Fluids

Fluid dynamics
An introduction to the fundamentals
Brian D. Storey
Olin College
© All
2018 Brian D. Storey
rights reserved
Contents
1 Introduction page 1
1.1 Scope of the text 3
1.2 What is a fluid? 4
1.3 Simple fluid properties 4
1.4 Other properties 5
1.5 What is a continuum? 7
1.6 Mathematical prerequisites 9
2 Dimensional analysis 11
2.1 Units 11
2.2 Buckingham Pi Theorem 12
2.3 Making equations dimensionless 25
2.4 Summary 29
3 The empirical approach 31
3.1 Drag revisited 32
3.2 Drag coefficients 35
3.3 Pipe flow 38
3.4 Summary 42
4 Vector calculus notation and review 43
4.1 Scalar and vector functions 43
4.2 Gradient, divergence, and curl 45
4.3 Normal vectors and flow through a surface 47
4.4 Volume integrals 48
4.5 Surface integrals 49
4.6 Surfaces and volumes in 2D 51
4.7 Divergence theorem 52
iv Contents
4.8 Summary 53
5 Diffusive mass transfer 55
5.1 Physical picture of mass diffusion 56
5.2 Continuum concentration 59
5.3 Equation in 1D 65
5.4 Role of simulation 73
5.5 Summary 74
6 Convective mass transfer 75
6.1 Convective mass transfer 75
6.2 Material point of view - be one with the fluid 76
6.3 The material derivative 77
6.4 Liebniz and Reynold’s transport theorem 81
6.5 Conservation of mass: the fluid 82
6.6 Dimensionless formulation 86
7 Conservation of momentum in a fluid 89
7.1 Transport theorem revisited 90
7.2 Linear momentum 91
7.3 F=ma 94
7.4 So what’s a tensor? 101
8 The Navier Stokes equations 103
8.1 Euler’s equation 104
8.2 Newtonian fluid & incompressible flow 106
8.3 Boundary conditions 109
8.4 Computing stress from flows 110
8.5 Comments on kinematics 111
8.6 Non-dimensionalization 117
8.7 The Reynolds number 118
8.8 Summary 119
9 Solutions to the Navier-Stokes equations 123
9.1 Flow between parallel plates - Poiseuille flow 124
9.2 Flow driven by a wall - Couette flow 129
9.3 Flow down a ramp 131
9.4 Combined Poiseuille and Couette flow 132
9.5 Slider bearing 134
9.6 Impulsively started Couette flow 138
9.7 Cylindrical Couette flow 140
Contents v
9.8 Poiseuille flow in a pipe 143

9.9 Comments on the stability of solutions 145
9.10 Computational Fluid Dynamics (CFD) 148
10 Inviscid flow, Euler’s equation and Bernoulli 151
10.1 Flow along a streamline: Bernoulli 152
10.2 General Bernoulli 154
10.3 Euler equations across a streamline 156
10.4 Vorticity 161
10.5 Irrotational flow 167
10.6 Lift on an airfoil 169
11 Boundary layers 173
11.1 Impulsively started plate 174
11.2 Boundary layer equations and laminar solution 176
11.3 Turbulent boundary layers 182
11.4 Boundary layer separation 183
11.5 Observations of drag on a sphere 186
12 Turbulence 191
12.1 Can we simulate turbulence? 191
12.2 Kolmogorov theory and the energy cascade 192
12.3 Numerical simulation of turbulence (CFD) 195
13 Control volume analysis 199
13.1 Control volume formulation 200
13.2 Examples 202
13.3 Some difficulties 212
Bibliography 215
1
Introduction
The text lays the foundations for the study of fluid dynamics. The
first question a student might ask is why should we be interested in
fluid dynamics? Let me address the question through some examples
of applications where these topics are relevant.
Biology: Our bodies are mostly water. Fluids are essential to life as
we know it. The understanding of the transport of oxygen and nutrients
throughout the body by fluids (at the level of lungs, heart, and blood
vessels as well a the cellular one) requires a basic understanding of the
topics herein. Locomotion of many species from fish to birds to small
microorganisms is determined by fluid flow.
Geophysical: The atmosphere and oceans are fluids. The currents,
the winds, the gulf stream, and the jet stream carry energy and matter
around the planet. Our basic understanding of the weather, the climate,
and the environment can only begin with a basic understanding of fluid
dynamics. The Earth’s magnetic field is generated by convection of the
mantle deep within the Earth.
Energy: Most of our current energy needs are met by the combustion
of fuel; the basic processes of combustion depend upon the details of
the fluid mixing of fuel and air and and the resulting heat release.
Advances in understanding of the details of these processes has led to a
number of developments over the years which made combustion cleaner
and significantly reduced the number of pollutants emitted from cars.
Many of the renewable sources of energy such as wind, geothermal,
ocean waves and tidal seems obvious to state that fluid flow plays a
prominent role. It is fair to claim that an understanding of all energy
systems must begin with the basic understanding of topics in this text.
2 Introduction
Climate The computer models used by climate scientists begin with

the basic equations we will develop here. Since we can only run a sin-
gle irreversible real-time experiment on our planet’s climate, computer
models are needed to understand what is likely to happen to the cli-
mate in the coming centuries. One cannot appreciate the complexity,
accuracy, and shortcomings of these computer models without under-
standing the basics of fluid transport.
Infrastructure and transportation: Fluid dynamics also plays a
role in our basic civil infrastructure; water delivery to our homes, water
resources management, sewage, building heating and cooling. Losses
due to aerodynamic drag on boats, cars, and planes set the efficiency
and energy use of most transportation systems.
Engineering: Many fields of engineering need some understanding
of flow transfer. Aerospace engineers need to understand fluid flow to
get planes to fly. Automotive engineers need to understand the pro-
cesses of combustion, the lift and drag forces on the vehicle, as well
as how to keep the engine cool. Civil engineers need to understand
water management in cities. Computer engineers need to understand
that dissipation of heat is one of the main limitations in making faster
computer chips. As argued before, our bodies are mostly water, thus it
seems obvious that bio-engineers need to understand fluids as well.
Astronomy: Despite the extraordinarily large scale, the formation
of galaxies, planets, and stars can be studied with the equations of
fluid dynamics. We cannot test experimental galaxies in a lab, thus
fluid dynamics simulations is an integral part of answering questions
about the formation of objects in our universe.
Mathematics: Fluid dynamics has also provided a rich field of study
for mathematicians. The basic equations of fluid dynamics, the Navier
Stokes equations, are notoriously difficult to solve. If you can provide
a general solution, you can win a million dollars (see the Millennium
Prize problems sponsored by the Clay Mathematics Institute). There
are many cases where in an attempt to solve a problem in fluid dynamics
has lead to new and powerful mathematical techniques. Computational
fluid dynamics, solving the Navier Stokes equations numerically, has
been one of the most powerful drivers in developing supercomputers
and other advances in high performance computing.
In short, while my viewpoint is biased, the foundations in this book
are critical to essentially all scientists and engineers.
1.1 Scope of the text 3
1.1 Scope of the text

Now that I have built up the number of applications this field is rele-
vant to, we must scale things back. A study this comprehensive would
overwhelm most mortals over their lifetime. We only begin our study
of these topics and this course will provide only a basic background.
This text will mostly provide the theoretical and mathematical foun-
dations. The book is not meant to stand-alone but should be part of
course which also involves a balance of working examples, conducting
some simple hands-on experiments, and learning how computer pack-
ages can help solve some of these problems.
It is expected that this text will provide the (somewhat) rigorous
background and derivations. You should read and work through many
of these results yourself. Even though the text is not too long, many of
the chapters are very dense in terms of content. In my course we will
review some of these high points in class, but I will leave it to you to
really read through the details. The details can sometimes be tedious,
and until you have some working knowledge (perhaps by the end of the
course), it is hard to see why one would care about them. These notes
should hopefully provide you a view that the mathematical foundations
for the field are relatively rigorous. In many cases, we can write the
equations down in a few short lines and we could in theory solve any
problem. We will soon see that matters are not so simple and that
while the equations are known, their solution can be extraordinarily
challenging.
Finally, there are a number of good books on these topics. I will never
provide some insight that was not thought of before. My treatment
will rely on a number of excellent books through which I learned the
topic. So why bother to write my own notes? The basic answer is that
professors can’t help themselves. The way everyone else explains stuff
is never quite the way you want it. However, the excellent books that
I use the most and those who influenced my view are,
• Fluid Mechanics by Kundu and Cohen (2004).

• Transport Phenomena by Bird et al. (1960).
• Div grad curl and all that by Schey (1997).
• Incompressible flow by Panton (1996).
• Album of fluid motion by VanDyke (1982).
4 Introduction
1.2 What is a fluid?

We have already talked about fluids a lot. The two fluids we will dis-
cuss most often are air and water. There are some real fundamental
differences between air and water - yet we will treat them on the same
footing as “fluids” in this book. So it would be reasonable to ask, what
is a fluid? A precise definition seems difficult. The most popular exam-
ple of a substance that is not clearly a fluid or a solid is Silly Putty. Put
it on the table, leave it, and it flows. Drop it and it bounces. Silly Putty
is just one example of a complicated material, though many more exist.
Substances which we normally think of as a solid can act like a fluid
(and vice versa) under extreme conditions. Hit water hard enough and
it can fracture. Wait long enough and a mountain can flow. You should
search the internet for “pitch drop experiment” to see an interesting
long term experiment. A funnel of pitch tar was placed in a funnel in
1927 and has been dripping ever since. To date nine drops have fallen,
so the material does flow. However, hit the material with hammer and
it easily shatters. Sometimes the difference between a fluid and solid is
just a matter of time scale.
The primary property that differentiates fluid from solid behavior is
that fluids cannot support shear forces at equilibrium. If I take a block
of solid material between my hands and shear it, the solid resists the
motion. The solid may deform, but the solid can resist and ultimately
stop the motion. Fluid on the other hand, will just flow and cannot
stop the shearing force. I can shear the fluid continuously.
We will be a little more precise later when we develop a mathematical
theory. We will see we need to define a constitutive law which relates
deformation to forces. The simplest fluid model is called a Newtonian
Fluid, which turns out to describe air, water, and oils quite nicely. It
does not describe Silly Putty. Later we will define fluids based on their
constitutive laws. For now just go with your intuition - if it flows, its a
fluid and not get caught up in a precise definition just yet.
1.3 Simple fluid properties

There are two common fluid properties that we will refer to throughout
this book, density and viscosity.
1.4 Other properties 5
Density which typically has the symbol, ρ, is nothing more than the
mass per unit volume. Everything we discuss in the development of the
theory is given on a per unit volume basis, thus the density will show
up throughout the course. Density of a liquid is really easy to measure,
just fill a beaker of known volume and weigh the mass of the fluid. The
ratio of mass divided by volume is density. When we write our fluid
version of Newton’s laws, density will take the role of mass when we
write F = ma.
Viscosity which typically has the symbol µ (though η is common
in chemical engineering) describes a substance’s resistance to flowing.
You probably have some intuition about this concept based on pouring
water and syrup. A simple device for quantifying viscosity comprises
two concentric cylinders with a very narrow gap between them. A fluid
of interest is placed in the gap. The inner cylinder is held fixed while the
outer one rotates at a constant angular speed, ω. The speed of the wall
at the outer cylinder is simply U = ωR, where R is the radius of the
cylinder. If we conducted this experiment for different gap sizes d and
different speeds, we would find the torque needed to spin at constant
speed would follow a law,
U
Torque ∝ µ .
d
We would find for all things equal, syrup is harder to spin than water
because the syrup has a higher viscosity. Not all fluids would follow the
law above. Only simple fluids such as water, air, syrup, and oil. Fluids
that follow a law such as that above are called Newtonian, a definition
we will be precise about later.
1.4 Other properties

There are many other properties of fluids that show up but will not
play much of a role in our initial study.
Surface tension is a usually given the symbol σ and has units of
force per unit length. Surface tension exists at the interface of a liquid
with a gas and it is the force that holds raindrops together and makes
them want to be spherical. Surface tension can hold relatively small
objects like water striders from sinking into the pond. Surface tension
6 Introduction
can hold water droplets to the ceiling tiles in your shower. Surface
tension also acts at the interface between two immiscible liquids like
oil and vinegar. If there is no free surface in the liquid then surface
tension does not play a role. For flows at a liquid interface for objects
the size of boats and people, surface tension is too small to matter
much. For objects that are measured in millimeters, surface tension
can be a dominant force.
Speed of sound is exactly what it says - it is the speed that sound
travels in the fluid. The speed of sound becomes important when flow
speeds start to approach it. The ratio of flow speed (or speed of the
object moving through the fluid) to sound speed is called the Mach
number, Ma. When the speed of the object becomes greater than the
sound speed, Ma > 1, then we have supersonic flow. Supersonic flow is
very interesting as much of our intuition about fluid flow goes out the
window. Supersonic flow is a topic for another course.
When the Mach number exceeds about 0.3, then the flow becomes
compressible meaning that the density can no longer be considered a
constant (this is a rule of thumb, not an exact criteria). It is the Mach
number that tells us when we need to start thinking differently about air
and water even though we called them both fluids. Gases are relatively
easy to compress when placed inside a closed container while liquids
are not. When we have flow, it is the Mach number that characterizes
whether the flow is compressible or not. Note that it is important to
distinguish between a compressible fluid and a compressible flow. Air is
an easily compressible substance, however most flows that we encounter
with air on a daily based are incompressible flows. It is the Mach number
that helps determine if we have a compressible flow.
Elasticity Normally we think of solids exhibiting elasticity. I pull
on a rubber band it will stretch. When I let go it will return to its
initial state. Simple fluids like water, air, syrup and simple oils do not
exhibit any significant elasticity. If I shear the fluid and then let go, the
fluid does not try to return to its original configuration. Some fluids
such as biological fluids or polymeric solutions exhibit both elasticity
and viscosity - typically called viscoelastic fluids. Silly putty is a classic
example, I can roll it in a ball and it will bounce or leave it on the table
and it will puddle up. In the case of silly putty, it is the time scale of
the flow that distinguishes between whether the flow is more viscous or
1.5 What is a continuum? 7
more elastic. Viscoelastic fluids have many interesting properties but

will not appear in our initial study.
Other viscosity parameters The basic behavior of simple fluids
such as water and air are the only one we will consider. Fluids that
obey our basic viscosity law are called Newtonian fluids, thus anything
that does not follow this law is generically called Non-Newtonian. The
term Non-Newtonian fluid is not very descriptive as it applies to any
behavior which is, well, not Newtonian. We will be more precise later
about what we mean by a Newtonian fluid.
Beyond viscoelasticity, another non-Newtonian behavior is called shear
thickening. Shear thickening means that the faster we stir the fluid, the
hard it becomes to stir. Cornstarch mixed with water is an example of
this behavior - you can find videos where people run across a pool of it
but when they stand still they sink. We can also have shear thinning,
meaning that the faster we stir the easier it is to stir. We can also have
fluids that act as a solid as long as the applied forces are small, but
once we exceed some threshold they start to flow. One could spend a
career (and many have) just studying Non-Newtonian fluids.
1.5 What is a continuum?

In this course we will develop a theory where materials can be consid-
ered to be a continuum. We will generally not be concerned that the
fluid or solid is made up of molecules or atoms. This assumption works
well in applications where the features we are interested in studying are
much larger than the distance between molecules. The mean free path
of air at standard conditions is approximately 0.1 µm. As long as the
smallest features of the problem are greater than this scale, the con-
tinuum approximation is a good one. In liquids the mean free path is
usually a few angstroms (an angstrom is 10−10 m), making cases where
we must account for non-continuum effects rare for an engineer.
To be more precise, let’s consider a measurement of the density of
air as an example. We could measure the density by taking a known
volume of air and weighing it. If our volume was the size of the room,
the density measure would be the average density in the room. As we
shrink our measuring volume, the value we measure for the density
might change. The density of warm air is less than cool air, so if we
8 Introduction
Height (z)
z
Density
Figure 1.1 Schematic of the concept of a continuum. The local mass

density of a gas varies with height in a room with less dense fluid
moving to the top. If we zoom in to sample the density we can obtain
a local value by using smaller and smaller measuring volumes. At
some point, we trapped so few molecules in the volume that the
notion of density at such a small scale loses meaning.
took our measuring volume near the ceiling we might expect a different
answer than near the floor. This idea is shown schematically in Figure
1.1, where the local measurement of density is shown to be linear with
height, with less dense fluid rising to the top. As we shrink the mea-
suring volume we obtain a more local measure of the density. At some
point we keep shrinking the measuring volume and we keep getting the
same answer. This constant value would be what we would consider
the local, continuum value of the density at that point in space. If we
continued to shrink the measuring volume we run into trouble with the
fact that air is comprised of many molecules. Eventually the measur-
ing volume would be so small that the mass we measure would depend
upon how many molecules we managed to trap. If there are only 10
molecules on average in our volume, then the answer starts to fluctuate
depending on if we randomly had 8, 9, 10, 11, or 12 molecules in the
volume at the time of measurement. The continuum limit is thus the
intermediate asymptotic, where we take the limit such that we can talk
about local quantities (i.e. density, temperature, pressure, velocity) at
a point in space, but we don’t take the limit too small that the molec-
ular nature of matter comes into play. As long as the geometry of the
problem is large compared to the molecular scale, then the continuum
assumption is good.
1.6 Mathematical prerequisites 9
1.6 Mathematical prerequisites

I will assume throughout that the reader is familiar with Newton’s
Laws, Calculus (integration and differentiation), and basic vector cal-
culus. If you know F = ma, can compute some simple integrals, and
have heard of the divergence, gradient, and curl, you are probably good.
If it has been some time since you took vector calculus, I would recom-
mend the book Div, Grad, Curl and All That as a complement to these
notes. A few common things that we will make use of are provided
below.
We will make use of the Fundamental Theorem of Calculus which
states
Z b
df
dx = f (b) − f (a).
a dx
We will use vector calculus notation such as the gradient of a scalar
function F (x, y, z) is a vector given as
∂F ∂F ∂F
∇F = i+ j+ k,
∂x ∂y ∂z
where i, j, k are the unit vectors in the x, y, z direction. We will use the
convention that bold face lower case letters are vectors. The divergence
of a vector field f (x, y, z) is a scalar given as
∂fx ∂fy ∂fz
∇·f = + + .
∂x ∂y ∂z
The primary vector calculus theorem we will use is the divergence the-
orem, which states,
Z Z
∇ · f dV = f · ndS.
This theorem equates the flux of a vector field through a closed surface
to the volume integral of the divergence of the vector field. Above, we
use the notation that dS is a unit of surface and dV is a unit of volume.
If if the above notation seems vaguely familiar to you but you have
forgotten the details, you will have no problem. We will review vector
calculus in a later chapter. If you have never seen these symbols before,
you might have some trouble understanding some of the later chapters.
2
Dimensional analysis
The result of any physical equation must remain unchanged regard-

less of the system of units that we select for the problem. While this
seems like a obvious statement the implications are quite interesting.
The simple fact that units are man-made and thus arbitrary can lead to
an extraordinary simplification in many problems. The number of free
parameters in any physical problem can be reduced by taking the units
out of the problem. This concept is not limited to the topics in this
book, however it seems that the fields of fluid dynamics and transport
phenomena have adopted this viewpoint to the extent that the discus-
sion of dimensional analysis is in every book. As we will soon show, the
reasons for adopting a dimensionless world-view are quite compelling.
Books of experimental data can be reduced to a single chart.
However, the dimensionless world-view can be overdone and only
working in a dimensionless world can obscure some intuition about the
physics of a particular problem. We will try and balance the dimensional
world-view and the dimensionless one throughout this book.
2.1 Units
We often refer to fundamental and derived units. Fundamental units in
this course will be mass (M ), length (L), time (T ), and temperature
(Θ). Derived units are things like velocity (L/T ), force (M L/T 2 ), or
density (M/L3 ). The distinction above is arbitrary. Of course we could
say that force is fundamental and mass is derived. If you want to do
12 Dimensional analysis
this, this is your right. I find it easiest to stick with the distinction
above.
We could relate temperature and energy and thus remove tempera-
ture as a fundamental unit. In this course, this path would not lead to
simplification. It would require us to bring in a new parameter, Boltz-
mann’s constant. Boltzmann’s constant bridges the statistical definition
of kinetic energy of molecules to the macroscopic temperature. If we are
not dealing with statistical mechanics, then introducing Boltzmann’s
constant just leads to confusion. In this course, everything is macro-
scopic and thus treating the temperature as its own fundamental unit
is always going to be the way to go.
2.2 Buckingham Pi Theorem

The Buckingham Pi Theorem is the starting point for any discussion of
dimensions. The Pi Theorem states that if a problem has N independent
parameters and there are R independent dimensions, then there are N-
R independent parameters in the problem. This simple theorem can
only be understood through example.
2.2.1 Pendulum
The first example on dimensional analysis every book uses is the pen-
dulum. I won’t be original. Imagine a point mass (m) hanging on a
massless string of length (`) in a gravitational field (g); shown schemat-
ically in Figure 2.1. There is no friction. All this is an idealization, but
it is one that we can easily realize in practice to a good degree of ap-
proximation. The question is what is the period of oscillation, t, for a
given starting angle θ and how does it depend upon the parameters in
the problem? There are five independent parameters (m, `, g, θ and
t). These parameters are made up of 3 independent dimensions (M , L,
and T ). Applying the theorem tells us that N = 5, R = 3 and thus
there are only 2 independent parameters.
There are a number different ways to formally approach the removal
of units from the problem to decide what the dimensionless parameters
are. Once you are comfortable, you can often just remove the units by
inspection. I like the table method as it is a provides a good framework
θ g
ℓ
m
Figure 2.1 Schematic of a simple pendulum.
for formally approaching the problem. We make a table with all the
parameters and their units in brackets. We put the primary thing we
want to know, in this case the period, in the first column. For this
problem, our table starts as,
t [T ] m [M ] ` [L] g [ TL2 ] θ
Note that the starting angle θ has no units. Angles are measured in
radians which is defined as the arc length divided by the radius, thus
the radian is already dimensionless.
Now start adding lines to the table, and in each new line remove one
of the independent dimensions and one dimensional parameter from
the problem. Let’s start by removing mass from the problem. Since
the dimension M only exists in the mass, there is nothing to do but
remove it from the problem. Mass cannot be a parameter to the problem
as there is no way to cancel that fundamental unit. With each new line
to the table it is useful to add yourself a little note to the right.
t [T ] m [M ] ` [L] g [ TL2 ] θ
t [T ] ` [L] g [ TL2 ] θ remove M
Now let’s remove length from the problem by defining a new param-
eter g/` that does not have length in it. We expand our table by one
line so it reads,
t [T ] m [M ] ` [L] g [ TL2 ] θ
g 1
t [T ] ` [T2 ] θ remove L
Finally, we can remove time from the problem by adding another line
to our table as follows,
t [T ] m [M ] ` [L] g [ TL2 ] θ
g 1
t [T ] ` [T2 ] θ remove L
t g`
p
θ remove T
p
The final result is that there are two parameters, t g/` and θ. We
can then state that the answer has the following functional form,
r s
g `
t = f (θ) or t = f (θ)
` g
At this point we don’t know what the function is, but it says the period
of any pendulum of any length in any gravitational field only depends p
upon its starting angle. We only need to measure time in units of g/`,
and all pendulums are the same. If we always started the pendulum
from the same angle, the function on the right hand side would then
just be a constant. The result tells if the length of the string is increased
by
p a factor of 4, then the period increases by a factor of 2 to keep
t g/`=constant. You can easily confirm this result experimentally.
2.2.2 Pythagorean theorem

Dimensional analysis can easily be used to prove the Pythagorean the-
orem, a2 + b2 = c2 for a right triangle. The area of the triangle A only
depends upon the angle θ and the length of the hypotenuse c. We don’t
even need the table, we could move directly to the result that the area
divided by c2 is a unknown function of theta, A/c2 = f (θ). The areas
of triangles 1 and 2 are thus A1 /a2 = f (θ) and A2 /b2 = f (θ). Since
A1 + A2 = A, then a2 f (θ) + b2 f (θ) = c2 f (θ). Since the function f must
be the same for all triangles, then we can divide it out and obtain the
desired result.
c A2
b
A1
θ
a
Figure 2.2 Schematic of a right triangle. It is easy to prove a2 +b2 =

c2 using dimensional analysis.
2.2.3 Caveats
There is a nice paper written by Lord Rayleigh on the “Principle of
Similitude” (Rayleigh (1915)); note that similitude is another name for
dimensional analysis. He starts the paper;
I have often been impressed by the scanty attention paid even by original
workers in physics to the great principle of similitude. It happens infrequently
that results in the form of “laws” are put forward as novelties on the basis
of elaborate experiments, which might have been predicted a priori after a
minutes’ consideration.
The Rayleigh paper then gives, in essentially 1 sentence each, 13

examples of answers to various problems in physics. The true analytical
answer to most of these problems is not trivial. However, the result
Rayleigh provides is obtained by essentially applying the Pi Theorem.
His results are all correct and some are as simple to obtain as the
pendulum example. For many of the examples, if you were given the
list of variables which the problem depends on, you could probably
reproduce Rayleigh’s answers quite readily. However, knowing which
parameters matter and which don’t is not always a simple task. If you
tried to replicate Rayleigh’s results from scratch, you would probably
have difficulty getting most of the results. You really need to understand
the problem before proceeding with dimensional analysis and the Pi
Theorem.
There is a textbook on Dimensional Analysis written by P.W. Bridg-
man in 1922 who makes this exact point (Bridgman (1922)). In Bridg-
man’s book he discusses an exchange that Rayleigh had with another
researcher in the journal Nature about Rayleigh’s original paper. D. Ri-
abouchinsky questions one of Rayleigh’s problem’s on heat conduction
arguing that temperature is not an independent dimension since tem-

perature is related to molecular kinetic energy. Rayleigh’s reply states
that “it would be a paradox if further knowledge of the nature of heat
afforded us by molecular theory put us in a worse position than before.”
Pointing to this full exchange, Bridgman says:
The reply of Lord Rayleigh is, I think, likely to leave us a little cold. Of
course we do not question the ability of Lord Rayleigh to obtain the correct
result by the use of dimensional analysis, but must we have the experience
and physical intuition of Lord Rayleigh to obtain the correct result also?
For the students, except in the simplest of problems, dimensional

analysis via the Pi Theorem, with no additional physical insight and
no model from which to extract the right parameters can lead to little
insight into the problem and not always a simplification. It usually takes
several attempts to understand even a simple problem. As Bridgman
says
The problem cannot be solved by the philosopher in his armchair... only by
someone at some time soiling his hands with direct contact.
Thus, when starting to learn dimensional analysis it is important to

realize that you will rarely be able to get the right result. You are
not Lord Rayleigh. Applying the technique is simple, getting the right
parameters is difficult.
As we will soon see, in many problems we know the equations and
through a proper formulation of the problem in dimensionless form,
we can often extract the right parameters without actually solving the
problem. Knowing the equations and being able to formulate a problem
mathematically can help us identify the parameters and dimensionless
groupings that matter. This approach is the one that will dominate this
book.
2.2.4 Drag on a sphere

A simple problem in fluid mechanics is the question, given a sphere of
diameter, D, moving at speed, U , through a fluid, what is the drag force,
F ? It seems reasonable to assume that the flow will depend upon some
fluid properties. We haven’t talked about fluid properties much, but we
introduced the two key properties in the opening chapter. Viscosity has
the symbol µ and is expressed in units of M/(LT ). Clearly it will be
harder for the sphere to move through the more viscous fluid all other
things being equal. It also seems reasonable to suspect that the fluids
density, ρ, might matter as well. These are the only two fluid properties
that really matter in this problem. We are starting to already see the
comments of Bridgman come in. We have to know something about the
physics of the problem to know that these are the only fluid properties
that matter. Why doesn’t the speed of sound in the fluid matter? Or its
surface tension? In later chapters as we develop the full theory of fluid
flow, we will see that these are the only two properties that emerge for
this problem. For now you will have to accept what I am suggesting.
Applying the Pi Theorem helps our problem immensely. There are
5 parameters, F , D, U , µ, and ρ. There are the 3 dimensions of M ,
L, and T . The Pi Theorem says there are only two parameters that
matter. We can proceed by constructing the table and writing all the
parameters and their units.
F [M L
T2 ] D [L] U [ TL ] ρ [ LM3 ] M
µ [ LT ]
Immediately we see some arbitrary choices to make. If we try to

eliminate mass, we can divide the drag force by the density or the vis-
cosity. According to the Pi Theorem, any choice is just as good. Which
choice is “better” is a matter of experience and even an experienced
person might try several combinations before deciding which choice
they prefer. We will proceed to apply the theorem making arbitrary
choices about how to remove the dimensions. (Of course the choices
are not arbitrary, they are ones that I selected.) Here is one possibility
in the table below. Note that with each line of the table I remove one
parameter and one of the independent dimensions.
F [M L
T2 ] D [L] U [ TL ] ρ [ LM3 ] M
µ [ LT ]
4 µ 2
F L
ρ [T2 ] D [L] U [ TL ] ρ [ LT ] Remove M
F 2 ρU 1
ρU 2 [L ] D [L] µ [L] Remove T
F ρU D
ρU 2 D 2 µ Remove L
The result says there are two parameters. One parameter is ubiqui-
tous in fluid dynamics and we will discuss it extensively in this course.
It is called the Reynolds number, Re = ρUµD . If you have the chance
to make a grouping that looks like (density·velocity·length)/viscosity,
Figure 2.3 Drag coefficient as a function of Reynolds number for

flow around a sphere.
you should do so. The other parameter is the dimensionless drag force,
known as the drag coefficient. Typically, the drag coefficient is defined
by convention to be Cd = 1 ρUF2 πR2 ; which is the same as we got from
2
the table only with a few different arbitrary constants; namely a factor
of 1/2 and πR2 instead of D2 . There are good reasons that we will
come to later in the course for using the constants in the classic form
of the drag coefficient. The dimensional analysis result is powerful as
the result says there is a single master curve, namely
Cd = f (Re).
This single master curve, the function f , captures the drag behavior of
all spheres in all fluids. It turns out the details of this curve can be quite
complicated and very difficult to calculate from the basic equations of
fluid dynamics. However, we can conduct experiments. The dimensional
analysis says that if we conduct the experiment of measuring drag as a
function of speed once - for one size sphere in a wind tunnel - we now
know the result for any other sphere in any fluid. A fit to experimental
data on a logarithmic scale is shown as the solid curve in Figure 3.1.
I want to emphasize that dimensional analysis let’s us represent data
that depends on multiple parameters in a simpler way. In dimensional
terms the drag force is given as F (U, D, µ, ρ). Parameterizing this func-
tion experimentally for all four variables requires a lot of experiments.
Lets say we would want 10 experimental data points to establish a
reasonable empirical functional relationship of one variable. A single

curve on a graph could display force as a function of velocity for a
given diameter and given fluid. We could vary sphere diameter to show
a family of 10 curves. One graph on one page could summarize 100
experiments. But now we need to vary viscosity. We are now up to
1000 experiments and 10 pages of graphs. Varying the density we need
10,000 experiments and 100 pages (a book). In dimensionless terms,
there is a universal master curve of one parameter and that curve is
good for all drag experiments with a sphere that have ever been done
and all that will be done in the future. The master curve works for all
fluids and spheres of any size. We will elaborate on this example in the
next chapter.
2.2.5 Pipe flow

Another classic example of a practical use of dimensional analysis is the
problem of flow in a pipe. A simple question to ask is given a length
of pipe `, with a constant diameter D, with a pressure difference ∆P
applied across it, what is the mean flow velocity U ? Here, we already
have four parameters. Clearly the fluid matters as well so as with the
previous problem we will assume that the fluids density and viscosity
matter as well.
Before the theory of fluid dynamics was firmly established and the
power of dimensional analysis understood, a problem emerged from this
seemingly simple example. It is clear that understanding the pressure-
flow relationship in a pipe is of great practical importance. While it
seems straight forward to establish an experiment to measure flow as
a function of applied pressure, there are a lot of parameters to vary in
order to obtain a complete model experimentally. Before dimensional
analysis was really understood and used, this extensive experimentation
is precisely what was done to describe this simple pipe flow problem.
There used to be page after page of numerous graphs showing pressure-
flow relationships for different pipes and different fluids.
Applying the Pi Theorem helps our problem. There are 6 parameters,
`, D, ∆P , U , µ, and ρ. There are the 3 dimensions of M , L, and T .
The Pi Theorem says there are only three parameters that matter. This
helps quite a bit. We can proceed by constructing the table and writing
all the parameters and their units.
U [ TL ] ` [L] D [L] M
∆P [ LT 2] ρ [ LM3 ] M
µ [ LT ]
As before, there are some arbitrary choices to make about how to
proceed. If we try to eliminate mass, we already see that we could
divide the pressure drop by either the density or viscosity. While any
choice is just as good as the other, one possibility is worked out below,
U [ TL ] ` [L] D [L] M
∆P [ LT 2] ρ [ LM3 ] µ M
[ LT ]
2 ρ
U [ TL ] ` [L] D [L] ∆P
ρ
[ TL2 ] µ
[ LT2 ] Remove M
ρU
µ
[ L1 ] ` [L] D [L] ∆P
ρ2 U 2
Remove T
ρU D ` ∆P
µ D ρU 2
Remove L
This result is one of many possibilities. My result says there are three
parameters. The first is again the Reynolds number - it will show up
in nearly every fluid mechanics problem. The second parameter is a
geometric parameter, the length to diameter ratio. The final parameter
is the dimensionless pressure drop. Now, conducting experiments for
three parameters is a lot easier than 6. We could fit all the results on a
single graph plotting the dimensionless pressure drop versus Reynolds
number for a family of curves for different `/D. We need a sheet of
paper rather than a book to display all results.
We can actually do even better, but the comment by Bridgman again
applies. We need a little physical reasoning and some knowledge of the
problem. This step might seem a leap now, however, as we learn more
throughout this course we will make this step more easily. The pressure
drop and pipe length are not really independent. It is the parameter
∆P/` that matters most, the pressure gradient. The assumption we
make here is that when the pipe is long that any section of pipe is
basically the same. If we look at the flow here or there in the pipe, it
is just a length of pipe and the flow doesn’t change. If the character
of the flow doesn’t change as a function of length, then doubling the
length should not change anything if we double the pressure drop. The
pressure drop per unit length is the parameter that really matters.
We could confirm this assumption experimentally by measuring the
pressure as a function of length and we would find that it is essentially
linear with distance. This assumption that pressure gradient matters
really only applies to pipes that are long relative to their diameters.
If we use the pressure gradient as our parameter our result is sim-
a) b)
Figure 2.4 Dimensional (a) and dimensionless (b) plots of the flow-
pressure drop data for the pipe flow problem. This data was taken
in a simple apparatus and consisted of 21 separate configurations
varying fluids (water and oil), pipe diameter, and pipe length. The
experimental dimensional data in (a) is collapsed to a single master
curve in (b). I took this data with a simple table-top device thus
the data at high Reynolds number show some significant scatter
that can be reduced with some care.
plified as the number of variables is reduced to 5 and the Pi Theorem

states we will only have two parameters that matter. Proceeding in the
same way to remove dimensions, we could obtain.
U [ TL ] D [L] ∆P M
` [ L2 T 2 ] ρ [ LM3 ] M
µ [ LT ]
ρ
U [ TL ] D [L] ∆P L
`ρ [ T 2 ] µ [ LT2 ] Remove M
ρU 1 ∆P 1
µ [L] D [L] `ρU 2 [ L ] Remove T
ρU D ∆P D
µ `ρU 2 Remove L
This result is powerful as this is a single master curve. This master

curve captures the pressure flow relationship for all pipe sizes, all fluids,
and all pressure drops. This is a very powerful result. The punchline
of the work is shown in Figure 2.4 where we show the dimensional
and dimensionless pressure-flow relationship for pipe flow. I took this
data myself in a simple apparatus. There are 21 different experiments
in these plots. The pipe diameter and length were varied and data
was taken with both water and oil to vary the fluid properties. In
dimensional form, the data spans 3 orders of magnitude in velocity. In
dimensionless form, all the data collapse to a single master curve. All
21 experiments are described by one curve.
An interesting feature of the data is the kink in the slope. This feature
Figure 2.5 Experimental images on pipe flow at different Reynolds

number. In the upper image the flow is laminar and injected dye just
goes right down the center in a straight line. At higher Reynolds
number the flow begins to fluctuate and become unstable. These
images were taken in Orbourne Reynold’s original experimental ap-
paratus but a century later. In Reynold’s original paper he included
hand sketches of the flow field. From Album of Fluid Motion.
of the curve was first described by Osbourne Reynolds in one of the

most well-known papers in all of fluid dynamics (Reynolds (1883)). At
low flow and low pressure, the relationship between pressure and flow
is linear. This linear regime is analogous to Ohms law for a resistor,
V = iR, where voltage drop is replaced by pressure drop and current is
replaced by flow. At higher flow rates (pressures) something happens.
We will discuss this transition in more detail later, but it represents a
transition from a smooth laminar flow to a turbulent one.
Visualization of this instability is shown in Figure 2.5 at three Reynolds
number from low to high. In the laminar flow case, the relationship be-
tween pressure and flow is easily described by a theory which we will
discuss later. In the turbulent case, at high flow, there is no theory
that can predict the behavior and even today’s fastest computers are
limited in calculating turbulent flows. Because the turbulent case is dif-
ficult to describe theoretically, the collapse of the experimental data is
important. It means from a limited set of experiments we could predict
turbulent flow in any pipe with any fluid. It is also important to real-
ize that he images seen in Figure 2.5 only depend upon the Reynolds
number. If we repeated the experiment in water, air, and oil - but we
changed the size of the pipe to keep the Reynolds number the same
across the three experiments - we would make the same visualizations.
2.2.6 Drag on a boat

Let’s consider the problem of drag on a boat. If my boat were a subma-
rine, then the analysis we conducted for the sphere immersed in a fluid
would be all we would need. That analysis said the only dimension-
less parameter that mattered was the Reynolds number. That result
means that if I build a model submarine of the same geometrical pro-
portions as the actual one and place it in an experiment that matches
the Reynolds number, then my model experiment is an exact replica of
the real thing. My model could be in air instead of water, as long as
the Reynolds number matched.
With a boat that sits on the surface of the water, the situation is a
little different. Not only is there drag due to the hull moving through the
fluid (like drag on the submarine), but there is clearly energy radiated
away in the waves of the wake. The waves carry energy and therefore
must be taking away energy that would otherwise be put toward moving
the object. For this problem we need to determine which parameters
matter, again unleashing our inner Lord Rayleigh.
Let’s consider a case where our model will have the same proportions
as the real ship and we will weigh down the model such that it sits at
the same depth as the real ship. Under these restrictions, the geometry
is captured by the single length of the boat, `, since all other length
scales would change in proportion. We would expect the boat’s drag to
at least depend upon the speed, U , fluid density, ρ, and fluid viscosity
µ. We could hypothesize that the air and it’s properties don’t matter.
Since the air is 1000 times less dense than the air, it seems reasonable to
neglect the influence of the air. One parameter we can think that might
play a role in the making of waves is the acceleration due to gravity, g.
In the limit of no gravity, the waves would be really weird - so clearly
g should play some role. The final parameter we might consider is the
surface tension of water, σ. While we could include this in our analysis,
once we knew more about surface tension we would find it is a very
small force for large objects on the size of boats with people in them.
So our physical intuition (that we are just building) says that g is the
only new parameter.
The drag force, F , is a dimensional function, F (`, U, ρ, µ, g). The Pi
theorem has 6 parameters expressed in 3 independent dimensions. We
expect one new dimensionless parameter over what we would find for
the submarine problem. Let’s proceed with the table and see what we
get,
F [M L
T2 ] ` [L] U [ TL ] ρ [ LM3 ] M
µ [ LT ] g [ TL2 ]
F L4 µ 2
ρ [T2 ] ` [L] U [ TL ] ρ [ LT ] g [ TL2 ] Remove M

F ρU 1 g
2
ρU 2 [L ] ` [L] µ [L] U2 [ L1 ] Remove T
F ρU ` g`
ρU 2 `2 µ U2 Remove L
Again, the exact choices at each step above are arbitrary and I could
have come up with a number of equivalent solutions. I took a path that
leads toward the standard way the result is presented. The new di-
mensionless parameter, is called the Froude number, and is historically
defined as
U
Fr = √ .
g`
The dimensionless drag force can be expressed as,

F ρU ` U
= f , √ = f (Re, Fr)
ρU 2 `2 µ g`
This result has an important implication for our idea of using model
experiments for testing drag on boats. We can’t change the dimension-
less parameters to arbitrary values in practice. We are pretty much
stuck with g, at least until we can colonize other planets. We also can-
not change ρ by much - most liquids tend to have similar densities as
water. We could change the density by 20% by adding salt to water.
We could take an extreme case and use liquid mercury. However, we
are limited in what choices we can make. We also can easily increase
µ, but decreasing it from a value much less than that of water becomes
challenging. It is not practical to be able to set the Reynolds number
and Froude number completely independently.
The things we can vary with some degree of control are the veloc-
ity and the length. The relationship between velocity and length are
different in the two dimensionless groups. If I shrink the boat, to keep

the Reynolds number the same I would need to increase the velocity in
proportion. On the other hand, to keep the Froude number the same
for a small boat model, I would need to decrease the velocity. So even
though the dimensional analysis has in theory made our lives simpler
by reducing our number of parameters from 6 to 3, we actually still
have some challenges ahead.
2.3 Making equations dimensionless

In many cases (in fact, most of them in this course), we can actually
write down in reasonably compact form the mathematical description
of the problem of interest. As we will learn, there are often many dif-
ficulties with mathematically solving the problem. In some cases it is
hard, in others it is impossible. Recall in the introduction that if you
could solve the basic equations of fluid dynamics, the Navier-Stokes
equations, in general form you would win a million dollars.
However, in cases where we know the equations but cannot solve
them, making the model equations dimensionless can help guide us to
pick the ”right” dimensionless groups. This equation based approach
can be equivalent to working the units through the table approach
but is a little bit different in methodology. The table approach will
always lead to the same result as the working through the equations.
When we have the equations at hand, we can ”see” the parameters that
matter and will often see where groupings of parameters only appear
together in certain ways, which is something that the table approach
cannot tell us. Working through the equations can remove a little of
the guess work as to what parameters should matter. However, working
with the equations does not fully absolve us from the responsibility of
knowing the physics of the problem - how do we know what equations
are appropriate?
The typical procedure is to define a new set of variables that have
no units by scaling the usual dimensional variables by a constant. This
sounds crazy but is no different than changing units. In an equation
if we have a variable time, t, we could express that time in seconds,
minutes, hours, or years. Similarly we can define a dimensionless time,
t̃ = t/t0 , where both t and t0 are measured in seconds. The constant t0
is just some constant that we pick to measure our unit of time in. There
is nothing special about the second or the minute, so we can change the
unit to whatever we feel like. Picking the ”right” scale for t0 can be a
bit of an art form that comes with experience. However, the equations
will serve as out guide us to a ”good” answer. Mathematically, t0 could
be anything.
Lets do a few examples to make this clear.
2.3.1 Pendulum
Since we started the chapter with the pendulum example, let’s repeat
this example using the equation approach. Newton’s law, F = ma, for
a simple pendulum with no friction is,
d2 θ
m` = mg sin(θ),
dt2
which is easily rewritten as
d2 θ g
2
= sin(θ).
dt `
We have already recovered the basic result of the Pi Theorem. We have
shown that the mass doesn’t matter and that the only parameter that
matters is g/`. However, the equation still has units so let’s proceed to
remove them all. The angle, θ, has no units so we leave it alone. Time,
t, has units so lets make the transformation t̃ = t/t0 (or equivalently
t = t̃t0 ).
d2 θ g
2
= sin(θ).
d(t̃ t0) `
Note how we made the derivative dimensionless. This step seems strange
to most students the first time they see it but we have to remember
that a second derivative in time has units of 1/seconds2 . Thus, the dt2
in the derivative is replaced by d(t̃ t0)2 . All we did was the substitution
t = t̃ t0 . Since t0 is a constant it can be moved in or out of the derivative
operator. This step is no different than changing the time derivative to
be measured in minutes instead of seconds.
Our equation can be rewritten as
d2 θ gt2
2
= 0 sin(θ).
dt̃ l
which is a dimensionless equation with one parameter. However, we

have left t0 unspecified. I am allowed to make it anything I want. It
seems convenient then to define t20 = `/g and our equation then be-
comes,
d2 θ
= sin(θ). (2.1)
dt̃2
This final equation has no units and no free parameters.
The solution to the equation requires two initial conditions for the
position and the velocity. Expressed mathematically our initial condi-
tions are
dθ V
θ(t = 0) = θ0 and =
dt t=0
l
where V is the velocity in units of L/T . We now must make these initial
conditions dimensionless in the same way as the equations. Under the
change of variables we obtain,

dθ V
θ(t = 0) = θ0 and =√ (2.2)
dt̃ t̃=0 gl
Thus our general problem comprises Equation 2.1 and the initial con-
dition 2.2. The final equation has no parameters and there are two pa-
rameters that enter into the initial condition. All pendulums (started
from rest and
p from the same angle) are the same when time is plotted
in units of `/g instead of seconds. The scaling for time is exactly as
we obtained from the Pi Theorem. We have no new information from
this approach than we got from the table.
2.3.2 Mass-spring-damper
As an example where we have additional parameters, consider a simple
mass-spring-damper system,
d2 x dx
m = −β − kx
dt2 dt
with the initial conditions that x(t = 0) = x0 and the velocity initially
is zero.
We will make the equations dimensionless by defining x̃ = x/x0 and
t̃ = t/t0 . Here, t0 is arbitrary, but x0 is the initial position. Using
the initial position seems like a good choice to make x dimensionless
because the solution for x will always be bounded in magnitude by 1.

Applying the change of variables we obtain
d2 (x̃ x0 ) d(x̃ x0 )
m = −β − kx0 x̃
d(t̃ t0 )2 d(t̃ t0 )
Be careful to note the units of d2 x/dt2 . Its units are length/time2 . Thus
t0 is squared whereas x0 is not. The placement of the squares when we
write d2 x/dt2 are literal; t is squared and x is not. Since x0 appears in
each term it cancels out and we can group t0 together to obtain,
d2 x̃ βt0 dx̃ kt20
2
=− − x̃.
dt̃ m dt̃ m
Since t0 is arbitrary we select t20 = m/k to obtain,
d2 x̃ β dx̃
2
= −√ − x̃. (2.3)
dt̃ mk dt̃
with the initial conditions that x(t = 0) = 1 and dx/dt = 0. Thus,
the all mass-spring-damper problems reduce to an equation and initial
condition with one free parameter, the dimensionless damping constant.
In the absence of damping, all mass-spring systems are reduced to a
single problem.
Let’s compare our result to what we would obtain by the Pi Theorem,
proceeding to remove units using our table method.
x [L] t [T ] x0 [L] k [ TM2 ] β [M
T ] m [M ]
k 1 β 1
x [L] t [T ] x0 [L] m [T2 ] m [T ] Remove M
x k 1 β 1
x0 t [T ] m [T2 ] m [T ] Remove L
q
x
t m k √β Remove T
x0 km
The result is consistent with what we obtained from making the

equations dimensionless. We can write the result from the table as
r !
x k β
=f t ,√ .
x0 m km
2.3.3 Derivatives have units too

An important point that came out in the previous two examples is that
derivatives have units too. Whenever we define a dimensionless time
2.4 Summary 29
variable based on some ”natural” time scale such as t̃ = t/t0 , then we

need to remember that the dimensionless time derivative will become,
∂ 1 ∂
= .
∂t to ∂ t̃
Whenever we define a dimensionless space variable based on some ge-
ometric length scale such as x̃ = x/L, then we need to remember that
the dimensionless spatial derivatives will become,
∂ 1 ∂
= ,
∂x L ∂ x̃
or
∂2 1 ∂2
= .
∂x2 L2 ∂ x̃2
It is also important to remember to apply this result to our vector
operators which will begin to appear soon. We will work with these
ideas in context in the coming chapters, but I want to highlight it now
so you don’t forget.
2.4 Summary
Throughout this book we will continue to return to the idea of work-
ing in dimensionless terms. We will periodically appeal to both the
table and equation based approach - realizing that we can get the same
results from both approaches. We will see many examples where the
mathematical description of a problem can be stated, but it is too dif-
ficult (or impossible) to solve the problem. In these cases we can still
obtain insight to the problem by appealing to dimensional analysis.
3
The empirical approach
In the last chapter we found that by using physical insight and know-
ing the important parameters in a problem we could use dimensional
analysis to dramatically reduce the number of parameters needed to
characterize a problem. We saw through the example of drag on a
sphere and flow in a pipe that the complexity of a problem in terms of
how many experiments need to be done could be dramatically reduced.
All the behavior is described on a single plot or with a single number.
After this current chapter, we will spend much of the rest of the
book developing the underlying theory for fluid flow. We will find that
this theory can be extremely accurate when compared to experiments
and has great predictive power. We will also find that in many cases
of practical interest even with great computational resources and so-
phisticated mathematical tools, many problems are just too difficult
to solve. Fortunately, many of these cases are pretty easy to measure
experimentally. We often find that dimensional analysis coupled with
experimental data can lead to practical solutions and accurate esti-
mates to real world problems. Many times there is no real need for a
time consuming or computationally expensive solution.
The focus of this chapter will be to show how dimensional analysis
plus experimental data can be put to practical use in an engineering
setting. We can think of this approach as purely empirical, While we will
only go through a few examples, this approach in a sense could be the
end of our practical study. We might not have a lot of physical insight or
deep understanding - but we could perform simple calculations, make
predictions, and design useful things.
While all the interesting stuff (to me, anyway) will come up in later
32 The empirical approach
chapters, this chapter may be the most useful when you just need to
solve problems. We will revisit two examples from the last chapter on
drag and pipe flow and explore them in more depth. The chapter here
is meant to be illustrative and not comprehensive. Many textbooks
on engineering fluid dynamics and have numerous details and practical
situations that have been studied. My aim here is to show this empirical
approach first, and then dig into the details a little later in the book.
3.1 Drag revisited

In the last chapter we described a simple problem (well, at least simple
to state); given a sphere of diameter, D, moving at speed, U , through
a fluid, what is the drag force, F ? We stated that the behavior will
depend upon two fluid properties - density and viscosity. There are
thus 5 parameters, F , D, U , µ, and ρ expressed in the 3 independent
dimensions of M , L, and T and thus the Pi Theorem says there are
only two dimensionless parameters. The table with all the parameters
and their units was found to be
F [M L
T2 ] D [L] U [ TL ] ρ [ LM3 ] M
µ [ LT ]
F L4 µ 2
ρ [T2 ] D [L] U [ TL ] ρ [ LT ] Remove M

F 2 ρU 1
ρU 2 [L ] D [L] µ [L] Remove T
F ρU D
ρU 2 D 2 µ Remove L
Defining the drag coefficient as Cd = 1 ρUF2 πR2 and the Reynolds

2
number as Re = ρU D/µ, dimensional analysis says there is a single
master curve, namely
Cd = f (Re).
This single master curve, the function f , is shown again in 3.1.
What is interesting is that we can find two simple limits that work
over a wide range of Reynolds number in this data. At values of Reynolds
24
number less than 1, the data is fit by a function Cd = Re and for the
3 5
range of 10 < Re < 10 the data is fit by a constant Cd ≈ 0.4. Between
1 < Re < 103 we see the data smoothly transitions from one limit to
the other. Above Re > 105 a strange dip in the drag coefficient happens
that we will discuss later. We can actually derive the two limits using
3.1 Drag revisited 33

flow around a sphere is shown as the solid curve. The dashed lines
are for the fit Cd = 24/Re and a constant of Cd = 0.4.
dimensional analysis. Low values of Reynolds numbers occur for small

objects, moving slowly, through a very viscous medium (i.e. look how
the terms D, U , and µ show up). These are flows that are dominated
by viscosity. The fact that the fluid has mass or inertia turns out to
be irrelevant. It takes much more effort to drag the fluid along at low
Reynolds number than it takes to push fluid out of the way. If fluid
inertia is irrelevant to the physics, then we can remove density from
our list of parameters and rework the dimensional analysis.
F [M L
T2 ] D [L] U [ TL ] M
µ [ LT ]
F L2
µ [ T ] D [L] U [ TL ] Remove M
F
µU [L] D [L] Remove T
F
µU D Remove L
Our result now says there is one dimensionless parameter, therefore

it doesn’t depend on anything and must be a constant,
F
= Constant.
µU D
Recasting in terms of the drag coefficient we can multiply both sides
of the equation by the inverse Reynolds number,
F µ F µ
= 2 2
= Constant
µU D ρU D ρU D ρU D
which states that the drag coefficient will be equal to a constant divided
24
by the Reynolds number. The data supports the function of Cd = Re .
The factor of 24 comes from a full calculation of the laws of fluid flow
and is also validated by experiments. Constants simply cannot come
from dimensional analysis.
The result I just “derived” probably seems suspicious to you. If it
doesn’t, it should. Why am I allowed to throw out density from my list
of parameters? My derivation again relies on this comment by Bridg-
man. You have to understand the physics of the problem to successfully
apply dimensional analysis. Since this is just the start of your study,
you should have no reason to accept that density is unimportant at low
Reynolds number. Remember that I have studied fluid mechanics since
graduate school, have read about this problem in numerous books, and
have run the experiment and plotted data myself. This problem is easy
for me at this point; however, faced with a new unfamiliar problem I
have to work hard to get the right answer. With a new problem, I would
also not typically believe an answer until is was backed up by a more
thorough analysis or experimental data. It will take some time until
we study fluid motion in more detail for this example to make physical
sense to you. However, the excellent fit to the experimental data at low
Reynolds number should convince you that we are on to something.
If we continued the graph in Figure 3.1 to lower and lower Reynolds
24
number we would see the fit to Cd = Re continue to high precision.
The other limit at higher values of Reynolds number where the drag
coefficient is approximately constant can also be “derived” from dimen-
sional analysis. In this limit the fact that the fluid has viscosity matters
less than the fact that the fluid has inertia. At high speeds we have to
move the fluid, which has mass, out of the way for the ball to go by.
This explanation is actually a little too simplistic and the truth is more
subtle, but let’s just assume for now that the viscosity doesn’t matter
and see what the analysis says.
F [M L
T2 ] D [L] U [ TL ] ρ [ LM3 ]
F L4
ρ [T2 ] D [L] U [ TL ] Remove M
F 2
ρU 2 [L ] D [L] Remove T
F
ρU 2 D 2 Remove L
Our result now says there is one dimensionless parameter and it must
be a constant,
F
= Constant.
ρU 2 D2
We see the above equation is the definition of drag coefficient. The
analysis says that drag coefficient is constant if the viscosity of the
fluid is irrelevant. Again, I should not have convinced you that it is
acceptable to ignore the fluid’s viscosity, however the data between a
Reynolds number of 103 and 105 supports this view. If we look at the
data we see a dip in the drag coefficient that we will later explain - but
it turns out the dip occurs because the fluid has viscosity.
In dimensionless terms, there is a universal master curve of one pa-
rameter (the Reynolds number) and that curve is good for all drag
experiments with a sphere that have ever been done and all that will
be done in the future. The master curve works for all fluids and spheres
of any size. However, we demonstrate here that if we understand even
a little something about the physics of our problem, we can often find
even simpler limits that can be good over a wide range of the parameter
space.
That the behavior changes dramatically with Reynolds number can
be seen through the flow visualization images in Figure 3.2. At low
Reynolds number (here Re=1.5 or lower) the flow looks very symmetric
from left to right. In this regime the flow is dominated be viscosity. At
Re=26, we see asymmetry and a recirculating wake forming behind
the cylinder. This is the effect of inertia - the fluid approaching the
cylinder has momentum that causes the flow to tend to “overshoot” on
the backside of the cylinder. As the Reynolds number increases to 2000
and 10,000 the wake becomes very complex and has no resemblance to
the flow at low Reynolds number. It is also interesting to realize that
the flow field is only dependent on Reynolds number. The experiments
in Figure 3.2 could be repeated in water, air, or oil and as long as the
Reynolds number was matched the flow would look identical.
3.2 Drag coefficients

In summary we have the following two simple formulas for the drag
force on a sphere. If the flow is dominated by viscosity then dimensional
a) b)
c) d)
Figure 3.2 Flow around a cylinder at different Reynolds numbers

(From Album of Fluid Motion). The Reynolds number is 1.5, 26,
2000, and 10,000 going from figures a to d.
analysis gave a force on a sphere to have a form
F = CµU D,
where C is constant. If the flow was dominated by inertia the force had
a form
F = CρU 2 D2 .
In both cases, the constant C (or Drag Coefficient) is unknown un-

til one conducts additional analysis, numerical simulations, or exper-
iments. Fortunately, for many common shapes drag coefficients have
been measured and can be found in books or the internet. When us-
ing drag coefficient data you must take care in noting the form of the
formula, and what area or length scale is used.
Most (but not all) drag formulas are written assuming inertia domi-
nated flow. The reason is that most ”human scale” problems are in the
inertia dominated regime. The transition between the inertia dominated
flows and viscosity dominated flows is determined by the Reynolds num-
ber, defined as
ρU D
Re = .
µ
The Reynolds number is usually large for things that you interact with
on a daily basis - a 1 mm object traveling at 1 mm/s in water will
have a Reynolds number of 1. What counts as high or low Reynolds
number depends on the situation, but usually greater than 1000 would
be considered well in the high Reynolds number regime. The exact
number really depends on the situation.
The most common way that the drag coefficient is characterized is,
1
F = Cd ρU 2 A,
2
where A is the area projected by the object to the flow. It is important
to note that this choice of area used in the formula is arbitrary. The
area used must be reported along with the drag coefficient to avoid any
ambiguity. If the drag coefficient is measured but the area that was
used to compute Cd is not reported, then the result is useless.
For many objects, the drag coefficient would be reported as a con-
stant. In some cases it will be given as a chart as a function of Reynolds
number. If the number is given as a constant, then most likely it is as-
sumed that the Reynolds number is large. Note that if it is given as
a constant in a table, it doesn’t mean that it is truly constant with
Reynolds number as seen in the sphere example. However, looking at
Figure 3.1, note how the assumption of constant drag coefficient is
pretty good over a wide range of conditions.
With a simple search will can find a lot of resources and tables of drag
coefficients. You will notice that most are around 1 for blunt objects
(sometimes around 2, but never 10) and as low as 0.04 for relatively
streamlined shapes. We will dig into the physics of this problem again
in a later chapter. For now, if you need to figure out the drag force on
an object it is pretty simple (at least conceptually),
• Compute (or estimate) the Reynolds number.
• Search for drag coefficients on the shape of interest.
• Understand what area was used on the reported drag coefficient.
Without the area defined, the coefficient is useless.
• If the coefficient is given as a function of Reynolds number, you can
just look get the coefficient for your Reynolds number.
• If a single number for drag coefficients is reported, it is most common

for that number to assume high Reynolds number. Be careful that
you are using the right Reynolds number range.
3.3 Pipe flow

Another useful example of dimensional analysis is the problem of flow
in a pipe which we introduced in the last chapter. Given a length of pipe
`, with a constant diameter D, with a pressure difference ∆P applied
across it, what is the mean flow velocity U ? This problem was discussed
previously and the dimensionless table is repeated below.
U [ TL ] D [L] ∆P M
` [ L2 T 2 ] ρ [ LM3 ] M
µ [ LT ]
ρ
U [ TL ] D [L] ∆P L
`ρ [ T 2 ] µ [ LT2 ] Remove M
ρU 1 ∆P 1
µ [L] D [L] `ρU 2 [ L ] Remove T
ρU D ∆P D
µ `ρU 2 Remove L
The final result is often written in the following way,
1 2 `
∆P = ρU f (Re)
2 D
where f is called the friction factor and is a function of Reynolds num-
ber only f (Re). As with drag coefficients, the friction factor must come
from experiments or theory. Note that the Reynolds number is defined
using the tube diameter as the appropriate length scale and uses the
average fluid velocity, U . Experimental friction factor data for a range
of Reynolds numbers is shown in Figure 3.3. Earlier I presented my
own data, but here I am taking the data from the professionals that
is very high quality. Note that the Reynolds number spans 6 orders of
magnitude in Reynolds number.
When the Reynolds number is less than 2300 (this is the most com-
monly reported transition number), the flow is smooth and laminar and
the friction factor can be found analytically,
64
f= for Re < 2300.
Re
We will compute this solution later in the course. When the Reynolds
number is larger than 2300 the flow becomes turbulent and f is found
3.3 Pipe flow 39
Figure 3.3 Experimental data for friction factor versus Reynolds

numbers. Plot is from from McKeon et al. Journal of Fluid Me-
chanics, 2004. The transition from laminar to turbulent flow at
Re ≈ 2000 is very clearly seen as the jump.
by fitting experimental measurements. There is no analytical solution

for the turbulent regime - however the overall friction factor is easy to
characterize experimentally. It is important to note that the transition
around Re ≈ 2300 is only seen this cleanly in a careful experiment. In
a practical application or more careless experiment (like my own data
that I presented in the last chapter) the transition would be seen at
a lower Reynolds number and may not be as neatly repeatable as in
Figure 3.3.
For turbulent flows, f is also experimentally determined to depend
upon roughness of the pipe. In Figure Figure 3.3 the data were taken
in a smooth pipe. However, in many practical situations the pipe is
not perfectly smooth and the roughness of the pipe is critical. The
dimensionless term is the relative roughness, /D, the size of the bumps
on the pipe wall relative to the diameter of the pipe. So the general
formula from dimensional analysis is
1 2 `
∆P = ρU f Re,
2 D D
Figure 3.4 Moody diagram for the friction factor in pipe flow. This
figure is a reproduction of the way the data are classically presented.
The diagram in it’s classical presentation is reproduced in Figure 3.4

and is called the Moody diagram.
The form of the result is essentially identical to the case of drag on
an object. The Reynolds number is the key parameter and the rest of
the complexity of fluid flow is wrapped up in the single parameter, f .
When the viscosity dominates at low Reynolds number, just as in the
drag problem, we have the friction factor varying as f = 64/Re. Note
that the form of this result can be obtained from dimensional analysis
(just as in the drag problem) though the factor of 64 can only come
from analysis.
3.3.1 Calculations with the Moody diagram

In order to solve practical problems, we can simply use the Moody
diagram as a tool along with the expression,
1 2 `
∆P = ρU f (Re) .
2 D
3.3 Pipe flow 41
If the velocity of the flow (or volumetric flow rate) is known, then the
solution procedure for the needed pressure to drive the flow is straight-
forward. The procedure is,
• For the given pipe size and fluid, compute the Reynolds number.
• Decide on a good estimate of the pipe roughness.
• Look up f (Re) on the diagram.
• Compute the pressure drop.
If the applied pressure is known, then the solution procedure for the
flow velocity is not as straightforward. The procedure is,
• Guess a Reynolds number (you can’t compute it since you don’t know
the velocity).
• Decide on a good estimate of the pipe roughness.
• Look up f (Re) on the Moody diagram.
• Compute the flow velocity as for the known pressure case.
• Compute the Reynolds number for this new flow velocity.
• Look up f (Re) on the Moody diagram.
• Compute the flow velocity again.
• Repeat (if needed) until you get a converged solution.
If at anytime in either procedure (or based on your intuition or

knowledge of the problem) you suspect that the flow is laminar with
Re < 2300, then the Moody chart is not needed. For laminar flow we
will later derive an analytical solution f = 64/Re, but for now you
could consider this to just as equally come from experimental data. For
laminar flow we have the simplified relation,
1 2 64 `
∆P = ρU .
2 Re D
or
32µ`
∆P = U.
D2
For laminar flow there is a simple linear relationship between pressure
and flow velocity.
3.4 Summary
These two examples show how we can combine dimensional analysis
with experimental measurements such that we can make accurate pre-
dictions. There is no satisfactory ”why” embedded in any of these re-
sults. However, the results can be used to design and build real devices.
Predictions can be made that will be really accurate, since they are
rooted in experiments. I again want to emphasize that we covered only
a few examples here and other books provide many more examples of
piping losses and drag on objects. We will revisit the physics of drag in
a later chapter.
4
Vector calculus notation and review
Starting in the next chapter we will begin to formulate the laws which
govern fluid flow. So far we have relied on observations, experiments,
and dimensional analysis to guide us. As we develop our more complete
theory, we will find laws will allow us to compute quantities such as the
temperature field which can vary both in time and space. Our laws
will be ”local” in that they can tell us quantities at a point (in the
continuum sense) in space. Inherent in our derivations of these local
laws will be volumes, surfaces, and integrals. Using the language of
vector calculus will become essential to maintain our results in a form
compact enough to provide some physical insight.
I am assuming that upon reading this that you have taken a course
in vector calculus but have likely forgotten many of the finer details.
This chapter is meant to review the notation and some of the useful
theorems. If you have never come across these topics then my coverage
here will not be sufficient. Our use of vector calculus will be as the
language we use to describe the physics. We will not need to carry
out complicated surface and volume integrals, however, we will need to
come to understand an intuitive and physical level what a surface or
volume integral is.
4.1 Scalar and vector functions

Many physical things we will want to describe are scalar functions. A
scalar function means that there is a single value at every location in
space. That value at each point can also change with time, but the
44 Vector calculus notation and review
field has no sense of direction. Two examples of scalar functions are

temperature, T , and pressure, P . When we are being explicit, we will
write scalar functions as T (x, y, z, t) to denote that temperature can
vary at every location in 3D space and evolve with time. In two dimen-
sional problems, we will most commonly visualize the scalar function
with a contour plot, where lines show values of constant temperature.
The contour plot is just like reading a topographic map; follow a con-
tour line and your elevation doesn’t change. Follow an contour line of a
temperature field and the temperature does not change along the path.
We will also want to describe quantities that have directionality to
them and these will be vector functions. We will use a boldface notation
for vector functions. The most common quantity we will discuss is a
fluid flow’s velocity field, v. If we are being explicit we will sometimes
write v(x, y, z, t). In two dimensional problems we will visualize the
velocity field with a set of arrows at many locations in space, showing
the velocity vector’s magnitude and direction. The velocity vector has
three components, thus it is important to always remember that the
vector function is comprised of three scalar functions for each compo-
nent. For velocity, we will commonly use the notation u, v and w for
the x, y, and z components respectively. We will sometimes write out
vector functions as a column vector
 
u
v =  v ,
w
or sometimes as components using unit vectors,
v = ui + vj + wk
Examples of scalar and vector fields in a 2D world are shown in Figure

4.1 in the context of weather. The contour map of atmospheric pressure
is shown at one instant in time during a classic “Nor’easter” snowstorm
that is common in New England. The low pressure contour off the coast
generates a strong counterclockwise wind that brings moisture in from
the ocean. The image of velocity vectors from Hurricane Katrina in
2005. The color image shows the magnitude of the velocity field (a
scalar field) and the vectors are superimposed on top.
4.2 Gradient, divergence, and curl 45
Figure 4.1 Examples of scalar and vector functions in weather appli-

cations. On the left are pressure contours during a classic Nor’easter
snowstorm (image from the National Weather Service archives). On
the right are velocity vectors from Hurricane Katrina in 2005 (from
Nasa’s Earth Observatory). The velocity vectors are placed on top
of an image of another scalar field, which is the magnitude of the
velocity at that point.
4.2 Gradient, divergence, and curl

There are three operations that we will use repeatedly in our study are
divergence, gradient and curl. These three operations that derivative
with respect to space. While all three have some similarities we will
think of them all as pretty distinct physical things. For all three, it is
convenient to use the notation for “del”, with the symbol ∇, written
as a column vector,
 
∂
∂x
 
∇= ∂ ,
 ∂y 
∂
∂z
or equivalently using unit vectors,
∂ ∂ ∂
∇= i+ j+ k,
∂x ∂y ∂z
where i, j, k are the unit vectors in the x, y, z direction.

When we apply ”del” to a scalar function (P in the example below)
we define the gradient which gives back a vector,

 
∂P
∂x
 
∇P =  ∂P ,
 ∂y 
∂P
∂z
or equivalently using unit vectors,
∂P ∂P ∂P
∇P = i+ j+ k.
∂x ∂y ∂z
Remember, the gradient of a scalar field gives a vector. We can also take
a gradient of a vector field, but let’s leave that issue to the side for now.
The most important property of the gradient is that the resulting vector
field always points perpendicular to the contours and points ”uphill”. If
you see closely spaced contour lines on a topographical map you know
the terrain is very steep.
When we take the dot product of ”del” with a vector field we define
the divergence which gives back a scalar field. The divergence of the
velocity vector is defined as,
∂u ∂v ∂w
∇·v = + + .
∂x ∂y ∂z
The divergence is a measure of how much a vector field is, well, diverg-
ing. If you see a point in the vector field where all the velocity vectors
were flowing out of a point, this would have positive divergence. If all
the vectors were flowing into a point, then that would have negative
divergence. If some vectors were flowing in and some out, and if they
were in perfect balance there would be zero divergence. The divergence
is shown schematically in Figure 4.2. We will see that divergence is con-
nected to the idea of conservation. In a fluid flow if there is non zero
divergence of the velocity field, then it means that mass is accumulat-
ing at that point. Non-zero divergence can only occur in a compressible
flow. Remember, the divergence of a vector field gives a scalar field. It
does not make sense to take the divergence of a scalar field.
Our final operator of interest occurs when we take the cross prod-
uct of ”del” with a vector. This operation is defined as the curl. For
4.3 Normal vectors and flow through a surface 47
v v v
v<0 v>0 v=0
Figure 4.2 Schematic of the divergence of velocity.
example, the curl of the velocity field is (known by the symbol ω),
 
∂w ∂v
−
 ∂y ∂z

ω =∇×v = ∂u ∂w  .
 ∂z − ∂x 
∂v ∂u
∂x − ∂y
The curl has to do with whether the vector field is spinning or not. We
need to be a little careful, we can have circular looking vector fields
that have zero curl. The physical interpretation of curl will be put to
the side for now and we will dig into a little deeper later in the book.
The curl of a vector results in a vector. It does not make sense to take
the curl of a scalar.
It is important to remember that del has units of 1/Length; for ex-
ample the units of ∇T are Kelvin/meter. If we work in dimensionless
terms we would very often (but not always!) decide to scale the x, y,
and z coordinate directions by the same length scale L. In this case our
dimensionless del would become,
1˜
∇= ∇.
L
Note that the same units apply to divergence, gradient and curl.
4.3 Normal vectors and flow through a surface

Consider an arbitrary surface. This could be a surface of a distinct
piece of material, i.e. the outer skin of a ball. We could also be taking
an imaginary slice or cross section through a solid body. Whenever
we define a surface, we also need to define a normal vector. A normal
n
ρv
Figure 4.3 Schematic of the normal vector and mass flux at a sur-
face. Here the vector field and the normal vector are not aligned so
only some fraction of the total mass flux at this location is cross the
surface.
vector is a unit vector (magnitude of 1) which points perpendicular to

a surface. When we have a closed surface, by convention we point the
normal vector outward. If we don’t have a closed surface, then we need
to define which direction the normal vector points.
We will commonly take the dot product of a vector field with the
normal vector of a surface. This product provides the the local rate
that “stuff” is flowing across the surface. For example, consider the
vector field ρv where ρ is the fluid density and v is the velocity field.
The units are kg/s · m2 . The meaning is the local mass flux, the rate
that fluid mass is being carried by the velocity field per unit area. The
dot product ρv · n gives the rate, per unit area, that mass is crossing
the surface at that point. If the mass flux vector field is aligned with
the normal vector of the surface, then all the matter that is moving
is also crossing the surface. If the flow is perpendicular to the surface
then ρv · n = 0 and nothing is crossing the surface even things are
flowing. The dot product gives us a measure of how well aligned the
flow is to crossing the surface. The surface normal and mass flux are
shown schematically in Figure 4.3.
4.4 Volume integrals

Imagine we have an arbitrary volume of material and we break the
volume up into a bunch of little discrete “bricks”. We weigh each brick,
P
sum them up and we have the total mass of the volume; M = mi .
We can also compute the mass of an individual brick by multiplying
the density by the volume, mi = ρi Vi . Now, you should be familiar with
4.5 Surface integrals 49
the idea that if we break the such a sum up into smaller and smaller
bricks that in the limit of lots of little pieces we have an integral,
Z
M = ρdV.
What we mean by this notation is that we integrate over x, y, and z;

ZZZ
M= ρ(x, y, z)dxdydz.
We will typically use the former notation for simplicity. If we actually

want to compute the integral above, we would need the density as a
function of x, y, and z as a known. If the density is a constant
R then we
can pull the density out of the integral in which case V = dV ; just
the total volume of the region.
We will need to remember that the element dV is a little volume
element, like a single brick in the wall. We also need to always keep
in mind that dV has units of volume. Above when we integrate mass
density (mass per unit volume) over the volume we are left with total
mass; the units of density turned to mass. Most often we will use the
volume integral to count up how much total “stuff” is contained in
a volume of material. By “stuff” we will mean total mass, energy, or
momentum. These type of total volume integrals are crucial in our
conservation laws.
We will also use the volume integral to give us a measure of the aver-
age value of some function, for example one useful average temperature
of an object could be computed as,
Z
1
Tave = T dV.
V
Volume integrals can be computed of either scalar or vector fields. The
volume integral of a scalar field is a single number and the the volume
integral of a 3D vector field is a single 3D vector.
4.5 Surface integrals

In addition to volume integrals, we will also very commonly describe
physical things using surface integrals. The concept is in many ways
similar to the volume integral only now we think about little patches
Figure 4.4 Latitude and longitude grid for the earth (from Encyclo-
pedia Britannica).
of surface area. Take for example the grid of latitude and longitude
on the globe shown in Figure 4.4. The grid patches are not equal in
area, thus we would need to weight any sort of sum by the area of each
patch. This weighting of the different area patches in a sum is what the
expression for the surface integral naturally does for us. We will use
surface integrals in many ways.
One example is the average of some scalar function on the surface.
If we wanted to compute the average surface temperature of the earth,
for example, we would express the calculation as,
Z
1
Tave = T dS,
A
where T varies at every point along the surface.

Very often we will compute the surface integral of a quantity which
has some dot product with the surface normal. These expressions will
be useful for telling us how much ”stuff” is crossing a surface. For
example ρv gives the total mass flux (in units of kg/s · m2 ), ρv · n gives
4.6 Surfaces and volumes in 2D 51
the how aligned the mass flux is with normal vector, and
Z
ρv · ndS,
would give the total mass flow (in units of kg/s) through the surface.
We will very often apply the surface integral to an entire closed sur-
face which encases some volume of material. A more formal notation
for the surface integral of a close surface would be,

ZZ
ρv · ndS.
This notation is explicit because it uses two integrals to highlight that

we are integrating over an area. The circle in the surface integral is to
remind us that we are integrating over a closed area which bounds the
volume. To keep our notation simpler (as with the volume integral) we
will not use the full notation with two integrals.
Can also compute surface integrals for vector quantities. Later in the
course we will discuss pressure a lot. In a fluid, pressure is an inward
internal force of the fluid. Pressure has units of force per unit area.
Pressure multiplied by the surface normal gives a stress vector, a force
per unit area acted at a point along the surface. If we add up all the
little stress vectors and weight them by the area over which they locally
act, we have the total force due to pressure acting on this surface. Thus,
the net force due to pressure.
Z
F = P ndS
The volume integral of a scalar field is a single number and the the
volume integral of a 3D vector field is a single 3D vector.
4.6 Surfaces and volumes in 2D

Through this course I will often use 2D problems as examples since
they are illustrative, easy to sketch, and easier to solve than 3D. I am a
poor artist and a little lazy, so 2D works great for me. It is often a point
of confusion as what is a 2D volume? What is a 2D surface? Simply
put, a 2D volume is a plane that we can sketch on a piece of paper
and a 2D surface is the line that encloses that plane. When we think
about what this means for our 3D world, just imagine that the thing
you draw extends outward and into the page forever. The 2D object
that we draw is a cross section of an infinitely long object. Working in
2D can sometimes confuse our units, so you have to remember that we
are computing things per units depth into the page.
4.7 Divergence theorem

The primary vector calculus theorem we will use is the divergence the-
orem, which states for any vector field v,
Z Z
∇ · vdV = v · ndS.
Even though we are using v as our vector field here, the above result
is a general result from calculus and has nothing to do with whether v
is a physical velocity.
We can derive this result by sticking to a simple 2D example using
vector field v with x and y components u and v. The components of the
vector field are functions of space; u(x, y) and v(x, y). For simplicity
we will take our region of interest to be a unit square. We will start by
evaluating the surface integral. On the left boundary v · n = −u and
on the right boundary v · n = u. The change in sign is that on the left
the normal vector and the positive x component of the velocity vector
point in opposite directions whereas on the right they point in the same
direction. The dot product means that the sign will be opposite for the
two terms. The same holds for the upper and lower surfaces, only here
it is the y component of the velocity that shows up from v · n. At this
point,
Z Z y=1 Z y=1
v · ndS = − u(x = 0, y)dy + u(x = 1, y)dy
y=0 y=0
Z x=1 Z x=1
− v(x, y = 0)dy + v(x, y = 1)dx
x=0 x=0
or by grouping terms,
Z Z y=1
v · ndS = (u(x = 1, y) − u(x = 0, y)) dy
y=0
4.8 Summary 53
Z x=1
+ (v(x, y = 1) − v(x, y = 0)) dx.
x=0
Recall that the fundamental theorem of calculus states that for any
function,
Z x=1
∂u
u(x = 1) − u(x = 0) = dx.
x=0 ∂x
Therefore, our first expression can be rewritten as

Z Z y=1 Z x=1 Z x=1 Z y=1
∂u ∂v
v · ndS = dx dy + dy dx.
y=0 x=0 ∂x x=0 y=0 ∂y
Since the order of integration doesn’t matter,

Z Z y=1 Z x=1
∂u ∂v
v · ndS = + dxdy.
y=0 x=0 ∂x ∂y
Therefore, we can write the final result in our vector calculus language,
Z Z
v · ndS = ∇ · vdV.
Note that we did this derivation for a unit square in 2D. We could have
been a little more general and the result would have held up. This is
one of the most useful theorems in this course. The theorem relates the
total flux of a vector field through a closed surface to the divergence of
the same field inside the volume.
4.8 Summary
As stated in the beginning, I am assuming that the notation is fa-
miliar to you and you are not completely new to vector calculus. If
you are new to the topic, then this “introduction” is going to be too
brief. You probably noticed that I focused more on the physical inter-
pretation of various surface and volume integrals rather than actually
computing surface and volume integrals. This focus on the meaning will
continue through the book. We will use the language of vector calculus
to describe our fundamental conservation laws and the notation is used
throughout the derivations.
5
Diffusive mass transfer
In this chapter we will consider an inert and dilute chemical species

mixed with a fluid; think of a little food coloring or dye added to water.
The notion of dye implies something passive (no chemical reactions) and
we have a way to measure and visualize the concentration. Starting with
the behavior a passive dye is a good place to begin our development of
theory since traces of dye is how we often visualize and measure fluid
flows in practice. Motion and transport of a dye is something we can
see, measure, and have some intuition for.
Through our discussion I will refer to the dye as comprised of molecules

since the microscopic physical picture is helpful for providing physical
intuition. Even though we will talk of molecules, it is important not
to forget that our core analysis is in the continuum realm. Over any
length scale of interest, there are many dye molecules in that volume -
there are just far fewer dye molecules than water molecules.
There are two mechanisms by which we can mix a dye with a fluid;
convection and diffusion. This chapter will only be concerned with
molecular diffusion in a stationary fluid. This chapter will set the stage
for the next chapter where we will start to see what happens when the
fluid starts moving. Starting with diffusion is also useful as it gives us
a simple case where we can start to connect our physical intuition with
the formal framework of vector calculus.
56 Diffusive mass transfer
5.1 Physical picture of mass diffusion

We will ultimately describe mass diffusion as a continuous problem and
will consider the dye concentration as a local continuum quantity. It is
useful, however, to start with a discrete picture where we follow a single
dye molecule. If we get our imaginary microscope we see dye molecules
are moving about even though the fluid is stationary. Each individual
dye particle undergoes a wayward path. The dye particle is constantly
colliding with the randomly jiggling and wiggling fluid molecules. All
this thermal motion of the fluid causes the dye particle to change direc-
tion and speed often. The path of any particle is random and cannot be
predicted. We can however, speak of a statistical average of the random
motion. This average motion of many wiggling and jiggling molecules
is what we call diffusion.
The random motion of a single particle is known as Brownian motion.
In 1828 Robert Brown noticed that small pollen grains in a liquid had
paths like that shown in Figure 5.1. If you ever look under a microscope
at very small particles a few microns in size, you can see this motion for
yourself. Einstein considered this problem in 1905 (the same year that
he wrote the theory of relativity) and put the motion of such particles
on a theoretical basis. From his analysis, he derived an expression that
allowed one of the first accurate measurements of Avogadro’s number
and other molecular properties. The Einstein analysis helped connect
the microscopic to the macroscopic (or continuum worlds). Brownian
motion is observed with particles, but the same type of motion would
be observed with our consideration of dilute dye molecules.
The random walk that a single particle takes is often called the
drunken sailor problem. The drunk sailor walks out of the bar has
no idea where he is going and does not remember where he has been.
Every step is taken at complete random. I do not know why we are so
quick to stereotype sailors, but that is beside the point. Some of his
steps will be small, some long; some to the left, some to the right. A
simulation of this situation is easy to do. Simply start at known x, y
coordinate location and with each time step, move a random distance
in both x and y. In doing so we use the Gaussian, or normal, distri-
bution to determine the direction and magnitude of each step. Every
time we run the simulation we will get something different, a sample is
shown in Figure 5.1.
5.1 Physical picture of mass diffusion 57
Figure 5.1 On the left, an example of a random path taken by the

drunken sailor. The sailor starts at zero,zero and at each instant in
time takes a step that has random direction and magnitude. On the
right, distance squared, r2 , as a function of time averaged for 100
random walking particles. The straight line is a linear relationship
for reference.
Since each simulation is different we can repeat the experiment many

times and ask the question, on average how far away from the bar is
the sailor? If we plot the total distance squared from the bar versus
time for the average of many simulations on log-log coordinates we
obtain a result as shown in Figure 5.1. We find that in this case that
an individual sailor gets further from the bar in a random way, but
that on average r2 increases linearly with time. To end up twice the
distance from the bar takes 4 times as long. This fact will come back
to us when we consider the continuous diffusion problem.
If we start a closed box filled with random walkers all concentrated
in the center, we expect after time passes for the distribution to spread.
If we wait a long time, the molecules should be evenly distributed
throughout the box since it becomes equally likely for a dye molecule to
be at any location within the box. Diffusion tends to make the dye con-
centration uniform. Statistically speaking, given the very large number
of 1023 molecules it is impossible for the system to evolve to any state
far in the future other than uniform distribution.
Imagine a 2D closed box and drawing a vertical line along the cen-
ter, Figure 5.2. It seems reasonable to believe that the rate that dye
molecules on average cross this imaginary line from left to right would
be proportional to the concentration of dye particles on the left side of
the line. Likewise, the rate that dye particles cross the imaginary line
Figure 5.2 Box where the left has higher concentration than the
right. .
from right to left is proportional to the concentration of dye particles on

the right side of the line. The net flux across this imaginary line is the
difference of the two values. Flux will be defined more formally soon,
but it is the number of dye molecules moving through an imaginary
surface per unit time per unit area.
If the concentration is the same on the left and right side of the
line, there is no net flux in any direction even though individual dye
molecules are constantly swapping sides. If the concentration is not
uniform as in Figure 5.2, then it should be clear that if each particle
is undergoing a random walk within the box that the left side will be
losing dye particles while the right side will be gaining them. All the
dye molecules are acting the same (on average) there are just more dye
particles on the left when we start. This difference in concentration
results in a net flux of molecules from the more concentrated to least
concentrated regions. Ultimately we would expect the particles to be
evenly distributed throughout the box, just based on probability, and
at that time the flux at the center of the box would be zero.
The important observation is that for their to be a net flux, we need
to have a concentration gradient. If the concentration is not uniform,
then the concentration at any location, just a small distance dx away,
is given by a Taylor series as
∂c
c(x + dx) = c(x) + dx.
∂x
where c(x) is the concentration on the left and c(x + dx) is the concen-
tration on the right. If the net flux at a point were to be proportional
to the difference in concentration just to the left and right side of an
imaginary line (i.e. c(x + dx) − c(x)), then the net flux would propor-
tional to the concentration gradient. The relationship between flux and
concentration gradient is in fact what is observed with the random walk
and is known as Fick’s law. We will formalize Fick’s law very soon.
5.2 Continuum concentration

Now we are going to move from thinking about individual dye molecules
to thinking about concentration as a continuum concept. We define the
continuum concentration, c, to be the number of dye molecules per unit
volume. Let’s take a 1 liter container of water and mix in 6.022 × 1020
dye molecules. When dye is evenly distributed, I have a concentration
of 6.022 × 1020 particles per liter, or 1 milli-molar (mM). This concen-
tration would be considered very dilute as there would be on the order
of 1 dye molecule per 100,000 water molecules. If we shrunk down and
took a sample volume the size of red blood cell, there would be on the
order of 10,000 dye molecules at that scale. These numbers should give
some sense by what we mean with a continuum value of concentration.
If we shrink down to describe a continuum “point” as something the
size of a single cell, we are fine talking about continuum concentration.
However, it is clear if we kept taking a smaller and smaller sampling
volume we would eventually have a small, countable number of dye
molecules. When the problem length scales is large compared to scale
where we would only find a few dye molecules, then we are pretty safe
working in the continuum realm.
If the concentration is uniform the total number of dye molecules,
N , is related to the concentration, c, simply through the volume,
N = cV.
Now let’s take a case where the dye is not evenly distributed. We could
imagine breaking the liter volume up into 1000 mL size samples and
count the molecules in each sub-volume. The total number of molecules
would then be the sum of the numbers in each element. In the contin-
uum limit when we can describe the concentration as a local scalar
field, c(t, x, y, z), then we formally find the total number of molecules
through volume integration,

Z
N= cdV. (5.1)
I keep referring to counting the number of dye molecules. In a real

experiment we would never count molecules. We could measure the
concentration many different ways, depending on the nature of the dye.
With dye or food coloring, we can visually see the concentration by the
darkness or lightness of the color.
5.2.1 Mass flux and conservation

Since the dye molecules do not chemically react we can (at least in
principle) track their numbers. If we consider any arbitrary volume
then the change in the number of molecules stored inside that volume
must equal the difference between how many dye molecules entered and
left the volume,
∆N = Nin − Nout .
This equation is nothing more than like counting the number of people
in a room. Stand at the door and count how many people come in
and out and you know by how much the total number of people in the
room has changed. In the subject of fluid dynamics, we typically work
in rate form, which we can obtain by taking the time derivative of the
expression,
dN
= Ṅin − Ṅout . (5.2)
dt
I should comment on the use of Ṅin to denote the rate of molecules
coming into our volume. It is common notation to use a dot above a
variable to denote a rate. This dot should not be confused with the
one often used in mathematics books where df /dt ≡ f 0 (t) ≡ f˙(t).
Making the analogy of the doorman, if I count 60 people entering into
a room in one minute, then I would have and average rate of Ṅin = 1
person/second.
In analyzing this equation in a continuum sense, we will need a way
to describe the rate that particles are leaving an arbitrary volume ele-
ment. To provide more formalism, we need to define a mass flux vector,
j. The boldface notation denotes a vector quantity. Mass flux has mag-
nitude and the direction points in the direction, on average, that the
dye molecules are moving. We saw previously that we expect the flux
to point from more to less concentrated areas of dye. The units of flux
are number of dye molecules per unit time per unit area. It is fine to
think of the flux as related to the average velocity of the continuum
collection of the dye molecules at that point in space. Remember, for
now we consider that the carrying fluid remains stationary.
If we take the surface of an arbitrary volume, and draw a surface nor-
mal vector n to point outward, then j · n is the rate that dye molecules
cross the surface at that point, per unit area. The dot product with the
surface normal vector is important since the mass flux is a vector. If
the mass flux is pointing tangential to a surface, then no dye molecules
are crossing the surface. If the flux vector and the normal vector are
aligned, then all the movement of the dye is such that it is cross the
surface. See the schematic in Figure 5.3.
The net rate that dye molecules flow across the entire boundary of a
closed volume is given by a surface integral. If we consider our surface,
S, which encloses our entire volume of interest the net rate that dye
molecules cross that surface is
Z
Ṅin − Ṅout = − j · ndS. (5.3)
The minus sign is there because the convention is to define the normal
vector to point outwards from a closed surface. The dot product of the
normal vector and the mass flux vector automatically gets the sign for
inward and outward flux from the volume correctly.
Making the analogy with people coming and going from a room,
the surface integral would be equivalent to giving 1000 bouncers, 1000
doors around the large room to watch. Each bouncer would count the
number of people coming and going. The surface integral would be the
summary report from all the bouncers. The surface integral counts the
amount of dye coming and going per unit area all along the surface,
and the adds it all up.
n
j
Figure 5.3 Schematic of mass flux and normal vector on an arbi-

trary, closed surface. The concentration of dye inside the volume
is sketched to be higher than outside. This case would have a net
flux outward from the volume, decreasing the total number of dye
molecules inside.
5.2.2 Conservation of mass

If we now consider a closed, arbitrary volume in the fluid, Equation 5.2
written in vector form is,
Z Z
d
cdV = − j · ndS.
dt
This equation makes the statement that the time rate of change of
the number of molecules inside the volume must equal the net rate
that molecules cross the surface. For now, we consider mass transfer in
a stationary supporting material and our volume of integration remains
fixed and does not move. If the volume is fixed, on the left hand side
the order of the derivative and integral can be switched,
Z Z
∂c
dV = − j · ndS.
∂t
Applying the divergence theorem to the right hand side converts the
surface integral to a volume integral. Grouping the two volume integral
terms together yields,
Z
∂c
+ ∇ · j dV = 0.
∂t
Putting the equation in this final integral form allows us to turn
the expression into a local differential equation. The equation above
says the integral of some quantity in parenthesis must equal to zero.
j j j
Cincreasing Cdecreasing Cconstant
j<0 j>0 j=0
Figure 5.4 Schematic of the divergence of the mass flux. The con-
centration at a point rises, falls, or stays the same depending on the
local divergence of the mass flux vectors.
An integral can be equal to zero when the thing you are integrating is
non-zero. Think about sin(x) integrated from 0 to 2π. The function is
positive some of the time and negative the rest and in perfect balance.
However, in our problem the volume of integration is arbitrary. I could
make the integration region bigger, smaller or move it to another place.
The only way the integral can always be zero is if the thing we are
integrating is zero. Everywhere. Therefore our integral equation can be
written in differential form as,
∂c
+ ∇ · j = 0. (5.4)
∂t
This equation is valid at every point.

The physical interpretation of the equation using what we know from
vector calculus is simple. If all the mass flux vectors point toward a sin-
gle location then dye is flowing toward that point and we expect the
local concentration to increase with time. Mathematically, when the
flux vectors point toward a single location, then the divergence is neg-
ative and thus our equation says that the rate of change of concentra-
tion at that location is positive. If all the mass flux vectors point away
from a point, then the concentration at that location decreases with
respect to time. At a point, the mass flux vectors could be converging
in the x-direction but diverging in the y direction, thus local amount of
dye flowing to that point on net zero and thus the concentration does
not change. The regimes of positive, negative, and zero divergence are
5.2.3 Fick’s law

So our equation is beautiful to look at,
∂c
+ ∇ · j = 0,
∂t
but it is not complete. There are four unknowns, 3 components of the
mass flux vector and concentration, and only one equation. Closure to
the problem is provided via Fick’s law. Ficks’s law is an example of a
constitutive law. Constitutive laws capture the microscopic behavior of
our continuum. The use of a constitutive law is always needed to close
conservation laws. The conservation law is based on the assertion that
the number of dye molecules is fixed. We will then require a constitutive
relationship to complete the problem.
Fick’s Law states that
j = −D∇c. (5.5)
The law, which is not really a fundamental law, states there is a linear
relationship between diffusive mass flux and concentration gradients.
Concentration has units of number of molecules per unit volume. The
mass flux has units of number of molecules per unit area per unit time.
The diffusivity D has units of m2 /s. We treat D as a material property
that we can look up. Typically the constant is quite small if the medium
is a liquid. For a small molecule in water D ∼ 10−9 m2 /s. For larger
molecules it can be an order of magnitude smaller. It is important to
remember that Fick’s law is a constitutive law which is not necessarily
valid for non-dilute systems. For the system of random walkers, Fick’s
law turns out to hold true.
5.2.4 Diffusion equation

The use of Fick’s law in our conservation equation yields,
∂c
= ∇ · (D∇c).
∂t
When the diffusivity is a constant, which is often but not always the
case, our equation simplifies to,
∂c
= D∇2 c. (5.6)
∂t
It is a single equation for the concentration at every point in space

and time. It can (at least in principle) be solved when provided an
initial condition and appropriate boundary conditions. The equation is
also sometimes called the heat equation as it is the same equation that
governs heat conduction.
In order to solve the above equation, we need boundary and ini-
tial conditions. The equation itself is good for the region of interest.
The boundary conditions are the mathematical expression of what is
happening just outside the region where we are solving the equation.
The initial condition represents the state of the system when we start
watching it.
There are some common classes of boundary conditions which are
show up in a number of problems. For diffusion of dye, the most com-
mon physically realizable condition is insulating. Here, we set the
mass flux through the surface to be zero, j · n = 0. This condition
would be realized at an impenetrable wall where the dye cannot leave
the system. Another possible condition is fixed concentration. This
condition simply means that we know and prescribe the concentration
along the boundary. This might be a condition at a solid surface where
the solid dye was dissolving into a liquid (or being generated through
a chemical reaction). We could also prescribe the mass flux at the
boundary, which would be stated as −D∇c · n = q.
The boundary conditions can be a great source of uncertainty. While
many conditions realize the insulating condition in practice, it is more
difficult to prescribe an exact concentration or flux at a surface. With
the exception of the insulating condition, many other boundary condi-
tions would be approximations and the source of the largest error when
comparing theory to experiments.
5.3 Equation in 1D
While we derived the equation in 3D, it is easiest to solve the equation
and gain some insight by focusing on a simpler 1D case. In 1D the
partial differential equation which relates changes in concentration with
respect to time and space is,
∂c ∂2c
= D 2.
∂t ∂x
Figure 5.5 Graphical interpretation of the mass diffusion equation.

We show three sample spatial concentration profiles at a particular
instant in time. Consider whether the concentration at the center
(x=0.5) increases, decreases, or stays the same (in time) based on
the spatial curvature.
The equation says the first derivative of concentration in time is propor-

tional to the second derivative in space. The second derivative in space
is the curvature. If the spatial concentration field is locally curved, then
the concentration at that point will change with time. This behavior
is shown graphically in Figure 5.5. Here we show three concentration
profiles in space at a snapshot in time. One profile is linear while the
others are concave up and concave down. Lets consider the concentra-
tion at the center point in space. In the concave upward case, curvature
at the center is positive. The center concentration therefore increases
with time. Physically this make sense because the material a little to
the left and a little to the right is more concentrated with dye than the
center, pulling the center concentration upwards. In the concave up sit-
uation, there are mass flux vectors pointing and converging toward the
center. The situation is reversed for the concave down example. When
the concentration gradient is linear, the point to the right is pulling
up and the concentration to the left is pulling down and thus the con-
centration remains constant with time. In the linear case, there is a
constant mass flux through the system. The flux is everywhere uniform
and thus no local accumulation.
Figure 5.6 Solution to the diffusion equation. The initial condi-

tion is that the left half of the domain starts at a concentration
of 1 and the right half is zero. Snapshots in time are shown at
t = 0.0001, 0.001, 0.01, 0.05, 0.1, 0.2, 1. Here D = 1.
Now let us consider a specific 1D problem. Consider a 1D region

that initially has dye concentrated on the left half and no dye in the
right half. A partition separating the two regions initially keeps the
left concentration uniform and allows nothing to the other side. We
then suddenly make the partition disappear and diffusion allows dye to
migrate from left to right. The end walls are insulated such that dye
cannot leave the system from either boundary. The region has a total
width of `, thus our domain will be 0 < x < `. The equation, initial
condition and two boundary conditions would be stated as
∂c ∂2c
=D 2 (5.7)
∂t ∂x
c(x < `/2, t = 0) = c0 ; c(x > `/2, t = 0) = 0,

∂c
= 0,
∂x x=0

∂c
= 0. .
∂x x=`
We want to understand the rate at which the dye spreads and overtakes
the entire area.
The mathematical solution (which we will soon walk through) is
shown in Figure 5.6. However, the solution probably makes some intu-
itive sense based in your experience. Pay attention to the curvature in
the solution. The left half always has curvature downward and the right
half is curvature upward. The curvature corresponds on the right where
the concentration always increases and the left it always decreases. The
center point is an inflection point with zero curvature and the value of
the concentration there doesn’t change. Note that the final state is uni-
form where the concentration field is flat. The equation says that steady
state (when things can no longer change in time) is reached when the
concentration field is uniform or linear. The boundary conditions re-
quire the concentration field to be flat at the boundary, thus the final
state can only be uniform concentration.
5.3.1 Pi Theorem
Without solving the equation, let’s use the Pi theorem and see what
that tells us. In dimensional terms we want to find c(x, t, D, `, c0 ). To
introduce the concentration we need a new independent dimension, the
number of molecules, N . The table for our dimensionless concentration
and the systematic removal of units is,
2
c [ LN3 ] x [L] t [T] c0 [ LN3 ] ` [L] D [ LT ]
2
c
c0
x [L] t [T] ` [L] D [ LT ] remove N
c x D
c0 `
t [T] `2
[ T1 ] remove L
c x tD
c0 ` `2
remove T
The Pi Theorem yields the dimensionless concentration is a function

of the dimensionless spatial coordinate and the dimensionless time,

c x tD
=f , 2 .
c0 ` `
Dimensionless numbers time and space coordinates should bother you
no more than the arbitrary definitions of the meter and the second.
5.3.2 Scaling the equation

The equivalent approach is to explicitly take the governing equations
and convert the units from man made ones to dimensionless ones. We
did this for a few simple problems in our first chapter on dimensional
analysis. To proceed, we introduce a non-dimensional spatial coordinate
x̃ = x/` so that the domain extends from 0 < x̃ < 1. We can also define
a non-dimensional time as t̃ = t/t0 where t0 is an arbitrary (for now)
time scale. Under this change of variables, the equation becomes,
∂c ∂2c
=D . (5.8)
∂(t̃ t0 ) ∂(`x̃)2
Since ` and t0 are constant we can pull it outside the derivatives to
obtain
∂c Dt0 ∂ 2 c
= 2 . (5.9)
∂ t̃ ` ∂ x̃2
Nothing is stopping us from arbitrarily setting t0 = `2 /D so that the
diffusion equation becomes simplified with no material or geometric
parameters,
∂c ∂2c
= . (5.10)
∂ t̃ ∂ x̃2
It should be clear that since the left and right side of the equation
are proportional to the concentration, we can easily divide the overall
concentration by c0 . Normalizing the concentration is really no different
that working with molar concentration units rather than with the actual
number of molecules.
The dimensionless equation and boundary conditions show that the
solution has to have a form

c x tD
= f (x̃, t̃) = f ,
c0 ` `2
The functional form is exactly that obtained via the Pi Theorem.
While it may not seem so, these results are very powerful and useful
even though we haven’t solved a problem yet. The units of the equation
says all diffusion problems (with this boundary condition) are the same.
We simply scale the geometry to have a length that ranges from zero
to one. The combination of the size and the diffusivity gives us the
appropriate time unit. On the scaled domain and in the proper time
units, problems of different size and material property will have the
same solution. Regardless of the size or material involved, if you solve
the diffusion equation for this problem you have solved it once and for
all. Once we have our solution we simply use the definitions above to
move from the dimensionless solution to physical units by multiplying

by the appropriate constants.
This scaling for time in the diffusion equation is very important. It
provides us with an estimate for how long things take to spread and
come to equilibrium. If we know the material and the length scale, then
equilibrium will be reached on a time scale of the order of `2 /D. This is
not a precise estimate, but a useful one nonetheless. The estimate tells
us the order of magnitude. If the estimate tells us the time scale is one
minute, we can’t be sure if it will take 1 minute or 3 minutes. However,
we can be confident that equilibrium will not occur in one second or
take hours. It is an order of magnitude estimate - that’s all. These types
of estimates based on arguments of units are useful because they are so
simple and can give some immediate intuition. Finally, note that the
scaling t ∼ `2 /D is the same as the average displacement we found in
the first section where we considered the particles undergoing a random
walk. In diffusion problems time scales with `2 - double the thickness
and it will take four times as long.
5.3.3 Analytical solution in 1D

In one dimension, the diffusion equation is usually solvable analytically.
In another book of mine on partial differential equations (well, it’s in
progress anyway), I work through these solutions in much more detail
than I do here. Here, I just want sketch out a few results for you to be
familiar with and you can consult other resources for more detail. Let’s
consider the equation in 1D in dimensionless terms where D = 1 and
0 < x < 1. The equation in 1D reads,
∂c ∂2c
=
∂t ∂x2
It turns out that the general solution has the form,
2
c(x, t) = e−λ t (A sin(λx) + B cos(λx)) .
While I am not providing how I arrived at this form, you can check by
inspection that it in fact satisfies the partial differential equation. The
constants A, B, and λ are unknowns until we apply the two boundary
conditions and the initial conditions.
To show how we go about determining the constants, let’s consider
the particular case where the two boundary conditions are insulating.
The boundary conditions were stated as,

∂c
= 0.
∂x x=0,1
Taking the first derivative in x of our solution form gives,

∂c 2
= e−λ t λ (A cos(λx) − B sin(λx)) .
∂x
Evaluating this derivative at x = 0 and applying the insulating bound-
ary condition gives,

∂c 2
= e−λ t λA = 0.
∂x x=0
This condition means that the solution should have A = 0. Applying
the boundary condition at x = 1 yields,

∂c 2
= −e−λ t λB sin(λ) = 1.
∂x x=1
While B = 0 or λ = 0 could be set such that the solution would satisfy
the boundary conditions, these choices would result in a trivial solution.
The only interesting solution that can satisfy the boundary condition
is when sin(λ) = 0. Thus λ = nπ where n is an integer. Therefore
on application of our two insulating boundary conditions, we have a
solution in the form of a sum of cosines with exponential decay in time,
∞
2
π2 t
X
c(x, t) = e−n Bn cos(nπx).
n=0
The constants Bn are determined by applying the initial condition.

The trick (and it really is a trick, at least until you have seen it once
before) to find the constants Bn relies on a fundamental property of
cosines,
Z 1
cos(nπx)cos(mπx)dx = 0; n and m are integers, n 6= m
0
Z 1
cos(nπx)cos(nπx)dx = 1/2; n is an integer
0
Let’s take an arbitrary initial condition, c(x, t = 0) = c0 (x). The initial

condition problem is to find Bn that satisfy,

∞
X
Bn cos(nπx) = c0 (x).
n=0
We multiply the entire initial condition equation by cos(mπx) and in-

tegrate the entire equation across the domain we obtain,
∞
Z 1 X !
Bn cos(nπx)cos(mπx) = c0 (x)cos(mπx) dx.
0 n=0
If you apply the properties of products of cosines integrated over the

domain, you will note that of all the terms in the infinite sum on the
left side of the equation, only the case where n = m gives something
other than zero. Therefore the constant Bn is set by,
Z 1
Bn = 2 c0 (x)cos(nπx)dx
0
If we perform the above integral for the initial condition used in

Figure 5.6, we would be able to easily determine the constants Bn . The
solution in this case would be
∞
1 X 2 sin nπ

2 2
c(x, t) = + 2
e−n π t cos(nπx).
2 n=1 nπ
While we can plot the solution to see what it looks like, let’s think
about it a little more. First, note that all the terms with even values
of n would be zero due to the sin(nπ/2) part of the equation. Now
imagine carrying out the sum for the first few odd values of n,
1 2 −π2 t 2 2 2 2
c(x, t) = + e cos(πx)− e−9π t cos(3πx)+ e−25π t cos(5πx)+...
2 π 3π 5π
Note in the above solution that for n > 1, once a little time has passed
the coefficient in front of these terms would be really small. Imagine
the solution at t = 1/π 2 ≈ 0.1. At this time e−9 << e−1 . It should be
clear that once t is no longer small then the solution is dominated by
the first cosine term,
1 2 2
c(x, t) ≈ + e−π t cos(πx)
2 π
When you look at Figure 5.6, notice that not much time passes before
you see this single cosine term in the solution.
5.4 Role of simulation 73
If this is the first time you have come across this type of solution to
a partial differential equation, the method probably seems mysterious
at this point. More detail on the origins and meaning of this solution
as well as examples for different boundary conditions can be found in
my other resources.
5.4 Role of simulation

The equation,
∂c
= D∇2 c, (5.11)
∂t
with appropriate boundary and initial conditions can provide the con-
centration as a function of time and all three spatial dimensions in a
stationary fluid. The equation is linear and well-behaved such that we
can usually obtain a solution to the problem of interest. If we make
some simplifying assumptions, we can often solve the problem analyti-
cally as in the last section.
In complicated geometries or with complicated boundary conditions
we can use computer simulation; usually finite element software to solve
the problem numerically. There are a number of free and commercial
codes that can be used to solve such problems. The equation is generally
well-behaved, thus numerical solutions are usually (but not always)
obtained without much difficulty with modern software.
If you use simulation, it will be important for you to keep the ba-
sic mathematical picture and physical problem formulation in mind.
In order to have a well defined problem you will need to specify the
following in software,
• The geometry of the physical domain. In this chapter, our geometry

was usually simple (e.g. a line) but in software you can easily have
any shape. The software will break your geometry up into a mesh of
a number of small elements. For mass diffusion, the default meshing
process will often be sufficient. In complex problems, meshing can
become an art form.
• The equation. While we have discussed the basic mass diffusion equa-
tion alone, the software might have additional terms which you can
activate, such as chemical reactions.
• The diffusivity of the dye species in the fluid, D. Alternately, you

may want to solve a dimensionless problem and you can set your
properties to 1.
• The initial concentration distribution (unless you are only interested
in steady state).
• Appropriate boundary conditions. All endpoints, edges, or surfaces
of your geometry (whether in 1D, 2D, or 3D respectively) will need
boundary conditions specified. The most common conditions would
be insulated (no normal gradient), fixed concentration, or fixed mass
flow.
Before you start a simulation you should perform some estimations
and try to obtain a picture of what to expect. For example, if you are
solving a transient problem, use the dimensional analysis result that
t ∼ `2 /D to determine approximately how long it will take to come
to equilibrium. Simple analytical solutions that are easy to evaluate,
provide the parameters that dominate a problem, or collapse all the
solutions to a single plot or equation are really handy to have. Even if
your problem has a complex geometry, having a simple 1D model can
still often be really handy.
5.5 Summary
Mass transfer by diffusion is generally a pretty slow process. If you
consider something like sugar mixed in water, the diffusivity is D =
6 × 10−10 m2 /s. We discussed that you could estimate the time it takes
diffusion to occur over some distance as t ∼ `2 /D. Using this argu-
ment and assuming your coffee cup is about ` ∼ 10 cm then it would
take about half a year for sugar to mix within your coffee cup. If you
wanted the molecules to diffuse over a length scale of your height, you
might need to wait 200 years. Diffusion is slow at the human scale. If
you consider a single cell, which might have a size of 10 microns (1
micron is 10−6 m), then we get diffusion times scales of on the order
of a second. Diffusion is not so slow if you happen to be really small.
When we observe fluid mixing in our everyday life, it is most certainly
dominated by convection. In the next chapter we will begin to describe
conservation laws in moving fluids.
6
Convective mass transfer
In this chapter we again consider an inert and dilute chemical species

mixed with a fluid. In the previous chapter we considered diffusion only
and we saw that diffusion is generally very slow (at least at the human
scale). A common analogy for mass diffusion is the example of opening
a bottle of perfume at the front of the room. If I did this, the front of
the class would smell it first, then it will spread across the room toward
the back. The behavior would seem “diffusion-like” but it is not really
mass diffusion. The dominant transport mechanism of perfume would
be driven by air motion. Even if the air in the room feels still, it is
not. Slight currents dominate over the very slow process of molecular
diffusion. In the perfume example we might have to wait years for
molecular diffusion to do the trick while convection in seemingly still
air takes less than a minute.
We will use the example of a passive dye to introduce how fluid flow
can dramatically change the behavior and analysis. This chapter will
introduce much of the formalism need to analyze fluid flows.
6.1 Convective mass transfer

The formulation derived in the last chapter is useful when the support-
ing medium is stationary. However, when fluids are involved it is often
not the case that things are still. If you drip a little food coloring in
water very carefully, you will probably see that after it sinks to the
bottom of the cup it just sits there. You would have to wait quite some
time for the dye to fully mix by molecular motion. Give the water a
76 Convective mass transfer
quick flip with a spoon and mixing occurs seemingly in an instant. In

general, at the scale of everyday things that you and I are intuitively
familiar with, mixing by convection in a fluid is much more efficient
than by molecular diffusion. When we observe mixing in our everyday
life, it is most certainly dominated by convection. In this section we will
start to describe how to describe conservation laws in moving fluids.
In describing fluid flow, we will refer to the carrying fluid’s velocity
field. We will describe the velocity vector field v(t, x, y, z) as a function
of space and time. By convention, we will refer to the x, y and z com-
ponents of the velocity vector as v = [u, v, w]. For the purposes of this
chapter we will assume by some magic that the fluid’s velocity field is
known. In the next chapter and for much of the rest of this book we
will discuss how to calculate that velocity field - a non-trivial matter.
6.2 Material point of view - be one with the fluid

Imagine a river which meanders through the countryside. The river has
some wide regions with very slow flow and some narrow regions with
faster flow along with some rapids and waterfalls. The river however, is
in a state that is constant with time and does change from hour to hour
or day to day. Imagine you are sitting on the bank of this river studying
the flow of water. Your reference frame is fixed and everything seems
constant and nothing changes if you come back the next day to the
same spot. If you measured any property of the river at your particular
location, and you would see the same value every day and might be
tempted to say nothing is changing. The river is static. Now imagine
you are on a raft studying the same river. The river seems anything but
steady and constant. You are constantly speeding up, slowing down,
twisting left and right, and generally being thrown about. From the
raft you would surely report that the river is dynamic. Whether some
property of a flow seems constant in time or not, is a matter of whether
you like to sit on the river bank or go with the flow.
In formulating our conservation laws, it is useful to take a point of
view of the fluid. In analyzing flow we often will refer to a “fluid parti-
cle”. This is a nomenclature that is common, but perhaps misleading.
There are no particles per se in our description of the flow. By a fluid
particle, we mean that we attach an imaginary massless membrane to
a fixed amount of fluid. We then watch where this fluid blob goes as
it flows and deforms. The fluid particle is considered to be composed
of the same material for all time. This consideration is not literal from
a molecular point of view, it is surely true that the random motion
of molecules means that individual molecules are exchanged from one
fluid particle to the next.
6.3 The material derivative

Let us continue with the river example and quantify the idea of change
from the fluids perspective. Imagine that upstream, someone is illegally
dumping some pollutants and you are tasked with figuring what is
going on. Let’s imagine a case where the river water flow is constant
in time. In principle, if you had enough sensors you could describe the
concentration field of the pollutant at every point in space and time;
i.e. c(x, y, z, t). This is difficult to do, so you set about to do some more
localized sampling.
First, you decide to fix your reference frame by standing in one spot
on a river bank. For this point of view we measure the rate of change
of pollutant by simply taking the time derivative of the measured con-
centration, ∂c/∂t. Since the river’s flow is constant, if you measure the
concentration changing in time we would assume this must have some-
thing to do with changes in the rate that pollutants are dumped in the
river upstream.
Now imagine that we decide to do some sampling of the river on a
small boat. We get on a small boat and drift down the river, just going
with the flow. If we monitored the concentration of pollutant from the
perspective of the flow, we would sense something different. If we are
upstream of the source, we sense no pollutant. As we pass the source,
we would presumably measure an increase in concentration. Now that
we have become a fluid particle the rate of change of concentration
with respect to time from the perspective of the fluid, is different than
in the fixed frame. We will find it useful to describe the rate of change
of something with respect to time from the perspective of the fluid.
Even though we measure a different rate of change depending on
the perspective, it is important to remember that there is only one
concentration field. If you are measuring from the river bank and me
from the boat, we better get the same instantaneous measurement at

the time I drift over your fixed sensor. We are making measurements
of the exact same thing and ultimately our perspectives must line up.
6.3.1 Derivation in 1D
Let’s start with a simplified example. Imagine the pollutants in a par-
ticular region of the river follow a linear and steady concentration field
that only varies in the downstream coordinate x. If you walk along
the river in this location and keep taking measurements of the concen-
tration field from the river bank you find a region that concentration
increases linearly, Figure 6.1. For this example, the concentration field
at the location of change is
x
c(x) = c0 + ∆c ,
L
where ∆c is the increase in concentration over distance L. The slope of
the concentration field in this region is constant,
dc ∆c
= .
dx L
The measurement from the bank shows the concentration field to be

fixed in time. If I return to the same point on the river bank tomor-
row, I will get the same measurement. However, this is a straight river
with constant flow from day to day and at every location. In a boat
we would just float right down the middle at constant velocity. If we
take measurements from the boat moving at constant velocity, u, then
from that perspective we would sense an increase in concentration with
respect to time. Since we traverse distance L in a time of ∆t = L/u
then from the boat’s perspective the time rate of change is,
dcboat ∆c ∆c
= =u ,
dt ∆t L
which can be written as,
dcboat dc
=u .
dt dx
The above expression should make sense, the faster we move through
the fixed concentration field the higher rate of change we will see from
the boat’s perspective.
c ∆c
Figure 6.1 Example concentration field, c(x), with a region of linear

change in concentration. The field is considered steady in time. If
you transport yourself through this fixed concentration field from
left to right you will sense an increase in concentration with respect
to time. The rate of change you sense will depend upon the speed
with which you move through this field.
We can also get the same result appealing to the chain rule. The boat
follows a simple path,
xboat (t) = xo + ut
The stationary concentration field, c(x) would be seen from the boat
as a function of time by c(xo + ut). Taking the derivative with respect
to time using the chain rule yields,
dc(x(t)) dc dx dc
= =u
dt dx dt dx
6.3.2 General derivation

To more formally describe change from the perspective of the fluid, we
take the variable a, to be the location of a fluid particle at time t = 0
and r(a, t) is the location of the same particle at some later time t; see
Figure 6.2. If we knew the function r(a, t) we would know everything
about the flow. Think about this for a minute, the notation above is
particle path
t=0
y
a
r(a,t)
Figure 6.2 Schematic of the material point of view. The vector a

describes where the chosen particle is at t = 0. The vector r describes
where the particle is a little later.
perhaps not intuitive at first glance. We are defining everything about

the flow based on tracking where each and every blob of fluid goes.
In this frame, the velocity of a fluid particle is defined as

∂r(a, t)
=v
∂t a=const
The notation above says we take the derivative of r and hold a constant.
It is perfectly valid for us to define the velocity as a function of a,
“material” or as a function of r “space”. Note that this expression is a
vector equation as denoted by the boldface notation.
To demonstrate the material derivative, consider the concentration
field, c that is a function of space and time. Normally, we would think to
write the concentration in our fixed frame as c(x, y, z, t). However, we
could equivalently write the concentration field through the mapping,
c(r(x, y, z, t)) = c(x(a, t), y(a, t), z(a, t), t)
where x(a, t) is the first component of the vector r. While this mapping
seems confusing written as an equation, intuitively the idea is simple. If
I am on the bank of the river sampling the concentration at a particular
point, and you pass through that point on your raft, at the instant you
pass through my point, our measurements must agree.
The derivative with respect to time for a fixed fluid particle is found
6.4 Liebniz and Reynold’s transport theorem 81
using the chain rule,

∂c(r(a, t), t) ∂c ∂c ∂r
= + .
∂t
a=const ∂t ∂r ∂t
Note that v = ∂r ∂t is the same as the fluid particle velocity. In vector
∂c
form, ∂r = ∇c, the gradient of the concentration field. Thus,

∂c(r(a, t), t) ∂c ∂ Dc
= + v · ∇c = +v·∇ c= (6.1)
∂t
a=const ∂t ∂t Dt
The notation D/Dt = ∂/∂t + v · ∇ is commonly used for compactness
and is called the material derivative. The material derivative’s purpose
in life is to tell us the rate of change of something in time from the per-
spective of the fluid. It captures the two effects; the field itself changing
with time and a fluid particle moving about in space where the field
changes from point to point. If we expand into component form we have
Dc ∂c ∂c ∂c ∂c ∂c
= + v · ∇c = +u +v +w .
Dt ∂t ∂t ∂x ∂y ∂z
If you got a little lost in the detail of the 3D derivation, don’t be
dismayed. The 1D example captures all the essential ideas, we are just
generalizing our derivations.
6.4 Liebniz and Reynold’s transport theorem

We can also discuss the material point of view in integral form. To do
so we need the Leibniz rule from calculus. The Leibniz rule allows us
to calculate the derivative of an integral when the limits are changing
with time. In 1D this rule is,
d b(t)
Z Z b
∂f db da
f (t, x)dx = dx + f (t, x = b) − f (t, x = a).
dt a(t) a ∂t dt dt
The three terms represent the change in the integral in time due to the
change in the function itself, the change in the upper limit of integration
and the change in the lower limit of integration. Here f is just some
arbitrary function - this is a mathematical statement.
In three spatial dimensions this theorem changes to;
Z Z Z
d ∂f
f (x, t)dV = dV + f v · ndS. (6.2)
dt V (t) V (t) ∂t S(t)
In this expression instead of using d/dt to denote the rate of change of

the location of the interface, we have used the velocity, v, of the surface
of integration. We use the subscript V (t) on the integral to remind us
that the volume of integration can be changing with time. The above
theorem says that the rate of change of a scalar field inside a volume,
V , is composed of two parts; the change due to the change of f inside
the control volume and the change due to the movement of the surface.
When Equation 6.2 is applied to a material fluid particle, then the
velocity on the surface of integration is equal to the fluid velocity.
Also note that use of the divergence theorem would allow us to write
the second term as a volume integral, namely,
Z Z
d ∂f
f (x, t)dV = + ∇ · (f v) dV. (6.3)
dt V (t) V (t) ∂t
When Equation 6.2 or 6.3 is applied in fluid mechanics it is often called

Reynold’s Transport Theorem. While it can also be derived a few dif-
ferent ways, we will treat it as simply a mathematical theorem from
calculus. The usefulness of this theorem will be seen in the next section
where we consider conservation laws. The connection to the material
derivative D/Dt will become apparent in the next chapter.
6.5 Conservation of mass: the fluid

Now lets consider conservation of mass; we consider the fluid itself and
the dye separately. Remember we will be interested in the limit where
the dye is so dilute it has no impact on the fluid it is dissolved in. In
this limit we just ignore the dye for deriving conservation of mass for
the fluid. The total fluid mass inside a material volume, V(t), is given
as
Z
M= ρdV,
V (t)
where ρ(x, t) is the fluid density field. The subscript V (t) reminds us
that the volume of integration is not fixed in time and may be changing
shape as the flow deforms the material. Our volume V (t) is going with
the flow and following the material for all times. Since our volume
of integration was taken to be a material volume (i.e. it encloses the
same amount of material at all times) and since mass is conserved then
dM
dt = 0. Using the Transport Theorem (Equation 6.3),
Z Z
dM d ∂ρ
= ρdV = + ∇ · (ρv) dV = 0
dt dt V (t) V (t) ∂t
Since the volume in our analysis was arbitrary and since the integral is
always equal to zero then the the integrand must be zero,
∂ρ
+ ∇ · (ρv) = 0. (6.4)
∂t
This equation must be true point-wise in the fluid and is the differential
statement of mass conservation. By rearrangement,
Dρ
+ ρ∇ · v = 0. (6.5)
Dt
This equation says that the density of a fluid particle can only change
if the velocity field has non-zero divergence. Divergence of the velocity
field is a measure of the net rate of fluid flowing into/out of a point.
If the velocity vectors are all pointing to one location, then the flow
must be accumulating mass at that point and the density should be
increasing. If the velocity vectors all point away from a point in the
flow, the density should be decreasing.
For an incompressible flow with constant density we have the simpli-
fied relation,
∇ · v = 0.
This assumption of incompressibility is one that we will use throughout
the course. For a liquid the assumption of incompressibility is a good
one. I challenge you to compress some water to any significant degree.
For a gas the assumption is questionable. It turns out that a rule of
thumb people use is that if the flow velocity is less than about 1/3
of the speed of sound in a gas, then the assumption of incompressible
flow is basically fine. For air this means we can safely treat cases with
flow velocities on the order of 100 m/s or less as an incompressible flow.
Flow around a car is incompressible, but flow around an airplane is not.
Compressible flows have many interesting features that will be beyond
the scope of our first course.
It is also important that we refer to a flow as incompressible and not
the fluid. For the flow of air, we have many examples of incompressible
flow despite the fact that air is quite compressible.
6.5.1 Conservation of mass: dilute species

Now let’s consider conservation of the dilute passive dye. We take a
closed, arbitrary volume moving with the fluid and consider the conser-
vation of a trace dye species within that moving volume. The conser-
vation law states that,
Z Z
d
cdV = − j · ndS,
dt V (t) S(t)
The subscripts V (t) and S(t) remind us that the volume of integration
is not fixed in time and is changing shape as the flow deforms the carrier
fluid. Using the Transport Theorem (Equation 6.3) for the left side of
the equation and the divergence theorem for the right side we obtain,
Z Z
∂c
+ ∇ · (cv) dV = − ∇ · jdV
V (t) ∂t V (t)
We can group all the integrands together and since the volume in our
analysis was arbitrary the integrand must always be zero,
∂c
+ ∇ · (cv) = −∇ · j.
∂t
Substituting in Fick’s law from the previous chapter and assuming a
constant diffusivity,
∂c
+ ∇ · (cv) = D∇2 c. (6.6)
∂t
Now, a new term has appeared in the equation ∇·(cv) which represents
convection of the dye molecules by fluid motion.
Let’s think about this equation for a minute. In the case where there
is no fluid motion, v = 0, we are back to the classic diffusion equation.
In another limit, let’s assume there is no molecular diffusion. As we
discussed earlier in the chapter, molecular diffusion is very slow and so
in many cases ignoring it is not an unreasonable assumption. In that
case the equation would be,
∂c
= −∇ · (cv).
∂t
The product cv is a vector which gives the convective flux or rate that
dye is carried by the fluid per unit time, per unit area at a point. If
you recall that c has units of number of molecules per unit volume,
you can check the units of cv yourself. Now recall that the divergence
operator gives us a measure of whether a vector field is converging or

diverging at a point. If the convective flux vectors are converging at a
point, then the concentration field must be locally increasing in time.
If the convective flux is flowing outward from a point in space, the local
concentration should be decreasing.
By rearrangement, Equation 6.6 could be written as,
∂c
+ v · ∇c + c∇ · v = D∇2 c.
∂t
Substituting in the conservation of mass from the previous section we
can replace the divergence of the velocity field as,
∂c c Dρ
+ v · ∇c − = D∇2 c,
∂t ρ Dt
which can be rewritten using the material derivative as
Dc c Dρ
− = D∇2 c. (6.7)
Dt ρ Dt
This last equation has an interesting interpretation. The material deriva-

tive D/Dt describes the rate of change of concentration from the per-
spective of the fluid. The operator is accounting for convection by fluid
motion. The right hand side is the diffusion term. The term on the
left ρc Dρ
Dt describes the concentration of dye (number per unit volume
of fluid) changing because the carrier fluid itself is being compressed
or expanded. If I have a container containing a fixed number of dye
molecules then I compress the closed contained to decrease the total
volume the dye concentration has increased even though the number
of molecules has not.
For incompressible flow (liquids are really pretty incompressible) the
density doesn’t change and thus the mass transfer equation is simplified
to
Dc
= D∇2 c, (6.8)
Dt
where the equation is “exactly” what we found for diffusive transport
with no fluid motion only we must remember that we have replace the
material derivative with the normal partial derivative with time on the
left side of the equation.
It is also interesting to note that in this form, if there were no molec-

ular diffusion then the equation would reduce to
Dc
=0
Dt
which means that the fluid particle cannot change its concentration,
the fluid particle keeps it’s initial value and the dye just moves it with
the fluid. The equation Dc/Dt = 0 for a non-diffusive dye is at it’s
essence the very definition of the material derivative - something that
is just carried by the flow has zero material derivative. It is this physical
interpretation of this mathematical operator that I want to stress. The
material derivative is simply the change with respect to time from the
perspective of the moving fluid.
6.6 Dimensionless formulation

Let’s consider a specific problem. Consider a solid cylinder of radius R
made of solid “dye” in a fluid flow. The incompressible flow of the fluid is
steady in time. Far from the object the flow has velocity U and near the
object the flow field would be more complex as the fluid skirts around
the cylinder. Assume, though, by some sorcery we have calculated the
complete velocity field, v. Upstream, the fluid approaching the cylinder
has a zero dye concentration. The solid dye as it dissolves sets the
concentration at the fluid/solid interface to be c0 . To solve the problem
for computing the dye concentration, we would state the formulation
as
∂c
+ v · ∇c = D∇2 c,
∂t
where
c(r = R) = C0 , c(r = ∞) = 0.
The initial concentration everywhere at time zero is c = 0. We define

the dimensionless coordinates to be x̃ = x/R and ỹ = y/R. The dimen-
sionless concentration is c̃/C0 and time is made dimensionless by a (for
now) arbitrary constant t̃ = t/t0 . Making these substitutions yields,
C0 ∂c̃ C0 U DC0 2
+ ṽ · ∇c̃ = ∇ c̃.
t0 ∂ t̃ R R2
6.6 Dimensionless formulation 87
a) b)
c) d)
Figure 6.3 a) Velocity field for flow around a cylinder. Concentration

field where fluid enters with concentration of zero on the left and
the cylinder surface sets a concentration in the fluid. b) Pe = 1, c)
Pe = 10, d) Pe = 100.
setting t0 = R/U we get,

∂c̃ D 2
+ ṽ · ∇c̃ = ∇ c̃. (6.9)
∂ t̃ UR
The chosen time scale was the time it takes to convect something the
distance of the cylinder radius. This time scale was a convenient choice
as it made the coefficient in the equation go to 1.
The parameter U R/D is called the Peclet number (pronounced, peck-
lay) and is a dimensionless measure of the strength of convection to
the strength of diffusion. A high Peclet number occurs when the flow is
fast around a large object. A high Peclet number means that the fluid
carries the dye downstream. We might expect at high Peclet number
the dye is mostly concentrated in a wake downstream of the cylinder.
For a small Peclet number diffusion dominates, and in this limit we
would expect the dye to work its way upstream against the incoming
flow. Physically, since D ∼ 10−9 m2 /s for small molecules in water, a 1
mm/s flow around a 1 mm object would have a Peclet number of about
1000. Thus, in water it is only flows truly at the microscale that have
low Peclet numbers.
7
Conservation of momentum in a fluid
The basic conservation law for momentum is developed in this chapter.

The law takes a form different than the law for particles that you may be
used to, but in the end the equations are nothing more than Netwonian
mechanics. The analysis in this chapter does not actually assume we
have a fluid, the laws work equally well for solids or materials that are
between fluid and solid. However, the fact that the fluid moves, flows,
and deforms makes the development of our laws seem different than
you might see in a solid mechanics course. We will want to use our
conservation laws to set up equations that (at least in principle) can be
used to calculate the velocity vector v(t, x, y, z) as a function of space
and time. It should be stressed that the velocity is not the velocity of
individual molecules, but is the averaged velocity for all the molecules
at a point. The velocity is that of the fluid in a continuum sense.
In the last chapter we discussed how fluid motion can carry dye
molecules around in a fluid flow. We applied conservation of mass and
our vector calculus theorems to work out how an equation for mass
being transported by a fluid flow. Now we will apply the same strategy
to conservation of momentum. One difference we find with momentum
transport is that the fluid motion that transports momentum around
also has momentum itself (think about that for a minute). This realiza-
tion introduces a non-linearity to the momentum equation that is the
source of much of the interesting behavior we find in fluid flows and
the source of the real mathematical difficulty with the equations.
Before we proceed, let’s remind ourselves of two important results
from the last chapter. We discussed the Transport theorem, a result
90 Conservation of momentum in a fluid
that is central to the development in this chapter,

Z Z
d ∂f
f (x, t)dV = + ∇ · (f v) dV. (7.1)
dt V (t) V (t) ∂t
It is important to keep in mind that this theorem is a mathematical

result from calculus. We also discussed conservation of mass for a fluid
which we derived as,
Dρ
+ ρ∇ · v = 0, (7.2)
Dt
which for incompressible flows became ∇ · v = 0.
7.1 Transport theorem revisited

Previously we stated the Reynolds Transport Theorem for an arbitrary
function f in Equation 7.1. Lets apply this theorem to some quantity
which is measured per unit mass, b. We would be interested in know-
ing about the behavior of the integral of ρb over a volume. Apply the
transport theorem and simply substituting f = ρb, we obtain,
Z Z
d ∂ρb
ρbdV = + ∇ · (ρbv) dV.
dt V (t) V (t) ∂t
We can expand the derivatives on the right side to write this expression
as
Z Z
d ∂ρ ∂b
ρbdV = b + ∇ · (ρv) + ρ + v · ∇b dV.
dt V (t) V (t) ∂t ∂t
This form is convenient since the first term in parenthesis is zero from
mass conservation, Equation 7.2. Our final result is then,
Z Z Z
d ∂b Db
ρbdV = ρ + v · ∇b dV = ρ dV, (7.3)
dt V (t) V (t) ∂t V (t) Dt
which is useful for subsequent analysis. Note that we made no assump-

tions about incompressibility here, so this relationship is one that is
always true. This relationship is convenient because it already sepa-
rates out conservation of mass.
7.2 Linear momentum

The total linear momentum, p, inside a material volume, V(t), is given
as
Z
p= ρvdV,
V (t)
where ρ(x, t) is the density field and v(x, t) is the velocity vector field.
Note that this is a vector expression as the momentum has an x, y,
and z component. In Newton’s law we know that we need to be able
to take the time derivative of momentum. To take the time derivative
wef can use Equation 7.3, simply replacing b with v,
Z Z
dp d Dv
= ρvdV = ρ dV (7.4)
dt dt V (t) V (t) Dt
This form might make you uncomfortable since this is a vector equa-
tion and we only discussed the transport theorem with scalar functions.
If it does you could proceed with only the x component of the momen-
tum instead,
Z Z
dpx d Du
= ρudV = ρ dV.
dt dt V (t) V (t) Dt
We could repeat the same procedure for py and pz . In working through

the three components, we would return to the compact vector form of
the equation in Equation 7.4.
Note that in our vector notation,
Dv ∂v
= + v · ∇v.
Dt ∂t
The term v · ∇v takes a little explanation. It is probably easiest to

remember what it means if you write the term as (v · ∇)v. Expand the
∂ ∂ ∂
part in parenthesis first, (v·∇) = u ∂x +v ∂y +w ∂z ; which is just the dot
product of the velocity vector and the gradient operator. Now apply
this operator to the velocity vector to get the components of (v · ∇)v.
∂ ∂ ∂
For example, the x component of (v · ∇)v is (u ∂x + v ∂y + w ∂z )u =
∂u ∂u ∂u
(u ∂x +v ∂y +w ∂z ). The entire operator expanded into component form
is a vector,
u ∂u ∂u ∂u
 
∂x + v ∂y + w ∂z
 ∂v ∂v ∂v 

v · ∇v = 
 u ∂x + v ∂y + w ∂z  ,
u ∂w ∂w ∂w
∂x + v ∂y + w ∂z
where the first, second, and third rows are the x, y, and z components
of the vector.
The expression 7.4 allows us to compute the rate of change of mo-
mentum of a material volume. Newton’s laws requires that this vector
must equal the sum of the forces acting on the control volume. Deter-
mining the forces acting on the fluid is not a simple matter and will
take some work. Before we turn to computing forces, let’s make sure
we really understand the idea of acceleration in fluid flows.
7.2.1 Interpretation of acceleration

The material derivative of the velocity,
Dv ∂v
= + v · ∇v,
Dt ∂t
is the expression for acceleration of a fluid particle. The expression says
there are two components to the acceleration in a fluid. The first term
∂v
∂t has to do with the velocity field changing with time. The second
term v · ∇v is acceleration from the perspective of the flow.
Imagine our river again. I go to the river every day to the same
point on the banks and measure the river’s flow velocity. Day to day,
nothing changes and I conclude that at this location that ∂v ∂t = 0. There
is no local acceleration in the flow due to the velocity field changing
with time. However, that does not mean the acceleration that the fluid
feels is zero. Downstream from our measurement point there is a set of
rapids. If I go down the river on the raft, from the perspective of the
fluid I certainly feel acceleration as I speed up and slow down through
the rapids. This acceleration arises from the v · ∇v term. Now imagine
that one day there is a big storm and when I go out to measure the
river, the velocity at my special point has increased. Thus it must be
true that ∂v∂t > 0, at least for some of the time between now and the
previous day. So there are two different types of acceleration, one is due
U(t=0) U(t=dt)
dx
=Udt
x=0 x = dx
Figure 7.1 Schematic of acceleration in a fluid flow from the per-

spective of the fluid. A person riding a raft through a narrowing
river feels acceleration even if the fluid flow is constant when ob-
served from the banks of the river.
to the velocity field changing in time and the other is from the velocity
field changing in space.
We can see that the spatial acceleration must follow a form like v·∇v
in a simple 1D example. Imagine a smooth river with a constant flow
in time. The river has a constant depth but there is a region where the
banks narrow down in width. Since the cross sectional area of the river
has been reduced, the flow velocity must increase in the narrow region,
Now let’s measure the acceleration of the fluid from the perspective
of the flow and the fixed river banks. Let’s assume that I am on a raft
and you are on the river bank. We are going to measure the acceleration
that a raft feels at the same location from our two different perspectives.
Even though we are observing from different points of view, in the end
our measurements should be the same. If I go right down the center of
the river on a raft, I sense the x component of the river flow velocity.
If I want to know my acceleration, I simply measure the raft velocity,
U , at two instances, t = 0 and a short time later, t = dt;
Uraft (t = dt) − Uraft (t = 0)
Acceleration =
dt
Now you are observing from the river bank and want to know the
fluid acceleration at the same point. If you watch me go past you could
measure my acceleration the same way. However, assume there is no
tracer for you to observe in the flow and you are stuck taking point
measurements of the river’s flow velocity. The previous formula for
acceleration is still valid, however from the bank you can only measure
the river velocity as a function of position. The equivalent measurement
of the fluid acceleration would would be,
u(x = dx) − u(x = 0)
Acceleration = .
dt
The distance dx that I would have moved on the raft over time dt is
related to the flow velocity as dx = u dt. Therefore substituting for dt,
u(x = dx) − u(x = 0)
Acceleration = u .
dx
Now the fraction in this equation is the definition of the spatial deriva-
tive of the flow velocity in the limit that dx goes to zero,
∂u
Acceleration = u .
∂x
The expression above is simply the 1D version of v·∇v, the acceleration
of a fluid particle due to changes in the velocity field in space.
7.3 F = m a
In particle mechanics, we usually think of Newton’s Law expressed as
the equation F = m a. However, this is for a constant mass particle
and a more general statement would be
dp
= F,
dt
where p is the linear momentum and F is the sum of the forces acting
on the particle. So we have developed a way to describe the rate of
change of the momentum for a fluid particle. Now we need to work
on the forces. We will account for body forces acting on the whole of
the material volume, fb and surface forces fs acting on the boundary
separately. Generally then, conservation of momentum can be stated
as,
Z
Dv
ρ dV = fb + fs . (7.5)
V (t) Dt
7.3 F = m a 95
7.3.1 Body forces

Obtaining an expression for the body forces is relatively straightfor-
ward. The only body force we will consider in this course is that due to
gravity. The gravitational body force is simply the continuous version
of f = mg, namely,
Z
fb = ρgdV (7.6)
V (t)
We could also consider electrostatic forces (if the fluid were charged
and subjected to an electric field) or Lorenz forces (if the fluid were
passing a current and subject to magnetic fields). These are interesting
topics for another day, however their inclusion into our theory would
not be difficult. If we know how to calculate the force per unit volume,
we only need to integrate this force density over our volume.
7.3.2 Surface forces

Describing the stress along the surface of an arbitrary material volume
inside the fluid takes a little work. First lets define a stress vector, s,
as
s(x, t).
The vector is interpreted as force per unit area at some location on our
surface of our arbitrary material volume. The total surface force that
is exerted on the enclosed volume is then given as,
Z
fs = sdS
Figure 7.2 shows a schematic where we need to add up, or integrate, all
the little surface forces acting over the entire material volume’s surface.
The general description of the state of stress within the fluid requires
us to introduce the idea of a tensor. A tensor has a structure like a ma-
trix. In three dimensions, a tensor is a 3x3 matrix with 9 components. A
tensor is needs to describe the state of stress in a material as there are 9
things that are important. The stress at a location has magnitude and
direction; it is a vector with 3 components. However, what happens to
the material under this force depends upon which face the stress acts.
A stress with only an x component which acts on the surface with the
s
s
s dS
s
s
s
Figure 7.2 The local stress vector, s, acts everywhere on the surface,
S, of an arbitrary region. Integrating that stress over the surface area
gives the net force acting on the surface of the volume.
Figure 7.3 Schematic of a square differential element of fluid. A force

in the x direction deforms the material differently depending on
which face the force acts upon. The element is stretched or sheared,
depending upon which face the force acts.
normal vector in the x direction will stretch a chunk of fluid. A stress

in the same direction but acting on a face with a normal vector in the y
direction will shear a chunk of fluid. This effect is shown schematically
in Figure 7.3. Thus the deformation of the material depends upon the
direction of the force and which face it acts upon. Thus the state of
stress requires more information than can be contained in a vector, so
we go to a tensor. We will follow the convention that tensors are bold-
face capital letters and vectors are boldface lowercase letters. Note that
we could spend a whole course on tensor calculus, thus we are covering
only the highest of high points here.
We can define a stress tensor with nine components that define the
state of stress on an arbitrary element within the fluid. The convention
is
Txz
7.3 F = m a 97
Tyy
Tyx
Tyz
Txy
Tzy
Txx
Tzx Txz
Tzz
Figure 7.4 Schematic the 9 components of the stress tensor, T.
where x is the direction of the normal plane and z is the direction of

the force associated with the stress. The tensor can be represented as,
 
Txx Txy Txz
state of stress at point = T = Tyx
 Tyy Tyz 
Tzx Tzy Tzz
with the components shown schematically in Figure 7.4. Each row of

the tensor is for a different face of the cube and each column is for the
three components of the force.
Just as we have seen scalar fields and vector fields, the stress tensor,
T, is a tensor field, i.e.
T(x, y, z, t).
The relationship between the stress tensor at a point, T, and stress

vector s is,
s=n·T
The dot product projects the tensor onto the surface with a normal
vector n to get the stress vector at that point on the surface. The dot
product of two vectors gives a scalar, while the dot product of vector
and a tensor gives a vector. To remember what n · T means, it is like
the vector-matrix operation,

 
Txx Txy Txz

s= nx ny nz  Tyx Tyy Tyz 

Tzx Tzy Tzz

You can test this yourself by taking a normal vector pointing in the x-
direction, n = [1 0 0]. Taking n · T with this normal vector returns the
first row of T which are the three forces acting on that face. Fortunately,
conservation of angular momentum will require the stress tensor to be
symmetric, thus if you forget the order of multiplying the tensor and
vector, you will get the right answer anyway. I still to this day confuse
(without looking it up) the what components go in the rows and which
go in the columns of T but since the tensor is symmetric I am saved.
Also, you might be tempted to confuse n · T as being Tn (i.e. the usual
matrix vector multiply), and again symmetry of the tensor will give you
the right answer. Since we won’t use the tensor notation much outside
of deriving equations and since this is likely the first class where you
have seen tensors, I’ll cut you some slack on keeping everything straight!
The total force exerted on an arbitrary volume element by the surface
forces, fs , is given as simply the integral of the stress vector around the
closed surface, i.e.
Z Z
fs = sdS = n · TdS.
By the divergence theorem, which also works for tensors just as it does
for matrices,
Z Z
fs = n · TdS = ∇ · TdV. (7.7)
In component form, the divergence of the stress tensor is,

 
Txx Txy Txz
h i
∂ ∂ ∂
∇ · T = ∂x  Tyx Tyy Tyz  =
 
∂y ∂z
Tzx Tzy Tzz
h i
∂Txx ∂Tyx ∂Tzx ∂Txy ∂Tyy ∂Tzy ∂Txz ∂Tyz ∂Tzz
∂x + ∂y + ∂z , ∂x + ∂y + ∂z , ∂x + ∂y + ∂z
Note that the divergence of the tensor returns a vector. So far we

haven’t done anything, really. We have just described a way in which we
7.3 F = m a 99
will compactly describe the internal state of stress of the fluid. That’s
all. This approach is not limited to the mechanics of fluids, but works
equally well in solid mechanics. In solid mechanics, we are often only
concerned with equilibrium and thus the commonly used equilibrium
statement in solid mechanics that ∇ · T = 0.
7.3.3 Meaning of ∇ · T
Above we used the definition of the stress tensor and our vector calculus
to show that the net surface force acting on an arbitrary volume is the
divergence of the stress tensor. We can get the same result using a free
body diagram for a infinitesimal element of fluid. Consider a small cube
that is dx, dy, and dz on each side. Consider the point P in the middle
of the cube. To find the forces on the face of the cube in the x-direction,
we must consider that the stress tensor varies in space.
The value of the stress at one point is related to the value at a
nearby point through a Taylor series. Don’t groan. Students always
groan at Taylor series. Taylor series is your friend. Taylor series says
any complicated crazy function, is just a line as long as you zoom in
enough. The Taylor Series expansion of the stress tensor on the right
face of the square relative to the center is,
dx ∂Txx
Txx + .
2 ∂x
Here Txx is the component of the stress tensor at the center of the cube.
Likewise, the stress on the left hand face of the cube in the x-direction
is,
dx ∂Txx
Txx − .
2 ∂x
A schematic of the x component of the forces is shown for a 2D square

in Figure 7.5.
Considering all the forces acting in the x-direction for a 3D cube, we
Tyx+ ∂Tyx dy
∂x 2
dy
y
x
Txx- ∂Txx dx Txx+ ∂Txx dx
∂x 2 ∂x 2
dx
Tyx- ∂Tyx dy
∂x 2
Figure 7.5 Schematic a square differential element of fluid where

the stress tensor is non-constant is space. Here, we show the x com-
ponent of the forces acting on the 2D differential square.
have

dx ∂Txx dx ∂Txx
dfsx = Txx + dydz − Txx − dydz +
2 ∂x 2 ∂x

dy ∂Tyx dy ∂Tyx
Tyx + dxdz − Tyx − dxdz +
2 ∂y 2 ∂y

dz ∂Tzx dz ∂Tzx
Tzx + dxdy − Tzx − dxdy
2 ∂z 2 ∂z
Which reduces to

∂Txx ∂Tyx ∂Tzx
dfsx = + + dxdydz.
∂x ∂y ∂z
Similarly, we would find

∂Txy ∂Tyy ∂Tzy
dfsy = + + dxdydz,
∂x ∂y ∂z

∂Txz ∂Tyz ∂Tzz
dfsz = + + dxdydz.
∂x ∂y ∂z
Using vector notation,
df s = (∇ · T)dV.
Thus the total force is Z
fs = ∇ · TdV.
and we get the same result as the previous section. The net surface
7.4 So what’s a tensor? 101
force at a point in space is given by the divergence of the stress tensor.

If the stress tensor is constant in space, then there is no net force on
the fluid.
7.3.4 Linear momentum, all together

Taking our definition of the rate of change of momentum and the ex-
pressions for the gravitational and surface forces we obtain,
Z Z Z
Dv
ρ dV = ρgdV + ∇ · TdV. (7.8)
V (t) Dt V (t) V (t)
Since all the integrals are volume integrals over any arbitrary volume
in the fluid, the equation holds at every point,
Dv
ρ = ρg + ∇ · T. (7.9)
Dt
While this looks like a beautifully simple equation, there is a seri-
ous problem. There are too many unknowns. The equation is actually
three equations for the 3 unknown components of the velocity vector.
However, there are 9 components of the stress tensor. If we considered
conservation of angular momentum (we will skip the derivation) we
would find that the stress tensor must be symmetric. Thus there are
only 6 unknown components of the tensor. This is a better situation
but still 6 equations short.
What is needed to close the problem is the same type of relationship
as Fourier’s law provided when we studied heat conduction. We need a
law that relates the unknown forces to the deformation of the material.
This is called a constitutive law and cannot be consider more than an
empirical relationship for the material of interest. The expression in
terms of the stress tensor is general and does not make any assumption
other than that of continuum. It works for fluids, solids, and other
crazy materials. The choice of constitutive law determines whether we
are working with fluids, solids, or something in between.
7.4 So what’s a tensor?

In this chapter we introduced the idea of a stress tensor, which may
seem a bit foreign to you, so it is worth emphasizing a few points again.
A scalar is used for representing things like temperature or pressure

which just have a single value at a point. Temperature is not direc-
tional, it’s just a number. A vector is used for representing things like
velocity or heat flux because the quantity has magnitude and direction.
It matters not just what the velocity is, but what direction the flow is
going. We describe vectors in an orthogonal basis, i.e. i, j, k.
The tensor is used to represent stress because with stress it matters
which direction the force is applied as well as whether that force is
a shear or tensile force. So we need 9 numbers to describe the three
components of stress which can act on three faces of the material. Just
like with vectors, we can project the state of stress onto any other
surface from the stress tensor, but we describe it on the three orthogonal
faces with their normal vectors aligned with the unit vectors. The stress
tensor, represented as a 3x3 matrix, is just a way of describing the
physical state of stress of the material.
Just like vectors are useful for describing a host of physical quan-
tities. Tensors can be used for things other than the state of stress,
though the most prevalent 3x3 tensors come from describing stress and
deformations of materials such as fluids and solids.
8
The Navier Stokes equations
So far, for a continuum we have conservation of mass and momentum

expressed as,
Dρ
= −ρ∇ · v,
Dt
Dv
ρ = ρg + ∇ · T.
Dt
These equations say nothing because we have not defined the stress ten-
sor, T. Defining the stress tensor requires knowledge of the material.
Its definition is essentially empirical or comes from a molecular model
of the material. The stress tensor is not a fundamental law of nature.
The situation is just like the case of heat conduction where our funda-
mental conservation law was written in terms of the heat flux vector
and Fourier’s law was needed to close the problem. The relationship
which defines the stress tensor is our constitutive law.
There are some constraints on what form the tensor must take. For
example, we stated in the previous chapter (without proof) that it is
symmetric due to conservation of angular momentum. It also can only
take limited forms if we want its definition to remain constant under
a change of coordinate systems (we always want this) or if we want it
to be isotropic (we often, but not always want this). We’ll save these
topics for a more advanced course on continuum mechanics and not
discuss them further.
104 The Navier Stokes equations
8.1 Euler’s equation

Euler’s equation (named after Leonhard Euler, one of the greatest
mathematicians in history) assumes a simple form for the stress ten-
sor. Euler’s equation assume that the only stress internal to the fluid
is pressure. Pressure acts only normal to a surface and thus the tensor
only has diagonal elements. The tensor has the simple form,
 
−P 0 0
T= 0 −P 0 
0 0 −P
With this stress tensor and the definition of the divergence we write all
the terms out,
−P
 
0 0
h i
∂ ∂ ∂
∇ · T = ∂x , ∂y , ∂z  0 −P 0 
 
0 0 −P

∂P ∂P ∂P
=− , , = −∇P
∂x ∂y ∂z
Substituting this expression into our conservation law would yield con-
servation of momentum to be,
Dv
ρ = −∇P + ρg.
Dt
Now we have 4 equations (1 mass and 3 momentum) and 5 unknowns,
the density, velocity vector, and the pressure. This is a bit better but
still a problem. The final equation to close the problem must come from
thermodynamics, which relates temperature, pressure, and density. To
completely define the problem we would need to introduce conservation
of energy to our formulation, which we consider energy later.
However, if the flow is incompressible, we do not need to consider
energy or thermodynamics. For incompressible flow, the density is con-
stant and taken as a material property. For an incompressible flow,
Euler’s equations reduce to,
∇ · v = 0, (8.1)
Dv
ρ = −∇P + ρg. (8.2)
Dt
8.1 Euler’s equation 105
where the density is taken as a known and constant material property.

Finally, we have 4 equations and 4 unknowns (velocity vector and pres-
sure) and a closed set of equations. To solve the equations we need
initial and boundary conditions. The commonly used boundary equa-
tion at the interface between a fluid and solid in Euler’s equation is
that the velocity normal to the surface must be zero; i.e. fluid cannot
go into a solid surface.
The above formulation looks pretty good though there are some com-
plexities hidden in that equation. Too bad that the Euler equations turn
out not to be very useful. The reasons for Euler’s equations to be of
limited value are subtle and not obvious. We will discuss their failing
in more detail as we progress.
The main feature of the constitutive law is that Euler’s equation as-
sumes zero viscosity. It seems in many cases the force due to viscosity
should be quite small and neglecting it in the equations doesn’t seem
like a bad approximation. For now, you can comfort yourself in that
the failings of Euler’s equations to provide many predictions of practi-
cal utility stumped many smart people for many years. About 100 years
spanned between Euler’s development with no viscosity and the devel-
opment of the Navier-Stokes equations with viscosity. It took perhaps
another 50 years for to truly understand and appreciate why Euler’s
equations can fail as an approximation even when the viscosity is small.
If some of these points seem difficult, you are in good company.
Before we get too dismissive of Euler’s equations, we will see that
they can be used for simple calculations, estimations, and qualitative
explanations of flow behavior. So Euler’s equations still have a lot of
value and we will discuss them a lot in coming chapters.
Finally, there is an interesting subtlety when we made the incom-
pressible assumption, pressure loses its meaning as a thermodynamic
quantity. In an incompressible flow, information propagates infinitely
fast. If I start to move, the flow adjusts everywhere instantaneously.
In reality, information in a fluid propagates at the speed of sound.
Thermodynamics connects the speed of sound to the thermodynamic
function, ∂P/∂ρ. Since the change in density, dρ is zero, then the speed
of sound is infinite. However, thermodynamics is now unhappy with
our definition of pressure. None of this is important for our discussion,
its just interesting.
8.2 Newtonian fluid & incompressible flow

In fluid mechanics the most common constitutive relationship that ac-
counts for viscosity is that of the Newtonian fluid. The law works well
for air, water, oils and other simple fluids. It does not work well for a
variety of complex fluids such as visco-elastic or polymeric liquids. The
experimental observation was discussed in the opening chapter. A sim-
ple device for quantifying viscosity comprises two concentric cylinders
with a very narrow gap between them. A fluid of interest is placed in
the gap. The inner cylinder is held fixed while the outer one rotates at a
constant angular speed, ω. The speed of the wall at the outer cylinder is
simply U = ωR, where R is the radius of the cylinder. If we conducted
this experiment for different gap sizes d and different speeds, we would
find the torque needed to spin at constant speed would follow a law,
U
Torque ∝ µ
d
In a more general form, the Newtonian assumption relates the shear
stress to the velocity gradient in the y-direction through the fluid vis-
cosity, µ,
∂u
τ =µ .
∂y
This is a very important concept to commit to memory: A Newtonian
fluid relates stress to velocity gradients.
The general form of the Newtonian stress tensor is stated as,

2
T = − P + µ∇ · v I + µ ∇v + ∇vT .

(8.3)
3
Here, the superscript T denotes the transpose and I the identity matrix.
You can start to notice a few things. First the pressure term is just
like in Euler’s equation, if I set the viscosity, µ, to zero then the tensor
limits to that in Euler’s equation. The other term added to the pressure
is the proportional to the divergence of the velocity field. Recall the
divergence of the velocity is a scalar. This term has to do with viscosity
resisting changes in volume of the fluid. We will typically only deal with
incompressible flows in this course and we can use the incompressible
form,
T = −P I + µ ∇v + ∇vT .

(8.4)
8.2 Newtonian fluid & incompressible flow 107
The terms with the velocity gradients is a little more complicated and
will be the subject of much discussion later in this chapter. Since ve-
locity is a vector, it’s gradient is a tensor. The velocity gradient tensor
is written out as,
 ∂u ∂v ∂w 
∂x ∂x ∂x
 
∂u ∂v ∂w
∇v = 
 
∂y ∂y ∂y 
 
∂u ∂v ∂w
∂z ∂z ∂z
By adding the velocity gradient tensor to it’s transpose we obtain a

symmetric tensor. If we worked through the details, we would find we
could express the divergence of this stress tensor as,
∇ · T = −∇P + µ∇2 v.
Substituting into the balance of linear momentum yields,

∂v
ρ + v · ∇v = ρg − ∇P + µ∇2 v. (8.5)
∂t
Coupled with mass conservation for an incompressible flow, our prob-
lem is closed;
∇ · v = 0. (8.6)
Equations 8.5 and 8.6 represent four equations for four unknowns; the
three components of velocity and the pressure. These equations are
known as the incompressible Navier Stokes equations and comprise the
mathematical foundation for describing fluid flow.
Just to be clear on the notation, if we expand the momentum equa-
tion in component form for a Cartesian coordinate system we obtain,
2
∂ u ∂2u ∂2u

∂u ∂u ∂u ∂u ∂P
ρ +u +v +w = ρgx − +µ + +
∂t ∂x ∂y ∂z ∂x ∂x2 ∂y 2 ∂z 2
∂2v ∂2v ∂2v

∂v ∂v ∂v ∂v ∂P
ρ +u +v +w = ρgy − +µ + +
∂t ∂x ∂y ∂z ∂y ∂x2 ∂y 2 ∂z 2
∂2w ∂2w ∂2w

∂w ∂w ∂w ∂w ∂P
ρ +u +v +w = ρgz − +µ + +
∂t ∂x ∂y ∂z ∂z ∂x2 ∂y 2 ∂z 2
The great thing about expressing the formulation in our vector calculus
notation is that we can change coordinate systems and simply look
up the operators in different coordinates. The most common alternate

coordinate systems are cylindrical and spherical.
Note that for incompressible flows where the density is constant,
conservation of momentum can be written as,

∂v
ρ + v · ∇v = −∇(P + ρgz) + µ∇2 v.
∂t
where (without loss of generality) we aligned the coordinate system
such that gravity is pointing in the negative z direction. We can intro-
duce the dynamic pressure as Pdyn = P + ρgz and rewrite the equation
as
∂v
ρ + v · ∇v = −∇Pdyn + µ∇2 v,
∂t
The dynamic pressure is the pressure that results from fluid motion
and subtracts out the hydrostatic component.
This form is convenient as it allows us to essentially ignore the effect
of gravity when computing fluid motion in a constant density fluid.
From the above equation we can see that if there is no fluid velocity
and we are at equilibrium then ∇Pdyn = 0. At equilibrium the dynamic
pressure (pressure due to fluid flow) is constant and does not impact the
flow. At equilibrium the measured pressure is given as ∇(P + ρgz) = 0,
meaning the pressure is only hydrostatic. In constant density flows,
gravity does not play a role in determining the flow field, but it does
effect the pressure. If the density is constant then gravity acts equally
on the whole fluid and thus there is no flow driven by gravity. If the
density of the fluid is not constant, then we could have buoyancy driven
flows.
8.2.1 So what does viscosity do?

If we look at the term in the equation that has the viscosity coefficient
in front of it, we might see something reminiscent of the heat equation.
In fact if we find a special class of constant density flows (which we will
discuss in the next chapter) which have no spatial acceleration, and
no pressure gradients, then we would have flows which are governed
µ 2
by the heat equation, i.e. ∂v ∂t = ρ ∇ v. Since this limited form of the
momentum equation looks like the heat equation, it will act in the
same way. Viscosity will act to smooth out the flow and “kill” spatial
8.3 Boundary conditions 109
curvature in the velocity field. While the full Navier Stokes equations
have other more complicated terms, we can think that viscosity will
always act to smooth the velocity field and pull the velocity to some
uniform or simple state - just like the heat equation.
The ratio of µ/ρ appears often that it is given its own symbol,
ν = µ/ρ. This parameter is called the kinematic viscosity and has units
of m2 /s. The material property, µ, is properly called the dynamic vis-
cosity. The kinematic viscosity plays a role like the thermal diffusivity,
α, or the diffusion coefficient D. In thermal problems, α, sets the rate
that heat propagated over some distance; the kinematic viscosity plays
a similar role. The kinematic viscosity sets the rate that momentum is
transferred from one sliding fluid layer to another.
8.3 Boundary conditions

In order to solve the Navier Stokes equations, boundary conditions are
needed. Boundary conditions tell you how the region of interest where
your equations are valid are coupled to the rest of the universe. When
there is a solid surface, we typically assume that all components of the
fluid velocity equal the velocity of that solid surface. For a stationary
object, the velocity vector goes to zero, v = 0. This statement means
that there is no flow through the solid surface (no normal velocity)
and no-slip at the surface (no tangential velocity). The condition of
no flow through the solid seems obvious. The no-slip condition, on the
other hand, is not a fundamental law with a very strong theoretical
backing. What boundary condition to use was a serious debate in the
fluid mechanics world up to the early 1900s. Today, many people still
study whether the condition is strictly true or only approximately true.
Recently, there have been experiments with superhydrophobic surfaces
with significant amounts of fluid slip in liquid systems. The no-slip con-
dition can also be violated in rarefied gas flows where the length scale
of the object is not that large compared to the molecules mean free
path. However, for most practical cases it is experimentally observed
that the no-slip condition is obeyed to a very high degree of approxi-
mation. At a free surface of liquid, we can also have a condition of no
stress. However, in this course we will usually consider problems of flow
around some object and thus will use the condition that v = 0.
u
n = [1 ]
0
y
x
Figure 8.1 A solid surface with no slip is oriented with a normal

vector in the vertical direction.
8.4 Computing stress from flows

In many cases we want to know the total force acting on an object
immersed in a fluid flow. If we know the velocity field we have a way of
computing the stresses at a solid surface from the stress tensor. To make
our lives simpler, lets consider a 2D world. Consider the solid surface
and we have aligned our coordinate system such that the normal vector
for that surface points only in the y direction n = [0, 1]; see Figure 8.1.
Assume that we have through some sorcery solved the Navier Stokes
equations for the full velocity field above that surface. Now we want to
use that solution to calculate the force exerted on a body. In a 2D world,
the Newtonian stress tensor in component form would be denoted as,
" #
2 ∂u ∂u ∂v

P 0 ∂x ∂y + ∂x
T=− + ∂u ∂v ∂v .
0 P ∂y + ∂x 2 ∂y
The stress vector at that surface would in general be,

" " ##
2 ∂u ∂u ∂v

P 0 ∂x ∂y+ ∂x
s=n·T= 01 − +µ ∂u ∂v ∂v .
0 P ∂y + ∂x 2 ∂y
However, at our surface the vertical velocity v is zero everywhere on

the surface (it is impenetrable). Therefore, ∂v/∂x is zero everywhere
on the surface. Also the horizontal velocity is zero everywhere on the
surface by the no slip condition, thus ∂u/∂x = 0 along the surface.
Conservation of mass for an incompressible flow states that

∂u ∂v
+ = 0.
∂x ∂y
At the surface, since ∂u/∂x = 0 then ∂v/∂y = 0. Using all the relations
above which hold at the surface the problem simplifies to
" " ##
0 ∂u

P 0
+ µ ∂u ∂y

s=n·T= 0 1 − .
0 P ∂y 0
Thus the stress vector at our point simplifies to,

h i
s = µ ∂u
∂y , −P
The result says that the x component of the stress is given by the
horizontal velocity gradient, just as we described when defining a New-
tonian fluid. This is the force that would cause a fluids viscosity to drag
a solid along with the flow. The y component of the stress is given by
the pressure.
8.5 Comments on kinematics

While the Navier-Stokes equation may look beautiful (or frightening),
our derivation based on the stress tensor might seem arbitrary. What
does the expression for the stress tensor mean and why should the stress
depend upon the velocity gradient in the following way,
T = −P I + µ ∇v + ∇vT .

(8.7)
In this section we will provide some interpretation of the velocity gra-
dient and the velocity gradient tensor. The development here follows
that of Kundu and Cohen (2004) quite closely.
However, we first need to emphasize again this Newtonian fluid con-
stitutive law is not a “law”. To derive it you need a molecular model
of the material. This can be done for simple gases, but not many other
fluids. So, in the end, we can consider this an empirical relationship
and nothing more. The general form of the tensor does have some con-
straints placed by more fundamental laws. For example, conservation of
angular momentum demands the tensor to be symmetric. Requiring the
tensor to not depend on our coordinate system places restrictions on
the form. However, in the end the constitutive relationship is something

that works well for many fluids, but it is not a law of nature.
8.5.1 Linear rate of strain

Consider a line element in a steady flow which is oriented in the x
direction. A schematic is shown in Figure 8.2. Lets assume there is
only an x component of the velocity field at this location. The element
is dx in length. The left point, A, at time zero, t = 0 is at position
XA (0). The right point B is at position XB (0). The positions XA and
XB are a material points that go with the flow and are only functions
of time. A short time later, dt, the material point A has gone with the
flow and moved to position,
XA (dt) = XA (0) + u(XA )dt.
For simplicity, the velocity field is only a function of space. The point
B has moved to position,
XB (dt) = XB (0) + u(XB )dt.
The distance XB − XA = dx is meant to be infinitesimally small. The
strain is defined as the change in length divided by the initial length,
thus
(XB (dt) − XA (dt)) − dx
Strain =
dx
XB (0) + u(XB )dt − XA (0) − u(XA )dt − dx
=
dx
(XB (0) − XA (0) − dx) + u(XB )dt − u(XA )dt
=
dx
u(XB )dt − u(XA )dt
=
dx
Therefore, in terms of a rate equation,
d Strain u(XB ) − u(XA )
= .
dt dx
The velocity at point B is related to A through a Taylor Series, since
the distance is assumed small. The expansion for the velocity at point
XB = XA + dx is
∂u
u(XB ) = u(XA ) + dx.
∂x
t=dt XA(dt) XB(dt)
U(XA) U(XB)
t=0
XA(0) XB(0)
t
x
Figure 8.2 Schematic of a material line in a 1D flow. The points A

and B are subjected to different velocities and thus move different
distances in one instant of time, dt.
Therefore the rate of change of linear strain is related to the velocity

gradient,
d Strain ∂u
= .
dt ∂x
The velocity gradient, ∂u∂x , provides the linear strain rate that a ma-
terial line elongates or shrinks. This makes sense. Imagine you and a
friend are in hot air balloons up above the earth. Since the length scale
of the earth is so big we can take dx at the human scale to be infinites-
imal. If the distance between you and your friend is growing, it means
there are velocity gradients in the flow. Your friend, even though close
by, is subjected to a different fluid velocity than you, thus you separate
from each other. The rate that you separate is a measure of the flows
velocity gradient.
8.5.2 Shear rate of strain

Now imagine a more complicated fluid motion in 2D. We have velocity
(and velocity gradients) in both the x and y components. Now we will
consider the motion of three material points, O (the origin), A, and
B. As in the previous section, the location of the material points are
only functions of time. However, since we are in 2D we need to consider
the motion in the plane, i.e. follow (XA , YA ), (XB , YB ), and (XO , YO ) .
The velocity field is taken as steady for simplicity of notation. Similar
dt UA
A
α
dt VA t=dt
A
β
o B
t=0 dt VB
dt Vo
o B
y dt Uo
dt UB
x
Figure 8.3 Schematic of material points O,A, and B in a 2D flow.

The points are subjected to different velocities and thus move dif-
ferent distances in one instant of time, dt. The sides of the square
are dx in length.
to the previous section, each point moves by an amount,
Xi (dt) = Xi (0) + u(Xi , Yi )dt; and Yi (dt) = Yi (0) + v(Xi , Yi )dt.
Here, the index i refers to point A, B or O. The above equation simply

says the points move in the flow and we update their position using
Euler integration in time. A schematic is shown in Figure 8.3. The
three points move to new locations due to the different fluid velocities.
Using these relations and a little bit of geometry, you can find the
tangent of the angles β and α in terms of the velocities at points A, B,
and O as,
(vB − v0 )dt
tan(β) = ,
dx + (uB − u0 )dt
(uA − u0 )dt
tan(α) = .
dx + (vA − v0 )dt
You can then use the Taylor Series to relate the velocities at point
A and B to that of point O, i.e. uB = u0 + ∂u∂x dx. Substituting the
Taylor Series results and taking the limit, these geometric relationships
become rate equations for the rate of change of the angles,
dβ ∂v dα ∂u
= ; = .
dt ∂x dt ∂y
Note that in the limit of small angles which occur over small times,
tan(α) ≈ α. Thus the sum of these two rates,
dβ dα ∂v ∂u
+ = + ,
dt dt ∂x ∂y
is the rate the angle between the two lines OA and OB close or open
up. This is the shear strain rate.
This relationship makes sense. Hold your hands together in front of
you, pointing upwards. Sliding your hands up and down relative to each
other creates shear. The right hand has a different vertical velocity than
∂v
the left hand. Therefore ∂x is not zero. Holding your hands horizontal
creates velocity gradient as ∂u/∂y. The shearing motion of your hands
would act to deform a material held between them.
8.5.3 Rotation
Using the results from the prior section we can also ask what is the
difference of the rate of change of the two angles. This difference is
twice the rate of solid body rotation of the line elements at that point,
dβ dα ∂v ∂u
− = − .
dt dt ∂x ∂y
In two dimensions, you may recognize this quantity is the same as the
curl of the velocity, which is referred to in the fluid mechanics world
as vorticity, ω. The curl is denoted by ω = ∇ × v. In component form,
the vorticity vector is
 
∂w ∂v
∂y − ∂z
ω = ∇ × v =  ∂u ∂w 
∂z − ∂x 
.

∂v ∂u
∂x − ∂y
The curl is the vector cross product of the gradient operator ∇ and the
velocity vector v. In two dimensions where the velocity is constrained
to the x-y plane, the vorticity only has a component normal to this
plane, the z plane.

∂v ∂u
ωz = − .
∂x ∂y
The curl of the velocity is equivalent to twice the angular velocity of
the fluid particle at a point. While we only made the argument in 2D,
the argument holds up in a more general 3D flow.
8.5.4 Generalized to 3D
In three dimensions, the results of this section are summarized as fol-
lows. The strain rate tensor is defined as
1
∇v + ∇vT .

S= (8.8)
2
In component form this tensor is
 
2 ∂u
∂x
∂u
∂y + ∂x
∂v ∂u
∂z + ∂x
∂w
1
S =  ∂u ∂y + ∂x
∂v ∂v
2 ∂y ∂v ∂w 
∂z + ∂y 
2 ∂u ∂w ∂v ∂w
∂z + ∂x ∂z + ∂y 2 ∂w
∂z
The strain rate tensor is symmetric and provides the rate of strain
of fluid elements in the different directions. The diagonal components
of the tensor are the linear strain rates in the three dimensions. The
off-diagonal components are the shear strain rates in the three planes;
x − y, x − z, and y − z.
To rotation tensor is defined as
 
0 −ωz ωy
1
R =  ωz 0 ωx  (8.9)
2
−ωy −ωx 0
where ω are three components of the vorticity vector, ω = ∇ × v.
The off diagonal elements of the rotation tensor provide the amount
of rotation in the three planes x − y, x − z, and y − z. The tensor is
anti-symmetric. It is fairly easy to see that the velocity gradient is the
sum of the strain rate and the rotation tensors,
∇v = S + R.
Velocity gradients lead to both strain of material elements and rotation
of those elements. The strain rate tensor provides the rate of strain of
the material. The rotation tensor provides the amount of rotation.
8.6 Non-dimensionalization 117
Using the above tensors, the Newtonian constitutive law for an in-
compressible flow states that
T = P I + 2µS.
The forces exerted on the fluid are proportional the strain rate, but not
the rotation of the fluid.
Please note that this discussion of kinematics is quite brief and more
detailed derivations can be found in the references. The main point is
to demonstrate that the form of the Newtonian stress tensor is con-
nected to the kinematics of rate of deformation of the material. The
stress is related to the velocity gradients. However, the use of the sym-
metric strain rate tensor subtracts our solid body rotation which does
not generate stress. In a Newtonian fluid only the rate of deformation
generates stress.
8.6 Non-dimensionalization
Consider the Navier-Stokes equations for incompressible constant den-
sity flow,

∂v
ρ + v · ∇v = −∇P + µ∇2 v,
∂t
∇ · v = 0.
Let’s say the geometry of the problem provides some characteristic
length, L, and some characteristic flow velocity, U0 . The equations can
be made dimensionless following the same procedure as in previous
chapters. Let’s define ṽ = v/U0 , [x̃, ỹ, z̃] = [x, y, z]/L, t̃ = tU0 /L,
and P̃ = P/P0 . Our time scale is picked so that it takes one time
unit for a fluid particle to travel the length L; i.e. our dimensionless
time is t̃ = tU/L. When making equations dimensionless, it is always
important to recall that derivatives have units. The derivative, ∂/∂x has
units of 1/Length, for example. Therefore, when we make our operator
∇ dimensionless, we make the substitution ∇ ˜ = L∇. The dimensionless
˜
version of the Laplacian is given as ∇ = L2 ∇2 . Making this change of
2
variables would yield,

ρU02 ∂ṽ

˜ P0 ˜ µU0
+ ṽ · ∇ṽ = − ∇ P̃ + 2 ∇2 ṽ,
L ∂ t̃ L L
˜ · ṽ = 0.
∇
Rearranging the momentum equation and defining P0 ≡ ρU02 we would
obtain,

∂ṽ ˜ ˜ P̃ + µ ∇2 ṽ,
+ ṽ · ∇ṽ =∇
∂ t̃ ρU0 L
˜ · ṽ = 0.
∇
Thus the problem has only one dimensionless parameter that matters,
the Reynolds number,
ρU0 L
Re ≡ .
µ
The Reynolds number came out in several of our problems early in this
book when we discussed dimensional analysis.
8.7 The Reynolds number

From the non-dimensionalization the Reynolds number shows up as the
parameter upon which the flow field depends. This parameter showed
up in the first chapter where we showed that it could be used to scale
all the drag coefficient and pipe flow data onto a single master curve.
The Reynolds number shows up as the most important parameter in
the Navier-Stokes equations. The Reynolds number is the parameter
when it comes to determining the character of incompressible flow.
Physically, the Reynolds number is the ratio of inertia to viscosity.
When the Reynolds number is large, then inertia dominates and when
it is small then viscosity dominates. Often in practice high Reynolds
number flows are unstable and subject to turbulence. Low Reynolds
number flows are smooth. What counts as a “high” or “low” Reynolds
number often depends on the situation.
The interesting thing is that since the flow only depends upon the
Reynold number, a model experiment is perfectly valid if you match
the Reynolds number. If I am interested in the character of the flow
around a tall building, I can create a small model, match the flow in
my wind tunnel to have the same Reynolds number as the thing I am
interested and I will get the same flow.
Just to get a feel for what the magnitude of the Reynolds number
8.8 Summary 119
is, lets consider a few hypothetical cases. For water, the density is ρ =
1000 kg/m3 and the viscosity is approximately µ = 0.001 Ns/m2 . So
lets say a person 2m tall who can swim at 1 m/s would have a Reynolds
number of 2 × 106 . A small organism that is 100 µm and can swim one
body length per second would have a Reynolds number of 0.01. For
you to match the Reynolds number of this organism you would need to
swim at about the same speed (100 µm/s) in molasses!
In a pipe flow, a standard rule of thumb is that the flow will transition
from laminar to turbulent at a Reynolds number of 2300. Here the
Reynolds number is defined on the diameter of the pipe. To get a sense
of where a Reynolds number of 2300 is - water flowing in a 1/4 inch pipe
(6.35 mm) at a velocity of 1 ft/s (0.3 m/s) we would be at a Reynolds
number of about 2000. The transition number is only a rule of thumb. If
you try this experimentally you will find the transition number changes
based on how carefully you set up the experiment. If you just hook up
some tubing and push water through it without worrying about keeping
everything smooth, straight, and well conditioned, you will probably
observe a transition at a much lower Reynolds number, probably under
1000. If you are extremely careful and take great pains to do a perfect
experiment the flow will transition at a much high Reynolds number. In
2011, the best theoretical determination of a true transition Reynolds
number to sustained turbulence was determined to be Re = 2040. While
a seemingly simple problem, the details are quite complex.
Reynolds original 1885 experimental apparatus survived at the Uni-
versity of Manchester, where he was a professor. A century later, his
exact experiments were recreated with his apparatus. The experiment
showed a transition Reynolds number much lower than he reported.
The difference was found to be due to cars in modern day Manchester,
the small amount of shaking from the streets perturbed the simple but
delicate experiments.
8.8 Summary
The starting point for much of our discussion in future chapters are the
Navier-Stokes equations for incompressible constant density flow,

∂v
ρ + v · ∇v = ρg − ∇P + µ∇2 v,
∂t
∇ · v = 0.
These equations are the formal statement of conservation of mass and

momentum (i.e. Newton’s laws of motion). The equations in this form
assume a Newtonian fluid constitutive law which relate the instanta-
neous rate of strain to stresses. Not all fluids follow the Newtonian
assumption but it is a good model of simple fluids such as air, water,
and oil. To achieve Euler’s equations simply take the limit that the
viscosity is zero and remove that term from the equations. This zero
viscosity limit is a funny one which we will address in a later chapter.
In the past two chapters we worked through the derivation of these
equations. The discussion of the past two chapters was not as complete
or detailed as can be found in many excellent textbooks. The aim of my
discussion was to try to give you a sense of where the equations come
from and how to think about the physical meaning of the different
terms. At this point in your study, I would expect that many of the
details seem at least somewhat confusing.
Of course the equations on their own don’t mean much. We will need
to understand the solution of these equations in order to connect to
physical phenomena. Even though you now have a complete mathe-
matical formulation for fluid flow in hand, your study (and mine) of
the subject is just beginning. To move forward in our understanding of
fluid flow we need to learn how to use these equations to explain phe-
nomena that we observe everyday. As explained previously, the Navier
Stokes equations are very difficult to solve, so we will need to use sim-
ple solutions to give us insight to the fundamentals and learn to use
the equations to guide our qualitative thinking. There is no one better
to make this point than Richard Feynman in his famous Lectures on
Physics. Italicized words are ones I changed to switch the context from
electromagnetism to fluid dynamics;
Mathematicians, or people who have very mathematical minds, are often led
astray when “studying” physics because they lose sight of the physics. They
say: “Look, these differential equations - the Navier-Stokes equations - are
all there is to fluid dynamics; it is admitted by the physicists and engineers
that there is nothing which is not contained in the equations. The equations
are complicated, but after all they are only mathematical equations and if
I understand them mathematically inside out, I will understand the physics
inside out”. Only it doesn’t work that way. ... They fail because the actual
physical situations in the real world are so complicated that it is necessary to
8.8 Summary 121
have a much broader understanding of the equations. What it means to really

understand an equation - that is, in more than a strictly mathematical sense
- was described by Dirac. He said: “I understand what an equation means if
I have a way of figuring out the characteristics of the solution without actu-
ally solving it”. A physical understanding is a completely unmathematical,
imprecise, and inexact thing, but absolutely necessary for a physicist and
engineer.
My hope with the subsequent chapters is to begin to help you “un-

derstand” the Navier-Stokes equations in this imprecise and inexact
sense.
9
Solutions to the Navier-Stokes equations
The Navier-Stokes equations are hard to solve. A general solution or

technique for all problems does not exist. Solutions to flow problems are
an active area of current research and fluid dynamics has continually
driven a large number of techniques in applied mathematics going back
all the way to the original work of Euler. There is no way we can
cover many of the varied mathematical techniques for analyzing fluid
mechanics in a first course. Going much further than we do in this text
would require a stronger background in partial differential equations,
a challenging topic in its own right. Computationally, the situation is
perhaps equally challenging. We can use modern computational fluid
dynamics (CFD) packages to solve numerous flow problems, however
in many cases these solutions must be approached with caution.
But do not despair. There are some simple (and practical) cases
where the Navier Stokes equations can be easily solved with hand cal-
culations. These simple problems are insightful as they can lead to
physical intuition and help connect the mathematical formulation to
observable phenomena. We will show some examples in this chapter of
some classical flow problems. It is important to realize that the Navier
Stokes equations are non-linear equations and their solutions are not
unique. If you find a solution, it might not be the one you observe in
the lab. Often the simple mathematical solutions turn out to be un-
stable ones. While these facts might seem annoying to you, it is these
facts exactly that give fluid flows some of the their beautiful patterns
and makes the subject still interesting even though it has existed in its
modern form for over a century.
In this chapter, we will ignore all these complexities for now and
124 Solutions to the Navier-Stokes equations
PHi u(y) H PLow

y
x
L
Figure 9.1 Schematic for Poiseuille flow between two parallel plates.
solve some simple problems. In all the examples that follow we will
study flows where there is no acceleration and force due to viscosity
will balance some driving force.
9.1 Flow between parallel plates - Poiseuille flow

The following problem is often referred to as Poiseuille flow after the
French physicist and physiologist who experimentally studied the flow
of liquid in narrow tubes to understand the flow of blood in the human
body. We will work this problem in great detail and then use what we
learned to quickly solve other related problems. We will consider here
the two-dimensional version of the problem. Imagine a flow in 2D slot
between solid walls with a steady, pressure driven flow. The channel
height is H and we will take y = 0 to be the lower wall. Gravity will
point in the negative y direction. We will assume that the channel is
infinite in the axial x-direction. We assume the flow is not only steady
in time and also flow is “steady” in the axial direction meaning that
since the channel is long and the flow has no variation in the x direction.
Since the channel is infinite in x we assume that there is no reason the
flow at any x location will be different than any other. Thus the axial
gradients are zero; ∂u/∂x = 0 and ∂v/∂x = 0. Making the assumption
of uniform, steady flow, we can rewrite the Navier-Stokes equations,
crossing all the terms with x derivatives and time derivatives as follows,
!
∂ 2
u ∂2u

∂u ∂u ∂u ∂P S
ρ S +u@ +v =− +µ 2
+ 2
∂t
SS ∂x
@@ ∂y ∂x ∂xSS ∂y
!
∂ 2 ∂2v

∂v ∂v ∂v ∂P S v
ρ A +u
@ +v =− − ρg + µ 2
+ 2
∂t
AA ∂x
@@ ∂y ∂y ∂xSS
∂y
+ ∂v = 0.
S∂u

∂x
S ∂y
S
Note that we keep the axial pressure gradient as the pressure drives the
flow and must drop continuously as we move down the channel.
From conservation of mass, ∂v/∂y = 0; an equation which states that
the vertical velocity is a constant in the y direction. Since v = 0 at the
wall, v = 0 everywhere. We can now cancel out all terms that have the
vertical velocity, v in them;
!
2 2

∂u ∂u ∂u ∂P S∂ u ∂ u
ρ S +u
@ +v =−
S +µ +
2
∂t
S
S ∂x
@@ ∂y SS ∂x ∂xSS ∂y 2
!
2 2

∂v ∂v ∂v ∂P ∂
S v S∂ v

ρ A +u
@ +v =−
S − ρg + µ +
2 S2
∂t
AA ∂x
@ ∂y
@ SS ∂y ∂xSS
∂y S
It makes sense that the left side of the Navier Stokes equations is zero.
The left side represents the acceleration of a fluid particle. There is
no acceleration in this system. There is neither acceleration due to
unsteadiness (change in velocity with respect to time) nor from the
perspective of a fluid particle. The channel is like a calm river whose
width doesn’t change; if you were on a raft you would translate at
constant velocity. Neither the observer on the raft or on the banks
would see any acceleration in this flow.
Thus, the momentum equations are simplified significantly to
∂P ∂2u
= µ 2. (9.1)
∂x ∂y
∂P
= −ρg. (9.2)
∂y
Integration of the second equation yields,
P = P (x, y = 0) − ρgy.
where P (x, y = 0) is the pressure along the lower wall, which is a func-
tion of x only. When we take the pressure gradient in the x direction,
the hydrostatic component (the term with ρg) does nothing to modify
the axial flow. Since the flow was assumed invariant in x, the only con-
sistent solution would be that the axial pressure gradient, dP/dx, is a
constant. Since the left side of Equation 9.1 is a constant the equation
can be easily integrated twice with respect to y,
dP 1 2
u(y) = y + C1 y + C2 .
dx 2µ
We are left with two constants of integration which are determined by
the no-slip boundary conditions at both walls. For the lower wall we set
u(y = 0) = 0. Substituting in this boundary condition states that the
constant C2 must be zero. The upper wall condition of u(y = H) = 0
allows us to determine C1 . The final result is,
dP 1
u(y) = − y(H − y).
dx 2µ
The velocity profile follows a simple parabola which is zero at y = 0
and y = H. Recall from our discussion on diffusion that the second
derivative in space is equivalent to the curvature. Therefore, the gov-
erning equation states that the curvature of the x component of the
velocity field is proportional to a constant (the pressure gradient). The
parabola is the only function which has constant curvature, thus we
could obtain the qualitative shape of the velocity profile without even
solving the equation. RH
The total flow, per unit width is Q = 0 u(y)dy. Integrating our
velocity profile gives the total flow as
dP H 3
Q=−
dx 12µ
Usually we use the fact that dP/dx is the same as the total applied
pressure difference divided by the total length of channel. We need to
be careful with the signs. Convention would usually define a positive
pressure difference, ∆P , as going from high pressure on the inlet to low
pressure on the outlet. With this convention, dP/dx = −∆P/L. Using
the total pressure drop notation gives,
12µL
∆P = Q .
H3
The expression says that pressure and flow rate are linearly related
- double the pressure and you double the flow. This expression is a
hydraulic version of Ohm’s law (V = iR) where pressure difference
acts like voltage and volumetric flow rate acts like current. The number
12µL/H 3 is the hydraulic resistance.
There are a few things we should check with our resistance formula.
First, we can check the units to make sure they agree with what we
expect and that we did not make an error. The other check is that
the trends all go in the right way. It makes physical sense that the
resistance increases with viscosity, channel length, and with a decrease
in the channel height. It is always good to look at a result and see if
you believe the basic facts about that result before proceeding.
Note that the mean flow velocity is ū = Q/H,
∆P H 2
ū =
12Lµ
and the maximum flow velocity (along the centerline) is found by eval-
uating the velocity profile at y = H/2,
∆P H 2

H 3
u y= = = ū.
2 8Lµ 2
Since there is no acceleration in this flow we can check our result
with a simple force balance If we draw a box (control volume) the net
force (per unit depth into the page) exerted on the box due to pressure
is
FP = ∆P H.
This force must be balanced by the shear stress exerted at the two
walls. Recall the method of finding the stress on a surface discussed
in the chapter on the Navier-Stokes equations. The shear stress acting
tangential to the wall is given by n·T where T is the total stress tensor.
The normal vector for the lower wall is n = [0 1]. The stress tensor at
the wall for this simplified flow is
" #
T
−P µ ∂u ∂y
T = −P I + µ ∇v + ∇v = .
µ ∂u
∂y −P
Thus the stress vector is,

" #
µ ∂u
∂y
s=n·T=
−P
The tangential stress is proportional to the velocity gradient ∂u/∂y

which for our velocity field is,

∂u ∆P 1
= (H − 2y).
∂y L 2µ
The tangential stress, or x-component of the stress vector, acting on
the lower wall is then

∂u ∆P H
sx (y = 0) = µ = .
∂y y=0 2L
By symmetry we would expect the stress on the upper wall to be the
same. If you carry through the operation as we did above you only need
to be careful with the signs. The sign of the slope of the velocity field
is reversed at the upper wall, but the sign of the normal vector is also
reversed, thus giving a force in the same direction as the lower wall.
To get the force, the tangential stress is integrated over the length of
the channel, which is just the constant shear stress multiplied by the
length of channel. The total shear force per unit depth into the page,
Fs , is,
Fs = (sx (y = 0) + sx (y = H))L = FP = ∆P H.
The total shear force equals the total pressure force.

Could we have estimated our result from the start without actually
solving the problem? We can actually get a good estimate using the
force balance. The applied pressure force was given simply above. The
shear stress at the surface is τ = µ∂u/∂y. If we did not solve the
equation, we could have guessed that the velocity gradient would be of
the same order of magnitude as the flow velocity divided by half the
channel height; namely τ = µ∂u/∂y ∼ 2µU/H. The total shear stress
exerted on both walls is,
U
Fs ∼ 4µ L.
H
Equating these two forces would yield an estimate of the maximum flow
velocity as
∆P H 2
umax ∼ .
4µL
which is not the precise formula but has the right scaling with the pa-
rameters. A simple estimate in a simple problem seems useless (and
perhaps it is), but the same type of scaling arguments can be used in
9.2 Flow driven by a wall - Couette flow 129
u(y) H
x
U
Figure 9.2 Schematic for Couette flow between two parallel plates.
more complicated problems to get a handle on the form of the answer,

the basic scaling, or the order of magnitude. It is very common prac-
tice to guess that the terms in the partial differential equation can be
approximated in simple ways just to get a handle on what the answer
might look like.
While the parabolic velocity profile is accurate and can be realized
in practice, the theoretical solution between parallel plates is only ob-
served when the Reynolds number is less than 1400 (or thereabouts). If
the Reynolds number is large, the flow becomes unstable and turbulent.
The simple laminar flow solution is lost. Instability is an issue we will
discuss later.
9.2 Flow driven by a wall - Couette flow

Another simple case is where the flow is driven by a moving wall, known
as Couette flow. We assume a 2D channel in cartesian coordinates. The
channel goes from 0 < y < H. Just like the previous example, we
will assume the channel is infinite in the x direction and is at steady
state (both with respect to time and distance along the channel). This
case can be realized in practice by placing fluid in a thin gap between
concentric cylinders and rotating the cylinders with respect to each
other. The fact that we use cartesian coordinates is akin to assuming
a flat earth (i.e. the curvature is so large compared to the length scale
across the channel). In this case we will assume that gravity acts normal
the x − y plane, and thus has no role in this problem.
We can write the full Navier-Stokes just as in the Poiseuille flow
example and cross out all the time derivatives and x derivatives. We can
then apply conservation of mass to tell us there is no vertical velocity,
v. Crossing out all these terms again yields the x and y momentum
equations as in Equations 9.1 and 9.2,
∂P ∂2u
= µ 2.
∂x ∂y
∂P
= 0.
∂y
However, in this case the x momentum equation simplifies further as
there is no applied pressure gradient. In the example of the concentric
cylinders there could be no pressure gradient around the closed loop
of the fluid gap, otherwise the pressure would be discontinuous as we
went around the circle. For Couette flow, the x momentum equation
simplifies to,
∂2u
0= .
∂y 2
The curvature of the velocity field is zero and thus we expect a lin-
ear velocity profile. Integrating this expression twice with respect to y
yields,
u(y) = C1 y + C2 .
The two boundary conditions used to determine the two constants of

integration are u(y = 0) = U and u(y = H) = 0. Applying these
conditions to get C1 and C2 , the full velocity field is,
H −y
u(y) = U .
H
The velocity varies linearly between the two plates. The total flow rate
between the plates is computed as
Z H
UH
Q= u(y)dy = .
0 2
That’s all there is to it.
Note the analogy with steady state diffusion. The equation is the
same as found in one-dimensional diffusion at steady state. The Couette
flow equation states that the velocity field has no curvature. The only
possible solution to this equation is a linear velocity profile.
9.3 Flow down a ramp 131
g
H
H
y
Figure 9.3 Schematic for gravity driven flow down a ramp.
9.3 Flow down a ramp

Another variation of the class of problems in simple plane parallel flow
is that of a thin liquid film flowing down a ramp. Here, the ramp is
inclined at an angle θ and the flow is driven by gravity. Fluid is poured
in continuously at the top of the ramp. The upper surface of the fluid
film is exposed to the air. We will make the same assumptions as before
that the flow is steady and not varying with distance down the ramp.
We orient the coordinate system with the inclination of the ramp such
that the y direction goes across the film of liquid, and x goes down the
ramp. The surface of the ramp will be y = 0 and the surface of the film
will be y = H.
The development of the basic equations is identical to what we have
seen. The fluid acceleration is zero, the y component of the velocity is
zero, and all gradients of the flow velocity with respect to x are zero.
Following the same procedure as before, we obtain for the x and y
momentum,
∂P ∂2u
= ρgsin(θ) + µ 2 .
∂x ∂y
∂P
= −ρgcos(θ).
∂y
Integrating the second equation across the thickness of the film and
using the boundary condition that the pressure at the surface of the
film is that of the air around, P∞ , we obtain the pressure everywhere.
P = P∞ + ρgcos(θ)(H − y)
The pressure does not depend on x, thus the x momentum equation is,
∂2u
−ρgsin(θ) = µ .
∂y 2
Integrating twice yields (since the left side is constant)
ρgsin(θ) 2
u(y) = − y + C1 y + C2 .
2µ
The two constants of integration are found from the boundary con-
dition. At the surface of the solid ramp, the no-slip condition holds,
u(y = 0) = 0. Applying this boundary condition shows that C2 = 0.
Therefore,
ρgsin(θ) 2
u(y) = − y + C1 y.
2µ
The second boundary condition is that there is no shear stress at the
free surface of the liquid. Nothing is there (other than a little air) to
exert a force on the upper surface of the free fluid film. Thus we need
to set ∂u/∂y = 0 at the upper surface y = H. Applying this condition
yields,
ρgsin(θ) y
u= y H− .
µ 2
You should always confirm at the end that the solution satisfies the
equation and the boundary conditions.
Also note that the solution here is half a parabola - or essentially one
half of the flow between two solid plates. The driving force in Poiseuille
flow is pressure whereas here it is gravity. Otherwise, the solutions look
very similar.
9.4 Combined Poiseuille and Couette flow

In cases where there is no fluid acceleration (or we can neglect it) the
Navier Stokes equations become linear equations. Linear equations are
wonderful since we can use the useful idea of superposition to con-
struct the solution to more complex problems. We can take solutions
to simple problems and add them up in clever ways to construct a new
solution. Rarely in fluid mechanics are we afforded the opportunity to
use superposition.
9.4 Combined Poiseuille and Couette flow 133
One example where we can use superposition is imagine a case where

we have flow in a two-dimensional gap but that the flow is driven by a
pressure gradient and a sliding wall. Let’s take the lower wall moving
at velocity U . For this problem, let’s ignore gravity and assume it is
pointing out of the plane. If we started with the Navier Stokes equations
in two dimensions, we would cross out all the neglected terms exactly
as we did in the section on Poiseuille flow arriving again at Equations
9.1 and 9.2. When we integrate the resulting equation, nothing changes,
dP 1 2
u(y) = y + C1 y + C2 .
dx 2µ
To find the constants of integration we would need to apply the bound-

ary conditions, u(y = 0) = U and u(y = H) = 0. Applying the first
condition yields,
u(y = 0) = U = C2 .
Applying the second condition yields,
dP 1 2 U dP 1
u(y = H) = 0 = H + C1 H + U → C1 = − − H.
dx 2µ H dx 2µ
The flow field is then given as
dP 1 H −y
u(y) = − y(H − y) + U .
dx 2µ H
The total flow rate

H
dP H 3
Z
UH
Q= u(y) = − + .
0 dx 12µ 2
If you look at the velocity profile and total flow rate derived for
Poiseuille and Couette flow independently, you will see the above result
is just the superposition of the two solutions. We could have guessed
this result from the start since under our assumptions the equations are
linear. When you have linear equations you can compose a solution to
a problem as the sum of other solutions. Superposition is a useful tech-
nique for solving more complex flow problems as we will demonstrate
in a few of the coming examples.
High Pressure
Pressure = 0 U1(y) Pressure = 0

H1 U2(y)
H2
L1 L2
Figure 9.4 Schematic for sliding block moving to the left over a
stationary surface. In the reference frame of the block the lower
wall appears to move to the right at constant velocity. While the
gap size is exaggerated here, the solution assumes the gap is very
thin relative to the length of the block.
9.5 Slider bearing

We can use the solutions derived in simple situations to piece together
what will happen in more complex situations. As an example, let’s
consider the motion of two solid surfaces sliding past each other as in
Couette flow. However, in this case the upper surface will have a step
change in the height of the gap. In Figure 9.4 we take the upper block of
length L1 + L2 moving to the left. We must realize that in the reference
frame of the stationary wall, the problem is unsteady. If we observe the
flow from the moving block the flow appears steady in time. Thus we
will take the perspective that rather than the object moving to the left,
the lower flat wall is moving to the right as shown in Figure 9.4.
We will assume the gap height is small relative to the length such
that we can ignore effects that might occur around the step. We will
assume the flow is essentially uniform in the x-direction other than a
small local region around the step. We might be tempted to think that
the flow would be a simple Couette flow in the two regions. However, if
that were true then the flow rate in the two regions (see the Couette flow
derivation) would be Q1 = U H1 /2 and Q2 = U H2 /2. A linear velocity
profile is not plausible since the total flow rate would be different in
the two regions. Since the flow is incompressible, the total volumetric
flow rate in each section under the slider must be the same. Thus, the
manner in which the problem rectifies itself is that a pressure develops
in the thin gap to push fluid out of the gap as shown schematically in
Figure 9.4. If the pressure is high at the step then the flow rate will be
enhanced in region 2 and decreased in region 1.
The velocity profile in each region can be thought of as the superposi-
tion of Couette and Poiseuille flow. Using the results derived previously,
the total flow rate in region 1 will be
dP H13

U H1
Q1 = − + ,
dx 1 12µ
2
and similarly for region 2. Overall conservation of mass would state
that Q1 = Q2 ,
dP H13 dP H23

U H1 U H2
− + =− + .
dx 1 12µ
2 dx 2 12µ
2
The expression gives us the relationship between the two pressure gra-
dients in regions 1 and 2. To close the problem, we need additional
information. We need to know what is the overall applied pressure
across the entire fluid gap. Let’s assume the upper slider is open to the
fluid through which it moves. Therefore, the pressure at the two ends
would be the same and there is no overall applied pressure. Since only
pressure differences matter, we can take the far left and right ends of
the slider to be zero pressure. If P is the pressure at the step then the
pressure gradient in region 1 would be P/L1 and the pressure gradient
in region two is −P/L2 . The difference in sign is because the pressure
rises from 0 to P in region 1 and falls from P to 0 in region 2. We now
have enough information to solve for the pressure at the step, P ,
1 H13

1 6µU
P + 3 = (H1 − H2 ).
L2 L1 H2 H23
Notice that when H1 = H2 the pressure is zero and we are back to
Couette flow. Also notice that the pressure under the slider is positive
when H1 > H2 and negative when the situation is reversed.
In order to understand the result a little easier, lets explore the case
where L1 = L2 = L/2 and H2 = H1 /2. Substituting in these parame-
ters yields a pressure at the step of,
4µU L
P = .
3H 2
The pressure grows as the gap gets smaller. The positive pressure in
the gap (when H1 > H2 ) provides a lift force on the sliding object.
Since the pressure grows as the layer is squeezed, the fluid can prevent
the two solids from coming into contact. The basic effect is used in
hydraulic bearings where the viscous fluid forces replaces those of solid-
solid contact.
Recall from the Poiseuille flow example that the stress acting on the
solid wall is,
" #
µ ∂u
∂y
s=n·T= .
−P
The total normal load that could be supported by the pressure in the
lubrication layer is the integral of the pressure (the y component of
the stress) under the slider. Since the pressure gradient is a constant,
the pressure varies linearly with x and the total normal force per unit
width would be,
4µU L2
FN = .
3H 2
The force is directed upward due to the high pressure in the gap and
thus there is a lift force exerted on the block.
The tangential force is that required to drag the block through the
fluid. The tangential force is found by evaluating the shear stress τ =
µ∂u/∂y at the wall and then integrating over the length. The shear
stress in region one is

∂u PH Uµ
τ1 = µ =− − ,
∂y y=0
L H
Substituting for P gives,
4µU Uµ 7µU
τ1 = − − =− .
3H H 3H
Likewise for region 2,

∂u P 2U µ µU
τ2 = µ = − =− ,
∂y y=0 2L H 3H
The total tangential force per unit width is then (τ1 L1 + τ2 L2 ), or
8µU L
FT = − .
3H
A variation of this problem is a classic one where instead of the step
change in the height, a smooth upper surface (with no step in height)
Figure 9.5 Schematic of the equilibrium position of a shaft in a lu-

bricated bushing. From Machinery Lubrication - a trade publication
(www.machinerylubrication.com).
slides at an angle relative to the lower one. This problem was solved
by Reynolds (yes, the same one) shortly after his work on pipe flow
and the transition from laminar to turbulent flow. The basic behavior
Reynolds found is comparable to the case here, it just requires a little
more analysis. When the upper surface angles upward (same as the
incoming section is thicker than the trailing one) a high pressure is
found in the gap which provides a lift force. This lift force can keep the
upper block from contacting the lower surface. As the gap gets thinner,
the pressure and thus the lift force increases so there is a stabilizing
tendency. You can observe this effect by tossing playing cards across a
table. Incline the card upward and give it a good flick and it will slide
a long distance. You probably know this if you play a lot of poker (or
other card games). If you try this experiment, you can also try to punch
holes in the card such that the pressure can’t build up in the gap. The
cards with holes will come to a quick halt when you try to slide them.
A shaft rotating through a bushing is an important component of one
of our most useful inventions; the wheel. It has been long known that
lubricating the gap between the shaft and bushing with oil or other
viscous fluid can dramatically reduce friction. The problem we worked
in this section is the beginning of the study of this classic journal bear-
ing problem. If there is load on the shaft, there will be a tendency for
the shaft to move and contact the bushing. However, when the shaft
is rotating, we now have flow in a thin gap of nonuniform height. A

lubrication pressure will build in the gap and push the shaft away from
the bushing. Since the smaller the gap the higher pressure, eventually
the system will come to an equilibrium position where the forces bal-
ances and the shaft is held off the bushing by the lubrication pressure.
It turns out that the lubrication force is not aligned with the load, so
the shaft will sit off center in the bushing as shown schematically in
Figure 9.5. The principle is also important in thrust bearings.
9.6 Impulsively started Couette flow

Now let’s consider a transient flow problem in the x − y plane. Imagine
the Couette flow problem where initially both plates and the fluid are at
rest. At t = 0, the plate is impulsively started to move with a constant
velocity U and we wish to know the velocity field at any time, t. There
is no applied axial pressure gradient in this example.
We can start with the reduction of the Navier-Stokes equations,
which will proceed along the same lines as all the previous examples.
Since the plate is infinite in x there is no reason to think that the flow
at any x location should be different than any other. Thus all terms
with ∂/∂x are set to zero. Applying this reduction to conservation of
mass yields,
+ ∂v = 0
S∂u

∂x
S ∂y
S
Therefore, ∂v/∂y = 0, and the vertical velocity is a constant in the y

direction. Since v = 0 at the plate, v = 0 everywhere. Even though the
flow is transient in time, the argument from conservation of mass about
why the vertical velocity is zero remains the same since conservation of
mass is satisfied instantaneously in an incompressible flow.
We can now turn to the x component of the momentum in the Navier-
Stokes equations and we can discard terms that are multiplied by v or
have x gradients. The complete x-momentum equation with discarded
terms is,
!
∂ 2
u ∂2u

∂u @∂u S∂u ∂P S
ρ +u +v = x−
ρg +µ + 2 .
H S

∂t ∂x 2
@@ ∂y SS H ∂x
S
S ∂xSS ∂y
9.6 Impulsively started Couette flow 139
The Navier-Stokes is simplified significantly to
∂u ∂2u
ρ = µ 2.
∂t ∂y
The reduction for the Navier-Stokes is precisely as we had for steady
Couette flow, only now we have retained the transient term.
Dividing by the density and recalling the µ/ρ is the kinematic viscos-
ity ν, we have something that looks suspiciously like the 1D diffusion
equation,
∂u ∂2u
= ν 2.
∂t ∂y
The kinematic viscosity plays the same role in this equation as diffu-
sivity in the diffusion equation. We expect that the behavior would be
the same as applying an instantaneous change in temperature to one
side of a block of material while holding the other side cold. In this flow
case, we are talking about diffusion of momentum instead of diffusion
of thermal energy. The kinematic viscosity ν has the same units as the
thermal diffusivity.
Let’s make this equation dimensionless. Setting a scale for velocity
and length seems obvious; ũ = u/U , ỹ = y/H. However, what should
the time scale be? Using an arbitrary scale t̃ = t/t0 would yield,
∂ ũ νt0 ∂ 2 ũ
= 2 2.
∂ t̃ H ∂ ỹ
So clearly the “right” choice is to set to ≡ H 2 /ν which yields the

problem,
∂ ũ ∂ 2 ũ
= .
∂ t̃ ∂ ỹ 2
with initial condition ũ(t̃ = 0, ỹ) = 0, and boundary conditions
ũ(t̃, ỹ = 0) = 1, and ũ(t̃, ỹ = 1) = 0.
Therefore the solution at steady state will be ũ(ỹ) = 1 − ỹ. The system
will approach this equilibrium state with a time scale on the order of
H 2 /ν. The dimensionless formulation of this problem is exactly as we
found for similar problems in diffusion.
θ
r
Ro Ri
Figure 9.6 Schematic for Couette flow between two concentric cylin-
ders. The outer cylinder with radius Ro rotates with angular velocity
Ω.
9.7 Cylindrical Couette flow

Previously we solved the flow between two sliding plates. This flow can
be approximated in the lab by spinning two concentric cylinders relative
to each other. Rather than making the cartesian assumption for a thin
gap, let’s solve the problem in cylindrical coordinates (r, θ, z). The
nice thing about using the Navier-Stokes equations written in vector
form is that we can keep the equations the same and easily transform
to another coordinate system such as cylindrical or spherical. We will
not derive the equations in different coordinate system but they can be
found in a number of references and easily looked up. Lets take a outer
cylinder of radius Ro rotating at constant angular velocity of Ω. The
inner cylinder is stationary with radius Ri . The fluid is infinite in the z
direction, and gravity acts in the z direction. We will assume that the
solution is axisymmetric (there is no variation in the θ direction) and
that there is no fluid motion in the z direction. This seems reasonable
to start. We will use the notation that ur is the radial component of
velocity and vθ is the θ velocity.
The momentum equations in the r and θ directions are stated below
(by looking up the vector operators in cylindrical coordinates). To sim-
plify matters, we ignore gradients in the z direction and assume that
everything is constant in z. The momentum and mass equation are
9.7 Cylindrical Couette flow 141
stated where we cancel unsteady terms and ones that disappear due to
the axisymmetric assumption. The equations with discarded terms are
for the r momentum,
v2

∂ur ∂ur vθ ∂u
r ∂P
+ − θ =−

ρ + ur +
∂t ∂r r ∂θ r ∂r
!
1 ∂ 2

∂ 1 ∂rur u
r 2 ∂u

θ
µ + 2 2 − 2 ,
∂r r ∂r r ∂θ
r ∂θ

θ momentum,

∂vθ ∂vθ vθ ∂v
θ
ur vθ 1 ∂P

ρ + ur + + = − +
∂t ∂r r ∂θ r r ∂θ
!
1 ∂ 2

∂ 1 ∂rvθ v
θ 2 ∂ur
µ + 2 2 + 2 ,
∂r r ∂r r ∂θ
r ∂θ
and conservation of mass,
1 ∂rur 1 ∂vθ
+ = 0.
r ∂r r ∂θ
Integration of conservation of mass tells us the rur =Constant. Since
the radial velocity at the surface of the cylinder is zero, the radial veloc-
ity is zero everywhere. This further simplifies the momentum equations
to
v2

∂ur ∂ur vθ ∂ur ∂P
+ ur + − θ = −

ρ +
∂t ∂r r ∂θ r ∂r
!
1 ∂ 2

∂ 1 ∂ru r u
r 2 ∂u

θ
µ + 2 2 − 2 ,
∂r r ∂r
r ∂θ
r ∂θ

θ momentum,

∂vθ ∂v
θ vθ ∂v
θ
ur vθ 1 ∂P

ρ + ur + + = − +
∂t ∂r r ∂θ r r ∂θ
!
1 ∂ 2

∂ 1 ∂rvθ v
θ 2 ∂u

r
µ + 2 2 + 2 ,
∂r r ∂r r ∂θ
r ∂θ

The equations are significantly simplified just as in the other exam-

ples of this chapter. One difference however, is that there is one accel-
eration term that shows up in the radial momentum balance. This is
the centripetal acceleration. This acceleration is balanced by the radial
pressure gradient. The radial pressure is analogous to the case where
you swing an object tied to the string in a circle above your head and
the tension in the string balances the object’s acceleration. The equa-
tions we will solve are the radial momentum balance,
vθ2 ∂P
−ρ =−
r ∂r
and the θ momentum,

∂ 1 ∂rvθ
0=µ
∂r r ∂r
Integrating the θ momentum equation once yields,
1 ∂rvθ
= C1
r ∂r
which can be integrated again,
C1 C2
vθ = r+ .
2 r
Solving for the constants of integration by applying the boundary con-
dition at the inner and outer cylinder gives the velocity field as,
r Ri
Ri − r
vθ = ΩRo R Ri
o
Ri − Ro
It is easy to confirm that this solution satisfies the boundary conditions.

If you plot the solution for cases where the inner cylinder has a radius
which is 90 percent of the outer cylinder, you will notice little difference
between the true solution and the linear “flat earth” approximation.
The flow between concentric cylinders is a classic problem called
Taylor-Couette flow. The general Taylor-Couette problem considers the
ability to spin the inner and outer cylinder wall independently. You
should check out some movies of this flow on youtube. The simple
flow with the velocity field above can be only observed in some regions
of parameter space. As the relative rotation speeds of the cylinders
are changed, the flow can transition through a number of complicated
9.8 Poiseuille flow in a pipe 143
r
θ z U( r )
Figure 9.7 Schematic for pressure driven Poiseuille flow in a pipe.
shapes and regular but complicated patterns. This behavior is indica-

tive of the non-linear behavior of the Navier-Stokes equations.
9.8 Poiseuille flow in a pipe

Analysis of flow in a circular pipe of radius R follows the same logic as
we discussed for Poiseuille flow in a slot. However, since in the labora-
tory or in application it is much easier and more common to have flow
in a circular pipe this is an important example. As in the slot example,
we will assume that the pipe is very long and we are interested in the
steady state behavior along the length of the pipe. We will assume that
we are only interested in flows that are axisymmetric. If we write the
Navier-Stokes in cylindrical coordinates, we can proceed by eliminating
terms based on the axisymmetric assumption and that the gradients in
the axial direction are zero. We will use the notation that ur is the
radial component of velocity, uθ is the θ velocity, and uz is the axial
velocity.
Writing conservation of mass in cylindrical coordinates and discard-
ing gradients in θ and z yields.
1 ∂rur 1 ∂u
θ ∂uz
+ + = 0.
r ∂r r ∂θ
∂z
Conservation of mass states that rur is a constant. Since the radial
component of the velocity is zero at the wall r = R, it is zero every-
where.
Since there is no radial or swirling (θ) velocity those components of
the momentum equation will all be zero. We only need to consider the
axial momentum equation,

!
∂uz ∂uz uθ ∂u

z ∂u
z ∂P
ρ + ur + + uz =− +
∂t ∂r r
∂θ ∂z ∂z
!
1 ∂ 2 ∂ 2 uz

1 ∂ ∂uz u
z
µ r + 2 2 + ,
r ∂r ∂r r ∂θ
∂z 2
Since there is only one component of the velocity we drop the subscript
on the uz and just use u to denote the axial velocity for simplicity.
The final reduced momentum equation is analogous to Equation 9.1 in
cylindrical coordinates,

dP 1 ∂ ∂u
=µ r . (9.3)
dz r ∂r ∂r
Integrating twice gives
1 dP 2
u(r) = r + C1 ln(r) + C2
4µ dz
Since we don’t want a singularity at r = 0, C1 = 0. Applying the no-slip
boundary condition that u(r = R) = 0 gives,
dP 1 2
u(r) = (r − R2 ).
dz 4µ
The total flow rate, Q, is found by
Z R Z R
dP 1
Q= 2πru(r)dr = 2π r(r2 − R2 )dr,
0 dz 4µ 0
Note that we need to remember our factors of r and π when integrating
over the surface of the pipe inlet in cylindrical coordinates. Performing
the integral we obtain
π∆P R4
Q= ,
8µL
or written as the common pressure-flow relationship,
128µL
∆P = Q
πD4
Notice the term 128µL/πD4 is the hydraulic resistance which relates
pressure and flow. The pipe resistance is a strong function of the di-
ameter. One effect is the area, there is less volumetric flow if the area
is reduced (for the same velocity). The other effect is that viscosity is
stronger when the diameter is small.
Taking the relationship for pressure and flow in a pipe and recasting
in terms of the velocity, we could rearrange the equation to obtain,
uµL
∆P = 32 .
D2
Here u would be the average velocity in the pipe, i.e. Q/(πD2 /4). Re-
member from an earlier chapter that pressure drop is pipe flow is char-
acterized as
1 L
∆P = ρu2 f (Re).
2 D
Equating the last two expression yield,
uµL 1 L 64µ 64
32 2
= ρu2 f (Re) → f = = .
D 2 D ρuD Re
the analytical solution for laminar flow friction factor that we saw on
the Moody diagram in Chapter 3. As we have described before, this
solution is only observed up to around a Reynolds number of 2300 and
then stability is lost and flow become turbulent.
9.9 Comments on the stability of solutions

In simple fluid flows as we have described in this chapter, we can often
find a solution to the Navier Stokes equations. As we have discussed,
these solutions are not guaranteed to be stable. Imagine trying to bal-
ance a pencil on its sharpened tip. While a free body diagram of a
vertical pencil on a table is a perfectly valid theoretical solution, you
can easily see that the situation is highly unstable and not realizable in
practice. If the pencil leans ever so slightly to one direction, the pencil
falls over. Now imagine an unsharpened pencil. You can balance this
one on its end, but if you hit the table it falls over. Both of these sit-
uations can occur in fluid flows where the solution to the equations is
valid, however it is not stable to either finite or infinitesimal pertur-
bations. There are many routes to instability in fluid flows, many of
which are an active area of research. Since the Navier-Stokes equations
are non-linear there are no guarantees of unique solutions to problems.
However it is the issue of stability and non-linearity that make fluid

mechanics a fascinating area of study.
The most straightforward method of stability analysis is to take a sit-
uation where a simple solution to the Navier-Stokes equations is known.
We can then consider very small perturbations to this solution and then
decide if those perturbations grow or decay in time. Mathematically we
make a substitution where we would say the the velocity is u = u0 + û
where u0 is the known simple base state and û is an infinitesimal per-
turbation to that base state. We can do this for all the variables and
plug this into the Navier-Stokes equations. Since the perturbations to
the base state are assumed small, we can discard all the non-linear
terms (products of hat terms) which comprise the product of two small
things. The result are linearized partial differential equations for the
perturbations to the base state. In practice, we can usually solve this
linear problem (at least numerically). We can then look at the solutions
and see if they either grow or decay with time. If the solutions to the
perturbed equation grow with time, this indicates that if we provide
a slight random kick, the solution will grow in size until it dominates
the problem, noting that the flow is unstable. This type of analysis is
called linear stability analysis.
This type of stability prediction works remarkably well in some cases
and remarkably bad in others. Two cases where this analysis works
well are called Rayleigh-Benard convection and Taylor-Couette flow.
Rayleigh Benard convection is what occurs when a channel is heated
from below and the hot fluid near the lower surface wants to rise up due
to buoyancy. If the temperature difference is small the fluid stays still
and heat is transported by conduction. As the temperature difference
increases, convection cells form. These convection cells can result in
interesting large scale patterns in the flow, such as stripes and spirals.
The initial onset of convection is well-predicted by a linear stability
analysis.
The Taylor-Couette problem is the flow between two concentric cylin-
ders rotating relative to each other. If the inner cylinder is rotating
slowly, we have a Couette flow in the gap between the cylinders that
we studied earlier. At higher speeds there is a centrifugal instability
that causes a set of convection cells to set in. So rather than a sim-
ple shearing motion, a set of vortices spontaneously emerge that wrap
around the cylinder. An example image is shown in Figure 9.8. If one
Figure 9.8 Image of the unstable vortices that wrap around the
cylinder in a Taylor-Couette device. On the right, the phase diagram
of the different types of flow observed if one can independently rotate
the inner and out cylinders in the device. Image from Album of Fluid
Motion. Phase diagram from Anderdeck, Liu, and Swinney, J. Fluid
Mech 1986.
controls the speed of both the inner and outer cylinder the flow is ex-
tremely complex and many different regimes of qualitative behavior
are seen in the phase diagram of Figure 9.8. The lower boundary of the
phase diagram from simple sheared Couette flow to the initial instabil-
ity is well predicted by linear stability analysis. The threshold and flow
patterns that one observes in experiment closely match those that are
predicted by linear analysis. The rest of the phase diagram beyond this
initial instability is much more complicated to predict. As you exceed
the threshold non-linear effects come into play and the patterns become
ever more complicated and beautiful. While the topic of linear stability
analysis is beyond what we will do in this class, it is not that difficult.
It is a topic found in a number of textbooks.
Flow in a pipe is a case where the linear stability analysis works
remarkably bad. In this case, the analysis predicts the flow is stable
at all Reynolds numbers, though we know in practice that is not true.
The mechanism for transition to turbulence in simple pipe flow is a
problem which is still worked on extensively to this day. While a lot is
understood about the laminar to turbulent transition in this problem,
it is still an area of research 125 years after Reynolds first made his
observations.
There are many complications with the Navier-Stokes equations but
the point I want to emphasize is simple. It is not sufficient to simply
solve the Navier Stokes equations. We always need to ask whether those
solutions are stable. Predicting stability is sometimes straightforward
and sometimes not. Just be aware that asking the question of stability
is a big deal in fluid mechanics and the complications of stability lead a
number of interesting and beautiful phenomenon. In the end, however,
any analysis or simulations you conduct must be brought before the
ultimate judge - experiments.
9.10 Computational Fluid Dynamics (CFD)

Since the 1970s, the desire to simulate fluid flow has been one of the
great drivers for development of simulation software, numerical meth-
ods, and supercomputers. CFD remains a very large endeavor in both
industrial and academic settings. Due to Moore’s law and dramatic in-
crease in computing power, the possibilities with CFD changes quite
rapidly. Problems that would have challenged the fastest computers
and the best researchers some years ago can be done routinely with
commercial software and a typical personal computer. Based on look-
ing at the sample problems that come with modern CFD packages, I
would estimate that a problem that had the complexity to have been
a PhD thesis 10-15 years ago can be done routinely with commer-
cial software. In terms of practical available computing power I would
also estimate about a 10 year lag between when a cutting edge super-
computer problem becomes feasible with a typical low-cost computer.
These two forces, improving hardware and software, continually lower
the financial cost and required expertise to using CFD.
There are different numerical methods which all approximate the
solution of the Navier Stokes equations as discrete chunks. The most
common for commercial CFD packages are based on either finite ele-
ment, finite volume, or finite difference methods. I suggest that if you
ever work with CFD for a job or in graduate school you learn the dif-
ference. For linear problems such as diffusion, a commercial package
can accurately solve the diffusion equation for any arbitrary 3D geom-
9.10 Computational Fluid Dynamics (CFD) 149
etry with very little effort. A novice user can quickly generate a good
simulation.
Due to the non-linear nature of the Navier-Stokes, we are not yet at
a state of “plug and play” for all problems. Laminar single phase flows
are at that point today. If we have a 2D or 3D flow where the Reynolds
number is low enough that the non-linear terms are not too “strong”,
then a modern CFD package can usually do a good job without the user
getting too involved in the details. This does not mean that laminar
flow is always easy, that there aren’t sometimes difficulties, that you
should run to the computer and shy away from analysis or that you
are always going to get a physical answer. But practically speaking, if
you are interested in a single phase laminar flow you could probably
simulate everything you need without too much trouble or expertise.
All the examples we worked in this section could be set up and run
very quickly using a commercial CFD package. However, as we dis-
cussed earlier in the diffusion chapter, when you just do a simulation
you can miss simple scaling laws, and simple formulas that are good
for basic understanding of trends and also very useful in design prob-
lems. Simulation is another tool that can work side by side with good
analysis, but is rarely the substitute.
We will discuss turbulent flows in a later chapter, but simulating
turbulent flows is a much trickier situation and one where you should
know what you are doing. There are fundamental limitations that make
high Reynold’s number flows challenging that we will address later. For
now, it is worth noting that if you use a CFD package and get good
results, turn up the Reynold’s number and it is guaranteed to stop
working.
10
Inviscid flow, Euler’s equation and Bernoulli
Previously we derived the Navier Stokes equations for a Newtonian

fluid with viscosity. If we take the limit of the Navier Stokes equations
of zero viscosity, we recover the Euler equations,
∇ · v = 0, (10.1)
Dv
ρ = −∇P + ρg. (10.2)
Dt
We previously found that for a human swimming at a reasonable pace,
the Reynolds number was on the order of a million. We can imag-
ine for a car going down the highway it is even larger. For weather
related phenomena a Reynolds number of 1010 to 1015 would not be
uncommon. Since the Reynolds number is a measure of the ratio of
inertia to viscosity, it seems that perhaps dropping the viscous terms
and turning to Euler’s equation might be a good approximation for
high Reynolds numbers. Euler’s equation looks simpler but it’s behav-
ior can be quite different than the Navier-Stokes even in the limit of
high Reynolds number. The mathematical difference is in the number
of boundary conditions required. For flow over a solid surface, Euler’s
equation would only require (and could only enforce) the boundary
condition that flow does not penetrate the surface. Fluid would slip
over the surface in Euler’s equation. Since the Navier Stokes equations
has higher spatial derivatives (in the viscous terms), it requires more
boundary conditions. The Navier Stokes requires that flow does not
penetrate the surface and that the flow does not slip at the surface.
This subtle difference can lead to quite different behavior. Mathemat-
ically, the zero viscosity limit is called a singular perturbation. This
152 Inviscid flow, Euler’s equation and Bernoulli
means that we take the limit of the small parameter (viscosity) going
to zero and we fundamentally change the problem at hand.
In the last chapter on Navier-Stokes solutions we studied examples
where the balance of forces was between viscosity and some driving
force such as applied pressure. In all the problems in the last chapter,
the fluid had no acceleration. When using Euler’s equation we are ig-
noring viscosity and the balance of forces is between acceleration and
pressure (or gravity). A good strategy for cases of high Reynolds num-
ber flows where fluid acceleration is important is to start a problem
by assuming that Euler’s equations work. We could then check what
the predictions and trends are versus what experiments and experience
say to decide if the approximation was reasonable. We will begin to
understand the breakdown of Euler’s equations better after the next
chapter when we discuss boundary layers.
10.1 Flow along a streamline: Bernoulli

One of the most useful (and simple) relations for flows with no viscosity
is known as Bernoulli’s equation. Let’s start by a simple demonstration
of the where the derivation comes from and then we will follow with a
more general derivation.
Consider the flow through a constricted tube as shown in Figure 10.1.
Let’s consider steady, incompressible, and inviscid flow. For now, let’s
further neglect gravity. Without gravity, Euler equations become
ρv · ∇v = −∇P.
The balance here is that acceleration of the fluid particle is balanced

by the net force due to the pressure gradient. In component form the
x momentum equation is

∂u ∂u ∂u ∂P
ρ u +v +w =− .
∂x ∂y ∂z ∂x
Now consider the flow just along the centerline. By symmetry, we could
argue that there is no component of velocity in the y or z direction;
v = 0 and w = 0. Along the centerline, there is nothing interesting
contained in the y and z momentum equations since all accelerations
in those directions are zero (you can write out all the components and
10.1 Flow along a streamline: Bernoulli 153
Figure 10.1 Venturi device. The cross sectional area at point 2 is

less than at point 1, thus the flow accelerates. Since the velocity at
the throat is faster than the upstream, Bernoulli’s equation predicts
the pressure is lower at the throat.
convince yourself this is true). Along the centerline the x momentum

equation simplifies to,
∂u ∂P
ρu =− .
∂x ∂x
Using the chain rule note that,
∂ u2 ∂u ∂u ∂u
=u +u = 2u .
∂x ∂x ∂x ∂x
Using this result we can rewrite the momentum equation as
∂ 12 ρu2 ∂P
=− ,
∂x ∂x
or
∂ 1 2
ρu + P = 0.
∂x 2
Along the centerline,
1 2
ρu + P = Constant,
2
which we will soon see is a specific version of a more general relationship
called Bernoulli’s equation.
For the device shown, by conservation of mass we know that the flow
velocity will increase at the throat. Our expression for Euler’s equation
along the centerline says that if the velocity increases, the pressure
must decrease. When the flow accelerates, the pressure decreases. When
the flow decelerates, the pressure increases. This general trend should
make sense from the perspective of the fluid particle. If you are flowing
through the device and are accelerating upon going into the throat, then
there has to be some net force acting on you pushing you from behind.
This net force is the high pressure behind you and lower pressure in
front of you.
10.2 General Bernoulli

Let’s consider a more general derivation of Bernoulli’s equation. Again,
we will consider steady incompressible, constant density, inviscid flow
where the conservation of momentum would be written as
ρv · ∇v = −∇(P + ρgz).
Here we assumed that gravity is pointing in the negative z direction,
without loss of generality. If we look up a table of vector calculus identi-
ties we are free to write the term of the left side in a number of different
forms. One identity would allow us to write,

1 2
ρ ∇v − v × (∇ × v) = −∇(P + ρgz).
2
which, when density is constant, is equivalent to

1 2
−ρv × (∇ × v) = −∇ ρv + P + ρgz .
2
We discussed this previously, the curl of the velocity vector ∇×v is also
known as the vorticity ω. Vorticity is a measure of the local solid body
rotation of a fluid particle. It can be useful in a number of settings, but
it is a little hard to wrap your head around what it means. Basically if
you measured a flow and plotted an image of the vorticity field, it would
be highest where the fluid was spinning (i.e. in tornadoes, toilets, and
whirlpools). It is a property of the curl operator, the vorticity always
points perpendicular to the velocity vector. If you have a 2D flow in the
plane of a sheet of paper, the vorticity will point out of the page. We
will discuss vorticity more in a coming section, for now you can consider
it a mathematical thing. Rewriting conservation of momentum with the
vorticity yields,

1 2
−ρv × ω = −∇ ρv + P + ρgz .
2
10.2 General Bernoulli 155
Note that at this point we really haven’t done anything, just rewrit-
ten the same equation in a different form using some vector calculus
identities.
Let’s take the whole equation and take the dot product with the
velocity vector itself,

1 2
−v · ρv × ω = −∇ ρv + P + ρgz .
2
Notice the term on the left side. The curl of the velocity and the vor-
ticity vectors will always be perpendicular to those two vector. Thus
the dot product with the velocity vector itself will be zero. This fact
can also be easily proven by carrying out all the terms in component
form. Since the left side is always zero, we obtain

1 2
0=v·∇ ρv + P + ρgz .
2
This fact means that the quantity in parenthesis does not change in the
direction of the velocity vector. In a fluid flow we define a streamline as
a line that follows the velocity vectors. In a steady flow a streamline
will correspond to the path a blob of injected dye would follow. Thus
in a steady flow,
1 2
ρv + P + ρgz = a constant along a streamline. (10.3)
2
This equation is known as Bernoulli’s equation. It is based on some
restricted assumptions, namely 1) incompressible, 2) steady, and 3)
inviscid flow. Despite these restrictions it is a powerful equation because
of its simplicity.
Bernoulli’s equation has a simple interpretation. The kinetic energy
per unit volume is 12 ρv2 , and the potential energy is ρgz. These expres-
sions should look familiar from particle mechanics only here we use the
mass density rather than the total mass. Pressure exerts a force per
unit area, thus a change in pressure between two locations indicates
that the pressure is doing work on the fluid. This work is related to the
familiar expression that work is equivalent to the integral of force over
distance. Bernoulli’s equation is saying that the work done by pressure
is equal to the change in energy (kinetic plus potential).
While simple and powerful, Bernoulli’s equation can also be mislead-
ing. There is a tendency to want to use it under conditions that are
not appropriate. We must always ask if the assumptions are met when
using Bernoulli, and ultimately determine whether this simple equation
provides predictions which match reality. The most common mistake is
using Bernoulli’s equation when viscosity is not negligible.
10.2.1 Example: Venturi

A venturi meter is a simple device for measuring flow rate shown in
Figure 10.1. The meter is inserted in a pipe and you measure the pres-
sure difference between an upstream location and at the center of the
contraction. Gravity acts normal to the flow, thus the potential energy
does not change for a fluid particle going through the meter. If the area
change is known, the velocity ratio at points 1 and 2 is known as
v2 A1
=
v1 A2
from conservation of mass. Here v1 and v2 are average velocities. Ap-
plying Bernoulli from point 1 to 2 yields

1 2 1 2
ρv + P = ρv + P .
2 1 2 2
Combining these expression yields,

2 !
1 1 1 A1
P1 − P2 = ρv2 2 − ρv1 2 = ρv12 −1
2 2 2 A2
If the pressure difference P1 − P2 is measured and the area ratio is
known, then we can figure out the fluid velocity. The venturi is one of
the simplest flow meters.
An old use of a venturi is the carburetor. The carburetor is a device
for mixing fuel and air prior to combustion in an engine. They are not
used in cars anymore, replaced by fuel injectors, however they are still
in use in low-cost engines. The carburetor has a venturi and makes use
of the low pressure in this region to draw fuel into the airstream.
10.3 Euler equations across a streamline

We can also use Euler’s equation to provide a simple relationship for
what happens across streamlines. The following derivation can be made
Vθ
Figure 10.2 Flow around a circular streamline. Euler’s equations can

tell us how the pressure changes as we jump across a streamline.
more general, but for simplicity lets consider that the streamlines at
some point in the flow are curved as perfect circles, such that we can
describe the flow in cylindrical coordinates. The nice thing about our
vector calculus approach is that we simply look up the different opera-
tors in different coordinate systems, and we are good to go. Looking up
the operators for cylindrical coordinates where the flow is in the r − θ
plane and gravity only points in the z direction the Euler equations in
steady flow become,
vθ2

∂ur vθ ∂ur ∂P
ρ ur + − =−
∂r r ∂θ r ∂r

∂vθ vθ ∂vθ 1 ∂P
ρ ur + =−
∂r r ∂θ r ∂θ
We use the subscripts θ and r just as reminders that u and v are

the components in these coordinate directions. Since at this point, the
streamlines follow the circle, ur = 0. There is no radial flow. Likewise,
since we are following a circle here ∂ur /∂θ = 0. Following the circu-
lar streamline at this point allows us to remove some terms from the
equation to obtain equations that are valid on the streamline,
vθ2 ∂P
ρ = ,
r ∂r

vθ ∂vθ 1 ∂P
ρ =− .
r ∂θ r ∂θ
H(r)
z
z=0 r
Figure 10.3 Schematic for the example of computing the free surface
shape of a cylindrical tank of water in solid body rotation.
The second equation for the θ component of momentum is equivalent

to

∂ 1 2
ρv + P = 0
∂θ 2 θ
which states that Bernoulli holds in the θ direction. Since this direction
is the streamline, we are satisfied.
The new equation is that the pressure changes across the streamlines
as
v2 ∂P
ρ θ = across a streamline (10.4)
r ∂r
Equation 10.4 should make some physical sense. If we have flow spinning
a circle, the pressure increases as we move outward in radius. The
pressure gradient is equivalent to the tension of a string that holds a
rock that you swing in a circle. If the streamlines are straight, then
there is no pressure jump across them.
10.3.1 Example: rotating tank

Consider a container of fluid on a turn table such that the whole system
is in solid body rotation. The velocity field of the fluid is vθ = Ωr where
Ω is the constant angular velocity and r is the coordinate. The axis of
rotation and gravity are aligned such that gravity does not appear in
the radial momentum equation. Using our expression from the previous
section, the radial component of the momentum equation is
vθ2 ∂P ∂P
ρ = → ρΩ2 r = .
r ∂r ∂r
Since everything on the left side is a constant we can integrate this
equation with respect to r (holding z = Z constant) from the origin to
obtain the pressure distribution as a function of r.
1 2 2
P (r, Z) − P (r = 0, Z) = ρΩ r
2
The pressure is higher at the wall of the container than the center. The
vertical momentum equation is simple since there is no vertical flow,
namely, at a constant radial location (r = R),
P (z, R) − P (z = 0, R) = −ρgz.
If we set z = 0 to the surface of the liquid at the center (r = 0) of

the container, we can set the pressure reference at this point to be 0.
We can follow at radial line outward to the wall to an arbitrary point
r,
1 2 2
P (r, z = 0) = ρΩ r .
2
Vertical balance provides the pressure at the same point to be
P (r, z = 0) = ρgH(r) − P (z = H, r)
Since the pressure at the surface of the liquid is zero, P (z = H(r), r) =

0, then we are left with a simple result that the height of the fluid
surface is,
1 Ω2 r 2
H(r) = .
2 g
You can confirm the result that the surface of the fluid is a parabola
by spinning a cup of fluid. You will see the shape of the surface looks
parabolic. We could have obtained this result using dimensional analy-
sis quite readily. We would not have known the factor of 1/2 if we had
done dimensional analysis.
r
r
1
2
P
Center
pressure
Wall pressure
Figure 10.4 Schematic of the pressure distribution in a venturi me-

ter comparing the center pressure to the wall pressure.
10.3.2 Example: Venturi revisted

Let’s now reconsider the pressure distribution in a venturi meter. The
pressure along the centerline will follow that predicted by the Bernoulli
equation along the streamline. The pressure at the wall will be dif-
ferent. In the first part of the contraction, the radius of curvature of
the streamline following the wall is such that the pressure at the wall
would be higher than the pressure along the center. Approaching the
nozzle throat, the wall streamline passes through an inflection point
and the radius of curvature changes sign. On the final approach to the
throat, the pressure along the wall will be l ower than the pressure along
the centerline. The pressure profile as a function of distance is shown
schematically in the dotted line of Figure 10.4. While a more detailed
calculation would be required to quantitatively describe the difference,
the qualitative result that pressure must increase as we move outward
along a radius of curvature is sufficient to explain the basic effect we
would see in experiment.
10.4 Vorticity 161
10.4 Vorticity
This section is going to rely on a fair amount of vector calculus and
identities. If you do not remember all these identities, then don’t get
too worried. Try to follow the derivation, but focus on understanding
the consequences. First let’s restate that vorticity is a vector defined as
the curl of the velocity,
ω = ∇ × v. (10.5)
Vorticity is a measure of solid body rotation of a fluid particle. Note

that flow can go in a circle, and still have no vorticity. A Ferris Wheel
is an example of a circular flow where the fluid particle (i.e. you) does
not undergo solid body rotation. You face the same direction as the
wheel spins. On a Merry-Go-Round, you the fluid particle, certainly
are rotating which is why you get very dizzy. In a fluid, water waves
is an example where fluid traces a circular path as a wave passes over,
however the fluid particles do not undergo solid body rotation (i.e. the
vorticity is zero).
You can construct a simple vorticity meter. You only need a little
paddlewheel that is free to spin on an axle. If you stick a paddlewheel in
a flow and it spins, then there is a component of vorticity pointing along
the wheel’s axis of rotation. If you move the paddlewheel’s orientation
around until it spins the fastest, then that would be the axis along
which the vorticity vector points.
The total vorticity over an area can be related to the velocity field
using another vector calculus theorem, Stokes’s Theorem. This theorem
says,
I Z
Γ = v · tdl = (∇ × v) · ndS. (10.6)
This theorem says that if I integrate the tangential velocity (t is the

tangent vector of the curve) around some arbitrary closed loop, where
the line element length is dl, this quantity must equal the net vorticity
flux coming through the area. The value of the integral is defined as
a property, the circulation Γ. The circulation tells you if the fluid is,
well, circulating. Stoke’s theorem is a general vector calculus theorem
for any vector, but here I have written it in the language of velocity
and vorticity.
You should look up a weather map for a Northern Hemisphere At-
lantic hurricane which has wind vectors drawn on it. You will notice
the vectors show the winds going in a circle in the counterclockwise
direction. To estimate the circulation, draw an arbitrary circle which
roughly follows the winds. To estimate the magnitude of the circulation
simply estimate the average wind speed on your circle, and multiply
that speed by the circumference. To compute the circulation more ac-
curately you could break your circle into little pieces to account for the
fact that the wind speeds will vary around the circle. Adding up these
little pieces is the approximation of the integral in the formal definition.
Now let us derive an equation for the vorticity. Lets take the curl of
Euler’s momentum equation in constant density, incompressible flow,

∂v
ρ∇ × + v · ∇v = −∇(P + ρgz) . (10.7)
∂t
Now we need to recall our vector calculus identities. You might re-
member that the curl of the gradient of something is zero. If you don’t
believe this, you could carry out the operation for each component to
find out. There are also identities that allow us to expand the ”v dot
grad v term”.
∂ω
+ v · ∇ω − ω · ∇v = 0 (10.8)
∂t
which is the same as
Dω
= ω · ∇v. (10.9)
Dt
What is interesting is that the pressure and gravity have disappeared.
Vorticity is a measure of solid body rotation and pressure and grav-
ity act through the fluid particles center of mass in constant density
flows, thus they can not change the rotation. If there are density gradi-
ents in the fluid, then these can couple with gravity to create vorticity.
This mechanism of vorticity production is common in oceanic and at-
mospheric flows. Now lets turn to trying to interpret these equations
starting with a simple example of 2D flow, remembering the restriction
that we have assume inviscid flow.
10.4.1 2D flows
Lets restrict our analysis to 2D flows to make things a little easier.
First of all in a 2D flow, velocity only has components in u and v and
10.4 Vorticity 163
Γ=0
Figure 10.5 Drag a spoon slowly through a bowl of water. You will
see two vortices form behind the behind of opposite sign but equal
strength. The two vortices will cancel each other such that the total
circulation around the outer dashed contour is zero.
gradients in x and y. Vorticity in a 2D flow only has a component in

the z direction; v = [u v 0]; ω = [0 0 ωz ]. Thus the term in Equation
10.9 on the right hand side is zero since the vectors are not in the same
plane. In a 2D flow,
Dω
=0 in 2D flows.
Dt
This equation states that vorticity is a material property. Whatever
vorticity the fluid particle has initially, it retains that amount and then
simply goes with the flow. Since the vorticity for a fluid particle is
conserved, if we draw a loop which is a material loop (a massless,
strechable string with no elasticity that can go with the flow), the
vorticity inside that loop is the same for all time. Thus the integral of
the vorticity or the circulation is constant,
DΓ
=0 around a closed material loop.
Dt
This statement is known as Kelvin’s circulation theorem. It is an inter-
esting theorem and is more a general theorem than the simple argument
I made here. A formal proof of the theorem would show that the the-
orem is valid if Euler’s equations hold along the contour around which
the circulation is calculated. Thus, there can be viscosity inside the
closed loop. The theorem also holds in three dimensions.
So how does this theorem manifest itself? Try the following experi-
ment. Take a bowl of water and sprinkle some tracer particles in the
water; fine black pepper works well. Take a spoon and give the sur-
face of the fluid a little flick. You should see two vortices generated.
They will be of opposite spin and propagate themselves for a bit before
decaying away. Now, imagine you draw your loop to calculate the cir-
culation around the spoon and far away from it. Since everything is at
rest intially, the circulation is zero. Now you flick the spoon and create
your vortices. So even though viscosity has done something inside the
loop (we could not create the vortices without it), the circulation the-
orem still applies far away from the loop. So the circulation is still zero
even though there is vorticity inside the loop. The circulation theorem
states that the circulation is zero and remains so. This means that the
positive and negative vorticity you generated must cancel each other
out via the definition of circulation as the vorticity integrated over the
area (via Stokes’s theorem). In this experiment you get two vortices of
equal strength and equal size, such that the total integrated vorticity
is zero. The theorem tells you nothing of the vortices decay, just that
the total circulation should always be zero.
You may have noticed that the equal sized but oppositely signed
vortices propagate themselves in a straight line. This behavior can be
understood by using an ideal vortex which is a solution to Euler’s equa-
tion. The ideal vortex has a velocity field given as
Γ
vθ (r) =
2πr
This velocity field is provided in cylindrical coordinates, thus the flow
is in the θ direction but only depends on r. This velocity field can be
plugged into Euler’s equation and would be found to be an acceptable
solution. It is easy to show that the circulation of this vortex is
Z 2π Z 2π Z 2π
Γ Γ
vθ (r)rdθ = rdθ = dθ = Γ.
0 0 2πr 0 2π
The circulation is a constant regardless of the radius of the circle we
choose to perform our calculation along. The vorticity, which only has
a component in the z direction is,
1 ∂rvθ 1 ∂ Γ
ωz = = =0
r ∂r r ∂r 2π
zero everywhere, except at the origin where it is undefined.
Now consider two vortices of equal circulation but opposite sign, sep-
arated by some distance, Figure 10.6. One vortex will induce a velocity
10.4 Vorticity 165
Image vortices
Solid
wall
Figure 10.6 Motion of two ideal vortices of opposite sign approach-

ing a wall. The velocity field of each vortex propagates the other.
As the vortices approach the wall, the introduction of an imagined
image vortex can meet the boundary condition at the wall of no flow
penetrating the surface. As the four vortices approach the wall, the
induced velocity fields cause then to spread apart.
field that will push the other forward. Since the vortices just move with
the fluid, that is what Dω/Dt = 0 says, then you can think about the
velocity of one vortex simply acting to move the other. Since the vor-
tices are the same strength they push each other in a straight line. If
you created two vortices of equal strength but the same direction of the
spin, the two vortices would orbit each other in a circle.
Now repeat your experiment with the spoon near the wall of your
bowl. You should be able to see that the vortices propagate themselves
to the wall, and then spread apart as they approach the wall. This
can be understood by imagining a set of image vortices on the other
side of the wall. These image vortices don’t really exist, but you should
be able to convince yourself with symmetry arguments that the flow
generated by the image vortices will always cancel the real vortices such
that there is no velocity penetrating the wall. The flow on the left side
of Figure 10.6 where the real vortices exist is the same whether there
was the wall or the set of symmetric image vortices. Figuring out the
qualitative flow is easier by visualizing the image vortices. Now as the
four vortices approach the wall, you can see that the image vortices
push each other outward if you consider the action of each vortex on
the other.
10.4.2 3D flows
In 3D, the above analysis just gets a little more complicated. First of all,
the circulation theorem remains, so without proof we will simply state
that in 3D the circulation is constant around a material loop. However,
simple applications of the circulation theorem (at least in this course)
are usually simple 2D approximations to give us some qualitative un-
derstanding of the flow.
The vorticity equation in 3D is,
Dω
= ω · ∇v.
Dt
and has a non-zero term on the right hand side. Vorticity for a material
point is not constant in 3D as there is an extra term in the equation.
This extra term accounts for things which cannot occur in a 2D flow.
One effect is vortex stretching. If a vortex is stretched out by a flow,
it intensifies, just like an ice skater speeding up as they pull their arms
in. If a vortex is squashed, the vorticity decreases. The stretching of
the vortex can be see in any bathtub drain. The vortex gets intensified
as it stretches down the drain. Another easy experiment to see vortex
stretching is to take a soda bottle and fill it with water. Turn it upside
down over the sink and as it is glugging, give it a strong swirl by hand.
You will set up a vortex that will be intensified as it is stretched and
pulled down out of the bottle. There are also effects that occur due
to vortex tilting. Tilting a vortex line can change a particles rotation.
This effect would be analogous to the physics demo where the instructor
takes a spinning bicycle wheel and tilts it while sitting on a stool which
is free to rotate.
In vortex dynamics we refer to vortex lines and tubes. A vortex line is
one which follows the vorticity vector, like a stream line. It would be the
axis of a tornado. A vortex tube would be a collection of vortex lines and
the vorticity vector is everywhere parallel to the surface of the vortex
tube. In 3D there are three laws of vortex motion that were derived by
Helmholtz in 1858. These laws are good for the approximations we have
been dealing with - inviscid and constant density. Helmholtz’s laws are
(Saffman (1992)).
• Fluid particles initially free of vorticity remain free of vorticity.

• Vortex lines and tubes move with the fluid.
10.5 Irrotational flow 167
• The strength of a vortex tube does not vary with time.
Due to the definition of vorticity as the curl of the velocity, it true that
∇ · ω = 0 since the divergence of the curl of any vector is zero (you
should quickly see if you can prove this fact to yourself). Since vorticity
is divergence free, then a vortex tube must have constant strength along
its length. While this last statement might not sound obvious, the proof
is precisely the same as showing that the total volumetric flow rate
through a pipe of varying cross section area must be the same at every
cross section if the flow is incompressible. If a flow is incompressible,
∇ · v = 0. The walls of the pipe are like the walls of the vortex tube
- since by definition there is no vorticity normal to the surface of the
vortex
R tube. At every cross section of Rthe tube the total flux of vorticity,
ω·n, is a constant just like in a pipe v·n is a constant. If the strength
of the vortex tube is constant everywhere along the length, then vortex
tubes must either close on themsleves (i.e. a smoke ring), go to infinity,
or end on solid boundaries. An example of vortex tubes are the long
white contrails seen behind airplanes on clear days when the weather
conditions are right. The two long white streaks are intense vortex
tubes which extend far across the sky. If you observe these streaks
closely you will notice that some distance away from the plane, the
individual tubes start developing a wavy character. These waves are
due to a fluid instability known as the Crow instability. Notice that
as the amplitude of the waviness increases the two tubes will interact,
cross, reconnect, and form a series of vortex loops. These circles are now
closed vortex tubes. How well you can make these observations depends
on the weather as the level of atmospheric turbulence can disrupt the
vortex visualization.
10.5 Irrotational flow

In a 2D flow we said that the vorticity equation says that Dω/Dt = 0,
the vorticity remains constant from a material point of view. Thus, if a
flow starts with no vorticity, none can be created. This statement was
Helmholtz’s first vortex law. You need viscosity to create and destroy
vorticity, thus you can’t make vorticity with Euler’s equations. You can
also see by the general 3D version of the vorticity equation, that if the
flow has zero vorticity initially, none can be created. Therefore, it is

common when dealing with Euler’s equations to consider irrotational
flow, that is a flow where ω = 0 everywhere.
If the flow has no vorticity, then you might remember a trick used
in electrostatics. The curl of the gradient of a scalar function is always
zero, this statement is an identity from vector calculus. Therefore, if
the flow has no vorticity (curl), then we can define a velocity potential,
v = ∇φ,
which is 100% analogous to the electric potential used in electrostatics.

If we can define a velocity potential, then conservation of mass ∇·v = 0
yields
∇2 φ = 0.
This equation is known as Laplace’s equation. At a solid surface we

would have the boundary condition that ∇φ · n = 0 which says that
the flow cannot penetrate the surface; v · n = 0. This equation is really
easy to solve, especially numerically if you have a complicated geome-
try. Once you have the velocity you can easily calculate the pressure,
especially in a steady flow. If you go back to our derivation of Bernoulli,
you will find that we wrote the momentum equation in the following
form,

1 2
ρv × ω = −∇ ρv + P + ρgz .
2
If the vorticity is zero, then it is clear that the left hand side is zero.
Therefore, in an irrotational flow Bernoulli’s equation holds everywhere,
not just along a streamline. Therefore,
1 2
ρv + P + ρgz = constant everywhere for irrotational flow.
2
So in irrotational flow, you solve Laplace’s and then plug the velocity
field into Bernoulli’s equation to solve for the pressure. This is a very
simple theory which is quite amenable to solution, which is in contrast
to the money you would earn for a general solution to the Navier-
Stokes equations. However, the utility of irrotational flow is quite lim-
ited. There are a number of good books that are filled with solutions to
potential flow and a number of good online calculators that will solve
potential flow problems for you. Potential flow finds a few niche appli-
cations, can many times (especially in aerospace applications) provide
a good qualitative picture of the flow and the theory is extremely im-
portant in the history of the development of the field of fluid dynamics.
However, the practical utility of potential flow is limited.
10.6 Lift on an airfoil

Euler’s equations can be used to predict lift forces on an airfoil. Eu-
ler’s equation, since it has no viscosity, cannot predict drag. Imagine
an airfoil with flow at uniform velocity coming at it. We can use our
equations for irrotational flow to solve this problem. We can assume the
flow is irrotational because imagine the airfoil is at rest. There is no
motion and thus no vorticity. Let’s accelerate our airfoil, and since this
is an Euler flow, we create no vorticity. The flow is irrotational and we
can then solve ∇2 φ = 0 around the airfoil shape to obtain the steady
flow patterns. The solution would show that streamlines approach and
leave the airfoil at two points near the leading and trailing edge. How-
ever the flow does not match what we observe experimentally. The flow
experimentally, shows that the streamlines leave the body at the trail-
ing edge. The only way we can make the calculated and observed flow
match is that if we add a net circulation to the flow. This circulation
effectively moves the point where the streamlines connect with the solid
airfoil. The behavior with and without circulation is shown in Figure
10.7 for a flat plate airfoil.
This seems arbitrary? Why would there be net circulation? The
source of the circulation is subtle. Imagine we start an airfoil from
rest. We draw loop around the airfoil and since everything is at rest,
the circulation is zero. If we impulsively start an airfoil, we would see
that a vortex gets shed from the airfoil at the instant that it starts.
This is just like moving the spoon in the water. However, the circula-
tion theorem would tell us that the circulation around the loop is still
zero. Therefore, this starting vortex would need to be compensated by
some other circulation of opposite sign. This is the circulation around
the airfoil. When we impulsively start the airfoil, viscosity acts to move
the point where the streamline separates from the airfoil. In doing so,
a) b)
Figure 10.7 A flat plate airfoil in potential flow. In (a) there is no

circulation and streamlines wrap around the rear of the airfoil. In
(b) there is circulation and the rear streamline leaves the airfoil at
the trailing edge.
vorticity is created and the counteracting circulation around the airfoil

must keep the net circulation at zero.
While the starting vortex shows that circulation must exist around
the airfoil, the idea that the circulation must exactly the amount that
moves the streamline to the trailing edge is not something that we can
prove. It is just a useful rule of thumb that turns out to work as long as
the airfoil is not at too steep of an angle of attack. This rule of thumb
is known as the Kutta condition. Once you know the flow field from
the solution of Euler’s equation, you can find the pressure from the
fact that in the absence of gravity effects Bernoulli holds everywhere,
P + 12 ρ(∇φ)2 = 12 ρU02 . Once you know the pressure, you you can find the
lift on the airfoil by integrating the pressure around the surface of the
airfoil and it turns out to actually agree pretty well with experiments.
For the same dimensional analysis arguments as with drag, when we
study lift we define the coefficient of lift in a similar way,
Force/width
CL = 1 2
,
2 ρU `
where ` is the length of the airfoil (which must be precisely defined

for any shape that is not a flat plate). If we dig into the mathematical
theory of potential (or irrotational) flow, it turns out we don’t need a
computer to solve for the flow field around a flat plate inclined to the
flow. The analysis (which you can find in a number of books if you are
-Γ Γ
Figure 10.8 Schematic of the starting vortex for an airfoil. The cir-
culation calculated around the outer oval must be zero, so the circu-
lation around the two inner loops must cancel each other out. The
airfoil must have circulation of the opposite sign as the starting vor-
tex. If the airfoil is suddenly stopped a vortex of the opposite sign as
the starting vortex is shed in order to maintain no net circulation.
interested) yields a surprisingly simple formula,

CL = 2π sin α,
where α is the angle the airfoil makes with the flow field, called the angle
of attack. This simple theoretical results compared to experimental
data is shown in Figure 10.9a. In the figure we see good agreement. At
higher angles we observe flow separation and a complex wake behind
the airfoil - the simple inviscid theory fails. At a high angle of attack,
flow separation can be so severe as to lead to stall.
You may have noticed that most planes you fly don’t have flat plates
for wings. In 1933 NACA (a predecessor of NASA) concluded a large
study of measuring lift and drag coefficients for numerous airfoil shapes.
A sample of that orignal data for a single airfoil is shown in Figure
10.9b. The figure shows the coefficient of lift, coefficient of drag, the
ratio of lift to drag (L/D), and the center of pressure (c.p.) as a function
of angle of attack. It might take a little study to read the plot. However,
if you look carefully you can see that the lift coefficient is well described
at modest angles of attack by the formula CL = 2π sin(α). Note that at
low angles of attack the drag coefficient is very small. The measurement
behavior is pretty close to what Euler equations predict; i.e. no drag.
It turns out that there is a mathematical tricks that allows us to
compute the full velocity field for different airfoil shapes in potential
flow. The trick allows us to transform the solution for flow around a
cylinder to flow around different airfoil shapes. Called conformal map-
a) b)
Figure 10.9 a) Coefficient of lift from experiments for an inclined

flat plate compared to the single equation from irrotational flow
theory. Data from “Low speed airfoil data” by Williamson et al. b)
Lift and drag data from the original 1933 NACA study for a single
airfoil.
ping, if we have a mathematical expression to squish a cylinder to a

new shape, we can use the same expression to squish the solution too (I
am oversimplifying things of course). While conformal mapping is well
beyond what we would cover in a first course, it is interesting to note
that airfoil theory is one area where our potential flow solutions work
quite well and a fully analytical approach provides useful predictions.
11
Boundary layers
The analysis of fluid flow via Euler’s equations seems relatively nice.
While we presented only a few highlights in this book you should have
gotten the impression that a number of problems could be readily solved
mathematically or computationally using Euler’s equations for incom-
pressible flow. So while we know viscosity is always with us, it is a small
force compared to the others in the problem in a number of applica-
tions. At the scale of humans, everything is high Reynolds number.
However, it has been observed for a long time that Euler’s equations
do an unusually bad job at making useful predictions. Euler’s equation
predicts no drag force, for example. It was not until the early 1900s
when Prandtl introduced the idea of the boundary layer, that people
began to appreciate why Euler’s equations were so poor and how to
rectify our understanding (Anderson (2005)).
If we have a high Reynolds number flow around a streamlined ob-
ject such as an airfoil, what is observed is that the fluid velocity just
away from the surface of the airfoil very closely matches the solution to
Euler’s equations. This part of the flow is not influenced by viscosity.
As we approach the surface, the flow velocity goes to zero because the
no-slip condition is obeyed. However, the size of this boundary layer is
extraordinarily thin. Flow over an airplane wing could have boundary
layers measured in the units of millimeters. Why should this thin region
play such a major role in determining flow? The answer is the topic of
this chapter.
We need to keep in mind all the forces at play in our problem. When
we looked at solutions to the Navier-Stokes equations we found we
could readily solve problems where there was no fluid acceleration - the
174 Boundary layers
balance was between viscosity and some driving force such as applied
pressure. When we discussed Bernoulli’s and Euler’s equation, we con-
sider the balance between pressure and acceleration. In this chapter we
will find that in many cases there is a complex interplay of fluid ac-
celeration, viscosity and pressure. While our analytical tools get more
complex as the flows do, all is not hopeless and there is a lot we can
do with simple problems.
11.1 Impulsively started plate

Let’s start with an idealized problem that has an analytical answer to
begin to explain this boundary layer. Imagine a flat plate that is infinite
in extent in the x direction. The plate and the fluid above it are initially
at rest at t = 0. The fluid above the plate extends to infinity. At t = 0,
the plate is impulsively started to move with a constant velocity U .
You should turn back to the chapter on Navier-Stokes solutions and
look at the problem of transient Couette flow. The stated problem here
is the same, only in the Couette flow the domain was bounded in the
y direction whereas here the domain extends infinitely in y. Since the
two problems are so similar, the development of the reduced form of
the Navier-Stokes is identical as transient Couette flow.
The reduced form of the Navier-Stokes in terms of the kinematic
viscosity, ν, is the 1D heat equation,
∂u ∂2u
= ν 2.
∂t ∂y
Recall that the kinematic viscosity plays the same role in this equation
as thermal diffusivity in the heat equation. We expect that the behavior
would be the same as applying an instantaneous change in temperature
to a semi-infinite solid.
We now want to make this equation dimensionless as we did with
Couette flow. We can set the velocity scale to ũ = u/U . However, what
should the time scale or spatial scale be? There is no natural geometric
length to scale length by. Our only choice is δ, the thickness of the fluid
region which is in motion. However this is an unknown scale. Using an
11.1 Impulsively started plate 175
arbitrary length ỹ = y/δ and t̃ = t/t0 would yield,

∂ ũ νt0 ∂ 2 ũ
= 2 .
∂ t̃ δ ∂ ỹ 2
This scaling indicates that the thickness of the region where the flow
transitions from the plate velocity, to being at rest grows with the
square root of time and depends on the kinematic viscosity of the fluid,
√
δ = νt.
This scaling of the growth of the viscous layer what we expect for all
diffusion problems - in the boundary layer it is diffusion of momentum.
The scaling argument is confirmed by the analytical solution to this
problem, which we will not derive but has the form,

u(y, t) y
= 1 − erf √ .
U 2 νt
where erf is known as the error function. If you are unfamiliar with
the error function, you can look up it’s definition and you will find
that most mathematical software will understand the definition and
generate a plot of the function for you. The dimensional solution is
shown in Figure 11.1. Since ν = 10−6 m2 /s in water, you can see that
the length scale of the boundary layer is pretty small. If we pulled the
plate for an hour in water, the thickness of the boundary layer would
be 6 cm.
What is interesting about the analytical solution is that the solution
√
is a single function of a variable y/(2 νt) which mixes the spatial and
time dependence of the problem into a single variable. This feature is
called a similarity solution. Similarity solutions are common in diffusion
problems where there is no natural length scale. In this problem
√ the
length scale for the problem is the boundary layer thickness, νt, which
depends on time. What is interesting about the solution is that the
shape of the boundary layer remains the same, it just gets thicker and
thicker with time.
A somewhat obvious but crucial point here is that the boundary
layer will exist regardless of the smallness of viscosity, as long as it is
not zero. Unless you are working with liquid helium near zero kelvin
and approach an exotic state known as the superfluidity, all fluids have
some viscosity. So even if viscosity is small, the boundary layer still
exists. If I am close to the surface I cannot ignore viscosity.
176 Boundary layers
Figure 11.1 Solution to the problem of an impulsively started plate

at different instances
√ in time. Here we set ν = 1. If the vertical axis
is plotted as y/(2 νt), then there is a single curve.
11.2 Boundary layer equations and laminar solution

Imagine flow over a stationary flat plate, a problem similar but different
from the previous section. The flow approaches the plate with uniform
flow velocity, U∞ . The plate is flat and thus far from the plate we expect
the flow not to be disturbed. At the surface of the plate, the velocity
must go to zero due to the no-slip condition. What is observed is that
a boundary layer region grows from the leading edge of the plate. The
boundary layer is simply the transition region where the velocity goes
from zero at the plate surface to the free stream value. The schematic
of the observed boundary layer is shown in Figure 11.2.
If we go to dimensional analysis and the Pi Theorem to ask what
is the boundary layer thickness, δ, at the end of a plate of length L.
The boundary layer thickness δ could then depend upon variables L,
U∞ , µ and ρ. The Pi Theorem would tell us that there are 5 variables
expressed in 3 dimensions so there should be only two dimensionless
parameters. The dimensional analysis would yield,
δ
= f (Re)
L
where the Reynolds number based on the plate length is ReL = ρU∞ L/µ.
U∞ U∞
Boundary
δ
Layer
Figure 11.2 Schematic of the laminar boundary layer which develops

as a uniform flow approaches a flat plate.
The boundary layer thickness can only depend upon the plate Reynolds
number.
We could estimate the dependence by a simple √ argument using the
solution to the impulsively started plate, δ ∼ νt. In the boundary
layer problem, there is no time. At a given downstream location, the
“time” is the time it takes fluid to reach that point, i.e. t ∼ L/U∞ .
√
r
νL
δ ∼ νt ∼ .
U∞
In terms of the Reynolds number this equation can be rewritten as
r r
δ ν 1
∼ ∼ .
L U∞ L Re
Since the flow progresses downstream, we the argument above works
for any x location. Thus we might expect,
r r
δ ν 1
∼ ∼ ,
x U∞ x Rex
where we use the Reynolds number based on the x location. This scaling
suggests that the boundary layer thickness grows as the square root
of the distance down the plate. To get a sense of scale, lets take air
ν = 10−6 flowing at U∞ = 1 m/s over a 1 m length plate. The boundary
layer would be 1 mm. If the velocity were increased to 100 m/s (an
airplane taking off), the boundary layer thickness would be 100 microns.
The shear stress on the plate due to viscosity can also be estimated.
For a Newtonian fluid, you can look back at the previous chapter and
178 Boundary layers
see that the shear stress (tangential to the surface) is τ = µ∂u/∂y. The
stress is given by the velocity gradient at the surface. The shear stress
would be approximately,
∂u U∞ U∞ µU∞ p
τ (x) = µ ∼µ ∼µ ∼ Rex
∂y δ xδ/x x
Making an analogy with drag, we would follow dimensional analysis to
define the coefficient of friction as,
τ (x)
Cf = 2 /2
.
ρU∞
Using the estimate we have already obtain for τ we have an estimate
for the scaling of the coefficient of friction,
r
µU∞ p 2µ p 1
Cf ∼ 2
Rex ∼ Rex ∼ .
xρU∞ /2 xρU∞ Rex
This simple estimate predicts that the shear stress should be highest
near the leading edge of the plate and continually decrease with down-
stream distance. The Rtotal drag force, F , for a plate of length L, is then
calculated from F = τ (x)dx. The total drag coefficient is thus given
as,
F 1
CD = 2 /2
∼√ .
LρU∞ ReL
From such simple arguments we have no way to compute what the
pre-factors in these estimates might be. However, we find that we can
get the basic scaling and qualitative behavior for the laminar boundary
layer.
11.2.1 Full solution to the boundary layer

The full boundary layer problem over a flat plate has a mathematical
solution. This solution makes use of the fact that δ/x is usually a pretty
small number. The basic idea is that when there is a vast difference in
scales, one can make the equations dimensionless using the different
scales. So we will scale the x coordinate
q by the length of the plate, L,
and the y direction is scaled by δ = UνL ∞
. We will also use different
scales for the u and v velocities; ũ = u/U∞ and ṽ = v/V0 .
U∞ U∞
δ
Figure 11.3 Schematic of control volume for the boundary layer

problem. The flow into the left side of the control exits from the
right side with a deficit on the order of U∞ δ. The extra flow that
came through the left side must exit the upper surface of the control
volume. The magnitude of this flow is on the order of V0 L.
Conservation of mass gives.

U∞ ∂ ũ V0 ∂ṽ
+ =0
L ∂ x̃ δ ∂ ỹ
This equation implies that V0 = U∞ δ/L would be a good choice. This
scaling implies that the vertical velocities are less than the horizontal
velocities by a factor of δ/L. If you think about this comment for a
minute, it should make some physical sense to you. The schematic in
Figure 11.3 shows a control volume picture of the scaling.
Applying this scaling for V0 and making the equation dimensionless
would yield,
2 2
δ ∂ ũ ∂ 2 ũ

∂ ũ ∂ ũ ∂ P̃ µ
ũ + ṽ =− + + 2
∂ x̃ ∂ ỹ ∂ x̃ ρU∞ L L2 ∂ x̃2 ∂ ỹ
and
δ2 δ2 µ δ 2 ∂ 2 ṽ ∂ 2 ṽ

∂ṽ ∂ṽ ∂ P̃
ũ + ṽ =− + 2 + .
L2 ∂ x̃ ∂ ỹ ∂ ỹ L ρU∞ L L2 ∂ x̃2 ∂ ỹ 2
Since δ/L is a small number and we have scaled all the partial derivative
terms in our equation to be on the order of magnitude of 1, we can take
the perhaps somewhat uncomfortable step of removing all terms with
small numbers out in front. This step seems odd the first time you see
it. However, it is pretty standard practice. It is a logical thing to do.
180 Boundary layers
We have supposedly scaled all the partial derivatives in the equation

to be on the order of 1. Thus the terms with (δ/L)2 simply go away
because they are small. Taking this leap of faith provides the boundary
layer equations,
∂ ũ ∂ ũ ∂ P̃ 1 ∂ 2 ũ
ũ + ṽ =− + (11.1)
∂ x̃ ∂ ỹ ∂ x̃ Re ∂ ỹ 2
and
∂ P̃
0= . (11.2)
∂ ỹ
The second equation simply states that the pressure across the bound-
ary layer is a constant. The fact that the pressure across the boundary
is constant is really useful. What you can do, in practice, is solve Eu-
ler’s equations for flow around your object. This solution gives you the
pressure at the surface of the object. You can assume that since the
boundary layer is so thin that the pressure at the surface of the ob-
ject and the pressure at the edge of the boundary layer would be the
same. Thus, you get the pressure and ∂∂P̃x̃ from the solution to Euler’s
equation, and then use the now known pressure gradient to solve the
boundary layer equations, Equation 11.1.
In the flat plate problem, the pressure “far away” from the plate (in
units of δ) is just a constant and there is no pressure gradient. Thus,
our final boundary layer equations for the two components of velocity
are,
∂ ũ ∂ ũ 1 ∂ 2 ũ
ũ + ṽ =
∂ x̃ ∂ ỹ Re ∂ ỹ 2
∂ ũ ∂ṽ
+ =0
∂ x̃ ∂ ỹ
The boundary layer equations describe a balance between viscosity and
fluid acceleration.
There is a mathematical solution to this flat-plate problem, and we
will skip the details and only show the result. The analytical solution
is not really a closed form answer. The mathematical solution simply
transforms this partial differential equation into an equivalent ordinary
differential equation. This ODE still needs to be solved numerically.
The solution to the velocity profile is shown in Figure 11.4.
Once the boundary layer equations are solved numerically, you can
Figure 11.4 Solution to the laminar boundary layer problem over a

flat plate. The figure shows the
q universal velocity profile near the
U∞
plate. The y axis variable is y xν
.
calculate everything you want to know about the boundary layer. The
thickness, defined as the distance when the velocity field is 99 % of the
free stream is,
δ 5
=√ .
x Rex
The factor of 5 comes from the solution, but the form was acquired
through dimensional arguments. The shear stress vector at a point in
a 2D flow is given by n · T. We can evaluate the stress at the point to
be in dimensionless terms,
τ (x) 0.664
Cf = 2 /2
=√ ,
ρU∞ Rex
where Cf is known as the coefficient of friction. The total drag force
(per unit width) is thus given as,
F 1.328
CD = 2 /2
=√ .
LρU∞ ReL
The numerical values come from the calculated numerical solution. The
really important point is that the basic scaling of the answer comes from
consideration of dimensional analysis and some physical arguments.
182 Boundary layers
11.3 Turbulent boundary layers

Any flow at high Reynolds number is subject to becoming unstable
and turbulent. If the boundary layer has become turbulent, there is
no mathematical solution to the problem and the laminar solution is
invalid. There is an extraordinary amount of theoretical, experimental,
and numerical work on the problem of the turbulent boundary layer.
The reason for all the interest is that the problem is important. Tur-
bulent boundary layers are seen in many applications and they play
a strong role in determining drag on boats, planes, and cars and thus
determines the energy use and efficiency of transportation. Boundary
layers also occur in the atmosphere and play a strong role in deter-
mining the weather and climate. It is difficult in a short introduction
like this to convey the importance of the turbulent boundary layer on
many problems of practical importance. There is no way I can ade-
quately convey both the amount that is known and unknown about
turbulent boundary layers.
The rule of thumb is that for a flat plate, the local Reynolds number,
Rex = ρU x/µ should be
Rex < 105 for laminar boundary layers.
In this definition, the local Reynolds number increases with distance
down the plate. Thus all boundary layers start laminar. For a sense of
scale, a 1 m/s flow in water would go turbulent around 10 cm from the
leading edge. The criterion for laminar flow is not an exact number.
It is a rule of thumb. The exact location of turbulence can be very
difficult to predict and is very sporadic, just like we have discussed in
other applications such as pipe flow. The exact location depends upon
whether an experiment is carefully done to remove all perturbations or
whether we are in a ”real world” application. For a carefully conducted
experiment where the upstream flow is very stable and you are careful
to make a perfectly smooth plate, the transition to turbulence occurs
at a higher Reynolds number.
Once the boundary layer is turbulent, we make use of the fact that
dimensional analysis tells us that
δ
= f (Rex ) .
x
Thus a small number of experiments can provide an empirical solu-
tion to the boundary layer thickness. Empirical fits to the data given
boundary layer thickness as,
δ 0.37
= ,
x Re1/5
x
and the coefficient of friction as

0.0592
Cf = .
Re1/5
x
These are not fundamental laws and there is no mathematical deriva-

tion, the expressions just fit the data well. Notice that the growth of
the layer with distance down the plate is
δ ∼ x4/5
which is much more rapidly than with the laminar boundary layer.
The turbulent velocity profile is thicker, but much blunter than the
laminar one. The random turbulent eddies pull high momentum fluid
from the free stream and bring it close to the solid surface. The viscous
drag is much higher in turbulent flow due to the steep velocity gradients
at the surface. The high momentum fluid sits closer to the plate as
shown in Figure 11.5a. So while the turbulent layer is thicker than the
laminar one, the velocity gradient at he surface is much steeper. The
boundary layer thickness as a function of distance down the plate are
shown in Figure 11.5b. There is a jump at Re = 105 in the two solutions.
In reality, there is not a jump but a transition regime from one solution
to the other. The thickness of the boundary layer rapidly expands when
the boundary layer becomes turbulent. Because the velocity gradient on
the blunt turbulent velocity profile is greater than the laminar one, the
shear stress at the solid surface is greater for the turbulent boundary
layer.
11.4 Boundary layer separation

Now consider 2D flow over a cylinder. Euler’s equations would predict
streamlines that approach and leave the centerline. The flow would
have fore-aft symmetry as shown in Figure 11.6a. This is not what is
observed. What is observed is an increasingly complicated asymme-
try. The asymmetry starts with the flow behind the cylinder to have
184 Boundary layers
a) b)
Figure 11.5 a) Experimental data for a turbulent boundary layer

in air. The data is openly provided by Princeton University Gas
Dynamics lab. Note the very steep velocity gradient at the plate
surface b) Boundary layer thickness as a function of distance down
the plate from the empirical correlations. The jump from the lower
laminar solution to the upper turbulent one would occur around the
location of the vertical line. For x < 0.05 we observe the laminar
boundary layer solution, for x > 0.05 m we observe the turbulent
one. In reality the transition happens over some distance and is not
a vertical jump.
some small vortices or recirculation. As the flow speed is increased,

an increasingly complicated structure of the wake emerges. What we
are seeing in the complex wake is a phenomena called boundary layer
separation.
The boundary layer equations derived in the previous section provide
a useful and practical way to calculate the flow using Euler’s equations
and the boundary layer equations. The basic idea is that you solve Eu-
ler’s equation for an arbitrary shaped object. This solution provides
the pressure everywhere along the surface. Since the pressure is con-
stant across the boundary layer, Euler’s equation provides ∂P/∂x in
the boundary layer equation.
Let’s apply this idea to flow over a cylinder. If we think about the
pressure, Bernoulli’s (Euler) equation would tell us that the pressure at
the stagnation point at the front side of the cylinder would be greater
than the ambient pressure by an amount 21 ρU 2 . Since the flow is sym-
metric by Euler’s equation, the pressure at the rear stagnation point
is also higher than the ambient. At the top and bottom of our cylin-
der, the pressure is low. The velocity will speed up to scoot around the
cylinder. Thus the pressure at the top and bottom are low. The pres-
sure along the surface of the cylinder as predicted by Euler’s equations
a) b)
Figure 11.6 a) Streamlines for irrotational, inviscid flow over a cylin-

der. b) The pressure along the surface of the cylinder as a function
of angle. The data are plotted as the coefficient of pressure which is
defined as Cp = P1 −P ∞
ρU 2
. The pressure at the forward and rear stag-
2 ∞
nation points is higher than the ambient pressure while the pressure
at the top of the sphere is the lowest. The forward and rear stagna-
tion points are at 0 and π.
are shown in Figure 11.6. The pressure distribution as predicted by the

equations follows our simple qualitative picture.
As flow reaches the front of the cylinder and moves toward the top,
the pressure gradient in the boundary layer, dP/dx, is such that the
flow is going from a region of high to low pressure. We saw the pressure
gradient is favorable in that it is pushing with the direction of flow. As
the flow proceeds from the top of the cylinder toward the rear stagna-
tion point, the pressure is increasing with the flow direction. We saw the
pressure gradient is adverse as it pushes against the flow. When solving
Euler’s equation this adverse pressure gradient is perfectly balanced by
the momentum of the fluid; the pressure gradient slows the fluid down.
However, the fluid near the surface of the cylinder, inside the boundary
layer, is deprived of its momentum due to the viscous boundary layer.
The adverse pressure gradient pushes back on the fluid in the boundary
layer and the fluid in this region has insufficient momentum to over-
come the adverse pressure gradient. This adverse pressure gradient can
cause reverse flow in the boundary layer on the back side of the cylin-
der. This reverse flow is unstable and creates what is called boundary
layer separation. This separation can be seen in the experimental image
in Figure 11.7.
186 Boundary layers
Figure 11.7 Boundary layer separation for flow over a cylinder

(From Album of Fluid Motion). The Reynolds number is 26. For
Reynolds numbers less than 1, the flow shows fore-aft symmetry.
11.5 Observations of drag on a sphere

Lets reconsider the problem of drag on a sphere moving at constant
speed. In theory we can solve the Navier-Stokes equations to obtain
the velocity field and pressure at every point. In practice we can create
the solution numerically using CFD software. Once the velocity and
pressure is known, we can compute the net force on the sphere by
integrating the stress vector around the surface of the sphere. The stress
vector is the dot product of the normal vector and the stress tensor,
n · T. The force is found by the the surface integral of the stress,
Z
F = n · (P + µ(∇v + (∇v)T ))dS
Note that the total force has a contribution from pressure and viscous
stresses - distinguishing between these two forces will be important to
our understanding soon. From dimensional analysis, we know that the
drag force is expressed as the drag coefficient which is only a function
of the Reynolds number
F
1 2 2
= Cd (Re).
2 ρU0 πR
Knowing that there is some functional relationship for the drag coef-

flow around a sphere.
ficient as a function of Reynolds number allows us to collapse experi-

mental results for all sphere sizes at all velocities and in all fluids onto
a single master curve, Figure 11.8.
Now that we know a little bit about boundary layers, let’s look more
closely at the data in Figure 11.8. At low Reynolds number there is an
analytical solution which provides CD = 24/Re - a solution supported
by the experiments. As the Reynolds number is increased, the drag co-
efficient limits to a constant of approximately 0.4. The low Reynolds
number solution is lost once boundary layer separation begins to set
in. These two behaviors were previously explained as the transition
from viscosity dominated flow to inertia (or density) dominated flow.
As the Reynolds number is further increased about 105 , the drag co-
efficient goes through a sudden decrease. This dip is very unexpected.
We might think that as the flow increases, the boundary layer would
eventually become turbulent and the drag should go up. Why did it
instead decrease?
The reason is that the total drag on the sphere is given by pressure
drag and friction (or viscous drag) along the surface. You already know
about different types of drag, even if you don’t know that you do.
Next time you are in the car, put your hand out the window. Turn
it horizontal with the ground and then turn it so it is vertical to the
ground. Feel the difference? The difference you feel is pressure and
friction. On a flat plate (your horizontal hand), there is only friction at
188 Boundary layers
the surface due to the velocity gradient. The shear stress at the surface
depends upon the viscosity of the fluid. On a sphere, the stress vector
at a point is given by the velocity gradient at the surface and the local
pressure. In Euler’s equation the pressure on the surface of the sphere
is symmetric, thus the net force due to pressure is zero. If we watch the
experimental streamlines at high Reynolds number where the boundary
layer around the sphere is still laminar but strong separation sets in, we
would see the streamlines move around the sphere, but do not recover
on the other side. If we measured the pressure on the surface we find
strong asymmetry when there is flow separation. Thus pressure at the
forward stagnation point is greater than the ambient by ρU 2 /2. On the
back side of the sphere, the wake is relatively quiet. If we look at the
streamlines at the top or bottom of the sphere, they would be straight
and thus the pressure would be the ambient. Thus the pressure in the
wake should be something close to the ambient pressure. Therefore, very
crudely, the force acting on the front half of the sphere due to pressure
is 21 ρU 2 πr2 . The force in the back half is zero. Thus the net force due
to pressure drag is on the order of 12 ρU 2 πr2 . Thus the drag coefficient
is about 1. Pressure drag dominates over friction at the surface in this
regime.
However, we said that separation occurs because momentum is de-

prived from the fluid in the boundary layer. When the boundary layer
becomes turbulent, the velocity profile gets blunter. The transition from
zero velocity at the wall to the free stream value occurs over an even
smaller distance than the laminar case. Thus, high momentum fluid is
even closer to the wall. The high momentum fluid near the wall allows
the turbulent boundary layer to push into the adverse pressure gradient
further than the laminar counterpart. Separation is delayed and occurs
further down the back of the sphere. This effect can be seen in the
experimental images in Figure 11.9. The low pressure wake acts over
a smaller area when the boundary layer is turbulent rather than lami-
nar. Pressure drag is decreased when the boundary layer is turbulent.
So even though the steeper velocity gradients in a turbulent flow in-
crease the friction at the surface, the net drag is dominated by pressure
drag and thus the overall drag force (coefficient) decreased when the
boundary layer becomes turbulent. The pressure drag is reduced due
to the delayed separation. Since the pressure drag dominates over the
Figure 11.9 Boundary layer separation for a laminar (upper picture)

and turbulent (lower picture) boundary layer. The delay of separa-
tion is clearly seen in the case of the turbulent boundary layer.
viscous drag, the net effect is that the turbulent boundary layer has
lower overall drag.
The role of boundary turbulence in reducing drag is the reason for
dimples on golf balls. The irregular surface induces turbulence to occur
at lower Reynolds number than a smooth sphere. The turbulent bound-
ary layer delays separation and the golf ball goes further. Imperfections
on the surface of a sphere and its effect on tripping the boundary layer
into a turbulent state can help explain why curveballs curve in baseball.
12
Turbulence
As we have repeatedly stated, the equations of fluid dynamics are non-

linear and challenging to solve. One of the key challenges is that many
of the engineering flows we are interested in are high Reynolds number,
and thus turbulent. Over the past century there have been countless
researchers who have devoted their careers to trying to “solve” the
turbulence problem. The first studies of turbulence are often credited
to Leonardo Da Vinci, who has a number of sketches of turbulent flows
in his work. Nobel physicist Richard Feynman called turbulence one of
the most important unsolved problems of classical physics. Even with
all this effort there are perhaps even more questions than answers. I do
not mean to say that because the problem is not ”solved” everything
is hopeless and unknown. There have been many modern experimental
and theoretical advances in physics, mathematics, and engineering that
have come as the result of all this effort. I simply want to stress that
this is a very complicated topic and my treatment in a short chapter is
going to be embarrassingly incomplete. There are a few concepts you
should know, that is the point of this chapter.
12.1 Can we simulate turbulence?

There is nothing about turbulence that is not embodied in the Navier-
Stokes equations. In theory we could just get a computer to solve the
Navier Stokes equations numerically and solve for all the details. This
approach is called Direct Numerical Simulation (DNS) and it is feasible
if the Reynolds number is not too large. However, turbulence can chal-
192 Turbulence
lenge the fastest supercomputers and we can quickly exceed the limits
of the worlds fastest computers. For many problems, the solutions span
too many orders of magnitude in time and space to be resolved. Fur-
ther, the outlook is such that for many important problems waiting for
computers to get faster doesn’t seem like a good prospect for the near
future, even if Moore’s law continues to hold.
The problem is that scaling arguments show that the total amount
of computational effort to solve a 3D turbulent flow scales very poorly
with Reynolds number. The total amount of computational effort scales
as
total computing power ∼ Re3 .
To get one order of magnitude increase in Reynolds number we need

1000 times the computing power. If you look at curves of supercom-
puter performance over the years, we historically get about a factor of
1000 improvement over 10 years. A similar though perhaps a bit slower
time scale holds for desktop computers. This fact is actually quite re-
markable, making things that were cutting edge when I was a graduate
student trivial for today’s computers. However, for something like at-
mospheric flows we are perhaps 10 orders of magnitude in Reynolds
number or more off in what we can calculate and what we would like.
In Figure 12.1 is an image of a water jet at relatively low Reynolds
number, yet the complexity and diverse length scales are readily ap-
parent. Even by the most optimistic estimates it will be at least several
decades before a full DNS of the atmosphere scale is feasible.
You might be tempted to simply truncate the size a DNS and solve
for only the large features. If you are interested in the dominant features
of the climate do you really need to resolve everything? If we are trying
to model the weather, can’t we make the grid scale on the order of
several miles such that we don’t capture every detail but the dominant
motions? The reason you cannot simply truncate the simulation is due
to the energy cascade, which we now explain.
12.2 Kolmogorov theory and the energy cascade

If we consider homogeneous turbulence, turbulence that is the same
in all directions like a big box of shaken up fluid, we can make some
12.2 Kolmogorov theory and the energy cascade 193
Figure 12.1 Experimental image from a water jet at Reynolds num-

ber ∼ 2000. Even at this Reynolds number, the complexity in the
number of length scales seen in the flow is striking. From Album of
Fluid Motion.
progress by considering only dimensional analysis and a simple model.

This analysis put forth by Kolmogorov in 1941 remains a centerpiece
of turbulence theory.
The classic picture of turbulence is that of the energy cascade. The
flow is considered to consist of a number of turbulent eddies of different
sizes. The model is that the largest eddies split into smaller eddies,
which split into smaller eddies, which split into even smaller eddies.
Every time the eddies split the kinetic energy is equally divided into
new, smaller sized eddies. There is no loss in this process. The cascade
continues to smaller and smaller scales until finally the local Reynolds
number is sufficiently small that viscosity can take the energy away.
The energy at the smallest scales are turned into heat and the kinetic
energy vanishes. This cascade of energy can occur over several orders
of magnitude in length.
From Kolmogorov’s scaling arguments it turns out that we could esti-
mate the length scale where viscosity removes energy from the system,
η, as
L
∼ Re3/4
η
194 Turbulence
where L is the length scale of the largest eddies and the start of the
cascade. For an atmospheric flow we can easily reach Reynolds numbers
of 1012 or higher; there are indeed many orders of magnitude of length
over which the turbulence cascade exists.
We can describe the flow via a spatial Fourier series u(x) = Ak eikx ,
P
where the wavenumber k is like a spatial frequency. Kolmogorov argued

from physical arguments and dimensional analysis that the energy per
wavenumber follows a scaling,
−5
E(k) ∝ k 3
where
Z
Total kinetic energy = E(k)dk.
This predicted energy spectrum comes from simple arguments of units

and a little physical interpretation. Note that as k increases the length
scale decreases. The picture is there is an inertial range where viscos-
ity is not important and the energy in the fluid flow follows the Kol-
−5
mogorov scaling. Energy progresses down the k 3 cascade until such
large wavenumbers (small length scale) that viscosity becomes impor-
tant and the scaling no longer holds. This energy spectrum has been
measured and generally experiments agree quite well with this simple
scaling law, as seen in Figure 12.2.
Imagine a bucket of water that you are stirring constantly with a big
spoon. Energy is constantly being put into the system with a length
scale on the size of the stir (let’s say that is the size of the bucket). The
big swirls you create break apart and that energy goes into smaller and
smaller swirls. This cascade of energy from large scales to small scales
follows the Kolmogorov scaling. Once the eddy size gets small enough
that viscosity becomes important the eddy motion is halted and the
energy goes into heating the fluid. You reach a steady state where the
energy you constantly put into the system at the large scale is turned
into thermal energy at the small scale. Just to give some numbers to
this problem, let’s say our bucket is a foot in diameter (≈ 30 cm) and we
are stirring at 1 diameter per second (30 cm/s). This yields a Reynolds
number of Re = 90,000 and a dissipation length scale is about η = 50
microns. Thus our inertial cascade spans nearly 4 orders of magnitude.
The picture of turbulence presented here, hardly captures all the
Figure 12.2 Experimental measurements of the energy spectra. The

−5
scaling law E(k) ∝ k 3 is clearly observed over many orders of
magnitude. Remember, large wavenumbers correspond to small fea-
tures. The rapid decline in energy for k > 10 cm−1 corresponds to
where viscous dissipation takes over. From Champagne, Journal of
Fluid Mechanics, 1978.
complexity that is observed in reality. However, the scaling law is ro-

bustly seen in experiments and thus contains some of the essential
physics of the problem.
12.3 Numerical simulation of turbulence (CFD)

While CFD is a great tool and works robustly for low and “moder-
ate” Reynolds number flows, at high Reynolds number flows become
turbulent much more care is needed. One issue with CFD in turbulent
flows comes from the picture of turbulence in the previous section and
our example of the stirred bucker. Let’s return to the 30 cm bucket
at Reynolds number of 90,000. In order to resolve all the length scales
with DNS we need computational grid that would require about Re3/4
or 6,000 discrete points in each spatial dimension. For a 3D flow this
would be 2.16 × 1011 total grid points. This resolution is outside the
range of current computing capability. So now let’s say that we de-
196 Turbulence
cide that we only care about the big swirls so we decide to make our
grid spacing 1 mm. Certainly this will suffice? Now we have 300 grid
points in each direction; still a lot but not outside of what a modern
computer and software could handle in 3D. However, 1 mm is 20 times
greater than the length scale where viscosity can remove energy from
the system. What happens is that energy will be put in at the large
scale, that energy will cascade to smaller scales, and when the energy
in the simulation reaches 1 mm (the grid spacing) it will just sit there.
There is no mechanism to remove energy from the simulation. Energy
will build up and the simulation will die a sad numerical death.
There are fixes to this problem by adding models which act as numer-
ical sponges to soak up all the grid scale energy from the simulation.
Methods called Large Eddy Simulations (LES) have this feature of be-
ing able to resolve our stirred bucket by modeling, but not resolving,
the smallest scale motions. The LES models are often called turbulence
closure models. There are many different models that have advantages
and disadvantages in different situations. There is no single model that
always works the best and it requires training and experience to do LES
of turbulent flows well. You can easily see that this turbulence closure
problem will be with us for a while. For an atmospheric flow which
effect the climate, the Reynolds number can be 1015 the large scales
structures are the sizes of continents. We are going to be continue to
need closure models in weather and climate modeling in my lifetime.
Most modern CFD packages have a number of different turbulence
closure models built in. The variety and character of these models is
too extensive to cover here. In the software, it can be just as easy as
clicking from one model to the next. However the different models have
different limitations and cases where they work better or worse. Just
be warned that you should be skeptical of the results unless you have
had more training and experience than you are getting from this brief
chapter.
Many people think that CFD is the answer to everything fluids. The
impulse for many people when confronted with a fluids problem is to
run to CFD. CFD can be powerful, but it is a tool that can be easily
misused. If you are ever working somewhere and your boss wants to do
CFD of a turbulent flow, be cautious going down this road unless you
have learned more about fluid dynamics than this course. It is very easy
to get a commercial code to ”solve” but give you an incorrect answer.
You just need to be aware that CFD of turbulent, unstable, or otherwise

non-linear flows is not yet plug and play and will probably remain this
way for some time to come. CFD can be powerful, but you really need
to know what you are doing. Like Stan Lee noted in Spiderman - CFD
is powerful, “but with great power comes great responsibility”.
13
Control volume analysis
In the previous chapters, we went through great effort to develop a

differential law for fluid flow based on conservation of momentum for
a material fluid particle. This conservation law holds at every point
in space and along with appropriate boundary information allows (in
principle at least) for the calculation of the velocity field at every point
in space. We have calculated such velocity fields, at least for some
simplified geometries. In many cases solving the velocity field is too
complicated of a problem, or overkill. Maybe we only need to know the
net force on an object and not the full velocity field. In many cases it is
useful to turn to an integral form of the conservation laws for a control
volume fixed in space. This control volume analysis is a macroscopic
balance, the analysis will allow for us to solve for things such as the
total force acting on an object in a flow due to the imbalance in the total
fluid momentum going in and out of the volume. The control volume
analysis does not allow us to calculate the details of the velocity field.
In order to get results out of a control volume analysis you usually
need to be given some additional information (perhaps from a mea-
surement) or you need to do some physical reasoning and make a some
simplifying approximations. This fact is an important point to empha-
size. You will often find that in order to make progress with a control
volume problem you will have to assume something that may not seem
obvious or you have difficulty justifying. For example, it will be com-
mon to assume that inlet flow fields are uniform in space with a single
velocity. It will also be common to neglect viscous forces. In some cases
these approximations will be great and in others, not so. What assump-
tions are reasonable come with experience. Don’t be fooled by textbook
200 Control volume analysis
problems (including the ones I’ll demonstrate) that tend to gloss over
the assumptions that are made as though you just believe they are
true. Like dimensional analysis, the procedure of control volume analy-
sis is straightforward, it is the proper use of it that takes some physical
insight that only comes with experience. The control volume is an excel-
lent tool, but keep in mind that there are also problems where doesn’t
really help us much.
Control volume analysis is an excellent tool for getting order of mag-
nitude estimates or simple scaling laws. Often in engineering situations
we care about the approximate magnitude of the total force or the pres-
sure, for example, and a simple control volume analysis can be more
useful that a complex simulation. Often we can get a useful and rea-
sonable answer with just a little effort. Control volume analysis is also
a quick way to check other more detailed calculations or simulations.
If you do a full analysis of the Navier-Stokes equations and solve for
the full velocity field, but that field does not satisfy the overall control
volume balance then you have done something wrong.
Many textbooks on fluid dynamics put the control volume analy-
sis very early in the development. A good reason for doing so is that
the analysis is much easier than analysis with the Navier-Stokes equa-
tions. However, when confronted with a real world problem that we
don’t find in a textbook executing a useful control volume analysis re-
quires some physical intuition. Similar to dimensional analysis, control
volume analysis is usually most powerful when coupled with good phys-
ical understanding and intuition about the important effects. Similar to
dimensional analysis, it can be straightforward to learn the technique
of control volume analysis, but using it wisely comes with experience.
13.1 Control volume formulation

Our general form of the conservation of momentum, integrated over an
arbitrary fixed volume is,
Z
Dv
ρ = ρg + ∇ · T dV
Dt
13.1 Control volume formulation 201
We can now move the integral on the left hand side to a different form
using the Reynolds Transport Theorem,
Z Z Z
Dv ∂(ρv)
ρ dV = dV + (ρv)v · ndS.
Dt ∂t
Using the divergence theorem we can also convert the stress tensor
term on the right side of the original equation back to integration over
a surface,
Z Z Z Z
∂(ρv)
dV + (ρv)v · ndS = ρgdV + n · TdS.
∂t
Substituting our constitutive law for a Newtonian fluid in incompress-
ible flow (T = −P I + 2µS), we have,
Z Z Z Z Z
∂(ρv)
dV + (ρv)v · ndS = ρgdV − P ndS + 2µ n · SdS.
∂t
Since we are now interested in a control volume fixed in space, the
derivative on the left hand side can be moved outside the integral, so
Z Z Z Z Z
d
ρvdV + (ρv)v·ndS = ρgdV − P ndS+2µ n·SdS. (13.1)
dt
The above expression holds for the fluid only. A more useful form in-
cludes an external force acting on an object in the flow. In this case,
you can draw a control volume around whatever you like, as long as
you remember to include the force which is holding the object in place.
Z Z Z Z Z
d
ρvdV + (ρv)v·ndS = ρgdV − P ndS +2µ n·SdS +Fext .
dt
(13.2)
Following the same ideas, the integral form of conservation of mass for
a fixed control volume is
Z Z
d
ρdV + ρv · ndS = 0. (13.3)
dt
The integral form of the momentum and mass conservation equations
for a fixed region can be very useful in a number of cases. The primary
use is in making estimations/calculations of net forces acting on an
object in a flow.
You should first note that conservation of momentum is a vector
equation - 3 components for x, y and z. On the left side of Equation
13.1 we have the rate of change of momentum inside the control volume
(first term) and the net flux of momentum coming into/out of the
control volume (2nd term). In the second term on the left side, note
that the momentum ρv is a vector quantity, while v · n is the rate
that the momentum crosses the surface. On the right side, the first
term is the net body force which acts on the whole volume. The second
term is the force due to pressure and the third term is the force due to
viscosity - both of these exert forces only on the surface of our volume.
What is nice about this expression is there are many cases where we
are only interested in the net force on an object, and in those cases we
can often use the integral form of the equation without worrying about
the details of the flow.
R
Getting the sign right of the term (ρv)v · ndS can be a little tricky.
You have be be careful that the velocity vector has a sign, as well as
the term v · n. Since we take normal vectors to point out of the volume,
then v · n is positive for outflow and negative for inflow.
13.2 Examples
The only way we can make sense of the complex looking equation is to
work through a few examples. The basic approach I usually take with
these problems is as follows. I start by drawing a sketch and picking
a control volume that seems convenient. A convenient control volume
often (but not always) has surfaces where the velocity inlets and outlets
are normal to the surface. I like to write out the entire control volume
equation and then systematically cross out terms that I have good
reason to ignore (say unsteady effects in a steady flow). I then would see
where that takes me. If I can get an interesting result without making
many assumptions, then great. If I don’t get much then I might start
making assumptions that I am less confident are true (say ignoring
viscosity in a case where it is not clear I can). I like to try and keep
track of my assumptions so when I get to the end of a problem, it is
easy for me to review what I had to do to get there. One of the most
common assumption you will tend to make in these problems is that
the velocity field is uniform over some region, an assumption we often
make because we have little choice.
13.2 Examples 203
A2 n
Fx
U A1 θ
n
Fy
Figure 13.1 Control volume for finding the force a water jet exerts
on an inclined block. The control volume is shown as the dashed
line.
13.2.1 Example 1: Inclined block

Let’s start with an hose spraying a block with a steady flow rate and the
block is inclined with an angle θ and we are interested in calculating
the total force exerted on the block. The mean velocity leaving the
hose is known to be U and the area of the jet leaving the hose is A. We
start by picking a control volume and experience will show you what
volumes work best. In general, it works well to draw the volume such
that it cuts through all fluid inlets and exits perpendicularly as shown
in Figure 13.1. After drawing the volume it is useful to draw the normal
vectors (pointing outward) where the fluid crosses the control volume
boundary.
Now in order to make some progress we need to assume a few things.
Let’s start by assuming that since the flow is exiting the hose at a
constant rate that we can neglect time dependent terms. Also, just
for simplicity, lets neglect gravity which can be done by taking our
apparatus to space, or by simply turning the device such that gravity
points into the page in Figure 13.1. Let’s start with conservation of
mass for
Z Z
HdH

HρdV + ρv · ndS = 0.
dt H
H
We have two locations where we need to evaluate the surface integral,
the inlet and outlet. We see that there is no way we can technically
evaluate the surface integral without knowing the velocity field of the
inlet and outlet jet. Let’s assume for now that the jet enters and exits
with a uniform velocity field, U1 and U2 respectively. On the inlet the
normal vector and the velocity are not aligned so the dot product is
negative and on the exit, the velocity and the normal vector are aligned
so the dot product is positive. Since the velocity at the inlet and exit
is assumed constant, then the surface integral is easy to do,
Z
ρv · ndS = −ρU1 A1 + ρU2 A2 = 0.
This statement simply says that the total mass flow rate in equals
the total mass flow out. Note that when we write the surface integral
above, it means that we have to evaluate ρv · n at every point along the
control volume surface and integrate it up. In this example it is only
the two little patches where fluid comes in and goes out that has any
fluid velocity.
Now let’s consider the momentum equation. Let’s start by writing
out the full integral equation and cross out the unsteady terms and
gravity which we previously stated we would neglect,
Z Z Z Z Z
Hd H Z
− P ndS +2µ n·SdS +Fext .
H
ρvdV + (ρv)v·ndS = ZρgdV
dt HH
H
Z
Z
Notice that the surface integral terms for pressure and viscosity are
evaluated only on the surface of the control volume. So while viscous
forces and pressure may do something inside the control volume on the
surface we have drawn it would be reasonable to assume that they have
no contribution. Since the control volume is open to atmosphere every-
where, the pressure around the surface of the control volume would be
atmospheric. If we integrate a constant pressure around a closed loop,
then the integral of P n around the loop results in zero net force. While
viscosity certainly acts along the surface of the block, along the sur-
face we have drawn the control volume boundary viscosity would not
seem to play a role since we are assuming uniform flow outward. Recall
that viscous stresses only occur when we have velocity gradients, so
a uniform free jet with uniform flow has no viscous stresses. You are
starting to see that in order for us to make progress we have to start
to use some physical insight. By making these physical arguments to
neglect the pressure and viscous terms on the control volume boundary,
13.2 Examples 205
we are left with Z

(ρv)v · ndS = Fext .
Since the momentum equation is a vector equation we have to con-

sider each component, being careful with the signs. Let’s start with the
x-component of momentum. On the inlet at surface 1 the x component,
Z Z
(ρv)v · ndS = (ρU1 )(−U1 )dS = −ρU12 A1 .
1 1
Note that the dot product of v · n = −U1 since the velocity vector and
normal vector are not aligned. Since the velocity at this surface is all
in the x-direction, then v = U1 . On the outlet at surface 2,
Z Z
(ρv)v · ndS = (ρU2 cos(θ))(U2 )dS = ρU22 A2 cos(θ).
2 1
Note that the dot product of v · n = U2 since the velocity vector and
normal vector are aligned. Since the velocity at this surface is inclined
at angle θ the x-component of the velocity is U2 cos(θ). Putting this
together gives us the external force as,
ρ U22 A2 cos(θ) − U12 A1 = Fx .

Substituting in conservation of mass (U1 A1 = U2 A2 ) yields,

−ρU1 A1 (U1 − U2 cos(θ)) = Fx .
So our final result still doesn’t tell us everything we want since U2 is
not yet determined. We presumably know U1 and A1 as that is given
by the size and flow rate of the hose. Without knowing U2 or A2 we
have no way to get the force. If we measured these quantities we could
proceed. We could also make an assumption that A1 ≈ A2 . While this
seems like a reasonable thing to assume, it need not be true. If the area
is the same then mass conservation says U1 = U2 and the x component
of the force would be,
−ρU12 A1 (1 − cos(θ)) = Fx .
In this case, the force is related to the deflection of x momentum out
of the control volume. The rate that momentum is carried into the
control volume is ρU12 A1 . The term (1 − cos(θ)) represents the fraction
of x momentum that is deflected. If the angle were zero, a flat plate,
then the force would be zero.
If we applied Bernoulli’s equation along a streamline between points

1 and 2, in the absence of gravity we would have,
1 1
P1 + ρU12 = P2 + ρU22 .
2 2
Since the pressure at points 1 and 2 would both be atmospheric, then
Bernoulli’s equation would tell us that U1 = U2 . Bernoulli’s equation
is consistent with the assumption that the area of the jet is the same
at the inlet and exit. Recall that Bernoulli’s equation assumes inviscid
flow, so our analysis with the assumption of equal area is consistent
with the limit where viscosity is small compared to inertia - the high
Reynolds’s number limit. Our control volume result is consistent with
the force on the block just due to changing the direction of inertia of
the incoming flow.
Now let’s consider y momentum. On the inlet at surface 1 the y
component,
Z Z
(ρv)v · ndS = (0)(−U1 )dS = 0.
1 1
As with the x component, the dot product of v · n = −U1 since the

velocity vector and normal vector are not aligned. However, the velocity
at this surface is all in the x-direction, then the y component of v is
zero. On the outlet at surface 2,
Z Z
(ρv)v · ndS = (ρU2 sin(θ))(U2 )dS = ρU22 A2 sin(θ).
2 1
Note that the dot product of v · n = U2 since the velocity vector and
normal vector are aligned. The dot product term is the same in the x
and y momentum equations. Since the velocity at this surface is inclined
the y-component is v = U2 sin(θ). The total y component of the force
is,
ρU22 A2 sin(θ) = Fy .
When we make our assumption that the area of the jet does not change,
ρU12 A1 sin(θ) = Fy .
Thus the final result for the case where the area of the jet does not
change is

cos(θ) − 1
Fext = ρU12 A1
sin(θ)
13.2 Examples 207
While this result assumes the constant velocity profile and the fact that
the jet area remains the same from inlet to exit, it seems reasonable
to think that scaling, order of magnitude and trends might be quite
reasonable.
13.2.2 Example 2: Flat plate

Let’s consider the same example but take the case where θ = 0, a flat
plate. We can use our results from the previous section for the force
which can only have an x-component,
Fx = −ρU1 A1 (U1 − U2 )
If we assume that the area is something we can measure and put our
result in terms of area, U2 = U1 A1 /A2 ,

2 A1
Fx = −ρU1 A1 1 − .
A2
Given that it only make sense given the direction of the flow for the force
to be in the negative x direction, this formula tells us that A1 < A2 ;
the jet grows toward the exit. If the area is the same (as we assumed
in the previous section) then the force would be zero.
Now let’s rework the same problem with a different control volume
that only includes the fluid. The same procedure would ensue as in the
previous problem only we will need to retain the viscous term and since
the plate is not in our control volume, there is no external force.
Z Z Z Z Z
Hd H Z
−HH
H
ρvdV + (ρv)v ·ndS = ZρgdV P ndS
+2µ n·SdS + HFext
H.

dt
H
HH
Z
Z
H
H
Nothing would change in the evaluation of the left side of the equation
with this new analysis, thus,
Z
2 A1
−ρU1 A1 1 − = 2µ n · SdS.
A2
Note that the result from this other control volume says the total force
on the plate is equal to the total viscous stresses on the plate. Since we
do not deflect any momentum, the only x component of force is due
to viscous drag. Any change in area of the jet would be related to the
viscosity. Viscosity removes momentum from the fluid slowing it down
and thus the area must grow accordingly to conserve mass. If this were
v(x)
U0
Fx H u(y)
U0 v(x) U0
L
Figure 13.2 Control volume for finding the drag on an object from
an experimentally measured wake.
a very viscous fluid coming from our hose, we would expect there to
be a more dramatic change in the increased area of the jet. Using the
control volume approach we can only compute the force on the plate if
we had a measurement of the exit area.
If the fluid had very low viscosity (high Reynolds number), we could
apply Benoulli’s equation. If we applied Bernoulli’s equation from the
inlet to the exit, we would see that just like in the last example the
velocity should not change. Since Bernoulli’s equation is only valid for
inviscid flow, our conclusion is consistent with the fact that at high
Reynolds number, we expect the inlet and outlet areas of the jet to be
approximately equal.
13.2.3 Example 3: Drag on an object

Let’s see how we can compute the total drag on an object from a mea-
surement of the surrounding fluid velocity field. The geometry and con-
trol volume are shown in Figure 13.2. We will consider four boundaries
the left, right, top and bottom. For simplicity let’s assume a symmet-
ric object so the top and bottom should act the same. Let’s consider
steady flow. The flow coming in on the left is uniform with a known
velocity U0 . The velocity on the right is measured as a function of y,
u(y).
13.2 Examples 209
Conservation of mass for steady flow in our control volume is,

Z Z
dH
H
HρdV + ρv · ndS = 0.

dt H
H
Let’s expand the surface integral into the four terms,
Z Z H Z H Z L
ρv · ndS = − ρU0 dy + ρu(y)dy + 2 ρv(x)dx = 0.
0 0 0
The order of the terms are left, right, and top/bottom (hence the factor
of 2). Note that on the inlet (left boundary) the velocity and normal
vector point in opposite directions hence the negative sign on that term.
Conservation of mass simply states that since there is a velocity deficit
in the wake of the object, there must be some flow out the top and
bottom surface.
Conservation of momentum is,
Z Z Z Z Z
Hd H Z
−HH XX
X · SdS +Fext .
ρvdV
H + (ρv)v·ndS = Z HH nX
ZρgdV P ndS
+ 2µ
dt HH
H Z
XX
We have crossed out terms since 1) the flow is steady, 2) we will ig-
nore the effect of gravity (or assume perpendicular to the page), 3) the
object is in an external flow field and thus the pressure is atmospheric
everywhere, 4) that along the boundary of control volume which is far
from the object there are no significant viscous stresses. Under our as-
sumptions, the net force on the object is given by the imbalance in the
momentum coming in and out of the control volume.
Z
(ρv)v · ndS = Fext .
Now, lets evaluate our terms for each surface for the x component of
the momentum,
Z H Z H Z L
2 2
− ρU0 dy + ρu(y) dy + 2 ρU0 v(x)dx = Fx .
0 0 0
The first term is the inlet. Since the flow and the normal vector are in
opposite directions, v · n = −U0 and the x component of the velocity
is U0 . The exit integral is computed similarly, only the sign is positive
since the velocity and flow direction are aligned. The top and bottom
terms should be treated with care. For the top and bottom, the outflow
is v · n = v(x). The x component of velocity is U0 since the surface is
far from the object.
A2
U1 A1
U2
Figure 13.3 Control volume for finding the net force required to
hold the nozzle on a pipe.
Substituting our conservation mass statement,

Z L Z H Z H
2 ρv(x)dx = ρU0 dy − ρu(y)dy,
0 0 0
into the x momentum equation we get,
Z H Z H Z H !
Z H
2 2
− ρU0 dy+ ρu(y) dy+U0 ρU0 dy − ρu(y)dy = Fext .
0 0 0 0
Canceling the offsetting terms we obtain,

Z H
−ρ u(y) (U0 − u(y)) dy = Fx .
0
If we have a measurement of the velocity field in the wake of an object
we can compute (or at least estimate) the total drag on the object.
13.2.4 Nozzle
Consider the tapered nozzle as shown in Figure 13.3. Lets consider a
steady flow rate and we want to estimate the force required to hold the
nozzle in place at the flange. We draw the control volume as shown.
We will assume steady flow, ignore gravity, assume the inlet and outlet
flow fields are uniform, and assume a high Reynolds number flow.
At this point, we can probably do the conservation of mass pretty
easily,
Z Z
HdH
ρdV + ρv · ndS = −ρU1 A1 + ρU2 A2 = 0.
H
dt H
H
13.2 Examples 211
Conservation of momentum is,

Z Z Z Z Z
Hd H Z
− P ndS +X
H
ρvdV + (ρv)v·ndS = ZρgdV 2µXX · SdS
n
X XX+Fext .
dt

H
H
H
Z
Z

Here we have neglected the viscous terms, not because we believe that
viscosity doesn’t matter inside the nozzle but that viscosity probably
doesn’t add much to on the surfaces of the control volume that we have
drawn. Since the flow is contained inside the nozzle, we must be careful
to use different pressure. Now, evaluating the x component of this force
and being careful with the signs,
−ρU12 A1 + ρU22 A2 = −(−P1 A1 + P2 A2 + P∞ (A1 − A2 )) + Fx .
In this example the momentum in and out of the control volume is sim-
ilar to previous examples. What is new is how we handle the pressure.
There are a few signs to keep track of. There is a negative in front of
the pressure term. We also must account for the direction of the normal
vector. Also, all of pressure at location 1 is acting on the area of the
inlet, while the pressure at point 2 is acting only smaller exit area. The
atmosphere is pushing on the right side of the control volume on the
area A1 − A2 . Since the pressure just outside the nozzle is in the open
air, it is reasonable to set the pressure at the exit to be P2 ≈ P∞ . In
this case our equation becomes,
−ρU12 A1 + ρU22 A2 = (P1 − P∞ )A1 + Fx .
Now we see in this form that the force to hold the nozzle is the difference
in momentum coming in and out of the volume and the pressure at the
nozzles inlet. We can use conservation of mass (U1 A1 = U2 A2 )to write
as

A1
ρU12 A1 − 1 − (P1 − P∞ )A1 = Fx .
A2
Note the sign of the terms. The net change in momentum alone would
lead to a force in the positive direction - meaning the nozzle is com-
pressed on the tube. Since the flow shoots fast out of the nozzle there
is a thrust like a rocket. The pressure term leads a negative force and
wants to pop the nozzle off. We can’t solve for the total force without
more information.
One assumption we could make is assume that the flow is high
Reynolds number and apply Bernoulli’s equation from the inlet to the
outlet. Bernoulli would yield,
1 1
P1 + ρU12 = P∞ + ρU22
2 2
or
1 2 1 2
P1 − P∞ = ρU − ρU1 .
2 2 2
Using conservation of mass gives,
A21

1
P1 − P∞ = ρU12 −1
2 A22
Using the result from Bernoulli and going back to our momentum
balance would give the nozzle force as
2
1 2 A1
Fx = − ρU1 A1 −1
2 A2
When A1 > A2 then the force acting on the nozzle is negative - meaning
without a bolt holding the nozzle on, there is a tendency of the nozzle
to pop off. The final result is only valid for high Reynolds number where
Bernoulli’s equation is reasonable.
13.3 Some difficulties

Sometimes, the control volume formulation does not lead us to a simple
conclusion. Let’s discuss a problem where the control volume approach
does not give a very clear result. The problem seems straightforward
and is known as the inverse sprinkler or Feynman’s sprinkler. This
problem is interesting because it seems that you should be able to get
the answer by applying a control volume.It is called Feynman’s sprinkler
after the great 20th century physicist Richard Feynman who wrote of
this problem in one of his books but never really provided what he
thought his answer was.
A “normal” simple sprinkler is shown in Figure 13.4. It is free to
pivot about the center point. If the entire sprinkler is submerged in
a pool such that the sprinkler emits water into a water environment,
most people would guess that the sprinkler spins clockwise. They would
be right. Now if we consider the sprinkler to work in inverse mode, i.e.
13.3 Some difficulties 213
Normal sprinkler Inverse sprinkler Control volume

analysis
u u
Which
way? n
F
u u u
Figure 13.4 Schematic for the Feynman sprinkler problem. We show

the normal sprinkler, which seems obvious to us should rotate clock-
wise. The direction of the inverse sprinkler is less clear. We also show
the control volume used for the analysis.
water intakes into the sprinkler, the question is which way does the
sprinkler spin? It seems the control volume analysis may be useful for
this problem because we just want to know which way it spins, not how
fast.
To analyze the problem, let’s look at one half of the sprinkler, imagine
that we hold the sprinkler fixed and calculate the force, F , needed to
hold the normal sprinkler in place. Let’s assume the water comes out
as a jet of uniform velocity u and area A, the flow is steady, the density
is constant, and viscosity is not important. Under these assumptions,
we have,
Z Z
(ρv)v · ndS = P ndS + F
The only tricky part of control volume analysis is getting the sign cor-
rect and since we are trying to get the direction correct, we need to treat
our signs with care. There are two signs to keep straight on the left side
of the equation, v · n and v. At the exit as drawn,R both v · n and v are
positive and thus we can estimate the term as (ρv)v · ndS ≈ ρAu2 .
Normally, we would assume that the pressure at the exit is the same
as the pressure in the environment. The force is positive and estimated
to have a magnitude of F = ρAu2 and be positive as drawn in the
schematic of figure 13.4. If we remove the force, the sprinkler spins.
Now consider the inverse sprinkler. The same integral expression
must hold true. Momentum enters the control volume at the same rate
as the sprinkler in forward motion. Since v · n and v are both nega-
tive, their product is positive; the momentum across the surface is the
same in magnitude and sign as the sprinkler in normal operation. If this
were all, the sprinkler should spin the same direction whether the fluid
comes in or out of the sprinkler. If one does the experiment (you can
try yourself with a bendable drinking straw) you find that the sprinkler
does not spin. If you did the experiment carefully you would actually
find the sprinkler spins very slowly in the reverse direction.
The reason the control volume approach does not work is that we
have not gotten the exit pressure right. In the sprinkler operating in
“normal” mode the pressure at the exit must be at least a little higher
at the exit than the environment since pressure gradients would be re-
sponsible for “pushing” the flow. When the sprinkler is operating in
reverse mode, the pressure at the exit must be lower than the environ-
ment for flow to go in that direction. It turns out that if viscosity is
not important then the low pressure at the sprinkler exit (which wants
to make the sprinkler spin counterclockwise) exactly cancels the mo-
mentum coming in (which wants to make the sprinkler spin clockwise).
I cannot derive the experimental result using control volumes. Control
volume analysis gives you no way to calculate the pressure in this prob-
lem. While a number of people have written about this problem I have
yet to find a solution that I find truly satisfying, nor have I been able
to work one out myself.
Bibliography
Anderson, J.D. 2005. Lundwig Prandtl’s boundary layer. Physics Today, 58,
42–48.
Bird, R.B., Stewart, W.E., and Lightfoot, E.N. 1960. Transport Phenomena.
Wiley.
Bridgman, P.W. 1922. Dimensional Analysis. Yale University Press.
Kundu, P.K., and Cohen, I.M. 2004. Fluid Mechanics. Elsevier Academic
Press.
Panton, R.L. 1996. Incompressible Flow. Wiley.
Rayleigh. 1915. The principle of similitude. Nature, 95, 66–68.
Reynolds, O. 1883. An experimental investigation of the circumstances which
determine whether the motion of water shall be direct or sinuous and
the law of resistance in parallel channels. Phil. Trans. R. Soc., 174,
935–982.
Saffman, P.G. 1992. Vortex Dynamics. Cambridge University Press.
Schey, H.M. 1997. div grad curl and all that. W.W. Norton and Co.
VanDyke, M.V. 1982. An Album of Fluid Motion. Parabolic Press.

Fluids

Uploaded by

Copyright:

Available Formats

Fluids

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

Fluids

Uploaded by

Copyright:

Available Formats

Fluid dynamics

An introduction to the fundamentals

9.8 Poiseuille flow in a pipe 143

Climate The computer models used by climate scientists begin with

1.1 Scope of the text

• Fluid Mechanics by Kundu and Cohen (2004).

1.2 What is a fluid?

1.3 Simple fluid properties

1.4 Other properties

more elastic. Viscoelastic fluids have many interesting properties but

1.5 What is a continuum?

Figure 1.1 Schematic of the concept of a continuum. The local mass

1.6 Mathematical prerequisites

The result of any physical equation must remain unchanged regard-

2.2 Buckingham Pi Theorem

Figure 2.1 Schematic of a simple pendulum.

2.2.2 Pythagorean theorem

Figure 2.2 Schematic of a right triangle. It is easy to prove a2 +b2 =

The Rayleigh paper then gives, in essentially 1 sentence each, 13

arguing that temperature is not an independent dimension since tem-

For the students, except in the simplest of problems, dimensional

Thus, when starting to learn dimensional analysis it is important to

2.2.4 Drag on a sphere

Immediately we see some arbitrary choices to make. If we try to

Figure 2.3 Drag coefficient as a function of Reynolds number for

reasonable empirical functional relationship of one variable. A single

2.2.5 Pipe flow

plified as the number of variables is reduced to 5 and the Pi Theorem

This result is powerful as this is a single master curve. This master

Figure 2.5 Experimental images on pipe flow at different Reynolds

of the curve was first described by Osbourne Reynolds in one of the

2.2.6 Drag on a boat

ρ [T2 ] ` [L] U [ TL ] ρ [ LT ] g [ TL2 ] Remove M

different in the two dimensionless groups. If I shrink the boat, to keep

2.3 Making equations dimensionless

which is a dimensionless equation with one parameter. However, we

because the solution for x will always be bounded in magnitude by 1.

The result is consistent with what we obtained from making the

2.3.3 Derivatives have units too

variable based on some ”natural” time scale such as t̃ = t/t0 , then we

3.1 Drag revisited

ρ [T2 ] D [L] U [ TL ] ρ [ LT ] Remove M

Defining the drag coefficient as Cd = 1 ρUF2 πR2 and the Reynolds

Figure 3.1 Drag coefficient as a function of Reynolds number for

dimensional analysis. Low values of Reynolds numbers occur for small

Our result now says there is one dimensionless parameter, therefore

3.2 Drag coefficients

Figure 3.2 Flow around a cylinder at different Reynolds numbers

analysis gave a force on a sphere to have a form

In both cases, the constant C (or Drag Coefficient) is unknown un-

• If a single number for drag coefficients is reported, it is most common

3.3 Pipe flow

Figure 3.3 Experimental data for friction factor versus Reynolds

by fitting experimental measurements. There is no analytical solution

The diagram in it’s classical presentation is reproduced in Figure 3.4

3.3.1 Calculations with the Moody diagram

If at anytime in either procedure (or based on your intuition or

4.1 Scalar and vector functions

field has no sense of direction. Two examples of scalar functions are