=2.54 cm and mean
flow was 10.3 m/s. Measurements were taken using hot wire anemometry at x/M=60 which is
x/L = 74 [31].
In the active grid configuration, the grid mesh was M=11.4 cm and a mean flow of 3.3 m/s was
used for the synchronous grid experiments and 7.0 m/s for the random grid experiments. It
produced a flow with Reλ =140 for the synchronous mode and Reλ =582 for the random mode of
operation. Measurements were taken using hot wire anemometry at x/M=62 which is x/L = 64
in synchronous mode and x/L = 16 in random mode [28].
The flow between counter-rotating disks is inside a 48.3 cm diameter cylindrical water tank.
The two counter-rotating disks are separated by 43.9 cm. At a propeller frequency of 3.5 Hz,
the flow reaches a high Reynolds number of Reλ =690. Nearly neutrally-buoyant polystyrene
particles 25 µm in diameter and 1.06g/cm3 in density were seeded as tracer particles. Their
trajectories were recorded using 3 cameras at a frame rate of 27,000 images per second. Two
pulsed frequency-doubled Nd:YAG lasers (wavelength 532 nm) with a combined power of roughly
150 W illuminated a 5 cm3 observation volume in the center of the tank. Particle velocities were
measured through 3D particle tracking techniques [7].
The oscillating grids are inside a 1 × 1 × 1.5 m3 octangular prism tank. The two grids were
spaced 56.2 cm apart with an 8 cm mesh spacing. A typical grid frequency was 3 Hz with a 12
cm peak to peak stroke. They produced a flow with Reλ = 260. Polystyrene tracer particles
of diameter 136 µm = 0.94η were recorded using 4 high-speed digital cameras recording 450
frames per second. The video data was filtered through real time image compression circuits,
allowing for continuous data recording. An approximately (5 cm)3 cubic detection volume was
illuminated in the center of the tank using a pulsed 50 W Nd:YAG laser (wavelength 532 nm).
Particle velocities were determined using 3D particle tracking techniques [20].
The Lagrangian Exploration Module (LEM) is a regular icosahedron with an edge length of 40
cm and a volume of 140 liters. A propeller is installed on each of the 12 vertices. The propellers
are driven by 12 independently controlled DC motors. Here we report data corresponding to a
motor frequency of 1.67 Hz. When the motor speeds were maintained constant, they generated a
water flow with Reλ =195 in the center of the apparatus. When the motor speeds were randomly
adjusted within ±40% of the nominal speed of 1.67 Hz, the Reynolds number of the flow was
Reλ = 210. Hollow glass spheres with diameter 60-70 µm and average density between 1 − 1.2
g/cm3 were used as tracer particles. A 15 × 10 × 10 cm3 detection volume was illuminated with a
30 W Nd:YAG laser (wavelength 532 nm). Tracer particle motions in the detection volume were
recorded with three high-speed cameras at a frame rate of 1,000 frames per second. Particle
velocities were measured from 3D particle tracking techniques [29].
5.2 Data
We focus on the second-order longitudinal Eulerian structure functions, labeled h(∆ur )2 i. From
the fluctuating velocity measured at two points separated by a distance r, the component of the
Chapter 5 - Conditional Structure Functions Compared in Various Turbulent Flows
Flow Reλ ε η L W/L Ref
Direct Numerical Simulation † 650 1.33 0.00141 3.08 12.8 [30]
Passive Grid Wind Tunnel 66 0.448 m2 /s3 295 µm 0.0207 m 44.2 [31]
2 3
Active Grid Wind Tunnel (Synchronous Driving) 140 0.042 m /s 550 µm 0.11 m 8.3 [28]
Active Grid Wind Tunnel (Random Driving) 582 0.94 m2 /s3 260 µm 0.43 m 2.1 [28]
2 3
Counter Rotating Disks 690 1.15 m /s 30 µm 0.071 m 6.2 [7]
Oscillating Grids 260 2.31 ×10−3 m2 /s3 144 µm 0.079 m 7.1 [20]
Lagrangian Exploration Module (Constant Driving) 195 2.10 ×10−4 m2 /s3 260 µm 0.094 m 6.4 [32]
Lagrangian Exploration Module (Random Driving) 210 2.70 ×10−4 m2 /s3 250 µm 0.10 m 6.0 [32]
Table 5.1: Flow parameters in the various apparatus. W represents the flow width, which is the smallest system dimension containing the flow.
† In DNS the choice of units is arbitrary.
76
Direct Numerical Simulation Passive Grid Wind Tunnel

2.5 a 2.5 b
Reλ = 650 Reλ = 66
2 2
<(Δur)2> 1.5 <(Δur)2> 1.5
<(εr)2/3> 1 <(εr)2/3> 1
0.5 0.5
0 0
100 101 102 103 104 100 101 102 103 104
r/η r/η
Active Grid Wind Tunnel (Synchronous Driving) Active Grid Wind Tunnel (Random Driving)
2.5 c Reλ = 140
2.5 d Reλ = 582
2 2
<(Δur)2> 1.5 <(Δur)2> 1.5
<(εr)2/3> 1 <(εr)2/3> 1
0.5 0.5
0 0
100 101 102 103 104 100 101 102 103 104
r/η r/η
Counter Rotating Disks Oscillating Grids
2.5 e Reλ = 690
2.5 f
Reλ = 260
2 2
<(Δur)2> 1.5 <(Δur)2> 1.5
<(εr)2/3> 1 <(εr)2/3> 1
0.5 0.5
0 0
100 101 102 103 104 100 101 102 103 104
r/η r/η
LEM (Constant Driving) LEM (Random Driving)
2.5 g Reλ = 195
2.5 h Reλ = 210
2 2
<(Δur)2> 1.5 <(Δur)2> 1.5
<(εr)2/3> 1 <(εr)2/3> 1
0.5 0.5
0 0
100 101 102 103 104 100 101 102 103 104
r/η r/η
Figure 5.1: Eulerian second order longitudinal velocity structure functions compensated by (εr)2/3 . a) Direct
Numerical Simulation b)Passive Grid Wind Tunnel c) Active Grid (Synchronous Driving) d) Active Grid (Random
Driving) e) Counter Rotating Disks f) Oscillating Grids g) Lagrangian Explorer Module (Constant Driving) h)
Lagrangian Exploration Module (Random Driving).
velocity difference parallel to the separation vector is extracted and the second moment gives
the longitudinal structure function [4]. Note that different experimental techniques access this
quantity in different ways. Hot-wire measurements in the wind tunnels use fixed probes, so
3 a 3
Reλ = 650 b Reλ = 66
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
3 c 3
Reλ = 140 d Reλ = 582
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3

3 e 3 f
Reλ = 690 Reλ = 260
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3

3 g 3 h
Reλ = 195 Reλ = 210
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Figure 5.2: (Color Online) The Eulerian second order longitudinal structure functions are conditioned on the
transverse velocity sum (Σu⊥ ), and plotted versus Σu⊥ . Symbols represents the following separation distances:
r/η:+ = 4, ◦ = 8, ∗ = 16, × = 32, = 64, ♦ = 128, △ = 256, ▽ = 512 ⊲ = 1024. a) Direct Numerical
Simulation b)Passive Grid Wind Tunnel c) Active Grid (Synchronous Driving) d) Active Grid (Random Driving)
e) Counter Rotating Disks f) Oscillating Grids g) Lagrangian Explorer Module (Constant Driving) h) Lagrangian
Exploration Module (Random Driving).
the separation vector has a fixed direction. Particle tracking measurements sample randomly
positioned tracer particles and average over separations with all directions. Figure 5.1 shows
Counter Rotating Disks

3 a
2.5
<(Δur)2|Σuz > 2
<(Δur)2>
1.5
0.5
-3 -2 -1 0 1 2 3
Σuz
√<u > 2
Oscillating Grids
3 b
2.5
2
<(Δur)2|Σuz >
<(Δur)2> 1.5
0.5
-3 -2 -1 0 1 2 3
Σuz
√<u >
2
Figure 5.3: (Color Online) Structure functions conditioned on the vertical component of the velocity sum, Σuz .
Symbols represents the following separation distances r/η:+ = 4, ◦ = 8, ∗ = 16, × = 32, = 64, ♦ = 128, △
= 256, ▽ = 512 ⊲ = 1024. a) Counter Rotating Disks b) Oscillating Grids .
the second-order longitudinal structure function for each of the flows we consider. The structure
functions are compensated by inertial range Kolmogorov (1941) scaling, (εr)2/3 , to better com-
pare the various flows. Most of the flows are at modest Reynolds numbers and show, at most, a
very small region of inertial range scaling. A larger separation of scales is clearly visible in the
higher Reynolds number data in Figs. 5.1a, 5.1d, and 5.1e.
To measure the effects of the large scales on each flow, we condition the structure functions
on a quantity that is representative of the large scales. We use a velocity sum, Σu, where
Σu = (u(x) + u(x + r))/2 because it is a convenient and measurable quantity that primarily
reflects the instantaneous state of the large scales. This velocity sum also can be decomposed
into longitudinal (Σuk ) transverse (Σu⊥ ) or directional (Σui ) components.
One possible concern is that the longitudinal velocity difference is kinematically correlated with
the longitudinal component of the velocity sum. Since the same measurements are used in
their calculation it is conceivable that the two values are not independent, either because of a
straightforward statistical correlation between velocity differences and velocity sums, or because
of a correlation intrinsic in Navier-Stokes turbulence. As we discuss in section 3.8, kinematic
3 a 3 b
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
3 c 3
d
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3

3 3
e f
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3

3 g 3
h
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Figure 5.4: (Color Online) The Eulerian second order longitudinal structure functions are conditioned on the
longitudinal velocity sum (Σuk ), and plotted versus Σuk . Symbols represents the following separation distances
r/η:+ = 4, ◦ = 8, ∗ = 16, × = 32, = 64, ♦ = 128, △ = 256, ▽ = 512 ⊲ = 1024. a) Direct Numerical
Simulation b)Passive Grid Wind Tunnel c) Active Grid (Synchronous Driving) d) Active Grid (Random Driving)
e) Counter Rotating Disks f) Oscillating Grids g) Lagrangian Explorer Module (Constant Driving) h) Lagrangian
Exploration Module (Random Driving).
correlations would not contribute in Gaussian random fields since the second moment of the
velocity difference is uncorrelated with the velocity sum, and we confirm this later in this chapter.
Gaussian Random Field

3
2.5
1.5
0.5
-3 -2 -1 0 1 2 3
Figure 5.5: Conditional structure functions for Gaussian random fields. The Eulerian second order longitudinal
structure function is conditioned on the longitudinal velocity sum. Symbols represents the following separation
distances r/η:+ = 4, ◦ = 8, ∗ = 16, × = 32, = 64, ♦ = 128, △ = 256, ▽ = 512 ⊲ = 1024.
However, Hosokawa [25] has shown that the Navier-Stokes equations require certain correlations
between longitudinal velocity sums and differences. There is no known relationship between the
second moment of the velocity difference and the even moments of the velocity sum, which
are the correlations that could dominate the data we present here. To minimize any possible
effects of kinematic correlations we use the longitudinal structure functions conditioned on the
transverse component of the velocity sum as our primary indicator of the state of the large scales.
The large differences between the flows in our results argues that the conditional dependence
cannot be explained by a universal kinematic correlation.
Figure 5.2 shows longitudinal structure functions for each flow conditioned on the transverse
component of the velocity sum. Figure 5.2b from the passive grid is nearly flat for all length
scales, indicating that there is almost no dependence on the large scales in the structure functions
at any r. Another flow with very weak dependence of the structure functions on the large-scale
velocity is counter-rotating disks in Fig. 5.2e. The other six flows all show significant dependence
on the large scales. The upward curvature indicates that large velocity differences preferentially
occur when the magnitude of the velocity sum is large.
Figure 5.2a shows that the DNS has a significant dependence of small scales on the large scales.
For velocities different from the mean by less than 2 standard deviations, the dependence is
very weak, but it becomes much larger for velocities 3 standard deviations from the mean. The
smallest scales show the strongest dependence on the large-scale velocity. In contrast, data
from the passive grid wind tunnel in Fig. 5.2b at much lower Reynolds number is essentially
flat. This DNS data differs from other published results [1] in which both DNS and passive
grid wind tunnel data show no dependence on the large-scale velocity. The reason for the
difference is still not clear, although we have considered several possible causes. It does not
seem to be Reynolds number since a different simulation at Reλ = 140 shows similar results
to Fig. 5.2a. The difference does not seem to be due to lack of statistical convergence since
individual snap-shots each give a similar result. Similar results also appear in data from a
simulation of decaying turbulence, indicating that the forcing mechanism is likely not the cause.
It appears that the small scales in these simulations are indeed correlated with the instantaneous
state of the large scales, and that this correlation is much smaller in turbulence from a passive
grid wind tunnel.
Figures 5.2c and 5.2d illustrate the difference in the effects of the large scales when the active
grid wind tunnel is driven in synchronous and random modes. In random mode, the curvature
is significantly larger even though the Reynolds number is much larger. A possible explanation
is that L is much larger in the random mode (See Table 5.1), rendering the flow less homoge-
neous and therefore more sensitive to large-scale effects. This explanation is consistent with the
differences between the data from the active grid wind tunnel and the passive grid wind tunnel
in Fig. 5.2b. Again, despite the lower Reynolds number, the flow with greater homogeneity has
less curvature.
A surprise in the data presented here is that data from counter rotating disks shown in Fig. 5.2e
show very little dependence on the large scales. This can be compared to the much stronger
dependence in the oscillating grid flow shown in Fig. 5.2f, even though the oscillating grid flow
is more homogeneous and isotropic. To investigate this further, we conditioned not on the
directionless transverse velocity sum but on different components of the velocity sum that are
fixed in the lab frame. Figure 5.3 reveals that the disparity between these two flows disappears
when the second order structure functions for both flows are conditioned on the component of the
velocity sum in the axial direction, Σuz which is vertical in the lab frame. Here the dependence
on the large-scale velocity for the counter rotating disk flow in Fig. 5.3a is approximately the
same as the oscillating grid data in Fig. 5.3b. Further evidence about effects of the direction
of the large-scale velocity comes from conditioning structure functions from each of these flows
on directions perpendicular to the axial direction (data not shown). Here the counter rotating
disks flow shows almost no dependence and the oscillating grids flow shows dependence similar
to the axial dependence. It appears that the direction of the large-scale conditioning can play a
significant role for systems where large-scale anisotropy is important.
Figures 5.2g and 5.2h show that the Lagrangian exploration module in constant and random
driving modes has structure functions with strong dependence on the large-scale velocity. Unlike
the different driving modes of the active grid, the different driving modes of the Lagrangian
exploration module show similar curvatures as well as similar Reynolds numbers.
A major point which is seen throughout Fig. 5.2 is that separation of scales alone is not sufficient
to produces small scales that are independent of the large scales. For most of the flows, there
is very little dependence of the conditional structure functions on the length scale, r, so the
dependence on the large scales does not diminish with decreasing scale. The DNS shows the
most dependence on the length scale, with the structure function at the smallest scales curving
upward more strongly than larger scales, exactly the opposite of what would be expected if
limited separation of scales was the cause. From this observation, we infer that the trend of the
conditional structure functions depending strongly on the large-scale velocity is likely to persist
to yet higher Reynolds numbers despite the wider range of scales that would be present.
Figure 5.4 shows longitudinal structure functions again, but this time conditioned on the longi-
tudinal component of the velocity sum. This data is largely similar to Fig. 5.2 conditioned on
the transverse sum, but conditioning on the longitudinal sum reveals a few new insights. The
similarity between Figs. 5.2 and 5.4 is evidence that kinematic correlation is not a large factor
in the observed conditional dependence. Data from the wind tunnels (b-d in Figs. 5.2 and 5.4)
shows the most differences between the conditions on the transverse and longitudinal velocity
sums. Note that the structure function minimum has been shifted towards negative large-scale
velocity by over one standard deviation. A major factor here is the orientation of the hot-wire
probes used for the wind tunnel measurements. They are fixed in space so that the positive
longitudinal direction is always downstream in the wind tunnel. The other data is taken from
particle tracking or simulations and the data is averaged over many different random longitudi-
nal directions. The wind tunnel data shows a correlation between larger downstream velocity
and larger structure functions. For the passive grid, this dependence is weak and independent
of scale, but for the active grid the structure functions conditioned on the downstream velocity
are somewhat different at different length scales.
We calculated conditional structure functions for Gaussian random fields which yielded some
additional interesting insights. Figure 5.5 shows the second order structure function conditioned
on the longitudinal velocity sum for Gaussian random fields. The Gaussian random fields are
constructed as vector fields where each component of the velocity vector is Gaussian, indepen-
dent and identically distributed. The velocity components are also subject to the condition of
being divergence free and having the same energy spectrum as the DNS velocity field at a given
Reynolds number. The result is almost perfectly flat showing no influence of the large-scale
velocity on any length scale [33]. This provides a baseline for evaluating conditional struc-
ture functions and confirms the argument presented in chapter 3 that the observed conditional
dependence is not a simple kinematic correlation between sums and differences of the same
measurements.
We also studied the effect of conditioning on the velocity at the midpoint between the two
points rather than the velocity sum. The midpoint velocity is inaccessible to particle tracking
experiments, but can be studied in the numerical simulations. Conditioning the DNS data on
the midpoint makes very little difference for small r as expected since here the two velocities are
nearly the same. At larger r there are some small quantitative differences between conditioning
on the velocity sum and the midpoint velocity, but the main features of the conditional structure
functions in Fig. 5.2a are unchanged. We also considered how the conditional dependence
changes if the conditioning point is selected away from the midpoint. For Gaussian random fields,
this introduces a kinematic correlation that is avoided by conditioning on the midpoint.
5.3 Discussion
We provide here a discussion of some possible interpretations of this data. The data clearly reveal
that the small scales have a different dependence on the large scale velocity in the different flows.
However, the causes of the differences are difficult to conclusively identify.
A central question is what distinguishes the flow that shows very little dependence of the con-
ditional structure functions on the large scales (passive grid wind tunnel turbulence) from the
other flows. The DNS is an intermediate case with very little dependence on the large scale
velocity when it is small (less than 2 standard deviations) but significant dependence beyond
this. One possible hypothesis is that the flows with weak dependence on the large scales are
homogeneous, not just across the detection volume but also well beyond it. Table 5.1 reports
the number of integral length scales that span the flow in its narrowest dimension, W/L. The
passive grid flow which has the largest W/L ratio also shows the least dependence on the large
scales. For DNS with periodic boundary conditions, W/L is defined using the size of the pe-
riodic box, and it is possible that the limited size of this domain contributes. The other flows
have smaller W/L and stronger dependence on the large scales. In highly homogeneous flows,
any fluid entering the observation volume would likely be statistically similar, and therefore
less likely to show a dependence on the large scales. In less homogeneous flows, fluid could be
swept into the detection volume from regions with different energy, leading to dependence of
the structure functions on the large-scale velocity that is responsible for the sweeping.
A related hypothesis is that structure functions depend on the large scales whenever significant
energy exists at scales much larger than L. In a passive grid wind tunnel, energy is injected at a
fairly well defined scale and then measurements are made before there is time for much energy
to be transferred up to larger scales. But in DNS and in the zero-mean-flow systems designed
for Lagrangian measurement (a and e-h in Figs. 5.1, 5.2 and 5.4), the flow is usually driven
continuously and the flow in the region studied is statistically stationary. In these circumstances
the largest length scales are dictated by the system dimensions which are typically much larger
than the integral scale, and these scales can accumulate energy. The existence of energy in these
large scales can lead to large-scale intermittency either through fluctuations in the energy flux
into and out of these scales or through transport of more energetic turbulence from near the
energy injection mechanism into the detection volume. Such energy fluctuations may result in
the dependence of all scales on the largest length scales shown in Figs. 5.2-5.4.
Another question is why the small scales in some flows show a stronger dependence on the large
scale velocity than larger scales (Fig. 5.2a and 5.2f), while in other flows all scales collapse (Fig.
5.2c, 5.2d and 5.2h). One possible consideration here is that large scale intermittency of the type
described in the previous paragraph will produce temporal variations in the energy dissipation
rate which would cause the Kolmogorov scale to fluctuate. If the Kolmogorov scale is smaller
when the large scale velocity is larger, it would increase the structure functions at small scales
when the large scale velocity is large. This is exactly what is seen in Fig. 5.2a and 5.2f. However,
more work is needed to understand this effect since it is not seen in all data sets.
We also note an intriguing connection between conditional Eulerian and Lagrangian statistics.
Lagrangian accelerations have been conditioned on the instantaneous large-scale velocity and
found to have a significant dependence in both a counter rotating disk experiment and isotropic
DNS [34–37]. The results are quite similar to the smallest scales of the Eulerian structure
functions presented here. However, here we also have data from the passive grid wind tunnel
where the conditional dependence is nearly absent. It would be very interesting to probe the
dependence of accelerations on the large scale velocity in a wider variety of flows to see if similar
flow dependence is observed there.
5.4 Conclusions
We have presented Eulerian structure functions conditioned on the large-scale velocity for eight
different turbulent flows. One flow (passive grid wind tunnel turbulence) shows very little de-
pendence on the large-scale velocity while the rest show a significant dependence that contains
information about how the unique large scales of each flow affect the small scales. This sys-
tematic comparison of conditional structure functions in different turbulent flows provides a
reference that can be used to compare the effects of the large scales in new flows as they are
developed.
Chapter 6
Conclusions
The bulk of this work has been the study of a turbulent flow between oscillating grids with
3D particle tracking and a novel real-time image compression system. Using this apparatus we
gathered a large amount of velocity statistics which have been thoroughly analyzed. We have
made significant progress in quantifying the effects of various properties of the non-universal
large scales on the inertial range and small scales.
We measured the mean and variance of the velocity as a function of distance from the grids. The
oscillating grid motion produced a weak mean flow as well as a region near the grid with high
velocity variance that falls off quickly to a very homogeneous, lower velocity variance, region in
the center. This profile has been key in the determination of the role inhomogeneity plays in
small scale statistics.
To determine the role of non-universal large scales, second order Eulerian velocity structure
functions were conditioned on the phase of the grid, an obvious source for periodic large scale
energy input. Results show little dependence of the structure functions in the center region and
surprisingly little even near the grid.
Eulerian and Lagrangian structure functions were also conditioned on the instantaneous large
scale velocity. A large dependence was found in the center of the tank, with the Eulerian
structure function increasing by a factor of 2 or more when the large scale velocity is large. The
dependence of the Lagrangian structure functions is somewhat larger. Conditioned structure
87
Chapter 6 - Conclusions 88
functions show that, in the center of the tank, all length scales are affected in approximately
the same way by the large scale velocity. The region near the grid was also analyzed and
compared with the region in the center. Near the grid, we found a much stronger dependence on
the instantaneous large scale velocity for both the Eulerian and Lagrangian structure functions
than we found in the center. Near the grid, there are clear signatures of the effects of large
scale inhomogeneity on the small scales. Fluid coming up from the energetic region nearer
the bottom grid has large structure functions, while fluid coming down from the relatively
quiescent region in the center has much smaller structure functions. The functional form of the
conditional structure functions are also different indicating the different histories of the different
fluid. These measurements provide a clear picture of the way inhomogeneity affects the small
scales of turbulence.
Although inhomogeneity was shown to be a significant factor in the dependence we observed

of the small scale structure functions on the large scale velocity, it is not the only factor, large
scale intermittency also plays a role. In order to evaluate the role large scale intermittency plays
we modulated the driving frequency and measured the structure functions in the center of the
tank. Adjusting the modulation parameters allowed us to augment the large scale intermittency
already present in the system and explore the extent to which small scales were affected by its
presence. The addition of energy fluctuations on a time scale longer than the eddy turnover
time (the hallmark of large scale intermittency) was found to greatly increase the dependence
of the small scale structure functions on the large scales.
Plotting the conditional structure functions versus the large scale velocity provides a powerful
method for visualizing the effects of the large scales on all length scales in turbulent flows.
This has been done for grid turbulence and homogeneous, isotropic DNS (Ref. [1] which show
almost no dependence of the structure functions on large scale velocity. Our oscillating grid
flow and high Reynolds number atmospheric boundary layer turbulence [1] show very similar
dependence.
To complement this comparison to previous work and further explore the effects of the large
scales in turbulence we have collaborated with other researchers in the field to compare 8 tur-
bulent flows. These flows include forced isotropic turbulence simulated on a periodic domain,
passive grid wind tunnel, active grid wind tunnel (in both synchronous and random driving
modes), counter-rotating disks, oscillating grids, and the Lagrangian exploration module (in
Chapter 6 - Conclusions 89
both constant and random driving modes). We observed that the dependence of the small scale
structure functions on the large scales is prevalent in all but one flow. The turbulent flow behind
a passive grid is unique in that no dependence on the large scales is observed..
By comparing these flows we have identified several possible causes for the dependence. The
causes include the presence of significant energy contained in the largest scales which can result
in energy flux into or out of the largest scales, and can also result in fluid transport from
regions of greatly different energies. Another possible cause is the energy of fluid adjacent to
the observation volume being significantly different from those in the observation volume, a
characteristic of the inhomogeneity of the flow. In less homogeneous flows, fluid could be swept
into the observation volume from adjacent regions with different energy, leading to dependence of
the structure functions on the large-scale velocity that is responsible for the sweeping. We found
this was best quantified by the ratio of the narrowest system dimensions to the integral length
scale W/L. In addition to identifying possible causes, the systematic comparison of conditional
structure functions in different flows provides a valuable reference that can be used to compare
the effects of the large scales in new flows as they are developed.
It is my hope that this work continues, tackling issues like the role anisotropy plays in the
dependence of the small scale structure functions on the large scales, as well as research into
finding statistics that are robust and universal, even in systems which currently show dependence
on the non-universal large scales.
Appendix A
Measurement Error Considerations
An important consideration in any experimental study is the measurement error. In experimen-

tal turbulence research a formal analysis of error propagation is generally not performed for the
higher level statistics such as the conditioned structure functions. This appendix will add on to
the measurement error discussed in the text and briefly touch on some of the error introduced
by our particle tracking algorithm, and examine statistical convergence.
A thorough study of our particle tracking algorithm was performed by undergraduate researcher
Nicholas Rotile. He compared our algorithm to the one used by the Lüthi research group at
Eidgenössische Technische Hochschule (ETH) in Zürich Switzerland. Two dimensional images
of particles in a flow were processed using the algorithm at ETH and here at Wesleyan (Wes).
Details of the comparison of these two algorithms can be found in Nicholas’ report; I will provide
here only the summary. In short, he found that our algorithm is sufficient, and under certain
circumstances more accurate then the algorithm used at ETH. The Wes algorithm does have
some distortion of the x and y axes, the distortion is minute and the velocity measurements
from both groups are very similar. The difference in particle position was most pronounced in
the x direction, and varied between the two groups by tens of microns. The difference in the
velocity measurements was also quite small, averaging only about 200 µm/s. Although this is
just the difference between two algorithms, and not the difference between our algorithm and
the true particle position, it does give us confidence that our measurements are so close to those
of an independent research group. It should also be noted that the position measurement error
90
is on the order of the known ray matching error, and is smaller than a typical Kolmogorov
length.
Another concern is for the statistical convergence in our data. In a typical calculation we take the
second moment of the velocity difference distribution as a function of r and Σu⊥ , a calculation
that requires a large amount of data. With this in mind we limited the number of r bins across
the range of interest to 8 and Σu⊥ bins to 16. Smaller bins would have lead to fewer samples
in each bin, and greater statistical error. To gain understanding of the statistical error in our
measurements we used a standard convergence determination. The entire data set was randomly
divided into 8 subsets. The conditioned structure functions were then calculated separately in
those subsets, and the standard deviation was found between the 8 subsets for each point in the
√
calculation. The standard deviation at each point was then normalized by 8 to account for
the error in the mean, and then plotted as error bars seen in figure A.1.
Figure A.1 shows the statistical error in a typical conditional structure function for various
separation distances at continuous 3Hz driving. As expected, the error is very small for all but
the largest Σu⊥ and the smallest r. We are able to measure r that is very small, but it seems
that at this particle seeding density that is relatively rare. However, this shows that in the range
of r and Σu⊥ where we are interested there is a high confidence level in the statistics.
2
1.5
0.5
2
1.5
0.5
2
1.5
0.5
2
1.5
0.5
-3 -2 -1 0 1 2 3
Figure A.1: Eulerian second order conditional structure function versus large scale velocity, plotted in a similar
manner to figure to 4.1a, with error bars representing the statistical error. The following separation distances
are shown in separate plots for clarity: r/η: ◦ = 5.33 to 10.67, × = 21.33 to 42.67, ⋄ = 85.33 to 170.67, ▽ =
341.33 to 682.67.
Appendix B
Tank Preparation
In this and the following appendices I will describe the tank apparatus used for turbulence
experiments and how to prepare it for experiments. This is not meant to replace a thorough
tutorial from an experienced researcher who has used the apparatus, and is comfortable with its
operation; this should be used as a supplement to such a tutorial. If a future researcher tries to
run the experiment based on these appendices alone he or she will almost certainly find them
inadequate, and I hasten to say it, blame the author. This and the following appendices are
meant to partially record the procedures for performing experiments and are of limited scientific
interest, instead they are intended for future researchers who will use the apparatus.
The tank was built at Wesleyan University in 2003-2004, primarily designed and by Greg Voth,
James Johnson, and Dave Boule, and constructed in the Wesleyan machine shop. Detailed
blueprints should be located in lab 039 above the desk by the optical table, with copies in the
machine shop. The technicians in the machine shop have been extremely helpful during my
time working on the tank, and I recommend using them as a resource, they are very skilled and
friendly.
First some obvious words of caution. There are several dangerous aspects to the operation of
the tank. When the motor is running the rods and rod supports oscillate with tremendous force.
Do not put any body part underneath the tank when the tank is running. As a safety feature
there is a shear pin in one of the cams which is a built in weak link that will hopefully break
and halt the oscillating movement before the oscillating parts cause too much damage. This
93
have never been tested, and hopefully never will. Another safety device is the mesh fencing that
can hang on the box beams to keep any flying debris inclosed in case of a failure. It is good
practice to keep two of these fences on when the motor is running (the other two cannot be put
on because they would block the laser beam. The tank is sturdily built and it is often necessary
to stand on the ledges, this is safe if one uses caution. The ledges are small, and tools tend to
accumulate on them. One researcher has fallen, although no injury resulted; be sure to wear
appropriate shoes when working on the tank.
There are several dangers that could harm the equipment as well. It is often necessary to
winch the structure that holds the rods (called the spider). Be sure to remove the winch before
powering on the motor, there is a note on the circuit breaker which reminds ’check spider’ for this
purpose. If the winch is fastened to the spider when the motor is running it may be destroyed.
A second point of caution is when filling the tank with water (for details on filling see appendix
D.1. Be sure the tank is open to atmosphere by opening a valve at the top of the tank. If the
tank is quickly filled while it is not opened to atmosphere the air pressure will build and the
tank may rupture. A third point of caution, there are a number of access points on the top of
tank, be sure they are all closed when filling with water. If one is left open water will spill from
the top of the tank and may damage the cameras or other equipment. An additional point of
caution when fastening the windows on to the tank. Be sure to use screws of the proper length.
If screws that are too long are used they could crack the Plexiglas. If a window screw socket
is too tight for a screw leave it without a screw. There is adequate redundancy, the windows
will not leak with one screw missing. Lastly a cautionary story, early in the development of the
tank the Plexiglas was split at a seam and all of the water was spilled. This occurred because
the tank was filled with cold water and the all-thread rods were tightened. The tank was then
filled with warmer water and thermal expansion caused the Plexiglas to split. The Plexiglas was
reglued and springs were put on the all-thread rods so the tank could thermally expand.
It is important to have a clean tank. Any debris in the tank could appear in the observation
volume and be tracked as a tracer particle even though it does not follow the flow. In addition,
dirty water can absorb light and lower the signal to noise ratio. Keeping the tank clean has been
a challenge that I have not satisfactorily met. There are several causes of debris in the tank.
Biological growth in the fill lines can enter into the tank after the filters. Hairs and dirt can
enter from open windows. Paper fibers can enter if paper towels are used to clean the tank. The
grids are corroding and aluminum oxide is contaminating the water. The seals contain metal
that can rust, and rust particles can enter the tank. The rods and linear bearing slowly wear
out and can add metal particles to the water. A good solution to remove the debris would be
to filter the water continuously for a long time before the particles are added. Unfortunately
in the current set up of pumps this adds air bubbles to the water. They may be created by
cavitation, or due to air entering the fill lines. Air bubbles should also be avoided because they
too could be mistaken as tracer particles when recording data. A future project to rearrange the
pumps to avoid the addition of bubbles could be fruitful. A program has already been written
that can occasionally stir the water for effective constant filtering (WashCycle.vi). Before the
experiment is run it is helpful to clean the tank. This can be done by vacuuming with the
shop vac, Kimwipes, or a sponge. It is up to the researcher to find the method they find most
effective. One note of caution, avoid using soap. It will be very difficult to remove all the soap,
especially from the fill lines, and it will cause bubbles which will be hard to get rid of. After
the experiment has been run it is good practice to clean the tank again. Keep a window open
to allow the water to dry quickly and avoid hard water stains on the windows, and vacuum up
the small amount of water that remains on the bottom of the tank. Any hard water stains can
be removed with vinegar, but are generally harmless. Be sure to vacuum the water from the 4
rod wells, stagnant water there maybe overlooked at biological growth may occur.
Appendix C
Calibration
The purpose of camera calibration is to set up a 3 dimensional coordinate system and locate
the cameras in that coordinate system. See section 2.3 for a brief outline.
C.1 Acquiring Calibration Images
The first step is to place the calibration mask in the tank at the location of the desired observation
volume. There are interlocking aluminum cylinders stored near the filters for this purpose. They
build off of a divot in the center of the bottom grid and hold the mask in the desired position.
Here is an example of how I set up the observation volume in the center of the tank: I place
several interlocking cylinders on the disk which fits into the divot in the center of the bottom
grid. I then locate the center of the tank by marking 75 cm from the bottom of the tank on a
post-it note stuck to the outside of the tank. I make the same 75 cm mark for the opposite side
of the tank. I then raise or lower the bottom grid via the winch attached to the spider until the
center screw of the mask is level with both of the 75 cm marks. I look through the tank and
try to make my eye level with the two 75 cm marks, and when the mask screw and the two 75
cm marks are all inline the mask is in the center of the tank. As an added check I will put the
meter stick in the tank and try to line it up next to the mask to make sure the mask screw is at
75 cm. I then look from the side of the tank and try to rotate the mask until it is parallel with
the bars of the grid, and therefore at 45 degree angles from the windows. The rear window is
96
then reinstalled and water is added into the tank. It is not necessary to fill the tank all of the
way, just to above the windows so that the refraction of the water is taken into account when
the calibration pictures are taken. Do not calibrate without water in the tank.
When the mask is in position and water is in the tank place a theater lamp behind the mask and
illuminate the mask. Turn on and adjust the camera positions and focus. This can be done by
moving the camera shelves as well as the position of the cameras on the shelves. It is important
to insure all of the cameras are focused on the same volume in the tank. To maximize this,
I open the XCAP (video5-C:\XCAP\program\xcapwxx.exe program and display a vertical line in
the middle of the screen (at 640 pixels) I align this vertical line with the middle of the mask. I
found it useful to align the vertical line on the camera with the vertical line in the ”D” printed
on the bottom of the mask. I followed a similar procedure with a horizontal line aligned with
the mask screw which I used to as a reference for half of the mask height. Once all the cameras
were pointing at the same volume in the tank I adjusted the focus of each of the cameras. I did
this by first increasing the aperture to its maximum (fstop of 2.8 this decreases the depth of
field), then rotating the focus ring until the center of the in-focus region was centered around the
vertical line displayed in XCAP. I repeated this for all of the cameras and return the aperture
to fstop 8 or 11 (note, all cameras do not necessarily need to have the same aperture).
With the cameras now in their proper positions and with proper focus take 10 .tiff still calibration
pictures in XCAP from each camera. One picture would probably be sufficient, but 10 are
normally taken so they can be averaged and the noise can be lowered. One helpful hint for
taking the calibration pictures, the theater lamp can be moved for each camera, there is no
reason why the lighting needs to be the same for calibration images of each camera. What is
important is that the mask be sufficiently bright and have even illumination across the mask
with few air bubbles on the mask.
C.2 Running the Calibration Programs
Now that the calibration images have been taken, the mask can be removed and the exper-
iment can be run (remember to take some data with a low particle density for the dynamic
calibration which will come later). I will here describe the procedure for running the cali-
bration programs which are used to determine the real space position of the cameras. First
Figure C.1: An example calibration image that is an average of 10 calibration images taken from camera 1.
the simple Matlab program ext_hd1-G:\Lab1backup\Matlab_local_lib\dbblum\Calibration\db_

calib_image_averaging_no_background_imgs.m is run, which simply averages all the calibration
images to reduce the noise. An example of such a picture is shown in figure C.1.
The averaged calibration images now need to be edited so they can be input into the calibration
program. The calibration program identifies the dots on the mask and finds their position. It
is important that no air bubbles or extraneous marks are present in the images. In Photoshop
I manually remove all parts of the images that are not dots. An example of such an edited
picture is shown in figure C.2. Note the dots that remain should be present in every calibration
image. It is possible, depending on the field of the view, that a large number of dots will be seen
from every camera. This is desirable, and can create an initial calibration that is as accurate as
possible, however it is also possible that having a large number of dots will cause convergence
problems in the calibration program. If this occurs I have edited the calibration images to lower
the number of dots into an array of 10 × 10.
Once the edited calibration images have been prepared for each camera they can be input into
the calibration program db_PTVsetupPrep_with_end_params.m. This program has a user friendly
interface and was created by Haitao Xu. Note, when running the program it is possible to
submit parameters for the dot recognition that are too lax. This will create a huge number of
Figure C.2: An example of a calibration image that has been edited so all that remains is the dots which are
viewed by each camera
recognized dots and may cause the program to stall, if this happens, restart Matlab.
It is difficult to understand the calibration algorithm; to understand the basics and the output
of the program I recommend first running and understanding gv_tsai_calibration_sandbox.m.
The basic premise is the position of the cameras are given as three angles and the distance from
the origin. The sandbox program demonstrates this and allows the user to convert from the
angles to Eulerian coordinates. When the calibration program is successfully completed the file
camParaCalib.mat will be created which contains all of the calibration parameters. The data is
now ready to be stereomatched, which will be described in appendix E.
One additional note, the original calibration program output a coordinate system which was left
handed. To convert to a right handed coordinate system the program db_left2right_coordsys2.m
was created.
C.3 Dynamic Calibration
Once initial stereomatching has successfully found particles on several thousand frames the
calibration can be adjusted using the ray intersection error found in the initial stereomatches.
Essentially dynamic calibration slightly adjusts the position of the cameras in the lab coordinates
so that the ray intersection error of the particles which have been stereomatched with the initial
calibration is minimized.
The program used to dynamically calibrate is called db_dynamic_calib2.m. There are several
relevant parameters and I recommend becoming familiar with the program before attempting
to run it (good advice for all programs in general). In order to insure the particles with which
we are dynamically calibrating are good matches they should be seen on each camera and have
a relatively low ray intersection error. Note there are versions of this program that have been
developed to correct for trapezoidal distortions inherent in viewing the particles through the
Plexiglas at an angle. Ultimately we decided not to use the trapezoidal corrections because the
differences were small.
When the dynamic calibration program has been successfully run the ray intersection error
should be reduced significantly. For an example, a typical system could have field of view with
55 µm per pixel and an initial mean intersection error of approximately 80 µm. This could be
reduced to approximately 25 µm for 4 cameras, or 14 µm for 2 cameras.
Appendix D
Data Acquisition
Data acquisition is a complex task, and I recommend recording test data with an experienced
researcher just to make sure all the procedures are clear. The set up for data acquisition can
take several hours and consists of tank preparation and camera calibration which were covered in
the previous appendices. Here I will describe the data acquisition along with all of the relevant
computer programs.
D.1 Filling the tank
Once the tank has been prepared and the cameras calibrated at least 300 gallons of degassed
water should be in the storage tank. I leave the vacuum pump on over night before the data
collection just to be sure any air bubbles captured in the process of transferring the water from
one tank to the other are removed. I have tried to make the water transfer process user friendly.
I numbered all of the 3-way valves in such a way that when the water is being transferred from
the storage tank to the turbulence tank the numbers are right side up, and when the water is
being transferred from the turbulence tank to the storage tank they are not. The 3-way valves
are also labeled with an ’L’ which shows how the water will flow through the valve. Despite
these attempts to make things user friendly it is still very easy to confuse things, so please trace
the intended path of the water before activating either pump, it is easy to dump water on the
floor, or overpressure a vessel.
101
With these words of caution make sure both tanks are open to atmosphere, all the windows are
in place, and begin pumping the water from the storage tank into the turbulence tank. If both
the coarse and fine water filters are inline (recommended for experiments) the water will enter
the turbulence tank somewhat slowly. It will take over an hour to fill. During this time I have
been able to begin other parts of the setup, but keep an eye on the water level, look for leaks
or when it reaches a few inches from the top. While the tank is filling I often fill the air bubble
suppression reservoirs at the top of the tank. Once the water line is within a few inches of the
top stop pumping and close the valve to the storage tank. The goal is now to fill the turbulence
tank with water and have no air bubbles remain at the top of the tank. Small air bubbles are a
problem because they could be entrained in the flow and be mistaken for tracer particles. Large
air bubbles are not as big of a problem, but should still be avoided if possible. Achieving the
goal of no large air bubbles in the top of the tank is not straightforward. A method I have used
with moderate success is to reduce the flow rate into the turbulence tank by partially closing the
valve near the storage tank and fill the turbulence tank until the fill line is slightly above the top
of the tank. This will cause water to rise through the valve to go to atmosphere and may cause
water to fall into the sink. This is acceptable, but be sure no water splashes on the mirrors or
lenses. If a large air bubble remain at the top of the tank it may be necessary to drain the tank
a few centimeters and try again. Keep in mind some small amount of air trapped at the top of
the tank is common, and can be difficult to remove. Once the experimenter is satisfied close all
the valves, including the one to atmosphere.
Note there is a white overflow tank resting on top of the turbulence tank. This was originally
designed to eliminate air bubbles in the tank, but I have not found it effective and do not fill it
with water.
D.2 Using the laser
I will briefly discuss the laser here. The laser is a dangerous piece of equipment and definitely
should not be run by anyone until they are familiar with it and the safety procedures. I will
give a run down of the steps to operate it here. First the cooling water valve should be turned
on, it is kind of hidden in the wall of the fume hood. Be sure power is being sent to the laser
by checking the circuit breaker. Then the laser key should be turned to ’unlock’ and the power
button pushed. This should cause the laser spring to life. The ’Cooler’ toggle switch should
then be moved to on. If the laser is being used frequently the green ’ok’ light should be on.
However, it is common for there to be a DI water fault. This occurs when the cooling water has
not been circulated through the DI filter for a long time. If this happens keep the laser on, the
cooling water pump should still circulate the water. After approximately 30 minutes shut down
the laser using turning the ’COOLER’ switch off then pressing the ’EMO’ button and try again.
If the problem persist it is possible the cooling water needs to be replaced (there is a stockpile
of steam distilled water by the filters that can be used). It is also prudent to check that there
is a proper amount of cooling water, and that no biological slime has grown on the sides of the
cooling water reservoir. The DI filter should also be replaced periodically according to the user
manual. Once the fault has been cleared and the fault light is green the lamp can be switched
on. After a few seconds the laser shutter can be switched open which will require waiting a few
moments. The display should indicate when the system is ready. At this time the system is
lasing but the light is being blocked by the laser aperture, a final check should be made that the
door to the lab is closed, safety goggles are on, and nothing obstructs the laser path into the
tank (make sure the bags are off the lenses). The ’APERTURE’ switch can be toggled up and
laser light will enter the tank. From here the beam properties can be adjusted, the current, the
pulse rate frequency, the triggering, etc.
A note about the laser, it was not mass produced like other lab equipment and may have design
or construction defects. As of now it has been returned to the manufacturer Quantronix on
Long Island twice. First the ’trans-orbs’on the power shelf malfunctioned. The second time the
whole system was returned, there were several wires that had come loose in the first repair and
biological growth had impeded the flow of cooling water. To inhibit biological growth the laser
should be run for at least an hour every month.
D.3 Using the image compression circuits
The image compression circuits must also be turned on, in a similarly involved process. First
be sure that the image compression circuits are properly cabled. By cable I mean that all
of the cables connecting the cameras to the image compression circuits and from the image
compression circuits to the computers are securely connected to the correct device. Improper
cabling happens surprisingly often. Once the cables have been checked attach the USB Blaster
to the data port on an image compression circuit and turn the image compression circuit on
using the custom made black power switch box. It is theoretically possible to damage the im-
age compression circuit by attaching the USB Blaster after the the image compression circuit
is on, so make sure it is attached beforehand. Upload the program video5-C:\NewICSFiles\
HardwareProgramming\NewImageCompressionVer2\Compression_Migration.qpf to the image com-
pression circuit using Quartus II video5-C:\altera\90\quartus\bin\quartus.exe. This can be
done through the programmer option on the tools menu. Once the program is uploaded press
both reset buttons on the image compression circuit. For the Mikrotron cameras, Impact
(video5-C:\NewICSFiles\Impact3.2\IMPACT3.exe) should now be opened. As a first test, set
the cameras on free run mode and shine a flashlight into the camera lens. Check to make sure
the image comes through (up to the maximum number of bright pixels) and moves smoothly
through the screen. It has happened before where a horizontal band will not respond to light
and the image compression circuit had to be turned off and reprogrammed. In general if some-
thing is awry turning off and reprogramming the image compression circuits is a good first step
for troubleshooting (after checking the cabling of course). Another troubleshooting tip is to
open XCAP with the image compression circuit running. The image will be encoded into what
should be vertical bands of pixels which are changing and unintelligible, but the simple test of
shining a flashlight into the lens will show a clear increase in signal. A similar procedure should
be followed for the Basler cameras, except that the option to have Impact run in free run mode
is absent. Instead the SetupB.fmt file must be swapped with one that was created for free run
mode. This should be more thoroughly described in the work of Dennis Chan.
D.4 Using external trigger
Once the image compression circuits are working in free run mode they can easily be changed over
to external trigger. For Basler cameras simply open Impact with the appropriate SetupB.fmt
file in the containing folder. Figure D.1 shows a diagram for the path of the triggering signal.
Note the pulse selector circuit. The cameras run at 450 Hz, but when the laser is running near
full power (around 13.5 - 14 A) at 450 Hz PRF it becomes unstable and the laser intensity
fluctuates. To remedy this we run the laser at 900 Hz, and the cameras at 450 Hz. To achieve
this the pulse selector circuit divides the signal by two so that the incoming signal is at 900 Hz
Function Generator
950 Hz Sync
2.5 V
EMO
Laser Control
Video7
Altera
Pulse
Selector
ICC 3 Cam 3
Sensor
Video8
DAQ
ICC 4
Cam 4
Video5
ICC 1 Cam 1
Video6
ICC 2 Cam 2
Figure D.1: A flowchart for the path of the camera trigger.

and the output is 450 Hz. As seen in figure D.1 the signal is output from the function generator
at 900 Hz and is split with one half going directly to the laser and the other half going to the
pulse selector circuit. The pulse selector circuit is nothing but a Stratix FPGA with a simple
frequency halving algorithm. Simply power the circuit and upload the program which is located
at video5-C:\altera\quartus60\Pulse_Selector\Stratix\Pulse_Selector_stratix.qpf. To test
to make sure everything is working I recommend turning off and on the trigger signal and
observing the display stop and start. I also recommend recording a small amount of data
It is important that all 4 cameras and the laser are working synchronously, early experiments
had frame skips which caused incorrect frame numbers and had to be corrected manually. This
problem was fixed by having the frame number input directly from the camera signal. However,
it is still a good idea to check for obvious frame skips. After stereomatching look for file sizes
that are drastically smaller than the others. This may be a sign that the frame numbers are
out of sync. I also recommend closing the laser aperture once or twice during a set of runs so
that a few seconds of recording will be blank. This will help identify any problems by showing
a marker when all the cameras should report zero particles found on the same frames.
For the normal frame rate of 450 Hz the exposure time for the Mikrotron cameras should be set
to 0.7 msec. This allows the cameras to capture one and only one laser pulse in a frame. The
exposure time for the Basler cameras should be set to 0.4 msec in order to capture one laser pulse
per frame. I believe the Basler cameras have the peculiar property that a 0.4 msec exposure
time will be interpreted in the following way: the shutter closed for the first 0.4 msec then open
for 0.4 msec and then closed again for the remainder of the frame period. This hypothesis was
drawn from carefully adjusting the exposure time and observing when the particles could be
seen. It is possible this hypothesis is wrong, but I do believe that these settings will garner one
(and the same) laser flash per frame in each of the cameras.
D.5 Running the motor
Once the tank is completely filled with water and all the valves are closed the motor can
be turned run. After double checking that the spider is free from the winch turn the motor
power on at the circuit breaker and open the motor control program (lab2-C:\ProgramFiles\
NationalInstruments\LabVIEW8.2\user.lib\dbblum\MotorControl.vi) in LabView. This program
controls the grid frequency as well as the acceleration, deceleration, and other parameters of
the motor. Note the frequency you enter differs from the grid frequency in the tank by a factor
of 180. For example entering 180 Hz will result in 1 Hz grid oscillation, 360 Hz will lead to 2
Hz grid oscillation and so on. To be precise the factor is closer to 167, however all previous
measurements have been made using the nominal number 180.
D.6 Water cooling
As turbulence theory tells us, stirring the water in the turbulence tank will lead to an increase
in temperature. To mitigate this a heat exchanger has been installed at the top of the tank. I
have simply connected a hose from the sink to the heat exchanger and allowed a small amount of
cold water to flow. Since so much water has a large heat capacity thermal changes will be very
slow and so such a method is effective, and really only (hypothetically) necessary for long runs.
Monitoring the temperature can easily be done using the temperature probe which attaches to
the top of the tank.
D.7 Adding particles
The particles are stored in water in a bulk container above the sink as well as some particles in
a smaller container. Be careful with the large bottle, particles are expensive, the large bottle
cost approximately $1,000. To add particles to the tank I have taken a small bottle of particles
to the top of the tank and carefully removed the temperature probe. Be sure not to lose the
small o-ring that fits around the probe and provides a water tight seal, and do not bend the
temperature probe. I then shake the small bottle to mix the particles in the water and use a
pipette to squirt the particles into the tank. The tank is probably a bit over pressurized so it
is common for water to come out of the port when the temperature probe is removed, losing
some water is acceptable but be sure to clean it up and be quick about adding the particles. I
have had the best results by inserting the pipette into the temperature probe port as far as it
will go and then squirting the particles in, and then sucking in some water from the tank and
squirting it out again into the bulk water. This insures fewer particles are left in the pipette.
As a very rough estimate 2 full pipettes corresponds to 50 particles in view, but do not rely
on this estimate. Add particles then stir with the motor and check how many particles are in
view. Keep in mind roughly 40% of particles seen in 2D are stereomatched. Remember to record
some data (several minutes) with a low particle density for the purposes of testing and dynamic
calibration before adding more particles.
D.8 Recording the grid position (optional)
Below one of the rods is a photogate that can record when it is broken as a function of frame
number. This can be used to calculate the position of the grids at any frame. Unfortunately
this system has proven to be needlessly complicated, especially when the grids are not driven
at one continuous frequency. However, if one wishes to record data using this device the Lab-
View program to do so is located at lab2-C:\ProgramFiles\NationalInstruments\LabVIEW8.2\
user.lib\dbblum\GridPhaseMeasurement\GateAndCameraRecorder.vi. The programs used to con-
vert these raw measurements to grid phase are located here ext_hd1-G:\Lab1backup\local_lib\
dbblum\lvm2gdf.
D.9 Recording data
Once everything is in place it is finally time to record data. Begin by setting the motor at
the desired frequency and allowing enough time for the turbulence to equilibrate. Setup each
computer so that Impact is ready to record data when the ’record’ button is pressed. Then stop
the trigger signal on the function generator. This will cause the laser to turn off too. Then press
reset on all of the image compression circuits, this will zero all of the frame numbers. Then
press record in Impact on all of the computers (press record as well on the photogate recorder
if it will be used). Now press the trigger button again on the function generator and the frames
should start recording. Check to see that the frame number is incrementing on all of the Impact
programs as well as the time remaining display. It has been a problem that one of the computers
will hang and Impact will not increment frames. The cause of this is not yet known, and the
experiment must be re-recorded. It is common for the Impact programs to finish at slightly
different times. As long as they all finish this is not a problem. This occurs because each
computer loses a different number of frames, and Impact will not report the recording complete
until the requested number of frames have been recorded. Keep in mind the frame numbers are
correct, so even if the elapsed time for each camera is different, all the cameras have recorded a
given frame synchronously.
To keep the file sizes manageable the maximum run duration I use is 90 minutes, which keeps the
cpv files to approximately 10 Gbyte. A few other miscellaneous tips: plan out a rough outline of
what data you must take and what data you would like to take. Things often take longer than
one would hope, and there must be an end at some point. In addition it has been my experience
that the water will become cloudy after enough run time. Also, taking data can be taxing, it
requires mental concentration in a noisy environment for a long time; try to account for this,
work in teams, and try to make sure the researcher is adequately cared for as well.
Appendix E
Data Processing
Once the data has been collected the researcher has only compressed video files (cpv files) and
calibration images. The goal is to process this raw data into 3D particles positions then to veloc-
ities then to higher order statistics (and eventually into a greater understanding of turbulence).
The computation time required for this process can be significant and there are many parts.
Here I will describe the largely automated procedure for processing the raw data.
Figure E.1 shows a flowchart which I will use to describe the data processing. We begin with
the two types of files mentioned above, the compressed video files (cpv) which Impact created
and the calibration images which were taken of the mask before the data was collected.
Begin by loading a cpv file into CPVplayer video5-C:\Impact3.3\IMPACT3.exe, and watching

some of the video. Play around some, make sure everything looks as it should, the particles are
a good size and shape, they are bright, they move smoothly, etc. It is also a good idea to check
several places in the file to make sure the data is consistent throughout the run. Try to develop
an intuition for what looks normal.
One of the features of CPVplayer is that it can recognize particles in a 2D image. It finds
the particle centers, total intensity, radii, eccentricity and frame number the particle was found
on, and writes this into an array called featdat.gdf. The user can change the parameters for
finding particles and display which pixels on the screen would be considered particles. Typical
parameters I have used are: a search perimeter of 30, this is how many pixels away from a given
110
Xcap.exe
Impact.exe
calib_image1-4_X.tif
run1-4_XX.cpv
db_calib_image_averaging_no_background_imgs.m
CPVplayer.exe
db_PTVsetupPrep_with_end_params.m
runjobs1-4_XX.sh
cpv2gdf_project_new_8-2007 camParaCalib.mat
run1-4_XX_featdat_YYY.gdf
OR
db_Stereomatch_Function.m db_Stereomatch_RUN_multiframe.m
parameterlog.mat vel3d2d_YYY.mat parameterlog.mat pos3d2d_YYY.mat vel3d2d_YYY.mat
db_dynamic_calib2.m
camParaCalib_DC.mat
db_smoothed_velocity_vector_field_creation_ML.pro
Rounded endges represent programs
Rectangles represent data files 3d_mean_vel_field_x.gdf
Shaded represents on
computer cluster
db_velocity_vector_field_subtraction_ML.pro
vel3d2d_SMS_YYY.gdf
Further Analysis
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 program must 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:
\Lab1backup\local_lib\dbblum\DataInspection\VelocityField\db_velocity_vector_field_subtraction_
ML.pro calculate the mean velocity at each voxel in 3D space and subtract that value from
each particle’s velocity so that what is left is the fluctuating velocity. Structure functions
were calculated using the program G:\Lab1backup\local_lib\dbblum\Collaboration\General\db_
stucture_function_conditional_fewrbins_uncond.pro which can create unconditioned or condi-
tioned structure functions. I have developed quite a few programs for analysis and display. Any
researcher is welcome to use and modify them, just be sure to copy the file into the user’s home
directory to retain a copy of the original.
Bibliography
[1] Katepalli R. Sreenivasan and Brindesh Dhruva. Is there scaling in high-reynolds-number

turbulence? Progress of Theoretical Physics, 130:103, 1998.
[2] Stephen B. Pope. Turbulent Flows. Cambridge University Press, 2000.
[3] L.F. Richardson. Weather prediction by numerical process. University Press,Cambridge,

1922.
[4] K. R. Sreenivasan and R. A. Antonia. The phenomenology of small-scale turbulence. Annu.

Rev. Fluid Mech., 29:435, 1997.
[5] A Arneodo, C Baudet, F Belin, R Benzi, B Castaing, B Chabaud, R Chavarria, S Ciliberto,

R Camussi, F Chilla, B Dubrulle, Y Gagne, B Hebral, J Herweijer, M Marchand, J Mau-
rer, JF Muzy, A Naert, A Noullez, J Peinke, F Roux, P Tabeling, W vandeWater, and
H Willaime. Structure functions in turbulence, in various flow configurations, at reynolds
number between 30 and 5000, using extended self-similarity. Europhys. Lett., 34:411, 1996.
[6] L. Biferale, E. Bodenschatz, M. Cencini, A.S. Lanotte, N.T. Ouellette, F. Toschi, and
H. Xu. Lagrangian structure functions in turbulence: A quantitative comparison between
experiment and direct numerical simulation. Physics of Fluids, 20:065103, 2008.
[7] NT Ouellette, HT Xu, M Bourgoin, and E Bodenschatz. Small-scale anisotropy in La-

grangian turbulence. New J Physics, 8:102, 2006.
122
BIBLIOGRAPHY 123
[8] Luca Biferale and Frederico Toschi. Anisotropic homogeneous turbulence: Hierarchy and
intermittency of scaling exponents in the anisotropic sectors. Physical Review Letters,
86:4831, 2001.
[9] X Shen and Z Warhaft. The anisotropy of the small scale structure in high reynolds number
turbulent shear flow. Phys. of Fluids, 12:2976, 2000.
[10] Michael Kholmyansky and A Tsinober. Kolmogorov 4/5 law, nonlocality, and sweeping
decorrelation hypothesis. Phys. of Fluids, 20:041704, 2008.
[11] A. Tsinober. An Informal Conceptual Introduction to Turbulence: Second Edition of An

Informal Introduction to Turbulence (Fluid Mechanics and Its Applications). Springer,
2009.
[12] K. R. Sreenivasan. On local isotropy of passive scalars in turbulent shear flows. Proc. Royal
Soc. London, 434:165, July 1991.
[13] Itai Arad, Victor S. L’vov, and Itamar Procaccia. Correlation functions in isotropic and
anisotropic turbulence: The role of the symmetry group. Phys. Rev. E, 59:6753, 1999.
[14] Itai Arad, Luca Biferale, Irene Mazzitelli, and Itamar Procaccia. Disentangling scaling
properties in anisotropic and inhomogeneous turbulence. Physical Review Letters, 82:5040,
1999.
[15] Dennis Chan, Dominik Stich, and Greg A. Voth. Real time image compression for high-
speed particle tracking. Review of Scientific Instruments, 78:023704, 2007.
[16] Florian Otto. Exploration of the flow field around an oscillating sphere using 3d particle
tracking velocimetry. Masters Thesis, Wesleyan University, 2004.
[17] Roger Y. Tsai. A versatile camera calibration technique for high-accuracy 3d machine
vision metrology using off-the-shelf tv cameras and lenses. IEEE Journal of Robotics and
Automation, RA-3:323–344, 1987.
[18] Kelken Chag, Nicolas T. Ouellette, Haitao Xu, and Eberhard Bodendschatz. Lagrangian
particle tracking in high reynolds number turbulence. EUROPEAN RESEARCH COM-
MUNITY ON FLOW, TURBULENCE AND COMBUSTION, 11:299–311, 2007.
BIBLIOGRAPHY 124
[19] Greg A. Voth, A. La Porta, Alice M. Crawford, Jim Alexander, and Eberhard Boden-
schatz. Measurement of particle accelerations in fully developed turbulence. Journal of
Fluid Mechanics, 469:121, 2002.
[20] Daniel B. Blum, Surendra B. Kunwar, James Johnson, and Greg A. Voth. Effects of
nonuniversal large scales on conditional structure functions in turbulence. Physics of fluids,
22:015107, 2010.
[21] Florian Otto, Emmalee K. Riegler, and Greg A. Voth. Measurements of the steady stream-
ing flow around oscillating spheres using three dimensional particle tracking velocimetry.
Physics of Fluids, 20:093304, 2008.
[22] Jacob Berg, Søren Ott, Jakob Mann, and Beat Lüthi. Experimental investigation of la-
grangian structure functions in turbulence. Phys. Rev. E, 80:026316, 2009.
[23] H.J.S. Fernando and I.P. DeSilva. Note on secondary flows in oscillating-grid, mixing-box,
experiments. Physics of Fluids, A5:1849, 1993.
[24] Alexander A. Praskovsky, Evgeny B. Gledzer, Mikail Yu Karyakin, and Ye Zhou. The
sweeping decorrelation hypothesis and energy-inertial scale interaction in high reynolds
number flows. Journal of Fluid Mechanics, 248:493, 1993.
[25] Iwao Hosokawa. A paradox concerning the refined similarity hypothesis of kolmogorov for
isotropic turbulence. Prog. Theor. Phys. Letters, 118:169, 2007.
[26] Michael Kholmyansky, V. Sabelnikov, and A Tsinober. New developments in field exper-
iments in asl: Kolmogorov 4/5 law and non-locality, 2008. http://ams.confex.com/ams/
18BLT/techprogram/paper_139408.htm.
[27] Anna von der Heydt, Siegfried Grossmann, and Detlef Lohse. Response maxima in modu-
lated turbulence. Phys. Rev. E, 67:046308, 2003.
[28] L. Mydlarski and Z. Warhaft. Passive scalar statistics in high-péclet-number grid turbu-
lence. Journal of Fluid Mechanics, 358:135–175, 1998.
[29] Robert Zimmermann, Haitao Xu, Yoann Gasteuil, Mickaël Bourgoin, Romain Volk, Jean-
François Pinton, and Eberhard Bodenschatz. The Lagrangian exploration module: An
apparatus for the study of statistically homogeneous and isotropic turbulence. Review of
Scientific instruments, 81:055112, 2010.
BIBLIOGRAPHY 125
[30] D.A. Donzis, P.K. Yeung, and K.R. Sreenivasan. Dissipation and enstrophy in isotropic
turbulence: resolution effects and scaling in direct numerical simulations. Phys. of Fluids,
20:045108, 2008.
[31] Armann Gylfason and Zellman Warhaft. On higher order passive scalar structure functions
in grid turbulence. Physics of Fluids, 16:4012–4019, 2004.
[32] R. Zimmermann. The lagrangian exploration module. Diploma Thesis, University of

Göttingen, 2008.
[33] A. Tsinobar, P Vedulat, and P.K. Yeung. Random taylor hypothesis and the behavior of
local and convective accelerations in isotropic turbulence. Physics of Fluids, 13:1974–1983,
2001.
[34] Greg Voth. Lagrangian acceleration measurements in turbulence at large Reynolds numbers.
Ph.D. Thesis, Cornell University, 2000.
[35] B.L. Sawford, P.K. Yeung, M.S. Borgas, P. Vedula, A. La Porta, A.M. Crawford, and E. Bo-
denschatz. Measurement of particle accelerations in fully developed turbulence. Physics of
Fluids, 15(11), 2003.
[36] N. Mordant, A.M. Crawford, and E. Bodenschatz. Experimentally lagrangian acceleration

probability density function measurement. Physica D, 193:245, 2004.
[37] A. M. Crawford, N. Mordant, and E. Bodenschatz. Joint statistics of the lagrangian accel-
eration and velocity in fully developed turbulence. PRL, 94:024501, 2005.

Thesis

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thesis

Uploaded by

Copyright:

Available Formats

Wesleyan University Physics Department

The Effects of Non-Universal Large Scales on

A dissertation submitted to the

Middletown, Connecticut April, 2011

2 Apparatus and Experiment 10

3 Effects of Non-Universal Large Scales on Conditional Structure Functions 18

3.5.2 Lagrangian Structure Functions Conditioned on the Large Scale Velocity:

4 Effects of Large Scale Intermittency on Conditional Structure Functions 47

5 Signatures of Non-Universal Large Scales in Conditional Structure Functions

A Measurement Error Considerations 90

D Data Acquisition 101

E Data Processing 110

5.1 Eulerian second order longitudinal velocity structure functions compensated by

C.1 An example calibration image that is an average of 10 calibration images taken

D.1 A flowchart for the path of the camera trigger. . . . . . . . . . . . . . . . . . . . 105

1.1 Basic Turbulence Theory

Big whirls have little whirls;

· Local Isotropy - Small scale turbulent motions are statistically isotropic.

An important note about language, it is commonplace to categorize all scales in a turbulent

1.1.1 Scaling Laws

We now use dimensional analysis to determine this relation

This directly leads to α = p/3. So now, by using dimensional analysis we have

h(∆ur )p i = Cp (εr)p/3 (1.2)

where Cp is a universal constant which is generally found experimentally. This is a testable

A traditional approach to determine if a small scale statistic is universal is to first categorize

2.1 The turbulence tank

2.4 3D Particle Finding

Camera 1’s view

2.5 Particle Tracking and Velocity

Mean Vertical Velocity (cm/s)

3.1 Characterizing the Flow

We define a characteristic velocity by U = (hui ui i/3)1/2 and a characteristic length scale by

< ∆ur > (cm/s)2

3.2 Structure Functions

h(∆ur )p i = Cp(E) (εr)p/3 (3.1)

< ∆ur > (cm/s)3

h(∆uτ )p i = Cp(L) (ετ )p/2 . (3.2)

3.3 Energy Dissipation Rate Measurement

3.4 Phase Dependence

3.5 Dependence on Large Scale Velocity

3.5.1 Eulerian structure functions conditioned on the large scale ve-

< ∆ur | Σuz > (cm/s)2

< ∆uτ | Σuz > (cm/s)2 10

10-1 100 τ/τη 101 102

3.5.2 Lagrangian Structure Functions Conditioned on the Large Scale

3.5.3 Eulerian Structure Functions Conditioned on the Large Scale

3.5.4 Third order Eulerian Structure Functions Conditioned on the

3.6 A Powerful Method for Plotting Conditional Struc-

3.6.1 Eulerian Structure Functions Conditioned on the Large Scale

< ∆ur3 | Σuz >

3.6.2 Eulerian Structure Functions Conditioned on the Large Scale

3.6.3 Lagrangian Structure Functions Conditioned on the Large Scale

3.6.4 Eulerian Structure Functions Conditioned on the Large Scale

3.6.5 Lagrangian Structure Functions Conditioned on the Large Scale

3.6.6 Third Order Eulerian Structure Functions Conditioned on the

3.6.7 Second Order Eulerian Structure Functions Conditioned on the

3.7 Properties of the Large Scales and Their Effects

3.7.1 Reynolds Number

in our flow with small shear.

3.7.5 Large Scale Intermittency

3.8 Kinematic Correlations

To make an experimental estimate of the effect of kinematic correlations, we conditioned the

4.1.1 Driving Frequency Modulation