INTEGRAL BOUNDARY LAYER METHODS IN PYTHON

A Thesis

presented to

the Faculty of California Polytechnic State University,

San Luis Obispo

In Partial Fulfillment

of the Requirements for the Degree

Master of Science in Aerospace Engineering

by

Malachi Edland

August 2021
 © 2021
Malachi Edland
ALL RIGHTS RESERVED

ii
 COMMITTEE MEMBERSHIP

TITLE: Integral Boundary Layer Methods in

Python

AUTHOR: Malachi Edland

DATE SUBMITTED: August 2021

COMMITTEE CHAIR: David D. Marshall, Ph.D.

Professor of Aerospace Engineering

COMMITTEE MEMBER: Aaron Drake, Ph.D.

Professor of Aerospace Engineering

COMMITTEE MEMBER: Paulo Iscold, Ph.D.

Professor of Aerospace Engineering

COMMITTEE MEMBER: Colleen Kirk, Ph.D.

Professor of Applied Mathematics

iii
 ABSTRACT

Integral Boundary Layer Methods in Python

Malachi Edland

This thesis presents a modern approach to two Integral Boundary Layer methods

implemented in the Python programming language. This work is based on two 2D

boundary layer methods: Thwaites’ method for laminar boundary layer flows and

Head’s method for turbulent boundary layer flows. Several novel enhancements im-

prove the quality and usability of the results. These improvements include: a common

ordinary differential equation (ODE) integration framework that generalizes compu-

tational implementations of Integral Boundary Layer methods; the use of a dense

output Runge-Kutta ODE solver that allows for querying of simulation results at

any point with accuracy to the same order as that of the solver; and an edge veloc-

ity treatment method using cubic spline interpolation that improves the simulation

performance using only points from an inviscid edge velocity distribution. Both the

laminar and turbulent methods are shown to benefit from smoothing of the edge ve-

locity distribution. The choice of ODE solver alleviates the need to artificially limit

step sizes. Comparisons against analytic solutions, experimental data and XFOIL

results provide a wide varity of verification and validation cases with which to com-

pare. The implementation of Thwaites’ method in this thesis avoids simplifications

made in other implementations of this method, which results in more robust results.

The implementation of Head’s method produces high-quality results typically found

in other implementations while utilizing the common ODE integration framework.

Utilizing the common ODE framework results in significantly less code needed to

implement Thwaites’ and Head’s methods. In addition, the boundary layer solvers

produce results in seconds for all results presented here. Boundary layer transition

iv
and separation criteria are implemented as a proof of concept, but require future

work.

v
 ACKNOWLEDGMENTS

Thanks to:

• My parents, whose financial support, patience, and love have carried me through

23 years of life, 4 educational institutions, 2 degrees, and endless experiments

and projects in the backyard

• My amazing girlfriend, for reassuring me when the hours are long and the work

gets tough that I’m capable enough to make it through

• Kari Mobley and ACI Jet for giving me two exhilarating years in the hangar

and the Python experience I needed to start this project

• Nick Brake and Cory Seubert for taking me on and providing invaluable work

experience at Empirical Systems Aerospace, concurrent and complimentary to

this project

• Finally, thank you to my wonderful advisor Dr. David Marshall for taking me

under your wing and encouraging me to find new things! I think back often

to when I walked into your office and first asked you about Python. Your first

words were “sure Malachi, take a seat.” The way you said it made me feel so

included, and I knew that I’d come to the right person to learn from. More

than a year later, I’ve learned so much, and am so glad that I took that seat.

vi
 TABLE OF CONTENTS

Page

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

NOMENCLATURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi

CHAPTER

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Boundary Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.3 The Blasius Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.4 The Falkner-Skan Solution . . . . . . . . . . . . . . . . . . . . . . . . 5

1.5 Integral Boundary Layer Methods . . . . . . . . . . . . . . . . . . . . 7

1.5.1 Thwaites’ Method for Laminar Boundary Layers . . . . . . . . 8

1.5.2 Head’s Method for Turbulent Boundary Layers . . . . . . . . 13

1.6 Transition Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

1.7 Runge-Kutta and Dense Output . . . . . . . . . . . . . . . . . . . . . 15

2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.1 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.1.1 Parent Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.1.2 User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.2 Thwaites’ Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.3 Michel Transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.4 Head’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.5 Laminar and Turbulent Separation . . . . . . . . . . . . . . . . . . . 25

3 Test Cases and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

vii
 3.1 Testing Thwaites’ Method Against Falkner Skan . . . . . . . . . . . . 27

3.1.1 Falkner-Skan Flow, m=0 . . . . . . . . . . . . . . . . . . . . . 28

3.1.2 Falkner-Skan Flow, m=1 . . . . . . . . . . . . . . . . . . . . . 31

3.1.3 Falkner Skan Flow, m=1/3 . . . . . . . . . . . . . . . . . . . . 31

3.2 Head’s Method on Flat Plate with Mild Pressure Gradient . . . . . . 37

3.3 Testing Against XFOIL . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.3.1 Laminar Results . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.3.2 Turbulent Results . . . . . . . . . . . . . . . . . . . . . . . . . 47

3.3.3 Combined Results . . . . . . . . . . . . . . . . . . . . . . . . . 49

4 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . 55

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

APPENDIX

A Falkner-Skan Flow Near Zero . . . . . . . . . . . . . . . . . . . . . . 60

viii
 LIST OF FIGURES

Figure Page

1.1 Falkner-Skan Wedge Flow . . . . . . . . . . . . . . . . . . . . . . . 6

1.2 F (λ) Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.3 F (λ) Approximation Error . . . . . . . . . . . . . . . . . . . . . . . 12

2.1 PyBL Class Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2 Separation and Transition Criteria Class Diagram . . . . . . . . . . 19

3.1 Falkner-Skan Flow, m=0 - Results . . . . . . . . . . . . . . . . . . 29

3.2 Falkner-Skan Flow, m=0 - Error . . . . . . . . . . . . . . . . . . . . 30

3.3 Falkner-Skan Flow, m=0 - Linear Error . . . . . . . . . . . . . . . . 30

3.4 Falkner-Skan Flow, m=1 - Results . . . . . . . . . . . . . . . . . . 32

3.5 Falkner-Skan Flow, m=1 - Error . . . . . . . . . . . . . . . . . . . . 33

3.6 Falkner-Skan Flow, m=1 - Linear Error . . . . . . . . . . . . . . . . 33

3.7 Falkner-Skan Flow, m=1/3 - Results . . . . . . . . . . . . . . . . . . 35

3.8 Falkner Skan Flow, m=1/3 - Error . . . . . . . . . . . . . . . . . . . 36

3.9 Falkner Skan Flow, m=1/3 - Linear Error . . . . . . . . . . . . . . . 36

3.10 Flat Plate with Mild Adverse Pressure Gradient - Velocity Inputs . 37

3.11 Flat Plate with Mild Adverse Pressure Gradient - Velocity Spline
Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.12 Flat Plate with Mild Pressure Gradient, Head’s Method . . . . . . 39

3.13 Flat Plate with Mild Pressure Gradient, Head’s Method - Error . . 40

3.14 Flat Plate with Mild Pressure Gradient, Head’s Method - Error
(Smoothed Velocity) . . . . . . . . . . . . . . . . . . . . . . . . . . 40

ix
3.15 Flat Plate with Mild Pressure Gradient, Head’s Method - Error
(Smoothed Velocity Derivative) . . . . . . . . . . . . . . . . . . . . 41

3.16 Flat Plate with Mild Pressure Gradient, Head’s method - Transpi-
ration Velocity Results . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.17 Fully Laminar XFOIL vs. Thwaites Results . . . . . . . . . . . . . 45

3.18 Fully Laminar XFOIL vs. Thwaites Results, Difference . . . . . . . 46

3.19 Laminar XFOIL Velocity Fluctuation . . . . . . . . . . . . . . . . . 46

3.20 Fully Turbulent XFOIL vs. Head Results . . . . . . . . . . . . . . . 48

3.21 Fully Turbulent XFOIL vs. Head Results, Difference . . . . . . . . 49

3.22 Full XFOIL Simulation Comparison . . . . . . . . . . . . . . . . . . 51

3.23 Full XFOIL Simulation Comparison - Difference . . . . . . . . . . . 52

3.24 Forced Transition XFOIL Simulation Comparison . . . . . . . . . . 53

3.25 Forced Transition XFOIL Simulation Comparison - Difference . . . 54

x
 NOMENCLATURE

Latin

Reθ Momentum thickness Reynolds number, page 2

cf Skin friction coefficient, page 3

F1 Head’s method experimental fit to H1 , page 14

H Shape factor, page 3

H1 Alternative shape factor used in Head’s Method, page 13

Htr Shape factor immediately after transition, page 24

m Falkner-Skan velocity distribution exponent, page 6

s Shear correlation, page 8

u Streamwise velocity component at a point in the flow field, page 5

u∞ Freestream velocity, page 4

ue Edge velocity, page 4

un transpiration velocity, page 42

v Transverse velocity component at a point in the flow field, page 5

Greek

β Falkner-Skan wedge angle parameter, page 5

δ∗ Displacement thickness, page 2

xi
δ Boundary layer thickness, page 2

θ Momentum thickness, page 2

ν Kinematic viscosity, page 2

xii
 Chapter 1

INTRODUCTION

1.1 Background

The exponential growth of computing power in recent years has made a powerful im-

pact in the development of aeronautics modeling techniques. Many steps in aircraft

design processes rely on computational tools that vary widely in precision and sim-

ulation times. Modern computational techniques for fluid flows have been driven by

the capabilities of the hardware on which they run, becoming more and more complex

and accurate, as computational costs decrease.

While high-accuracy solvers are the focus of research labs, lower fidelity models are

appreciated in many design shops as a way of obtaining quick, accurate results. In

design environments where simulation time savings, workflow, and modularity are

extremely valuable, Integral Boundary Layer methods offer valuable results and can

quickly return important boundary layer parameters needed for the analysis of a

design.

1.2 Boundary Layers

In aerodynamics, the concept of boundary layers has been very important for un-

derstanding viscous flows, since the velocity of a low-viscosity fluid flowing around

an object changes rapidly near the object’s surface. Prandtl was the first to define a

Grenzschicht, literally boundary layer, as the thin layer of viscous influence near a wall

in 1904 [15]. While the Navier-Stokes equations for fluid flows already existed at this

1
point, their complexities had limited aerodynamicists to inviscid solutions before this

development [17]. Prandtl’s definition would result in a simplification of the Navier-

Stokes equations applicable within the boundary layer. The result allowed the flow

to be treated as inviscid outside the boundary layer and simplified the Navier-Stokes

equations within the boundary layer.

Boundary layers are often characterized by their thicknesses. The boundary layer

thickness, δ, is the distance at which the local velocity is equivalent to the freestream

velocity, u∞ . The displacement thickness, δ ∗ , characterizes the displacement of the

inviscid streamline from the surface and is calculated as the amount of freestream

mass leaving the boundary layer in the direction normal to the surface:

Z ∞  
∗ u(x, y)
δ (x) = 1− dy (1.1)
0 ue x

The momentum thickness, θ, characterizes the amount of momentum lost within the

boundary layer, and is calculated as the amount of equivalent freestream momentum

lost:
Z ∞  
u u
θ(x) = 1− dy (1.2)
0 ue ue

Other parameters are also used to characterize the boundary layer flow.

The momentum thickness Reynolds number is defined as:

ue θ
Reθ = (1.3)
ν

where ν is kinematic viscosity. Reθ is a common term used to characterize the

boundary layer, including in the methods used for predicting transition. The shape

2
factor H is the ratio of displacement thickness over momentum thickness:

δ∗
H= (1.4)
θ

It is typically used as a criteria for separation and for transition, and also appears

in boundary layer approximation methods. The skin friction coefficient cf quantifies

the viscous stresses on the surface. It is defined as the wall shear stress τw nondimen-

sionalized by dynamic pressure q = (1/2) ρu2 :

τw
cf = (1.5)
q

The velocity used to calculate q has not always been defined consistently. Some

references use the freestream velocity (far upstream of the surface) and some use

the velocity at the edge of the boundary layer at the current streamwise location.

In this work, cf is nondimensionalized using the boundary layer edge velocity, ue to

calculated q.

Boundary layers in fluid flows undergo specific changes as they move along the surface.

Boundary layers are described as laminar when flow travels smoothly and steadily [17].

Laminar boundary layers are prevalent in low Reynolds number flows. In contrast,

boundary layers that are turbulent occur at higher Reynolds numbers and are charac-

terized by random fluctuations in the instantaneous flow properties. They are often

described on statistical or time-averaged bases [17].

As a flow moves along a surface, the boundary layer near the surface may change from

having laminar qualities close to the stagnation point, to becoming turbulent further

downstream. The process by which a laminar flow becomes turbulent is known as

transition. Laminar or turbulent flows may also separate from the surface. These

changes occur because of the growth of instabilities along the surface [17].

3
1.3 The Blasius Solution

In 1908, the work of Blasius presented a solution for boundary layer growth on a flat

plate with no streamwise pressure gradient [4]. The inviscid velocity distribution for

this type of flow is represented as:

ue = u∞ (1.6)

Where at any x point along the surface, the velocity at the edge of the boundary

layer, ue , is equal to the freestream velocity, u∞ . This solution utilizes the similar-

ity property of this type of boundary layer. The boundary layer equations can be

transformed into a third order, non-linear ODE that represents a similarity solution

for the laminar laminar flow field anywhere within the boundary layer, except at the

start of the plate, x = 0. The solution at any streamwise location in the boundary

layer can be found from solving for the dimensionless stream function f (η):

f f 00 + 2f 000 = 0 (1.7)

with boundary conditions

f (0) = 0, f 0 (0) = 0; f 0 (∞) = 1 (1.8)

where η is a dimensionless coordinate η ∼ y/δ∗ :

r
u∞
η=y (1.9)
νx

At the surface, η = 0. A value of η = 1 corresponds to the displacement thickness δ ∗ .

Blasius’ work described the velocity distribution along the surface, as well as through

4
the boundary layer, with the following relations:

u = u∞ f 0 (η) (1.10a)

r
1 ν∞
v= (ηf 0 − f ) (1.10b)
2 x

where u and v are streamwise and transverse velocity, respectively. This solution

provides the following expressions for displacement thickness, momentum thickness,

and skin friction coefficient:

r
∗ νx
δ = 1.7208 (1.11a)
u∞
r
νx
θ = 0.664 (1.11b)
u∞
0.664
cf = √ (1.11c)
Rex

1.4 The Falkner-Skan Solution

In 1930, Blasius’ solution was generalized by the work of Falkner and Skan that

applied a boundary layer to a wide variety of flows [2]. Potential flow theory was

used to find the edge velocity expressions and the physical flows that they represent.

One flow that may be represented using the Falkner-Skan equations is a wedge flow.

In a wedge flow, a flat surface is positioned at an angle πβ/2 to the flow, as shown in

fig. 1.1.

The edge velocity, ue , for Falkner-Skan flow is defined by the following expression:

 x m
ue = u∞ (1.12a)
L

5
 y x

u∞

πβ
2

Figure 1.1: Falkner-Skan wedge flow. Note that if β = 0, the wedge is

equivalent to a flat plate.

where
β
m= (1.12b)
2−β

and where x is the distance along the surface from the stagnation point [18]. This

velocity distribution can represent a number of different flows beyond wedge flows, so

it is treated in this project as a general Falkner-Skan flow.

Once again, the boundary layer equations are transformed into a non-linear ODE:

f 00 + f f 00 + β[1 − (f 0 )2 ] = 0 (1.13a)

with the boundary conditions:

f (0) = 0, f 0 (0) = 0; f 0 (∞) = 1 (1.13b)

A numerical solution must be used to obtain an approximation of f . Similarly to

Blasius’ solution, f and its derivatives may then be used to obtain δ ∗ , θ, and cf . The

displacement thickness, momentum thickness, and skin friction coefficient are given

by:
 1/2  1/2 Z ∞
∗ 2 νx
δ = (1 − f 0 ) dη (1.14a)
m+1 u∞ 0
 1/2  1/2 Z ∞
2 νx
θ= f 0 (1 − f 0 )dη (1.14b)
m+1 u∞ 0

6
 q
2 m+12
cf = √ (f 00 |0 ) (1.14c)
Rex

1.5 Integral Boundary Layer Methods

A different approach to analyzing the boundary layer equations came from Theodore

Von Karman in 1921 [19]. He integrated the boundary layer equations through the

boundary layer (essentially averaging over the thickness) [19] to arrive at:

dθ θ due 1
+ (2 + H) = cf (1.15)
dx ue dx 2

This is known as the momentum integral equation. This equation forms the basis

for the Integral Boundary Layer (IBL) methods. Empirical relations are needed to

complete models based on this equation. Equation (1.15) is valid for both laminar

and turbulent boundary layers with the appropriate empirical relations; however, to

model the relationships between H, cf , and θ, different empirical relations are needed

based on the boundary layer characteristics. At a given streamwise location on the

surface, these methods return the shape factor, momentum thickness, skin friction,

and displacement thickness. Two specific Integral Boundary Layer methods were

used in this project: Thwaites’ method for laminar flows, and Head’s method for

turbulent flows. These methods are well studied and are generally simpler than other

IBL methods, as their ODE’s are mostly derived from experimental data or analytic

solutions [8].

7
1.5.1 Thwaites’ Method for Laminar Boundary Layers

A method for estimation of laminar boundary layers was presented by Bryan Thwaites

in 1949 [3]. This method is 2-dimensional, but may be applied to axisymmetric bodies

with the Mangler transformation [8]. In this work, it is used exclusively as a 2D

method. Thwaites defined a parameter λ, where:

θ2 due
λ= (1.16)
ν dx

By arranging the momentum integral equation as follows:

cf dθ θ due
= + (2 + H) (1.17)
2 dx ue dx

and by multiplying both sides by ue θ/ν :

τw θν ue θ dθ θ2 due
= + (2 + H) = s(λ) (1.18)
ue ν dx ν dx

He found that two functions, a wall shear correlation, s, and shape factor, H, were

approximated fairly well as functions only of λ.

Tabulated values correlating s and H to λ were also provided by Thwaites and were

obtained from a number of solutions to the laminar boundary layer equations. Various

fits to this tabulated data have been presented, such as those presented in Cebeci and

Bradshaw [8] and White [17].

Solving eq. (1.16) for θ2 :

λν
θ2 = due/dx
(1.19)

8
and given that:
dθ2
   
dθ 1
θ = (1.20)
dx 2 dx

Equation (1.18) may be rewritten as:

dθ2 θ2
   
ue due
s(λ) = + (2 + H) (1.21a)
2ν dx ν dx

 
d λ
2s(λ) = ue + 2λ(2 + H) (1.21b)
dx due/dx
 
d λ
ue = 2[s(λ) − λ(2 + H)] = F (λ) (1.21c)
dx due/dx

To simplify solving this, Thwaites recognized that if F (λ) was of the form F (λ) =

a − bλ, then the following exact solution for θ2 may be obtained:

x
θ2
Z 
= au−b
e ub−1
e dζ +c (1.22a)
ν 0

Where the integration constant c must be zero so that the momentum thickness at

the start of the surface, x = 0, is not infinite, c/0. Thwaites found the best fit values

to be:

a = 0.45, b = 6.0 (1.22b)

This yields the final expression of:

Z x
2.45ν
θ = 6 u5e dx (1.23)
ue 0

Implementations of Thwaites’ method typically carry out this integral on the velocity

distribution (possible analytically with a known velocity distribution, or numerically

with a simple quadrature technique). With θ calculated at a given point and due/dx

either solved for analytically or approximated in order to determine λ using eq. (1.16),

s and H may be estimated using the aforementioned empirical relationships.

9
While the assumed relationship for F in equation eq. (1.21c) makes this method easier

to solve, it also sacrifices the accuracy of the tabulated values Thwaites provided for

s and H (as well as the fits provided by others) by introducing potential error in θ

from the linearization of F (λ).

A new method, using the empirical fits for s(λ) and H(λ), was derived during this

project to avoid this linearization and the error it involves.

Presented in fig. 1.2 are the values of F (λ) using the linear assumption and two

different fits. The associated errors are shown in fig. 1.3. The linear assumption (the

most common) produces the most error, as shown in fig. 1.3. The lowest error in

F is achieved with the fits to s and H presented by Cebeci and Bradshaw [8] and

eq. (1.21c) to solve for F :


0.22 + 1.57λ − 1.8λ2

 for 0 ≤ λ ≤ 0.1
s(λ) ≈ (1.24a)
0.018λ

0.22 + 1.402λ +

0.107+λ
for −0.1 ≤ λ < 0


2.61 − 3.75λ + 5.24λ2

 for 0 ≤ λ ≤ 0.1
H(λ) ≈ (1.24b)
 0.0731 + 2.088


0.14+λ
for −0.1 ≤ λ < 0

This model, while accurate, applies to a specifically limited range compared to the

White fits [17]:

s(λ) ≈ (λ + 0.09)0.62 (1.25a)

H(λ) ≈ 2 + 4.14z − 83.5z 2 + 854z 3 − 3337z 4 + 4576z 5

(1.25b)
where z = (0.25 − λ)

Figure 1.3 shows that by using a linear approximation for F , significant error is

introduced at certain values of λ. While splicing the different fits may have shown

slight improvement, this project instead implements a cubic spline between Thwaites’

10
 1 Linear
Splined
Cebeci and Bradshaw
White
0.5
Thwaites

−0.5

−1

−0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25

Figure 1.2: Linearized F (λ), (F (λ) = .45 − 6λ), splined F (λ) from Thwaites’
original values, and F (λ) using fits to s(λ) and H(λ) presented by Cebeci
and Bradshaw [8] and White [17].

original values to return s and H as functions of λ. This ensures zero error in F at

Thwaites’ points, and a smooth derivative through the entire distribution.

Using a Runge-Kutta solver, this project’s implementation of Thwaites’ method elim-

inates the need to assume that F is linear. The details of this change are described

more in section 2.2. For comparison with known analytical results, Thwaites’ method

may be applied analytically to Falkner-Skan flow. Given the Falkner-Skan velocity

distribution described by eq. (1.12a), Thwaites’ integral may be applied to calculate

θ2 at a given point:
Z x
2 .45ν  x 5m
θ (x) = u5∞ (1.26)
x 6m L


u6∞ L
0

which may be simplified to

s
.45Lm νx1−m
θ(x) = (1.27)
u∞ (5m + 1)

11
 10−1
F (λ) Relative Error

10−2

10−3
Linear
Cebeci and Bradshaw
White

−0.05 0 0.05 0.1 0.15 0.2 0.25

λ

Figure 1.3: Error introduced using linearized F (λ) (F (λ) = .45 − 6λ), fits
presented by Cebeci and Bradshaw [8], and fits presented by White [17],
against Thwaites’ original tabulated values. Note that that no error for the
spline (presented later) is shown, as its error is 0 at every one of Thwaites’
points.

12
For various values of m, the error of this calculation is within 10% for the momentum

thickness for Falkner-Skan flow, eq. (1.14b) [17].

1.5.2 Head’s Method for Turbulent Boundary Layers

M. R. Head first described his method for solving for a turbulent boundary layer in

1958 [6]. This is a 2-dimensional method, but can be applied to axisymmetric bodies

with the appropriate transformation. His method used a new shape factor, H1 , based

on the entrainment velocity [14]. Similar to the definition of H in eq. (1.4),

δ − δ∗
H1 = (1.28)
θ

Given the momentum integral equation:

dθ θ due 1
+ (2 + H) = cf (1.15 revisited)
dx ue dx 2

a relation for the entrainment velocity and the new shape factor, H1 , such as the

experimental data created by Cebeci and Bradshaw:

1 d
F1 (H1 ) = (ue θH1 ) ≈ 0.0306(H1 − 3)−0.6169 (1.29)
ue dx

3.3 + 0.8234(H − 1.1)−1.287

 for H ≤ 1.6
H1 (H) ≈ (1.30)
3.3 + 1.5501(H − 0.6778)−3.064

 for H > 1.6

and the Ludwieg-Tillman relation for cf , θ, and H:

cf ≈ 0.246 × 10−.678H Reθ−.268 (1.31)

13
a system of partial differential equations with 2 unknowns, θ and H, may be obtained

if ue and due/dx are known as functions of x. This process is described further with

the implementation of Head’s method in section 2.4.

Cebeci and Bradshaw [8] presented a Runge-Kutta implementation for solving Head’s

Method, using θ and H as the simulation variables. Given the relationship presented

by Ludwieg and Tillman [5] in eq. (1.31), skin friction coefficient can be found as a

function of θ and H to eiliminate it from the momentum integral equation, eq. (1.15).

Finally, the derivative of H with respect to x can be determined from experimental

fits, F1 as a function of H1 in eq. (1.29), and H1 as a function of H in eq. (1.30).

Cebeci and Bradshaw used a simple approximation for due/dx using the provided values

of the edge velocity in order to obtain dH/dx from eq. (1.29). Their implementation

used the form of the fits to F1 and H1 to calculate the derivatives needed to implement

the solution to the system of ODE’s.

Note that dH1/dH may be obtained analytically when H1 (H) is estimated as in eq. (1.30).

By applying the chain rule to eq. (1.29) and using the fits to F1 (H1 ) and H1 (H), the

derivative of H with respect to x may be found:

  
dH1 dH1 dH
= (1.32a)
dx dH dx

d due dθ dH1
ue F 1 = (ue θH1 ) = θH1 + ue H1 + ue θ (1.32b)
dx dx dx dx
  
dH1 dH1 dH due dθ
ue θ = ue θ = ue F1 − θH1 − ue H1 (1.32c)
dx dH dx dx dx
dθ due
dH ue F1 − ue dx H1 − dx
θH1
= dH
(1.33)
dx 1
dH e
uθ

14
1.6 Transition Modeling

Near the location of transition, the flow becomes unstable and undamped Tollmien-

Schlichting waves appear [17]. They develop within the boundary layer until the flow

becomes fully turbulent and transitions. As described in White [17], the transition

from laminar flow to turbulent flow and the determination of its location is not built

of fundamental theory. Rather, tests and fits try to predict the location of transition

as a function of various parameters. While many parameters such as pressure gradient

and wall roughness may be chosen for functional relationships to determine transition,

many of the approximations use only one or two.

One such method is that of Michel [10]. This method uses only one parameter, the

local momentum thickness Reynolds number (defined in eq. (1.3)). Reθ is compared

to the Michel criteria, defined by:

 0.4
xue (x)
2.9 (1.34)
ν

According to this model, transition has occurred at the location where Reθ surpasses

the Michel criteria:

 0.4
xue (x)
Reθ = 2.9 (1.35)
ν

1.7 Runge-Kutta and Dense Output

Runge-Kutta solvers provide a numerical approximation for the solution to a first

order ordinary differential equation (or system of first order differential equations)

of the form y 0 = f (t, y), where f is a function returning the derivative of y at the

current value of the independent variable t and the simulation variable y [9]. Runge-

15
Kutta solvers are often referred to by their order, such as RK4, which indicates a 4th

order accurate solver. Solvers may have multiple schemes of different orders running

concurrently to maintain an error tolerance between the separate orders.

Runge-Kutta solvers are a natural fit for use in IBL methods. Surface distance x

may be used as the independent variable for the simulation. The y vector may be

composed of one or more variables specific to each IBL method.

For example, a Runge-Kutta implementation of Thwaites method could implement

the following expression for y:

y = θ2
 
(1.36)

resulting in the derivative expression

dθ2
 
0
y = (1.37)
dx

These values can be found given ue , due/dx, and values of s and H. A Runge-Kutta

implementation of Head’s method could implement the following expression for y:

 
θ
y=  (1.38)
H

resulting in the y 0 vector:  

dθ
dx
y0 =  (1.39)
 

dH
dx

where these relations may be obtained from eq. (1.15), eq. (1.29), eq. (1.30), and

eq. (1.31).

Runge-Kutta simulations often determine results in as few solution points as possi-

ble to reduce simulation times, as long as the error remains within a user defined

16
tolerance. However, fewer simulation points result in fewer locations with known

values of y when examining simulation results. Consider the IBL example above.

The user may need a boundary layer parameters at several specific x points. Setting

specific simulation points increases the time to run, and requires points of interest to

be determined before running each simulation. Interpolating between result points

introduces error unless an interpolant is used that is the same order of accuracy as

the Runge-Kutta solver. This undermines the accuracy of the solver itself. In order

to find accurate results at specific points without recomputing the solution, some

solvers have a property known as dense output. Dense output returns a polynomial

fit between each pair of simulation points. This polynomial has the same degree of

accuracy as the solver. This feature allows for quick querying of results with minimal

computational costs and without loss of accuracy.

Chapter 2 presents implementations of Thwaites’ and Head’s methods using a com-

mon Runge-Kutta solver between both implementations. The solver used in this work

is a Runge-Kutta of the order 5-4, a 5th order solver that uses a 4th order solution

as an error estimate in order to adjust the step size [21]. This chapter also covers

various implementation details, such as a common IBL parent class intended for use

with a variety of IBL methods, transition modeling, and separation. Chapter 3 dis-

cusses the results of the solvers compared against analytical solutions, experimental

data, and XFOIL results. Finally, Chapter 4 offers takeaways from the models and

opportunities for future work.

17
 Chapter 2

IMPLEMENTATION

2.1 Structure

While Thwaites’ Method and Head’s method have both been implemented using com-

putational methods (Cebeci and Bradshaw [8] present implementations of both), this

work provides a Python implementation that makes use of a common parent class and

user interface between the two that can be used as the basis for the implementation

of other Integral Boundary Layer methods. Figure 2.1 shows the overall structure of

the module. In addition, fig. 2.2 shows the classes implementing laminar separation,

boundary layer transition, and turbulent separation. The structure shown in fig. 2.1

serves multiple purposes. There are two major divisions between the object classes

in fig. 2.1 that may be made in order to better understand the implementation. The

first is a vertical division between Thwaites’ and Head’s method. The second is a

lateral division between the SimData classes and Sim classes.

2.1.1 Parent Classes

The division between Thwaites’ and Head’s methods is the more trivial of the two.

Thwaites’ and Head’s methods are separate methods for laminar and turbulent analy-

sis, so this division shows the classes associated with each method. This division also

highlights their parallel implementations and common inheritances from the IBLSim

and IBLSimData classes.

18
 IBLSimData
+char_length
+nu
+u_e_vec
+u_inf
+x_vec
-x_u_e_spline
+y0
-ue(x)
-duedx(x)

ThwaitesSimData HeadSimData
+theta0 +theta_0
+s(lam) +h_0
+h(lam) +c_f_0

<<use>> <<use>>

ThwaitesSim HeadSim
-__init__(ThwaitesSimData) -__init__(HeadSimData)
+c_f(x) +c_f(x)
+lam(x) +h(x)
+h_x(x) +rtheta(x)
+s_x(x) +theta(x)
+theta(x)

IBLSim
-sim_x_vec
-dense_output_vec
-derivatives(t, y)
-__init__(IBLSimData)
+y(x)

Figure 2.1: Structure of the object classes within PyBL.

TransitionModel SeparationModel
-iblsim -iblsim
+x_tr +x_sep
+h0 +separated
-transitioned -buffer
-buffer -criteria(x)
-criteria(x) -__init__(iblsim, criteria, buffer)
-__init__(iblsim, criteria, h0calc, buffer)

Michel ThwaitesSeparation HeadSeparation

-__init__(iblsim, buffer) -__init__(thwaitessim, buffer) -__init__(headsim, buffer)

Figure 2.2: Separation and transition object classes within PyBL.

19
Integral Boundary Layer methods share common attributes; they manipulate the

momentum integral equation, eq. (1.15), and relationships using an edge velocity dis-

tribution and other boundary layer parameters to arrive at a system of differential

equations that may be solved for over x. When pursuing the problems computa-

tionally, it follows that certain details can be abstracted from any Integral Boundary

Layer method and reused for other methods of the same category, in order to avoid

having to implement the same features more than once.

Implementing these two methods concurrently demanded that this relationship be

abstracted into a parent class. Both simulations need velocity distributions as func-

tions of x and one or more boundary conditions to run, and need to return results.

A user running a simulation through transition would make use of both models and

should not have to learn two separate conventions in order to implement them. At

the very least, a parent class maintains that the user interfaces between models are

consistent. In addition, updates and revisions that would benefit both IBL models

(or future models using the same parent classes) may be implemented in the parent

classes, as opposed to updating every model.

As shown in section 1.7, both simulations require current estimates of ue and due/dx

at any simulation point in order to complete their calculations of f 0 . In order to

accomplish this when given only points from a velocity distribution, a cubic spline

was implemented in the IBLSimData class [22]. The spline not only returns both the

edge velocity and due/dx as functions of x, but also ensures a smooth first derivative

for optimal behavior of the Runge-Kutta solver. It uses an edge condition that sets

the first and second segment at each curve end with the same polynomial. Internal

testing showed that this edge condition was appropriate for these results.

20
2.1.2 User Interface

The division between the Sim and SimData classes is a user interface feature that

uses object classes to differentiate between the setup and results of each simulation.

The SimData classes receive and process the data, creating the velocity distribution

spline and saving the appropriate initial conditions for access by the simulation. Any

modification made from one simulation to the next can be made on this class. For

example, changing the momentum thickness at the stagnation point does not require

creating simulation data from scratch. This class does not represent any simulation

results however, only the appropriate inputs.

The Sim classes accept only a SimData class as an input from the appropriate sim-

ulation. Creating a Sim class runs the simulation, storing dense output polynomials

from between each pair of solution points. In order to query the results, the IBLSim

parent class contains a method that takes one or more x values, finds the appro-

priate polynomial fit, and returns the appropriate y value or vector at that point.

Other boundary layer parameters may be extracted from y by the simulation-specific

objects.

Because the simulation inputs and the simulations are implemented as objects instead

of dictionaries or arrays, their attributes have permissions that keep them from direct

access to the user. This prevents values being edited erroneously, and tracks values

that are left modifiable by the user.

2.2 Thwaites’ Method

This implementation of Thwaites’ method presents substantial modifications to the

traditional implementations. An implementation using a Runge-Kutta solver to solve

21
Thwaites’ method improves the predictive results by eliminating assumptions that

were previously needed to find a solution, with similar computational costs.

Having defined the Thwaites parameter λ: eq. (1.16),

θ2 due
λ= (1.16 revisited)
ν dx

the momentum integral equation can be written as:

ue dθ2
= 2(−[H(λ) + 2]λ + s(λ)) = F (λ) ≈ .45 − 6λ (2.1)
ν dx

Although s and H are treated as functions of λ when querying results and are calcu-

lated from empirical fits, traditional implementations of Thwaites’ method approxi-

mate the relationship 2(−[H(λ) + 2]λ + s(λ)), or F (λ), as a linear relationship. This

linear approximation allows for the Thwaites integral, eq. (1.23) (restated below) to

be solved by hand for known velocity distributions:

Z x
2.45ν
θ = 6 u5e dx (1.23 revisited)
ue 0

When arranging this method for the Runge-Kutta integrator, eq. (2.1) may be solved

for dθ2/dx instead:

dθ2 ν(.45 − 6λ)
f0 = = (2.2)
dx ue

Using θ2 as the simulation variable y, λ can be obtained from θ2 and the velocity

spline, and this can be propagated downstream. Implemented in this way, this method

solves the same problem as traditional Thwaites method implementations.

22
Using the same method to estimate λ, and using fits to s and H, the assumption of

linearity for F may be instead removed so that eq. (2.1) then becomes:

dθ2 2ν(s(λ) − (2 + H(λ))λ)

f0 = = (2.3)
dx ue

This uses more detail on the front end, solving for θ using s and H.

In addition, a cubic spline was used to estimate s and H using Thwaites’ original

tabulated values. Fits introduced by Cebeci and Bradshaw [8] and White [17] intro-

duce up to 10% error at some points and are limited in their applicable ranges. In

addition, their piecewise nature does not ensure a smooth derivative for the entirety

of λ. Fitting a cubic spline to Thwaites’ values eliminates both of these issues.

With more care taken in modeling θ, Thwaites’ method was implemented in a more

consistent form to its derivation, and with more direct influence from Thwaites’ values

of s and H.

The module assumes that the starting point for ThwaitesSim, x_vec[0], is a stag-

nation point, unless theta0 is given. The momentum thickness is determined using

the following stagnation condition presented in Moran [14]:

s
.075ν
θ0 = due
(2.4)
dx 0

Note that the slope of the edge velocity distribution is used here and can have a signif-

icant impact on the boundary layer predictions. In addition, this relationship requires

the edge velocity derivative at zero to be infinite for zero momentum thickness.

For the special case of a flat plate, θ0 may be set equal to 0. If the velocity distribution

does not begin at the stagnation point, θ0 must be known and set with theta0. The

23
user may specify a starting point, x0, if not starting at the first point of the velocity

distribution, but must also specify theta0.

2.3 Michel Transition

In order to provide transition analysis for Thwaites’ method and accompanying initial

conditions for Head’s method, this implementation utilizes the Michel transition cri-

terion introduced in section 1.6. The Michel criteria is a satisfactory approximation

and is a sufficient proof of concept to show that the laminar and turbulent solvers

could be tied together to simulate transition. In this respect, more focus was given

to the laminar and turbulent solvers themselves.

The transition model uses a parent class, TransitionModel, in order to make some

of the transition features available for other models. Michel is a class that takes a

completed simulation (SimData) as an input, and has methods and attributes that

find and describe the transition point. Taking advantage of the laminar solver’s dense

output, the implementation of the Michel criteria inherits an iterative method from

its parent class to find the point of transition, where

 .4
xue (x)
Reθ = 2.9 (1.35 revisited)
ν

After this point, the Michel class will return an estimated value of the shape factor

after transition, Htr , . The method uses the following relationship presented by Coles

[11]:
1.4754
Htr = + .9698 (2.5)
ln(Reθ |xtr )

The Michel object returns this value after solving for xtr .

24
2.4 Head’s Method

This implementation of Head’s method follows the implementation presented by Ce-

beci and Bradshaw [8] fairly closely, using the same empirical fits. H1 (H) is fit with

a piecewise function, restated below:


3.3 + 0.8234(H − 1.1)−1.287

 for H ≤ 1.6
H1 (H) ≈ (1.28 revisited)
3.3 + 1.5501(H − 0.6778)−3.064

 for H > 1.6

In order to determine dH1/dH , it is useful to note that both segments are of the form:

H1 (H) = 3.3 + a(H − b)c (2.6)

and that their derivatives are of the form:

dH1
= ac(H − b)c−1 (2.7)
dH

dθ/dx and dH/dx may then be obtained as described in section 1.5.2.

2.5 Laminar and Turbulent Separation

Two separation models were implemented in this project for comparison with test

cases, based on the possibility that nonphysical or inconsistent results could be the

product of separation. A SeparationModel parent class was created similarly to the

TransitionModel class, with methods to optimize for the x value of separation using

the solvers’ dense output capabilities.

25
The criteria used for laminar separation using Thwaites method is a λ value of:

λ < −.0842 (2.8)

Which corresponds to a value of s(λ) = 0. This value was noted by Moran [14] along

with correlation formulas for λ, and the assertion that the wall shear stress is 0 at

separation.

For Head’s method, the criteria used for separation was instead the value of H. This

was due to the Ludwieg-Tillman relation for cf LT = 0.246 ∗ 10−.678H Rθ−.268 tending to

0 as H grows. Shape factor was shown to have sufficient correlation with turbulent

separation when:

1.8 < H < 2.4 (2.9)

The specificity of the value was not seen as critical since the derivative of H was also

very high at separation. This relationship was also prescribed by Moran [14].

No criteria was implemented to simulate reattachment. If the boundary layer was

found to have separated at any point, results downstream of that point were said

to be invalid, since the solvers had no implementation to account for a separated

boundary layer.

26
 Chapter 3

TEST CASES AND RESULTS

3.1 Testing Thwaites’ Method Against Falkner Skan

Because Falkner-Skan flow can be solved analytically with Thwaites’ method, it was

chosen as a test case for the current method. This series of tests allowed for com-

parison between the Falkner-Skan solution, Thwaites’ method solved analytically, a

Runge-Kutta implementation with linearized F (λ), and finally, a Runge-Kutta im-

plementation with non-linearized F (λ), interpolating Thwaites’ tabulated values to

determine s and H as functions of λ. For these flows, dθ2/dx is infinite at zero, so

these simulations begin at an offset from zero. Chapter A describes the reason for

this choice in more detail.

These various cases produce results with accuracies consistent with Thwaites’ method.

Thwaites’ method is known to be within 10% of exact results before separation [17].

The analytic solution using Thwaites’ method matches the Falkner-Skan solution

closely, but not exactly. Section 1.5.1 shows how Thwaites’ method may be used to

calculate the momentum thickness for Falkner-Skan flow. In addition, the simulations

with non-linearized F (λ) were expected to differ from the numerical solutions using

the linearized F (λ), since different equations are be solved.

Even the solution between the numerical and analytical solutions using linearized

F (λ) (posed to solve mathematically equivalent cases) were expected to differ, as a

result of using a numerical solver.

27
The solver was tested at various values of m. Three are presented here. As noted

earlier, Falkner-Skan flow with a value of m = 0 reduces to Blasius flow. Each of the

numerical methods was given a starting value of θ0 from the analytic Falkner-Skan

result.

The three simulations are meant to compare against known results and validate the

performance of the algorithm in controlled conditions. The last solution in this section

tested the resilience of this implementation and made a naive attempt at solving

without the analytical solution for θ0 , instead implementing the Moran stagnation

condition. While it is important that this algorithm perform well with specific initial

conditions, the fourth case is intended to show its robustness in cases where the initial

conditions are not known.

The results were compared against the Falkner-Skan solution and the analytical

Thwaites solution for the corresponding m and u∞ values. While the derivative

of the velocity profile may be derived analytically, the Runge-Kutta solver was given

only points from the velocity distribution in order to show the ability of the cubic

spline to sufficiently represent the situation.

3.1.1 Falkner-Skan Flow, m=0

At m = 0, the Falkner-Skan solution simplifies to the Blasius result. This results in a

uniform edge velocity distribution ue = u∞ . Figure 3.1 demonstrates that the solver

results match very closely with the analytical solution using Thwaites.

28
 ·10−3 ·10−3
6
2
4
θ (m)

δ (m)
1
2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)
·10−2
2.8

2
2.6
H
cf

1
2.4

0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

x(m) x(m)
Falkner-Skan
Thwaites (Analytical)
Thwaites
Thwaites (Linear)

Figure 3.1: Falkner-Skan Flow, m=0 Results. For this flow, the analytical,
linear, and interpolated Thwaites method are very consistent.

29
 θ
0
10 δ
cf
H
10−1
Relative Error
10−2

10−3

10−4

10−5
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x(m)

Figure 3.2: Falkner-Skan Flat Plate Flow Error.

θ
100 δ
cf
H
10−1
Relative Error

10−2

10−3

10−4

10−5
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x(m)

Figure 3.3: Falkner-Skan Flat Plate Flow - Linear Method Error.

30
3.1.2 Falkner-Skan Flow, m=1

For this flow, the Falkner-Skan velocity distribution simplifies to:

x
ue = u∞ (3.1)
L

with the velocity derivative simplifying to:

due u∞
= (3.2)
dx L

This simulation’s results are presented in fig. 3.4, and the errors for each simulation are

presented in fig. 3.5 and fig. 3.6. This result shows that the linearized and interpolated

Thwaites produce similar results for this condition to the analytic solution. While the

solution begins at the given value of θ0 , it quickly gravitates towards the Thwaites

result and has similar error.

3.1.3 Falkner Skan Flow, m=1/3

While the m = 0 and m = 1 examples both gave useful results for quantifying the

effectiveness of the various new implementations of Thwaites, they were both fairly

special cases of Falkner-Skan flow. The case with m = 0 results in constant edge

velocity, and both m = 0 and m = 1 have constant derivatives.

A value of m = 1/3 was chosen as an in-between value to display testing against a

more average Falkner-Skan flow. The results for a simulation at this condition are

displayed in fig. 3.7.

31
 ·10−4 ·10−4
9

3.5 8.5
θ (m)

δ (m)
8
3
7.5
2.5
7
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)
0.3

0.2 2
H
cf

0.1
1.5

0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)
Falkner-Skan
Thwaites (Analytical)
Thwaites
Thwaites (Linear)

Figure 3.4: Results for Thwaites’ method on Falkner-Skan Flow, m=1.

This result shows more fluctuation in the results.

32
 θ
100 δ
cf
H
10−1

Relative Error
10−2

10−3

10−4

10−5
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x(m)

Figure 3.5: Falkner-Skan Flow Error, m=1. This result was generated
using the full estimation of F (λ) with cubic fits for s and H.

θ
100 δ
cf
H
10−1
Relative Error

10−2

10−3

10−4

10−5
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x(m)

Figure 3.6: Falkner-Skan Flow Error. This result was generated using the
linear estimation of F (λ) with the cubic fits for s and H used only after θ
had been solved.

33
The spline for the m = 0 and m = 1 cases was given a relatively simple velocity

distribution to approximate for the other two cases, given that their m values both

eliminated the exponential qualities of their velocity distributions. In comparison,

this simulation left an exponential function to be estimated by a cubic spline.

Figure 3.7, fig. 3.8 and fig. 3.9 show the total results and associated errors for this

simulation. These results show similar errors to the m = 0 and m = 1 cases, regardless

of the increased complexity of the edge velocity distribution. Testing showed that the

interpolated values from Thwaites’ paper maintained similar error to the linearized

method.

34
 ·10−3 ·10−3
1.5
3
1
θ (m)

δ (m)
2
0.5 1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)

0.1 2.4
H
cf

5 · 10−2 2.35

0 2.3
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)
Falkner-Skan
Thwaites (Analytical)
Thwaites
Thwaites (Linear)

Figure 3.7: Falkner-Skan Flow Results, m=1/3. These results are the result
of a more complex velocity distribution.

35
 θ
0
10 δ
cf
H
10−1

Relative Error 10−2

10−3

10−4

10−5
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x(m)

Figure 3.8: Falkner-Skan Flow Error, m=1/3. This error was calculated
using the Falkner-Skan result as reference. These results are well within
the 10% error reported for the analytic solution.

θ
0
10 δ
cf
H
10−1
Relative Error

10−2

10−3

10−4

10−5
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x(m)

Figure 3.9: Falkner-Skan Flow Error, m=1/3, using linearized F. This sim-
ulation was analogous to the analytical solution, which also had an offset
error from the Falkner-Skan solution.

36
 ue
34
Smoothed ue
Smoothed derivative
32

30

ue (m/s)
28

26

24

0.5 1 1.5 2 2.5 3 3.5 4 4.5

x(m)

Figure 3.10: Comparison of different velocity inputs to Head’s method for

a flat plate with mild adverse pressure gradient

3.2 Head’s Method on Flat Plate with Mild Pressure Gradient

In order to test the accuracy of the Head’s method implementation, simulations were

run to compare results with a test case presented by Ludwieg and Tillman [5]. The

test case was a flat plate with a mild adverse pressure gradient.

Verification against this specific set of experimental results demonstrated the rela-

tionship between the smoothness of the Head implementation’s inputs and its ability

to produce useful outputs. Experimental data presented by Ludwieg and Tillman

[5] (and, notably, tabulated in [7]) was used to run the turbulent simulation. The

newer, tabulated data was presented with two plotted results: the edge velocity, ue

as a function of x, and the derivative of edge velocity, due/dx. Both papers note that

this data had been smoothed, but the exact method by which it was smoothed was

not described.

37
 −5

−4

−3

due −2
dx
−1

0
ue
1 Smoothed ue
Smoothed derivative
2
0.5 1 1.5 2 2.5 3 3.5 4 4.5
x(m)

Figure 3.11: Comparison of different velocity derivatives for each spline

input to Head’s method for a flat plate with mild adverse pressure gradient

In order to run this simulation, points for edge velocity were used from the tabulated

data, extracted from the smoothed plot, and extrapolated from due/dx values taken

from the smoothed derivative plot. Figure 3.10 shows the three splines that were

generated from points taken from the tabulated values and smoothed ue and derivative

plots. Note that the smoothing omitted a small discrepancy near x = 3.3.

While the splines in fig. 3.10 seem to match fairly closely, the derivatives of the splines

(shown in fig. 3.11) show significant differences. The splines from the tabulated ue

and smoothed ue were generated fitting only the points without consideration for

the smoothness of the derivative. The third spline, because it was built from the

smoothed, derivative-extrapolated points, provided an additional degree of smooth-

ness. Its antiderivative spline (and appropriate integration constant) were used to

run a third simulation.

38
These different velocity conditions had significant impacts on the boundary layer

results. Comparing a variety of results for different boundary layer parameters in

fig. 3.12, it is clear that smoothing the derivative made a significant difference in the

accuracy of the program and produced results more closely matching the experimental

results compared to the other edge velocity treatments. Boundary layer data shown

are the reported values from Kline et al [7].

·10−2 ·10−2
4
2
3
θ (m)

δ (m)
2
1
1

0 0
1 2 3 4 1 2 3 4
x(m) x(m)
·10−3
1.5
2.5
1
H
cf

2
0.5
1.5
0
1 2 3 4 1 2 3 4
x(m) x(m)
Tabulated Values
ue
Smoothed ue
Smoothed derivative

Figure 3.12: Flat plate with mild adverse pressure gradient, Head’s
Method. The simulation results using different velocity distrubutions,
alongside tabulated experimental results.

39
 100 θ
δ
cf
H
10−1
Relative Error

10−2

10−3
1 1.5 2 2.5 3 3.5 4 4.5
x(m)

Figure 3.13: Head’s method results for a flat plate, mild adverse pressure
gradient. Error for each boundary layer parameter. Without smoothing
implemented, the simulation failed near x = 3.6.

θ
δ
10−1 cf
H
Relative Error

10−2

10−3

1 1.5 2 2.5 3 3.5 4 4.5

x(m)

Figure 3.14: Head’s method results for a flat plate, mild adverse pressure
gradient with smoothed ue data. Error for each boundary layer parameter.

40
 θ
δ
10−1 cf
H

Relative Error

10−2

10−3

1 1.5 2 2.5 3 3.5 4 4.5

x(m)

Figure 3.15: Head’s method results for a flat plate, mild adverse pressure
gradient using smoothed velocity derivative distribution. Error for each
boundary layer parameter. The error improves slightly from the smoothed
ue result.

0.4
ue
0.35 Smoothed ue
Smoothed derivative
0.3

0.25
un (m/s)

0.2

0.15

0.1

0.05

0
1 1.5 2 2.5 3 3.5 4 4.5
x(m)

Figure 3.16: Flat Plate with mild adverse pressure gradient, Head’s
method - transpiration velocity results.

41
This case shows that in some instances, providing edge velocity data may result in

non-physical results when due/dx is not sufficiently smooth. When possible, smoothed

data should be used.

This test shows the capabilities of this Head’s method implementation to model tur-

bulent boundary layers. When practical, smoothing of the edge velocity derivative

seems to return more accurate results. For example, in a case where an inviscid

solver is to be paired with Head’s method, the transpiration velocity, un , would be

very important in generating new panel strengths and running a new inviscid solution.

Figure 3.16 shows calculated transpiration velocities from each simulation’s results.

Transpiration velocity may be calculated using the following relationship:

due dH dθ
un = θH + ue θ + ue H (3.3)
dx dx dx

Clearly, these results show that the smoothness of the velocity derivative has a signif-

icant impact on the resulting transpiration velocity. Smoothness of the data in a use

case like this would be very important. However, in a case where the user is looking

for the momentum thickness at a certain point on the surface, smoothness of the edge

velocity profile or its derivative may not be an issue.

3.3 Testing Against XFOIL

In order to better demonstrate the results of the different methods incorporated into

the module, the results from this work were compared to XFOIL against the current

method. Note that since XFOIL tightly couples the Integral Boundary Layer solu-

tion to its panel strength solution iteration and uses a more sophisticated transition

42
criteria, these results will not match exactly1 . This tight coupling normally involves

several iterations to obtain the final edge velocity and boundary layer results. To

minimize the effects of this tight coupling, care was taken to run only one viscous

iteration of XFOIL for comparison of PyBL and XFOIL results. Even with this con-

sideration, XFOIL’s viscous method is admittedly different and equal results between

these methods should not be expected, nor is XFOIL the gold standard for boundary

layer results. Integral Boundary Layer methods stand in validity on their own [8],

but XFOIL provides another tool to compare against.

XFOIL was used to create the inviscid flow over the airfoil to be analyzed. From this

solution the edge velocities and stagnation point were found and used as the initial

conditions for PyBL. XFOIL was then run in viscous mode, taking care to limit viscous

iterations to one, to obtain the results needed for the comparisons. This process

allowed the ability to control iterations of XFOIL’s viscous solver and settings for

transition method and location. For comparison against the fully turbulent solution

with Head’s method, this was important for forcing transition as early as possible.

In order to minimize the influence XFOIL’s tightly coupled viscous solver has on its

results, its number of iterations was set to one. However, this does not eliminate the

effect of the boundary layer solution being closely coupled with its panel algorithm.

This results in the data returned from XFOIL not being a true one-way coupling of the

inviscid solution and the Integral Boundary Layer method as in this work. However,

the comparisons are still instructive for examining the predicition capabilities of this

work. Finally, it is recognized that the Michel transition criteria used for this work is

not as capable as XFOIL’s implementation of the eN method; however, its inclusion

in this work is presented here as a proof of concept. The introduction of a transition

1
Since the Integral Boundary Layer solution affects the inviscid solution, via the transpiration
velocity term, XFOIL uses an interative solution process to converge on the solution and solves for
the panel strengths and Integral Boundary Layer terms simultaneously.

43
criteria parent class provides a building block for future authors to implement their

own transition criteria using the spline and dense output capabilities of the module.

3.3.1 Laminar Results

In order to compare laminar results for as much chord length as possible, the laminar-

only test case was chosen at a low Reynolds number, with an airfoil at 0◦ angle of

attack. The airfoil chosen was a NACA 0003 airfoil, which is an airfoil with zero

camber and maximum thickness of 3% of the chord length. This yielded an XFOIL

prediction that did not transition or separate until nearly the trailing edge. The

results are displayed in fig. 3.17. Figure 3.18 shows the associated differences against

XFOIL.

This case did not trigger the module’s transition criteria, but did trigger the sepa-

ration criteria at x = .956. This chord location is where there is very high velocity

derivative.

As seen most prominently in the displacement thickness and shape factor, an insta-

bility in the result arises near x = 0.6 and grows downstream. This fluctuation can

be attributed to the non-smooth edge velocity profile in the XFOIL inviscid results at

the same location, as shown in fig. 3.19. While this may not provide a large percent

difference for velocity, the velocity derivative is significantly affected, also shown in

fig. 3.19. This is another indication, along with the Head’s method example, that

smoothing of the edge velocity and its derivative should be investigated further.

44
 ·10−2 ·10−2
8
2
6
θ (m)

δ (m)
4
1
2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)

3.5
2

3
H
cf

2.5
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8
x(m) XFOIL x(m)
PyBL

Figure 3.17: Comparison against the fully laminar PyBL result and
XFOIL. PyBL reported laminar separation at x = .956.

45
 θ
2 δ
10
cf
101 H

Relative Difference
100

10−1

10−2

10−3

10−4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
x(m)

Figure 3.18: Difference between the fully laminar PyBL result and XFOIL.
Note that PyBL reported separation at x = .956, so the large spike in
difference near x = 1 is not guaranteed accuracy by the method.

19.6 0
−20
19.4
−40
19.2 −60
(m/s2 )
ue (m/s)

−80
19
−100
due
dx

18.8 −120

18.6 −140
−160
18.4
−180
0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1
x(m) x(m)

Figure 3.19: The velocity distribution from XFOIL provided difficulty in

derivative approximation due to its jagged nature.

46
3.3.2 Turbulent Results

For the turbulent results, the choice of airfoil and conditions was less restrictive. In

order to emulate a fully turbulent boundary layer, XFOIL was forced to transition

near the leading edge. A value of x = 0.01 was used as the point at which to transition.

Because x = 0 was the point chosen for initial conditions for the turbulent solver,

this introduced a difference in the result at the leading edge, as seen in fig. 3.20 and

fig. 3.21. Note that there is a discrepancy at the leading edge because of XFOIL’s

transition criteria, which smooths out the transition into a range of chord locations.

After this initial difference, the turbulent results show strong agreement with XFOIL

for the rest of the domain, up until the trailing edge. For this case, the turbulent

separation criteria was met by the Python module very close to the trailing edge

at x = 1.02, due to the large edge velocity velocity derivative at that point. This

value corresponds to the sharp increase in difference with XFOIL for all values at the

trailing edge.

47
 ·10−3 ·10−2

1
4
θ (m)

δ (m)
2 0.5

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)
·10−2

8 2

6
H
cf

4 1
2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)
XFOIL
PyBL

Figure 3.20: Results from fully turbulent XFOIL (forced transition at

x = .01) and fully turbulent PyBL simulation (Head’s Method only) using
XFOIL’s initial conditions. XFOIL’s more gradual transition implementa-
tion bled remnants of laminar shape factor into the turbulent results.

48
 100

10−1

Relative Difference 10−2

10−3
θ
δ
cf
10−4 H

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

x(m)

Figure 3.21: The differences of Head’s method with respect to XFOIL

results where transition was forced at x = 0.01.

3.3.3 Combined Results

This simulation was a comparison against XFOIL with a full implementation of stag-

nation criteria, the laminar Thwaites solver, transition criteria, the turbulent Head

solver, and separation criteria for both the laminar and turbulent simulations. A

different airfoil and flow conditions were tested in order demonstrate the versatility

of the solver. The following result is a NACA 0009 airfoil, run at a Reynolds number

of 2 × 106 and angle of attack of 0◦ . The boundary layer parameters for this result

are displayed in fig. 3.22. The Python module predicted transition at x = 0.492.

This differed from the XFOIL prediction of x = .66. The results showed noticeable

differences in this region, with the largest differences occurring at PyBL’s transition

location. The discrepancies propagated downstream, leaving the relative difference

fairly high for this result. Figure 3.23 shows the differences for variables of interest

for this simulation. While the differences are significant, this is a challenging test case

49
where errors will accumulate as they are propogated downstream. Note that the skin

friction results matched once the XFOIL boundary layer transitioned. Figure 3.22

shows that the trend of the shape factor was not dissimilar to the XFOIL result,

except for noticeable differences near transition.

Overall, there is much better matching between the XFOIL and PyBL results com-

pared to the natural transition case. The laminar region matches quite well, and

noticeable differences arise at the transition location because of XFOIL’s treatment

of the transition. The smoothing of the the transition region impacts the predicted

properties at the transition location as well as downstream. However, this effect is

much less pronounced than in the natural transition case.

Figure 3.25 shows the difference between XFOIL and PyBL where the differences in

the laminar region are much less than 10%. The differences in displacement thickness,

skin friction coefficient, and shape factor are highest at the transition location and

then generally decrease downstream. For the momentum thickness, those differences

do not spike at the transition location because both implementations keep the mo-

mentum thickness constant at the transition location because both implementations

keep the momentum thickness constant at the the transition location. However, the

XFOIL smoothing does impact the differences downstream as expected.

While XFOIL results are not to be accepted as truth, they provide valuable compar-

ison for implementation of IBL methods.

50
 ·10−3 ·10−3
3
4
2
θ (m)

δ (m)
1 2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)
·10−2
6 3

4 2.5
H
cf

2
2
1.5
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)
XFOIL
PyBL

Figure 3.22: Results for full implementation of Michel transition criteria,

laminar and turbulent solvers, against XFOIL. The curves showed simi-
larities, even with transition location disagreement. While sufficient as a
proof of concept for transition criteria, major differences existed between
the two solutions.

51
 θ
101
δ
cf
100 H
Relative Difference

10−1

10−2

10−3

10−4

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

x(m)

Figure 3.23: Full simulation result difference against XFOIL. Note that the
largest difference occurs between the two solvers’ estimates of xtr . The Htr
value estimated by Michel did not represent the same gradual transition
as XFOIL’s eN method.

52
 ·10−3 ·10−3
3
4
2
θ (m)

δ (m)
1 2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)
·10−2
6

2.5
4
H
cf

2
2
1.5
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x(m) x(m)
XFOIL
PyBL

Figure 3.24: Results after forcing XFOIL to transition at the point pre-
dicted by Michel criteria. The instantaneous change in shape factor de-
fined by the Michel method showed a clear difference from the gradual
change in the XFOIL result.

53
 θ
δ
100 cf
H
Relative Difference

10−1

10−2

10−3

10−4

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

x(m)

Figure 3.25: Simulation result difference after forcing XFOIL to transition

at the same point as PyBL. A significant offset was still present in the
results near transition.

54
 Chapter 4

CONCLUSION AND FUTURE WORK

This project shed new light on implementations for Thwaites’ and Head’s methods.

The results have shown to have sufficient agreement for various test cases, and were

implemented such that a user could query an accurate solution value at any point

along the surface. Cases run naively with only points from an inviscid velocity dis-

tribution matched very closely to other viscous simulations, and the dense output

functionality resulted in a streamlined implementation of transition and separation

criteria.

In terms of future work to be done, more accurate transition criteria should be im-

plemented in order to produce more realistic results for transition and downstream.

Further, a deeper change to Head’s method (similar to removing the linear assump-

tion of Thwaites’ method for the s and H functions) could add to its fidelity. Better

fits to the empirical relations of Head’s method could improve the quality of the re-

sults. More complex integral boundary layer models can be implemented to improve

the results and provide a wider selection of modeling options. This could include

transformations for axisymmetric bodies or by integrating along streamlines for 3-

dimensional geometries.

Not all features of the flow field may be well represented by the cubic spline method

used in this work. For example, surfaces with abrupt steps or gaps may not be

accurately represented by a smooth spline. This framework can be enhanced to

encompass these cases. Finally, more work may be done to investigate the error

introduced by using a cubic spline to interpolate within a 5-4 order Runge-Kutta

55
solver. It is possible that lower or higher order solvers may be a better fit, or that

large step sizes decrease accuracy since portions of the spline may be stepped around

entirely. More work should be done to ensure that this grouping of tools does not

present inaccuracies by promoting instability and oscillations in the results.

As a general lesson from this project, the empirical and experimental fits and splines

were less accurate in the extreme cases and reduced the quality of the results from

the solver in those regions. This also affected the solution process with the ODE

solver taking large spatial steps, and resulted in nonphysical values or leaving the

domain for the fits. This typically resulted in poor quality solutions or the solutions

diverging.

Finally, creating more comparisons against experimental data would greatly improve

the definition of limits for these solvers and provide a valuable resource for additional

models to be developed. Testing done in this work provided a good baseline com-

parison, but there are a wide range of cases to be covered that are of importance to

aerodynamicists.

Revisiting these methods decades later has proven to be fruitful in refining their

implementations. Only by continued effort will they reach their fullest adoption.

56
 Bibliography

[1] H. Blasius. “Grenzschichten in Flussigkeiten mit kleiner Reibung, 2. angew”.

In: Math. Phye 56 (1908).

[2] V.M. Falkner and S.W. Skan. Aeronautical Research Council Report 1314. 1930.

[3] B. Thwaites. “Approximate Calculation of the Laminar Boundary Layer”. In:

Aeronautical Quarterly 1.3 (Nov. 1, 1949), pp. 245–280. issn: 0001-9259. doi:

10.1017/S0001925900000184.

[4] H Blasius. The Boundary Layers in Fluids with Little Friction. NACA technical

memorandum 1256. Washington, Feb. 1950, p. 58.

[5] Hubert Ludwieg and W. Tillmann. Investigations of the Wall-Shearing Stress

in Turbulent Boundary Layers. Report 1285. Number: NACA-TM-1285. May

1950.

[6] M. R. Head. Entrainment in the Turbulent Boundary Layer. Aeronautical Re-

search Council Report 3152. London: Her Majesty’s Stationary Office: Ministry

of Aviation, Sept. 1958.

[7] Hubert Ludwieg and W. Tillmann. “Ludwieg and Tillman, mild adverse pres-

sure gradient (1100)”. In: Computation of turbulent boundary layers: 1968 AFOSR-

IFP-Stanford Conference. Ed. by S. J. Kline et al. Vol. II. Meeting Name: Con-

ference on Computation of Turbulent Boundary Layers. [Stanford] Calif., Ther-

mosciences Division, Stanford University, 1969, pp. 55–70.

[8] Tuncer Cebeci. Momentum transfer in boundary layers. In collab. with P. Brad-

shaw and Long Beach Faculty publications California State University. Series in

thermal and fluids engineering. Washington: Hemisphere PubCorp, 1977. isbn:

978-0-07-010300-9.

57
 [9] J. R. Dormand and P. J. Prince. “A family of embedded Runge-Kutta formu-

lae”. In: Journal of Computational and Applied Mathematics 6.1 (Mar. 1, 1980),

pp. 19–26. issn: 0377-0427. doi: 10.1016/0771-050X(80)90013-3.

[10] R. Michel, D. Arnal, and E. Coustols. “Stability Calculations and Transition

Criteria on Two- or Three-Dimensional Flows”. In: Laminar-Turbulent Tran-

sition. Ed. by V. V. Kozlov. International Union of Theoretical and Applied

Mechanics. Berlin, Heidelberg: Springer, 1985, pp. 455–461. isbn: 978-3-642-

82462-3. doi: 10.1007/978-3-642-82462-3_55.

[11] Alec David Young. Boundary Layers. Google-Books-ID: n1GAAAAYAAJ. Amer-

ican Institute of Aeronautics and Astronautics, 1989. 296 pp. isbn: 978-0-

930403-57-7.

[12] Stephen J. Pollard. “Evaluation of the CMARC panel code software suite for the

development of a UAV aerodynamic model”. Accepted: 2013-04-30T21:56:07Z.

Thesis. Monterey, California. Naval Postgraduate School, 1997. url: https:

//calhoun.nps.edu/handle/10945/31947 (visited on 08/16/2021).

[13] Giovanni P. Galdi. “An Introduction to the Navier-Stokes Initial-Boundary

Value Problem”. In: Fundamental Directions in Mathematical Fluid Mechanics.

Ed. by Giovanni P. Galdi, John G. Heywood, and Rolf Rannacher. Advances in

Mathematical Fluid Mechanics. Basel: Birkhäuser, 2000, pp. 1–70. isbn: 978-3-

0348-8424-2. doi: 10.1007/978-3-0348-8424-2_1.

[14] Jack Moran. An introduction to theoretical and computational aerodynamics.

Mineola, N.Y: Dover Publications, 2003. 464 pp. isbn: 978-0-486-42879-6.

[15] John D. Anderson. “Ludwig Prandtl’s Boundary Layer”. In: Physics Today

58.12 (Dec. 1, 2005). Publisher: American Institute of Physics, pp. 42–48. issn:

0031-9228. doi: 10.1063/1.2169443.

58
[16] Tuncer Cebeci and Jean Cousteix. Modeling and computation of boundary-layer

flows. 2., rev. and extended ed. Berlin Heidelberg: Springer, 2005. 502 pp. isbn:

978-0-9668461-9-5 978-3-540-24459-2.

[17] Frank White. Viscous Fluid Flow. 3rd edition. New York, NY: McGraw Hill,

Jan. 1, 2011. isbn: 978-1-259-00212-0.

[18] Hermann Schlichting and Klaus Gersten. Boundary-Layer Theory. 9th ed. Berlin

Heidelberg: Springer-Verlag, 2017. isbn: 978-3-662-52917-1. doi: 10.1007/978-

3-662-52919-5.

[19] Arthur E. P. Veldman. “Entrainment and boundary-layer separation: a model-

ing history”. In: Journal of Engineering Mathematics 107.1 (Dec. 2017), pp. 5–

17. issn: 0022-0833, 1573-2703. doi: 10.1007/s10665-017-9930-x.

[20] Tuncer Cebeci and Jean Cousteix. Modeling and Computation of Boundary-

Layer Flows. isbn: 0-9668461-9-2.

[21] scipy.integrate.RK45 — SciPy v1.7.0 Manual. url: https://docs.scipy.org/

doc/scipy/reference/generated/scipy.integrate.RK45.html (visited on

07/02/2021).

[22] scipy.interpolate.CubicSpline — SciPy v1.7.0 Manual. url: https : / / docs .

scipy.org/doc/scipy/reference/generated/scipy.interpolate.CubicSpline.

html (visited on 07/07/2021).

59
 APPENDICES

Appendix A

FALKNER-SKAN FLOW NEAR ZERO

While Falkner-Skan flow is a well-established laminar flow case, it provided difficulty

for numeric implementation of Thwaites’ method starting at zero due to its behavior

near zero. For this reason, the results provided in this thesis avoid starting near zero,

starting instead at x = 0.05. Depending on values of m ∈ [0, 1], the derivative of

θ2 with respect to x can take a variety of values near zero. These values might be

well-behaved analytically, but this term is often problematic numerically. For these

cases, the assumptions of both the cubic spline derived from the velocity distribution

and the Runge-Kutta solver leave the program with no ability to accurately represent

the behavior of the initial conditions. For some cases, the slope is infinite, and the

cubic spline cannot represent thst. This results in the polynomial fits being poorly

suited for the solver.

At each time step, this Runge-Kutta implementation calculates λ:

θ2 due/dx
λ= (1.16 revisited)
ν

then solves for the derivative of θ2 with respect to x:

dθ2 2ν(s − [2 + Hλ)]

= (A.1)
dx ue

60
At the first solver step, the values of ue and θ (given that the exact value of θ from the

Falkner-Skan solution is used) are exact values. Note that if ue = 0, then this term

will usually be ∞ (depending on the value of the numerator), and the ODE solver will

not be able to proceed. Even if the derivative is finite for this case, special treatment

would be needed for the numerical implementation to handle the indeterminate 0/0

result. In addition, due/dx is a cubic spline estimate, and this makes the evaluation of

λ problematic in some situations. The Falkner-Skan flow edge velocity can directly

calculated as an input given:

 x m
ue = u∞ (1.12a revisited)
L

The analytic edge velocity derivative, which the solver is only estimating, has the

form:
due mu∞  x m−1
= (A.2)
dx L L

At x = 0, ue has the values:






 u∞ for m = 0


ue (0) = 0 for 0 < m < 1 (A.3a)





0
 for m = 1

Immediately after x = 0, ue has the values:






 u∞ for m = 0


ue (x > 0) = x m (A.3b)


 u∞ L
for 0 < m < 1



x


u∞
 for m = 1
L

61
The value of due/dx at x = 0 (found analytically) take the forms:






 0 for m = 0

due 
(0) = ∞ for 0 < m < 1 (A.3c)
dx 



 u∞


L
for m = 1

Immediately after x = 0:





0 for m = 0

due 
(x > 0) = mu∞ x m−1 (A.3d)


dx  L L
for 0 < m < 1



 u∞


L
for m = 1

Examination of eq. (A.1) and eq. (A.3a)-eqs. (A.3a) to (A.3d) shows that for x = 0,
dθ2/dx is not finite for m > 0. Further, eq. (A.3c) and eq. (A.3d) show that due/dx is

infinite for some values of m at x = 0. For most cases, a cubic spline cannot accurately

represent this function. For these reasons, the Falkner-Skan flows presented in this

thesis begin at x = 0.05.

62