Investigating the Entropy in the Frequency Spectra

of Spheres in Laminar to Turbulent Flow

Matthew Lee
July 2023

Abstract

The transitional regime between laminar and turbulent flow comes up in systems from aerosol particles
moving at high speeds to the trajectories of golf balls through the air. However, the mechanics of
turbulence are not trivial to model computationally with many modern schemes undergoing continuous
development over the last few decades. In this project, we focus on external shear flow, or fluid that
flows in a generally one-dimensional manner over the exterior of an object. Using the OpenFOAM 6
software, drag and lift force data is extracted from the simulation of shear flow at different Reynolds
numbers. A Fourier transform is then applied to examine the corresponding frequency distributions.
With a simulated flow field and relevant data, the following objectives can be addressed: determining
the values of Reynolds number at which the system transitions from laminar to turbulent flow, how the
drag coefficient changes with Reynolds number, and at what rate the number of dominant frequencies in
the variation of force coefficients increases in the transitional regime. Understanding these behaviors is
highly relevant within the field of fluid dynamics and will inform the modeling that is crucial for aircraft
design, particle clouds, and fluid transport systems.

1
Contents
1 Background 2
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Fluid Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Fluid Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Procedure 6
2.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Results 8
3.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4 Acknowledgements 15

1 Background

1.1 Introduction

Understanding turbulence is important in accurately modeling all sorts of naturally occurring aerodynamic
and hydrodynamic systems. From atmospheric convection currents to ocean tides, fluids are everywhere and
often do not move with predictable laminar flow. Weather forecasting, for instance, involves the analysis of
an inherently chaotic system and requires precise consideration of the behavior of turbulent flows in order
to come up with accurate predictions. It is also important to understand the development of turbulence in
man-made systems, such as in generation of lift on the wing of an airplane. Without proper simulations of
air flow that account for turbulence, there would be a catastrophic lack of control over all sorts of aerial
vehicles. Additionally, due to a concept known as the universality of chaos, further observation of reference
systems could prove vital in advancing our current models for fluid systems in general [6].
In this project, turbulence will be simulated computationally with a sphere obstructing a constant in-
coming flow. This is reminiscent of a ball moving through air or a particulate in a large flowing pipe. Such a
system can be used as a model for more complicated systems, such as the front hull of a vehicle, or generalized
for use in other fields involving external flow, such as aerodynamics. To draw conclusions about this fluid
system, the vorticity field can be used to visually identify the behavior of eddies in the flow as they shed off
the object. The net force on the object can be extracted from the simulation as well and analyzed through
time averaging and Fourier analysis. These observations will be used to better understand how the flow tran-
sitions between laminar and turbulent flow for this particular system in two main ways. The primary goal is
to see how the average drag and lift coefficient vary as the Reynolds number increases. The secondary goal

2
is to apply Fourier analysis to the force coefficient data and examine how the frequency amplitude spectra
change as the Reynolds number increases. This data analysis can be supplemented with visual observation
of the simulation under various post-processing tools such as streamlines, threshold subsetting, and cross
sections.

1.2 Fluid Mechanics

All fluids are assumed to be Newtonian and incompressible. This implies that the viscosity is constant
throughout the fluid and the divergence of fluid velocity is zero everywhere.
When a fluid flows around an object, it applies a force to it at every point along its surface. The net
force F is referred to as the aerodynamic force when air is the fluid in the system. The component of the
aerodynamic force in the direction of flow is called drag while the component normal to flow is called lift.
For slow-moving flows and simple geometries, the force distribution can be determined using Bernoulli’s
principle. However, complex flows are too sensitive to be analyzed this way and experimental testing is
usually required to measure the aerodynamic force on an object.
Using dimensional analysis, an equation for the drag and lift forces can be determined. Doing so reveals
that F ∝ ρv 2 A, and to construct the full equation we must introduce a constant of proportionality. This is
why we define the following dimensionless quantities, the drag coefficient and the lift coefficient:

Fd 2Fd
Cd = 1 = (1)
2 ρv 2A ρv 2 A

Fl 2Fl
Cl = 1 = (2)
2
2 ρv A
ρv 2 A

where 21 ρv 2 is known as the dynamic pressure, ρ is the fluid density, v is the fluid speed, and A is a reference
area scale (such as the projection of the object normal to the overall flow direction). The drag and lift
coefficients Cd and Cl are essentially dimensionless representations of the drag and lift force Fd and Fl [2].
The primary characteristic of a fluid flow is its Reynolds number, which can be derived by nondimension-
alizing the Navier-Stokes equation for momentum conservation in an incompressible fluid as

∂Ω η
+ ∇ × (Ω × u) = ∇2 Ω (3)
∂t ρV D

Doing so collapses all independent parameters in the equation into a single factor, with the convention being
to define Re = ρV D/η where ρ is the fluid density, V is a reference velocity scale (such as the freestream
velocity), D is a reference length scale (such as the chord length of an airfoil), and η is the dynamic viscosity
of the fluid.
Another relevant quantity to define is vorticity, Ω, given by Ω = ∇ × v. Vorticity describes how much
a fluid tends to rotate around a given location in the velocity field and is helpful in identifying vortices in
the flow. Often synonymous with eddies, these pockets of rotating fluid are objects of interests in turbulent

3
flows due to how they hold kinetic energy and their stability as a persistent structure that can move without
breaking apart for some time [5]. Their most important role, however, is transferring their kinetic energy to
smaller and smaller length scales until the energy can be dissipated as heat due to molecular viscosity. This
process is known as an energy cascade and is crucial to the understanding of turbulence.

1.3 Fluid Modeling

The typical description of fluids relies on the Navier-Stokes equation. However, it is astronomically inefficient
to computationaly simulate at a high enough resolution to accurately capture the volatility of turbulent flows
using just the Navier-Stokes equation. Such a simulation is called a Direct Numerical Simulation (DNS)
and is hardly ever used in practice. Typically, a more statistical approach known as Reynolds-Averaged
Navier-Stokes (RANS) is taken by decomposing the fluid velocity like

u(x, y, z, t) = u(x, y, z) + u′ (x, y, z, t).

Here, u is the actual flow velocity, u is the time-averaged flow velocity, and u′ is the difference between
u and u. This decomposition of fluid motion into a mean flow and its turbulent fluctuations allows us to
time-average and simplify the governing equations into the RANS equations. However, this decomposition
is an important conceptualization beyond RANS since it allows us to make reference to the contributions of
eddies. Notably, we define the turbulent kinetic energy as k = 12 (u′x u′x )(u′y u′y )(u′z u′z ), which is an important
idea in modeling the energy cascade.
Contrary to RANS, there is another type of simulation known as Large Eddy Simulation (LES). Rather
than finding the mean flow from the time-averaged Navier-Stokes equation, an LES simulation aims to
directly solve the original Navier-Stokes equation up to a certain length scale. This does consume more
computational resources, but will result in the appearance of finer turbulent structures and time-varying
flow rather than a steady average flow.
Many computational fluid dynamics (CFD) software applications operate using a finite volume method
in which a finite spatial domain is split up into a three-dimensional grid of small volumes to which the
governing equations can be applied. This configuration of cells is known as a mesh and is the backbone of a
CFD simulation. However, eddies smaller than the size of the mesh will not be rendered since there are not
enough cells to capture the full motion of the rotating flow. This also means that eddies just large enough
to be rendered by the mesh will be unable to disperse their kinetic energy and instead linger unnaturally
rather than breaking down into smaller eddies [1].
The underlying idea is to artificially increase the turbulence dissipation rate, commonly denoted by ϵ. It
quantifies the rate at which turbulent kinetic energy is converted to heat and therefore has units of (J/kg)/s.
Increasing ϵ should allow these barely rendered eddies to dissipate as they should, but doing so will have
to be done indirectly since ϵ does not appear in any of the governing equations as they stand. In a purely

4
numerical model, ϵ is defined as

µ ∂ui ∂ui ∂ui ∂ui

ϵ= =ν
ρ ∂xj ∂xj ∂xj ∂xj

where ν is called the kinematic viscosity and is defined as ν = µ/ρ. The approach taken in most LES models
is to augment ν and increase it by νsgs , the sub-grid scale viscosity, in order to increase ϵ.[1].
Mathematically, this argument can be similarly achieved by filtering the Navier-Stokes equation, which
is almost entirely analogous to the time-averaging process used to produce the RANS equations. Instead
of taking an average over the entire time period, a local average is taken across the time domain using a
convolution with some filter function such as the box filter, Gaussian filter, or sinc filter. By applying such a
filter to the Navier-Stokes equation, we end up with an equation that describes the mean flow with turbulent
fluctuations above the filter size. This is suitable for application to a mesh around the filter size. The filtered
Navier-Stokes equation looks like

∂u
ρ( + u · ∇u) = −∇p + µ∇2 u − ∇ · τ sgs , (4)
∂t

with τsgs = ρ(ui uj − ui uj ), the sub-grid scale stress or residual stress that represents the contributions from
sub-grid scale fluctuations on the rendered scales of motion. An eddy viscosity model is typically applied to
τsgs :

∗ 2
τsgs = 2ρνsgs Sij − ρksgs δij . (5)
3

Plugging this in to equation 4 gives

∂ui ∂ ∂p∗ ∂ ∗
ρ( + (ui uj )) = − + 2ρ ((ν + νsgs )Sij ),
∂t ∂xj ∂xi ∂xj

which agrees with the intuition of augmenting ν from earlier. p∗ is a modified pressure that accounts for the
second term in equation 5 with the sub-grid scale turbulent kinetic energy [3].
Now all that is left is to model νsgs . In the Smagorinsky model, this is done by first recognizing a
dimensional argument that νsgs ∼ U0 l0 for some representative velocity U0 and representative length l0 .
p
The velocity scale is derived from the strain rate tensor and is given by U0 = l0 2Sij Sij . The length scale
is derived from the mesh size and is given by l0 = Cs ∆ with the empirical coefficient Cs known as the
Smagorinsky coefficient and the cell size ∆. Altogether, this gives

νsgs = (Cs ∆)2

p
2Sij Sij (6)

where Cs is given a value between 0.1 and 0.2 based on software. Though this does close the system, this
model was developed for homogeneous isotropic turbulence far from walls and overestimates the sub-grid
stress in the region close to a solid surface. One method to fix this issue is Van Driest damping, which limits

5
the length scale l0 based on a continuous model for the velocity profile near a solid surface with

l0 = min(κy · (1 − exp(y + /A+ )), Cs ∆) (7)

where κ is the von Kármán constant, y is the distance to the wall, y + is the dimensionless distance to the
wall, and A+ is an empirical coefficient [1].

2 Procedure

2.1 Model

The OpenFOAM software is used to simulate and extract data from the fluid systems. First, a two-
dimensional cylinder case is configured and used to validate the procedure. The vorticity field and Fourier
spectrum from the simulation of the cylinder are compared with [4] in hopes of achieving similar results.
These two-dimensional cases are simulated on a personal laptop with OpenFOAM 10. It should be noted
that [4] uses a Lagrangian vortex particle method approach rather than a finite volume method approach,
so differences in simulation results are expected. Then, a three-dimensional spherical case is configured by
modifying the geometry of the cylinder case. The spherical cases presented in this paper are simulated using
the Bridges-2 supercomputer at the Pittsburgh Supercomputing Center, which uses OpenFOAM 6.
All meshes are generated using the built-in blockMesh utility found in OpenFOAM. Since the cylindrical
case is used for introductory testing only, the configuration details will be omitted in this paper and the
remainder of this section will refer to the spherical case. The general geometry for the spherical case is a
solid sphere placed in the center of a bounding cube with two bounding rectangular prisms extending from
opposite sides of the cube. To maximize the number of cells near the surface of the sphere, a fictitious sphere
is set around the solid sphere so that the distribution of cells can be set to be finer within the fictitious
sphere rather than throughout the entire bounding cube. The radius of the solid sphere is 1 meter, while
the radius of the fictitious sphere is 3 meters. The bounding cube has a side length of 12 meters and the
rectangular prisms have the same square cross sections as the cube. The prism in front of the sphere, where
the fluid approaches, is 2 meters long while the prism behind the sphere, in the wake, is 18 meters long.
There are a total of 619680 cells in the entire computational domain, with the cells in the bounding cube
highly concentrated near the solid sphere and concentrated to a lesser extent towards the fictitious sphere.
The cells are uniformly distributed in each of the extending prisms.
Each simulation will use the same mesh and boundary conditions summarized in table 1. Second-order
numerical schemes are used for time and spatial derivative terms. The PIMPLE algorithm is used for
pressure-velocity coupling since it provides a blend between the steady-state SIMPLE and the transient
PISO algorithm. For turbulence modeling, the Smagorinsky LES model with Van Driest damping is used
with the default coefficients found in the OpenFOAM 6 software. To adjust the Reynolds number, the

6
 Figure 1: Cross section of mesh used in spherical cases rendered in Paraview

U (velocity) p (pressure) nut (νt )

sphere 0 zeroGradient nutkWallFunction
walls zeroGradient 0 zeroGradient
inlet 2 zeroGradient calculated
outlet zeroGradient 0 calculated

Table 1: Boundary conditions as implemented in OpenFOAM

kinematic viscosity will be varied accordingly in order to preserve the velocity and time scale. For instance,
to achieve a Reynolds number Re = 100 a kinematic viscosity of ν = 0.04 is needed since the inlet velocity
is 2 m/s and the diameter of the sphere is 2 m. All simulations are run for 1200 simulated seconds with
a dynamic time step to keep the Courant number less than 1. We simulated at the following Reynolds
numbers: 1, 10, 100, 250, 300, 350, 375, 400, 425, 450, 500, 550, 700, 850, 1000, 104 , 105 , 106 , and 107 .

2.2 Analysis

The force coefficients are recorded in a separate text file at every time step using the forceCoeffsIncompressible
tool. The text files containing the force coefficient data are then imported into Python using the pandas
module. Raw plots of this data can be seen in Figures 2 and 4 as signals over time, made using the
mathplotlib module. Due to the nature of the computational model, it takes some time before the force
coefficients settle into fluctuations around a steady mean value. In order to analyze the portion of the signal
where the mean is steady, each signal must be trimmed appropriately. To be consistent across simulations,
each force coefficient signal is clipped to be between 600 and 1200 seconds, covering the second half of the
total simulation time. Plots of these clipped signals can be seen in Figures 3 and 5. The mean drag coefficient
is calculated over each clipped drag coefficient signal and plotted in Figure 7.
Due to the dependence of the dynamic time step on velocity gradients, the number of data points
varies between simulations. To avoid discrepancy due to this, the clipped force coefficient data from each

7
simulation is recast to 35000 evenly spaced time values within the clipped range. This was achieved by
linearly interpolating between consecutive data points to form a continuous function of force coefficients over
time, and then resampling along this function. With this recasting and clipping of data, we subtract off the
mean drag coefficient from each drag coefficient signal before applying the Fourier transform found in the
scipy module to each. Some of the resultant spectra are plotted in Figure 6.
For each recast and clipped drag coefficient frequency spectrum, a quantity known as entropy can be
calculated. The motivation is to develop a measure of periodicity in the force signals in order to quantify
their transition from periodic to aperiodic, presumably indicating a transition from laminar to turbulent
flow. Entropy is typically calculated for a discrete probability distribution with the formula

X
E=− pi logb (pi ), (8)

for some base b. We used a base of 2, a base of 2 is natural when considering chaotic systems that exhibit
period-doubling or a similar descent into chaos. Additionally, a base of 2 gives entropy in units of bits, which
are a familiar concept. Note that since our frequency spectra are not normalized to have their integrals
equal to 1, we must first divide every amplitude by the sum of amplitudes in the spectrum before calculating
entropy. Finally, each calculated entropy is divided by logb (N ) where N is the number of data points which
is 35000 in this case. The scaling factor logb (N ) gives the entropy for a uniform distribution, which we take
to be the ”maximum” entropy for N data points. These normalized entropies for the frequency spectra are
plotted in Figure 8.

3 Results

3.1 Discussion

Figures 3 and 5 illustrate the development in the behavior of this system as we increase Reynolds number.
For low Re less than 100, drag and lift coefficients are nearly constant. For Re between 300 and 425, we see
periodic variation in our force signals. From Re = 450 onwards, the signals exhibit some regular fluctuation,
but it is not nearly as periodic as previously. This is an interesting result that is further supported by Figure
6, which shows single stems at particular frequencies for Re < 425. This suggests that there are a finite
number of frequencies that dominate the variation of drag, which is an indication of periodicity. On the other
hand, for Re > 450, there are more clusters or smears of frequencies that dominate the signal, which is an
indication that our drag coefficient signal is aperiodic and the flow is turbulent or approaching turbulence.
Examining Figure 8 we notice a very abrupt transition in entropy around Re = 400. Before this critical
value, entropy is relatively high. This is unusual: we would usually expect laminar flow to have zero
fluctuation and therefore a frequency spectrum that is practically zero everywhere. Such a spectrum would
have zero or nearly zero entropy, which is the opposite of what actually happens. However, the general

8
Figure 2: Drag coefficient over 0 to 1200 seconds at different Reynolds number

9
Figure 3: Drag coefficient from 600 to 1200 seconds at different Reynolds number

10
Figure 4: Lift coefficient over 0 to 1200 seconds at different Reynolds number

11
Figure 5: Lift coefficient from 600 to 1200 seconds at different Reynolds number

12
Figure 6: Fourier spectra of drag coefficient for simulations close to the laminar to turbulent
transition regime

13
Figure 7: Drag coefficient averaged over last 600 simulated seconds for each simulation plotted
against their Reynolds numbers

Figure 8: Relative entropy of drag coefficient frequency spectra for each simulation plotted
against their Reynolds numbers

14
upwards trend in entropy following the critical value does seem to make sense physically since the higher the
Reynolds number, the more turbulent the flow and perhaps the more ”aperiodic” its behavior is. Around
the critical value, there is a sharp decrease in entropy that could be linearly fit in order to determine some
sort of ”rate of transition” from laminar to turbulent flow.

References
[1] Fluid Mechanics 101. Large Eddy Simulation. 2021. url: https : / / youtube . com / playlist ? list =
PLnJ8lIgfDbkoPrNWatlYdROiPrRU4XeUA.

[2] Aerodynamic Forces. May 2021. url: https://www.grc.nasa.gov/www/k-12/rocket/presar.html.

[3] Peter Davidson. Turbulence: An Introduction for Scientists and Engineers. Oxford University Press,
June 2015. isbn: 9780198722588. doi: 10.1093/acprof:oso/9780198722588.001.0001. url: https:
//doi.org/10.1093/acprof:oso/9780198722588.001.0001.

[4] D. Durante et al. “Numerical simulations of the transition from laminar to chaotic behaviour of the
planar vortex flow past a circular cylinder”. In: Communications in Nonlinear Science and Numerical
Simulation 48 (2017), pp. 18–38. issn: 1007-5704. doi: https://doi.org/10.1016/j.cnsns.2016.
12.013. url: https://www.sciencedirect.com/science/article/pii/S100757041630507X.

[5] Richard P Feynman. 1963. url: https://www.feynmanlectures.caltech.edu/info/.

[6] Robert C. Hilborn. Chaos and nonlinear dynamics: An introduction for scientists and Engineers. Oxford
University Press, 1984.

4 Acknowledgements
I would like to thank Dr. Jonathan Bennett for his continued guidance throughout my experience in the
Research in Physics program, Dr. Michael Falvo for his supervision and insight during the Summer Re-
search in Physics program, and Mr. Bob Gotwals with his training and support with using the Bridges-2
supercomputer. This work used Bridges-2 at Pittsburgh Supercomputing Center from the Advanced Cy-
berinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by
National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.

15