A SURFACE VORTICITY METHOD

FOR WAKE–BODY INTERACTIONS

A Dissertation
Presented to
The Academic Faculty

By

David Joyner Pate

In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy in the
School of Aerospace Engineering

Georgia Institute of Technology

May 2017

Copyright
c David Joyner Pate 2017
A SURFACE VORTICITY METHOD FOR WAKE–BODY INTERACTIONS

Approved by:

Professor Brian J. German, Advisor Professor Graeme Kennedy

School of Aerospace Engineering School of Aerospace Engineering
Georgia Institute of Technology Georgia Institute of Technology

Professor Lakshmi Sankar Dr. Erik D. Olson

School of Aerospace Engineering Aeronautics Systems Analysis Branch
Georgia Institute of Technology NASA Langley Research Center

Professor Marilyn Smith Date Approved: March 31, 2017

School of Aerospace Engineering
Georgia Institute of Technology
 ACKNOWLEDGEMENTS

While this thesis represents about two years of work, my journey really started a long

time ago, and there are many people I would like to thank. Thanks to my parents for all

their support, especially Mom for taking an active role in my education. My wife, Cindy,

for gently reminding me that seven years is a long time to spend as a grad student. My

brother, Michael, for guiding the countless hours of music I listened to during grad school.

And, of course, my thesis advisor, Prof. Brian German, for taking a chance on me and

helping me become the researcher I am today.

Thanks Matt Daskilewicz for helping me get started and for raising the bar. Youngjun

Choi for being a little bit crazy. Michael Patterson for helping me through those first few

years—I think we complemented each other well. Andrea and Giada for teaching me how

to speak proper Italian. Marc Canellas for always being that ray of sunshine when I’m

feeling blue. Xiaofan Fei for playing a few games with me every now and then. And

thanks to the rest of the past and present GRG members. We’ve had some good times.

Of course, I must also acknowledge the teachers and professors I’ve had over the years

who took the time to get to know me, and, as a result, helped me immensely.

iii
 TABLE OF CONTENTS

Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x

Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Research Hypothesis and Contributions . . . . . . . . . . . . . . . . . . . 6

Chapter 2: Free Vortex Sheets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.1 Governing Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.1.1 Velocity–Vorticity Kinematic Relations . . . . . . . . . . . . . . . 10

2.1.2 Vortex Sheet Evolution Equations . . . . . . . . . . . . . . . . . . 11

2.2 Literature Review for Modeling Vortex Sheet Evolution . . . . . . . . . . . 14

2.2.1 Vortex Sheets in Two Dimensions . . . . . . . . . . . . . . . . . . 14

2.2.2 Vortex Sheets in Three Dimensions . . . . . . . . . . . . . . . . . . 18

2.3 Numerical Method for Regularized Vortex Sheet Evolution . . . . . . . . . 22

2.3.1 Discretization with Triangular Panels . . . . . . . . . . . . . . . . 23

2.3.2 Vorticity Update via Half-Edge Circulations . . . . . . . . . . . . . 25

2.3.3 Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.3.4 Adaptive Paneling . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.3.5 Fast Summation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

iv
 2.4 Evolution of a Sphere in a Uniform Freestream . . . . . . . . . . . . . . . 41

Chapter 3: Bound Vortex Sheets . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3.1 Governing Equations for the Surface Vorticity Model . . . . . . . . . . . . 47

3.1.1 Bound Vortex Sheet Definition and Continuous Problem Formulation 48

3.1.2 Unsteady Pressures . . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.1.3 Lifting Flow and the Unsteady Kutta Condition . . . . . . . . . . . 52

3.2 Triangular Panel Representation . . . . . . . . . . . . . . . . . . . . . . . 55

3.2.1 Panel Perimeter Circulation Constraint . . . . . . . . . . . . . . . . 58

3.3 Numerical Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . 59

3.3.1 Normal Flow Constraint in Matrix Form . . . . . . . . . . . . . . . 60

3.3.2 Panel Perimeter Circulation Constraint in Matrix Form . . . . . . . 61

3.3.3 Numerical Non-Lifting Problem Formulation . . . . . . . . . . . . 61

3.4 Non-Lifting Comparison to Analytical Solution for Ellipsoid . . . . . . . . 63

Chapter 4: Combined Bound and Free Vortex Sheets . . . . . . . . . . . . . . . 69

4.1 Problem Formulation for Unsteady Lifting Flows . . . . . . . . . . . . . . 70

4.2 Wake Shedding Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . 74

4.3 Algorithm for Combined Bound and Free Vortex Sheets . . . . . . . . . . . 75

4.4 Comparison to Rectangular Wing Experiment . . . . . . . . . . . . . . . . 76

4.5 Comparison to Wind Turbine Experiment . . . . . . . . . . . . . . . . . . 84

4.6 Example Simulation: Two Aircraft in Close Formation . . . . . . . . . . . 89

Chapter 5: Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

v
 5.1 Free Vortex Sheets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

5.2 Bound Vortex Sheets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

5.3 Combined Bound and Free Vortex Sheets . . . . . . . . . . . . . . . . . . 100

5.4 Realizability Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

5.5 Summary of Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 102

5.6 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

Chapter A: Evaluation of the Biot–Savart Integral . . . . . . . . . . . . . . . . . 106

A.1 Geometric Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

A.2 Specification of Domains of Integration . . . . . . . . . . . . . . . . . . . 107

A.3 Coordinates for Integration . . . . . . . . . . . . . . . . . . . . . . . . . . 108

A.4 Integral solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

Appendix B: Treecode Description and Performance . . . . . . . . . . . . . . . 112

Appendix C: Potential Flow Around an Ellipsoid . . . . . . . . . . . . . . . . . 118

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

vi
 LIST OF FIGURES

2.1 Notation for vector quantities associated with a vortex sheet. . . . . . . . . 11

2.2 The notation for calculating the circulation around closed loop C. This is a
side view of the vortex sheet. . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.3 Notation for the triangular panels. . . . . . . . . . . . . . . . . . . . . . . 24

2.4 Half-edge circulations around a triangle. . . . . . . . . . . . . . . . . . . . 26

2.5 Initial and final triangulation of the vortex sheet in the z = 1/2 plane. . . . . 29

2.6 Comparison of surface vorticity at t = 0.25. . . . . . . . . . . . . . . . . . 31

2.7 Example of the velocity field induced by a vortex sheet with smoothing and
without smoothing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.8 Example of triangle splitting. . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.9 Example of edge removal. . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

2.10 Example of Delaunay edge flipping. . . . . . . . . . . . . . . . . . . . . . 34

2.11 The triangulated sheet as it evolves at iterations 0, 100, and 1000. Also
indicated are the infinite vortex filaments. . . . . . . . . . . . . . . . . . . 36

2.12 The results of the adaptive paneling algorithm test, clockwise from top left:
the final triangulation, the computed volume at each iteration normalized
by the exact volume of a sphere, a histogram of the edge lengths, and a
histogram of the vertex degrees. . . . . . . . . . . . . . . . . . . . . . . . 36

2.13 Shadowgraph of the deformed sheet at the end of the adaptive paneling test. 37

2.14 Comparison of computational time for several approaches to evaluating

Equation 2.65. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.15 The evolution of a spherical vortex sheet over 1000 iterations. The shad-
owgraphs reveal the shape of the sheet while the radiographs reveal the
vorticity carried by the sheet. . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.16 Contour of integration for calculating the circulation. . . . . . . . . . . . . 44

vii
2.17 The time history of the simulated evolution of a spherical vortex sheet. . . . 44

3.1 Cross section of a wing illustrating the interpretation of the trailing edge. . . 52

3.2 Cross section of a wing illustrating paths along the surface to a point on the
trailing edge. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3.3 Notation for the triangular panels. . . . . . . . . . . . . . . . . . . . . . . 56

3.4 The vector γd in the vertex plane with normal vector d3 must be related
to some vector γc in the triangle plane with normal vector c3 . All vectors
outside of the (c1 , c2 ) plane are accompanied by thin gray lines to indicate
their height (except c3 ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.5 Aggregate metrics indicating the error associated with the relative weight-
ing w for ellipsoids of varying resolution. . . . . . . . . . . . . . . . . . . 65

3.6 Convergence of different calculations with the proposed numerical method. 66

3.7 Pathlines of surface vorticity, pathlines of surface velocity, and surface

pressure coefficient for the numerical flow solution. . . . . . . . . . . . . . 67

3.8 Histograms of the residuals for the numerical calculation of surface vortic-
ity on the ellipsoid with 5120 triangles. . . . . . . . . . . . . . . . . . . . . 67

3.9 Streamsurfaces around an ellipsoid calculated from the numerical solution. . 68

4.1 A sheet of fluid encasing a body before and after motion. . . . . . . . . . . 70

4.2 The tip section of a sample triangulated wing demonstrating the topological
split along the trailing edge. Note that the trailing edge vertex pairs are
actually collocated. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4.3 Triangulated body and wake vortex sheets illustrating the wake shedding
procedure. Note that the groups of orange vertices are actually collocated
(the wake is attached to a closed trailing edge). . . . . . . . . . . . . . . . 75

4.4 Shadowgraph and radiograph of the simulated wake. . . . . . . . . . . . . 79

4.5 Comparison to Figure 22(c) in McAlister [157]. . . . . . . . . . . . . . . . 81

4.6 Comparison to Figure 39 in McAlister [157]. . . . . . . . . . . . . . . . . . 82

viii
4.7 Comparison to Figure 37(a) in McAlister [157]. . . . . . . . . . . . . . . . 83

4.8 Blade triangulation and pressure measurement locations juxtaposed with

Figure 9 in Hand et al. [159]. . . . . . . . . . . . . . . . . . . . . . . . . . 84

4.9 Wake visualizations for 5 m/s freestream. . . . . . . . . . . . . . . . . . . 86

4.10 Wake visualizations for 7 m/s freestream. . . . . . . . . . . . . . . . . . . 87

4.11 Location of tip vortex compared to the prediction of Lynch et al. [160] for
the 7 m/s case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

4.12 Section pressures for 5 m/s (left) and 7 m/s (right) cases compared with
NREL UAE measurements [159]. Note the different vertical axis scales. . . 88

4.13 Two notional flying wings arranged in a close formation. . . . . . . . . . . 89

4.14 Individual computing times for each calculation during an iteration. . . . . 90

4.15 Radiograph of the simulated wakes. Darker regions indicate strong vorticity. 93

4.16 Shadowgraph of the simulated wakes. Darker regions indicate many layers
of the vortex sheet. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

4.17 A shadowgraph with a closer view of the aircraft, which are not transparent. 95

4.18 Surface pressures for each side of each aircraft. The bottom surfaces are
left–right mirrored. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

A.1 Geometric relations between triangle (P1 , P2 , P3 ) and point PA measured

relative to origin O. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

A.2 Three surfaces, S1 , S2 , and S3 , used to cover exactly S. In this case, S3 has
opposite orientation of S1 and S2 . . . . . . . . . . . . . . . . . . . . . . . . 108

A.3 Cartesian coordinates (u, v) and polar coordinates (ρ, φ) for integration on
S1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

B.1 Example quadtree. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

B.2 Results for the treecode parameter experiments colored by each parameter. . 117

ix
 NOMENCLATURE

The following list describes the symbols used in this thesis. Bold notion indicates a vector
in 3-space. The overhead arrow notation indicates a column matrix.

α vortex particle strength

γ surface vorticity

ω vorticity

a1 , a2 , a3 fixed inertial reference frame and orthonormal basis vectors

b1 , b2 , b3 body-fixed reference frame and orthonormal basis vectors

c1 , c2 , c3 triangle orthonormal basis vectors

d1 , d2 , d3 vertex orthonormal basis vectors

n normal vector

p vertex position vector

r relative vector

v velocity

x position

δ smoothing parameter; Dirac delta function

Γ circulation

Γ1a , . . . , Γ3b half-edge circulation strengths

ν kinematic viscosity

φ scalar potential field

ρ density
~λ list of Lagrange multipliers
~b list of right-hand-side terms

~x list of degrees of freedom

~z list of degrees of freedom in nonlinear optimization scheme

A matrix

x
AΓ circulation loops matrix

Aeq equality constraints matrix

Anf normal flow matrix

Appc panel perimeter circulation matrix

a00 , . . . , b01 linear surface vorticity coefficients

C path

Cp pressure coefficient

f, g surface vorticity scalare functions

M equality-constrained quadratic program matrix

Ndof number of degrees of freedom

NT number of triangles

NV number of vertices

P vertex

p pressure

S surface

s path parameter; wing semispan

t time

u, v triangle Cartesian coordinates

V volume

v∞ freestream velocity

W weighting matrix

w weighting factor

x, y, z Cartesian coordinates

xi
 SUMMARY

The objective of this dissertation research is to develop a surface vorticity method for

simulating high Reynolds number incompressible aerodynamic flows with strong unsteady

interactions between wakes and lifting bodies. Examples of these types of flows include

rotors in hover, propeller/wing installations, and impingement of vortex cores shed from

wing strakes or flaps on downstream surfaces. Although higher-order panel codes provide

good representation of potential flow around lifting bodies, their treatment of wakes is

inadequate for our purpose. In the absence of significant boundary layer separation, the

vorticity in these flows concentrates into thin shear layers. Therefore, vortex sheets are a

natural mathematical representation of these flows.

We leverage and extend rigorous methods from the vortex methods literature to model

a wake as a free vortex sheet discretized as a triangulation of panels with linearly varying

surface vorticity. The vorticity evolution equation is solved approximately by maintaining

constant circulation along each half-edge in the triangulation, an approach that generalizes

current methods for constant-strength elements. The vortex sheet is regularized with a

smoothing parameter which provides an apparent thickness that mimics the limited viscous

mixing in high Reynolds number flow. An adaptive paneling algorithm is implemented to

maintain the desired level of detail as the wake triangulation stretches and deforms. The

induced velocities from the wake vortex sheet are computed with a treecode implemented

on a graphics processing unit (GPU) to allow computations with millions of panels.

Lifting bodies are modeled with bound vortex sheets that are also triangulated with

linear strength panels. These higher-order vorticity elements provide accurate velocity pre-

xii
dictions on and near the surface, allowing for high resolution streamline tracing. Surface

vorticity is determined by enforcing flow tangency constraints at each triangle centroid,

zero circulation around each panel perimeter, and the unsteady pressure matching Kutta

condition. These constraints result in an overdetermined system that is solved in a least

squares formulation. Thus, our method is a second-order surface vorticity boundary el-

ement method that combines both solid bodies and wakes in a rigorous and consistent

manner.

The results of the method are shown to compare favorably to wind tunnel experimental

results, including wake profiles, for a rectangular wing in a steady freestream, and for a

horizontal axis wind turbine. Finally, we demonstrate the capabilities of our method in the

context of strong wake–body interactions by simulating two flying wing aircraft in close

formation, with the wake from the leading aircraft impacting the tailing aircraft.

xiii
 CHAPTER 1

INTRODUCTION

Aerodynamic flows in which lifting bodies are tightly coupled with the wakes they shed

are particularly challenging to simulate. A few common examples include: a hovering

helicopter rotor shedding a wake that remains very close to the rotor; multiple propellers

closely distributed along a wing shedding wakes that flow around the wing; and a wing

with deflected flaps shedding wakes that form strong vortex cores and then pass near a tail

surface. In each of these cases, there is a strong mutual interaction between the bodies and

the wakes, which implies that the accurate simulation of one (e.g., unsteady pressures acting

on a body) requires the accurate simulation of the other (e.g., wake strength and location).

These types of flows are also inherently unsteady and do not admit a relaxation-based

solution. The purpose of this thesis is to develop a numerical method capable of simulating

these interacting lifting systems for incompressible, high Reynolds number flows.

Problems of this nature have traditionally been solved by Navier–Stokes methods based

on Eulerian grids. Difficulties with these methods arise from numerical dissipation of vor-

ticity and the challenges with creating intricate volume grids. Panel methods based on

Laplace’s equation have not been a viable alternative for high-fidelity simulations because

the typical techniques for wake modeling are insufficient at adapting to the complex wake

structures in these flows.

However, Lagrangian methods based on the Helmholtz vorticity transport equation have

matured to become a realistic alternative to grid-based Navier–Stokes methods for simulat-

1
ing vorticity-dominated flows [1]. Indeed, incompressible flow can be described by vortic-

ity alone, and in high Reynolds number flow, the vorticity is confined to thin layers because

it diffuses much slower than it advects. Thus, the flow may be described by the distribution

of vorticity concentrated in thin sheets instead of by distributions of primitive flow variables

in a volume grid throughout the entire flow field. Additionally, Lagrangian methods have

the benefit of automatically adapting the computational elements to concentrate where they

are needed [2].

One challenge with Lagrangian methods has been the interaction with solid boundaries.

Also, these methods have historically suffered from the high computational cost of calculat-

ing the mutual influence of elements with the Biot–Savart law. Fortunately, fast summation

techniques originally developed for N -body problems in gravitation and electrostatics have

recently alleviated this limitation.

1.1 Literature Review

Among the first numerical models for simulating unsteady wakes is that of Belotserkovskii

and Nisht [3]. They use vortex filaments arranged in a quadrilateral-based lattice that is

algebraically equivalent to vortex rings. The vortex rings convect under their own mutual

influence while maintaining a constant circulation strength. A wing is treated as thin and

also represented by vortex rings with a no-normal-flow condition enforced at the centroids.

New vortex rings are shed from the trailing edge and are assigned a circulation strength to

satisfy the unsteady Kutta condition. This method, most often called the unsteady vortex

lattice method, was further developed [4–12], including in PMARC [13], in which the thick

wing is modeled. Katz and Plotkin [14] codified the standard form of this method, and it

2
remains popular due to its simplicity [15–20]. Versions of this method have also been

applied to helicopter rotors, including the work of Quackenbush et al. [21], Wachspress et

al. [22], and others [23–25]. Unfortunately, the unsteady vortex lattice method does not

model vortex sheet evolution in a sufficiently rigorous and robust manner to simulate the

motion of a vortex sheet as it approaches the immediate neighborhood of a body.

Several authors have paired high-order panel methods with wakes represented by a

lattice of vortex filaments or vortex particles [26–34]. For example, Willis [35, 36] im-

plemented this approach with a fast multipole tree algorithm to accelerate the vortex par-

ticle induced velocity computations. Similar to the unsteady vortex lattice method, these

works also use a fixed grid-based representation of the wake. By not refining the wake

as it evolves, the panels are stretched beyond any meaningful physical representation. For

particle methods, this behavior manifests in unacceptably large gaps between nearby par-

ticles. As stated by Voutsinas [37], “the stretching of the wake surface grid is not totally

unphysical . . . what is unphysical is to keep the same grid and disregard the changes due to

deformation.”

Low-order panel methods have also been paired with particle-based wakes, as exempli-

fied by the work of Calabretta and McDonald [38]. In contrast with the heretofore discussed

works, Martin [39] adaptively refined the wake based on a vortex particle splitting scheme

introduced by Mansfield et al. [40]. In this scheme, a newly created particle is assigned

three mutually orthogonal vectors that evolve alongside the particle strength vector. When

the length of a vector exceeds a chosen threshold, the particle is split into two particles

placed in-line with the vector. Unfortunately, Martin experienced numerical stability is-

sues when applying this method to a boat propeller. The velocities induced by the constant

3
doublet panels are singular at the panel edges; thus, these panels produce large, nonphys-

ical velocities and velocity gradients nearby. This behavior adversely affects the particles

that recirculate back toward the propeller hub, especially with the inclusion of the parti-

cle splitting scheme. Also, Martin did not incorporate fast-summation techniques and was

therefore limited to a relatively small number of particles. Opoku et al. [41] employed a

thin wing low-order panel method and vortex particle wakes in a particle–mesh method.

Zhao and He [42–44] paired a vortex particle method with lifting-line theory to represent

a helicopter rotor in hover and forward flight, and included the adaptive particle splitting

scheme of Mansfield et al. [40]. Lastly, Park et al. [45] paired a vortex lattice method with

vortex particles to represent a helicopter rotor in hover.

To simulate helicopter rotor–wake interactions, some authors have coupled a primitive-

variable-based Navier–Stokes methods, which capture the compressible and viscous effects

near the rotor, with a vortex particle method, which captures the vorticity in the wake.

Stock et al. paired OVERFLOW, a Navier–Stokes method with Chimera overset grids, with

a Lagrangian vortex particle method [46]. By including a fast multipole treecode and exe-

cuting their programs on an Intel/Tesla cluster computer, they simulated a helicopter rotor

in forward flight with over 55M particles and 46M grid points. Zhao and He [47] also

coupled OVERFLOW with a vortex particle method and saw improved predictions for a

UH-60A rotor as compared to a solitary OVERFLOW simulation. Additionally, Rajmo-

han et al. [48] implemented a coupled method with OpenFoam, an unstructured finite vol-

ume Navier–Stokes flow solver, and He and Rajmohan [49] used FUN3D, an unstructured

Reynolds-averaged Navier–Stokes (RANS) method, to model a helicopter fuselage.

Several authors have noted the difficulty faced by traditional Navier–Stokes CFD meth-

4
ods when modeling vorticity-dominated flows, such as the blade–wake interactions with

a helicopter rotor. Caradonna [50] states: “. . . it is clear that resolving scales of the order

of the vortex core diameter imposes an enormous burden on such a solution,” and that the

main reason for this grid resolution is to avoid the numerical dissipation of vorticity inher-

ent in a finite-differencing method. Similarly, Strawn and Djomehri [51] state: “. . . current

numerical schemes and computational grids are unable to preserve the vortices long enough

to predict accurately the noise and airloads caused by blade–vortex interactions.” For more

recent work in this area, see Eshcol et al. [52], and for a review of the literature see Strawn

et al. [53] and Hariharan et al. [54].

However, the numerical dissipation of vorticity can be mitigated by solving the incom-

pressible Navier–Stokes equations in vorticity–velocity form (see Equation 2.4). As noted

by Speziale [55] and Gatski [56], an additional benefit of this approach is a simpler handling

of the far field boundary conditions. Brown [57] applied the vorticity–velocity approach to

capture blade–wake interactions for a helicopter rotor. An efficient solution on a volume

grid is obtained via a fast Poisson solver. Brown noted that one of the main challenges in

these simulations is that the vehicle dynamics of interest are on a much larger timescale

than the aerodynamics. They applied the fast multipole method to calculate velocity on

the volume grid, and they included an adaptive grid generation technique that created and

destroyed grid nodes to follow the regions of intense vorticity. Brown and Line model a

helicopter fuselage with vortex rings (i.e., vortex filaments), and they represent the rotor

with an unsteady lifting line formulation. Additional applications of the vorticity transport

method include Scheurich and Brown [58], and Whitehouse and Boschitsch [59]. Smith

et al. [60] compared the vorticity transport method to the primitive-variable Navier–Stokes

5
solver FUN3D.

Finally, the vorticity transport method has also been coupled with traditional CFD meth-

ods to solve the compressible, viscous flow near the blade with primitive variables, and

then to solve the external vorticity-dominated flow in vorticity–velocity form. Whitehouse

et al. [61] paired the vorticity transport method with OVERFLOW, FUN3D, and RSA3D

(an unstructured RANS method) to model helicopter aerodynamics. Similarly, Whitehouse

and Tadghighi [62] paired a vorticity–velocity formulation with OVERFLOW.

This overview of related works is not comprehensive. Indeed, these and similar meth-

ods for modeling wings and wakes have been combined in numerous ways to varying de-

grees of success. Nonetheless, at present, the aerodynamic simulation of tightly coupled

wakes and lifting bodies remains challenging, regardless of the computational approach.

1.2 Research Hypothesis and Contributions

As evidenced by this diverse set of literature, there are many methods and hybrid com-

binations of methods for modeling wings and wakes, each with certain advantages and

disadvantages, such as accuracy, computational expense, and ease of use. A gap within

these prior works is the lack of a higher-order surface element representation of the vortic-

ity within the flow.

The fundamental hypothesis of this research is that interacting aerodynamic flows with

strong couplings between wakes and lifting bodies can be modeled efficiently and effec-

tively as vortex sheets discretized into triangular panels with linearly varying surface vortic-

ity. Accordingly, the prime contribution of this thesis is a method for accurate simulations

of wake–body interactions, which is enabled by multiple individual mathematical and nu-

6
merical contributions.

To this end, we develop a boundary element method for which the surface of an arbitrary

body is represented by a bound vortex sheet discretized into a triangular mesh with a linear

variation of surface vorticity along each triangle. This approach is an original contribution,

as compared to previous low-order or thin-surface methods. Additional contributions com-

prise a zero panel perimeter circulation constraint to achieve a solenoidal vorticity field, a

vector rotation scheme that provides a fully continuous surface vorticity field, and a least

squares solution approach.

A wake is represented by a free vortex sheet, also discretized in the same manner.

This higher-order representation of the wake elements and the corresponding vorticity up-

date scheme is a novel extension of the constant strength elements used by Stock [63] and

Feng [64]. An additional contribution is a novel Delaunay edge flipping scheme, which

helps maintain a more regular triangulation. Finally, our GPU-based implementation of

a treecode is also an important contribution, enabling the large number of computational

elements required for high resolution simulations.

In Chapters 2 and 3, the free and bound vortex sheet formulations and simulation meth-

ods are derived independently from the Helmholtz decomposition of velocity, with incre-

mental validation tests presented along the way. In Chapter 4 we present our full numerical

method, combining bound and free vortex sheets to model complex wake–body interac-

tions. Then, we compare predictions from this method to two sets of wind tunnel experi-

ments. Finally, we demonstrate the application of this method with an example simulation

in which two aircraft are flying in close formation with the wake of the lead aircraft im-

pinging upon the tailing aircraft.

7
 CHAPTER 2

FREE VORTEX SHEETS

Vortex sheets are often used to model incompressible fluid flows in which vorticity is con-

fined to thin shear layers [65]. A vortex sheet is a mathematical representation of a shear

layer as a surface across which the tangential velocity is discontinuous but the normal ve-

locity is continuous [66]. In three dimensions, a vortex sheet is defined by a distribution

of vorticity concentrated along a continuous surface. Vortex sheets often arise in high

Reynolds number flows in which viscous diffusion is small at the reference velocity.

The numerical simulation of vortex sheets has been an extensive area of research in the

field of vortex methods. Cottet and Koumoutsakos [1] state: “Vortex methods have reached

today a level of maturity, offering an interesting alternative to finite difference and spectral

methods for high-resolution numerical solutions of the Navier–Stokes equation.” As com-

pared to grid-based Navier–Stokes and Euler methods, Lagrangian vortex methods do not

suffer from the numerical dissipation of vorticity [67] or the need to create an appropriate

volume mesh.

In the next section, we present the fundamental equations for vortex sheets in exact

form. In Section 2.2 we discuss the relevant literature—general concepts in the field of

vortex methods as well as similar works. In Section 2.3 we present our numerical method

for vortex sheet evolution, including an adaptive paneling algorithm and a treecode for

fast summation. Finally, Section 2.4 concludes with a common example from the vortex

methods literature.

8
2.1 Governing Equations

In three dimensions the constant-density Navier–Stokes equations comprise conservation

of momentum
∂v 1
+ v · ∇v = − ∇p + ν∇2 v (2.1)
∂t ρ

and conservation of mass

∇ · v = 0, (2.2)

where v is velocity, t is time, ρ is density, p is pressure, and ν is kinematic viscosity.

Vorticity is defined as the curl of velocity,

ω := ∇ × v, (2.3)

and by taking the curl of Equation 2.1, an alternate form is obtained in terms of vorticity:

∂ω
+ v · ∇ω = ω · ∇v + ν∇2 v. (2.4)
∂t

This equation is often referred to as the vorticity evolution equation [1, page 6],[14, page

24].

Helmholtz’s second vorticity theorem requires that vortex lines move through a fluid

as material lines [68–70], enabling a Lagrangian description of the flow. The position of a

Lagrangian marker x starting at x0 is therefore given in differential form as

dx(x0 , t))
= v(x0 , t), x0 := x(x0 , 0), (2.5)
dt

which may be solved by integrating the velocity of the fluid as an ordinary differential

equation.

9
2.1.1 Velocity–Vorticity Kinematic Relations

Equation 2.3 provides vorticity in terms of velocity; however, we will need velocity in

terms of vorticity, which is obtained from velocity–vorticity kinematics.

The Helmholtz decomposition stipulates that any vector field may be expressed as the

sum of a solenoidal vector field and an irrotational vector field [69, page 84],[63, page 4].

Then, for any general velocity field v, we may specify

v = vω + vφ (2.6)

with the solenoidal vector field vω satisfying

∇ × vω = ω, ∇ · vω = 0, (2.7)

and the irrotational vector field vφ satisfying

∇ × vφ = 0, ∇ · vφ = σ, (2.8)

where σ is the local volumetric rate of expansion and ω is the local vorticity. The irrota-

tional field can be related to a scalar function φ as vφ = ∇φ.

When combined, the equations governing vω produce a Poisson equation

∇2 vω = −∇ × ω (2.9)

that can be solved using the infinite-medium Green’s function [63, page 4],[65, page 307]

as
1 (x0 − x) × ω(x0 )
Z
vω (x) = dV (x0 ), (2.10)
4π ||x0 − x||3

10
which is often called the Biot–Savart law. Equation 2.6 may now be rewritten as

1 (x0 − x) × ω(x0 )
Z
v(x) = 0 3
dV (x0 ) + vφ , (2.11)
4π ||x − x||

which provides velocity in terms of vorticity and an irrotational velocity field.

2.1.2 Vortex Sheet Evolution Equations

A vortex sheet is represented by an orientable surface S paired with a vector field γ in the

local tangent space of S. This vector field is called the surface vorticity and determines the

strength of the vortex sheet. As described by Stock [67] and Sohn [71], surface vorticity is

defined as

γ := n × (v + − v − ), (2.12)

where n is the unit vector normal to S oriented outward, v + is the velocity immediately

above (or outside) S, and v − is the velocity on the opposite side. The sheet moves with the

average of the velocity on either side

vsheet = 21 (v − + v + ). (2.13)

This notation is demonstrated in Figure 2.1.

n
v + ∆v
vsheet
s v−
γ
sheet
s
Figure 2.1: Notation for vector quantities associated with a vortex sheet.

Pozrikidis [72] derived the vortex sheet evolution equation by applying the inviscid

11
momentum equation to v + and v − and relating the results via Equations 2.12 and 2.13:

∂γ
= − v · ∇γ + γ · ∇v − γ(∇ · P · v) . (2.14)
∂t | {z } | {z } | {z }
advection stretch dilation

P is the tangential projection operator, which is a dyadic (second order tensor) defined as

P := I − nn, (2.15)

where I is the identity dyadic, n is the sheet normal vector, and the juxtaposition of two

vectors indicates the dyadic product [73, page 67]. The advection and stretch terms are

common to the vorticity evolution equation; the dilation term is unique to vortex sheets and

accounts for changes in surface vorticity parallel to the direction of its strain [63].

The surface vorticity is related to vorticity via the Dirac delta function as ω = γδ(n),

where n is the distance normal to the sheet. This relation allows Equation 2.11 to be restated

in terms of surface vorticity as

1 (x0 − x) × γ(x0 )
Z
v(x) = dS(x0 ) + vφ . (2.16)
4π ||x0 − x||3

Next, the circulation around a closed path is defined as

I
Γ := v · dl. (2.17)

To evaluate the circulation along a portion of a vortex sheet, consider a path C0 along the

sheet from point P1 to point P2 and a corresponding closed loop C that parallels C0 and

pierces the sheet at points P1 and P2 , as demonstrated in Figure 2.2. Parameterize path C0

as C0 (s) with C0 (s1 ) = P1 and C0 (s2 ) = P2 .

12
 n
∆v
+
v
vsheet

P2 s
C v−
γ sheet
P1
Figure 2.2: The notation for calculating the circulation around closed loop C. This is a side
view of the vortex sheet.

Then, using the relation

v + − v − = ∆v = γ × n, (2.18)

the circulation around C can be expressed in terms of surface vorticity as

I
ΓC = v · dl
C
Z Z
= +
v · dl + v − · dl
−C0
ZCs02 Z s1
= +
v (s) · dl(s) + v − (s) · dl(s)
Zs1s2 s2

= (v + (s) − v − (s)) · dl(s)

Zs1s2
= ∆v · dl(s)
s1
Z s2
= (γ(s) × n(s)) · dl(s). (2.19)
s1

This result leads to a fundamental relation between circulation and surface vorticity.

Consider an arbitrary unit vector m in the tangent plane of the surface at an arbitrary point

x. Then, by differentiation under the integral sign, the rate of change of circulation in the

direction of m at x is

dΓm (x)
= ∆v · m = (n × m) · γ(x). (2.20)
ds

13
Furthermore, if the tangent space at x is spanned by orthonormal basis (b1 , b2 ) with re-

spective Cartesian coordinates (u, v), the total differential of circulation is

dΓ = b2 · γdu − b1 · γdv. (2.21)

2.2 Literature Review for Modeling Vortex Sheet Evolution

We now review the body of literature broadly categorized as vortex methods, with a limita-

tion to methods in which discrete Lagrangian elements carry vorticity and represent a thin

vortex sheet. The three-dimensional methods with sheet elements share the most similari-

ties with our method.

2.2.1 Vortex Sheets in Two Dimensions

Point Vortex Method

Numerical approximations of a vortex sheet originated with the point vortex method (also

called the discrete vortex method), which can be traced to the early works of Rosenhead

[74] and Westwater [75]. In this method, vortex sheets are represented by discrete point

vortices with a specified value of circulation, which is constant with time for inviscid flow.

The general procedure for time-stepping point vortices with ordinary differential equations

can be traced back at least as far as Rosenhead [74]. As described by Kuwahara and Takami

14
[76], the equations of motion in Cartesian coordinates (x, y) for the ith particle are

dxi −1 X yi − yj
= Γj 2 ,
dt 2π j6=i ri,j
dyi 1 X xi − xj
= Γj 2
,
dt 2π j6=i ri,j
2
ri,j = (xi − xj )2 + (yi − yj )2 .

These early calculations were done by hand and limited to a small number of particles.

As a fortuitous result, the calculated roll-up patterns appeared realistic and matched the

expectations of Prandtl [77]. When Birkhoff and Fisher attempted to replicate the results

with a computer and with many more particles, irregularities became apparent, with the

point vortices “tending toward randomization of position” [78]. Further improvements

to the point vortex method were made, including requirements for convergence, and the

method eventually became codified as the Birkhoff–Rott equation [79–85].

Regularization and the Vortex Blob Method

Chorin and Bernard [86] took a different approach to resolving the irregularities associated
with point vortices:

It is fairly obvious that a point vortex approximation to a vortex sheet cannot be

taken too literally, since a point vortex induces a velocity field which becomes
unbounded, and cannot approximate a bounded field in any reasonable norm . . .
However, we conjecture that as soon as the velocity field of the point vortices is
smoothed out and made bounded, i.e., the point character of the point vortices
is not taken too literally, the approximation becomes reasonable.

15
In their method, the circumferential velocity vθ at a radial distance r from a point vortex

with strength Γ0 is modified as


Γ
 0

if r ≤ rc ,


vθ (r) = 2πrc
 Γ0


if r > rc .

2πr
They suggest this spreading of vorticity is analogous to artificial viscosity, and the method

produced improved results.

Around the same time, Kuwahara and Takami [76] took a similar approach. They also

applied a viscous core model to “suppress the unessential irregularity by introducing an

‘artificial viscosity.’” They modified the velocity induced by a point vortex with an exact

solution to the Navier–Stokes equations for a single vortex in a viscous fluid introduced by

Oseen [87] and further developed by Rott [88]. In this viscous vortex model, the circum-

ferential velocity at a radial distance r from a point vortex with strength Γ0 is

 2 
−r

Γ0
vθ (r, t) = 1 − exp ,
2πr 4νt

and the corresponding vorticity is a Gaussian distribution

−r2
 
Γ0
ω(r, t) = exp .
4πνt 4νt

They note that applying this viscous core model to multiple point vortices does not retain

exactness because the nonlinearity of the Navier–Stokes equations does not permit super-

position of independent solutions. Further implementations of this approach include Bloom

and Jen [89] and Leonard [65]. In contrast to the viscous core model used by Chorin and

Bernard, the Gaussian core function used by Kuwahara and Takami spreads with time.

16
 The concept of regularizing the point vortex method by desingularizing the vorticity

is generally referred to the vortex blob method and is attributed to Chorin and Bernard

[86]. Krasny [90] and Anderson [91] formulated an implementation of this method which

includes a smoothing parameter δ > 0 to mollify the induced velocity as

Γ0
vθ (r) = √ ,
2π r2 + δ 2

and the corresponding equations of motion are referred to as the “δ-equations.” The smooth-

ing parameter, which was originally introduced by Rosenhead [92], controls the rate at

which the vortex sheet rolls up [93].

Smoothing is often insufficient to prevent the eventual formation of irregularities. Fink

and Soh [81] reposition the point vortices after each iteration such that they are equally

spaced and adjust the strength of each point vortex accordingly. Krasny [84, 90] applied a

spectral filter to suppress noise incited by computational machine precision errors.

Also, as the vortex sheet strains, gaps form between the vortex blobs (i.e., the cut-

off radii cease to overlap). Indeed, Hald [94, 95] relies on this overlapping in his proof

of convergence for the vortex blob method. If these gaps coincide with high curvature,

the discrete representation of the vortex sheet may also intersect itself, which should be

avoided.

Krasny [96] discusses two possible solutions to avoid these gaps: start with a large

enough number of vortices, or adaptively insert new vortices. The most efficient solution,

and often the most practical, is the latter. Winckelmans et al. [97] refer to a more general

form of this adaptive point insertion as particle redistribution which is conducted either

locally or globally.

17
 Further detailed descriptions of the vortex blob method may be found in references [1,

7, 65, 97, 98].

Singularity Formation

From the beginning of the development of the vortex blob method, it was soon realized

that vortex sheets tend to roll up into singularities [99–102]. As phrased by Krasny [84],

“the singularity which appears is an infinite jump discontinuity in the vortex sheet’s curva-

ture.” While the singularity poses a computational challenge, a more general challenge is

the exponential growth of the arc length of the vortex sheet. With the use of adaptive point

insertion, this causes a rapid increase in the number of vortices and subsequent computa-

tional burden. Singularity formation and rapid growth of surface area is also a characteristic

of 3-D vortex sheets [103].

2.2.2 Vortex Sheets in Three Dimensions

In three dimensions, the inviscid vorticity transport equation is

∂ω
+ v · ∇ω = ω · ∇v. (2.22)
∂t

The term on the right-hand side is the vortex stretching term, which is absent in the 2-D

form of this equation. Cottet and Koumoutsakos [1, page 55] state “this term fundamentally

affects the dynamics of the flow; it is [particularly] responsible for vorticity intensification

mechanisms that make long-time inviscid calculations very difficult.” The solution tech-

nique for this term is driven by the degree of connectivity between computational elements,

and it is this connectivity which characterizes 3-D vortex methods.

18
Vortex Particles

One of the first vortex particle methods was proposed by Beale [104, 105], in which a

collection of disconnected particles are arranged to approximate a vortex sheet. The vec-

tor strength αp of particle p is initially assigned by the vorticity ωp and the volume vp

represented by the particle as

αp = vp ωp . (2.23)

Because the flow is incompressible, vp is constant in time, even though the shape of the

particle could deform. Similar to the vortex blob method, each particle serves as a basis

function of vorticity as
N
X
ω(x, t) = αp δ(x − xp ), (2.24)
p=1

where δ(x) is the 3-D Dirac delta function [106]. The strength of each particle changes

with time according to

dαp
= αp · ∇v, (2.25)
dt

and each particle moves with the local velocity.

Various implementations often differ by their approaches to the calculation of velocity,

the calculation of the velocity gradient tensor, the vorticity distributions (regularization

functions), and remeshing schemes [65, 106].

Vortex particles are not considered to be connected. This makes local remeshing tech-

niques difficult to implement, but allows advanced techniques for modeling viscous diffu-

sion [63, 106].

19
Vortex Filaments

In vortex filament methods, a vortex sheet is represented by a collection of filaments in

which vorticity is concentrated along one-dimensional curves and the direction of vorticity

is locally parallel to the tangent of the curve [107]. Vortex filaments are the natural gen-

eralization of the 2-D vortex blob method [65]. The space curve representing a filament

may be piecewise linear or have a higher order of continuity, but it is generally defined by

a set of Lagrangian markers. Each filament maintains a scalar value of circulation that is

constant along its length as well as constant with time.

One difficulty with vortex filaments is that they must be initialized according to the

topology of the vortex lines [107, page 70].

Sheet Elements

A vortex sheet can also be represented by sheet elements (i.e., panels). These sheet ele-

ments are often triangles, but quadrilaterals have also been used. Agishtein and Migdal

[108] represent a vortex sheet with flat triangles connected as a triangulated mesh. In their

method, vortex sheet strength is constant along each triangle and velocity is calculated

with an exact solution to the Biot–Savart law, which is modified to include a smoothing

parameter to provide apparent thickness to the sheet. The motion of the vortex sheet is then

described by the motion of the vertices of the triangulated mesh, which move as Lagrangian

markers (i.e., with the local fluid velocity). They re-triangulate the sheet at every time-step

with a new Delaunay triangulation.

Brady et al. [109] extended the work of Agishtein and Migdal. A scalar circulation

20
function is defined on the vortex sheet that is constant with time, and then the surface

vorticity (which they call “sheet vorticity”) is determined from spatial derivatives of circu-

lation. They use cubic Bézier triangular patches to provide C 1 continuity for the mesh and

circulation and C 0 continuity for the sheet vorticity. They calculate induced velocities by a

Gaussian quadrature approximation to the Biot–Savart integral, which is regularized with

a smoothing parameter.

Stock [63] devised a method for evolving regularized vortex sheets in which a vortex

sheet is represented by a triangular mesh with vortex sheet strength defined as constant

along each flat triangle. Each edge carries a scalar circulation value that is used to update

vorticity at each time-step to satisfy Equation 2.14. The triangular panels adapt to excessive

curvature and strain with schemes for edge splitting and adjacent-node merging. Each tri-

angle serves as a smoothed basis function for vorticity which is interpolated onto a regular

Cartesian grid, and then the Poisson equation

∇2 v = −∇ × ω (2.26)

is solved with a fast Poisson solver to obtain velocity at the triangle vertices (Lagrangian

markers).

Feng [64] uses triangular panels each with constant surface vorticity. The velocity

induced by a panel is calculated with a quadrature scheme that considers point vortices at

each triangle vertex. He implemented a tree-code based on that of Lindsay and Krasny

[110], which is based on Cartesian Taylor expansions of the Biot–Savart kernel. Citing

stability considerations, Feng includes a smoothing parameter in the Biot–Savart kernel to

regularize the vortex sheet. The method was further developed and applied to vortex rings

21
by Feng et al. [111].

Kandil et al. [112] devised an unsteady lifting surface method in which both the wing

and wake comprised panels with linearly varying surface vorticity. Though this is actually

published in the aeronautics literature, it is worth mentioning here as the only method we

are aware of that used wake panels with linearly varying surface vorticity. It was not de-

veloped further, probably because the computational requirements imposed an intractable

burden; a computation with 25 wing panels and 22 time steps took 10 minutes on a CYBER

175 computer.

Additional methods based on sheet elements include [72, 113–115]. As compared to

particle methods, the computational elements are fully connected. This connectivity pro-

vides for a simple solution to the stretch term in the vorticity transport equation but ob-

structs the generalization of sheet methods to rigorously model viscous diffusion [63].

2.3 Numerical Method for Regularized Vortex Sheet Evolution

In this section, we detail the equations and algorithmic approach comprising our method

for numerically simulating the evolution of a regularized vortex sheet.

A vortex sheet is represented by connected triangular panels that each carry a linear

distribution of surface vorticity. The vorticity transport equation is satisfied by maintaining

circulation along the edges of the triangulation as the vertices move with the local flow

velocity in a Lagrangian manner, which is similar to Stock [63, 67]. Induced velocities are

calculated directly from the Biot–Savart law to avoid computational grids used with Poisson

solvers, because that could become cumbersome for wakes that grow far downstream. At

each time-step: (1) surface vorticity is updated, (2) the velocity at each vertex is calculated,

22
(3) the vertices are displaced according to the chosen algorithm for solving ODEs, such

as the Runge–Kutta method, and (4) the triangular panels are locally refined to maintain

sufficient detail.

2.3.1 Discretization with Triangular Panels

The vortex sheet surface is discretized into flat, triangular, connected panels. The resulting

triangulation is a union of triangles such that the intersection of any two triangles is either

empty, a shared edge, or a shared vertex1 .

Each panel serves as a basis for a linearly varying vector function γ representing the

surface vorticity, which lies in the plane of the triangle. The surface vorticity is expressed

in a local coordinate system constructed from the three edge vectors belonging to a triangle.

For a triangle comprising vertices (P1 , P2 , P3 ) with corresponding locations p1 , p2 , and p3 ,

the edge vectors are defined as

r1 := p2 − p1 , r2 := p3 − p2 , r3 := p1 − p3 . (2.27)

Each triangle is assigned a basis (c1 , c2 , c3 ), with

r1 r1 × r2
c1 := , c2 := c3 × c1 , c3 := , (2.28)
||r1 || ||r1 × r2 ||

in which case c3 is the triangle normal vector. This notation is illustrated in Figure 2.3a. A

local Cartesian coordinate system may then be established on the triangle with coordinate

u along c1 , coordinate v along c2 , and the origin at p1 . Next, as illustrated in Figure

2.3b, the surface vorticity γ varies linearly along the triangle with two scalar functions
1
This definition is accompanied by a few more characteristics described in [116, page 31]

23
 γ3
c3 c3
v (u3 , v3 ) f 3 c1
c2 c2 f c g 3 c2
r3 P3 1 1
P3
(0, 0)
P1 c1 P1 γ1
r2 γ(u, v)
c1
u g 1 c2
r1 g 2 c2
(u2 , 0) γ2
P2 P2 f c
2 1

(a) Panel vertices, edges, and coordinates (b) Panel surface vorticity

Figure 2.3: Notation for the triangular panels.

f, g : R2 → R as

γ(u, v) = f (u, v)c1 + g(u, v)c2 , (2.29)

f (u, v) = a00 + a10 u + a01 v, (2.30)

g(u, v) = b00 + b10 u + b01 v. (2.31)

These coefficients are uniquely determined by the surface vorticity at each of the three

vertices γ1 , γ2 , γ3 according to the linear system

    
1 0 0  a00 b00  f1 g1 
    
    
  10 b10  = f2 g2  , (2.32)
1 u 0  a   
 2
    
    
1 u3 v3 a01 b01 f3 g3

where these vertex values are given by

u 2 = r 1 · c1 f 1 = γ 1 · c1 g1 = γ1 · c2

u3 = −r3 · c1 f 2 = γ 2 · c1 g2 = γ2 · c2

v3 = −r3 · c2 f 3 = γ 3 · c1 g3 = γ3 · c2 ,

24
as illustrated in Figures 2.3a and 2.3b. The solution to Equation 2.32 yields

a00 = f1 b00 = g1 (2.33)

f2 − f1 g2 − g1
a10 = b10 =
u2 u2
u3 − u2 u3 1 u3 − u2 u3 1
a01 = f1 − f2 + f3 b01 = g1 − g2 + g3 .
u2 v3 u2 v3 v3 u2 v3 u2 v3 v3

2.3.2 Vorticity Update via Half-Edge Circulations

The vortex sheet evolution equation governs the manner by which surface vorticity changes

over time—that is, updates. Unfortunately, this equation does not welcome a direct solu-

tion. An alternate approach to update surface vorticity at each time step is to leverage an

advantageous property of circulation in the Euler limit as viscosity vanishes [1, page 9],[63,

page 37],[14, page 31]:

DΓ
= 0. (2.34)
Dt

Then, by carefully selecting many distinct paths along which circulation must remain

constant, we can update surface vorticity at each time step by relating it to circulation.

Stock, who used triangular sheet elements each with constant surface vorticity, maintained

constant circulation along each edge of the triangulation [63, 67]. However, in our case,

three scalar circulation strengths governing a triangle leaves too many degrees of freedom

for γ to vary linearly along it. Instead, we consider the circulation along each half of each

edge.

We can apply Equation 2.19 to calculate the circulation around each half-edge, as illus-

trated in Figure 2.4. Without loss of generality, consider the first half of the edge from P1

25
 Γ3a P3
Γ2b
Γ3b
γ
P1 Γ2a
Γ1a
Γ1b P2

Figure 2.4: Half-edge circulations around a triangle.

to P2 with edge vector r1 . The differential path length is

dl = r1 ds, (2.35)

the surface vorticity varies linearly along the edge as

γ(s) = (1 − s)γ1 + sγ2

= (1 − s)(f1 c1 + g1 c2 ) + s(f2 c1 + g2 c2 ), (2.36)

and the normal vector n is the third element of basis (c1 , c2 , c3 ). Then, the circulation

along the first half-edge is

Z 1/2

Γ1a = (γ(s) × n) · r1 ds
0
Z 1/2

= [(1 − s)(f1 c1 + g1 c2 ) × n + s(f2 c1 + g2 c2 ) × n] · r1 ds

0
Z 1/2

= [(1 − s)(g1 r1 · c1 − f1 r1 · c2 ) + s(g2 r1 · c1 − f2 r1 · c2 )]ds

0

= 83 (g1 r1 · c1 − f1 r1 · c2 ) + 81 (g2 r1 · c1 − f2 r1 · c2 ), (2.37)

which is a linear expression in f1 , g1 , f2 , and g2 . All six half-edge circulations for a triangle

26
may be expressed in matrix form as
    
Γ1a  −3r1 · c2 −r1 · c2 0 3r1 · c1 r1 · c1 0  f1 
    
    
Γ   −r · c −3r · c 0 r1 · c1 3r1 · c1 0   f2 
 
 1b   1 2 1 2
    
    
Γ  
 2a  1  0 −3r2 · c2 −r2 · c2 0 3r2 · c1 r2 · c1   f3 
 
 =   
  8  
Γ 
 2b 

 0 −r2 · c2 −3r2 · c2 0 r2 · c1 3r2 · c1   g1 
 
    
    
Γ −r · c 0 −3r · c r · c 0 3r · c 1  g2 
    
 3a   3 2 3 2 3 1 3
    
    
Γ3b −3r3 · c2 0 −r3 · c2 3r3 · c1 0 r3 · c1 g3
(2.38)

Therefore, the following sets of six scalars are interchangeable for a given triangle: the

half-edge circulations Γ1a , . . . , Γ3b , the vertex surface vorticity scalars f1 , . . . , g3 , and the

multilinear polynomial coefficients a00 , . . . , b01 .

As the vortex sheet evolves, the surface vorticity for a given triangle at a particular

time-step is determined from the half-edge circulations, which are held constant, via Equa-

tion 2.38. This vorticity update scheme automatically satisfies the vortex sheet evolution

equation [67].

Validation of the Vorticity Update

To validate this approach for updating surface vorticity from the half-edge circulations, we

conduct two tests—the same tests introduced and used by Stock [63, page 63]. In these

tests, a vortex sheet is immersed in a prescribed and constant velocity field, and the sub-

sequent evolution of this vortex sheet is determined analytically (both position and surface

vorticity). The initial conditions also establish a discretized vortex sheet, the half-edge

circulations of which are calculated from the surface vorticity prescribed at each vertex

27
with Equation 2.38. The vertices of the discretized vortex sheet move exactly according to

the prescribed velocity field, and then the surface vorticity calculated from the half-edge

circulations may be compared to the analytic solution.

For both tests, a solenoidal, periodic, and steady velocity field is prescribed as

v(x, y, z) = cos(2πy) cos(2πz)a2 + sin(2πy) sin(2πz)a3 (2.39)

with orthonormal basis (a1 , a2 , a3 ) and corresponding Cartesian coordinates (x, y, z). A

planar vortex sheet is initiated at the z = 21 plane, where v is zero along a1 and a3 , which

implies the vortex sheet remains planar and strains along a2 . The solution to Equation 2.39

provides the position y at time t given the initial location y0 as

e4πs(t+c) − 1
 
1 1
y(y0 , t) = tan−1 + round(2y0 ), (2.40)
2π 2e2πs(t+c) 2

where

s = sgn (− cos(2πy0 )) , (2.41)

 
s
q
2
c= ln tan(2πy0 ) + tan (2πy0 ) + 1 , (2.42)
2π

and round(·) indicates rounding to the nearest integer. The spatial evolution of the discrete

vortex sheet is demonstrated in Figure 2.5, which depicts the initial triangulation as well as

the triangulation at t = 0.25.

The surface vorticity evolution equation is invoked to solve for the exact evolution of

surface vorticity with time, and it is restated here for convenience:

∂γ
= − v · ∇γ+ γ · ∇v− γ(∇ · P · v). (2.43)
∂t

28
 t=0 t = 0.25
1

0.75

0.5

y 0.25

-0.25

-0.5
0 0.25 0.5 0 0.25 0.5
x x

Figure 2.5: Initial and final triangulation of the vortex sheet in the z = 1/2 plane.

P is the tangential projection operator, which is a dyadic defined as

P := I − nn, (2.44)

where I is the identity dyadic, n is the sheet normal vector, and the juxtaposition of two

vectors indicates the dyadic product.

The variables are expanded in the working basis to

γ = γx a1 + γy a2 + γz a3 , (2.45)

v = vx a1 + vy a2 + vz a3 , (2.46)
∂ ∂ ∂
∇ = a1 + a2 + a3 , (2.47)
∂x ∂y ∂z

and, accordingly, the terms of Equation 2.43 are expanded to

v · ∇γ = (v · ∇γx )a1 + (v · ∇γy )a2 + (v · ∇γz )a3 , (2.48)

γ · ∇v = (γ · ∇vx )a1 + (γ · ∇vy )a2 + (γ · ∇vz )a3 , (2.49)

∇ · P · v = ∇ · v − [(n · ∇vx )a1 + (n · ∇vy )a2 + (n · ∇vz )a3 ] · n (2.50)

29
Convenient choices of γ along with n = a3 will lead to substantial simplification of the

evolution equation.

For the first test case, the initial surface vorticity γ0 is parallel to the direction of strain:

γ0 := a2 . (2.51)

This results in a constant surface vorticity:

γx = 0, γy = 1. (2.52)

In this simple case, the numerical scheme correctly matches this solution to machine pre-

cision.

For the second test case, the initial surface vorticity is transverse to the direction of

strain:

γ0 := a1 , (2.53)

which simplifies the evolution equation to

∂γx ∂γx ∂vy

= −vy − γx , (2.54)
∂t ∂y ∂y

with the solution

2 [1 + sin(2πy)] e2πt
γx (y, t) = . (2.55)
[1 + sin(2πy)]2 e4πt + cos2 (2πy)

Figure 2.6 depicts the surface vorticity predicted by the numerical update scheme as com-

pared to the analytic solution at t = 0.25. The error remains small compared to the initial

value even for this coarse discretization (16 vertices between peaks).

Therefore, we can conclude that the half-edge circulation method is effective at resolv-

ing a vortex sheet in the presence of planar strain in directions parallel or transverse to

30
 5
analytic
4 numeric
error
3 initial
γx
2

0
-0.5 -0.25 0 0.25 0.5 0.75 1
y

Figure 2.6: Comparison of surface vorticity at t = 0.25.

vorticity. As noted by Stock, the vortex blob method requires remeshing of the particles in

both of these cases and the vortex filament method requires remeshing when the strain is

transverse to vorticity [63, page 67].

2.3.3 Smoothing

As mentioned previously, vortex sheets have a tendency to form singularities in the center of

a rolled-up region. Additionally, numerical imprecision tends to provoke instabilities [84].

These features may be limited by regularizing the vortex sheet, which we implement by

including a smoothing parameter in the Biot–Savart integral [64, 90]. Smoothing provides

the vortex sheet an apparent thickness that more closely resembles a high Reynolds number

flow rather than an inviscid one, which is desirable [117, 118].

We employ the smoothing parameter introduced by Anderson [91] and Krasny [90] so

as to maintain the same method of calculating the velocity induced by the vortex sheet (the

31
Biot–Savart law), which is now expressed as

1 (x0 − x) × γ(x0 )
Z
v(x) = dS(x0 ). (2.56)
4π S (||x0 − x||2 + δ 2 )3/2

The effect of the smoothing parameter is demonstrated in Figure 2.7, which depicts the

velocity field induced by a triangular vortex sheet element for two cases: δ = 0 and δ = 0.2.

The two induced velocity fields are sampled along the z-axis, which is normal to the trian-

gle and pierces it at the black marker. The δ = 0 case, which lacks smoothing, is discontin-

uous at this intersection; however, it remains bounded, which would not be the case for a

filament or particle. In the smoothed case, δ = 0.2, the induced velocity exhibits a smooth

transition across the triangle and asymptotically approaches the δ = 0 case away from the

triangle. Note that for this particular surface vorticity field, the induced velocities include

a component normal to the triangle.

z
v
δ = 0.2

δ=0

Figure 2.7: Example of the velocity field induced by a vortex sheet with smoothing and
without smoothing.

The smoothed vorticity generated by this vortex sheet element can be expressed as the

curl of Equation 2.16. Larger values of δ spreads the vorticity outward from the panel, and

this limits the peak vorticity reached at the center of a rolled-up vortex core.

32
2.3.4 Adaptive Paneling

In Section 2.3.2 we demonstrated that maintaining constant half-edge circulation passively

satisfies Equation 2.14 (the vortex sheet evolution equation) in the presence of planar strain.

However, no such guarantee exists in the presence of large curvature, such as in a rolled-up

region, and so the resolution will become insufficient to adequately resolve the vorticity

[110]. To avoid this eventuality, it is necessary to adaptively refine the resolution of the

triangulation, which we accomplish by splitting long edges, removing short edges, and

flipping selected edges to satisfy the Delaunay condition [119].

First, any edge longer than a specified threshold is split into two edges, and a new

vertex is placed at the geometric midpoint as demonstrated in Figure 2.8. We considered a

spline-based point insertion technique, but we found it to be unstable, as did Kaganovskiy

[113].

Figure 2.8: Example of triangle splitting.

Second, any edge shorter than a specified threshold is removed, thereby also remov-

ing the two triangles that share it. This step serves the opposite purpose of the previous

step—to remove detail where it is not required and save computational cost. The two ver-

tices comprising the removed edge are merged and located at their geometric midpoint, as

demonstrated in Figure 2.9.

Lastly, we enforce the Delaunay condition on each edge, which stipulates that the pair

33
 Figure 2.9: Example of edge removal.

of angles opposite an edge should sum to less than or equal to 180◦ . Any edge not meeting

this condition is flipped as shown in Figure 2.10. This step helps maintain a more regular

triangulation where the degree2 of each vertex is close to six.

Figure 2.10: Example of Delaunay edge flipping.

The Delaunay edge-flipping procedure we implement is novel, whereas the first two

procedures are common to other methods based on triangulations. An edge is prevented

from being split, removed, or flipped if the degree of a neighboring vertex would exceed

the minimum or maximum limits. Also, edges on the boundary of a sheet are protected

from removal to prevent the boundary from degenerating.

To assess this adaptive paneling algorithm, we again leverage a test introduced by Stock

[63]. A sphere of radius 0.4 is immersed in a prescribed and constant velocity field, which is

constructed by placing three infinite vortex filaments with a smoothed Biot–Savart kernel.

The filaments are each parallel to one of the basis vectors a1 , a2 , a3 , respectively, and they
2
The degree of a vertex is the number of edges at the vertex, or, equivalently, the number of neighboring
vertices [120].

34
pass through

p1 = 0.2a2 + 0.2a3 , (2.57)

p2 = 0.2a1 + 0.2a3 , (2.58)

p3 = 0.2a1 + 0.2a2 , (2.59)

each with strength Γ = 4π. The resulting velocity at a point p is

3
X ri × ai
v(p) = (2.60)
i=1
ri · ri + δ 2

ri := (pi − p) − (pi − p) · ai ai (2.61)

The test was conducted with δ = 0.1 and ∆t = 0.001, and it was time-stepped with

the fourth-order Runge–Kutta method; the triangulation began with 20,480 triangles and

ended with 111,586 triangles after 1000 iterations. Figure 2.11 illustrates the shape of the

sheet as it evolved.

The metric for assessing the accuracy of the algorithm is the volume enclosed by the

triangulation, which should be conserved because the velocity field is solenoidal. The re-

sults are presented in Figure 2.12. The histogram of edge lengths suggests that the edge

splitting and edge removal schemes successfully maintained the 0.005 and 0.02 threshold

values chosen for this test. Similarly, the histogram of vertex degrees indicates that the

edge flipping scheme kept the degree of each vertex near six, which corresponds to a tri-

angulation with mostly regular triangles. Finally, the ratio of the volume enclosed by the

triangulation to the exact value remained very close to unity.

To visualize the complex geometric forms attained by a vortex sheet, we render it as

35
Figure 2.11: The triangulated sheet as it evolves at iterations 0, 100, and 1000. Also
indicated are the infinite vortex filaments.

Figure 2.12: The results of the adaptive paneling algorithm test, clockwise from top left: the
final triangulation, the computed volume at each iteration normalized by the exact volume
of a sphere, a histogram of the edge lengths, and a histogram of the vertex degrees.

a shadowgraph, as originally inspired by the captivating figures presented by Stock [63].

Borrowing an analogy from Schlieren photography, a vortex sheet can be imagined to affect

light in a similar manner as a shockwave, which is approximately a thin surface. Each

36
panel is treated as semitransparent plate that obstructs light in proportion to the distance it

travels through the plate. Therefore, the image appears darker in regions where the sheet

is perpendicular to the viewing direction. The shadowgraph of the final configuration of

the surface in this adaptive paneling test is illustrated in Figure 2.13. In this example, the

sheet did not carry vorticity and was thus not a vortex sheet; however, this Schlieren-like

technique will be used throughout this thesis to visualize vortex sheets.

Figure 2.13: Shadowgraph of the deformed sheet at the end of the adaptive paneling test.

2.3.5 Fast Summation

The calculation of the induced velocities comprises nearly all of the computational effort

spent while time-stepping a vortex sheet. For a triangulation with M triangles and N

vertices, the number of computations scales as O(M N ), and a typical triangulation will

have approximately twice as many triangles as vertices. Fast calculations will therefore be

37
essential for complex multi-body problems with interacting wakes

The computational efficiency is improved in three ways. First, a triangle is approxi-

mated as a point vortex of equal strength when the triangle and point of velocity computa-

tion are sufficiently distant. The contribution of a triangle T to the induced velocity at x0 is

described by the regularized Biot–Savart law, which is then approximated as

1 (x0 − x) × γ(x0 )
Z
v(x) = 3/2
dS(x0 )
4π T (||x − x|| + δ )
0 2 2

1 r
Z
≈ × γ(x0 )dS(x0 ), (2.62)
4π (r · r + δ 2 )3/2 T

where r = pC − x, and pC is the centroid of the triangle. This simplified integral provides

the strength of the particle approximating the surface vorticity carried by the triangle:

Z
α := γ(x0 )dS(x0 ), (2.63)
T

which is solved in terms of the polynomial coefficients presented in Section 2.3.1 as

3
1 1 1

α= 2
u 2 v3 a 00 + 6
(u 2 + u 3 )u 2 v3 a 10 + 6
u 2 v3 a 01 c1

3
1 1 1

+ 2
u 2 v3 b 00 + 6
(u2 + u 3 )u 2 v 3 b 10 + 6
u2 v 3 b 01 c2 . (2.64)

Then, for a set of M vortex particles—approximating triangular panels—the induced ve-

locity at the ith vertex is

M
1 X rij × αj
vi = , rij = pC,j − pi . (2.65)
4π j=1 (rij · rij + δ 2 )3/2

Equation 2.65 requires many fewer floating point operations than the exact solution for

a triangular panel provided in Appendix A. However, the computational expense of this

direct summation will still scale as O(N 2 ) and remains prohibitively expensive for large

38
problems. Therefore, to further improve computation efficiency we invoke a vortex particle

treecode. A treecode provides a summation that scales as O(N log N ) by representing the

influence of a cluster of particles on a far away point with an approximation to a specified

tolerance. Along with the related fast multipole method, treecodes are common in N -body

problems seen electrostatics, celestial mechanics, and vortex methods [97, 110, 121–129].

The celestial mechanics problem [121], in which a set of point masses experience a

mutual gravitational pull described by Newton’s law of gravitation, provides a convenient

and intuitive example: the influence of a dense cluster of point masses on a distant point can

be approximated by replacing the cluster by a single point mass with the same total mass

placed at the mass center of the cluster. Better approximations are produced by including

higher-order mass moment of inertia terms. The fast multipole method goes further by

including approximations for cluster–cluster interactions.

In this work, we construct the tree in the same manner as Barnes and Hut [122], with

the exception that we allow multiple particles per cell instead of only one. We leverage the

work of Lindsay [128] to compute the cell moments as a Cartesian Taylor series expan-

sion of Equation 2.65. The details of the implementation and testing of the treecode are

presented in Appendix B.

Lastly, to improve computational efficiency even further, we implemented the treecode

in parallel on a graphics processing unit (GPU) in the CUDA programing language, which

is a recent advance in the vortex methods literature [46, 67, 125, 130, 131]. The large

number of processors on a GPU, while individually slower than a CPU, operate in parallel,

making a GPU well suited to compute these N -body interactions.

To demonstrate these efficiency improvements, the total computational time was mea-

39
sured for several summing approaches for a varying number of panels for a structured

toroidal mesh. The results depicted in Figure 2.14 compare CPU vs GPU, direct sum-

mation vs treecode, and single precision vs double precision. The gray lines provide a

reference to the canonical growth rates of O(N 2 ) and O(N log N ), and the data reveal that

each summation algorithm asymptotically approaches the expected growth rate.

This example, as well as all of the numerical simulations in this thesis, was conducted

on a 3.5 GHz Intel Core i7-3770K and a GeForce GTX TITAN Z with 5760 CUDA cores

at 705 MHz clock speed, and the algorithms were written in C/C++ and CUDA.

103
CPU,double,full sum
N2
CPU,double,treecode
GPU,double,full sum
2 GPU,double,treecode
10
GPU,single,full sum
GPU,single,treecode

101
computational time (s)

0
10

10-1
N log N

10-2

10-3
103 104 105 106
number of panels

Figure 2.14: Comparison of computational time for several approaches to evaluating Equa-
tion 2.65.

40
2.4 Evolution of a Sphere in a Uniform Freestream

To demonstrate the efficacy of the method presented in this chapter for simulating the evo-

lution of a free vortex sheet, we conduct the following typical example [63, 72, 103, 127,

132]. Consider a spherical vortex sheet of unit radius immersed in an otherwise incom-

pressible uniform freestream of unit velocity. The surface vorticity initially establishes

potential flow around the sphere and then evolves according to its own influence starting at

time t = 0. This simulation began 20,480 triangles and ended with 408,590 triangles after

1000 time-steps with ∆t = 0.01, δ = 0.1, and using the fourth-order Runge–Kutta method.

The simulated evolution of the spherical vortex sheet is presented in Figure 2.15. The

left column contains shadowgraphs of the vortex sheet (the type of imaging introduced via

Figure 2.13). These images demonstrate the sheet’s continuous growth in surface area and

complexity with time. The majority of the sheet rolls into a ring but some escapes down-

stream. The right column contains radiographs, which depict the vorticity carried by the

vortex sheet. This imaging technique was also inspired by Stock [46], and is constructed

by computing the vorticity ω on a volume grid throughout the flow and then correlating

||ω|| to a grayscale, similar to the way in which dense bone obstructs radiation in a medical

radiograph. Recall that the smoothing parameter δ has the effect of spreading the vorticity

away from the vortex sheet. Most of the vorticity can be seen to concentrate in a toroidal

ring, with only a small amount escaping downstream. Both sets of images reveal the in-

stabilities natural to vortex sheets; these particular instabilities can be attributed to small

perturbations arising from numerical roundoff error, as discussed by Krasny [84].

The accuracy of this simulation may be assessed by tracking the invariant (conserved)

41
properties of the flow [1, 109]. In this case we examine volume and circulation. The volume

is calculated directly from the triangulation, as was done in Section 2.3.4. The circulation

is calculated with Equation 2.17 by integrating velocity along the closed path illustrated

in Figure 2.16, which extends to ±∞ along the x-axis (the freestream direction) and then

connects via a circular arc. In the limit as this path extends to infinity, the velocity along the

circular arc becomes v∞ , and so only the velocity along the straight segment contributes

to the integral. The exact solutions are V ∗ = 4π/3 and Γ∗ = 3 for volume and circulation,

respectively. These metrics are displayed in Figure 2.17a; there is a slight initial error in

volume which then grows but remains within about 2%, while circulation error remains

within about 3%. The short period noise observed in the circulation history is due to the

treecode, which can produce small but abrupt jumps in velocity as slight changes in vertex

location cause topological changes in the tree. Figure 2.17b demonstrates the growth in the

number of triangles over time, which is approximately quadratic.

42
 0 shadowgraphs radiographs

100

200

300

400

600

800

1000

Figure 2.15: The evolution of a spherical vortex sheet over 1000 iterations. The shadow-
graphs reveal the shape of the sheet while the radiographs reveal the vorticity carried by
the sheet.

43
 contour of
integration
y
vortex
sheet

−∞ x ∞

Figure 2.16: Contour of integration for calculating the circulation.

1 5
V /V ∗
0.995 4
0.99
NT × 105

3
0.985
2
0.98
Γ/Γ∗ 1
0.975

0.97 0
0 2 4 6 8 10 0 2 4 6 8 10
t t
(a) Volume and circulation error ratios. (b) Number of panels in the sheet triangulation.

Figure 2.17: The time history of the simulated evolution of a spherical vortex sheet.

44
 CHAPTER 3

BOUND VORTEX SHEETS

The numerical calculation of potential flow around three-dimensional arbitrary bodies orig-

inated with the work of Smith, Hess, and others at the Douglas Aircraft Company [133–

139]. They conceived the panel method in which the body is represented with a source sin-

gularity distribution on the surface, leveraging classical potential theory [136, 140]. Lifting

bodies are modeled by including doublet distributions both on the body surface and in a

shed wake. The singularity distributions are discretized with elementary basis functions

defined on panels, thereby enabling the scalar potential function to be calculated. The

resulting set of linear algebraic equations is solved numerically to yield the source and

doublet distribution strengths [14, page 241]. Variations of the panel method have been

developed into several successful computer programs including VSAERO [141], PANAIR

[142, 143], and PMARC [13].

An alternative approach to a source–doublet formulation is to represent a body with a

surface vortex sheet and to solve directly for the surface vorticity. Martensen [144] de-

veloped this approach into the surface vorticity method for two-dimensional bodies, also

leveraging classical potential theory [140] as well as analytical work by Prager [145]. The

surface vorticity method was further developed and applied to airfoil cascades and annular

airfoils for use in turbomachinary flow problems [146–150].

Lewis and Ryan [149] introduced a general analytical form of the surface vorticity

method in three dimensions. They took special note of the restriction that surface vorticity

45
must have zero divergence and called this restriction a “continuity equation for the vortex

sheet.” This constraint poses a unique challenge to the surface vorticity method in three

dimensions that is not present in the source panel method [151]. For general body shapes,

the most advanced implementations of surface vorticity methods have been the ring vortex

and horseshoe vortex methods [150], variations of which are presented by Katz and Plotkin

[14]. Unfortunately, these low order approaches prevent the surface vorticity method from

competing with higher-order panel methods for three-dimensional flow in terms of accu-

racy. A high order surface vorticity method for three-dimensional flow has remained an

elusive challenge in part because of the divergence obstacle.

In this chapter, we develop a new higher-order surface vorticity panel method for three-

dimensional bodies. The approach is to determine the strength of the vortex sheet on

the body surface such that the surface velocity is tangential and the surface vorticity is

divergence-free. The only modifications for lifting flows are the shedding of a wake, which

is also modeled as a vortex sheet, and the enforcement of the unsteady Kutta condition.

The equations are discretized via a surface triangulation with linear variation of surface

vorticity along each triangle. This discretization provides a fidelity analogous to second

order source–doublet panel methods. The discretized problem formulation produces an

overdetermined linear system which is solved approximately by minimizing the sum of

the squares of the residuals. We assess the convergence of the method by comparing it to

the analytical (exact) solution to the potential flow over an ellipsoid. Some of the equa-

tions in this section were also presented in Chapter 2, but they are repeated to provide an

uninterrupted derivation.

46
3.1 Governing Equations for the Surface Vorticity Model

As described by Batchelor [69, page 147],[14, page 13], the motion of an incompressible,

inviscid fluid is governed by the Euler momentum equation

∂v 1
+ v · ∇v = − ∇p (3.1)
∂t ρ

and the continuity equation

∇ · v = 0. (3.2)

The Helmholtz decomposition stipulates that any vector field may be expressed as the sum

of a solenoidal vector field and an irrotational vector field [69, page 84],[63, page 4]. Then,

for any general velocity field v, we may specify

v = vω + vφ (3.3)

with the solenoidal vector field vω satisfying

∇ × vω = ω, ∇ · vω = 0, (3.4)

and the irrotational vector field vφ satisfying

∇ × vφ = 0, ∇ · vφ = σ, (3.5)

where σ is the local volumetric rate of expansion and ω is the local vorticity. The irrota-

tional field can be obtained from a scalar function φ as vφ = ∇φ.

When combined, the equations governing vω produce a Poisson equation

∇2 vω = −∇ × ω (3.6)

47
which can be solved using the infinite-medium Green’s function [63, page 4],[65, page

307] as
1 (x0 − x) × ω(x0 )
Z
vω (x) = dV (x0 ), (3.7)
4π ||x0 − x||3

which is often called the Biot–Savart law. Equation 3.3 may now be rewritten as

1 (x0 − x) × ω(x0 )
Z
v(x) = dV (x0 ) + vφ , (3.8)
4π ||x0 − x||3

which is guaranteed to satisfy Equation 3.2 for any vorticity field

ω = ∇ × v. (3.9)

3.1.1 Bound Vortex Sheet Definition and Continuous Problem Formulation

A vortex sheet is represented by an orientable surface S paired with a vector field γ rep-

resenting the vortex sheet strength, which we call surface vorticity, existing in the local

tangent space of S. As described by Stock [67] and Sohn [71], the surface vorticity is

defined as

γ := n × (v + − v − ), (3.10)

where n is the unit vector normal to S oriented outward, v + is the velocity immediately

above (or outside) S, and v − is the velocity on the opposite side. The surface vorticity

is related to vorticity via the Dirac delta function as ω = γδ(n), where n is the distance

normal to the sheet. Equation 3.8 may then be restated in terms of surface vorticity as

1 (x0 − x) × γ(x0 )
Z
v(x) = 0 3
dS(x0 ) + vφ . (3.11)
4π ||x − x||

Consider a solid body with boundary surface S in a high Reynolds number flow. In

48
this case, vorticity is confined to a thin region around S referred to as the boundary layer

[69, 152]. In the limit as Reynolds number tends to infinity (the Euler limit), the boundary

layer collapses to an infinitely thin shear layer along S, which is equivalent to a vortex

sheet [153, page 194] with strength γ(x), x ∈ S. The velocity jumps from v − at the

surface to v + immediately outside, and it follows directly from the main assumption that

v + is irrotational. Thus, our goal is to devise a set of equations in terms of γ such that the

solution provides a valid vortex sheet bound to the surface S.

For non-lifting flow, there are two requirements for a valid bound vortex sheet. The first

requirement comes from the no-slip condition, v − = 0 (relative to the body), which persists

in the Euler limit. If S is closed and simply connected, it will suffice to enforce this only

in the normal direction. Then, we require n · v = 0, where the “sheet” designation is not

needed because n · v − = n · vsheet = n · v + .

The second requirement for a valid bound vortex sheet is that the velocity resulting

from the surface vorticity is irrotational. From a vector calculus identity, on S we have

∇ · γ = ∇ · (n × v + )

= v + · (∇ × n) − n · (∇ × v + )

= v+ · 0 − n · 0

= 0.

Thus, for velocity to be irrotational, surface vorticity must be solenoidal.

The problem formulation to determine the incompressible, infinite Reynolds number

flow around a solid body with closed boundary surface S moving with velocity vS and

49
immersed in a uniform freestream v∞ can now be stated:

Determine γ on S such that (3.12)

v(x) · n(x) = vS (x) · n(x), x ∈ S (3.13)

∇ · γ(x) = 0, x ∈ S (3.14)
1 (x0 − x) × γ(x0 )
Z
v(x) = v∞ + 0 3
dS(x0 ) (3.15)
4π S ||x − x||

This is the continuous boundary integral form of the problem formulation, and the corre-

sponding discrete form, which is the actual numerical problem that is solved, is presented

subsequently.

3.1.2 Unsteady Pressures

The problem formulation does not invoke Equation 3.1, the Euler momentum equation,

which is consistent with traditional potential flow problem formulations based on Laplace’s

equation [136, page 6],[14, page 33]. Instead, we use Equation 3.1 in the form of the

Bernoulli equation to determine pressure from velocity. The following derivation of the

Bernoulli equation for unsteady flow is summarized from the presentation by Katz and

Plotkin [14, page 28]. Equation 3.1 is reformulated to include vorticity as

∂v 1
− v × ω + 12 ∇(v · v) = − ∇p. (3.16)
∂t ρ

This equation will be applied on the body immediately outside the boundary layer, where

v is assumed to be irrotational, which implies v × ω = 0; the v + notation is dropped for

convenience. The time derivative of velocity is converted to a time derivative of a scalar

50
potential function Φ defined by
Z P
Φ := v · dl (3.17)
P0

with an arbitrary reference point P0 . This line integral is path independent because v is a

conservative vector field. Substituting

∂v ∂Φ
=∇ (3.18)
∂t ∂t

and factoring out the gradient operator from Equation 3.16 produces

 
p 1 ∂Φ
∇ + 2v · v + = 0. (3.19)
ρ ∂t

This implies the scalar function inside the parentheses varies only with time, giving

p 1 ∂Φ
+ 2v · v + = C(t). (3.20)
ρ ∂t

Finally, the pressure coefficient may be formulated in reference to freestream pressure and

velocity as
p − p∞ v·v 2 ∂Φ
Cp := 1 2
= 1 − 2
− 2 ∂t
. (3.21)
2
ρv ∞ v∞ v∞

While the solution to the problem formulation provides surface vorticity, velocity can

be recovered from surface vorticity by the vector triple product

γ × n = (n × v) × n

= −(n · v)n + (n · n)v

= v.

This result demonstrates one of the incentives for formulating the flow solution in terms of

surface vorticity: velocity on the surface is immediately obtained from the solution and to

51
the same order of accuracy.

3.1.3 Lifting Flow and the Unsteady Kutta Condition

Lifting flow is characterized by the inclusion of a wake emitted from a separation path on

the body surface. Following the logic presented previously in Section 3.1.1, the wake is

concentrated into an infinitely thin shear layer and is also represented by a vortex sheet.

The body surface, renamed as Sb , is topologically split along the separation path (e.g., the

trailing edge) so as to admit two distinct values of surface vorticity, γu and γl , at any point

along the path.

This topological description is illustrated in Figure 3.1, where we note that the wing

is not physically open at the trailing edge. In the geometric description the black line

represents the separation path, and in the topological description the black lines represent

the shared boundaries of Sb and Sw .

geometric description

body
wake

separation path

topological description
upper
Sb
Sw
shed
lower

Figure 3.1: Cross section of a wing illustrating the interpretation of the trailing edge.

The introduction of a separation path produces a boundary for the bound vortex sheet,

which requires the inclusion of a boundary condition in the problem formulation. The Kutta

52
condition is imposed when the separation path is a sharp (i.e., wedge-shaped) trailing edge.

The Kutta condition stipulates that the flow leaves the trailing edge smoothly with a finite

velocity [154],[69, page 436],[14, page 103].

Hess describes three distinct applications of the Kutta condition in a 3-D boundary

element method, two of which are relevant to our surface vorticity method [155, page 300].

Application (a): “a stream surface of the flow leaves the trailing edge with a direction that

is known, or at least can be approximated.” This requires a no-normal-flow constraint at

a specified point aft of the trailing edge, typically along the trailing-edge bisector. The

selection of this distance aft requires special attention. Application (b): “as the trailing

edge is approached, the surface pressures (velocity magnitudes) on the upper and lower

surfaces have a common limit.” This requirement on surface pressure is applied at the

control points nearest the trailing edge and is thus subject to the paneling resolution near

the trailing edge. Hess concludes from a numerical experiment that the pressure equality

condition (b) is more accurate and robust than the flow tangency condition (a). However,

condition (a) produces linear equations while condition (b) produces quadratic equations.

In this thesis, we implement the pressure equality condition because it will not add

significant computational expense. At a point P on the trailing edge, consider point Pu

assigned to the upper surface and point Pl assigned to the lower surface, both collocated

53
with P . The unsteady pressure difference between Pu and Pl is

v(Pu ) · v(Pu )
 
2 ∂Φ(Pu )
Cp (Pu ) − Cp (Pl ) = 1 − 2
− 2
v∞ v∞ ∂t
v(Pl ) · v(Pl )
 
2 ∂Φ(Pl )
− 1− 2
− 2
v∞ v∞ ∂t
v(Pl ) · v(Pl ) − v(Pu ) · v(Pu )
= 2
v∞
2 ∂
− 2 [Φ(Pu ) − Φ(Pl )] . (3.22)
v∞ ∂t

The difference of potentials can be reformulated as the circulation around the wing. Con-

struct paths Cu and Cl from reference point P0 to Pu and Pl , respectively, as illustrated in

Figure 3.2. The path integrals for potential combine to give the closed path integral for

circulation as

Z Z
Φ(Pu ) − Φ(Pl ) = v · dl − v · dl
Cu Cl
I
= v · dl
Cl→u

= Γ(P ). (3.23)

In this example, paths Cu and Cl were purposefully constructed to combine to produce

Cu
Cl→u P
Cl
P0

Figure 3.2: Cross section of a wing illustrating paths along the surface to a point on the
trailing edge.

path Cl→u for which Γ is intuitive, but any arbitrary paths would suffice because v is a

conservative vector field on the body surface.

54
 Therefore, the unsteady Kutta condition requiring pressure equality at the trailing edge

simplifies to

v(Pl ) · v(Pl ) − v(Pu ) · v(Pu ) 2 ∂

0 = Cp (Pu ) − Cp (Pl ) = 2
− 2 Γ(P ). (3.24)
v∞ v∞ ∂t

The application of the unsteady Kutta condition to the problem formulation is discussed in

Chapter 4, in which the bound vortex sheets and free vortex sheets are combined.

3.2 Triangular Panel Representation

To discretize the continuous problem formulation into a numerical problem formulation

that can be solved by a computer, the solid body surface Sb is triangulated into surface Sb4

composed of flat triangular panels. These panels are then used for the subsequent calcu-

lations of influence velocities, circulation, and surface pressure. For a triangle comprising

vertices (P1 , P2 , P3 ) with vertex locations p1 , p2 , and p3 , the edge vectors are defined as

r1 := p2 − p1 , r2 := p3 − p2 , r3 := p1 − p3 . (3.25)

Each triangle is assigned a basis (c1 , c2 , c3 ), with

r1 r1 × r2
c1 := , c2 := c3 × c1 , c3 := , (3.26)
||r1 || ||r1 × r2 ||

in which case c3 is the triangle normal vector. This notation is illustrated in Figure 3.3a.

A local coordinate system may then be established on the triangle with coordinate u

along c1 , coordinate v along c2 , and the origin at p1 . Next, as illustrated in Figure 3.3b, the

surface vorticity γ is confined to vary linearly along the triangle with two scalar functions

55
 γ3
c3 c3
v (u3 , v3 ) f 3 c1
c2 c2 f c g 3 c2
r3 P3 1 1
P3
(0, 0)
P1 c1 P1 γ1
r2 γ(u, v)
c1
u g 1 c2
r1 g 2 c2
(u2 , 0) γ2
P2 P2 f c
2 1

(a) Panel edges and coordinate system (b) Panel surface vorticity

Figure 3.3: Notation for the triangular panels.

f, g : R2 → R as

γ(u, v) = f (u, v)c1 + g(u, v)c2 , (3.27)

f (u, v) = a00 + a10 u + a01 v, (3.28)

g(u, v) = b00 + b10 u + b01 v. (3.29)

These coefficients are uniquely determined by the surface vorticity at each of the three

vertices γ1 , γ2 , γ3 according to the linear system

    
1 0 0  a00 b00  f1 g1 
    
    
  10 b10  = f2 g2  , (3.30)
1 u 0  a   
 2
    
    
1 u3 v3 a01 b01 f3 g3

where these terms are given by

u 2 = r 1 · c1 f 1 = γ 1 · c1 g1 = γ1 · c2 (3.31)

u3 = −r3 · c1 f 2 = γ 2 · c1 g2 = γ2 · c2

v3 = −r3 · c2 f 3 = γ 3 · c1 g3 = γ3 · c2 ,

56
as illustrated in figures 3.3a and 3.3b. The solution to Equation 3.30 yields

a00 = f1 b00 = g1 (3.32)

f2 − f1 g2 − g1
a10 = b10 =
u2 u2
u3 − u2 u3 1 u3 − u2 u3 1
a01 = f1 − f2 + f3 b01 = g1 − g2 + g3 .
u2 v3 u2 v3 v3 u2 v3 u2 v3 v3

Next, each vertex is assigned a normal vector d3 , which may be calculated as the aver-

age of the normal vectors from neighboring triangles, or supplied as exact from analytical

calculations on Sb . A basis (d1 , d2 , d3 ) may then be formed. As stated previously, the sur-

face vorticity on a triangle is determined by the strength at its vertices γ1 , γ2 , γ3 . However,

when a vertex normal vector d3 is not parallel to normal vector c3 of a neighboring triangle,

these strengths will likely not be in the plane of the triangle, as demonstrated in Figure 3.4.

Therefore, we must decide on a transformation to convert vector γd in the (d1 , d2 ) plane to

a vector γc in the (c1 , c2 ) plane. The vector projection

γc = γd − (γd · c3 )c3 (3.33)

is an obvious choice, which, unfortunately, leads to inaccurate calculations because the

projection does not preserve the magnitude of γd . Alternatively, a vector rotation does

preserve the magnitude of γd ; rotating a vector around an edge is analogous to the parallel

transport of a vector along a smooth curve.

To rotate a vector γd in the (d1 , d2 ) plane onto the (c1 , c2 ) plane, construct q := (d3 ×

c3 )/||d3 × c3 || and θ := sin−1 ||d3 × c3 ||, as depicted in Figure 3.4. Then, γc is obtained

57
 d3 c3 vertex tanget
plane
triangle plane d2
θ q
c2

θ γd
γc
d1
c1
line of
intersection

Figure 3.4: The vector γd in the vertex plane with normal vector d3 must be related to
some vector γc in the triangle plane with normal vector c3 . All vectors outside of the
(c1 , c2 ) plane are accompanied by thin gray lines to indicate their height (except c3 ).

by rotating γd about q through the angle θ via Rodrigues’ rotation formula

γc = γd cos θ + (q × γd ) sin θ + q(q · γd )(1 − cos θ), (3.34)

which is linear in γd .

3.2.1 Panel Perimeter Circulation Constraint

The linear representation of surface vorticity expressed in Equation 3.27 provides a straight-

forward expression to discretize the zero-divergence constraint

∂f (u, v) ∂g(u, v)
∇·γ = +
∂u ∂v
= a10 + b01 (3.35)

However, we found that better solutions are produced by directly requiring the surface

velocity to be irrotational, as discussed in Section 3.1.1. Because v is irrotational, it must

be conservative on any simply connected subset of the domain, i.e. the circulation must be

zero:
I
Γ= v · dl = 0. (3.36)

58
The perimeter of each panel provides a convenient set of loops to implement this constraint,

which we call the panel perimeter circulation constraint.

3.3 Numerical Problem Formulation

The surface triangulation Sb4 comprises NV vertices and NT triangles; a typical triangu-

lation will have NT ≈ 2NV . Each vertex possesses two degrees of freedom defining the

local surface vorticity, and the resulting total number of degrees of freedom is Ndof = 2NV .

The surface vorticity at the ith vertex is expressed via the two scalars fi and gi and the local

basis vectors d1,i and d2,i as γi = fi d1,i + gi d2,i . The scalars fi and gi , i = 1, . . . , NV are

arranged into column matrices f~ and ~g and are then concatenated as

 
f~
~x := 
 
 (3.37)
~g

to form one complete list of unknowns. The notation ~· indicates a column matrix (or

column vector) and is not to be confused with physical vectors in R3 indicated by bold

characters.

Both divergence and induced velocity are calculated on a triangular panel in terms of

the surface vorticity at its three vertices, which are given by f1 , f2 , f3 , g1 , g2 , g3 , as indicated

in Figure 3.3b. Therefore, we need access to f~1 , f~2 , etc., which requires the introduction

of matrices Af1 x , Af2 x , etc., the elements of which are populated by applying Equation

3.34 to the basis vectors of every pair of associated vertices and triangles. For example,

f~1 = Af1 x~
x, (3.38)

59
which means Af1 x is a sparse matrix of size NT × Ndof ; together, f~1 and ~g1 are the com-

ponents of the surface vorticity at the first vertex of every triangle.

3.3.1 Normal Flow Constraint in Matrix Form

The normal flow constraint (Equation 3.13) is applied at the centroid of each triangle

pc = (p1 + p2 + p3 )/3, and the velocity normal to the surface at the centroid of the ith

triangle is expressed as vc,i = vc,i · c3,i . The change in induced velocity at pc on the ith tri-

angle due to a change in f1 on the jth triangle is expressed as ∂vc,i /∂f1,j and is calculated

from the solution to Equation 3.15 presented in Appendix A. These partial derivatives are

arranged in a Jacobian matrix

 
∂vc,1 ∂vc,1
· c3,1
··· · c3,1
∂f1,1 ∂f1,NT
 
 
∂~vc  .. .. ..

= , (3.39)
 
~ . . .
∂ f1 



 ∂vc,NT ∂vc,NT 
· c3,NT · · · · c3,NT
∂f1,1 ∂f1,NT

and likewise for ∂~vc /∂ f~2 , ∂~vc /∂ f~3 , ∂~vc /∂~g1 , etc. Then, the matrix Anf is computed as

∂~vc ∂~vc ∂~vc

Anf = A f1 x + Af2 x + ··· + Ag x (3.40)
∂ f~1 ∂ f~2 ∂~g3 3

and the corresponding vector ~bnf is

 
 (vS,1 − v∞ ) · c3,1 
 
~bnf = 
 .
.

, (3.41)
 . 
 
 
(vS,NT − v∞ ) · c3,NT

where vS,i is the velocity of the body at the centroid of the ith triangle.

60
3.3.2 Panel Perimeter Circulation Constraint in Matrix Form

The constraint for zero panel perimeter circulation has the form

Appc~x = ~bppc = ~0. (3.42)

To construct matrix Appc , consider a set of directed edges comprising the triangulation.

Each triangle is an ordered assembly of three edges, either with matching or reversed orien-

tation. This allows us to construct the triangle–edge adjacency matrix AT E , where element

i, j is unity if triangle i comprises edge j, negative unity in the case of reverse orientation,

and zero otherwise.

Next, construct a matrix of edge velocity integrals Aevi in which element i, j is the

contribution from the jth degree of freedom to the velocity integral along the ith edge.

Then, the perimeter panel circulation matrix is constructed as

Appc = AT E Aevi . (3.43)

3.3.3 Numerical Non-Lifting Problem Formulation

With the degrees of freedom arranged in column vector ~x as defined in Equation 3.37, the

numerical boundary element form of the problem formulation presented in Section 3.1.1 is

Determine ~x such that (3.44)

Anf~x = ~bnf (3.45)

Appc~x = ~bppc . (3.46)

61
These two constraints are concatenated as
   
 Anf   ~bnf 
A :=  , ~b :=   . (3.47)
   
Appc ~bppc

The constraint matrices Anf and Appc both have have size NT × Ndof , and so A has approx-

imate size 2Ndof × Ndof . This produces an overdetermined system that cannot be solved

exactly. Our solution strategy is to minimize the sum of the squared residuals

min (A~x − ~b)T W (A~x − ~b) (3.48)

~
x

subject to Aeq~x = ~beq

where W is a weighting matrix and Aeq and ~beq are optional equality constraints that will be

introduced subsequently for lifting flow. A perfectly determined system is often preferable

to an overdetermined one, and a square matrix could be obtained by omitting certain resid-

uals. But there is no guarantee that the remaining residuals will be linearly independent or

that the resulting system will be well-conditioned.

Because the panel perimeter circulation is an integral expression, the residuals in Equa-

tion 3.46 are intrinsically scaled by the panel areas, which are typically small. To balance

the scale of these residuals with the normal flow residuals, Equation 3.46 is scaled by the

reciprocal of the mean panel area.

Equation 3.48 is a quadratic program with linear equality constraints, which may be

solved with the following linear system [156]

    
G ATeq   ~x∗  −d~
   =  , (3.49)
    
Aeq 0 −~λ∗ ~beq

62
where

G := 2(AT W A), d~ := −2AT (W~b). (3.50)

The matrix on the left-hand side of Equation 3.49 is Hermitian and dense.

To further control the relative importance of the sources of residuals, the weighting

matrix W in Equation 3.48 is expressed simply as scalar weighting factor w ∈ (0, 1). The

quadratic and linear terms of the objective function are computed as

G = 2[wATnf Anf + (1 − w)ATppc Appc ] (3.51)

d~ = −2[wATnf~bnf + (1 − w)ATppc~bppc ]. (3.52)

The relative weighting factor, w is further explored in the next section.

3.4 Non-Lifting Comparison to Analytical Solution for Ellipsoid

To assess the errors associated with the least-squares formulation, consider a tri-axial el-

lipsoid with semi-major axis lengths a = 3, b = 2, c = 1, corresponding to the body-fixed

basis (b1 , b2 , b3 ) with Cartesian coordinates (x, y, z). The triangulations are formed from

a regular convex icosahedron by recursively dividing each triangle into four new triangles

and projecting the new vertices on the circumscribed sphere. This results in a triangulated

sphere which is then stretched along b1 , b2 , b3 by a, b, c. The number of triangles resulting

from N4 recursive divisions is

NT = 20 · 4N4 . (3.53)

Two metrics are considered to compare the numerical approximation to the analyti-

cal solution (i.e., exact solution) for the potential flow around an ellipsoid immersed in a

63
uniform freestream that is presented in Appendix C.

The first metric is the relative error of the aerodynamic pitching moment. For a freestream

velocity of v∞ = b2 + 81 b3 , the resulting net force is zero but the nonzero angle of attack

creates a pitching moment about b1 as

Z
b1 · −x × n(x)Cp (x)dA. (3.54)
Sb4

The analytic solution is sampled at the vertices of Sb4 and then used in the same manner

as the numerical solution to determine Cp ; that is, it defines the linear variation of surface

vorticity along each panel.

The second metric focuses on the local difference in surface pressure by integrating the

error over the surface of the ellipsoid:

Z
 
Cp (x) − Cp,0 (x)dA. (3.55)
Sb4

This metric captures noise in the solution that could be hidden in an otherwise accurate

pitching moment prediction.

The errors for a full range of values for the relative weighting factor w and triangulation

resolution N4 are depicted in Figures 3.5a and 3.5b. Note that resolutions of N4 = 2, 3, 4

correspond to triangulations with NT = 320, 1280, 5120. These figures reveal two promis-

ing results: the errors converge to zero as the resolution increases, and they are insensitive

to a broad range of w—from w = 0.2 to w = 0.9.

To further study the convergence behavior of the numerical method, Figure 3.6 presents

several error metrics as the triangulation resolution is refined: first (blue), the relative error

of the pitching moment calculated from sampling the analytic solution at the vertices as

64
 0.01 0.05

integrated surface pressure error

pitching moment relative error
N4 = 2 N4 = 2
N4 = 3 0.04 N4 = 3
0.005 N4 = 4 N4 = 4

0.03
0
0.02

-0.005
0.01

-0.01 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
w w

(a) Pitching moment relative error. (b) Integrated surface pressure error.

Figure 3.5: Aggregate metrics indicating the error associated with the relative weighting w
for ellipsoids of varying resolution.

compared to the exact pitching moment; second (red), the relative error of the pitching

moment calculated from the numerical prediction; third (yellow), the relative error of the

volume of the triangulated surface compared to the exact ellipsoid; lastly (purple), the

integrated error expressed in Equation 3.55. The horizontal axis variable in Figure 3.6 is

the mean edge length of each triangulation, and, therefore, represents the length scale of

the surface mesh. A gray line is included to indicate a second order rate of convergence,

which suggests we can conclude this boundary element method is second-order accurate.

As a demonstration of this numerical method, the flow solution for an ellipsoid with

5120 triangles is illustrated in Figure 3.7. The pathlines of surface vorticity form closed

loops, whereas the pathlines of velocity all travel from the forward stagnation point to the

aft stagnation point. As expected, the surface vorticity and surface velocity pathlines are

mutually perpendicular. To inspect how closely the normal flow and panel perimeter cir-

culation constraints were satisfied for this solution, histograms of the residuals from Equa-

tions 3.45 and 3.46 are presented in Figures 3.8a and 3.8b, respectively. These residuals

65
 100
moment, analytic
moment, numeric
10-1 volume
integrated error

relative error
10-2

10-3

10-4
10-1 100
mean edge length

Figure 3.6: Convergence of different calculations with the proposed numerical method.

closely match those of the analytic solution sampled at the vertices.

Also, as a further demonstration with this predicted flow solution, Equation 3.11 is

applied to calculate streamsurfaces in the external flow around the ellipsoid, as depicted

in Figure 3.9. The blue curves and the yellow curves are streamlines, and the red curves

and purple curves indicate fluid particle positions at constant time intervals. The blue/red

streamsurface remains in the vertical plane of symmetry and demonstrates that the stream-

lines passing near the stagnation points lag behind their neighboring streamlines that do

not. In contrast, the yellow/purple streamsurface travels around the tip without slowing

down, but it is sheared such that the upper portion is displaced inward.

The accurate prediction of streamlines near the surface is the main reason for using

panels with linearly varying surface vorticity. Our method is well-suited as an inviscid

solver to couple with an interactive boundary layer model to better simulate high Reynolds

number flow.

66
 γ v

Figure 3.7: Pathlines of surface vorticity, pathlines of surface velocity, and surface pressure
coefficient for the numerical flow solution.

(a) Normal flow residuals. (b) Panel perimeter circulation residuals.

Figure 3.8: Histograms of the residuals for the numerical calculation of surface vorticity
on the ellipsoid with 5120 triangles.

67
 choi
jun
young

Figure 3.9: Streamsurfaces around an ellipsoid calculated from the numerical solution.

68
 CHAPTER 4

COMBINED BOUND AND FREE VORTEX SHEETS

The previous two chapters detailed the equations and algorithms for modeling free vortex

sheets and bound vortex sheets independently. Of course, bound and free vortex sheets

abide by the same physical laws—their separate Eulerian and Lagrangian specifications

are for convenience.

As a demonstration, consider Figure 4.1, which depicts a sheet of fluid adjacent to a

body before and after it is set in motion. In the absence of viscous mixing, this sheet

continues to encase the body as it stretches downstream, away from the sharp trailing edge.

Thus, the bound and free vortex sheets we employ to describe this flow are actually two

parts of the same sheet. Furthermore, because the wake is assumed to be infinitely thin, the

two vortex sheets bounding the wake are actually collocated, and so they can be replaced

by a single vortex sheet. In Figure 4.1 this corresponds to replacing the two adjacent red

curves in the wake with the single yellow curve emanating out of the trailing edge.

In this manner, the free vortex sheet is imagined to be created at the trailing edge, but it

is simply an accounting technique to switch from an Eulerian specification to a Lagrangian

specification. This chapter is devoted to combining bound and free vortex sheets to produce

a comprehensive method capable of modeling complex aerodynamic flows.

69
 Figure 4.1: A sheet of fluid encasing a body before and after motion.

4.1 Problem Formulation for Unsteady Lifting Flows

In Section 3.1.3, we described the unsteady Kutta condition in continuous form, and we

now resume this discussion by presenting the final version of the numerical problem formu-

lation for lifting flow. Solving for the strength of the bound vortex sheet will fortuitously

provide the amount of circulation assigned to the free vortex sheet as it is created at the

trailing edge.

As previously mentioned, the body surface Sb is topologically split along a prescribed

wake separation path, such as a trailing edge. This splitting is implemented in the numer-

ical problem formulation by creating new vertices and edges along the separation path to

topologically divide the upper and lower surfaces. The two endpoint vertices of the sepa-

ration path are not duplicated; this modification to the surface triangulation is illustrated in

Figure 4.2. The circulation corresponding to each trailing edge vertex pair is calculated by

integrating the velocity along the loops indicated by the purple paths. This example depicts

a structured triangulation for which the paths comprising these loops can be determined

easily. Fortunately, appropriate paths can also be found on an unstructured triangulation

by applying shortest path algorithms from graph theory [120].

70
 topological description
integration
loops

closed tip
trailing edge
separation path
vertex pairs
endpoint

Figure 4.2: The tip section of a sample triangulated wing demonstrating the topological
split along the trailing edge. Note that the trailing edge vertex pairs are actually collocated.

Similar to calculating the circulation around each panel perimeter, the circulation around

these loops can be calculated by multiplying a sparse matrix Aloop , which assigns oriented

edges to each integration loop, and the matrix of edge velocity integrals previously in-

troduced in Section 3.3.2, Aevi . These circulations are arranged in a column vector and

expressed as

~Γ = AΓ~x, AΓ := Aloop Aevi . (4.1)

Unfortunately, matching the pressures at the trailing edge vertex pairs requires the so-

lution to Equation 3.24,

v(Pl ) · v(Pl ) − v(Pu ) · v(Pu ) 2 ∂

0 = Cp (Pu ) − Cp (Pl ) = 2
− 2 Γ(P ),
v∞ v∞ ∂t

which is nonlinear and thus cannot be included in the linear solution to the quadratic pro-

gram, Equation 3.49. Instead of solving Equation 3.24 directly, we modify the problem

formulation and implement a nonlinear iterative optimization procedure that leverages the

linear solution to the quadratic program.

In this nonlinear optimization procedure, the degrees of freedom are the scalar compo-

71
nents of surface vorticity at each of the trailing edge vertices, which are arranged in column

vectors as f~u , ~gu , f~l , ~gl and then concatenated into a single column vector
 
f~u 
 
 
~g 
 u
~z := 
 .
 (4.2)
 f~ 
 l
 
 
~gl

The number of degrees of freedom in ~z is quadruple the number of trailing edge vertex

pairs in the triangulation, and they are enforced in the quadratic program via an additional

set of linear equality constraints

Aeq,opt~x = ~z. (4.3)

The matrix Aeq,opt is a sparse matrix with one element per row in the appropriate column to

provide a surjective mapping from a subset of ~x onto ~z.

These new constraints are appended to the old linear equality constraints, which are

renamed Aeq,0 :  
 Aeq,0 
Aeq = 

,
 (4.4)
Aeq,opt

thereby allowing ~x and ~z to be related via the solution to the modified quadratic program:
   
 −d~   −d~ 
  
T
 G Aeq   ~x  
   
  
    = ~b  = ~b  + Az ~z, (4.5)
     eq,0   eq,0 
Aeq 0 −~λ 


 
 

~z ~0

where Az is a sparse matrix that places ~z at the bottom of the right-hand side vector. Then,

72
~x can be determined from ~z as

~x = Ax M −1 (~y0 + Az ~z), (4.6)

where  
 −d~ 
 
 G ATeq 
 
 
M := 

,
 ~y0 := ~
 eq,0  ,
b  (4.7)
Aeq 0 



~0

and Ax is a sparse matrix that removes the Lagrange multipliers ~λ present after applying

M −1 .

Therefore, the optimization problem statement to minimize the sum of the squared

residuals while enforcing the unsteady pressure matching Kutta condition is:

min F (~x) (4.8)

~
z

subject to ~ p (~x, ~z) = ~0

∆C

~x = Ax M −1~(~y0 + Az ~z)

with

F (~z) = 12 ~xT G~x + ~xT d~ + ~bT~b (4.9)

~ p (~x, ~z) = (f~u )2 + (~gu )2 − (f~l )2 − (~gl )2 + 2 (~Γnew − ~Γold )

∆C (4.10)
∆t
~Γnew = AΓ~x. (4.11)

The vector square notation in Equation 4.10 refers to the Hadamard product, e.g., the ith

element of vector (f~u )2 is fu,i

2
.

This modified optimization problem formulation has a quadratic objective function with

73
a set of quadratic equality constraints, and, therefore, it can be solved very efficiently by

an optimizer that leverages the first and second derivatives of the objective function and

constraints:

∇F (~z) = (ATz M −1T ATx GAx M −1 Az )~z + ATz M −1T ATx (d~ + GAx M −1 ~y0 ) (4.12)

∇2 F (~z) = ATz M −1T ATx GAx M −1 Az (4.13)

 
 2diag(f~u ) 
 
 
 2diag(~g ) 
 + 2 ATz M −1T ATx ATΓ
u 
~ p (~x, ~z)i ) = 

∇~z (∆C (4.14)
−2diag(f~ ) ∆t
 
 l 
 
 
−2diag(~gl )
 
 2~λ 
 
 
 2~λ 
~ p (~x, ~z)) = diag 
X
λi ∇~z2~z (∆C
 


 (4.15)
i   ~
−2λ

 
 
−2~λ

4.2 Wake Shedding Procedure

A new set of wake elements is shed at the wing trailing edge at the end of each time-step, as

will be discussed in the following section. The vortex sheet strength of these new elements

is determined from the solution to Equation 4.12 and corresponds to the circulation that

was shed during the time-step.

Figure 4.3 illustrates the wake shedding. The newly created edges that are purple in

color are assigned a circulation equal and opposite the corresponding element in ~Γnew −~Γold ;

this is analogous to the “bound circulation” often discussed in lifting-line theory. The newly

74
created edges that are dark red in color are assigned a circulation equal to the sum of the

corresponding edge circulations along the trailing edge (note the matching colors); this

is analogous to the “shed circulation” in lifting-line theory. Finally, the newly created

edges that are green in color are assigned a circulation such that the circulation around the

perimeter of the new wake panels sums to zero. The gray edges are those remaining from

previous iterations, and thus their circulation is unchanged.

topological description
integration
loops

new edges
old edges

Figure 4.3: Triangulated body and wake vortex sheets illustrating the wake shedding proce-
dure. Note that the groups of orange vertices are actually collocated (the wake is attached
to a closed trailing edge).

4.3 Algorithm for Combined Bound and Free Vortex Sheets

We now have all the equations and algorithms necessary to describe the overarching pro-

cedure for the simulation, presented in Algorithm 4.1.

75
Algorithm 4.1 Main simulation

1. Create the wing: Create the wing triangulation and calculate the local triangle and
vertex basis vectors.

2. Initialize: Initialize the bound vortex sheet by calculating Anf , Appc , AΓ , and setting
up the quadratic program.

3. First solution: Calculate the first bound vortex sheet solution with Equation 4.8,
letting ~Γold = ~0.

4. Shed first wake row: Shed the first row of wake panels from the trailing edge using
the recently calculated ~Γnew and the shedding procedure described in Section 4.2.

5. Iterate:

(a) Free influence: Obtain the surface vorticity on the free vortex sheet from the
stored edge circulations as described in Section 2.3.2. Then, calculate the influ-
ence of the free vortex sheet on the other vortex sheets with the treecode.
(b) Solve: Solve for the bound vortex sheet strength. This should only require a
couple iterations when using the previous solution as the initial guess.
(c) Bound influence: Calculate the influence of the bound vortex sheet on the free
vortex sheet.
(d) Displace: Displace the vertices defining the free vortex sheet according to the
chosen numerical ODE algorithm.
(e) Shed: Shed a new row of wake panels from the trailing edge.
(f) Adaptive Paneling: Apply the adaptive paneling algorithm described in Sec-
tion 2.3.4 to the free vortex sheet.

4.4 Comparison to Rectangular Wing Experiment

As a first test of our method, we compared it to a set of experiments conducted by McAl-

ister and Takahashi in the NASA Ames 7 ft x 10 ft subsonic wind tunnel [157]. Because

they were interested in observing the formation of the vortex near a wing tip, the authors

measured wake velocities and on-body surface pressures for various rectangular wings in

a uniform freestream. This provides a straightforward comparison to our simulation. In

the specific experiment to which we compared, McAlister and Takahashi used a straight

76
rectangular wing with a NACA 0015 airfoil, a flat (square) tip, a 1.7 ft chord, and a 5.59 ft

semispan (aspect ratio 6.576), operating at a 12◦ angle of attack and a Reynolds number of

2,500,000.

For our simulation we used the same wing and flow parameters, except that the freestream

velocity and wing chord were both normalized to unity. The simulation progressed 1500

iterations with a time step of 0.02, which means the starting vortex ended approximately 30

chord-lengths or 4.5 span-lengths downstream. The wing comprised 11,424 triangles, and

the wake comprised 228,547 triangles at the last iteration. Lastly, the smoothing parameter

for the free vortex sheet was 0.1. The shadowgraph and radiograph of this simulation are

presented in Figure 4.4, which indicate the presence of a well-formed vortex core composed

of two trailing vortices and the starting vortex.

The comparison of measured and predicted surface pressures is presented in Figure 4.5,

in which the data has been digitized from Figure 22(c) in McAlister [157]. McAlister noted

that wall corrections were not applied to the pressure data. The chordwise pressure distribu-

tions are plotted individually for different spanwise locations indicated by y/s, where s is

the semispan, and the predicted pressures are interpolated over the structured triangulation

to match the pressure tap locations. In general there is strong agreement between measured

and predicted values, except at the tip. McAlister deduced that part of the wake is shedding

along the tip upstream of the trailing edge and forming a vortex, which provides the ob-

served suction peak near the tip. The wake in our simulation was prescribed to shed from

the trailing edge—and not upstream along the tip—thus, our simulation does not replicate

this behavior. Another difference is observed at the trailing edge, where the predicted pres-

sures approach stagnation but the measured pressures do not. This is due to our inviscid

77
flow assumption, which is invalid at the trailing edge because the boundary layer thickness

is not small compared to the airfoil thickness. Note that the pressure at x/c = 1 provided

by McAlister is actually extrapolated from the neighboring pressure measurements.

Next, a comparison of wake velocities near one of the trailing vortices is provided in

Figure 4.6, in which the data has been digitized from Figure 39 in McAlister [157]. The

experimental velocities were measured with a two-color laser velocimeter in scans along

the span direction (y) at different locations downstream of the trailing edge. The velocities

predicted at the end of the simulation, which are sufficiently steady, are clearly well short

of the measured values. The smoothing parameter δ = 0.1 is providing too much smoothing

and is overly restricting the peak velocity in the vortex core. If we instead use a smaller

δ = 0.015 with the same predicted wake and the circulation therein, a much closer match is

observed. To be clear, the yellow curves in Figure 4.6 are calculated from the same wake

half-edge circulations, only with less spreading of vorticity. Unfortunately, conducting the

simulation with such a small smoothing parameter is quite difficult because of the rapid

formation of small eddies and other chaotic features [71, 111, 158]. Note that in Figure 4.6

the origin for the x, y, z coordinates has been translated to the left tip trailing edge.

Lastly, a comparison between the measured and predicted spanwise circulation distri-

bution is depicted in Figure 4.7, in which the data has been digitized from Figure 37(a)

in McAlister [157]. This experiment used the same parameters as the previous two, ex-

cept with a Reynolds number of 1,500,000. McAlister calculated circulation from velocity

measurements taken around a chordwise rectangular path encasing the wing at different

spanwise stations. For a direct comparison, we calculated circulation in the same manner,

using off-body velocities calculated at the end of the simulation with δ = 0.1, and these

78
Figure 4.4: Shadowgraph and radiograph of the simulated wake.

79
predictions are indicated by the red square markers. Also, the yellow curve indicates the

circulation calculated as ~Γ = AΓ~x (Equation 4.1), which involves integrating the surface

velocity along the bound vortex sheet (the purple lines in Figure 4.3). The predicted cir-

culation closely matches the data except near the tip, where the simulated wake spreads

vorticity more than is measured win the wind tunnel experiment.

80
 -5
y/s = 0.994 y/s = 0.974 y/s = 0.944
-4
McAlister
-3 numerical

Cp -2

-1

1
-5
y/s = 0.899 y/s = 0.843 y/s = 0.773
-4

-3
Cp -2

-1

1
-5
y/s = 0.692 y/s = 0.597 y/s = 0.491
-4

-3
Cp -2

-1

1
-5
y/s = 0.370 y/s = 0.238 y/s = 0.094
-4

-3
Cp -2

-1

1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/c x/c x/c

Figure 4.5: Comparison to Figure 22(c) in McAlister [157].

81
 1 1
x/c = 0.1 McAlister x/c = 2.0
numerical, δ=0.1
0.5 numerical, δ=0.015 0.5
vz /v∞

0 0

-0.5 -0.5

-1 -1
-0.1 0 0.1 0.2 -0.1 0 0.1 0.2
1 1
x/c = 0.2 x/c = 4.0
0.5 0.5
vz /v∞

0 0

-0.5 -0.5

-1 -1
-0.1 0 0.1 0.2 -0.1 0 0.1 0.2
1 1
x/c = 0.5 x/c = 6.0
0.5 0.5
vz /v∞

0 0

-0.5 -0.5

-1 -1
-0.1 0 0.1 0.2 -0.1 0 0.1 0.2
1
y/s
x/c = 1.0
0.5
vz /v∞

0 z

-0.5 y
McAlister
x,v ∞ numerical, δ=0.1
numerical, δ=0.015
-1
-0.1 0 0.1 0.2
y/s

Figure 4.6: Comparison to Figure 39 in McAlister [157].

82
 0.6

0.5

0.4
Γ/(cv∞ )

0.3

0.2

0.1 McAlister
external field loops
0 wing surface paths

0.4 0.5 0.6 0.7 0.8 0.9 1 1.1

y/s

Figure 4.7: Comparison to Figure 37(a) in McAlister [157].

83
4.5 Comparison to Wind Turbine Experiment

As a second test of our method, and to more closely examine a complex wing–wake inter-

action, we compared to the results of the National Renewable Energy Laboratory (NREL)

unsteady aerodynamics experiment (UAE) phase VI, which examined a horizontal axis

wind turbine in the NASA Ames 80 ft x 120 ft wind tunnel. As documented by Hand et

al. [159], the two-blade wind turbine has a 10 m diameter, straight taper, nonlinear twist,

and uses the S809 airfoil.

The blade triangulation used in our simulation is depicted in Figure 4.8 alongside the

line drawing from Figure 9 in Hand et al. [159]. We did not include the region near the root

in which the blade transitions from a circular cross-section to the S809 airfoil because the

blade twist reduces to 0◦ at the root, which results in a very large local angle of attack and

would likely incur fully separated flow. Figure 4.8 also indicates the radial position of the

pressure measurements used for comparison. The structured triangulation for each blade

comprised 6502 triangles, including 402 triangles for the flat tips.

0.5
r/R = 0.300 0.466 0.633 0.800 0.950

-0.5

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Figure 4.8: Blade triangulation and pressure measurement locations juxtaposed with Figure
9 in Hand et al. [159].

84
 We compared to a subset of the experiments in which the blade rotated at 71.63 rpm

in freestreams of 5 m/s and 7 m/s, with 0◦ yaw, and with a blade tip angle of 3◦ . The tip

Reynolds number is on the order of the one million. We used a time step of 0.0025 seconds,

which corresponds to 335 time steps per blade revolution, and the simulation progressed

for a total of four revolutions, or 1340 iterations. At the last iteration, the wakes together

comprised 650,126 triangles for the 5 m/s case and 1,165,176 triangles for the 7 m/s case,

and these wakes are presented in Figures 4.9 and 4.10, respectively. Strong vortices form

behind both the root and tip of each blade, with the vortices at the tips carrying more

vorticity.

An unexpected phenomenon that arose in these simulations was the lasting effect of

the starting vortex. When a blade passed over the starting vortex (after 180◦ ), there was

a reduction in circulation around the blade, followed quickly by a restoration thereof, as

would be expected. However, this transit past the starting vortex deposited a significant

amount of vorticity into the wake which then affected the passing of the next blade (after

another 180◦ ), and this process cascaded through each subsequent revolution of the rotor.

The radial and axial positions of the tip vortex, as defined by the peak concentration of

vorticity, are presented in Figures 4.11a and 4.11b for the 7 m/s simulation. The persistent

effect of the starting vortex can be seen in these figures at regular intervals of 180◦ . These

tip vortex locations are compared to predictions from Lynch et al. [160], who used FUN3D,

the unstructured Navier–Stokes code from NASA Langley. The small fluctuations in the

radial position prediction are due to the adaptive paneling scheme.

Finally, the predicted section pressures after four rotor revolutions are compared to the

measurements from the NREL experiments. Figure 4.12 displays the comparisons for both

85
the 5 m/s and 7 m/s cases. In general, the predictions are quite close to the measurements.

As mentioned previously, the predicted pressures at the trailing edge tend toward stagna-

tion, whereas the measured pressures stay closer to the ambient pressure. Note that for

these rotating blades the pressure coefficient (Equation 3.21) is modified as

v·v 2 ∂Φ
Cp := 1 − 2 +v 2
− 2 +v 2
, (4.16)
v∞ loc v∞ loc ∂t

where vloc is the local speed of the measurement point, which depends on the radial position

along the blade. Integrating these surface pressures gives a predicted shaft torque of 988

Nm for the 7 m/s case, as compared to the 801 Nm reported by Simms [161]. This is an

overprediction of 23%, which is likely due to not including the effect of viscous drag.

(a) Shadowgraph (b) Cross-section (slice) of vorticity magnitude

Figure 4.9: Wake visualizations for 5 m/s freestream.

86
 (a) Shadowgraph (b) Cross-section (slice) of vorticity magnitude

Figure 4.10: Wake visualizations for 7 m/s freestream.

0 6
Lynch (FUN3D) Lynch (FUN3D)
-2 current method current method
radial position (m)

5.5
axial position (m)

-4
5
-6

4.5
-8

-10 4
0 90 180 270 360 450 540 0 90 180 270 360 450 540
vortex age (deg) vortex age (deg)
(a) Vortex axial position (b) Vortex radial position

Figure 4.11: Location of tip vortex compared to the prediction of Lynch et al. [160] for the
7 m/s case.

87
 v∞ = 5 m/s v∞ = 7 m/s
-2 -4
NREL UAE r/R = 0.300 r/R = 0.300
-3
numerical
-1
-2
Cp
-1
0
0
1 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
-2 -4
r/R = 0.466 r/R = 0.466
-3
-1
-2
Cp
-1
0
0
1 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
-2 -4
r/R = 0.633 r/R = 0.633
-3
-1
-2
Cp
-1
0
0
1 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
-2 -4
r/R = 0.800 r/R = 0.800
-3
-1
-2
Cp
-1
0
0
1 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
-2 -4
r/R = 0.950 r/R = 0.950
-3
-1
-2
Cp
-1
0
0
1 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x

Figure 4.12: Section pressures for 5 m/s (left) and 7 m/s (right) cases compared with NREL
UAE measurements [159]. Note the different vertical axis scales.

88
4.6 Example Simulation: Two Aircraft in Close Formation

As a demonstration of an intended application of our numerical method, consider two air-

craft flying in a close formation, with the tailing aircraft slightly lower and to the right of

the leading aircraft, as indicated in Figure 4.13. The example aircraft is a notional flying

wing created in OpenVSP [162] and triangulated with the OpenVSP CFD unstructured

meshing tool, resulting in 6390 triangles (each). The reference span length is 12.5 ft and

the wing section is a NACA five-series reflexed airfoil with five degrees of washout at the

tip. The aircraft are both prescribed an 11.5◦ angle of attack and their positions are held

fixed during the simulation.

Figure 4.13: Two notional flying wings arranged in a close formation.

The simulation proceeded for 800 iterations with a time-step of ∆t = 0.04s, for a total

duration of 32 seconds, and with a freestream velocity of v∞ = 1 ft/s. The smoothing

parameter used for the wake is δ = 0.2, which is in the same proportion to the reference

89
length as the smoothing parameter used in the validation studies. The results of this simu-

lation are presented in Figures 4.14–4.18. By the last iteration, the two wakes comprised

3,253,078 triangles.

The computing times for each step of Algorithm 4.1 are displayed in Figure 4.14. As

expected, the influence velocity calculations dominate the computing time spent during

each iteration, especially as the number of wake panels grows more rapidly near the end of

the simulation. The adaptive paneling algorithm, which is executed entirely on the CPU,

also consumes a significant amount of computing time refining the wake shed from the

leading aircraft.

102

101
computing time (s)

100

10-1

free influence
10-2
bound influence
leading wake refinement
10-3 trailing wake refinement
bound solution
10-4
0 200 400 600 800
iteration

Figure 4.14: Individual computing times for each calculation during an iteration.

The radiograph, which shows the concentration of vorticity carried by the wake, is

illustrated in Figure 4.15. In this figure, the wings are transparent so any vorticity passing

beneath the wing can also be observed. Most of the wake from the leading aircraft passes

beneath the trailing aircraft, except for the right tip vortex, which passes above the trailing

90
aircraft. The tip vortices emerging from the right tips of each aircraft interact in an unstable

and chaotic manner. The transition to chaos appears to happen shortly downstream from

the trailing aircraft. The vorticity shed by the trailing aircraft lacks left–right symmetry—

the right side is shedding far more vorticity than the left. This is likely due to the left tip

vortex from the leading aircraft providing a beneficial upwash, while the right tip vortex

alters the lift distribution much less favorably.

The shadowgraph, which reveals the shape of the vortex sheets, echoes the observations

from the radiograph. The dark regions correspond to very large numbers of triangles, and

suggest that the chaotic interactions of the two wakes lead to a much greater computational

burden than an isolated plain wing (e.g., Figure 4.4). A second shadowgraph is presented

in Figure 4.17 to clarify the path of the wake from the leading aircraft as it passes both

above and below the trailing aircraft.

Finally, the surface pressures are presented in Figure 4.18. The images of the bottom

surface of each wing are mirrored such that the tips are aligned with the corresponding top

views (as if the bottom surface could be viewed from above). Both the highest and lowest

pressures throughout the flow field are observed around the right midspan of the trailing

aircraft. In fact, the peak positive pressure coefficient exceeds 1.0, indicating a significant

unsteady motion. Of course, this is a direct result of the tip vortex from the leading wing.

The asymmetric pressure distribution on the trailing aircraft results in a rolling moment

toward the left. In fact, this rolling moment would be reversed if the impinging vortex

passed beneath the right wing instead of above it.

The prediction of these surface pressures relied on an accurate representation of the

wake shed by the leading aircraft, as well as accurate near-body velocities to carry this

91
wake closely around the trailing aircraft. This provides the principal motivation for this

thesis.

92
Figure 4.15: Radiograph of the simulated wakes. Darker regions indicate strong vorticity.

93
Figure 4.16: Shadowgraph of the simulated wakes. Darker regions indicate many layers of
the vortex sheet.

94
Figure 4.17: A shadowgraph with a closer view of the aircraft, which are not transparent.

Figure 4.18: Surface pressures for each side of each aircraft. The bottom surfaces are
left–right mirrored.

95
 CHAPTER 5

CONCLUSIONS

In this thesis, we described a new surface vorticity boundary element method to predict

intricate wake–body interactions for incompressible high Reynolds number flow. We were

motivated by the impressive detail and rigor of vortex sheet simulations in the vortex meth-

ods literature [1], as compared to traditional panel code wake models. In a traditional

force-free wake model [10, 13, 14], the wake is represented by a lattice of vortex filaments

that progress through the flow in a Lagrangian manner while maintaining constant circula-

tion. Plausible wake evolution behavior was achieved in these models by simply applying

a viscous core model to the Biot–Savart equation to limit the effects of the velocity singu-

larity near a filament. However, these wake models quickly become non-physical as they

evolve. Thus, we expected that significant improvement could be achieved by incorporating

concepts from the vortex methods literature.

On the other hand, the source–doublet panel method for representing potential flow

around a body has been well developed into effective high-order models, but the surface

vorticity method has not received similar attention. In this work, we recognized that bound

vortex sheets and free vortex sheets are two separate descriptions of the same object, and,

therefore, we hypothesized that triangular elements with linearly varying surface vorticity

could be used to effectively described the entire flow, including not only wake elements

but also elements on aerodynamic surfaces. These higher-order elements would provide

accurate streamlines around a bound vortex sheet and also underpin the well-behaved evo-

96
lution of a free vortex sheet. Therefore, we pursued a boundary element method that used

the same surface vorticity elements for both wings and wakes. Many lessons have been

learned along the way.

5.1 Free Vortex Sheets

High-order wake panels provide a more realistic induced velocity field than singular parti-

cles or filaments. However, we found that they are still insufficient for well-behaved wake

evolution because thin vortex sheets naturally roll up into a curvature singularity [90]. In-

deed, this had been discovered in the vortex methods literature [86], and the solution is to

regularize the vortex sheet. We adopted a common regularization technique of including a

smoothing parameter in the Biot-Savart kernel, which acts as a surrogate for viscosity in

high Reynolds number flows and bounds the peak concentration of vorticity.

Yet, this regularization is still insufficient to provide an accurate representation of a

vortex sheet. As the elements strain, they may become too large or too irregular to ade-

quately resolve the sheet. The discretization must be refined to maintain an appropriate

level of detail; the lengths of the edges in the triangulation should not significantly exceed

the regularization length.

The main benefit of representing a free vortex sheet with linear surface vorticity pan-

els is the ability to resolve vorticity accurately over a relatively large space, which would

otherwise require many vortex particles for a similar level of accuracy, as demonstrated

in Section 2.3.2. However, this advantage only applies when the vortex sheet is mostly

flat, which corresponds to regions with little vorticity. In regions of concentrated vorticity,

which cause the sheet to roll up into a spiral with many layers, higher-order triangular pan-

97
els lose their advantage because large numbers of panels are still required. Furthermore,

when the layers of the rolled-up sheet are much closer together than the regularization

length (the smoothing parameter δ), a sheet representation becomes unnecessary and po-

tentially misleading.

Our adaptive paneling algorithm included a Delaunay edge flipping scheme that helped

maintain more regular triangles. It also had a beneficial effect of reducing the number of

edges that needed to be split because the flipped edges tended to become shorter. Our

implementation of this algorithm required the sheet to be a locally planar graph, which

requires an edge to neighbor exactly two triangles, unless it is on the boundary. Unfor-

tunately, maintaining a locally planar graph was very challenging, and it required many

specialized and detailed numerical procedures to avoid erroneous results.

5.2 Bound Vortex Sheets

We pursued a higher-order surface vorticity representation of bound vortex sheets for two

reasons: (1) a convenient congruity with the free vortex sheets, for which the equations

governing the panels had already been developed and implemented, and (2) to provide

accurate predictions of velocity, both on and near the surface of a wing. Fortunately, our

expectations were met. Our method could also be considered a velocity-based boundary

element method because surface vorticity and surface velocity are interchangeable via the

surface normal vector.

Defining two degrees of freedom for surface vorticity at each vertex of a triangulation

was a natural choice to ensure C 0 continuity of γ over the entire body, which is critical

for accurate surface velocities. Additionally, the panel centroids are a natural choice to

98
enforce no normal flow because the velocity induced by a panel is indeterminate at the

edges and vertices. This should come as no surprise, because a triangular vortex sheet

could not physically exist in isolation unless the strength dropped to zero at the edges.

Selecting surface vorticity instead of doublet or source strength requires the imposition of

the panel perimeter circulation constraint to enforce ∇ · γ = 0. Fortunately, this calculation

contributes only a negligible cost to the solution and is relatively easy to implement.

As a result of these choices for constraints and degrees of freedom, there are approx-

imately twice as many constraints as degrees of freedom. We obtain a surface vorticity

solution by formulating a least squares minimization problem. Unfortunately, this requires

the multiplication of two large and full matrices. At the same time, the least squares formu-

lation is also liberating. Additional constraints can be freely included; for example, some

panels could have multiple collocation points. Also, the least squares approach gives the

freedom to control which residuals are the most important to minimize.

Our boundary element method is especially accurate and robust for closed surfaces,

such as an ellipsoid, including with random or irregular triangulations. Though small in

contribution, the vector rotation scheme was critical to the accuracy of the bound vortex

sheet method (Section 3.2). However, the lifting solution for a wing was more sensitive

and required more careful weighting of the residuals.

It is also worth clarifying the difference between the surface vorticity method and a tra-

ditional source–doublet panel method. Aside from their shared classification as boundary

element methods, there are several fundamental differences. The potential formulation as-

sumes the flow to be irrotational and then arrives at a Laplace equation for a scalar potential

field. The surface vorticity method, however, assumes the vorticity to be concentrated as a

99
vortex sheet at the surface of the body, and the problem formulation is arranged to achieve

a valid bound vortex sheet (stagnant internal flow and a solenoidal surface vorticity field).

Instead of solving a Laplace equation, a Poisson equation is solved via the Biot–Savart law.

The surface vorticity method has an additional burden of constraining the degrees of free-

dom to produce a solenoidal surface vorticity field, but the resulting benefit is that surface

velocity is directly obtained from the solution without requiring the gradient computations.

In some versions of the panel method, wakes are prescribed to extend into the body; this is

not necessary for the surface vorticity method. Indeed, the flow is stagnant inside a closed

body and the wakes emanate externally from a prescribed shedding path (e.g., a trailing

edge).

5.3 Combined Bound and Free Vortex Sheets

The bound and free vortex sheets are combined into a complete simulation via the wake

shedding scheme at the wing trailing edge, which enforces the unsteady pressure matching

Kutta condition. This constraint requires tracking the time rate of change of circulation

around the wing. Ideally, any path starting from the lower surface and ending at the cor-

responding vertex on the upper surface would correctly predict the circulation, which is

equivalent to the potential jump at the trailing edge. However, this path independence

relies on the panel perimeter circulations being zero, which is not perfectly achieved. Un-

fortunately, the circulation calculation appears to be sensitive to the errors accumulated

along these paths.

The large matrix M , inside the bound vortex sheet problem formulation 4.8, can be

computed and inverted once, and then stored to be accessed at each iteration. Accordingly,

100
computing the solution for the strength of the bound vortex sheet is inexpensive compared

to the other computing tasks, as demonstrated in Figure 4.14. However, this computation

is only possible when the aerodynamic bodies are not in relative motion. For example, M

remains constant for isolated rotating helicopter blades, but M would vary with time if a

non-rotating component were included (e.g., a fuselage). This limitation could be remedied

with a least-squares GMRES approach [163–165].

5.4 Realizability Constraints

Throughout this thesis, several parameters, thresholds, and other choices were presented

without clear guidance for their values. However, these choices can affect the quality of the

simulation.

For non-lifting flow, the solution is insensitive to the quality of the triangulation, but this

is not the case for lifting flow. To achieve a quality solution for a wing, it is important to

refine the triangulation close to the tips and to smoothly continue the resolution all the way

around the tip, whether the tip is round or square. Also, there is not an obvious choice for

the vertex normal vector at the trailing edge of the tips, but this normal vector determines

the plane in which surface velocity must reside. Placing the normal vector in the plane of

the wing chord and facing outward typically produced the best results.

For lifting flow, it is important to choose the weighting factor (Equation 3.51) carefully.

When the panel perimeter circulation residuals are normalized by the average panel area,

the weighting factor should be between 0.85 and 0.98.

The smoothing parameter δ is analogous to the thickness of a vortex sheet, and is, there-

fore, connected to the length scale of the simulation. The spans of the wings used in this

101
thesis were close to 10 (feet or meters), and δ = 0.1 offered a reasonable balance between

accuracy (vortex core size) and computational cost (number of computational elements).

Below δ = 0.05, the free vortex sheet tended toward rapid instability. Experimental mea-

surements could be used to tune δ to achieve the measured peak vorticity.

Similarly, the threshold lengths for the adaptive paneling edge splitting and removal

schemes must also achieve a balance between cost and accuracy. In most simulations, we

used a maximum edge length of 0.125 and a minimum of 0.01. Also, we found that newly

created panels should forgo any adaptive paneling for several iterations to prevent any noise

from feeding back into the bound vortex sheet solution.

Lastly, the time-step is also an important parameter. It contributes to the unsteady terms

in the bound vortex sheet solution and the vertex displacements of the free vortex sheets

(e.g., a forward Euler step). For the simulations in this thesis with a freestream velocity

of unity, a time-step between 0.01 and 0.02 provided sufficient detail while maintaining

stability.

5.5 Summary of Contributions

The principal contribution of this thesis is the accurate and detailed simulation of complex

wake–body interactions, unaccompanied by any volume meshes.

The wake is represented by triangular panels with linearly varying surface vorticity,

which maintain constant circulation along every half-edge, a generalization of the constant

strength elements used by Stock [63]. The triangular panels retain a nearly regular shape via

the adaptive paneling algorithm, which includes a novel Delaunay edge flipping scheme.

Our higher-order surface vorticity boundary element method for determining the bound

102
vortex sheet around a lifting body is a significant generalization of existing low-order sur-

face vorticity methods, and is based on several enabling numerical methods that were novel

to this work.

The analytic solution to the velocity induced by the linear strength elements allowed

for efficient Biot–Savart calculations without the need for a quadrature scheme. The panel

perimeter circulation constraint provides a simple approach to guaranteeing a solenoidal

surface vorticity field, which was considered one of the main obstacles facing a surface

vorticity method [151]. The vector rotation scheme ensures a fully continuous surface

vorticity field across edges and vertices.

For these models and algorithms to successfully provide a detailed depiction of the

flow, a large number of computational elements is required. The associated computations

would not have been practical without the treecode and its implementation on a graphics

processing unit.

The hypothesis stated in the introduction asserted that interacting aerodynamic flows

with strong couplings between wakes and lifting bodies can be modeled efficiently and

effectively as vortex sheets discretized into triangular panels with linearly varying surface

vorticity. We have provided significant evidence to support this hypothesis, but we have

also documented several areas for improvement.

5.6 Future Work

There are several remaining tasks that are worth pursuing. A well known challenge faced by

panel methods (and by extension, boundary element methods) is the accurate calculation

of induced drag [10, 166]. A common remedy when using straight prescribed wakes is

103
to calculate induced drag in the Trefftz plane [167, 168]. Unfortunately, a Trefftz plane

calculation is not available for unsteady wakes and a new technique may be required.

The least squares minimization formulation balanced the normal flow residuals with

the panel perimeter circulation residuals via a weighting factor. We showed that the results

were insensitive to this weighting factor for non-lifting flow, but for lifting flow the weight-

ing factor had to be chosen more carefully. A more intricate weighting scheme could help

produce better results for lifting flows and for unstructured triangulations.

Regions of strong vorticity tend to cause a free vortex sheet to roll up into a spiral

with many layers, requiring a large number of panels. If fine detail in this region is not of

interest, such as a starting vortex that has traveled far downstream, the smoothing parameter

could increase with time, thereby slowing the growth rate.

In the adaptive paneling scheme, the decision to split or remove an edge is based entirely

on the edge length, as compared to a global threshold. Of course, there are many other

criteria that could be used, and including the circulation strength along the edge could

be advantageous. For example, edges that carry little circulation should be favored for

removal.

104
Appendices

105
 APPENDIX A

EVALUATION OF THE BIOT–SAVART INTEGRAL

The analytic solution to the Biot–Savart integral in Equation 3.15 is critical to the imple-

mentation of the proposed method, particularly for the self-induction terms. Solutions to

various forms of the Biot–Savart integral are presented by Suh [169]; however, the solution

approach employed in this work is derived from that of Johnson [142] in the development

of PANAIR. This method is valid for any polygon, but is described here only for a triangle.

A.1 Geometric Description

Consider the velocity vA induced at point PA from a sheet of vorticity γ linearly distributed

over surface S is given by the Biot–Savart law as

1 r × γ(r)
Z
vA = dS(r) (A.1)
4π S ||r||3

with r directed from PA to the differential dS. To solve this integral, we will decompose it

into several more manageable integrals and introduce several coordinates.

Surface S is defined by the triangle (P1 , P2 , P3 ), as indicated in Figure A.1. Plane P

contains S, the first two vectors of orthonormal basis (b1 , b2 , b3 ), and point PB , which is the

projection of PA onto the plane. The position vectors pA , p1 , etc. describe the position of

their respective points from a common origin O, and vector r describes the relative position

from PA to the differential element dS. The projection of pA onto plane P defines point PB

as pB := pA −(pA −p1 )·b3 b3 . The height of PA above PB is z := (pA −pB )·b3 = −r ·b3 ,

which gives ρ = r + zb3 , where ρ is an additional vector directed to dS.

106
 PA vA

r
pA
b3
b2 P3
PB dS
b1 ρ γ
pB p S
3 P1
P2 P
a3 p1
p2

a1O a2

Figure A.1: Geometric relations between triangle (P1 , P2 , P3 ) and point PA measured rel-
ative to origin O.

A.2 Specification of Domains of Integration

Instead of integrating on S directly, three new areas S1 , S2 , and S3 are created to cover

exactly S as shown in Figure A.2, where S1 := triangle (PB , P1 , P2 ), S2 := triangle

(PB , P2 , P3 ), and S3 := triangle (PB , P3 , P1 ). With vectors qi := pi − pB , i = 1, 2, 3,

the orientation of triangle Si is given by the sign of

b3 · (qi × qi+1 ), i = 1, 2, 3 and q4 := q1 . (A.2)

In the example illustrated in Figure A.2, S3 has a negative orientation while S1 and S2 have

a positive orientation, and, therefore, S = ∪3i=1 Si . The proof of this equality for a general

triangle follows from the application of Equation A.2 to compute the area of S.

Then, the integral in Equation A.1 is rewritten as

3 Z
1 r×γ 1 X r×γ
Z
3
dS = dSi . (A.3)
4π S ||r|| 4π i=1 Si ||r||3

In the next section, these integrals will be further decomposed.

107
 P2
S2
P3 q2
q3 S1

P1
b2 q1
S3

PB b1

Figure A.2: Three surfaces, S1 , S2 , and S3 , used to cover exactly S. In this case, S3 has
opposite orientation of S1 and S2 .

A.3 Coordinates for Integration

Without loss of generality, this section will focus on the integration of surface S1 , which

corresponds to triangle (PB , P1 , P2 ) and vectors q1 := p1 − pB and q2 := p2 − pB . Define

a new orthonormal basis (c1 , c2 , c3 ) as

c3 := b3 , (A.4)
q2 − q1
c2 := , (A.5)
||q2 − q1 ||

c1 := c2 × c3 , (A.6)

which is illustrated in Figure A.3. Three parameters are established to set the bounds of

integration in the subsequent calculations: a := q1 · c1 , l1 := q1 · c2 , and l2 := q2 · c2 .

For the purpose of integration, a Cartesian coordinate system is created on plane P to

describe ρ with independent variables u := ρ · c1 , v := ρ · c2 . Recall the height coordinate

z, which gives r = uc1 + vc2 − zc3 and ρ = uc1 + vc2 . The coordinates u, v are the

independent variables of integration, and so dS1 = dudv. Additionally, we also introduce

a pair of polar coordinates (ρ, φ), where φ is measured from the line along c1 toward the

108
 P2
l2 c2

S1
P1 l1 c 2

v u ac1
φ2 φ φ1
b2
c2
c1
PB b1

Figure A.3: Cartesian coordinates (u, v) and polar coordinates (ρ, φ) for integration on S1 .

direction of c2 , ρ = ||ρ||, and the substitution relations are: u = ρ cos φ, v = ρ sin φ, and

dS1 = ρdρdφ.

Therefore, surface vorticity strength is expressed in terms of two scalar functions as

γ(u, v) = f (u, v)c1 + g(u, v)c2 which vary linearly as

f (u, v) = a00 + a10 u + a01 v, (A.7)

g(u, v) = b00 + b10 u + b01 v. (A.8)

The numerator in Equation A.3 may be expanded as

r × γ = zg(u, v)c1 − zf (u, v)c2 + [ug(u, v) − vf (u, v)] c3 (A.9)

= z(b00 + b10 u + b01 v)c1 − z(a00 + a10 u + a01 v)c2

+(b00 u + b10 u2 + b01 uv − a00 v − a10 uv − a01 v 2 )c3 (A.10)

One difficulty in solving for the antiderivatives arises from the possibility of the de-

nominator in the integrand to reach zero. To avoid this, a parameter δ > 0 is introduced to

109
force the denominator to be positive. Accordingly, the height of PA above PB is modified
√
to h := z 2 + δ 2 . The solutions subsequently obtained will be exact in the limit as δ → 0.
√
Along with ρ = u2 + v 2 we have

√ p
r := ||r|| = u2 + v 2 + z 2 + δ 2 = ρ2 + h2 . (A.11)

The integral may now be expressed as

r×γ 1
Z Z
dS1 = z(b00 + b10 u + b01 v)c1 − z(a00 + a10 u + a01 v)c2
S1 ||r||3 S1 r3
+(b00 u + b10 u2 + b01 uv − a00 v − a10 uv − a01 v 2 )c3 .

(A.12)

Because each power of u and v involve antiderivatives of different forms, we define the H

integrals in a form similar to Johnson [142]:

um v n dudv
Z
Hmn := , (A.13)
S1 r3

which gives

r×γ
Z
dS1 = z(b00 H00 + b10 H10 + b01 H01 )c1 − z(a00 H00 + a10 H10 + a01 H01 )c2
S1 ||r||3
+(b00 H10 + b10 H20 + b01 H11 − a00 H01 − a10 H11 − a01 H02 )c3 . (A.14)

A.4 Integral solutions

The H integrals are solved by switching to the polar coordinates presented in the previous

section

φ2 ρ
um v n dudv (cosm φ)(sinn φ)(ρm+n+1 )dρdφ
Z Z Z
Hmn = = . (A.15)
S1 r3 φ1 0 (ρ2 + h2 )3/2

110
The inner integral is solved first. There are three cases corresponding to ρ, ρ2 , and ρ3

ρ
p
ρdρ ρ2 + h2 − h
Z
2 2 3/2
= p (A.16)
0 (ρ + h ) h ρ2 + h2
ρ
ρ2 dρ ρ
Z p
2 2 3/2
= ln(ρ + ρ2 + h2 ) − p − ln h (A.17)
0 (ρ + h ) ρ2 + h2
Z ρ
ρ3 dρ ρ2 + 2h2
2 2 3/2
= p (A.18)
0 (ρ + h ) ρ2 + h2

Next, the outer integrals are solved by switching coordinates one last time from φ back

to a rectilinear coordinate l (see Figure A.3) with the relations

a
cos φ = √ tan φ = l/a (A.19)
l 2 + a2
l adl
sin φ = √ dφ = √ (A.20)
l 2 + a2 l 2 + a2

The solutions are

  l 2
1 al
= tan−1

H00 √ 
h a2 + h2 + h l2 + a2 + h2 l1
√ l
l √   √  l ln h  2
H10 =√ ln l + a + h + l + a − ln l + l + a + h − √
2 2 2 2 2 2 2 2
l 2 + a2 l 2 + a2  l 1
a h √ √ i l2
H01 = √ ln h − ln 2 2 2
l +a +h + l +a 2 2 
2
l +a 2 
l1
2 2 2
√
 l 2
al l + a + h − h l + a + h 2 2 2

−1 l −1 hl 
H20 = √ − h tan + h tan √ 
(l2 + a2 ) l2 + a2 + h2 a a l2 + a2 + h2 l1
√
−a2 l2 + a2 + h2 − h l2 + a2 + h2 l2


H11 = √
(l2 + a2 ) l2 + a2 + h2 
l1
  l 2
l hl
H02 = −h tan−1 + h tan−1

√ 
a a l2 + a2 + h2 l1

111
 APPENDIX B

TREECODE DESCRIPTION AND PERFORMANCE

A treecode is an algorithm used to efficiently approximate the influence of a set of particles,

and it proceeds in two major steps: first, a tree is constructed over a set of particles, and

then it is applied to a set of influence points. A tree is a hierarchy of cells, starting with the

root cell. In an octree, which is used for three-dimensional space, each cell can have up to

eight child cells, which, in turn, can also each have up to eight child cells, etc. Parent cells

are internal to the tree, whereas cells without any children are external, and these external

cells contain the particles. In keeping with the botany analogy, external cells are also called

leaves.

In this thesis, we construct a tree in the same manner as Barnes and Hut [122], with

the exception that we allow Nmax particles per external cell instead of only one. The in-

structions for constructing a tree are presented in Algorithm B.1. The steps that are serial

in nature are performed on the central processing unit (CPU), whereas the steps that are

parallel in nature are performed on the graphics processing unit (GPU).

In Step 3 of Algorithm B.1, the particle moments are calculated by the Cartesian Taylor

series expansion of the Biot–Savart law, Equation 2.65, as described by Lindsay [128]. The

order of the Taylor series approximation, p, requires p(p + 1)(p + 2)/6 moment coefficients

to be calculated for each cell; thus, p must be chosen carefully.

Then, as described in Algorithm B.3, the influence at a point p is calculated by starting

at the root and traversing the tree, until a cell is considered sufficiently far from p, or until

112
an external cell is reached and the exact particle–point calculations are used. A point is

considered sufficiently far from a cell if the following inequality is true:

L < θd, (B.1)

where L is the characteristic length of the cell, and d is the distance from the center of the

cell to the point. Therefore, higher values of θ encourage more approximations, whereas

lower values are more conservative.

Algorithm B.1 Construct a tree

1. Root Cell: The root cell is the first cell of the tree from which the first set of children
are created. The root cell is the entry point into the tree as it is stored as a linked list.
Create the root cell as a cube slightly larger than the extent of all of the particles.

2. Create Tree: Loop through every particle and place them in the root cell with Algo-
rithm B.2. This will create the tree structure as cells fill with particles and split into
children.

3. Particle Moments: Loop over every particle and calculate the moment relative to
the origin. Perform this computation on the GPU and copy the result back to the
CPU.

4. Aggregate Moments: Aggregate all the moments of the children of each cell, start-
ing with the root cell. Descend down the tree until the external cells are reached, and
then recursively wind back up to the root cell.

5. Copy: Copy the tree from the CPU memory to the GPU memory.

6. Translate Moments: Translate the moments of each cell from the origin to the cell
centers. Perform this calculation on the GPU.

An example two-dimensional quadtree is illustrated in Figure B.1a to demonstrate the

structure of a tree. In this example, the maximum number of particles permitted in a cell is

Nmax = 2. Note that external cells are not all of the same size, and some contain fewer than

Nmax particles. The application of this example tree to a particular point is demonstrated

113
Algorithm B.2 Put particle P into cell C
if C is external then
if C is not over capacity then
P is now in C
else
Divide C into eight child octants, which are empty and external cells
Put P and any pre-existing particles in C into the appropriate children of C
C is now an internal cell
end if
else
Put P into the appropriate child of C
end if

Algorithm B.3 Calculate v, the influence of cell C at point p

if C is internal then
if p is sufficiently far from C then
calculate v as the point–cluster interaction from C on p
else
for each octant child O of C do
calculate the influence of O at point p
end for
end if
else
for each particle Q in C do
v = v+ the direct influence of Q on p
end for
end if

114
in Figure B.1b. The cells in which an approximation is used are colored according to their

size. The gray cells are closer than the approximation threshold, and, therefore, the exact

particle–point influence is used for the particles within them. In this example, the influence

of 48 particles is calculated by 16 direct particle summations and 11 cluster approximations.

(a) Quadtree structure for a set of particles. (b) Approximate calculation of velocity at a point.

Figure B.1: Example quadtree.

To determine appropriate values for the parameters Nmax , p, and θ that efficiently bal-

ance cost and accuracy, a set of experiments were conducted for a full factorial table with:

Nmax = 5, 10, 20, 40, 80, 160 (B.2)

p = 1, 2, 3, 4, 5, 6 (B.3)

θ = 0.01, 0.04, 0.09, . . . , 0.81, 1.00 (B.4)

The results are presented in Figures B.2a–B.2c, in which each figure presents the computa-

tional time and an aggregate error metric for a set of 750, 000 particles arranged in a torus.

115
The separate figures are each colored by a different parameter to reveal their independent

effects. The θ parameter closely follows the cost–accuracy Pareto frontier (the gray line),

while p and Nmax trend in opposite directions. The parameter values chosen for all applica-

tions of the treecode in this thesis are Nmax = 160, p = 2, and θ = 0.5, as indicated by the

black circle.

116
 θ p
0 0.2 0.4 0.6 0.8 1 1 2 3 4 5 6

100 100

10-2 10-2

10-4 10-4
error

error
10-6 10-6

10-8 10-8

10-1 100 101 10-1 100 101

computational time (s) computational time (s)

(a) Colored by distance parameter. (b) Colored by Taylor series approximation order.

Nmax
5 10 20 40 80 160

100

10-2

10-4
error

10-6

10-8

10-1 100 101

computational time (s)

(c) Colored by leaf cell capacity.

Figure B.2: Results for the treecode parameter experiments colored by each parameter.

117
 APPENDIX C

POTENTIAL FLOW AROUND AN ELLIPSOID

The solution for the potential flow around an ellipsoid originates with Lamb [66], and was

popularized by Munk [170]. This section is a summary of the derivation presented by

Clarke [171]. To begin, consider a tri-axial ellipsoid with semi-principal axes lengths a, b,

and c and assume without loss of generality that a > b > c. The ellipsoid is described in

Cartesian coordinates (x, y, z) with basis (b1 , b2 , b3 ) as

x2 y 2 z 2
+ 2 + 2 = 1. (C.1)
a2 b c

The ellipsoid can also be described by confocal ellipsoidal coordinates (λ, µ, ν) which

relate to the Cartesian coordinates as

(a2 + λ)(a2 + µ)(a2 + ν)

x2 = , (C.2)
(a2 − b2 )(a2 − c2 )
(b2 + λ)(b2 + µ)(b2 + ν)
y2 = , (C.3)
(b2 − c2 )(b2 − a2 )
(c2 + λ)(c2 + µ)(c2 + ν)
z2 = , (C.4)
(c2 − a2 )(c2 − b2 )

where −a2 < ν < −b2 < µ < −c2 < λ < ∞, and the ellipsoid is obtained when λ = 1.

The outward normal vector on the ellipsoid surface is generated by the partial derivatives

118
of the coordinates (x, y, z) with respect to λ

∂x x
= 2
, (C.5)
∂λ 2(a + λ)
∂y y
= 2
, (C.6)
∂λ 2(b + λ)
∂z z
= . (C.7)
∂λ 2(c2 + λ)

The normal vector n is

 
1 ∂x ∂y ∂z
n= b1 + b2 + b3 , (C.8)
hλ ∂λ ∂λ ∂λ

where
x2 y2 z2
 
1
hλ = + + . (C.9)
4 (a2 + λ)2 (b2 + λ)2 (c2 + λ)2

Also, the parameters α0 , β0 , and γ0 are computed with the following elliptic integrals

∞
dλ
Z
α0 = abc p , (C.10)
0 (a2 + λ) (a2 + λ)(b2 + λ)(c2 + λ)
∞
dλ
Z
β0 = abc p , (C.11)
0 (a2 + λ) (b + λ)(b2 + λ)(c2 + λ)
2

∞
dλ
Z
γ0 = abc p . (C.12)
0 (a2 + λ) (c2 + λ)(b2 + λ)(c2 + λ)

Next, consider a freestream with velocity v∞ defined component-wise with respect to

basis (b1 , b2 , b3 ) as

v∞ = U∞ b1 + V∞ b2 + W∞ b3 . (C.13)

Finally, the velocity v at any point on the ellipsoid surface is given in terms of the surface

119
normal vector

2U∞
vU∞ = [b1 − (b1 · n)n] , (C.14)
2 − α0
2V∞
vV∞ = [b2 − (b2 · n)n] , (C.15)
2 − β0
2W∞
vW∞ = [b3 − (b3 · n)n] , (C.16)
2 − γ0

v = vU∞ + vV∞ + vW∞ . (C.17)

120
 REFERENCES

[1] G.-H. Cottet and P. D. Koumoutsakos, Vortex Methods: Theory and Practice. Cam-
bridge University Press, 2000.

[2] J. R. Mansfield, O. M. Knio, and C. Meneveau, “Dynamic LES of Colliding Vortex

Rings Using a 3D Vortex Method,” Journal of Computational Physics, vol. 152,
no. 1, pp. 305–345, 1999.

[3] S. M. Belotserkovskii and M. I. Nisht, “Nonstationary Nonlinear Theory of a Thin

Wing of Arbitrary Planform,” Fluid Dynamics, vol. 9, no. 4, pp. 583–589, 1974.

[4] J. Katz, “Lateral Aerodynamics of Delta Wings with Leading-Edge Separation,”

AIAA Journal, vol. 22, no. 3, pp. 323–328, 1984.

[5] C. P. Mracek, “A Vortex Panel Method for Potential Flows with Applications to
Dynamics and Control,” PhD thesis, Virginia Polytechnic Institute and State Uni-
versity, 1988.

[6] Z. Mittelman and I. Kroo, “Unsteady Aerodynamics and Control of Delta Wings
with Tangential Leading-Edge Blowing,” in 16th Atmospheric Flight Mechanics
Conference, AIAA, Boston, MA, Aug. 1989.

[7] T. Sarpkaya, “Computational Methods with Vortices—The 1988 Freeman Scholar

Lecture,” Journal of Fluids Engineering, vol. 111, no. 1, pp. 5–52, 1989.

[8] D. Mook and A. Nayfeh, “Numerical Simulations of Dynamic/Aerodynamic Inter-

actions,” Computing in Science & Engineering, vol. 1, no. 2, pp. 461–482, 1990.

[9] S. C. Smith and I. M. Kroo, “Computation of Induced Drag for Elliptical and
Crescent-shaped Wings,” Journal of Aircraft, vol. 30, no. 4, pp. 446 –452, 1993.

[10] S. C. Smith and I. M. Kroo, “Induced Drag Computations on Wings with Accu-
rately Modeled Wakes,” Journal of Aircraft, vol. 34, no. 2, pp. 253 –255, 1997.

[11] S. Preidikman, “Numerical Simulations of Interactions among Aerodynamics, Struc-

tural Dynamics, and Control Systems,” PhD thesis, Virginia Polytechnic Institute
and State University, 1998.

[12] G. Bramesfeld, “A Higher Order Vortex-Lattice Method with a Force-Free Wake,”

PhD thesis, Pennsylvania State University, 2006.

121
[13] D. L. Ashby, M. R. Dudley, S. K. Iguchi, L. Browne, and J. Katz, “Potential Flow
Theory and Operation Guide for the Panel Code PMARC,” National Aeronautics
and Space Administration, Ames Research Center, Technical Memorandum TM-
102851, Jan. 1991.

[14] J. Katz and A. Plotkin, Low-Speed Aerodynamics, Second. Cambridge University

Press, 2001.

[15] R. A. Roccia, S. Preidikman, and J. C. Massa, “Modified Unsteady Vortex-Lattice

Method to Study Flapping Wings in Hover Flight,” AIAA Journal, 2013.

[16] D. Maniaci, “Wind turbine design using a free-wake vortex method with winglet
application,” PhD thesis, The Pennsylvania State University, 2013.

[17] A. Dhruv, C. J. Blower, and A. M. Wickenheiser, “A Three Dimensional Iterative

Panel Method for Bio-Inspired Multi-Body Wings,” in Proceedings of the ASME
2014 Smart Materials, Adaptive Structures and Intelligent Systems, ASME, New-
port, RI, Sep. 2014.

[18] R. Paul, J. Murua, and A. Gopalarathnam, “Unsteady and Post-Stall Aerodynamic

Modeling for Flight Dynamics Simulation,” in AIAA SciTech, AIAA, National Har-
bor, MD, Jan. 2014.

[19] H. Hesse and R. Palacios, “Reduced-Order Aeroelastic Models for Dynamics of

Maneuvering Flexible Aircraft,” AIAA Journal, vol. 52, no. 8, pp. 1717–1732, Aug.
2014.

[20] Y. Hirato, M. Shen, and S. Aggarwal, “Initiation of Leading-Edge-Vortex Forma-

tion on Finite Wings in Unsteady Flow,” in AIAA SciTech, AIAA, Kissimmee, FL,
Jan. 2015.

[21] T. R. Quackenbush, A. H. Boschitsch, D. A. Wachspress, and K. Chua, “Rotor De-

sign Optimization Using a Free Wake Analysis,” National Aeronautics and Space
Administration, Tech. Rep., 1993.

[22] D. A. Wachspress, T. R. Quackenbush, and A. H. Boschitsch, “First-Principles

Free-Vortex Wake Analysis for Helicopters and Tiltrotors,” in American Helicopter
Society 59th Annual Forum, American Helicopter Society, vol. 59, Phoenix, AZ,
May 2003, pp. 1763–1786.

[23] M. J. Bhagwat and J. G. Leishman, “Time–Accurate Modeling of Rotor Wakes Us-

ing a Free-Vortex Wake Method,” in 18th AIAA Applied Aerodynamics Conference,
Denver, CO, 2000.

122
[24] S. Y. Wie, S. Lee, and D. J. Lee, “Potential Panel and Time-Marching Free-Wake-
Coupling Analysis for Helicopter Rotor,” Journal of Aircraft, vol. 46, no. 3, pp. 1030–
1041, 2009.

[25] H. Abedi, “Development of Vortex Filament Method for Aerodynamic Loads on

Rotor Blades,” Master’s thesis, Chalmers University of Technology, 2013.

[26] S. C. Smith, “A Computational and Experimental Study of Nonlinear Aspects of

Induced Drag,” PhD thesis, Stanford University, Jun. 1995.

[27] J. Conway and F. Tezok, “A Time-marching Scheme for the CANAERO Three-
dimensional Vortex Sheets Panel Method,” in 38th Aerospace Sciences Meeting &
Exhibit, AIAA, Reno, Nevada, Jan. 2000.

[28] O. Levinski, “Prediction of Buffet Loads on Twin Vertical Tails Using a Vortex
Method,” Defense Science & Technology Organization, Austrailia, Tech. Rep. DSTO-
RR-0217, Jul. 2001.

[29] M. Roura, A Cuerva, A Sanz-Andrés, and A. Barrero-Gil, “A Panel Method Free-

Wake Code for Aeroelastic Rotor Predictions,” Wind Energy, vol. 13, no. 4, pp. 357–
371, 2010.

[30] J. Tan and H. Wang, “Simulating Unsteady Aerodynamics of Helicoptor Rotor

with Panel/Viscous Vortex Particle Method,” Aerospace Science and Technology,
vol. 30, pp. 255–268, 2013.

[31] J. P. Moore IV, “An Arbitrarily High-Order, Unstructured, Free-Wake Panel Solver,”
Master’s thesis, Massachusetts Institute of Technology, Sep. 2013.

[32] C. R. Satterwhite, “Development of CPANEL, An Unstructured Panel Code, Us-

ing a Modified TLS Velocity Formulation,” Master’s thesis, California Polytechnic
State University, San Luis Obispo, CA, Aug. 2015.

[33] C. Sousa, “Unsteady Panel Code Utilizing a Vortex Particle Wake,” Master’s thesis,
California Polytechnic State University, San Luis Obispo, CA, Dec. 2016.

[34] N. Ramos-Garcı́a, J. N. Sørensen, and W. Z. Shen, “Three-Dimensional Viscous-

Inviscid Coupling Method for Wind Turbine Computations,” Wind Energy, vol. 19,
no. 1, pp. 67–93, 2016.

[35] D. J. Willis, “An Unsteady, Accelerated, High Order Panel Method with Vortex
Particle Wakes,” PhD thesis, Massachusetts Institute of Technology, Jun. 2006.

123
[36] D. J. Willis, J. Peraire, and J. K. White, “A Combined pFFT–Multipole Tree Code,
Unsteady Panel Method with Vortex Particle Wakes,” International Journal for Nu-
merical Methods in Fluids, vol. 53, pp. 1399–1422, 2007.

[37] S. G. Voutsinas, “Vortex Methods in Aeronautics: How to Make Things Work,”

International Journal of Computational Fluid Dynamics, vol. 20, no. 1, pp. 3–18,
2006.

[38] J. S. Calabretta and R. A. McDonald, “A Three Dimensional Vortex Particle-Panel

Method for Modeling Propulsion-Airframe Interaction,” in 48th AIAA Aerospace
Sciences Meeting, AIAA, Orlando, FL, 2010, pp. 1–16.

[39] E. H. Martin, “Assessment of Panel and Vortex Particle Methods for the Modelling
of Stationary Propeller Wake Wash,” Master’s thesis, Memorial University of New-
foundland, 2015.

[40] J. R. Mansfield, O. M. Knio, and C. Meneveau, “Towards Lagrangian Large Vortex

Simulation,” in ESAIM: Proceedings, EDP Sciences, vol. 1, 1996, pp. 49–64.

[41] D. G. Opoku, D. G. Triantos, F. Nitzsche, and S. G. Voutsinas, “Rotorcraft Aero-

dynamic and Aeroacoustic Modelling Using Vortex Particle Methods,” in Proc. of
the ICAS 2002 Congress, The International Council of the Aeronautical Sciences,
Toronto, Canada, 2002.

[42] J. Zhao and C. He, “A Viscous Vortex Particle Model for Rotor Wake and Interfer-
ence Analysis,” in American Helicopter Society 64th Annual Forum, AHS, Mon-
treal, Canada, May 2008.

[43] C. He and J. Zhao, “Modeling Rotor Wake Dynamics with Viscous Vortex Particle
Method,” AIAA Journal, vol. 47, no. 4, pp. 902–915, 2009.

[44] J. Zhao and C. He, “A Viscous Vortex Particle Model for Rotor Wake and In-
terference Analysis,” Journal of the American Helicopter Society, vol. 55, no. 1,
pp. 12 007–12 007, 2010.

[45] S. Park, Y. Chu, and D. Lee, “Investigation of Main Rotor-Tail Rotor Interaction
in Hover Flight using Vortex Particle Method,” in 5th Asian/Australian Rotorcraft
Forum, Asian/Australian Rotorcraft Forum, 2016.

[46] M. J. Stock, A. Gharakhani, and C. P. Stone, “Modeling Rotor Wakes with a Hy-
brid OVERFLOW-Vortex Method on a GPU Cluster,” in 28th AIAA Applied Aero-
dynamics Conference, AIAA, Chicago, IL, Jun. 2010.

124
[47] J. Zhao and C. He, “A Hybrid Solver with Combined CFD and Viscous Vortex
Particle Method,” in 67th Annual Forum of the American Helicopter Society, AHS,
Virginia Beach, VA, May 2011.

[48] N. Rajmohan, J. Zhao, and C. He, “A Coupled Vortex Particle/CFD Methodology

for Studying Coaxial Rotor Configurations,” in Fifth Decennial AHS Aeromechan-
ics Specialists Conference, San Francisco, California, 2014.

[49] C. He and N. Rajmohan, “Modeling the Aerodynamic Interaction of Multiple Ro-

tor Vehicles and Compound Rotorcraft with Viscous Vortex Particle Method,” in
American Helicopter Society 72nd Annual Forum, AHS, West Palm Beach, FL,
May 2016.

[50] F. Caradonna, “Developments and Challenges in Rotorcraft Aerodynamics,” in 38th

Aerospace Sciences Meeting and Exhibit, AIAA, Reno, NV, Jan. 2000, p. 109.

[51] R. C. Strawn and M. J. Djomehri, “Computational Modeling of Hovering Rotor

and Wake Aerodynamics,” Journal of Aircraft, vol. 39, no. 5, pp. 786–793, 2002.

[52] R. M. Eshcol, C. Zhou, J. Kim, and L. N. Sankar, “A Comparative Study of Two

Hover Prediction Methodologies,” in 54th AIAA Aerospace Sciences Meeting, AIAA,
San Diego, CA, Jan. 2016.

[53] R. C. Strawn, F. X. Caradonna, and E. P. Duque, “30 Years of Rotorcraft Com-

putational Fluid Dynamics Research and Development,” Journal of the American
Helicopter Society, vol. 51, no. 1, pp. 5–21, 2006.

[54] N. S. Hariharan, T. A. Egolf, and L. N. Sankar, “Simulation of Rotor in Hover: Cur-

rent State, Challenges and Standardized Evaluation,” in 52nd Aerospace Sciences
Meeting, AIAA, National Harbor, MD, Jan. 2014.

[55] C. G. Speziale, “On the Advantages of the Vorticity–Velocity Formulation of the

Equations of Fluid Dynamics,” Journal of Computational Physics, vol. 73, pp. 476–
480, 1987.

[56] T. B. Gatski, “Review of Incompressible Fluid Flow Computations Using the Vorticity–
Velocity Formulation,” Applied Numerical Mathematics, vol. 7, pp. 227–239, 1991.

[57] R. E. Brown, “Rotor Wake Modeling for Flight Dynamic Simulation of Helicopters,”
AIAA Journal, vol. 38, no. 1, pp. 57–63, 2000.

[58] F. Scheurich and R. E. Brown, “Modelling the Aerodynamics of Vertical-axis Wind

Turbines in Unsteady Wind Conditions,” Wind Energy, vol. 16, no. 1, pp. 91–107,
2013.

125
[59] G. R. Whitehouse and A. H. Boschitsch, “Innovative Grid-Based Vorticity–Velocity
Solver for Analysis of Vorticity-Dominated Flows,” AIAA Journal, vol. 53, no. 6,
pp. 1655–1669, Jun. 2015.

[60] M. J. Smith, R. Shenoy, A. R. Kenyon, and R. E. Brown, “Vorticity-transport and

Unstructured RANS Investigation of Rotor–fuselage Interactions,” in 35th Euro-
pean Rotorcraft Forum, 2009.

[61] G. R. Whitehouse, A. H. Boschitsch, M. J. Smith, C. E. Lynch, and R. E. Brown,

“Investigation of Mixed Element Hybrid Grid-based CFD Methods for Rotorcraft
Flow Analysis,” in 66th Annual Forum of the American Helicopter Society, 2010.

[62] G. R. Whitehouse and H. Tadghighi, “Investigation of Hybrid Grid-Based CFD

Methods for Rotorcraft Flow Analysis,” in American Helicopter Society Aerome-
chanics Specialists Meeting, Jan. 2010.

[63] M. J. Stock, “A Regularized Inviscid Vortex Sheet Method for Three Dimensional
Flows with Density Interfaces,” PhD thesis, University of Michigan, 2006.

[64] H. Feng, “Vortex Sheet Simulations of 3-D Flows Using an Adaptive Triangular
Panel/Particle Method,” PhD thesis, The University of Michigan, 2007.

[65] A. Leonard, “Vortex Methods for Flow Simulation,” Journal of Computational

Physics, 1980.

[66] H. Lamb, Hydrodynamics. Cambridge University Press, 1895.

[67] M. J. Stock, W. J. Dahm, and G. Tryggvason, “Impact of a Vortex Ring on a Density

Interface Using a Regularized Inviscid Vortex Sheet Method,” Journal of Compu-
tational Physics, vol. 227, pp. 9021–9043, 2008.

[68] H. von Helmholtz, “On Integrals of the hydrodynamical equations, which express
vortex-motion,” The London, Edinburgh, and Dublin Philosophical Magazine and
Journal of Science, vol. 33, no. 226, pp. 485–512, 1867.

[69] G. K. Batchelor, An Introduction to Fluid Dynamics. Cambridge University Press,

1967.

[70] J. Casey and P. M. Naghdi, “On the Lagrangian Description of Vorticity,” Archive
for Rational Mechanics and Analysis, vol. 115, pp. 1–14, 1991.

[71] S.-I. Sohn, “Two Vortex-Blob Regularization Models for Vortex Sheet Motion,”
Physics of Fluids, vol. 26, 2014.

126
[72] C. Pozrikidis, “Theoretical and Computational Aspects of the Self-induced Mo-
tion of Three-Dimensional Vortex Sheets,” Journal of Fluid Mechanics, vol. 425,
pp. 335–366, 2000.

[73] T. R. Kane and D. A. Levinson, Dynamics, Theory and Applications. McGraw Hill,
1985.

[74] L. Rosenhead, “The Formation of Vortices from a Surface of Discontinuity,” Pro-

ceedings of the Royal Society of London, vol. 134, no. 823, pp. 170–192, Nov.
1931.

[75] F. L. Westwater, “The Rolling Up of the Surface of Discontinuity behind an Aero-

foil of Finite Span,” Aeronautical Research Council, Tech. Rep. R&M 1692, Aug.
1935.

[76] K. Kuwahara and H. Takami, “Numerical Studies of Two-Dimensional Vortex Mo-

tion by a System of Point Vortices,” Journal of the Physical Society of Japan,
vol. 34, pp. 247–253, 1973.

[77] L. Prandtl, “Applications of Modern Hydrodynamics to Aeronautics,” National Ad-

visory Committee for Aeronautics, Technical Report 116, 1923.

[78] G. Birkhoff and J. Fisher, “Do Vortex Sheets Roll Up?” Rendiconti del Circolo
Matematico di Palermo, vol. 8, no. 1, pp. 77–90, 1959.

[79] G. Birkhoff, “Helmholtz and Taylor Instability,” in Proceedings of Symposia in

Applied Mathematics, 1962, pp. 55–76.

[80] R. R. Clements and D. J. Maull, “The Representation of Sheets of Vorticity by

Discrete Vortices,” Progress in Aerospace Sciences, vol. 16, no. 2, pp. 129–146,
1975.

[81] P. T. Fink and W. K. Soh, “A New Approach to Roll-Up Calculations of Vortex

Sheets,” Proceedings of the Royal Society A: Mathematical, Physical and Engi-
neering Sciences, vol. 362, pp. 195–209, 1978.

[82] P. G. Saffman and G. R. Baker, “Vortex Interactions,” Annual Review of Fluid Me-
chanics, vol. 11, pp. 95–122, 1979.

[83] G. R. Baker, “A Test of the Method of Fink and Soh for Following Vortex-Sheet
Motion,” Journal of Fluid Mechanics, vol. 100, pp. 209–220, 1980.

[84] R. Krasny, “A Study of Singularity Formation in a Vortex Sheet by the Point-Vortex

Approximation,” Journal of Fluid Mechanics, 1986.

127
[85] D. W. Moore, “On the Point Vortex Method,” SIAM Journal on Scientific and Sta-
tistical Computing, 1981.

[86] A. J. Chorin and P. S. Bernard, “Discretization of a Vortex Sheet with an Example

of Roll-Up,” Journal of Computational Physics, vol. 13, pp. 423–429, 1973.

[87] C. W. Oseen, “Uber Wirbelbewegung in Einer Reibenden Flssigkeit,” Arkiv fr Matem-

atik, Astronomi och Fysik, vol. 7, no. 14, pp. 14–21, 1911.

[88] N. Rott, “On the Viscous Core of a Line Vortex,” Zeitschrift für Angewandte Math-
ematik und Physik (ZAMP), vol. 9, no. 5, pp. 543–553, 1958.

[89] A. M. Bloom and H. Jen, “Roll-Up of Aircraft Trailing Vortices using Artificial
Viscosity,” Journal of Aircraft, vol. 11, no. 11, pp. 714–716, Nov. 1974.

[90] R. Krasny, “Desingularization of Periodic Vortex Sheet Roll-Up,” Journal of Com-

putational Physics, vol. 65, pp. 292–313, 1986.

[91] C. R. Anderson, “A Vortex Method for Flows with Slight Density Variations,” Jour-
nal of Computational Physics, vol. 61, pp. 417–44, 1985.

[92] L Rosenhead, “The Spread of Vorticity in the Wake behind a Cylinder,” Proceed-
ings of the Royal Society of London. Series A, Containing Papers of a Mathematical
and Physical Character, pp. 590–612, 1930.

[93] R. Krasny, “Vortex Sheet Roll-Up Due to the Motion of a Flat Plate,” in Interna-
tional Symposium on Nonsteady Fluid Dynamics, J. A. Miller and D. P. Telionis,
Eds., vol. 92, 1990.

[94] O. Hald and V. M. Del Prete, “Convergence of Vortex Methods for Euler’s Equa-
tions,” Mathematics of Computation, vol. 32, no. 143, pp. 791–809, 1978.

[95] O. H. Hald, “Convergence of Vortex Methods for Euler’s Equations. II.,” SIAM
Journal on Numerical Analysis, 1979.

[96] R. Krasny, “Computation of Vortex Sheet Roll-Up in the Trefftz Plane,” Journal of
Fluid Mechanics, 1987.

[97] G. Winckelmans, R. Cocle, L. Dufrense, and R. Capart, “Vortex Methods and their
Applications to Trailing Wake Vortex Simulations,” C. R. Physique, vol. 6, pp. 467–
486, 2005.

[98] G. Tryggvason, “Simulation of Vortex Sheet Roll-Up by Vortex Methods,” Journal

of Computational Physics, 1989.

128
 [99] D. W. Moore, “The Spontaneous Appearance of a Singularity in the Shape of an
Evolving Vortex Sheet,” Proceedings of the Royal Society of London, vol. 365,
pp. 105–119, 1979.

[100] M. J. Shelley, “A Study of Singularity Formation in Vortex-Sheet Motion by a Spec-

trally Accurate Vortex Method,” Journal of Fluid Mechanics, vol. 244, pp. 493–
526, 1992.

[101] J. W. Rottman and P. K. Stansby, “On the ‘δ-Equations’ for Vortex Sheet Evolu-
tion,” Journal of Fluid Mechanics, vol. 247, pp. 527–549, 1993.

[102] S. J. Cowley, G. R. Baker, S. Tanveer, and M. Page, “On the Origin of Singularities
in Vortex Sheet Motion,” in 26th AIAA Fluid Dynamics Conference, AIAA, San
Diego, CA, Jun. 1995.

[103] M. Nitsche, “Singularity Formation in a Cylindrical and a Spherical Vortex Sheet,”

Journal of Computational Physics, vol. 173, no. 1, pp. 208–230, 2001.

[104] J. T. Beale and A. Majda, “Vortex Methods. I: Convergence in Three Dimensions,”

Mathematics of Computation, vol. 39, no. 159, pp. 1–27, 1982.

[105] J. T. Beale, “On the Accuracy of Vortex Methods at Large Times,” in IMA Work-
shop on Computational Fluid Dynamics and Reacting Gas Flows, Springer–Verlag,
1988.

[106] G. S. Winckelmans and A. Leonard, “Contributions to Vortex Particle Methods for

the Computation of Three-Dimensional Incompressible Unsteady Flows,” Journal
of Computational Physics, vol. 109, no. 2, pp. 247–273, 1993.

[107] G.-H. Cottet, P. Koumoutsakos, and M. L. O. Salihi, “Vortex Methods with Spa-
tially Varying Cores,” Journal of Computational Physics, vol. 162, no. 1, pp. 164–
185, 2000.

[108] M. E. Agishtein and A. A. Migdal, “Dynamics of Vortex Surfaces in Three Dimen-

sions Theory and Simulations,” Physica D, vol. 40, pp. 91–118, 1989.

[109] M. Brady, A. Leonard, and D. I. Pullin, “Regularized Vortex Sheet Evolution,”

Journal of Computational Physics, vol. 146, pp. 520–545, 1998.

[110] K. Lindsay and R. Krasny, “A Particle Method and Adaptive Treecode for Vor-
tex Sheet Motion in Three-Dimensional Flow,” Journal of Computational Physics,
vol. 172, pp. 879–907, 2001.

129
[111] H. Feng, L. Kaganovskiy, and R. Krasny, “Azimuthal Instability of a Vortex Ring
Computed by a Vortex Sheet Panel Method,” Fluid Dynamics Research, vol. 41,
2009.

[112] O. A. Kandil, L.-C. Chu, and T. Tureaud, “A Nonlinear Hybrid Vortex Method for
Wings at Large Angle of Attack,” AIAA Journal, vol. 22, no. 3, pp. 329–336, 1984.

[113] L. Kaganovskiy, “Adaptive Panel Method for Particle Simulation of Three Dimen-
sional Vortex Sheet Motion,” PhD thesis, University of Michigan, 2006.

[114] L. Kaganovskiy, “Adaptive Panel Representation for 3D Vortex Ring Motion and
Instability,” Mathematical Problems in Engineering, vol. 2007, 2007.

[115] L. Kaganovskiy, “Adaptive Panel Representation for Oblique Collision of Two Vor-
tex Rings,” International Journal of Non-Linear Mechanics, vol. 46, no. 1, pp. 9–
13, 2011.

[116] J. Jost, Compact Riemann Surfaces: An Introduction to Contemporary Mathemat-

ics. Springer Science & Business Media, 2006.

[117] G. Tryggvason, W. J. A. Dahm, and K. Sbeih, “Fine Structure of Vortex Sheet

Rollup by Viscous and Inviscid Simulation,” Journal of Fluids Engineering, vol. 113,
pp. 31–36, Mar. 1991.

[118] M. Nitsche and R. Krasny, “A Numerical Study of Vortex Ring Formation at the
Edge of a Circular Tube,” Journal of Fluid Mechanics, vol. 276, pp. 139–161, 1994.

[119] H. Edelsbrunner, Geometry and Topology for Mesh Generation. Cambridge Uni-
versity Press, 2001.

[120] R. Diestel, Graph Theory. Springer-Verlag, 2000.

[121] A. W. Appel, “An Efficient Program for Many-body Simulation,” SIAM Journal on
Scientific and Statistical Computing, vol. 6, no. 1, pp. 85–103, Jan. 1985.

[122] J. Barnes and P. Hut, “A Hierarchical O(N log N ) Force-calculation Algorithm,”

Nature, vol. 324, pp. 446–449, 1986.

[123] L. Greengard and V. Rokhlin, “A Fast Algorithm for Particle Simulations,” Journal
of Computational Physics, vol. 73, no. 2, pp. 325–348, 1987.

[124] F. Zhao, “An O(N ) Algorithm for Three-Dimensional N-Body Simulations,” PhD
thesis, Massachusetts Institute of Technology, 1987.

130
[125] J. K. Salmon and M. S. Warren, “Fast Parallel Tree Codes for Gravitational and
Fluid Dynamical N-Body Problems,” International Journal of High Performance
Computing Applications, vol. 8, no. 2, pp. 129–142, 1994.

[126] C. I. Draghicescu and M. Draghicescu, “A Fast Algorithm for Vortex Blob Interac-
tions,” Journal of Computational Physics, vol. 116, no. 1, pp. 69–78, 1995.

[127] G. S. Winckelmans, J. K. Salmon, M. S. Warren, A. Leonard, and B. Jodoin, “Appli-

cation of Fast Parallel and Sequential Tree Codes to Computing Three-Dimensional
Flows with the Vortex Element and Boundary Element Methods,” in ESAIM: Pro-
ceedings, EDP Sciences, vol. 1, 1996, pp. 225–240.

[128] K. Lindsay, “A Three-Dimensional Cartesian Tree-Code and Applications to Vor-

tex Sheet Roll-Up,” PhD thesis, University of Michigan, 1997.

[129] J. Board and K. Schulten, “The Fast Multipole Algorithm,” Computing in Science
& Engineering, vol. 2, no. 1, pp. 76–79, 2000.

[130] M. J. Stock and A. Gharakhani, “Graphics Processing Unit-Accelerated Boundary

Element Method and Vortex Particle Method,” Journal of Aerospace Computing,
Information, and Communication, vol. 8, no. 7, pp. 224–236, 2011.

[131] J. Choi, A. Chandramowlishwaran, K. Madduri, and R. Vuduc, “A CPU–GPU Hy-

brid Implementation and Model-driven Scheduling of the Fast Multipole Method,”
in Proceedings of Workshop on General Purpose Processing Using GPUs, ACM,
2014, p. 64.

[132] Q. Nie and G. Baker, “Application of Adaptive Quadrature to Axi-symmetric Vor-

tex Sheet Motion,” Journal of Computational Physics, vol. 143, no. 1, pp. 49–69,
1998.

[133] A. M. O. Smith and J. Pierce, “Exact Solution of the Neumann Problem: Calcula-
tion of Non-circulatory Plane and Axially Symmetric Flows about Or Within Arbi-
trary Boundaries,” Douglas Aircraft Co., Long Beach, CA, Tech. Rep. E.S. 26988,
Apr. 1958.

[134] A. M. O. Smith, “Incompressible Flow about Bodies of Arbitrary Shape,” in IAS

National Summer Meeting, 1962.

[135] J. L. Hess and A. M. O. Smith, “Calculation of Nonlifting Potential Flow about Ar-
bitrary Three-Dimensional Bodies,” Journal of Ship Research, vol. 8, no. 2, pp. 22–
24, 1964.

[136] J. L. Hess and A. M. O. Smith, “Calculation of Potential Flow About Arbitrary

Bodies,” Progress in Aerospace Sciences, vol. 8, pp. 1–138, 1967.

131
[137] J. L. Hess, “Calculation of Potential Flow About Arbitrary Three-Dimensional Lift-
ing Bodies,” McDonnell Douglass Corporation, Long Beach, CA, Tech. Rep. MDC
J5679-01, Oct. 1972.

[138] J. L. Hess, “Panel Methods in Computational Fluid Dynamics,” Annual Review of

Fluid Mechanics, vol. 22, no. 1, pp. 255–274, 1990.

[139] A. M. O. Smith, “The Panel Method: Its Original Development,” in Applied compu-
tational aerodynamics, American Institute of Aeronautics and Astronautics, vol. 1,
Washington, DC, 1990, pp. 3–17.

[140] O. D. Kellogg, Foundations of Potential Theory. Frederick Ungar Publishing Com-

pany, 1929.

[141] B. Maskew, “Program VSAERO Theory Document,” National Aeronautics and

Space Administration, Tech. Rep., 1987.

[142] F. T. Johnson, “A General Panel Method for the Analysis and Design of Arbitrary
Configurations in Incompressible Flows,” National Aeronautics and Space Admin-
istration, Contractor Report NASA-CR-3079, 1980.

[143] R. L. Carmichael and L. L. Erickson, “PANAIR—A Higher Order Panel Code for
Predicting Subsonic or Supersonic Linear Potential Flow About Arbitrary Config-
urations,” in 14th Fluid and Plasma Dynamics Conference, AIAA, Palo Alto, CA,
Jun. 1981.

[144] E. Martensen, “The Calculation of the Pressure Distribution on a Cascade of thick

Airfoils by means of Fredholm Integral Equations of the Second Kind,” National
Aeronautics and Space Administration, Tech. Rep., 1971.

[145] W Prager, “The Pressure Distribution on Bodies in Plane Potential Flow,” Physik.
Z, vol. 29, pp. 865–869, 1928.

[146] K. Jacob and F. W. Riegels, “The Calculation of the Pressure Distribution Over
Aerofoil Sections of Finite Thickness with and without Flaps and Slats,” Z. Flug-
wiss, vol. 11, no. 9, pp. 357–367, 1963.

[147] D. H. Wilkinson, “Numerical Solution of the Analysis and Design Problems for the
Flow Past One or More Aerofoils on Cascades,” Aeronautical Research Council,
Tech. Rep. R&M 3545, 1967.

[148] D. H. Wilkinson, “A Numerical Solution of the Analysis and Design Problems for
the Flow Past One or More Aerofoils or Cascades. Part 1- The Analysis Problem.
Part 2- The Design Problem (numerical Solution of Analysis and Design Problems

132
 for Flow Past One or More Airfoils or Cascades),” Aeronautical Research Council,
Tech. Rep., 1968.

[149] R. I. Lewis and P. G. Ryan, “Surface Vorticity Theory for Axisymmetric Potential
Flow Past Annular Aerofoils and Bodies of Revolution with Application to Ducted
Propellers and Cowls,” Journal of Mechanical Engineering Science, vol. 14, no. 4,
pp. 280–296, 1972.

[150] V. Hill, “A Surface Vorticity Theory for Propeller Ducts and Turbofan Engine
Cowls in Non-Axisymmetric Incompressible Flow,” Journal of Mechanical En-
gineering Science, vol. 20, no. 4, pp. 201–219, 1978.

[151] M. Drela, Flight Vehicle Aerodynamics. MIT Press, 2014.

[152] H. Schlichting, Boundary-Layer Theory, Seventh, F. J. Cerra, Ed. McGraw-Hill

Book Company, 1976.

[153] L. Prandtl and O. G. Tietjens, Fundamentals of Hydro- and Aeromechanics, H. W.

Craver, Ed. Dover, 1957.

[154] W. M. Kutta, “Uber eine mit den grundlagen des flug problems in beziehung ste-
hende zweidimensionale strömung,” 1910.

[155] J. L. Hess, “The Problem of Three-Dimensional Lifting Potential Flow and Its Solu-
tion by Means of Surface Singularity Distribution,” Computer Methods in Applied
Mechanics and Engineering, vol. 4, no. 3, pp. 283–319, 1974.

[156] J. Nocedal and S. J. Wright, Numerical Optimization, P. Glynn and S. M. Robinson,

Eds. Springer, 1999.

[157] K. W. McAlister and R. K. Takahashi, “NACA0015 Wing Pressure and Trailing

Vortex Measurements,” National Aeronautics and Space Administration, Ames Re-
search Center, Moffett Field, California, Technical Paper NASA TP-3151, Nov.
1991.

[158] R. Krasny and M. Nitsche, “The Onset of Chaos in Vortex Sheet Flow,” Journal of
Fluid Mechanics, vol. 454, pp. 47–69, 2002.

[159] M. Hand, D. Simms, L. Fingersh, D. Jager, J. Cotrell, S Schreck, and S. Larwood,

“Unsteady Aerodynamics Experiment Phase VI: Wind Tunnel Test Configurations
and Available Data Campaigns,” National Renewable Energy Laboratory, Golden,
CO, Technical Report NREL/TP-500-29955, 2001.

133
[160] C. E. Lynch, D. T. Prosser, and M. J. Smith, “An Efficient Actuating Blade Model
for Unsteady Rotating System Wake Simulations,” Computers & Fluids, vol. 92,
pp. 138–150, 2014.

[161] D. Simms, S. Schreck, M. Hand, and L. J. Fingersh, “NREL Unsteady Aerody-

namics Experiment in the Nasa-Ames Wind Tunnel: A Comparison of Predictions
to Measurements,” National Renewable Energy Laboratory, Golden, CO, Technical
Report NREL/TP-500-29494, Jun. 2001.

[162] A. S. Hahn, “Vehicle Sketch Pad: A Parametric Geometry Modeler for Conceptual
Aircraft Design,” in 48th AIAA Aerospace Sciences Meeting, AIAA, Orlando, FL,
Jan. 2010.

[163] K. Hayami, J.-F. Yin, and T. Ito, “GMRES Methods for Least Squares Problems,”
SIAM Journal on Matrix Analysis and Applications, vol. 31, no. 5, pp. 2400–2430,
2010.

[164] Y. Saad and M. H. Schultz, “GMRES: A Generalized Minimal Residual Algorithm

for Solving Nonsymmetric Linear Systems,” SIAM Journal on Scientific and Sta-
tistical Computing, vol. 7, no. 3, pp. 856–869, 1986.

[165] D. C.-L. Fong and M. Saunders, “LSMR: An Iterative Algorithm for Sparse Least-
Squares Problems,” SIAM Journal on Scientific Computing, vol. 33, no. 5, pp. 2950–
2971, 2011.

[166] M. J. Smith, “A Fourth Order Euler/navier–stokes Prediction Method for the Aero-
dynamics and Aeroelasticity of Hovering Rotor Blades,” PhD thesis, Georgia Insti-
tute of Technology, Feb. 1994.

[167] J. A. Blackwell, Jr., “Numerical Methods to Calculate the Induced Drag or Op-
timal Span Loading for Arbitrary Non-planer Aircraft,” Lockheed-Georgia Com-
pany, Tech. Rep., 1976.

[168] D. J. Pate and B. J. German, “Improved Computation of Induced Drag for Wakes
of Arbitrary Shape,” in AIAA Aviation, AIAA, Los Angeles, CA, Aug. 2013.

[169] J.-C. Suh, “The Evaluation of the Biot–Savart Integral,” Journal of Engineering
Mathematics, vol. 37, pp. 375–395, 2000.

[170] M. M. Munk, “Remarks on the Pressure Distribution over the Surface of an Ellip-
soid, Moving Translationally Through a Perfect Fluid,” National Advisory Com-
mittee for Aeronautics, Technical Note NACA-TN-196, Jun. 1924.

134
[171] D. B. Clarke, “Experimental and Computational Investigation of Flow about Low
Aspect Ratio Ellipsoids at Transcritical Reynolds Numbers,” PhD thesis, University
of Tasmania, 2009.

135