Académique Documents
Professionnel Documents
Culture Documents
by
Copyright
c 2012 by Charles Alexander Mader
Abstract
A set of techniques is developed that allows the incorporation of flight dynamics metrics
for including static stability constraints and handling qualities constraints in a high-fidelity
aerodynamic optimization are demonstrated. These constraints are developed from stability
derivative information calculated using high-fidelity computational fluid dynamics (CFD). Two
techniques are explored for computing the stability derivatives from CFD. One technique uses
full set of static and dynamic stability derivatives from a single steady solution. The other tech-
nique uses a linear regression method to compute the stability derivatives from a quasi-unsteady
time-spectral CFD solution, allowing for the computation of static, dynamic and transient sta-
bility derivatives. Based on the characteristics of the two methods, the time-spectral technique
is selected for further development, incorporated into an optimization framework, and used to
tion framework is then used to conduct an optimization study of a flying wing configuration.
This study shows that stability constraints have a significant impact on the optimal design of
flying wings and that, while static stability constraints can often be satisfied by modifying the
airfoil profiles of the wing, dynamic stability constraints can require a significant change in the
ii
Dedication
To my parents:
For nurturing my curiosity and giving me the tools to succeed
iii
Acknowledgements
The end product of a doctoral program is a thesis with a single name on it. However, the
creation of that thesis would not be possible without the support of many, many people. I
would like to take this opportunity to thank the various people who have supported me in this
endeavor.
First and foremost I would like to thank my supervisor, Professor Joaquim Martins. When
this project first started, it was, in many ways, out in left field. Professor Martins allowed me
the freedom to pursue the research where it led. While perhaps costly in the short term, I think
that in the end this proved very fruitful. Further, Professor Martins’ enthusiasm for research in
general and optimization research in particular provided a constant impetus pushing this work
along. Finally, Professor Martins’ advice and guidance has been invaluable over the course of
my degree. Thank you.
I would also like to thank Professor Zingg and Professor Damaren, my other committee
members, for their time and effort. Their insightful questions pushed me to better my under-
standing of various subjects and to fill in holes in certain areas of my research. This has led
to significantly better final product. I would also like to thank Professor Kyle Anderson, my
external examiner, for taking the time to review my thesis and provide constructive feedback
on my work.
I would like to express my gratitude to Dr. Ruben Perez, whose insights into aircraft design
and in particular, stability and control have been instrumental in my work.
Finally, I would like to thank all of my colleagues in the MDO lab. We have succeeded
in creating an excellent collaborative atmosphere that has been a pleasure to work in. The
impromptu brainstorming sessions and theoretical discussions are invigorating and the positive
energy in the lab helps keep things going on those occasional days when things don’t go exactly
as planned. Thanks Everyone!
Charles Alexander Mader
University of Toronto Institute for Aerospace Studies
February 2012
iv
Contents
1 Introduction 1
1.1 The Case for Unconventional Configurations . . . . . . . . . . . . . . . . . . . . . 2
1.2 The Need for Stability Constraints in High-Fidelity Optimization . . . . . . . . . 2
1.3 Overview of the Stability Constraint Approach . . . . . . . . . . . . . . . . . . . 3
1.4 Thesis Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Original Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
v
4.3.1 Single-Cell Routine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3.2 ∂R/∂ζ and ∂R/∂x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.3 ∂I/∂ζ and ∂I/∂x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.3.4 Solution of the Adjoint System . . . . . . . . . . . . . . . . . . . . . . . . 56
4.4 Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.4.1 Test case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.4.2 Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.4.3 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5 Auxilliary Analyses 64
5.1 Center of Gravity Calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.2 Moment of Inertia Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3 Root Bending Moment Constraint . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.4 Stability Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4.1 Static Longitudinal Stability Definition . . . . . . . . . . . . . . . . . . . 70
5.4.2 Static Margin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.4.3 Control Anticipation Parameter . . . . . . . . . . . . . . . . . . . . . . . . 72
5.5 Geometry Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
7 Conclusions 116
vi
8 Future Work 118
References 120
Appendix 129
vii
List of Tables
3.1 Comparison of lift and moment coefficients for the NACA 0012 at α = 0 for
various values of q̂ (results from Limache [55] in parentheses) . . . . . . . . . . . 22
3.2 Sensitivity verification: NACA 0012 test case, Mach = 0.5, 131,072 cells: 10−12
relative convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 ADjoint stability derivatives for a NACA 0012 airfoil at α = 0.0 degrees . . . . . 25
3.4 Effect of point location on time-spectral derivative location . . . . . . . . . . . . 36
3.5 SACCON test case conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.1 Sensitivity verification: ONERA M6, pitching motion, 917,000 cells: 10−12 rela-
tive convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2 Sensitivity verification: ONERA M6, plunging motion, 917,000 cells: 10−12 rel-
ative convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.3 Time-spectral ADjoint computational cost breakdown for ONERA M6 (normal-
ized with respect to a total flow solution cost of 160.3 seconds) . . . . . . . . . . 61
B.1 NACA 0012 wing: planform only optimization results summary: 1107k cells,
M = 0.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
B.2 NACA 0012 wing: shape optimization results summary: 1107k cells, M = 0.5 . . 135
viii
B.3 NACA 0012 wing: planform only optimization results summary: 1107k cells,
M = 0.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
B.4 NACA 0012 wing: shape optimization results summary: 1107k cells, M = 0.7 . . 137
B.5 NACA 0012 wing: shape optimization results summary: 1107k cells, M = 0.85 . 138
ix
D.25 FFD coordinates for CAP constrained case with shape variables, M = 0.7, part 1 170
D.26 FFD coordinates for CAP constrained case with shape variables, M = 0.7, part 2 171
D.27 FFD coordinates for bending moment constrained case with shape variables,
M = 0.85, part 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
D.28 FFD coordinates for bending moment constrained case with shape variables,
M = 0.85, part 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
D.29 FFD coordinates for Cmα constrained case with shape variables, M = 0.85, part 1174
D.30 FFD coordinates for Cmα constrained case with shape variables, M = 0.85, part 2175
D.31 FFD coordinates for Kn constrained case with shape variables, M = 0.85, part 1 176
D.32 FFD coordinates for Kn constrained case with shape variables, M = 0.85, part 2 177
D.33 FFD coordinates for CAP constrained case with shape variables, M = 0.85, part 1178
D.34 FFD coordinates for CAP constrained case with shape variables, M = 0.85, part 2179
x
List of Figures
xi
5.2 Discrete surface mesh with lumped masses at the cell centroids . . . . . . . . . . 66
5.3 Cell area calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.4 Bending moment calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.5 Geometry manipulation setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.6 Deformed geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
xii
6.26 Kn constrained optimization: planform variables only: M=0.7 . . . . . . . . . . . 103
6.27 Kn constrained optimization: planform and shape variables: M=0.7 . . . . . . . 104
6.28 Kn constrained optimization: planform and shape variables: M=0.85 . . . . . . . 105
6.29 XDSM for the CAP constrained problem . . . . . . . . . . . . . . . . . . . . . . . 106
6.30 CAP constrained optimization: planform variables only: M=0.5, e = 0.942 . . . . 107
6.31 CAP constrained optimization: planform and shape variables: M=0.5, e = 0.970 108
6.32 CAP constrained optimization: planform variables only: M=0.7 . . . . . . . . . . 108
6.33 CAP constrained optimization: planform and shape variables: M=0.7 . . . . . . 109
6.34 CAP constrained optimization: planform and shape variables: M=0.85 . . . . . . 110
xiii
Nomenclature
Introduction:
Cp coefficient of pressure
Chapter 2:
g gravitational acceleration
Ii moment of inertia components
L, Li roll moment in the body frame, stability derivative with respect to parameter i
m mass
M , Mi pitch moment in the body frame, stability derivative with respect to parameter i
N , Ni yaw moment in the body frame, stability derivative with respect to parameter i
p, ṗ roll rate in the body frame, time derivative of p
q, q̇ pitch rate in the body frame, time derivative of q
r, ṙ yaw rate in the body frame, time derivative of r
u, u̇ velocity in x direction in the body frame, time derivative of u
v, v̇ velocity in y direction in the body frame, time derivative of v
w, ẇ velocity in z direction in the body frame, time derivative of w
X, Xi force in x direction in the body frame, stability derivative with respect to parameter i
Y , Yi force in y direction in the body frame, stability derivative with respect to parameter i
Z, Zi force in z direction in the body frame, stability derivative with respect to parameter i
α angle of attack
ω short-period frequency
φ vehicle Euler angle
ψ vehicle Euler angle
θ vehicle Euler angle
ζ short-period damping ratio
Chapter 3:
b span
c chord
Dt spectral derivative operator
et total energy
f generic function
xiv
fi flux term in direction i
Fi force in the ith coordinate direction
h step length
√
i −1
I identity matrix (Section 3.3.1), function of interest (Section 3.3.2)
k thermal conductivity of fluid (Section 3.3.1)
k frequency index (Section 3.4.1)
k reduced frequency (Section 3.4.3)
l time instance index
Mi moment about the ith coordinate axis
n time instance index
N number of spectral time intervals
p pressure (Section 3.3.1), roll rate (Section 3.3.2)
q pitch rate
r yaw rate
s source terms
t intermediate variable in automatic differentiation (Section 3.2), time
T temperature (Section 3.3.1), time period of periodic flow problem (Section 3.4.1)
u fluid velocity in fixed frame
v velocity of fluid with respect to moving grid
V aircraft speed, cell volume (Section 3.4.1)
w velocity of the moving grid
x grid coordinates (Section 3.3.1), design variables (Sections 3.3.2, 3.2)
R flow residual
p̂ normalized roll rate (pb/2V )
q̂ normalized pitch rate (qc/2V )
r̂ normalized yaw rate (rb/2V )
α angle of attack
β side slip angle
δi ith control surface deflection (Section 3.3.2)
δi elements of the identity matrix (Section 3.3.1)
Ψ adjoint vector
γ specific heat ratio
Ω rotational rate of the moving grid
ρ density
σ viscous stress tensor
xv
τ generalized time variable
ζ flow states for entire mesh
ζcell flow states for a single cell
CD aircraft drag coefficient
CL aircraft lift coefficient
Cl airfoil lift coefficient (Section 3.3.1), aircraft roll moment coefficient (Section 3.4.2)
Cm airfoil pitch moment coefficient (Section 3.3.1)
Cm aircraft pitch moment coefficient (Section 3.4.2)
Cn aircraft yaw moment coefficient
Cp coefficient of pressure
CY aircraft side force coefficient
Ci0 initial value of coefficient Ci
Cij stability derivatives (derivative of Ci with respect to j)
Chapter 4:
xvi
Chapter 5:
A area
c̄ref reference chord length
CG% location of the center of gravity expressed as a percentage of the distance
between the two wing spars, referenced from the leading edge spar.
g gravitational acceleration
hCG location of the center of gravity normalized by the MAC
hNP location of the neutral point normalized by the MAC
I second moment of area
Izz pitch moment of inertia
Kn static margin
L generic length scale
m mass
M Mach number
M bending moment
∆n normal acceleration of the aircraft
nα change in normal acceleration of the aircraft with α
q dynamic pressure
q̇ pitch acceleration of the aircraft
r distance
s cell area
t thickness
V velocity
x, y, zCG center of gravity location
x, y, zmass point mass location
xbendingi root location of bending axis
xrefi moment coefficient reference point
y beam height for stress computation
α angle of attack
Ω domain of integration
ρ density
σ bending stress
ωn short-period natural frequency
Cb total coefficient of bending
xvii
Cbendingx coefficient of bending moment about x axis
Cbendingz coefficient of bending moment about z axis
CL coefficient of lift
Cfx force coefficients in the x direction
Cfy force coefficients in the y direction
Cfz force coefficients in the z direction
CLi stability derivatives (derivative of CL with respect to i
CmCG moment coefficient about the CG location
CmNP moment coefficient about the NP location
Cmx moment coefficients about the x axis
Cmy moment coefficients about the y axis
Cmz moment coefficients about the z axis
Cmi stability derivatives (derivative of Cm with respect to i)
Chapter 6:
A area
b span
croot , ctroot root chord length, target root chord length
CG% location of the center of gravity expressed as a percentage of the distance
between the two wing spars, referenced from the leading edge spar.
e span efficiency factor
Kn static margin
M Mach number
MACt target mean aerodynamic chord
PIzz multiplication factor on computed Izz
xCG , xtCG location of the CG, target location of the CG
y coupling variables
α angle of attack
λ sweep
θi section twist at station i
∆FFD change in FFD volume control points
ζsp short-period damping ratio
Cb total coefficient of bending
t
CD , CD drag coefficient, target drag coefficient
Cl section lift coefficient
xviii
CL lift coefficient
Cm , Cmz pitch moment coefficient
CmCG moment coefficient about the CG location
CmNP moment coefficient about the NP location
Cp coefficient of pressure
CLi ,CLt i stability derivatives (derivative of CL with respect to i), target stability derivatives
Cmi ,Cm t stability derivatives (derivative of Cm with respect to i), target stability derivatives
i
Abbreviations:
AD Automatic differentiation
BLISS Bi-level integrated system synthesis
BWB Blended-wing-body
CAP Control anticipation parameter
CFD Computational fluid dynamics
CG Center of gravity
CO Collaborative optimization
CSSO Concurrent subspace optimization
DATCOM United States air force data compendium
DES Discrete eddy simulation
ESDU Engineering sciences data unit
FFT Fast Fourier transform
FFD Free form deformation
ICAO International civil aviation organization
IDF Individual discipline feasible
MAC Mean aerodynamic chord
MDF Multidisciplinary feasible
MDO Multidisciplinary design optimization
NACA National advisory committee for aeronautics
NATO RTO North Atlantic treaty organization research and technology organization
NLFD Nonlinear frequency domain
NP Neutral point
ONERA Office National d’Études et de Recherches Aérospatiales
RANS Reynolds averaged Navier-Stokes
SACCON Stability And Control CONfiguration
SAND Simultaneous analysis and design
xix
UCAV Unmanned combat aerial vehicle
XDSM Extended design structure matrix
xx
Chapter 1
Introduction
Demand for aircraft transportation is increasing at a tremendous rate around the world. Ac-
cording to Boeing and Airbus projections, global airline passenger traffic will increase at an
average annual rate of close to 5% for the next 10 to 20 years [2, 12]. In particular, Boeing
expects there to be over 35,000 aircraft in passenger service by 2028, of which 29,000 will have
been newly constructed between 2008 and 2028 [12]. At the same time, the public is demanding
increased accountability regarding the environmental impact of all carbon consuming industries,
air transportation included. To put aviation’s place in this area into perspective, at 2006 levels,
aviation contributed approximately 2% of global CO2 emissions [45]. While this contribution
is relatively minor at 2006 levels, ICAO projects that with the aforementioned growth in air
traffic demand, aviation CO2 emissions will grow from the 2006 level of 632 Mega tonnes(Mt)
to between 1,422 and 1,738 Mt in 2036 [45]. This is more than a factor of two increase from
2006 levels.
In an attempt to mitigate the effects of this growth, the commercial aviation industry is
investigating a variety of possibilities to reduce its environmental footprint. Some of these
possibilities include the use of alternative fuels, improvements to operational practices and
technical innovation to improve the efficiency of the world fleet. While each one of these areas
promises to provide significant efficiency gains, the magnitude of the expected increase in traffic
is such that no one field on its own will be able to offset all of the increase. Therefore, all possible
avenues of improvement must be explored.
This thesis relates primarily to the improvement in efficiency of the aircraft itself. Two main
areas will be considered, the use of unconventional designs to improve the efficiency of aircraft
and the use of design optimization to facilitate the design of these unconventional concepts. In
particular, this thesis will focus on the design and optimization of flying wing configurations
and show that high-fidelity aerodynamic shape optimization can be used to produce stable and
flyable, flying wing designs.
1
Chapter 1. Introduction 2
The aviation industry’s track record for technological improvement is impressive. According to
Lee [52], average fuel consumption per revenue passenger mile improved by approximately 70%
between 1959 and 1995. This improvement has come from a variety of new technologies — such
as high-bypass ratio turbofans, supercritical airfoils and improved structural materials — as well
as improvements in design and analysis methods. While these improvements are impressive, the
consensus in the design community seems to be that most of the benefit available from these
improved technologies and design methods has been realized. Given the expected increase in
demand for air travel, the incremental improvement still available from these approaches will
not be sufficient to achieve the performance targets set for the industry. In response to this,
a variety of alternative technologies are being explored, with the goal of finding solutions that
will provide significant improvements in efficiency.
One of the major ideas being advanced as a solution to this issue is the use of unconven-
tional airframe designs. For each of these alternative concepts — such as the blended-wing-
body (BWB) [53], the joined wing [113, 28] or the strut-braced wing [36, 29] — there is some
physical characteristic that has been shown, at least on paper, to provide significant efficiency
improvements over conventional aircraft, for a given technology level. For example, Liebeck [53]
demonstrates a blended-wing-body design that, through aerodynamic and structural improve-
ments, has approximately 30% lower fuel burn per passenger than an equivalently sized A380
aircraft. Therefore, while these configurations may be relatively untried, they certainly deserve
to be developed further in order to identify what is needed to take advantage of their potential.
While the blended-wing-body promises improved aerodynamic and structural performance [53,
13, 90, 85], the improved efficiency does not come for free. Because of the unconventional
layout of the passenger cabin, the configuration raises questions about passenger comfort, mar-
ket acceptance and emergency egress that would have to be addressed in any commercialized
design [53, 13]. Also, because the configuration is essentially a single, large, lifting surface, it
introduces additional stability and control considerations into the design problem [103, 53, 109].
Further, there is a strong coupling between the aerodynamic performance of the configuration
and its trim and stability characteristics. For example, Liebeck notes that there is a conflict
between the high aft camber characteristics common to the class of transonic airfoils needed
for good high speed performance and the need to maintain a low pitch moment airfoil in order
Chapter 1. Introduction 3
to be able to trim a BWB at cruise [53]. More specifically, transonic airfoils are designed to
limit the peak Cp values generated by the airfoil to a value below the critical Cp that will cause
shock waves to form on the wing. This tends to force the generation of lift further back on the
airfoil. Low moment airfoils are trying to balance the lift generation fore and aft of the quarter
chord of the foil. Given the ratio of area fore and aft of this point, this fundamentally requires
relatively large Cp values forward of the quarter chord. This effect is clearly demonstrated in
the Mach = 0.85 cases in Chapter 6. While multidisciplinary coupling such as this is inher-
ent in all aircraft design, issues such as these tend to become significantly more restrictive in
unconventional configurations.
To further complicate matters, by their very nature, unconventional designs are well outside
the typical range of experience for most designers. As a result, there is not a large body
of experiential knowledge to be drawn upon in the design of such aircraft. In response to
this, the design of these configurations tends to gravitate towards the use of higher fidelity
analysis methods and more powerful design techniques such as design optimization, as these
methods allow for accurate and efficient exploration of the design space. This is also true
of this thesis. The core analysis involved in this work is a high-fidelity computational fluid
dynamics (CFD) solver, capable of solving Euler and RANS flows for steady, unsteady and time-
spectral problems. This solver is used to model the aerodynamic performance of the various
configurations considered. However, as shown by Martins et al. [64], simply having a highly
accurate single discipline optimization is not sufficient to achieve a good design. Optimization
of a single discipline on its own often leads to a suboptimal design of the full multidisciplinary
system. Indeed, as will be shown in Chapter 6, in the context of flying wing design, pure
aerodynamic shape optimization can lead to unstable designs if stability characteristics are not
considered. Thus, the goal of this work is to develop techniques that can be used to calculate
stability and control information from the high-fidelity CFD solver that can then be used in
conjunction with high-fidelity aerodynamic shape optimization techniques.
To compute the stability and control metrics required for the aircraft, and to keep the cost of
analysis reasonable for numerical optimization, linear flight dynamics will be used to model the
stability of the aircraft. This linear flight dynamics model will be used to formulate a variety
of flight dynamics constraints, which will ensure that the optimization produces a stable design
with the desired handling characteristics. While not as accurate as a fully nonlinear flight
dynamic model, this approach will provide sufficient information to examine the effects of
stability considerations on the optimal design of flying wing configurations.
Chapter 1. Introduction 4
Linear flight dynamics requires information about three characteristics of the aircraft: how
its aerodynamic characteristics change with the dynamic states (stability derivatives), how much
mass it has (mass) and how its mass is distributed (moments of inertia). A primary focus of this
work is examining methods for the computation of stability derivatives and their sensitivities
for use in optimization. Two approaches, both based on three-dimensional CFD, are introduced
and discussed. The first approach is based on the automatic differentiation adjoint (ADjoint)
sensitivity analysis method developed by Mader et al. [60]. This method is very efficient at cal-
culating the full suite of static (α, β, V ) and dynamic (p, q, r) derivatives, but can not calculate
the transient (α̇, q̇, etc.) derivatives. It is also not suited to large scale optimization, as the
cost of computing the gradients of the stability derivatives with respect to the design variables
scales with the number of design variables. Therefore, a second approach, based on unsteady,
forced oscillation techniques was developed. This method uses a time-spectral CFD method to
compute a forced oscillation solution, which is then used to generate the stability derivatives
algebraically. This whole system can then be differentiated using the ADjoint technique to
provide the sensitivities required for gradient-based optimization. Furthermore, because the
stability derivatives are based on an unsteady solution, estimation of the “dot” derivatives is
possible. The inclusion of the “dot” derivatives allows for the computation of handling qualities
estimates, which are not possible with only the static and dynamic derivatives.
The second two characteristics — mass and moment of inertia — receive a rather cursory
treatment in this work. The calculation of these parameters are based on relatively simple,
preliminary and conceptual level methods as described in Chapter 5. The author acknowledges
that the results of this work would benefit from more rigorous treatment of these calculations.
However, as the focus of this work was the aerodynamic side of the problem, that portion of
the analysis was left for future research.
The work presented in this thesis is laid out in five parts. Chapter 2 introduces the basics of
the linear flight dynamics models used to generate the dynamic stability constraints for the
optimization problems. Chapter 3 focuses on the computation of the stability derivatives, with
sections on both the ADjoint approach and the time-spectral approach. Each approach is intro-
duced, discussed and verified against benchmark results. Chapter 4 introduces and discusses the
sensitivity analysis methods used to compute the sensitivities required for gradient-based opti-
mization. In particular, it introduces and discusses the time-spectral ADjoint method, which
is a key element in enabling efficient stability-constrained optimization. Chapter 5 outlines the
simple methods used to calculate the center of gravity (CG) location and moment of inertia of
Chapter 1. Introduction 5
the aircraft, as well as the specific methods used to implement the various stability constraints
in the optimization. Finally, Chapter 6 presents the optimization results obtained during the
study.
Previous work has shown that the inclusion of flight dynamics and control in MDO can have a
significant effect on the optimal design of an aircraft [88, 40]. While inclusion of stability and
control information has been accomplished at the low-fidelity level [110, 88, 40], the effects of
this discipline have not been explored at the high-fidelity level. The work in this thesis seeks
to start addressing this deficiency.
Since a linear flight dynamic model has been selected to describe the flight dynamics of the
aircraft, the question of computing the aerodynamic characteristics of the aircraft for the stabil-
ity discipline is reduced to one of computing the stability derivatives for the aircraft. Therefore,
the first step in addressing this deficiency is developing a method for the computation of aircraft
stability derivatives that is compatible with CFD based aerodynamic optimization. Several re-
searchers — for example, Park et al. [83, 82], Murman [74] and Limache and Cliff [54] — have
explored the use of CFD in the computation of stability derivatives. However, all of these
methods focus strictly on the computation of the stability derivatives and make no considera-
tion of the implications associated with using these methods in optimization. The work in this
thesis contributes to this area in two ways. First, the ADjoint approach to the computation
of stability derivatives is introduced and demonstrated. This method exhibits extremely good
numerically accuracy and is an efficient method for computing full sets of static and dynamic
stability derivatives from a single CFD solution. Further, because it is based on an adjoint
method, the method is well suited to the computation of derivatives in cases involving large
numbers of control effectors. However, because this method is itself based on advanced sensi-
tivity analysis techniques, it is not well suited to use in optimization. For this task, the second
major contribution of this work — the time-spectral stability derivative method — is much more
useful. Like the methods of Murman [74] and DaRonche et al. [94], the time-spectral stability
derivative approach is based on a forced oscillation technique. However, it has some unique
qualities. First, two unique motions are used that allow the α derivatives and q derivatives
to be computed independently. This allows the α̇ derivatives to be computed independently
of the q derivatives, which is useful in the assembly of the flight dynamic constraints. Also,
because this method uses relatively simple algebraic operations on top of a periodic unsteady
solution, it is relatively straight forward to apply an adjoint technique to the problem to obtain
the sensitivities necessary for gradient-based optimization. This leads to the third contribution
Chapter 1. Introduction 6
in this thesis, the time-spectral ADjoint. Previous work has demonstrated the use of adjoint
methods for spectral flow solution techniques. Thomas et al. [104] have demonstrated an ad-
joint technique for the two dimensional harmonic balance technique, Nadarajah and Jameson
have implemented an adjoint for the Nonlinear Frequency Domain method [76] and Choi et
al. [20] have demonstrated an adjoint method for the time-spectral equations. However, the
time-spectral adjoint method of Choi et al. contained approximations for the derivatives of the
spectral radius and artificial dissipation computations, thereby limiting its numerical accuracy.
The time-spectral ADjoint approach introduced in this work avoids those approximations by
using automatic differentiation to compute the components of the adjoint system. Further,
because the time-spectral equations are formulated entirely in the real domain, the complex
step method can be used to verify the sensitivities. As is shown in Section 4.4, this allows
for verification of the derivatives to extremely small tolerances. These two techniques — the
time-spectral stability derivative method and the time-spectral ADjoint — can be combined
to enable the final major contribution of this thesis, the application of stability constraints in
high-fidelity aerodynamic shape optimization. As is shown in Chapter 6, the inclusion of stabil-
ity constraints has a significant impact on the optimal shape for an aircraft across a wide range
of Mach numbers. To the best of the author’s knowledge, this is the first time this capability
has been demonstrated in the context of high-fidelity aerodynamic shape optimization.
Chapter 2
The fundamental idea pursued in this thesis is that, in order to properly optimize an aircraft,
one needs to account for both the stability characteristics and handling characteristics of the
aircraft during the optimization process. As shown by the results in Chapter 6, without this
information, the optimization produces an aerodynamically optimal design that may or may
not be stable and flyable. Such a design would then require follow-up modifications to achieve
the desired level of stability and handling, yielding a suboptimal design. This is of particular
importance with unconventional designs, such as a BWB or a flying wing, because of the strong
coupling between aerodynamic performance, trim and stability in these designs. However, while
this is a simple enough goal to describe, it actually turns out to be challenging to implement
in practice.
To provide a basis for the following discussion, consider the equations of motion for an
aircraft, as written by Etkin [27]:
X − sin θ (u̇ + wq − rv)
Y + mg sin φ cos θ = m (v̇ + ru − pw) (2.1)
Z cos φ cos θ (ẇ + pv − qu)
L Ix −Ixy −Izx ṗ 0 −r q Ix −Ixy −Izx p
M = −Ixy Iy −Iyz q̇ + r
0 −p −Ixy
Iy −Iyz q
(2.2)
N −Izx −Iyz Iz ṙ −q p 0 −Izx −Iyz Iz r
φ̇ 1 sin φ tan θ cos φ tan θ p
θ̇ = 0 cos φ − sin φ (2.3)
q
ψ̇ 0 sin φ sec θ cos φ sec θ r
These equations assume that the aircraft is rigid, that the effect of spinning rotors is negligible,
that the mass of the aircraft is constant and that the aircraft is flying through a still air mass.
7
Chapter 2. Linear Flight Dynamics 8
Note that the I variables represent the various aircraft moments of inertia, m represents the
aircraft mass and φ, ψ, and θ are the vehicle Euler angles.
As can be seen from these equations, the relations governing the motion of an aircraft are
quite complex. Both sets of equations (2.1) and (2.2) are fully nonlinear. In addition, the forces
— X, Y , Z — and moments — L, M , N — are dependent on the current as well as the past
values of the motion variables in the equations, namely the velocities — u, v and w — and the
rotational rates — p, q, and r. Therefore, to accurately evaluate this system, one would need
to run a time accurate evolution of the motion equations, evaluating the forces on the aircraft
at each instance in time. Given that computationally intensive CFD methods are being used
to compute the forces on the aircraft, this would be prohibitively expensive in the context of
an optimization problem.
A common methodology for reducing the complexity of this system of equations is to formu-
late a linearized form of the equations. This is accomplished in two stages. First, using small
disturbance theory and assuming small angle theory to be valid, equations (2.1) through (2.3)
can be rewritten as:
X0 + ∆X − sin θ0 − ∆θ cos θ0 ∆u̇ + w0 ∆q − v0 ∆r
Y0 + ∆Y + mg ∆φ cos θ 0
= m ∆v̇ + u0 ∆r − w0 ∆p) (2.4)
Z0 + ∆Z cos θ0 − ∆θ0 ∆ẇ + v0 ∆p − u0 ∆q)
L0 + ∆L Ix −Ixy −Izx ∆ṗ
M0 + ∆M = −Ixy Iy −Iyz ∆q̇
(2.5)
N0 + ∆N −Izx −Iyz Iz ∆ṙ
∆φ̇ 1 0 tan θ0 ∆p
∆θ̇ = 0 1 0 ∆q . (2.6)
∆ψ̇ 0 0 sec θ0 ∆r
where values with the subscript 0 represent initial values. One can then express the linearized
forces as a sum of their derivatives with respect to the motion variables:
∆X Xi ∆i Xj ∆j XControl ∆Control
∆Y Yi ∆i Yj ∆j YControl ∆Control
∆Z Zi ∆i Zj ∆j ZControl ∆Control
= + + (2.7)
∆L Li ∆i Lj ∆j LControl ∆Control
∆M M ∆i M ∆j M
i j Control ∆Control
∆N Ni ∆i Nj ∆j NControl ∆Control
After substituting in the initial conditions, this allows the flight dynamic system to be
rewritten as a state-space system of the form:
ẋ = Ax + bu. (2.8)
This is a set of coupled ordinary differential equations (ODEs) in state-space form, which can
be solved in a standard fashion. The full derivation of this system is shown in Appendix A.
If one considers the aircraft to be symmetric about the mid-plane, both the cross moment of
inertia elements — Ixy and Iyz — and the change in lateral forces with respect the longitudinal
degrees of freedom are zero. Further, one can also assume that the change in longitudinal
forces and moments with respect to lateral motions is small enough to neglect. This allows the
longitudinal degrees of freedom to be separated from the lateral degrees of freedom.
Continuing with the derivation above, by setting Ixy and Iyz to zero and removing the cross
dependence of the forces and moments, one can write the longitudinal set of equations as:
(Xu̇ − m) Xẇ Xq̇ 0 ∆u̇
Zu̇ Zẇ − m Zq̇ 0 ∆ẇ
∆q̇ +
−Iy + Mq̇
Mu̇ Mẇ 0
0 0 0 1 ∆θ̇
Xu Xw (Xq − mw0 ) −mg cos θ0 ∆u XControl
Zu Zw muo + Zq −mg sin θ0 ∆w ZControl
∆q + M ∆Control = 0. (2.9)
Mu Mw Mq 0 Control
0 0 −1 0 ∆θ 0
These equations can be solved for the time derivative terms to yield:
−(Xq −mw0 )
−Xu −Xw
g cos(θ0 )
∆u̇ m m m ∆u
Zu Zw (Zq +mu0 ) (gm sin(θ0 ))
∆ẇ (Zẇ −m) (Zẇ −m) (Zẇ −m)
− (Z −m) ∆w
∆q̇ =
(Mẇ Zu )
−M u (Mẇ Zw )
−M w (Mẇ (Zq +mu0 )) M
− Iq
ẇ
(M gm sin(θ0 ))
− (I ẇ∗(Z −m))
+
∆q
(Iy (Zẇ −m)) Iy (Iy (Zẇ −m)) Iy (Iy (Zẇ −m)) y y ẇ
∆θ̇ 0 0 −1 0 ∆θ
XControl
ZControl
M ∆Control. (2.10)
Control
0
Again, this is an equation in standard state-space form, which can be solved using standard
differential equation solution techniques.
Chapter 2. Linear Flight Dynamics 10
While in general, one could solve the full flight dynamics model to determine stability and
control characteristics for an aircraft, for this work, a simplified short-period approximation is
used. This approximation assumes that the variation in forward velocity (u) is negligible and
that the short-period characteristics of the aircraft can be modeled by changes in pitch rate (q)
and angle of attack (α) alone. In this case, vertical velocity (w) is used in place of the angle of
attack where:
w
tan α =. (2.11)
u
To facilitate this approximation, the second and third equations are extracted from equa-
tions (2.10). Further, it is assumed, as per McRuer et al. [71] that:
Ultimately, as will be seen in Chapter 5, the handling qualities analysis requires the fre-
quency and damping ratio of the system. These values can be determined from the characteristic
equation of the 2 × 2 system in equation (2.13). The characteristic equation of this system can
be calculated as det(sI − A) or:
s − −Z w
−u0
m
−Mẇ Zw − Mw
−Mẇ u0 Mq
(2.14)
Iy m Iy s− Iy − Iy
s2 + 2ζωs + ω 2 (2.19)
yields: q
ω= Zˆw M̂q − Mˆw uo (2.20)
or
M̂ẇ u0 + M̂q + Zˆw
ζ= . (2.22)
2ω
where ζ is the short-period damping ratio.
2.3 Summary
By making the simplifications demonstrated above, the complex, nonlinear flight dynamics
problem is reduced to a set of coupled, ordinary differential equations which are much simpler to
solve. Further, rather than having to compute time dependent forces, the forces are represented
by a set of aerodynamic derivatives, typically called stability derivatives [27], which can be
computed about a single operating point. Using these stability derivatives as a basis for the
model, simple calculations can then be performed to find the frequency and damping ratio of
the system, which can in turn be used to quantify the handling qualities of the aircraft. Now
the key part is to find an efficient way of computing stability derivatives for use in optimization,
which is the subject of Chapter 3 and one of the goals of this thesis.
Chapter 3
Having reduced the aerodynamic problem from one of computing a full set of time accurate
forces for a given maneuver to that of computing the stability derivatives about a given flight
condition, one can now examine the methods available for computing these derivatives. His-
torically, these derivatives have been estimated using empirical methods (e.g., DATCOM [106],
ESDU [26]), wind tunnel testing, or lower fidelity CFD methods such as panel codes. While
each of these methods is useful, they each have their own disadvantages with respect to various
factors such as cost, range of applicability, and computational time. Given that this work is
focused on conducting CFD-based aerodynamic shape optimization, it is necessary to be able
to compute accurate stability derivatives over the same range of flight conditions for which the
flow is being analysed. This leads to the idea of computing the stability derivatives with the
same high-fidelity methods that are being used to compute the flow.
3.1 Background
Over the last 15-20 years, several research groups have examined the idea of computing sta-
bility derivatives using high-fidelity CFD. Some early work in this area was conducted by
Charlton [19], who conducted simple α and β sweeps to get the force and moment information
required for falling-leaf predictions for tailless aircraft. The study concluded that, in most in-
stances, the required stability data could be computed accurately, but that highly nonlinear
regions of the flow required larger test cases and longer run-times than desired. It is these
highly nonlinear cases which are often critical in aircraft design. While not explored in detail in
this work, the methods proposed here extend to those flow regimes, provided the CFD methods
used provide accurate solutions in those cases.
Godfrey and Cliff [33] explored the use of analytic sensitivity methods — in particular, the
direct method — for the computation of static stability derivatives. They show the computation
12
Chapter 3. CFD-Based Stability Derivatives 13
of the α and β derivatives of the YB-48 flying wing using a three dimensional inviscid flow solver.
Their results match existing semi-empirical methods to between 5% and 30%, depending on the
derivative. However, no comment is made regarding the computational efficiency of the method.
Limache and Cliff [54] followed up this work by examining the use of the same method for the
computation of dynamic stability derivatives. Limache and Cliff show the computation of the
dynamic pitching derivatives of an airfoil using a two-dimensional Euler CFD code. This study
shows the significant promise of analytic sensitivity methods for the computation of stability
derivatives, showing that multiple stability derivatives — including the rotational derivatives
— can be computed from a single steady solution. The study also highlights the importance of
using, at the minimum, Euler CFD to capture the shock waves seen in transonic flow.
Another approach that has been pursued for the computation of stability derivatives is
the use of automatic differentiation (AD). Park et al. [83, 82] applied ADIFOR [8] — an
automatic differentiation (AD) tool — to a three-dimensional viscous flow solver to compute
the static and dynamic derivatives of various configurations. The results from this work showed
promise, providing accurate results across a variety of flight conditions. However, the additional
computational cost of running the differentiated code significantly impacted the computational
efficiency of the approach, leading to an eight-fold increase in the computational time required
to compute the derivatives when compared to a single flow solution.
The previous two approaches to dynamic derivative computation — those of Limache and
Park — rely on a non-inertial reference frame CFD formulation to handle the rotational rates
needed for the dynamic derivatives. Babcock and Arena [5] handle the dynamic derivatives
by modifying the boundary conditions in a finite element based Euler CFD solver to separate
the velocity and position boundary conditions. With this approach they are able to perturb
the static states (α, β) and the dynamic states (p,q,r) independently to determine the stability
derivatives using finite differences. The results from this approach compare relatively well with
theoretical, empirical and experimental results, confirming the validity of the method. However,
no computational cost details are included in the study.
A separate approach for computing the dynamic derivatives — one that has been used in the
experimental community for many years — is the forced oscillation approach. In this approach,
the test aircraft is driven thorugh a prescribed oscillatory motion. The resulting oscillatory
aerodynamic data can then be used to determine various characteristics of the aircraft. A
number of papers from the recent NATO RTO Task Group AVT-161 have explored the use
of this technique in conjuction with a variety of CFD solution methods [72]. In particular,
it has been demonstrated with RANS [96, 95], DES [22] and harmonic balance [94] solution
techniques. These techniques were shown to correlate well with experimental data.
In an independent study, Murman [74] also presents a method for computing stability deriva-
Chapter 3. CFD-Based Stability Derivatives 14
tives based on the traditional forced oscillation approach. In this work, Murman uses a fre-
quency domain method to produce periodic data for the forced oscillation of the configuration
of interest. This data can then be analyzed with the same techniques used to produce stability
derivatives from forced oscillation wind tunnel data, which allows the method to take advantage
of the large body of knowledge that has been developed in that field. The study demonstrates
good accuracy for a variety of configurations and flight conditions, however because of the time
periodic nature of the solution, the computational cost is higher than would be the case for an
equivalent steady-state solution of the same configuration.
In the following chapter, two approaches for computing stability derivatives using CFD
are introduced. The first, based on adjoint sensitivity methods, is an efficient method for
calculating static and dynamic stability derivatives about a single operating point. The second
approach, based on time-spectral CFD, includes unsteady information, thereby allowing for
the computation of the “dot” derivatives. It is also more conducive to optimization than the
ADjoint approach. The details of these two methods are explained in the following sections.
Before proceeding to desribe the methods developed for computing stability derivatives from
CFD, it is worth taking a brief aside to talk about sensitivity analysis for CFD in general.
Sensitivity analysis is typically performed using one of four main methods: finite-difference,
complex-step, analytic methods or automatic differentiation. The characteristics of each method
are as follows:
Finite-Difference: Finite-difference methods [15] are probably the most common methods
for calculating derivatives. The method comes in many forms, but the basic form of the method
can be expressed as:
df (x) f (x + h) − f (x)
= . (3.1)
dx h
this method is popular because it requires no specialized knowledge of the code being differen-
tiated. However, the method is subject to subtractive cancelation errors, which limit its overall
accuracy [102]. Further, the basic forward method shown above requires N + 1 function evalu-
ations to compute a full gradient, where N is the number of design variables x. This leads to
very poor scaling as the number of design variables increases.
Complex-Step Method: The complex-step method [57, 102, 65, 80] is analagous to the
finite-difference method, but uses a complex perturbation rather than a real perturbation. This
leads to a formulation of the form:
df (x) Im [f (x + ih)]
= . (3.2)
dx h
Chapter 3. CFD-Based Stability Derivatives 15
The use of a complex perturbation eliminates the subtractive cancellation issues inherent in
finite-difference methods and allows for extremely high numerical accuracy. Unfortunately, like
finite-difference methods, the cost of this method scales with the number of design variables.
Analytic Methods: For the purposes of this discussion, this class of methods consists of
the adjoint method and the direct method. This class of methods is primarily for improving the
efficiency and accuracy of the computation of derivatives for problems that use iterative solvers
to satisfy a set of residual equations. These methods involve rewriting the total derivative of the
function as a series of partial derivatives and linear solutions. A full derivation of the adjoint
method is shown in Section 3.3.2, but the essence of the method comes down to the following
equation:
∂f (x) ∂f (x) ∂R −1 ∂R
df (x)
= − . (3.3)
dx ∂x ∂ζ ∂ζ ∂x
In this expression, total derivatives, denoted by the d operator, involve a solution of the residual
system R, while partial derivatives, denoted by the operator ∂ are evaluated at the current value
of the states. By rewriting the total derivative as shown above, the necessity of solving the
full system of residual equations for each derivative component is eliminated, leaving instead
a set of four partial derivatives, and a series of linear solutions. The adjoint method involves
solving the set of linear systems formed by the second and third terms, which requires a linear
solution for each f . The direct method involves solving the set of linear systems formed by the
last two terms, which requires a linear solution for each design variable x. These methods can
be extremely efficient if either the number of functions, f , or the number of design variables, x,
are small. However, as should be evident even from this brief discussion, this class of methods
is significantly more complicated than either of the previous two methods. Further, this class
of methods requires access to the underlying code in order to be able to compute the necessary
partial derivatives, which are typically hand differentiated versions of the underlying evaluation
code. A small variation on these methods, called semi-analytic methods, consists in evaluating
the necessary partial derivatives using numerical techniques.
Automatic Differentiation: Automatic differentiation is a method that involves using
computational tools to perform a systematic application of the chain rule to a piece of code.
In essence, almost every computational algorithm can be broken down into a set of very simple
operations, each of which is simple enough to have a simple analytic derivative. By combining
these simple derivatives together, using the appropriate application of the chain rule, the correct
derivative can be determined for the entire algorithm [37]. If the operations in a given algorithm
are defined as:
Then the differentiation of the algorithm, using the chain rule can be written as:
i−1
∂ti X ∂fi ∂tk
= , j = 1, 2, . . . , n. (3.5)
∂tj ∂tk ∂tj
k=1
Note that the propagation of the derivative can be performed either from the design variables
forward or from the objective function backwards, a property that this method shares with
analytic methods. Many tools exist to implement this technique in a variety of programming
languages [17, 84, 10, 38, 9]. Once again this method requires access to the underlying analysis
code.
The ADjoint sensitivity method, mentioned frequently throughout this thesis, is a combi-
nation of the last two methods. Automatic differentiation techniques are used to compute the
partial derivatives in the adjoint equations. The combination of these two techniques produces
a method which has the computational efficiency of the adjoint method, and the numerical
accuracy and relative ease of implementation that come with automatic differentiation. A more
detailed discussion of this method is presented in Section 3.3.2.
The ADjoint approach to stability derivatives presented below builds on the work of Limache
and Cliff [54], and Park et al. [83, 82]. The ADjoint sensitivity method, developed by Mader et
al. [60] is applied to a moving grid CFD formulation — which is equivalent to the non-inertial
formulation used by both Limache and Cliff [54] and Park et al. [83, 82] — to compute the stabil-
ity derivatives. As in previous work, this CFD formulation allows both the static and dynamic
derivatives to be computed from a single, steady, flow solution. The main advantage of this
approach is that it combines the computational efficiency of analytic sensitivity methods with
the relatively straightforward implementation of automatic differentiation. This has enabled
the rapid development of an efficient method for the computation of stability derivatives.
3.3.1 Theory
The stability derivative formulation described in this section is based on two key methods. The
first is a CFD code that can compute solutions for rotating geometries. This can be accom-
plished with either a non-inertial reference frame formulation or a moving grid formulation. The
non-inertial formulation is well described by Limache and Cliff [54], and Park and Green [82],
so no further discussion on this topic is presented here. The other formulation, the moving
grid formulation, is used in this thesis and is discussed later on in this section. The second
key method is an efficient, robust and accurate sensitivity analysis method for the CFD code.
Chapter 3. CFD-Based Stability Derivatives 17
For the current work, this is accomplished with the ADjoint method. A brief summary of
this method is provided in Section 3.3.2, with more details available in previous work by the
author [60].
In order to compute derivatives with respect to both the static (α, β, V ) and dynamic (p, q, r)
parameters for a given configuration from a single steady-state flow solution, one needs a flow
solver that can compute steady-state solutions for constant, nonzero values of each of the
parameters. For the case of the static parameters, this is accomplished in most standard CFD
solvers. However, in the case of the dynamic parameters, this is usually not the case. The
method used to handle this motion in the following work is a moving grid formulation. The
flow solution is computed using the global velocities as the states. These velocities are expressed
in terms of the moving grid base vectors. This transformation introduces additional terms into
the equations that account for the moving coordinates of the grid, but eliminates the need to
add the centripetal and Coriolis forces as source terms to the momentum equations, as required
in the non-inertial formulation. The moving grid formulation is derived below.
u = v + w, (3.6)
where u is the velocity of the fluid in the fixed frame, v is the velocity of the fluid with respect
to the moving grid and w is the velocity of the moving grid. Using the approach of Warsi [112],
a general form of the conservation law can be written for moving coordinates as
∂A
− (∇A) · w + ∇ · F = C, (3.7)
∂τ
Where, for the conservation of mass, momentum and energy, F takes the form F = Au+B, and
A, B, C and τ can represent various quantities, depending on what quantity is being conserved.
In addition, Warsi [112] uses the identity:
which can be derived by applying the product rule to the left hand side and rearranging the
components. This identity can then be rearranged as follows,
Using this relationship in the conservation law (3.7), one can write
∂A
− ∇ · (Aw) + A (∇ · w) + ∇ · F = C. (3.10)
∂τ
Substituting F as defined above yields,
∂A
− ∇ · (Aw) + A (∇ · w) + ∇ · (Au + B) = C. (3.11)
∂τ
Rearranging the above equation, one obtains:
∂A
+ ∇ · (Au − Aw) + A (∇ · w) + ∇ · (B) = C. (3.12)
∂τ
Since (∇ · w) = 0 — i.e., the grid is incompressible — this equation simplifies to
∂A
+ ∇ · (Au − Aw) + ∇ · (B) = C. (3.13)
∂τ
Following the work of Warsi [112] and Ghosh [31], the conservation equations for mass, mo-
mentum and energy can be written as follows.
Energy Conservation: To obtain the energy conservation equations, one sets A = ρet ,
B = (pI − σ) · u + k∇T , C = 0 and τ = t, which yields
∂ρet
+ ∇ · (ρet u − ρet w) + ∇ · ((pI − σ) · u + k∇T ) = 0. (3.21)
∂t
Rearranging this equation, one obtains
∂ρet
+ ∇ · (ρet u − ρet w + (pI − σ) · u + k∇T ) = 0, (3.22)
∂t
where et = p/(γ − 1) + 1/2ρ|u|2 .
Final Formulation: Combining Equations (3.15), (3.20) and (3.22) yields the following set
of equations:
∂ρ
+ ∇ · (ρ(u − w)) = 0 (3.23)
∂t
∂ρu
+ ∇ · [ρu ⊗ u − ρu ⊗ w + pI − σ] + ρ[Ω × u] = 0 (3.24)
∂t
∂ρet
+ ∇ · [ρet u − ρet w + pu − σ · u + k∇T ] = 0. (3.25)
∂t
Note that the body forces have been neglected in this derivation. Inclusion of body forces is
accomplished by setting a nonzero value of C in each case.
Flux Vector Form: Now, restricting the deriviation to just the inviscid portion of the equa-
tions, the flux vector form of the governing equations can be written as,
∂ζ ∂fi
+ + s = 0, (3.26)
∂t ∂xi
where xi are the coordinates in the ith direction, and the state and the fluxes for each cell are
ρ ρui − ρwi 0
ρu ρu u − ρw u + pδ ρΩ u − ρΩ u
1 i 1 i 1 i1 2 3 3 2
ζ= ρu2 ,
ρui u2 − ρwi u2 + pδi2 ,
fi = s= ρΩ3 u1 − ρΩ1 u3 ,
(3.27)
ρu3 ρui u3 − ρwi u3 + pδi3 ρΩ1 u2 − ρΩ2 u1
ρet ρui (et + p) − ρwi et 0
and where w = w0 + Ω × x.
This is the formulation implemented in SUmb [107] and used in this thesis.
Having developed the ability to compute solutions for rotating configurations, one must now
consider how to compute the required derivatives. To accomplish this, one must develop a series
Chapter 3. CFD-Based Stability Derivatives 20
of motions that allow for the variation of a single parameter while holding all other parameters
constant. Consider pitch rate, q. As described by Limache and Cliff [54], a constant α loop
at a given q generates a steady solution. The radius of the loop is inversely proportional to q.
Thus, as q reduces to zero, the radius approaches infinity and steady level flight is recovered.
Similar ideas can be applied to roll rate, p, and yaw rate, r. However, in these two other
cases, achieving a steady rotating flow is more complicated. When the body axis of the config-
uration coincides with the wind axis, the logic is the same as for the pitching motion, but if we
incline the body axes at an angle of attack, α, relative to the wind, rotation about the body
yaw and roll axes no longer yields a steady state solution. In this case, rotation about the wind
axis is required to generate a steady solution. For small angles of attack, this distinction is not
of great importance. On the other hand, for large angles of attack, this should be taken into
account.
One further distinction must be drawn between the moving grid formulation and the non-
inertial reference frame formulation. To calculate the correct derivatives, the grid motion must
be specified in such a way that the magnitude of the grid velocity is independent of the rotational
rate. More specifically, the grid motion must be specified such that the velocity of the flow at
the nominal center of rotation is that of the desired free stream velocity. This falls out naturally
with the non-inertial reference frame formulation, because the velocity is specified in terms of
the local grid. For the moving grid formulation, this is not so obvious. To implement this, one
needs to specify a grid velocity that is linked to the rotational velocity such that the velocity
of the grid’s center of rotation is independent of the rotational speed.
To verify the implementation of the moving grid formulation, results are compared for a NACA
0012 airfoil rotating at a finite q̂ to those produced by Limache [55]. In this comparison, inviscid
flow is simulated around a NACA 0012 airfoil at Mach = 0.2 with q̂ = 0.0, 0.01, 0.03, 0.05. The
comparison is performed using a pseudo two-dimensional mesh, which has two cells in the
spanwise direction, with symmetry planes at the ends of the wing segment to reproduce two-
dimensional flow. Each slice of the mesh contains 65,536 cells for a total mesh size of 131,072
cells. Comparisons are made for the Cp distribution of the solution near the airfoil, as well for
the streamlines of relative velocity in the solution.
The Cp distributions shown in Figs. 3.1a–3.1d provide visual verification that the mov-
ing grid formulation implemented in SUmb is consistent with the non-inertial reference frame
Chapter 3. CFD-Based Stability Derivatives 21
Y
1
Z X
pressure
CoefPressure 0.9
0.9
0.5 0.7
0.7 0.5
0.5
0.3 0.3
0.1 0.1
0.05 0.05
0.01
-0.01 0.01
-0.05 0 -0.01
-0.1 -0.05
-0.2
-0.3 -0.1
-0.4 -0.2
-0.6 -0.3
-0.5 -0.4
-0.6
-1
1
Z X
pressure
CoefPressure
0.9
0.9 0.7
0.7 0.5
0.5 0.5
0.3 0.3
0.1 0.1
0.05
0.01 0.05
-0.01 0.01
-0.05 0 -0.01
-0.1
-0.2 -0.05
-0.3 -0.1
-0.4 -0.2
-0.6
-0.3
-0.5 -0.4
-0.6
-1
Figure 3.1: Cp contour and streamline comparison, rotating NACA 0012: M=0.2,α=0.00
Chapter 3. CFD-Based Stability Derivatives 22
30
20
Cp 20
0.9
pressure
0.7
0.9
10 0.5
0.3 10 0.3
0.1 0.1
0.05 0.05
0.01 0.01
y
0 0.005
y
-0.01 0
-0.05 -0.005
-0.1 -0.01
-0.2 -0.05
-0.3 -0.1
-10 -0.4 -10 -0.3
-0.6 -0.6
-20 -20
x x
(a) q̂=0.05: SUmb (b) q̂=0.05: Limache [55]
Figure 3.2: Flowfield streamline comparison, rotating NACA 0012: M=0.2, α=0.00
formulation used by Limache [55]. For each value of q̂, the Cp distribution matches the one
presented by Limache. In addition, the relative velocity contours also match. In particular,
note that for the q̂ = 0.05 case, in Figs. 3.2a and 3.2b, the point of zero relative velocity occurs
at the expected coordinates, (0, −20), in both cases.
Table 3.1: Comparison of lift and moment coefficients for the NACA 0012 at α = 0 for various
values of q̂ (results from Limache [55] in parentheses)
Finally, in Table 3.1 values of Cl and Cm are compared to reference values from Limache [55].
As can be seen in this table, the two implementations match very well over a range of q̂ values.
The largest discrepancy is 0.011 in Cl at Mach = 0.8 and q̂ = 0.05. This close correlation is
further confirmation that the formulation is correct.
Chapter 3. CFD-Based Stability Derivatives 23
Having modified the CFD code to handle rotating geometries, the code can now be differentiated
to get stability derivatives. To do this efficiently, the ADjoint method, previously outlined by
Mader et al. [60], is used. In this approach, automatic differentiation techniques are combined
with the adjoint method to generate the sensitivities for the CFD equations. The application
of this method to stability derivatives is as follows.
Start by considering the functions of interest, I, which may be either the forces (CL , CD , CY )
or moments (Cl , Cm , Cn ) acting on the aircraft. These are functions of both the states of the
system (ζ) and the values of the design variables (x), which in this case would be the states of
the dynamic model (α, β, V, p, q, r, h, δi ,etc.). The function of interest can be expressed as
The solution of the CFD equations can be represented as a set of governing equations that
are also functions of the states (ζ) and the design variables (x). These equations can be written
as
R (x, ζ (x)) = 0. (3.29)
To generate the adjoint equations for this system, one first writes the total derivative for
both the function of interest (3.28) and the residuals (3.29), which yields,
dI ∂I ∂I dζ
= + (3.30)
dx ∂x ∂ζ dx
and
dR ∂R ∂R dζ
= + = 0. (3.31)
dx ∂x ∂ζ dx
Where Eq. (3.31) must equal zero, because R = 0 when the governing equations are satisfied.
Note that a distinction is made between total and partial derivatives. In the context of this
thesis, a partial derivative is defined as a derivative evaluated for a constant set of states and
a total derivative is a derivative evaluated including a solution of the governing equations to
determine a new set of states, ζ, that satisfy R = 0. In the above equations, the derivatives
dI/ dx is the total derivative that one would obtain by performing a standard finite-difference
calculation over the entire flow solver. ∂I/∂x and ∂I/∂ζ are partial derivative vectors of size Nx
— the number of design variables — and NζT — the total number of flow states — respectively.
These derivatives are evaluated for a fixed set of states, ζ. dζ/ dx is another total derivative,
in this case representing the total derivative of the states with respect to the design variables.
This is a matrix term and is of size NζT × Nx . Similarly, in Equation (3.31), dR/ dx is the
total derivative of the residuals, including the solution of the governing equations. ∂R/∂ζ is
simply the flux Jacobian, a partial derivative of size NζT × NζT . ∂R/∂x is the partial derivative
Chapter 3. CFD-Based Stability Derivatives 24
of the residuals of the governing equations with respect to the design variables, x and is of size
N ζT × N x .
Therefore, the total derivative from Eq. (3.31) can be rewritten
∂R −1 ∂R
dζ
=− . (3.32)
dx ∂ζ ∂x
∂I ∂R −1 ∂R
dI ∂I
= − . (3.33)
dx ∂x ∂ζ ∂ζ ∂x
Now, the total derivative dI/ dx is expressed in terms of four partial derivatives that do
not require a solution of the residual equations in their computation. Instead, to compute the
total derivative dI/ dx, a series of linear solutions must be performed.
In this work, the adjoint system is solved,
∂R T
∂I
Ψ=− , (3.34)
∂ζ ∂ζ
which requires a separate linear solution for each component of I. The other option, the direct
method, involves solving:
∂R dζ ∂R
=− , (3.35)
∂ζ dx ∂x
which requires a separate linear solution for each component of x. The relative efficiency of the
two approaches depends on the relative number of components in I and x. In this case, the
number of components in I is six and the number of components in x is six or more, depending
on the number of control derivatives required. Therefore, it is slightly favorable to use the
adjoint method, but the relative numbers are close enough that it is not overly important. In
this thesis, the adjoint approach is used because it has already been implemented to compute
the derivatives for design optimization [58]. In that case, the number of input variables generally
exceeds the number of output variables, making the adjoint approach the more efficient of the
two approaches.
The other consideration associated with Eq. (3.33) is how to calculate the four partial
derivatives that make up the expression. This is where automatic differentiation is used. One
of the most significant drawbacks of the adjoint method outlined above is that the calculation
of the partial derivatives making up the expression can be very complex. In many cases,
such as those involved in complex CFD schemes, the effort required to differentiate the code
used to compute the residuals is significant. By using automatic differentiation to compute
these derivatives, the amount of effort required to complete the differentiation is significantly
reduced. Also, no approximations are made in the differentiation, and as a result, the derivatives
computed with the ADjoint method are extremely accurate.
Chapter 3. CFD-Based Stability Derivatives 25
To verify the ADjoint implementation for the computation of stability derivatives, two com-
parisons are performed. In Table 3.2, the values of Clα , Cmα , Clq and Cmq are compared
against derivatives computed using the complex-step derivative approximation [65]. Using the
complex-step approach, the sensitivity is computed by
dI(x) Im [I(x + ih)]
= + O(h2 ), (3.36)
dx h
√
where i = −1. As mentioned in Section 3.2, this approximation avoids the subtractive
cancellation issues inherent in finite differences. Therefore, with a sufficiently small step size
— in this case 10−20 — the method is able to produce derivatives with the same accuracy as
the flow solution, making it an excellent benchmark for verifying derivatives.
Table 3.2: Sensitivity verification: NACA 0012 test case, Mach = 0.5, 131,072 cells: 10−12
relative convergence
Table 3.3: ADjoint stability derivatives for a NACA 0012 airfoil at α = 0.0 degrees
Table 3.2 shows that the ADjoint and complex-step results match to between 9 and 11
digits. This is extremely accurate, far beyond the accuracy of the underlying physical model
used. Furthermore, given the iterative nature of the solvers used, the accuracy is consistent
with the 10−12 relative convergence tolerance that was achieved.
Chapter 3. CFD-Based Stability Derivatives 26
In Table 3.3, the values of Cmq and Clq are compared against those computed by Limache
and Cliff [54] for the same case. This comparison is done to show that the dependencies of the
coefficients on q are accurately captured with this method. The results in this table show that
the code is very accurate relative to the reference implementation of Limache and Cliff [54]. For
the Mach = 0.1 case and the Mach = 0.5 case, the difference between the two implementations
is less than one percent. In the Mach = 0.8 case, Clq is within one percent of the reference
results, while Cmq differs by approximately five percent. This difference can be attributed
to the fact that there is a shock wave in the solution of the Mach = 0.8 case. The precise
location of the shock has a significant impact on the value of the moment coefficient and hence
the moment coefficient derivatives. Given that SUmb is a structured, multi-block code, and
that the reference results were computed with an unstructured code, slight variations in the
prediction of the shock location are not unexpected. Based on these results, it can be concluded
that, for the NACA 0012 airfoil case, the stability derivatives predicted by the code are correct.
A complete demonstration of the three dimensional capabilities of this method can be found in
Mader and Martins [59].
However, despite the efficiency of this method for computing the static and dynamic deriva-
tives, it has two inherent drawbacks. First of all, it is based on a steady-state flow solution.
While this enhances its desireability from a performance perspective, it means that the flow
solution contains no knowledge of the time history of the flow and is therefore unable to provide
any insight towards the “dot” derivatives for a given configuration. Since these “dot” deriva-
tives are necessary to complete the linear flight dynamic model, this is a significant drawback.
Secondly, since the adjoint method was used to generate the stability derivatives, a second
derivative method would be necessary in order to include these derivatives in an optimization.
Possible methods for achieving this with a certain amount of efficiency have been shown by
Ghate and Giles [30] and by Rumpfkeil and Mavriplis [97]. However, even these methods re-
quire the solution of an adjoint system for all of the output variables as well as a direct method
solution for each design variable in the problem. For simple problems, this would work, but for
aerodynamic shape optimization problems with large numbers of design variables, such as the
ones demonstrated in this work, the cost of this method would quickly become prohibitive.
The time-spectral stability derivative formulation presented in the following section is similar
to the methods presented by Murman [74] and DaRonch et al. [94] and is motivated by the
complex number α̇ derivative methodology outlined by Etkin [27]. There is also a similar
method, outlined in the context of analyzing experimental results outlined in Rohlf et al. [93].
Chapter 3. CFD-Based Stability Derivatives 27
The underlying core of this approach to computing stability derivatives is the time-spectral
CFD method. This method is one of a family of methods that discretize the time derivative in
the unsteady flow equations in a spectral form for periodic problems. These methods exploit
the periodic nature of the problem by expressing the states of the system as a Fourier series in
time. The entire periodic solution can then be recovered from a small number of state instances
spanning the time period, or frequency spectrum, of the problem. These state instances can
then be solved for directly, using standard steady-state convergence acceleration techniques,
thereby eliminating the need to iterate through the startup transients typical of unsteady CFD
problems.
These spectral method computations can be performed in either the time-domain, the fre-
quency domain, or a combination of the two. In the time-domain, the state instances represent
discrete snapshots of the solution in time, while in the frequency domain, the state instances
represent distinct frequencies present in the solution. In each case, the spectral solution is
capable of representing a fundamental frequency as well as a number of higher harmonics. The
number of resolved harmonics is related to the number of time or frequency instances present
in the solution.
Early work on time nonlinear spectral solution techniques was conducted by Hall et al. [41],
who derived a spectral formulation for the two-dimensional Navier–Stokes equations. This
derivation of the spectral equations is conducted in the frequency-domain. However, to facilitate
computation, the flow equations are transformed back to the time-domain. This allows a
typical, time-domain residual formulation to be used for the computation of the solution in
each of the spectral instances. This residual is augmented by a spectral term that couples the
various solution instances. In this case, while the residuals are computed in the time-domain,
both the spectral operator and the boundary conditions are applied in the frequency domain,
yielding a mixed time-domain/frequency-domain approach. In an extension of this work, Ekici
Chapter 3. CFD-Based Stability Derivatives 28
and Hall [24] apply this technique, known as the harmonic balance technique, to multistage
turbomachinery applications where a variety of frequencies may be present. A mixed time-
domain/frequency domain approach is also used in this extension.
Another nonlinear spectral solution technique is the nonlinear frequency domain (NLFD)
method developed by McMullen et al. [69, 68, 70]. In this technique, the solution process takes
place primarily in the frequency domain. The states of the system are stored as frequency
domain Fourier coefficients and the solution steps are generated from the frequency domain
residual and spectral operator. To simplify the implementation, the residual is evaluated in the
time domain, where the states are transformed from the frequency domain to the time domain.
Then, the residual is transformed from the time-domain to the frequency domain using fast
Fourier transform (FFT) techniques.
The time-spectral method introduced by Gopinath and Jameson [34, 35] is similar to the
harmonic balance method of Hall et al. [41]. However, the time spectral method is derived
completely in the time domain. This yields a purely real spectral operator, and allows for the
use of the time-domain residual operator in its original form, including boundary conditions.
The time-spectral method is the method that has been used in this thesis. While the original
derivation of this method is due to Gopinath [35], the basics of the method have been included
here to put the remainder of the discussion in context.
To derive the time-spectral equations, one starts by writing the governing equations for
unsteady flow. These can be written as:
∂ζ ∂fi
V + = 0, (3.37)
∂t ∂xi
where xi are the coordinates in the ith direction. Once again, as in Section 3.3, grid motion
will need to be modeled, therefore the moving grid formulation, derived before, is used again
here. However, in this case, since the motions are more general and not intended to produce
a steady solution, the general form of the momentum equations shown in Equation (3.17) is
used. Based on these equations, for inviscid flow, the states and the fluxes for each cell are:
ρ ρui − ρwi
ρu ρu u − ρw u + pδ
1 i 1 i 1 i1
ζcell = ρu2 ,
fi = ρui u2 − ρwi u2 + pδi2
. (3.38)
ρu3 ρui u3 − ρwi u3 + pδi3
ρet ρui (et + p) − ρwi et
dζ
V + R(ζ) = 0. (3.39)
dt
Chapter 3. CFD-Based Stability Derivatives 29
where R represents the spatially discretized residual operator implemented in the flow solver.
For SUmb [107], the flow solver used in this thesis, this is a second order, cell-centered finite-
volume scheme. This operator includes all of the boundary conditions and artificial dissipation
operators in the flow solver.
Now, the spectral time-derivative operator can be derived as follows:
One assumes that the states of the system, ζ, can be expressed as a Fourier series. Thus, a
Fourier transform of the system states can be written as,
N −1
ˆ 1 X n −ikxn
ζk = ζ e , (3.40)
N
n=0
Note that in Equations (3.40) and (3.41) there are two representations of the state vector, ζ and
ζ̂. These vectors represent the time-domain and frequency domain representations of the state
vector respectively. The time interval for the series is xn = 2πn/N , where N is the number
of time intervals, and n is the index of the current time interval. In the frequency domain, k
represents the frequency component index of the state vector.
Combining Equations (3.40) and (3.41) to express ζ n explicitly in the time domain yields:
N −1
2 N −1
n
X 1 X l −ikxl ikxn
ζ = ζe e . (3.42)
N
k=− N 2−1 l=0
Noting that xln is a function of time, Equation (3.47) can be differentiated with respect to time
to give:
N −1
" #
n 1 X l N cos( N2 xln ) sin( N x2ln ) cos( x2ln ) dxln
Dt ζ = ζ − (3.48)
N 2 sin( x2ln ) 2 sin2 ( x2ln ) dtln
l=0
Now consider the derivative dxln / dtln from Equation (3.51). Substituting in the value of xln
and evaluating gives
2π(n − l) 2π T (n − l) 2π 2π 2π
xln = = = ∆t(n − l) = (tn − tl ) = tln . (3.52)
N T N T T T
Therefore, the derivative of xln with respect to time is
dxln 2π
= (3.53)
dtln T
and this relationship can now be used in Equation (3.51) to get
N −1
" #
π X (−1) (n−l)
Dt ζ n = ζl . (3.54)
T sin( x2ln )
l=0
Thus, Dt is an operator that spans all of the time instances in the solution. By solving the N
coupled time instances represented in the equation,
V Dt ζ n + R(ζ n ) = 0, (3.57)
where n represents each of the N time instances, one ends up with a coupled set of solutions
that represent the periodic, steady-state solution to the given problem.
Note that while the Dt ζ n operator can be thought of as a matrix vector product, it is
implemented on a cell by cell basis. Thus, for a single state in a single cell, the matrix operator
can be expressed as
0 d1,0 ··· dN −1,0
d0,1 0 ··· dN −1,1
Dt = . (3.58)
.. .
.. .. ..
. .
d0,N −1 d1,N −1 · · · 0
This operator is applied across the spectral instances of each state on a cell by cell basis to
form a time-spectral residual
RT S = V Dt ζ n + R(ζ n ) = 0. (3.59)
∂wn
V + RT S = 0. (3.60)
∂t∗
Where the pseudo-time iterations are solved using a mixed Runge-Kutta, Newton-Krylov ap-
proach. The startup iterations are solved using a full multigrid, five stage Runge-Kutta method.
After two or three orders of magnitude of convergence, the iterations switch over to a precondi-
tioned Newton-Krylov approach. The configuration of the preconditioner for this solver is the
same as described for the ADjoint in Chapter 4.
Having described the time-spectral formulation for the flow solver, the time-spectral approach
to the computation of stability derivatives can now be outlined. To start, consider the linear air
reaction theory outlined by Etkin [27]. This theory states that, for a general motion, the force
Chapter 3. CFD-Based Stability Derivatives 32
and moment coefficients of an aircraft — for example, the lift coefficient — can be approximated
as:
CL = CL0 + CLα ∆α + CLα̇ ∆α̇ + CLα̈ ∆α̈ + ... + CLβ ∆β + ..., (3.61)
where all motion states α, β, V , p, q, r, ze and their associated time derivatives are included.
CL0 is the value of the coefficient for the steady-state reference flight condition about which
the motion occurs. However, if a simple motion is specified that consists of a single dynamic
state — for example α — the value of the remaining motion states are zero and they, and their
associated derivatives, drop out of the equation. In the case of a pure α motion this leads to:
One can then make the further assumption, justified further on in this section, that the higher
order derivatives are small and neglect them, which gives:
6.0e-02
CL
CL : 3 Point Solution
CL : 5 Point Solution
4.0e-02 α
α: 3 Point Solution
α: 5 Point Solution
2.0e-02
Value
0.0e+00
-2.0e-02
-4.0e-02
-6.0e-02
0.0e+00 5.0e-01 1.0e+00 1.5e+00 2.0e+00
Time (s)
6.0e-02
3 point Solution
5 point Solution
7 point Solution
Derivative estimate
4.0e-02
solution
2.0e-02
0.0e+00
CL
-2.0e-02
-4.0e-02
-6.0e-02
-1.0e-02 -5.0e-03 0.0e+00 5.0e-03 1.0e-02
α (rad/s)
are shown in Figure 3.5. To compute stability derivatives from this solution, one computes a
linear least-squares fit of the output coefficient (e.g., CL ) with respect to the primary motion
variable from the solution (e.g., α). The slope of the line resulting from this fit is the stability
derivative (CLα ) while the y-intercept of the line is the value of the coefficient (CL ) at the zero
value if the motion perturbation ∆α = 0. To demonstrate this process, Figure 3.6 shows a
time-spectral solution plotted against the corresponding ∆α motion. As can be seen from the
figure, the general trend of the relationship between the coefficient and the motion parameter
is linear. The linear regression generates the line shown in the figure which quantifies the slope
and intercept of the trend. These two values represent the steady-state value of the coefficient
and its derivative with repect to the motion variable. For example,
However, there is also a distinct hysteresis in the solution, as demonstrated by the gap between
the coefficient on the upstroke and downstroke of the oscillation. This hysterisis relates to
the “dot” derivatives. As discussed by Etkin [27] with respect to the α̇ derivatives, the “dot”
derivatives represent the time lag in the development of the coefficient resulting from a change
in the flow. To estimate this quantity, one subtracts the value of the linear regression line from
the solution. This process eliminates the bulk dependence of the solution on the main motion
Chapter 3. CFD-Based Stability Derivatives 35
1.0e-02
3 point solution
5 point solution
7 point solution
Derivative Estimate
5.0e-03
Residual CL
0.0e+00
-5.0e-03
-1.0e-02
-3.0e-02 -2.0e-02 -1.0e-02 0.0e+00 1.0e-02 2.0e-02 3.0e-02
α̇ (rad/s2 )
variable, leaving just the variation associated with the hysteresis. This yields another strongly
linear relationship, as shown in Figure 3.7. Taking the slope of this trend yields the value of
Ciα̇ . Further, as this plot shows, the linear approximation is a very good representation of the
α̇ trend, indicating that any dependence on the higher order time derivatives is small. This is
sufficient to justify the assumption made for Equation (3.63).
Note that in Figure 3.6, all three solutions — 3, 5 and 7 time instance — produce the same
elliptical solution. This indicates that for the small amplitude (< 0.5 deg oscillations), inviscid
flow solutions used in this work three time instances are sufficient to accurately represent the
periodic solution. Therefore all of the subsequent time-spectral results in this thesis are base
on three time instance solutions.
Also note that, generally speaking, linear regressions will be biased towards large clusters
of data. Thus, if a random set of points is selected on the ellipse shown in Figure 3.6, the linear
regression analysis would produce inconsistent results. However, because the time-spectral CFD
approach generates solution points that are equally spaced around the ellipse, the solution turns
out to be independent of which three evenly space points are chosen. Data demonstrating this
fact is shown in Table 3.4.
Chapter 3. CFD-Based Stability Derivatives 36
Time Instance
1 2 3 Cl0 Clα
α 0.0 0.007558 -0.007558
Set 1 6.218 -0.000717
Cl -0.009895 0.050868 -0.043125
α 0.005130 0.003550 -0.008679
Set 2 6.218 -0.000717
Cl 0.023756 0.029739 -0.055646
α 0.008300 -0.001814 -0.006485
Set 3 6.218 -0.000717
Cl 0.048058 -0.003023 -0.047187
To verify and validate this approach, two test cases will be examined. The first, a NACA 0012
airfoil undergoing an oscillating plunging motion, will be compared against a theoretical thin
airfoil theory result. The second, the Stability And Control CONfiguration (SACCON) UCAV,
will be compared against wind tunnel results.
The NACA 0012 case is a pseudo three-dimensional case. The meshes used are two cells thick,
with symmetry planes on both sides to create two dimensional flow. The mesh is a C-mesh
topology split into 16 blocks for efficient parallel balancing. Samples of the meshes are shown
in Figure 3.8. To assess the numerical accuracy of the solutions, a mesh convergence study
was conducted with meshes containing 4,096, 16,384, and 65,536 cells per slice. Plots of these
results are shown in Figure 3.9. The values of the coefficients (CL0 , Cm0 ) and their derivatives
with respect to α show excellent numerical accuracy, with relative errors less than 1% on the
finest mesh for both sets of coefficients and derivatives. The accuracy of the derivatives with
respect to α̇ are not quite as good, with relative errors on the order of 5% on the finest mesh.
However, that level of error is acceptable for the purposes of this work.
Etkin [27] presents theoretical results for an oscillating plunging airfoil based on the work of
Theodorsen. Using the Theodorsen function as a basis, Etkin presents the theoretical values of
the lift and moment coefficients as:
CLα = 2πF (k) (3.66)
Chapter 3. CFD-Based Stability Derivatives 37
Y Y
X
Z X
Z
(a) Mesh topology (b) Near field mesh: 65536 cells per slice
100 101
∆CL/CL ∆Cm/Cm
∆CLα /CLα ∆Cmα /Cmα
100
10−1 ∆CLα̇ /CLα̇ ∆Cmα̇ /Cmα̇
10−1
−2
10
10−2
10−3
10−3
10−4 −3 10−4 −3
10 10−2 10−1 10 10−2 10−1
N −1/2 N −1/2
G(k)
CLα̇ = π + 2π (3.67)
k
1
Cmα = 2πF (k)(h − ) (3.68)
4
1 G(k) 1
Cmα̇ = π(h − ) + 2π (h − ) (3.69)
2 k 4
where F (k) and G(k) are the real and imaginary parts of the Theodorsen function,
Since the theoretical result varies with frequency, the following comparison is performed over
a variety of reduced frequencies. In addition, because the theoretical result is for an infinitely
thin airfoil, numerical results are computed for the full thickness airfoil as well as half and
quarter thickness airfoils. As shown in Figure 3.10, the numerical results match the theoretical
results well over the full range of reduced frequencies. There is a small offset between the
numerical results and the theoretical results. However, for the α̇ derivatives, this discrepancy
reduces significantly for the half and quarter thickness airfoils, indicating that the bulk of this
discrepancy for the α̇ derivatives comes from the thin airfoil theory assumptions of the reference
results. For the α derivatives, the trend is less conclusive. At lower frequencies, the thin
results match the theoretical result more closely than the full thickness results, while at higher
frequencies the offset is larger for the thin airfoils than the thick airfoils. However, the shape
of the curve for the thinner results matches the theoretical result more closely than the shape
of the thicker airfoils. Note that the theoretical results for the α̇ derivatives become undefined
as the reduced frequency tends to zero. This singularity is clearly apparent in Equations (3.67)
and (3.69), and prevents the calculation of a definitive value for the derivative. However, given
the excellent agreement between the theoretical and numerical results, this comparison confirms
that the method is valid for this simple two-dimensional case.
The second test case considered is a full, flying-wing, UCAV configuration. The mesh is con-
structed with an H-H topology and 32 blocks. The geometry of the case is based on the
information provided in Schütte et al. [99] and the surfaces of the wing are modeled as inviscid
walls with a symmetry plane imposed at the root. The off wall spacing for the 1,179,000 cell
mesh is 1 × 10−3 root chords. Sample meshes are shown in Figure 3.11
Once again, solutions were computed on a series of meshes to assess numerical accuracy.
Specifically, meshes with 147,000, 1,179,000 and 9,430,000 cells were used to compute the solu-
tions. Assessments of the error associated with CL and its derivatives are shown for both an α
motion and a q motion in Figure 3.12.
Chapter 3. CFD-Based Stability Derivatives 39
6.5 5
Theoretical
Time-Spectral:Full Thickness
Time-Spectral:Half Thickness
6.0 Time-Spectral:Quarter Thickness 0
5.5 −5
5.0 −10
C Lα
CLα̇
4.5 −15
4.0 −20
6
Theoretical
Time-Spectral:Full Thickness
Time-Spectral:Half Thickness
5
Time-Spectral:Quarter Thickness
−1.0
4
−1.2
C mα
Cmα̇
1
−1.4
Time-Spectral:Full Thickness
−1.6 Time-Spectral:Half Thickness −1
Time-Spectral:Quarter Thickness
Theoretical
−2
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Reduced Frequency (k) Reduced Frequency (k)
Y X Z X
(a) Surface mesh: 147,000 cell mesh (b) Near field mesh: 147,000 cell mesh
101 100
∆CL/CL ∆CL/CL
∆CLα /CLα ∆CLq /CLq
100 10−1
∆CLα̇ /CLα̇
10−1 10−2
10−2 10−3
10−3 −3 10−4 −3
10 10−2 10−1 10 10−2 10−1
1/N 1/3 1/N 1/3
The meshes, again, show useful mesh convergence as the grids are refined. The numerical
accuracy of CL and CLα is better than 1% for the α motion on the finest mesh. The value of
CL for the q motion shows similar accuracy, better than 1%, while the value of CLq is slightly
less accurate, on the order of 5% for the finest mesh.
Using the 1,179,000 cell grid, a series of comparisons were run against derivatives computed from
experimental data by Rohlf et al. [93]. The goal of this comparison is to place the results from
the time-spectral computation in context with physical, real world results, thereby validating
the method. The flow conditions used are shown in Table 3.5. These conditions are taken
from Tormalm and Schmidt [105] and are intended to match the conditions of the experimental
results from Loeser et al. [56]. The results of these comparisons are show in Figures 3.13
and 3.14.
Parameter Value
Static pressure (Pa) 97767
Static temperature (K) 291.0
Density (kg/m3 ) 1.185
Mach number 0.149
α 0 - 10 deg.
Half body reference area (m) 0.385
Reference half span (m) 0.769
Reference chord (m) 0.479
Frequency 6.28 rad/s (1 Hz)
Reference moment center (m) 0.6
Reference rotational center (m) 0.8554
The derivatives in Figures 3.13 and 3.14 match the experimental results surprisingly well,
especially considering that the numerical results are computed using an Euler solver. The
results capture the correct order of magnitude for each of the derivatives. As one might expect,
the CLα derivatives are the most accurate, giving results within 10% of the experimental values.
The CLq+α̇ and Cmq+α̇ derivatives have errors of 30% and 40% respectively. Given the lack of
viscous effects in the numerical solution, this is amount of error is unsurprising. The error in
Cmα is on the order of 80%. This is again most likely due to the missing physics in the Euler
model used for the computation. This is consistent with the accuracy of the coefficients CL
and Cm — shown in Figure 3.15 — where the lift prediction is very good and the moment
Chapter 3. CFD-Based Stability Derivatives 42
3.0 0.0
CLq+α̇ Exp.
CLq+α̇ TS
2.5 −0.5
2.0 −1.0
1.5 −1.5
1.0 −2.0
0.5 −2.5
CLα Exp.
CLα TS
0.0 −3.0
0 2 4 6 8 10 0 2 4 6 8 10
α α
0.0
Cmα Exp. Cmq+α̇ Exp.
Cmα TS Cmq+α̇ TS
0.3
−0.5
0.2
0.1 −1.0
0.0
−1.5
−0.1
−0.2 −2.0
0 2 4 6 8 10 0 2 4 6 8 10
α α
0.5 0.045
CL Exp. Cm Exp.
CL TS α Cm TS α
0.4 CL TS q 0.040 Cm TS q
0.035
0.3
0.030
0.2
0.025
0.1
0.020
0.0
0.015
−0.1 0.010
0 2 4 6 8 10 0 2 4 6 8 10
α α
(a) CL (b) Cm
prediction is poor. That said, the values predicted by the method are accurate enough to give
a sense of what the stability characteristics of the aircraft are and are sufficiently accurate to
be useful in demonstrating the effect of stability constraints on an aerodynamic optimization.
As a final consideration, one needs to consider the frequency dependence of the derivatives. As
was shown in the two dimensional test case, there can be a significant frequency dependence
in the derivatives. To examine these effects in the three-dimensional case, a frequency sweep
was conducted on the 1,179,000 cell mesh, the results of which are shown in Figures 3.16
through 3.18.
Considering Figure 3.16 first, one can see that while there is a definite frequency dependence
in the α derivatives, the values tend to a constant value as the frequency is reduced. The same
can not be said for the α̇ and q derivatives. As shown in Figures 3.17 and 3.18 both of these
quantities vary continuously as the frequency is varied. This variation is most pronounced in
the CLα̇ derivative, which varies almost linearly over the range of frequencies tested, although
there does appear to be a slight flattening of the trend at the lower end of the frequency range.
The remainder of the α̇ and q derivatives are relatively constant over a range of frequencies
between k = 0.03 and k = 0.3. Therefore, frequencies in this range should be used for the
computation of derivatives using this method unless other considerations dictate otherwise. Of
note, the 1 Hz frequency (k = 0.03) used in the experimental comparisons falls at the lower end
of this range.
Chapter 3. CFD-Based Stability Derivatives 44
3.0 0.20
2.5
0.15
2.0
1.5 0.10
1.0
0.05
0.5
CLα TS
CLα Steady
0.0 0.00
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.00 0.05 0.10 0.15 0.20 0.25 0.30
k k
1.0 1.0
0.5 0.5
0.0 0.0
−0.5 −0.5
−1.0 −1.0
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.00 0.05 0.10 0.15 0.20 0.25 0.30
k k
0.0 0.0
CLq TS Cmq TS
CLq Steady Cmq Steady
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1.0 −1.0
−1.2 −1.2
−1.4 −1.4
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.00 0.05 0.10 0.15 0.20 0.25 0.30
k k
The ultimate goal of this work is to be able to perform gradient-based optimization of a given
aircraft configuration with stability considerations included. In the previous chapter, a time-
spectral method for the computation of stability derivatives was presented. This method is
capable of computing all of the necessary stability derivatives for the longitudinal stability
constraints described in Chapter 5. However, as time-spectral CFD solutions are relatively
expensive to compute, there is a need for an efficient method for computing the gradient of
the CFD solution with respect to the large number of design variables typical of aerodynamic
shape optimization. In the field of steady-state aerodynamic optimization, adjoint methods
are known to provide sensitivities at a cost that is practically independent of the number of
design variables. Based on this known characteristic, the ADjoint method of Mader et al. [60]
is extended to compute the time-spectral adjoint of the aerodynamic parameters needed for the
optimizations. The details of this computation are presented in this chapter.
4.1 Background
Adjoint methods have become well known in the aerodynamic shape optimization research
community. The method was introduced in the context of fluid dynamics by Pironneau [89]
in the mid 1970s, where it was used to demonstrate how to minimize the drag over bodies
immersed in laminar viscous flows. Angrand [4] introduced the use of an adjoint for the opti-
mization of airfoils in potential flow in the early 1980s. It was then applied in the context of
aerodynamic shape optimization of airfoils and wings with the Euler equations by Jameson [46]
in the late 1980s. Since then it has been applied to the optimization of airfoils — including
viscous effects [78, 3, 101], laminar-turbulent transition prediction [23] and multi-point opti-
mizations [101, 79] — as well as three-dimensional wings — for inviscid flows [48, 44], viscous
flows [47, 81] and multi-point problems [16] — and full wing-body configurations [91, 92]. It has
46
Chapter 4. Time-Spectral ADjoint Method 47
also been applied to noise related applications such as sonic-boom reduction [77], hypersonic
flows including magneto-hydrodynamics [63, 62] and has been generalized to multidisciplinary
systems by Martins et al. [64, 66]. In each of these cases, the adjoint method allowed for
the efficient optimization of the design in question with respect to a large numbers of design
variables.
However, while the use of adjoint techniques has become relatively commonplace in the con-
text of steady-state optimizations — at least in the research community — its use is still rela-
tively uncommon in time-dependent problems. Examples in two-dimensions have been demon-
strated by Rumphkeil and Zingg [98], Nadarajah and Jameson [75], Mani and Mavriplis [61],
and Wang et al. [111] and a three-dimensional example has been demonstrated by Mavriplis [67].
These adjoint methods are a significant improvement over finite difference methods, but they
still have a high computational cost. The unsteady adjoint computation requires a reverse
integration in time from the final solution back to the initial condition [61, 75]. Thus, a time
dependent adjoint requires the full forward solution of the unsteady problem, storing the flow
states for each time step along the way, followed by a reverse sweep of the solution process
to find the adjoint solution. While this process is still more efficient than computing a full
unsteady solution for every design variable in the problem, it is still an expensive proposition.
Various methods for reducing the computational resources required for the computation have
been suggested; for example, writing the solution history to disk rather than storing the solu-
tion in memory [61] or evaluating only a periodic portion of the time history for the adjoint
problem [75, 98]. However, even with these additions, the computational cost of full unsteady
adjoint methods is still significant.
In the same way that the time-spectral CFD method has a lower computational cost than a
full, unsteady flow solution for periodic problems, the time-spectral adjoint method offers a lower
computational cost than an equivalent unsteady adjoint approach for the same problem. Just
as the spectral solution technique modifies a single unsteady CFD problem into a set of coupled
steady CFD problems, the time-spectral adjoint technique converts a full unsteady adjoint
problem into a single, large, steady adjoint problem. Coupling this with the efficient solution
of large, sparse linear systems provided by modern software packages — such as PETSc [6] —
allows for the rapid implementation of an adjoint technique for periodic unsteady problems.
It is worth noting that these potential advantages apply to all spectral methods, not just
the time-spectral method. To date, adjoint methods have been developed for each of the
spectral methods mentioned in Section 3.4.1. An adjoint for the two dimensional, viscous
harmonic balance equations was developed by Thomas et al. [104]. In this work, the authors
use a combination of forward and reverse mode automatic differentiation to generate the terms
necessary for the adjoint. Further, they use the adjoint to compute mesh sensitivities for
Chapter 4. Time-Spectral ADjoint Method 48
4.2 Theory
To derive the time-spectral adjoint, as with the steady case, one starts by writing the vector-
valued function of interest, I, as:
I = I(x, ζ n (x)), (4.1)
where x represents the vector of design variables and ζ n is the state variable vector for the
nth time instance and n = 1, ..., N , with N representing the number of time instances in the
solution.
For a given vector x, the solution of the governing equations of the system yields a vector
ζ n. As shown in Section 3.3.2, when deriving the adjoint equations, one generally expresses the
governing equations as:
R (x, ζ (x)) = 0. (4.2)
However, since the flow solution is time-spectral in nature, from van der Weide et al. [108], the
governing equations are redefined as:
where R (x, ζ n (x)) is a normal steady-state residual for the nth time instance, n = 1, ..., N , and
Dt is the spectral operator defined in Equation (3.55). This modified residual:
can now be treated in the same fashion as the steady-state residual would be treated in a
normal adjoint formulation. Thus, as before, the total sensitivity of the vector-valued function
of interest, I, is written as:
dI ∂I ∂I dζ n
= + n . (4.5)
dx ∂x ∂ζ dx
Once again, because the governing equations must always be satisfied at a converged solu-
tion, the total derivative of the residuals in Equation (4.4) with respect to any design variable
must also be zero. Writing the total derivative of the governing equations with respect to the
design variables gives:
dRT S ∂RT S ∂RT S dζ n
= + = 0. (4.6)
dx ∂x ∂ζ n dx
This expression provides the means of eliminating the total derivative dζ n / dx from the total
sensitivity computation for I. Moving the first term of this equation to the right-hand side
gives:
∂RT S dζ n ∂RT S
n
=− . (4.7)
∂ζ dx ∂x
∂I ∂RT S −1 ∂RT S
dI ∂I
= − . (4.8)
dx ∂x ∂ζ n ∂ζ n ∂x
The adjoint approach consists in factorizing the ∂RT S /∂ζ n matrix with the term to its left,
yielding the adjoint system:
T
∂RT S ∂I
ψ= . (4.9)
∂ζ ∂ζ
This solution is then used in Equation (4.8) to obtain the total sensitivity:
dI ∂I RT S
= − ψT . (4.10)
dx ∂x ∂x
Just as with the steady-state case, this yields a sensitivity method with a computational cost
that is essentially independent of the number of design variables. Note that because the time-
spectral system is N times the size of the steady-state system, the adjoint system is also N
times larger than the equivalent steady-state system.
Chapter 4. Time-Spectral ADjoint Method 50
4.3 Implementation
Having discussed the theory behind the time-spectral adjoint, it is now possible to discuss the
details of the implementation. Much of the implementation is the same between the steady-
state and time-spectral adjoint methods, therefore the implementation of the two methods will
be discussed at the same time.
The basis for the residual derivatives in the ADjoint approach is a single cell residual routine,
developed from the original residual routines. This single cell residual routine containes all
of the functionality of the original block based routines — including dissipation terms and
boundary conditions — but is designed to operate on a 5 × 5 × 5 cell cube, which is the smallest
block of cells that encloses the second order inviscid flux stencil, shown in Figure 4.1. This
2.5
1.5
0.5
0
k
−0.5
−1
−1.5
−2
−2.5 2
−3 1
−2 −1 0
0 1 −1
2 3 −2
j
i
single cell residual routine is generated by cutting and pasting the residual computation from
the original residual code, modifying the indices to limit the evaluation to a single stencil. This
process, while not ideal, is less onerous than coding a derivative scheme from scratch. Further,
by reducing the residual computation to that associated with a single cell at a time, the size
of the differentiation problem is reduced significantly, making reverse mode differentiation less
memory intensive. The ability to use the reverse mode is advantageous in this context, as is
shown in Section 4.3.2, because of the large ratio of inputs to outputs in the residual stencil.
An outline of the single cell routine used in this work is shown in Figure 4.2. Note that the
routine includes several boundary conditions, multiple dissipation schemes, the metric terms
Chapter 4. Time-Spectral ADjoint Method 51
necessary for the mesh derivatives, as well as the loops required for the time-spectral implemen-
tation. Figure 4.2 also shows the locations of the time instance loops needed to compute the
Outer Time Spectral Loop
ComputeResidual
xHalo
CheckInputParameters GridVelocitiesFineLevel
ApplyAllBCs BC Farfield
Residual
Left/Right State
VolumeNormalization
Riemann Flux
Figure 4.2: Call-graph: current residual computation: steady and time-spectral cases
residual for all time instances in the solution. The actual spectral operator from Equation (3.55)
is implemented inside the initializeResidual subroutine.
All of the derivatives used in the ADjoint formulation are computed using reverse mode AD.
More specifically, the Tapenade [84, 42] AD tool is used to perform reverse mode source trans-
formation AD on the necessary routines. For the residual derivatives — ∂R/∂ζ and ∂R/∂x
— reverse mode differentiation is performed on the single cell residual routine described above.
To properly understand this choice, it is necessary to discuss, in some detail, the structure of
the single cell residual routine.
For the steady case, the residual computation shown in Figure 4.2 computes the value of the
residual in a single cell. This evaluation produces a result of length Nζ — the number of states
per cell — one value for each of the governing equations in that cell. The number of inputs
required to generate that result is much higher than this. The variables of interest here are the
system states, ζ, and the mesh coordinates, x. For the second order, finite volume discretization
of the Euler equations used here, this requires the states in the nearest neighbor cells for the
central flux and the nearest neighbor states, as well as the next nearest neighbor states, for the
Chapter 4. Time-Spectral ADjoint Method 52
dissipation fluxes. Therefore a total of 13 × Nζ states are required for the evaluation of the
residual. The residual evaluation also depends on the coordinate locations of the corner nodes
of each of the nearest neighbor cells. This includes a total of 32 × 3 independent spatial degrees
of freedom. Therefore, for the five states in the Euler equations, we have an overall ratio of
13 × 5 + 32 × 3 = 161 : 5 — more than 32 : 1 — input variables to output variables. Even
considering the fact that a single reverse mode calculation is about 4.5 times more costly than
the equivalent forward evaluation of the single cell routine, the reverse mode calculation is still
about seven times faster than a forward computation for the single cell routines.
In the time-spectral case, it is necessary to consider the temporal dependence of the operator
in addition to the its spatial dependence. In describing this dependence, this discussion refers
to both on-time-instance states and off-time-instance states. On-time-instance states are those
states that exist in the same time instance as the current residual evaluation. Off-time-instance
states are those states that exist on a time instance other than the one associated with the
current residual evaluation. In the case of a time-spectral solution, the residual is dependent
on the same on-time-instance states and coordinates as described above. In addition to that, it
is dependent on the states and coordinates of the current cell on each of the off-time-instances.
Thus, there are now Nζ ×N residual values and (13 × Nζ + 32 × 3)×N +(Nζ + 8 × 3)×(N − 1)
input states and coordinates. This leads to a ratio of 36:1 input variables to output variables
for the Euler equations with three time instances, a ratio that increases as the number of time
instances increases.
Note that both of these ratios include the state variable derivatives of ∂R/∂ζ n as well as the
coordinate derivatives of ∂R/∂x. This fact is made possible because reverse mode AD has been
for the derivative calculation. The reverse accumulation, shown in Equation (3.5), allows the
computation to start with a single residual and, accumulating backwards through the routines,
calculate the derivative of all of the inputs at once. This turns out to be a significant advantage
in this case.
An important aspect of the expanded routine in the context of the time-spectral adjoint is
the need for additional loops to account for the extra time instances. As is shown in Figure 4.2,
there are two time instance loops in the spectral computation. The outer loop accounts for
the time instances in the residual, while the inner loop accounts for the time instances of the
states and coordinates. Therefore, we can see that the inner part of the computation scales
with the number of time instances squared, while the outer part of the computation scales
only with the number of time instances. However, this is a somewhat naive implementation
of the single cell routine. Examining Equation (3.57) more closely, one can see that it is only
the spectral operator, V Dt ζ n , that contains states from all N time instances at one time.
Therefore it is possible to reduce the number of terms inside the inner time-spectral loop to
Chapter 4. Time-Spectral ADjoint Method 53
only those necessary for this term. This simplified routine is diagramed in Figure 4.3. With
ComputeResidual
AdjustInflowAngle
CheckInputParameters
ReferenceState
SetFlowInfinityState
xHalo
Metric VolPym
GridVelocitiesFineLevel
NormalVelocitiesAllLevels
ComputePressure BC Symmetry
ApplyAllBCs BC Farfield
TimeStep BC EulerWall
Residual
this improved implementation, the number of computations inside the inner time-spectral loop
is significantly reduced and the derivative computation now scales, essentially, with the number
of time instances, N , rather than the number of time instances squared, N 2 . Timing results
demonstrating this fact are presented in Section 4.4.3.
For most aerodynamic shape optimization problems, the objectives of interest are the forces
and moments — or the corresponding coefficients — acting on the body being optimized.
Computing the partial derivatives of these quantities with respect to the states and mesh
coordinates, ∂I/∂ζ and ∂I/∂x, is significantly simpler than the computation of the residual
partial derivatives. This becomes evident when comparing the routines required to compute the
residual (Figure 4.3) to the routines required to compute the forces and moments (Figure 4.4).
As these figures show, for the inviscid equations considered here, the force computation simply
ComputeForces
GetSurfaceCoordinates
ComputeSurfacePressures
ComputeSurfaceNormals
ApplyAllBCs BC EulerWall
ComputeForcesAndMoments
AdjustInflowAngle
ComputeCoefficients
requires an integration of the pressure over the surface of the body in question. This requires
an application of the boundary conditions as well as a surface normal computation, but this is
significantly simpler than the complicated flux computation required for the residual. Further,
the ratio of input variables to output variables strongly favors a reverse mode technique for
this computations. For a typical optimization problem, one might be interested in a few, force
Chapter 4. Time-Spectral ADjoint Method 55
and moment coefficients — for example, CL and CD for a lift constrained drag minimization
— while the surface needed to compute those coefficients may require hundreds, thousands
or even tens of thousands of surface cells to accurately discretize the surface. This yields an
extremely high ratio of input variables to output variables and thus strongly favors the reverse
mode approach. As with the residual routines, the force routines have been differentiated using
the reverse mode, source transformation capabilities of Tapenade [84, 42]. This differentiation
yields a routine that computes all of the state and coordinate derivatives of a specified force or
moment coefficient in a single pass.
The extension of this concept to the time-spectral case is relatively straightforward. In this
thesis, simple spectral objectives, such as the average lift, drag and moment coefficients are
considered. In these cases, the spectral objective is based on simple algebraic combinations of
the corresponding coefficient values at each of the discrete time instances in the solution. For
example, the average drag may be computed as:
1 1 1
CD = CD1 + CD2 + · · · + CDN (4.11)
N N N
where each of the coefficients, CDi is computed directly from the states of time instance i. While
the time instances of the residual computation are coupled through the spectral time derivative
of Equation (3.57), once the solution is computed, the computation of the coefficients at each
instance in time are independent. Therefore, so long as the objective function is a function
of these independent coefficients, the time instance based coefficients and their derivatives can
be used to form the time spectral objective derivatives. Using the chain rule, this idea can be
expressed as:
∂I ∂I ∂CDi
n
= . (4.12)
∂ζ ∂CDi ∂ζ n
In this case, the derivative of interest, ∂I/∂ζ n , a vector of size Ni × [Nζ × Ncells ], is computed
as a matrix-matrix product of two matrices of size Ni × N and N × [Nζ × Ncells ] respectively.
However, a reverse mode computation for the second term, ∂CDi /∂ζ n already exists from
the steady-state case. Because the coefficient computation on each of the time instances is
independent, once the first derivative matrix, ∂I/∂CDi , is computed it can be multiplied by the
result of the ∂CDi /∂ζ n computation, time instance by time instance, to generate the necessary
derivatives without ever storing the second matrix. This approach yields identically the same
ratio of inputs to outputs as the general steady case discussed above. The only distinction is
that in the spectral computation, the coefficient derivatives are run N times, once for each time
instance, while for the steady case, the derivatives are run once.
Chapter 4. Time-Spectral ADjoint Method 56
Based on the characteristics of the adjoint equations as well as the fact that both of the residual
sensitivity matrices can be computed at the same time, the matrices from Equation (4.8) are
both stored explicitly in memory for the solution of the adjoint equations. The matrices are
sparse for both the steady case and the time-spectral case. However, the time-spectral solution
has a slightly different sparsity pattern than the steady case. The sparsity patterns of the steady
and time-spectral cases for a 24 block H-H mesh are shown in Figures 4.5 and 4.6 respectively.
The two figures show that the overall sparsity pattern for the two cases are similar. Further,
the close up views show that what was a single block in the steady case has become a grouping
of N blocks in the spectral case. Note that there is an additional set of off diagonal entries
present in each row for each time instance in the solution. These off diagonal entries come
from the spectral operator that couples the time instances together. The grouping apparent in
Figure 4.6 is indicative of the decision to order the matrix such that all time instances of a given
block are adjacent in the matrix. For SUmb, this ordering provides much better performance
than the alternative, where all blocks of a given time instance are adjacent. This stems from
the fact that SUmb is parallelized by block, not by time instance. Thus, the derivatives for
all time instances of a given block are computed on the same processor. Grouping the cells in
the matrix by block instead of by time instance requires an immense amount of communication
Chapter 4. Time-Spectral ADjoint Method 57
Figure 4.6: Time-spectral sparsity pattern: 24 block H-H mesh: 3 time instances
Once the various terms are generated, the Portable, Extensible Toolkit for Scientific Com-
putation (PETSc) [6] is used to store the sparse derivatives and solve the linear system of
equations. More specifically, the linear solution is computed using a restarted GMRES iterative
solver — restarted after 150 subspace vectors — with an additive-Schwarz global preconditioner
and ILU(1) local preconditioning. The preconditioning matrix is computed using the lumped
dissipation technique of Hicken and Zingg [43]. This approach reduces the bandwidth of the pre-
conditioning matrix, thereby reducing the cost of generating the matrix as well as reducing the
memory requirements for storing and factoring the matrix. In addition, the off-time-instance
entries of the matrix are ignored when forming the preconditioner. This further reduces the
fill-in of the ILU preconditioner and contributes to a reduction in the memory required to solve
the system. Note that since the preconditioning matrix contains derivatives with respect to the
states only, the preconditioning matrix is computed with a reduced residual computation that
does not include the metric terms.
Chapter 4. Time-Spectral ADjoint Method 58
4.4 Verification
To verify the implementation of the time-spectral ADjoint, derivatives from the ADjoint are
compared to derivatives computed using the complex-step method [65]. For the complex-step
method, the sensitivity of a function, I(x), is computed as:
dI Im(I(x + ih))
= + O(h2 ) (4.13)
dx h
√
where i = −1 and h is an extremely small step — in this case 10−20 . Because the perturbation
is carried through the code in the complex portion of the variable, the normal subtractive
cancellation issues associated with the finite difference derivative technique are not present.
Thus, h can be made very small, reducing the O(h2 ) truncation error to negligible levels and
yielding numerically exact benchmark derivatives.
To benchmark the time-spectral adjoint for accuracy, derivatives are shown for a 917,000 cell
ONERA M6 wing mesh. The wing is simulated with Euler wall surfaces and a symmetry plane
at the root. The mesh has an H-H topology with a nominal off wall spacing of 0.002. The near
field symmetry plane of the mesh and a sample solution at Mach=0.8395 and α = 3.06 deg.
are shown in Figure 4.7.
4.4.2 Accuracy
The design variables considered for this case are those for a simple planform optimization,
specifically a simple linear twist distribution, a single value of sweep, as well as angle of attack.
Derivatives are considered for both pitching and plunging motions. Table 4.1 shows the values
for a pitching motion and Table 4.2 shows the values for a plunging motion.
As can be seen from Tables 4.1 and 4.2, the time-spectral ADjoint implementation is ex-
tremely accurate. For both the pitching and plunging motions, the derivatives of the coefficients
— CL , CD , Cm — match to between 8 and 12 digits, which is consistent with the accuracy of
the flow solution and adjoint solution used to generate them. The derivatives of the derivative
quantities — Cmq , Cmα , and Cmα̇ — are slightly less accurate, showing between 7 and 9 digits
of agreement. This is likely attributable to the fact that they are derivatives of derivatives.
None the less they are certainly accurate enough for the numerical optimizations conducted
in this work, which are typically converged to O 10−6 . Note also that the very high accuracy
of the results helps to ensure that there are no errors or omissions in the linearization of the
CFD code. This is important for gradient-based optimizations as errors in the gradients of the
Chapter 4. Time-Spectral ADjoint Method 59
Z X Y X
CoefPressure: -1.2 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
(a) ONERA M6 mesh: root slice, near field (b) ONERA M6 solution: M = 0.8395, α = 3.06 deg.,
917,000 cell mesh
Table 4.1: Sensitivity verification: ONERA M6, pitching motion, 917,000 cells: 10−12 relative
convergence
Chapter 4. Time-Spectral ADjoint Method 60
Table 4.2: Sensitivity verification: ONERA M6, plunging motion, 917,000 cells: 10−12 relative
convergence
Chapter 4. Time-Spectral ADjoint Method 61
objective and constraints can have a significant impact on the performance of gradient-based
optimization techniques.
4.4.3 Performance
To assess the computational cost of the method, the cost of the time-spectral ADjoint is com-
pared against the cost of the time-spectral flow solver, as well as the steady-state ADjoint. The
flow solver in this case is a Newton–Krylov solver implemented using the nonlinear solvers in
PETSc. The steady-state ADjoint is also solved using PETSc and implemented with the same
methods as the time-spectral ADjoint. The results of these comparisons are shown in Table 4.3.
Time-Spectral
Steady-State N =3 N =5 N =7 N =9
Number of processors 24 80 112 176 224
Number of flow states 4587520 13762560 22937600 32112640 41287680
Average # of flow states per processor 191147 172032 204800 182458 184320
Percent load imbalance 7.14 7.15 0 17.86 0
Flow solution time: (10−10 convergence) 1.08 1.56 1.87 2.46 2.34
Jacobian matrix assembly time 0.26 0.46 0.58 0.68 0.66
Preconditioner assembly time 0.03 0.04 0.08 0.1 0.13
Algebraic volume/surface sensitivity time 0.14 0.14 0.14 0.14 0.14
−10
Adjoint solution time: (10 convergence) 1.89 2.89 2.53 2.98 2.75
Total sensitivity time 0.23 0.23 0.23 0.23 0.23
Total adjoint time 2.54 3.76 3.56 4.12 3.92
Table 4.3: Time-spectral ADjoint computational cost breakdown for ONERA M6 (normalized
with respect to a total flow solution cost of 160.3 seconds)
Table 4.3 shows that the time-spectral ADjoint implementation is cost competitive with the
steady-state case and that the overall cost of the ADjoint method is reasonable. Comparing
on a case by case basis, the adjoint solution ranges between 1.6 and 2.4 times the cost of a
flow solution. Further the ratio of the steady case and the three time instance case are the
highest ratios, indicating that the implementation scales well. To quantify this more exactly,
the log of the total computational cost (normalized time multiplied by the number of flow
states) is plotted versus the log of the number of flow states in the problem (Ncell × Nζ × N ) in
Chapter 4. Time-Spectral ADjoint Method 62
Figure 4.8. In an ideal case, this plot would have a slope of one, indicating that the amount of
work required to solve the problem is scaling exactly with the problem size. Figure 4.8 shows
that the value of this slope for the current implementation of the flow solver is 1.38 and the
slope for the implementation of the time-spectral adjoint is 1.21. While not perfect, these values
show that the problem is scaling well. Looking at the performance from an absolute sense, the
4.5
Log of Total Computational Cost
3.5
Slope =1.21
3
Slope = 1.38
2.5
1.5
1
Flow Solution
Adjoint Solution
0.5
0
6.8 7 7.2 7.4 7.6
Log of number of flow states
implementation of the time spectral ADjoint allows for the solution of problems — including
both the flow solution and the adjoint solution — with over 41 million flow states in under
30 minutes on 224 Intel Nehalem processors. As will be shown in Chapter 6, at this level of
computational cost, optimization of periodic unsteady problems are manageable.
4.5 Summary
In this chapter, the time spectral ADjoint has been introduced, developed and verified. The the-
ory has been demonstrated on a meaningful sized case and has been shown to be both extremely
accurate and cost effective. Having demonstrated this technique for computing the gradients
neccesary for aerodynamic shape optimization with stability derivatives, the discussion can now
Chapter 4. Time-Spectral ADjoint Method 63
move to the remaining disciplines neccesary for the various stability-constrained optimizations
— such as center of gravity and moment of inertia calculations — and to a discussion of the
optimization results.
Chapter 5
Auxilliary Analyses
Aside from the CFD-based stability derivative calculations, which form the major component
of the development in this thesis, there are some additional analyses which are necessary for
the various stability constraints. The sections in this chapter outline each of these, why they
are necessary, and what is involved in each of the calculations.
The center of gravity (CG) location is fundamental for most of the calculations in this thesis.
Firstly, it is the point about which all aerodynamic moments are calculated. Therefore, the
moment coefficient, Cm , and all of its derivatives — in this case Cmα , Cmα̇ and Cmq — are
strongly dependent on the CG location. Secondly, the mass moment of inertia calculation uses
the CG location as a reference point. All of these quantities are necessary for the computation
of the stability parameters used as constraints. Therefore, having an appropriate CG location
is necessary to produce meaningful design optimization results.
For this work, a relatively simple wing CG calculation was implemented. The method is
derived from the work of Chai et al. [18]. In that work, the authors state that the wing CG
for a normal transport wing is located between the fore and aft spars along the wing mean
aerodynamic chord (MAC). Since flying wings are being examined in this study, it is assumed
that this estimate of the wing CG is a reasonable estimate of the CG for the entire aircraft.
The starting point for the calculation is the computation of the wing MAC and the location
of its quarter chord. These calculations are based on the methods presented in ESDU item
76003 [25]. One then determines the location of the intersections of the spars with the MAC. In
this work, the fore and aft spars are assumed to be at 25% and 75% of the MAC, respectively.
The longitudinal location of the CG can then be determined as a percentage — CG%M AC —
of the distance between the spars at the MAC — ∆xspar . Thus, a value of zero for CG% would
64
Chapter 5. Auxilliary Analyses 65
Wing Root
Forward Spar
deltaX spar
Rear Spar
CG Range:
% of Dist. Between spars
Wing Tip
place the CG at the intersection of the forward spar and the MAC, a value of one for CG%
would place the CG at the intersection of the rear spar and the MAC, and a value of negative
one for CG% would place the CG a distance of ∆xspar in front of the intersection of the forward
spar and the MAC. Figure 5.1 illustrates the key parameters in the approach.
The moment of inertia plays a key role in the dynamic stability constraints. Just as mass relates
force to acceleration, moment of inertia relates moment to rotational acceleration. As a result,
the pitch moment of inertia — Izz — shows up in the normalization of all of the moment terms
in the short-period approximation described in Chapter 2. Thus, an estimate for the moment
of inertia is required for the dynamic-stability constrained optimizations.
The calculations for the moments of inertia are based on first principles. The definition of
moment of inertia is: Z
I= r2 dm (5.1)
Ω
To compute this value, the domain of integration, Ω, is set to be the surface of the aircraft.
In this case, the surface is defined by a B-spline surface within the pyGeo geometry engine of
Chapter 5. Auxilliary Analyses 66
Figure 5.2: Discrete surface mesh with lumped masses at the cell centroids
Kenway et al. [49]. Having defined the surface in this fashion, a discrete surface mesh can be
created by evaluating the spline over a uniform distribution of values in u,v parameter space. A
sample wing mesh is shown in Figure 5.2. These discrete surface cells are then given a thickness
(t) and a density (ρ), and the area of each cell is computed using the following:
where v1 and v2 are the diagonal vectors of the cell, s is the projected area of the cell in
each coordinate direction and sTotal is the total area of the cell. The key variables in this
computation are illustrated in Figure 5.3. The area and thickness of the cell are multiplied
to get the volume of the cell, which when multiplied by the density gives the total mass for
Chapter 5. Auxilliary Analyses 67
S_Total
i,j = 2,2
i,j = 2,1
v1
v2
i,j = 1,2
i,j =1,1
each of the discrete cells on the surface mesh. These lumped masses are illustrated as spheres
in Figure 5.2. Considering this mass to be located at the centroid of the cell, the moment of
inertia can the be computed as:
X
I= mr2 . (5.6)
In this case, r2 is the distance from the lumped mass to the CG, perpendicular to the rotational
axis of interest. Therefore, for Ixx :
for Iyy :
r2 = (xmass − xCG )2 + (zmass − zCG )2 , (5.8)
In this thesis, the coordinate system used has x in the flow direction and z out the wing.
Therefore the moment of inertia of interest for the pitching calculations is Izz or:
X
m (ymass − yCG )2 + (xmass − xCG )2 .
I= (5.10)
Chapter 5. Auxilliary Analyses 68
Because this value approximates the moment of inertia of the wing itself, an additional multiplier
is added to the computation to account for the remaining mass distribution in the aircraft. This
parameter is given to the optimizer as a variable to allow it to satisfy the constraints in the
dynamic stability constrained optimization.
The root bending moment constraint is included to give some consideration to the structural
implications of the varying wing shapes. There are well known trade-offs between aerodynamic
and structural performance for wings, the most prominent of which has to do with the effects
of span. Increasing the span increases the aspect ratio of the wing, thereby reducing induced
drag. However, the same span extension also increases the bending moment at the root of the
wing, requiring a heavier structure to support the aerodynamic load. A similar effect is caused
by wing sweep. For shearing sweep — where individual sections of the wing are translated
in the flow-wise direction to sweep the wing — the effective structural length of the wing
increases as the wing is swept. This can be counteracted by reducing the span of the wing,
which introduces a trade-off between induced drag and wave drag at transonic Mach numbers.
Therefore, if span and sweep are used as design variables, it is necessary to have some method
to account for these trade-offs between aerodynamic performance and structural weight. In
this thesis, the root bending moment is used as a proxy for the structural performance of the
wing. The assumption implicit in this approach is that two wings with the same root bending
moment will require the same amount of material to support the load on the wing and thus
have the same weight. This is a relatively simplistic assumption, but serves as a useful metric
for including the effects of structural considerations in the optimization.
To compute the bending moment coefficient, the pressure is integrated over the aircraft to
get the force and moment coefficients about a reference point, xref . This computation yields
the values of Cfx , Cfy , Cfz , Cmx , Cmy and Cmz , which are the force and moment coefficients in
the three principle Cartesian axes. Note that x is the flow direction, y is the vertical direction
and z is the span-wise coordinate. A bending reference point, xbending , is then specified at the
root of the wing. This is the point about which the net bending moment is calculated. The
bending moment is then calculated as:
z
Wing Tip
Wing Root
X_ref
Cmz
C_bending_z
C_bending_x
C_b
where c̄ref is the reference chord length, Cb is the total root bending moment coefficient and
Cbendingx and Cbendingz are the Cartesian components of the root bending moment about the
bending reference point a the wing root.
A diagram illustrating the various components of this calculation is shown in Figure 5.4.
Note that since the torsional component of the wing moment is non-zero, the effective bending
moment is not necessarily aligned with the wing. Also note that as the wing is swept, this
torsional component increases, causing the combined effective bending moment to sweep with
the wing. Therefore, it is assumed that the wing can be supported normal to the effective
total bending moment, regardless of the wing sweep and pressure distribution, Thus, only the
total magnitude of the bending coefficient, Cb , is considered. The reference value of Cb used to
constrain the wing can have a significant effect on the optimal design. In this study, the value
from the elliptical optimum — computed at M = 0.5 — is used as the reference value. This
value is then scaled with the root cross section to ensure that as the optimizer increases the
aspect ratio of the wing, the allowable bending moment is reduced to account for the reduced
second moment of area at the wing root. The reasoning behind this scaling is shown below.
The definition of bending stress is:
My
σ= (5.12)
I
Rearranging for the moment gives:
σI
M= (5.13)
y
In this work, scaling the chord of the wing also scales its thickness, so both the thickness
and chord effects can be considered together. Therefore, doing a dimensional analysis on the
moment we get:
σL4
M∝ = σL3 . (5.14)
L
Chapter 5. Auxilliary Analyses 70
However, because the wing is assumed to have a constant t/c ratio, c̄ is also of dimension L.
Also, A is constrained to be constant in the optimizations, therefore
σL3
Cbending ∝ (5.17)
qAL
or,
σL2
Cbending ∝ (5.18)
qA
Thus, with q and A kept constant, the allowable bending moment must be scaled with L2 to
enforce a constant allowable bending stress at the wing root. L is the scaling factor applied to
the root chord of the wing.
As stated previously, the purpose of this work is to develop a practicable means of including
stability constraints based on three-dimensional CFD methods into an aerodynamic optimiza-
tion problem. The approach adopted here is to model the stability characteristics of the aircraft
with a linear flight dynamic model, which allows the aerodynamic portion of the problem to be
modeled with linear stability derivatives. The remainder of this section outlines, more specifi-
cally, how these stability derivatives, and the resulting linear flight dynamics model, are used
to formulate stability constraints for the optimization problem. Three stability constraints are
considered, two based on static stability theory and one based on dynamic stability theory.
A discussion of other stability criteria applied to a flying wing configuration can be found in
Agenbag [1]
The first static constraint is based on the definition of longitudinal pitch stiffness. To develop
this criteria, consider first the definition of equilibrium. For an aircraft to be in equilibrium,
the sum of the forces acting on the CG and the sum of the moments about the CG must be
zero. That means that lift must equal weight — requiring a CL > 0 — and CmCG must also be
zero. Further, the aircraft should return to this equilibrium condition when subject to minor
perturbations. Making the assumption that significant perturbations occur only in angle of
Chapter 5. Auxilliary Analyses 71
attack, α, and that a nose up pitching moment is positive; then to return to equilibrium, a nose
down — or negative — pitching moment must be developed when α increases. This requires
that the derivative ∂CmCG /∂α be negative. Therefore, the conditions for static stability in pitch
are:
CL > 0
CmCG = 0 (5.19)
∂CmCG
< 0
∂α
This set of conditions enforces the limiting case of static pitch stability and is implemented as
shown in Section 6.3.
As will become evident from the results in Chapter 6, the limiting case of the definition does
not always yield a useful result. If the constraint is active, the design ends up on the limiting
bound of the stable envelope, which is not generally a desirable quality for an aircraft. This
raises the question of how to quantify a desirable value for Cmα . For this, one can turn to the
definition of static margin. First, consider the definition of the moment coefficient,
Where hCG and hNP are the streamwise locations of the CG and neutral point respectively,
normalized by the MAC. Differentiating with respect to α yields,
∂CmCG ∂CL
= (hCG − hNP ) (5.21)
∂α ∂α
or, using stability derivative notation,
Defining the static margin, Kn , as the distance between the CG and the neutral point normalized
by the MAC,
Kn = hNP − hCG (5.23)
and substituting this relationship back into Equation (5.22) results in,
Thus, by specifying a static margin, one can determine a meaningful value for Cmα .
To this point the discussion has been completely general and applicable to any aircraft. By
examining Equations (5.20) and (5.22), one can gain insight as to how the various stablility
conditions may be satisfied for a flying wing. Equation (5.20) shows that the moment coefficient
is dependent on the moment about the neutral point — which does not change with angle of
attack — and the location of the neutral point with respect to the CG. Thus, the optimizer
can trim the aircraft by altering one of three parameters. It can alter the moment coefficient
about the neutral point, accomplished primarily with airfoil shape — adding reflex to the airfoil
— or by twisting down the wing tips on a swept wing. It can alter the neutral point of the
aircraft, accomplished primarily by sweeping the wing. Finally, it can alter the CG location
of the aircraft, which is also affected by wing sweep, as well as payload location. However,
Equation (5.22) shows that these last two quantities also impact Cmα and hence the static
margin. Thus, conceptually, the optimizer must first adjust the sweep and payload location to
create an acceptable static margin, then it must trim the aircraft using airfoil shape and section
twist. Practically speaking the optimizer will adjust the parameters simultaneously to satisfy
all of the necessary constraints. These various elements will be discussed along with the results
in Chapter 6.
The control anticipation parameter (CAP) is a method used to quantify the handling qualities
of an aircraft based on its short-period characteristics. The fundamental idea behind this
approach is that a pilot’s ability to fly an aircraft precisely along a given flight path is related
to the pilot’s ability to anticipate response of the aircraft. In this approach, Bihrle [7] relates
the pilot’s ability to anticipate the aircraft’s response to a ratio between the instantaneous
pitch acceleration of the aircraft, q̇, and the steady-state normal acceleration of the aircraft,
∆n. This can be expressed as:
q̇
CAP = (5.26)
∆n
However, these quantities are not necessarily simple to evaluate for an aircraft. Therefore, with
a little rearranging, Bihrle [7] provides the following expression:
ωn2
CAP = (5.27)
nα
The United States Military has specified acceptable limits for CAP and damping ratio for
various combinations of aircraft and flight conditions [73]. For an aircraft at cruise (Category
B), those limits are listed in Table 5.1.
Table 5.1: Stability bounds for damping ratio and CAP parameter
The geometry for the optimizations detailed in Chapter 6 is manipulated using a multi-layered
approach. The top level geometry definition is handled by pyACDT [86]. This geometry definition
is based on a conceptual design methodology, so the wing is defined in terms of individual linear
segments, each with a defined area, span, taper ratio, sweep and twist. The design variables
presented in Section 6.2 are configured such that each section comprises one ninth of the area
and span specified by the optimizer. The sweep of all nine sections are defined to be the
same such that the leading edge of the wing is always straight. Finally, the twist variables are
configured such that the tip twist of one section is the same as the root twist of the section
immediately outboard of it, yielding nine independent values of section twist. The root twist
angle of the wing is fixed at zero degrees, to avoid duplicating the design degree of freedom
of the angle of attack. Note that the conceptual planform parameters from this layer are also
used in the CG computation outlined in Section 5.1.
The construction of the detailed geometry is handled by pyPSG as presented by Kenway et
al. [49]. This tool is used to create a watertight surface representation of the aircraft, a set of
spline volumes enveloping the aircraft and a set of reference axes inside the aircraft. The set
of volumes enveloping the aircraft are used in conjunction with a free form deformation (FFD)
technique to handle the detailed geometry manipulations during the optimization [49]. The
surface points on the CFD mesh are embedded parametrically in these spline volumes, such
that when the volumes are moved or deformed, the CFD surface mesh is modified as well. The
reference axes are used to tie together the control points of the volumes enveloping the aircraft.
By tying the control points together with a reference axis, the effects of physical design variables
— such as sweep, twist, taper, area, span and chord — can be recreated. To implement these
Chapter 5. Auxilliary Analyses 74
design variable groupings, the pyACDT layer has been modified to compute the appropriate
coefficients required by the reference axis to transfer the motion from the conceptual geometry
to the FFD volumes, and onto the CFD surface. A graphic depicting the surface, FFD and
reference axes is shown in Figure 5.5. Note that the lines between the spheres represent the links
between the FFD volume control points and the reference axis, the large rectangular volume is
the FFD volume and the line along the trailing edge marked by the small cubes is the reference
axis.
Figure 5.6 further demonstrates the approach, showing a deformed geometry. This defor-
mation includes a span extension from 3m to 3.5m, a sweep of two degrees, a twist of negative
ten degrees near mid-span and a tip twist of positive ten degrees.
Mesh Deformation
Once the CFD surface mesh is deformed, those deformations need to be propagated to the
remainder of the CFD volume mesh. This is handled using a hybrid, finite element – algebraic
mesh deformation technique, as outlined in Kenway et al. [49]. In this technique, the CFD
mesh is represented by a very coarse finite element mesh. This mesh is deformed using the
perturbations of the coarse mesh surface nodes as input values. When the finite element system
is solved, the large surface deformations are propagated to the remainder of the volume mesh.
Once the large deformations are handled, an algebraic mesh deformation scheme — based on
trans-finite interpolation — is used to handle the detailed surface deformations on the blocks
immediately adjacent to the surface. This completes the two stage mesh deformation process
and provides a fully updated volume mesh to the CFD Solver.
Chapter 5. Auxilliary Analyses 75
To demonstrate the effectiveness of the methods developed in this thesis, a series of design
optimization problems is solved with a simple rectangular wing as the baseline design. The wing
has a half span of three meters and a chord of one meter, giving the wing an aspect ratio of six.
The wing has a taper ratio of one and a leading edge sweep angle of zero degrees. The initial
airfoil section has a NACA 0012 profile. Details of the geometry are summarized in Table 6.1
and a perspective view of the wing is shown in Figure 6.1. This test case is based on the induced
drag validation case proposed by Hicken and Zingg [44]. Hicken and Zingg demonstrate that
using this wing, with sections twisted about the trailing edge, one can reproduce the elliptical
distribution outlined by lifting line theory. In particular, they highlight the use of a straight
trailing edge, with spanwise sections twisted about the trailing edge, to minimize the impact
of nonplanar effects in the wake. This configuration is used as the initial geometry in this
set of optimizations so that the twist-only optimization outlined in Section 6.3.1 reproduces
the elliptical result and can be used as a point of comparison for the remaining optimizations.
This will provide a means of quantifying the effect of the various stability constraints on the
optimal solution. All of the optimization results presented are computed on an 1,105,920 cell
mesh. The mesh has a C-O topology and is split into 32 blocks. The farfield boundary is
approximately 15 chords from the wing and the off wall spacing of the mesh is 1 × 10−3 m at
the leading edge and 5 × 10−4 m at the trailing edge. Note that all optimizations conducted in
this thesis use the pyOpt [87] interface to SNOPT [32]. pyOpt is a Python based optimization
framework that wraps a variety of commercial and open source optimization packages with a
common Python interface [87], allowing for a straightforward interchange of various optimizers
for a given problem. SNOPT is an optimization package based on the SQP method for the
solution of large-scale non-linear optimization problems [32]
76
Chapter 6. Design Optimization Results 77
Parameter Value
Half wing area (m) 3.0
Half wing span (m) 3.0
Chord (m) 1.0
Leading edge sweep (deg.) 0.0
Taper ratio 1.0
Wing tip washout (deg.) 0.0
Wing dihedral 0.0
In order to better understand the optimization problems presented in this chapter and the choice
of variables used herein, it is necessary to provide a brief overview of multidisciplinary design
optimization (MDO) architectures. Generally speaking, an MDO architecture is a method used
to handle the coupling between disciplines in a multidisciplinary optimization problem. Numer-
ous architectures have been developed to handle this coupling and the various methods can be
divided into two broad categories: monolithic architectures and decompositional architectures.
Monolithic architectures — such as the multidisciplinary feasible (MDF) architecture [21], the
individual discipline feasible (IDF) architecture [21] and the simultaneous analysis and design
(SAND) [39] architecture — set up the problem as a single level optimization and handle the
objective function optimization and interdisciplinary coupling in one optimization problem.
Chapter 6. Design Optimization Results 78
x yt
0, 7→1:
x∗ Optimization
2 : x0 , x1 3 : x0 , x2 4 : x0 , x3 6:x
1, 5→2:
2 : y2t , y3t 3 : y3t
MDA
2:
y1∗ 5 : y1 3 : y1 4 : y1 6 : y1
Analysis 1
3:
y2∗ 5 : y2 4 : y2 6 : y2
Analysis 2
4:
y3∗ 5 : y3 6 : y3
Analysis 3
6:
7 : f, c
Functions
x, y t
0,3→1:
x∗ 2 : x, y t 1 : x0 , x1 , y2t , y3t 1 : x0 , x2 , y1t , y3t 1 : x0 , x3 , y1t , y2t
Optimization
2:
3 : f, c, cc
Functions
1:
y1∗ 2 : y1
Analysis 1
1:
y2∗ 2 : y2
Analysis 2
1:
y3∗ 2 : y3
Analysis 3
of the force and moment coefficients, as well as the stability derivatives for the aircraft, the
structures discipline computes the bending coefficient and the stability discipline computes the
static margin, CAP and damping ratio for the aircraft. In this thesis, the task of coupling
the disciplines is simplified greatly by the fact that, in this case, there is no feedback between
disciplines. Thus, the tightly coupled MDF architecture amounts to a sequenced evaluation of
the disciplines. However, when using the adjoint methods described in this work, coupling the
derivatives of the various disciplines is not as straightforward. This motivates the use of the IDF
architecture. In this context, the IDF architecture allows the individual discipline sensitivity
analyses to function unmodified and greatly simplifies computation of accurate gradients.
Despite this apparent advantage, there are some disciplines for which tight coupling makes
more sense. First of all, the coupling between the geometry and the aerodynamics requires a
large number of coupling variables. The number of mesh points on the surface can be in the
order thousands or more. In an IDF architecture, this would add thousands of variables and
constraints to the optimization problem, which would increase the complexity of the optimiza-
tion problem unnecessarily. By tightly coupling the geometry surface to the aerodynamics,
those coupling variables do not become design variables and the optimization is simplified sig-
nificantly. The other section where tight coupling makes sense is between the aerodynamics and
structures. The bending coefficient calculation is based on six force and moment coefficients, of
which only one , Cmz , is already present in the optimization. Separating these two disciplines
would require the solution of five additional adjoint problems per optimization iteration. By
tightly coupling the aerodynamics and structures, this can be reduced to a single additional
adjoint problem for the bending coefficient. A detailed overview of each optimization problem
is presented with the results for each formulation in the following sections.
The study presented in this chapter involves optimizing the baseline wing design by solving
different optimization problems to demonstrate the usefulness of the proposed approach. The
various optimization problems — outlined in Section 6.3 — are studied at three different Mach
numbers: 0.5, 0.7 and 0.85. This allows the effects of the various constraints to be considered
in both the subsonic and transonic regimes.
Two sets of design variables are used: one involving only planform variables and one involving
planform variables as well as 280 surface shape design variables. These surface shape design
variables — the control points of the FFD volume — modify the detailed surface shape of the
Chapter 6. Design Optimization Results 81
wing and affect both the stream-wise and span-wise profile of the wing. The planform-only
results were run at the two lower Mach numbers, while the optimizations including the surface
shape variables were run at all three Mach numbers. The design variables and their limits are
outlined in Tables 6.2 and 6.3. Table 6.2 shows the primary design variables that alter shape
and planform of the wing as well as the variables that affect the flow condition of the test case,
while Table 6.3 shows the compatibility design variables introduced as a result of the hybrid
IDF/MDF architecture used to solve the problems.
The relationship between each variable and the optimization problems can be described as
follows:
• Primary Variables:
– Section twist, θi : Variables to modify the twist values for individual sections
along the wing. The sections rotate about the reference axis, which for this
study is located at the wing trailing edge. These variables primarily affect the
spanwise lift distribution, which in turn affects the induced drag created by
the wing.
– Area, A1 : The planform area of half the wing. This value is divided equally
among the sections specified in pyACDT. Note that this value remains essentially
constant. It varies to allow the calculated value of area, A2 , meet the constraint.
This is necessary because A2 varies slightly with wing twist and sweep.
– Span, b: The span of the half wing. This value is divided equally among the
sections specified in pyACDT. Since the area is essentially fixed this variable also
determines the chord of the wing. The span has a strong impact on both the
induced drag and the root bending moment coefficient of the wing.
– Sweep, λ: The leading edge sweep of the wing. This variable affects both the
wave drag and the root bending moment coefficient of the wing.
– Center of gravity variable, CG% : A variable that controls the x location of
the aircraft center of gravity. This value represents the value of the location of
the center of gravity in terms of a percentage of the distance between the front
and rear spars at the MAC. Modifying this variable is equivalent to modifying
xCG and is used to provide a nicely scaled range of values for the optimizer to
work with.
– Izz modifier, PIzz : A factor that multiplies the pitch moment of inertia. The
base value of moment of inertia is calculated using the method described in
Section 5.2. This variable is a direct multiplier of that calculated value.
– FFD control points: y-offset, ∆FFD : Variables controlling the vertical
motion of the FFD control points. These variables modify the shape of the
wing surface and have a significant impact on the pitching moment of the wing
as well as the wave drag caused by the wing.
– Moment reference location, xtCG : An IDF target value that provides the
x location of the reference point used to calculate the moment coefficient, Cm ,
and all related quantities. This value is constrained to be equal to xCG at the
optimal solution.
Chapter 6. Design Optimization Results 83
– Rotation point, xtCG : An IDF target value that provides the x location of
the rotation point for the flow solutions that involve a pitching motion. This
value is constrained to be equal to the xCG at the optimal solution.
– Chord, ctroot : An IDF target variable for the root chord length. This variable
is used to scale the allowable root bending moment coefficient, Cbref , and is
constrained to be equal to the value of root chord calculated in pyACDT.
– Target MAC, MACt : An IDF target variable for the reference chord length
used in the CFD solver. Constrained to be equal to the MAC computed from
the conceptual geometry definition at the optimal solution.
t
– Target drag coefficient, CD : A variable resulting from the IDF architecture.
The variable is used as an input to the various stability disciplines. The value
is constrained to be equal to the drag coefficient predicted by the CFD solver
at the optimal solution.
– Target lift curve slope, CLt α : A variable resulting from the IDF architecture.
The variable is used as an input to the various stability disciplines. The value
is constrained to be equal to the CLα value predicted by the CFD solver at the
optimal solution.
t
– Target moment curve slope, Cm α
: A variable resulting from the IDF ar-
chitecture. The variable is used as an input to the various stability disciplines.
The value is constrained to be equal to the Cmα value predicted by the CFD
solver at the optimal solution.
t
– Target α̇ derivative, Cm α̇
: A variable resulting from the IDF architecture.
The variable is used as an input to the various stability disciplines. The value
is constrained to be equal to the Cmα̇ value predicted by the CFD solver at the
optimal solution.
t
– Target pitch derivative, Cm q
: A variable resulting from the IDF architec-
ture. The variable is used as an input to the various stability disciplines. The
value is constrained to be equal to the Cmq value predicted by the CFD solver
at the optimal solution.
Note that not all design variables are used for all cases. A number of the variables listed
are present as a function of the IDF architecture mentioned previously and are, therefore, only
present when the associated disciplines are included in the problem statement. The specific sets
of variables used in each problem are shown in the problem statements presented in Section 6.3.
Chapter 6. Design Optimization Results 84
Figure 6.4: Thickness constraints: constraint locations and initial values represented by the
vertical line segments marked by the black spheres
There are two additional constraints imposed for cases that include shape design variables.
The first constraint is a thickness constraint, included to keep the optimizer from making the
wing too thin. This is implemented by computing thicknesses at a variety of locations around
the aircraft and constraining them. At the optimal solution, they must be no smaller than
their initial value. The locations where these thicknesses are evaluated are defined by the
user. The user specifies an enclosed portion of the planform, inside which the thicknesses
are to be constrained. The user then defines the number of spanwise and chordwise sections
to be considered inside this area, which generates a grid of points at which the thicknesses
are constrained. Figure 6.4 demonstrates this process for the current test case. The vertical
line segments marked by the black spheres indicate the location and initial thicknesses of the
thickness constraints.
The second constraint concerns the control points of the FFD at the leading and trailing
edges of the wing. If, at either of these locations, the control points are allowed to move in
the same direction vertically, they have the same impact as a twist variable, leading to an
ill-defined problem. This is demonstrated in Figure 6.5. In this figure, the top left image shows
the unperturbed geometry, the top right image shows a geometry with the tip twisted upwards,
the bottom left image shows the geometry with the leading edge control points perturbed in the
same direction, and the bottom right image shows the geometry with the leading edge control
points perturbed in opposite directions as well as a second set of control points perturbed
upwards. As can be seen from the second and third images, both twisting the wing and moving
Chapter 6. Design Optimization Results 85
the leading edge control points in tandem produce a change in the effective twist angle of the
section. In the fourth figure, where the leading edge points are moved in equal and opposite
directions and the interior control points are moved together, there is a significant change in
the section shape, but the effective twist of the section does not change. Therefore, the control
points at the leading edge and trailing edge are constrained to move in equal and opposite
directions.
To assess the effects of stability constraints on the aerodynamic shape optimization results,
a series of optimization problems is considered. The following section describes the various
formulations. The three optimization problems in Section 6.3.1 are a set of reference optimiza-
tions, designed to provide a basis for comparison with the stability constrained cases. The
two optimization problems in Section 6.3.2 are those based on static longitudinal stability con-
straints and are designed to show the impact that even basic stability constraints can have on
the optimal design. The final optimization, shown in Section 6.3.3 is based on dynamic longi-
tudinal stability constraints and shows the full impact of the methods developed in this work.
The problem statements are interspersed with figures showing the results of the optimizations.
These figures are used to demonstrate the effects of the different formulations for the different
combinations of design variables and Mach number. More detailed quantitative results and the
associated discussion are presented in Section 6.4.
Baseline Problem
As a first step in the study, a problem is formulated to find the angle of attack, α, and CG
location, xCG , that yield a trimmed aircraft at the target value of CL for each Mach number.
These problems are formulated as:
minimize CD
w.r.t. x = α, xCG (6.1)
subject to CLref − CL ≤ 0
Cm = 0
or in graphical form as shown in Figure 6.6. Note that the lift coefficient constraint is formu-
lated as an inequality constraint rather than an equality constraint. Because induced drag is
Chapter 6. Design Optimization Results 86
Figure 6.5: FFD control point constraints: initial configuration (upper left), perturbation in
section twist (upper right), unconstrained perturbation of leading edge control points (lower
left), constrained perturbation of leading edge control points as well as a perturbation of mid-
chord control points (lower left)
Chapter 6. Design Optimization Results 87
x : α, xCG
0,3→1:
α∗ , x∗CG 1 : α, xCG
Optimization
2:
f : CD
3 : f, c
c : CLref − CL
Cmz
1:
CL∗ , CD
∗ ∗
, Cm 2 : CL , CD , Cm Aerody-
namics
proportional to CL2 , minimizing CL also minimizes drag, therefore the upper bound on the CL
constraint is unnecessary.
The solutions for Mach = 0.5, 0.7 and 0.85 are shown in Figures 6.7 through 6.9. These
solutions show that the solutions at the three different Mach numbers each have different
characteristics. The solution at M = 0.5 is fully subsonic, with a correspondingly peaky Cp
distribution on the airfoil. Both the neutral point of the wing and the required center of gravity
to trim are just slightly forward of the wing quarter chord. The M = 0.7 case is just starting
into the transonic regime and the inner half of the wing is exhibiting a weak shockwave near
the leading edge. The location of the center of gravity to trim the aircraft is slightly further
forward than in the M=0.5 case, but is still near the quarter chord. The neutral point has
also shifted slightly forward. The M = 0.85 case is fully in the transonic regime and exhibits a
strong shockwave on the aft half of the airfoil. As a result, the neutral point and the center of
gravity location to trim are significantly farther back, near the half chord.
Twist Optimization
As a second metric for comparison, a simple, twist-only optimization has been completed for
the subsonic — Mach = 0.5 — case. This problem is designed to reproduce the elliptical result
Chapter 6. Design Optimization Results 88
minimize CD
w.r.t. x = α, θi , CG% , xtCG (6.2)
subject to CLref − CL ≤ 0
Cm = 0
xtCG − xCG = 0
The data dependencies and solution process are shown in Figure 6.10. In this optimization,
twist variables θi are added to control the lift distribution, and the full CG calculation is added
to control the location of the CG. Note that the aircraft is trimmed during the optimization
by forcing the pitch moment coefficient about the CG, Cm , to be zero. However, since the
location of the CG is allowed to move — through the variable CG% — the aircraft can be
trimmed without affecting the aerodynamic shape. This formulation yields the solution shown
in Figure 6.11.
The lift distribution achieved with this solution clearly shows an optimal solution that
closely approximates the elliptical solution. The exception to this is the area near the tip. This
location consists of complex three-dimensional flows that violate the assumptions used to derive
Chapter 6. Design Optimization Results 90
0,4→1:
α∗ , θi∗ , CG∗% 3 : xtCG ,MACt 1 : θi , CG% 2 : α, xtCG
Optimization
3:
f : CD
c
4 : f, c, c
c : CLref − CL cc : xCG − xtCG
Cmz MAC − MACt
1:
x∗CG ,MAC∗ 3 : xCG ,MAC 2: Surface
Geometry
2:
CL∗ , CD
∗ ∗
, Cm z
3 : CL , CD , Cmz Aerody-
namics
the elliptical result; therefore, it is expected that the lift distribution in this region does not
match the elliptical distribution.
minimize CD
w.r.t. x = α, θi , A1 , b, λ, CG% , ∆FFD , ctroot , xtCG , MACt (6.3)
subject to CLref − CL ≤ 0
Cm = 0
Cb − Cbref = 0
A2 − Aref = 0
xCG − xtCG = 0
MAC − MACt = 0
croot − ctroot = 0
Figure 6.12 depicts the data and process flow for this formulation in graphical form. The
root bending moment constrained optimizations have the same design variable flexibility as the
stability constrained cases described in Sections 6.3.2 and 6.3.3, but are constrained only by the
root bending moment. Therefore they are able to produce shapes with lower drag than those
cases with stability constraints. These optimizations provide the best indication of how much
performance is sacrificed to ensure that the stability constraints are satisfied. The results for
the planform-only optimizations are shown in Figures 6.13 and 6.14, while the results including
shape variables are shown in Figures 6.15 and 6.16. Note that there is no shape optimization
result for the M=0.5 case as the airfoil shape has very little impact on the drag of a subsonic
wing in inviscid flow.
Comparing the elliptical optimum solution shown in Figure 6.11 to the bending constrained
case at M=0.5, shown in Figure 6.13, there are some subtle yet important differences. In the
bending constrained case, the optimizer has taken advantage of the span variable to reduce the
induced drag. There is a visible extension in the span of the wing as compared to the original
Chapter 6. Design Optimization Results 92
x : α, θi , A1 , b,
λ, CG% , ∆FFD ,
MACt , xtCG ,
ctroot
4:
f : CD
5 : f, c, cc c : CLref − CL
cc : xCG − xtCG
Cmz
MAC − MACt
Cb − Cbref (ctroot )
croot − ctroot
A2 − Aref
1:
x∗CG ,M AC ∗ ,ctroot 4 : xCG ,MAC,croot 2: Surface
Geometry
3:
Cb∗ 4 : Cb Bending
Analysis
Figure 6.12: XDSM for the root bending moment constrained problem
Figure 6.13: Bending moment constrained optimization: planform variables only: M=0.5,
e = 0.964
Chapter 6. Design Optimization Results 93
Figure 6.14: Bending moment constrained optimization: planform variables only: M=0.7
Figure 6.15: Bending moment constrained optimization: planform and shape variables: M=0.7
Chapter 6. Design Optimization Results 94
Figure 6.16: Bending moment constrained optimization: planform and shape variables: M=0.85
wing, which is shown as a black outline. This increase in span is limited by the allowable
root bending moment and comes with a corresponding movement away from an elliptical lift
distribution.
Looking at the M=0.7 case, shown in Figure 6.14, the optimizer has added a significant
amount of sweep to the wing. This causes a trade-off between induced drag and wave drag.
The added sweep reduces the effective Mach number normal to the wing leading edge, thereby
reducing the strength of the leading edge shock. With sufficient sweep, the Mach number
normal to the leading edge of the wing is reduced below the critical Mach number, eliminating
the shock wave and the corresponding wave drag. However, adding sweep increases the effective
root bending moment. Thus, as sweep is added, the span must be reduced to meet the bending
constraint. This is evident in the difference in span between the Mach = 0.5 case shown in
Figure 6.13 and the Mach = 0.7 case shown in Figure 6.14.
When shape variables are added, the need for sweep to reduce wave drag is eliminated and,
as shown in Figure 6.15, the optimizer does not sweep the wing, allowing for a larger span.
Further, there is now a significant change in the Cp distribution on the wing. The individual
sections now show a rooftop pressure profile at the leading edge, allowing the maximum section
Cp to stay below the critical Cp for the wing, thereby eliminating the wave drag on the wing.
Note that the planform is now similar to the optimum M = 0.5 planform in Figure 6.13.
Chapter 6. Design Optimization Results 95
At M = 0.85, the optimizer has produced the features typical of supercritical transonic
airfoils for the individual airfoil sections. The sections show a rooftop Cp profile and are highly
aft loaded. However, even with these transonic airfoils, the wave drag is not entirely eliminated,
so the optimizer has introduced some sweep in the design to mitigate this effect. This added
sweep has caused a corresponding reduction in span to maintain the required bending moment.
Note that because of the aft loaded nature of these airfoils, the neutral point of the wing is
significantly forward of the required CG location for trimmed flight, leading to a statically
unstable configuration.
The first stability constrained case is based on the traditional definition of static-stability. As
described in Section 5.4, this definition requires the aircraft to have positive pitch stiffness, at
a trimmed state, with positive lift. The formulation for this case is shown in Figure 6.17 an
can be written mathematically as:
minimize CD
w.r.t. x = α, θi , A1 , b, λ, CG% , ∆FFD , ctroot , xtCG , MACt (6.4)
subject to CLref − CL ≤ 0
Cm = 0
Cmα ≤ 0
Cb − Cbref = 0
A2 − Aref = 0
xCG − xtCG = 0
MAC − MACt = 0
croot − ctroot = 0
In this problem, the previous trim constrained case is made stricter by requiring that Cmα < 0.
As a result, the center of gravity variable is no longer as free to move to trim the aircraft. It
now has to satisfy both the Cm constraint and the Cmα constraint. As is shown in the results,
this creates a trade-off for the optimizer to deal with and frequently leads to changes in the
optimal configuration.
As shown in Table 6.4, the optimal wing at Mach = 0.5 is very slightly unstable. In order
to satisfy the Cmα constraint, the optimizer adds sweep to the wing, resulting in the wing
shown in Figure 6.18. As discussed in Section 5.4, stabilizing the wing requires manipulating
Chapter 6. Design Optimization Results 96
x : α, θi , A1 , b,
λ, CG% , ∆FFD ,
MACt , xtCG ,
ctroot
4:
f : CD
c : CLref − CL
5 : f, c, cc Cmz cc : xCG − xtCG
Cb − Cbref (ctroot ) MAC − MACt
A2 − Aref croot − ctroot
Cmα < 0
1:
x∗CG , M AC ∗ ,c∗root 4 : xCG ,MAC,croot 2: Surface
Geometry
CL∗ , CD∗ ∗
, Cm , 4 : CL , CD , Cmz , 2: 3 : Cfx , Cfy , Cfz ,
z
Cm∗
, A∗2 Cmα , A2 Aerody- Cmx , Cmy , Cmz
α
namics
3:
Cb∗ 4 : Cb Bending
Analysis
the relative locations of the CG and the neutral point. Sweeping the wing shifts the neutral
point of the wing further aft. This will nominally increase the stability of the aircraft, if the
CG stays fixed. However, sweeping the wing also shifts the CG further aft. In this thesis, the
extra degree of freedom provided by the CG% variable allows the optimizer to shift the CG
forward to compensate for shift caused by the sweep. This simulates moving the positions of
the internal systems and payload to provide a feasible CG range. Unfortunately, as shown in
Equation (5.20) shifting the distance between the CG and the neutral point also affects the
trim of the aircraft. However, the addition of sweep to the wing also allows the optimizer to
use washout of the wing tip to provide the pitch moment needed to trim the aircraft in its new
state. Once shape variables are added, this additional sweep is not necessary to stabilize the
aircraft. As shown in Figure 6.19, the optimizer is able to use the shape variables to modify the
Cp to achieve the same effects. In this case, the Cp distribution, especially near mid-span, has
developed negative lift at the trailing edge, which alters the CmNP of the wing. This allows the
optimizer to shift the CG forward — stabilizing the wing — using CG% while still maintaining
a trimmed state. Because the wing is in a stable trimmed state without sweep, there is no
increase in the bending moment associated with sweep, therefore the optimizer is again able to
increase the span from the baseline value of three meters, thereby increasing the aspect ratio
of the wing and reducing the resulting induced drag.
Moving to the Mach = 0.7 case, the results in Table 6.6 show that for the planform only
case, the Cmα constraint is not active. Because of the amount of sweep necessary to counteract
the wave drag present at the transonic Mach number, the optimal wing from the bending
Chapter 6. Design Optimization Results 97
Figure 6.18: Cmα constrained optimization: planform variables only: M=0.5, e = 0.963
Figure 6.19: Cmα constrained optimization: planform and shape variables: M=0.5, e = 0.968
Chapter 6. Design Optimization Results 98
constrained case has a positive static margin. Thus, the solution with the Cmα constraint,
shown in Figure 6.20, is essentially the same as the bending constrained solution shown in
Figure 6.14. The same is not true of the shape optimization cases at Mach = 0.7. The bending
constrained case from Figure 6.15 has a negative static margin. Thus, Cmα constrained case
shown in Figure 6.21, the optimizer uses the shape variables to alter the Cp distribution of the
wing to trim the wing in a stable state. In this case, in addition to adding the negative lift
at the trailing edge, the optimizer has flattened the Cp distribution at the leading edge of the
wing, eliminating the pressure peak and thereby reducing the wave drag. As a result, the wing
is once again unswept, allowing for the maximum span extension possible.
The Mach = 0.85 case, shown in Figure 6.22, is physically a more challenging problem for
the optimizer to solve. As in the previous bending constrained case, the optimizer is attempting
to reduce the wave drag with the airfoil shape. However, in this case, the optimizer is unable
to produce the heavily aft loaded airfoils that are optimal for this case because they lead to an
unstable design. As a result, the optimizer is forced to compromise between the reducing the
drag and maintaining Cmα < 0. These compromises show up in different ways at the different
sections of the wing. At the wing root, the Cp profile is spread over the entire chord of the
section with significant lift generated at both the leading and trailing edges with relatively
little lift generated mid chord. The mid-wing section exhibits a rooftop Cp distribution over
Chapter 6. Design Optimization Results 99
Figure 6.21: Cmα constrained optimization: planform and shape variables: M=0.7
the front half of the foil, helping to reduce the wave drag of the wing. However, the addition of
the Cmα constraint prevents the optimizer from extending this trend over the entire foil. As a
result there is a fairly constant amount of lift generated over the forward two thirds of the foil
with negative lift generated at the trailing edge. The optimizer has also added more sweep to
the wing than in the comparable bending constrained case. As discussed earlier, this reduces
the effective Mach number the wing sees and allows tip washout to contribute to wing trim.
In addition to this general discussion, there are two specific secondary characteristics that are
worth highlighting. First, the optimizer is adding a significant loading to the bottom surface
of the leading edge of the wing at the root. This forward loading helps reduce the moment of
the root section. Also, the optimizer has developed an interesting inflection in the camber of
the mid-wing airfoil. At the trailing edge, the airfoil starts to develop the high camber shape
typical of transonic airfoils, but partway to the trailing edge the foil develops reflex to help
reduce the moment of the wing.
As discussed in the previous chapter, the Cmα constrained formulation does not accomplish
exactly what is desired. This formulation of the constraint is only able to guarantee a neutrally
stable result. As shown in the Mach = 0.5 planform-only optimization case, when the constraint
Chapter 6. Design Optimization Results 100
Figure 6.22: Cmα constrained optimization: planform and shape variables: M=0.85
is active, it produces a Cmα of zero, giving a neutrally stable aircraft. To generate designs that
have positive static-stability, a formulation based on static margin is used. That formulation is
Chapter 6. Design Optimization Results 101
as follows:
minimize CD
w.r.t. x = α, θi , A1 , b, λ, CG% , ∆FFD
ctroot , xtCG , MACt , CLt α , Cm
t
α
(6.5)
subject to CLref − CL ≤ 0
Cm = 0
Knref − Kn ≤ 0
Cb − Cbref = 0
A2 − Aref = 0
xCG − xtCG = 0
MAC − MACt = 0
croot − ctroot = 0
CLα − CLt α = 0
t
Cmα − Cm α
=0
Note that in the above formulation, Kn is the static margin and that additional constraints
have been added for CLt α and Cm
t
α
corresponding to the additional IDF target variables that
have been added for the computation of the static margin constraint. A graphical depiction of
solution process for this formulation is shown in Figure 6.23.
The planform-only, static margin constrained optimization at Mach = 0.5, shown in Fig-
ure 6.24, shows an increase in sweep, just as the Cmα case did. In this case, however, the
increase in sweep is much more significant. Just as before, the optimizer is adding sweep and
wash-out to trim the aircraft for a more forward CG position. In this case, the magnitude of the
changes is larger to compensate for the correspondingly large separation between the neutral
point and CG for a 5% static margin. As with the Cmα case, adding shape variables to the
problem allows the optimizer to satisfy the stability constraint without sweeping the wing. As
a result, the static margin constrained optimal solution is similar to the optimal solution for
the Cmα constrained case.
The solution for the planform-only optimization at Mach = 0.7 continues the trend of the
Cmα constrained case. The optimizer adds sweep and twist to create separation between the
neutral point of the wing and the CG location of the aircraft while maintaining a trimmed
state. As can be seen in Figure 6.26 there is now visible separation between the neutral point
and the CG location.
Again, the addition of shape variables allows the optimizer to un-sweep the wing without
Chapter 6. Design Optimization Results 102
x : α, θi , A1 , b,
λ, CG% , ∆FFD ,
MACt , xtCG ,
ctroot , CLt α , Cm
t
α
5:
f : CD
1:
x∗CG , M AC ∗ ,c∗root 5 : xCG ,MAC,croot 2: Surface 4: MAC
Geometry
CL∗ , CD∗
, Cm ∗
, 5 : CL , CD , Cmz , 2: 3 : Cfx , Cfy , Cfz ,
z
CL∗ α , Cm∗
, A∗2 CLα , Cmα , A2 Aerody- Cmx , Cmy , Cmz
α
namics
3:
Cb∗ 5 : Cb Bending
Analysis
4:
Kn∗ Stability
5 : Kn
and
Control
Figure 6.25: Kn constrained optimization: planform and shape variables: M=0.5, e = 0.968
sacrificing performance. In this case the optimal solution has a very small amount of sweep.
This is possibly due to the increased difficulty of satisfying the trim constraint at the higher
Mach number.
The Mach = 0.85 case shows similar characteristics to the previously discussed Cmα con-
strained case. Once again the root section has a distributed Cp profile with lift split between
the fore and aft portions of the airfoil. The mid-section foils are show a rooftop profile on the
forward portion of the foil and exhibit some reflex at the trailing edge, while the tip foil also
exhibits some reflex. Note that once again the optimizer has added a significant amount of
sweep to the wing and reduced the span accordingly.
The dynamic stability formulation used in this work is based on the CAP parameter described
in Section 5.4.3. This formulation builds on the previously described static margin formulation
Chapter 6. Design Optimization Results 105
minimize CD
w.r.t. x = α, θi , A1 , b, λ, CG% , ∆FFD , ctroot , xtCG , MACt ,
t
CD , CLt α , Cm
t
α
t
, Cm α̇
t
, Cm q
(6.6)
subject to CLref − CL ≤ 0
Cm = 0
Knref − Kn ≤ 0
0.085 ≤ CAP ≤ 3.6
0.3 ≤ ζsp ≤ 2.0
A2 − Aref = 0
xCG − xtCG = 0
MAC − MACt = 0
croot − ctroot = 0
t
CD − CD =0
CLα − CLt α = 0
t
Cmα − Cm α
=0
t
Cmα̇ − Cm α̇
=0
t
Cmq − Cm q
=0
Chapter 6. Design Optimization Results 106
x : α, θi , A1 , b,
λ, CG% , ∆FFD ,
MACt , xtCG ,
ctroot , CD
t
, CLt α ,
t t t
Cm α
, Cm α̇
, Cm q
5:
f : CD
cc : xCG − xtCG
c : CLref − CL
MAC − MACt
Cmz
6 : f, c, c c croot − ctroot
Knref − Kn t
CD − CD
Cb − Cbref (ctroot )
CLα − CLt α
A2 − Aref t
Cmα − Cm
CAP t
α
Cmα̇ − Cm
ζsp t
α̇
Cmq − Cm q
1:
x∗CG , M AC ∗ ,c∗root 5 : xCG ,MAC,croot 2: Surface 4: MAC
Geometry
5 : CL , CD , Cmz , 2:
CL∗ , CD∗
, Cm ∗
, 3 : Cfx , Cfy , Cfz ,
z
CLα , Cmα , Cmα̇ , Aerody-
CL∗ α , Cm∗
, A∗2 Cmx , Cmy , Cmz
α
Cmq , A2 namics
3:
Cb∗ 5 : Cb Bending
Analysis
4:
Kn∗ , CAP∗ , ζsp
∗
5 : Kn , CAP, ζsp Stability
and
Control
Relative to the static margin optimization, two primary constraints have been added, a CAP
constraint and a ζsp constraint. These constraints define the appropriate limits for the dynamic
stability parameters. The range of allowable CAP values constrains the allowable short-period
frequency, and the damping ratio, ζsp , is constrained directly. Also, IDF constraints have been
t , C t , and C t to reflect the addition of these variables for the computation of the
added for CD mα̇ mq
dynamic stability parameters. A graphical depiction of the solution process for this formulation
is shown in Figure 6.29.
The CAP constrained optimizations add the dynamic stability constraints to the problem.
This requires the consideration of extra stability derivatives as well as the mass moment of
inertia of the aircraft during the optimization. The optimal solution for the CAP constrained
planform-only optimization at Mach = 0.5 — shown in Figure 6.30 — is essentially the same
as the static margin result from Figure 6.24. This results from the fact that the static margin
constraint is still active and that moment of inertia multiplier is able to raise the moment
of inertia sufficiently to satisfy the CAP and damping constraints. If the moment of inertia
multiplier were limited to a smaller value, the optimal solution would likely be more highly
swept with a larger static margin.
An interesting result from the addition of the dynamic constraints is that the addition of
the shape variables no longer produces an unswept wing. As shown in Figure 6.31, the optimal
solution now has almost 20 degrees of sweep. This results largely from the need to maintain a
Chapter 6. Design Optimization Results 107
Figure 6.30: CAP constrained optimization: planform variables only: M=0.5, e = 0.942
high enough moment of inertia to satisfy the damping requirements. Another interesting side
effect of this result is that the section profiles no longer exhibit any reflex at the trailing edge.
This comes from the fact that the optimizer is now able to use sweep and tip washout to trim
the aircraft without any additional penalty, so the Cp distribution is not required to add reflex
to do so.
The planform-only Mach = 0.7 CAP constrained case produces a result similar to static
margin constrained case. As with the Mach = 0.5 case, the moment of inertia multiplier is
able to increase the moment of inertia to the point where the CAP constraint and damping
constraints are satisfied.
As with the previous CAP constraint cases, at Mach = 0.7 the shape variables are not suffi-
cient to allow the optimizer to reduce the sweep of the wing to zero. Again the optimal results
from the bending constrained case and the static-stability constrained cases have insufficient
damping to satisfy the dynamic constraints. In this case, however, the added sweep causes the
static margin constraint to be inactive, though the mid-span section still develops a significant
amount of reflex.
At Mach = 0.85 the CAP constrained result is essentially the same as the static margin
constrained result. Because the shape variables were not able to completely eliminate the sweep
at this higher Mach number, the static margin case has enough sweep to ensure dynamic stability
Chapter 6. Design Optimization Results 108
Figure 6.31: CAP constrained optimization: planform and shape variables: M=0.5, e = 0.970
Figure 6.33: CAP constrained optimization: planform and shape variables: M=0.7
given a sufficient moment of inertia. Thus, the optimal result exhibits the same combinations
of Cp distributions and planform variables as the static margin constrained case.
To compare the results in a more quantitative fashion, the results of the optimizations are
presented in tabular form below. For the simpler cases where the various stability parameters
were not included in the optimizations, the remaining parameters have been calculated for
comparison purposes. A short summary of the most significant parameters is shown in Tables 6.4
through 6.8 to support the following discussion. A full compilation of the results for each
test case is shown in Appendix B. Note that the drag values presented are from steady-state
simulations of the optimal shapes for each case. The remainder of the values are computed
using the time-spectral methods used in the optimizations.
The key results for the Mach = 0.5 cases are shown in Tables 6.4 and 6.5. The first parameter
to consider is the drag for each of these cases. The drag for the planform-only cases — shown
in Table 6.4 — show precisely what one would expect. The optimal elliptical solution from the
Chapter 6. Design Optimization Results 110
Figure 6.34: CAP constrained optimization: planform and shape variables: M=0.85
twist-only optimization has a lower drag than the baseline case. The predicted value of drag for
this case is 0.00489, within 2.5% of the theoretical value of 0.00477, leading to a span efficiency
of 0.977. The remaining error in this case can be attributed to numerical errors — artificial
dissipation for example — as well as the three-dimensional flow at the wing tip.
The added flexibility of the span variable, introduced in the bending moment constrained
case, allows the optimizer to further reduce the drag in the baseline aerostructural case. Con-
sidering the problem from this perspective allows an additional 1.4% reduction in drag. Once
the stability constraints are included in the optimization, the drag increases. This increase in
drag comes from the changes necessary to make the aircraft stable. In the case of the Cmα
constraint, the changes are relatively minor, so the increase in drag is correspondingly small.
However, for the static margin and CAP constrained cases, the necessary increase in sweep is
much more significant, causing a much larger increase in the drag.
Also of note are the span efficiency values for the various optimal solutions. In each case,
the span efficiency is lower than for the elliptic solution. This is unsurprising, as the lift
distributions in these case deviate from elliptical. While the span efficiencies are lower for the
bending moment and Cmα constrained cases, the increased aspect ratio of these designs makes
up for the loss of span efficiency, allowing for a slightly lower overall drag coefficient. In the case
of the static margin and CAP constrained cases, both the aspect ratio and the span efficiency
Chapter 6. Design Optimization Results 111
are lower than in the elliptical case, leading to a higher drag coefficient.
Looking at the values of the various stability constraints, the results show that the static-
stability constraints are active in each of the stability constrained cases. This leads to a neutrally
stable result in the Cmα constrained case and results with the desired static margin in the last
two cases. Also note that the three cases without the stability constraints are statically unstable.
Finally, the CAP constrained case is essentially identical to the static margin constrained case.
This is primarily because the inertia modifier is able to increase the moment of inertia sufficiently
to satisfy the damping constraint without modifying the geometry.
Table 6.4: NACA 0012 wing: planform only optimization results: 1107k cells, M = 0.5
The results for the shape optimization cases in Table 6.5 show a very different picture. In
these cases, both of the static-stability constrained cases have essentially the same drag as the
planform-only bending constrained case. This comes from the fact that both the aspect ratio
and span efficiency are similar to the planform-only bending constrained case. Further, the
results show that neither of the static stability constraints are active for this case. However,
the Cp distributions from Figures 6.18, 6.20, 6.24 and 6.26 clearly show that the solution
is different for the planform-only cases and the shape optimization cases. This leads to the
conclusion that, for subsonic cases, airfoil shape can be used to trim the aircraft for a wide
variety of CG locations with little penalty in terms of drag. This supports the idea that a
model with more stringent limits on CG could be used in the context of this case and still allow
the optimizer to find feasible solutions, a fact that is not guaranteed for the planform-only
cases.
The same is not true for the dynamic stability constrained case. The results for this case
show that both the damping constraint and the static margin constraint are active. The addition
Chapter 6. Design Optimization Results 112
of the dynamic constraint pushes the optimizer to increase the moment of inertia of the aircraft,
which causes an increase in sweep of the wing. Because of the root bending moment constraint,
a reduction of the span ensues, causing a 7% increase in drag even though the span efficiency is
similar to the two static-stability constrained cases. Note that the moment of inertia multiplier
is also at its upper bound. Therefore, an increase in this limit may reduce the required sweep
and thus the amount of drag increase. This simply emphasizes the importance of developing
an improved moment of inertia prediction technique for any future work in this area.
Table 6.5: NACA 0012 wing: shape optimization results: 1107k cells, M = 0.5
The planform-only cases at Mach = 0.7 — shown in Table 6.6 — are not as interesting as the
Mach = 0.5 cases. The optimal bending constrained result has a static margin of roughly 0.7%,
satisfying the Cmα constraint. As a result, the optimal results for the bending constrained case
and the Cmα constrained case are essentially the same. The use of the static margin constraint
requires an increase in sweep to provide the necessary static margin. However, this additional
sweep, along with a relatively large moment of inertia are sufficient to satisfy the CAP and
damping constraints. Thus, the CAP constrained case has essentially the same optimal solution
as the static margin constrained case.
Examining the shape optimization cases at Mach = 0.7 — shown in Table 6.7 —, the results
once again show what one would expect. The bending constrained case produces a drag result
that is within one count of the optimal solutions at Mach = 0.5. This highlights the ability
of the shape optimizations to modify the airfoil shapes to eliminate the wave drag caused
Chapter 6. Design Optimization Results 113
Table 6.6: NACA 0012 wing: planform only optimization results: 1107k cells, M = 0.7
by transonic flow. While the bending constrained case is statically unstable, the addition of
the static stability constraints does not cause any significant change in the drag. Thus, the
optimizer is able to find airfoil profiles that are able to both eliminate the transonic wave drag
and satisfy the moment criteria necessary for static stability at the same time. As with the
Mach = 0.5 case, there is enough flexibility in the airfoil shape to allow a wide variety of CG
locations.
As with the Mach = 0.5 optimization cases, the CAP constraint forces the optimizer to
add sweep to the wing to increase the damping ratio of the aircraft. As a result, there is a
corresponding increase in the amount of drag produced by the wing. Note also that the inertia
modifier is double what it was for the Mach = 0.5 case. Further, the fact that it is at its upper
limit once again highlights the need for more investigation into methods for estimating this
parameter.
The Mach = 0.85 results, shown in Table 6.8, exhibit some interesting characteristics. Once
again, the bending constrained optimal solution is massively unstable, with a static margin
of -23.5%. The addition of the static stability constraints remedies this deficiency, with both
constraints producing optimal solutions at the limiting value of the constraints. However, in
this case the drag does not increase as the additional stability constraints are added. In fact,
the most constrained case produces the solution with the lowest drag. This might be explained
by the larger spans obtained in the Kn and CAP optimal solutions. Finally, as with some of the
previous cases, the moment of inertia multiplier is sufficient to satisfy the CAP and damping
Chapter 6. Design Optimization Results 114
Table 6.7: NACA 0012 wing: shape optimization results: 1107k cells, M = 0.7
constraints based on the static margin constrained optimal solution, therefore the static margin
constrained optimum and the CAP constrained optimum are essentially the same.
Table 6.8: NACA 0012 wing: shape optimization results: 1107k cells, M = 0.85
6.5 Summary
This chapter presented an optimization study exploring the effects of static and dynamic sta-
bility constraints on the optimal shape of flying wings. The results showed that the stability
constraints can have a significant impact on the optimal shape of the wing, usually causing a cor-
responding increase in drag. However, for cases in the subsonic regime and the lower end of the
Chapter 6. Design Optimization Results 115
transonic regime, the study showed that airfoil shape can be used to satisfy the static-stability
constraints without significant degradation in performance. As a result, in these flow regimes,
the optimizer tends to prefer using airfoil shape rather than wing sweep and twist to satisfy the
stability requirements. At higher transonic Mach numbers, the study showed that degradation
in performance seems to be unavoidable and that the addition of sweep is necessary, regardless
of the airfoil shape, to achieve satisfactory results. The same can not be said for the dynamic
stability constrained cases. In these cases, the study showed that for the subsonic and low-end
transonic cases, the optimizer was forced to add sweep to the wing, regardless of the airfoil
shape, in order to raise the damping ratio of the aircraft to acceptable levels. At the higher
end of the transonic regime, the static stability constrained results required sufficient sweep to
be feasible for both the static and dynamic stability constraints. Thus, the results of this study
underline the importance of considering both static and dynamic stability considerations in the
design of flying-wing aircraft.
Chapter 7
Conclusions
The work described in this thesis demonstrated a method for conducting stability-constrained
aerodynamic shape optimization of aircraft configurations using three-dimensional computa-
tional fluid dynamics techniques. To facilitate this method, two approaches for computing
stability derivatives using CFD were developed. The first approach, based on the ADjoint sen-
sitivity analysis method, was shown to be efficient for accurately computing full sets of static
and dynamic stability derivatives. While this technique is very effective for analyzing the stabil-
ity derivatives of a given configuration, the use of the ADjoint method in the computation of the
stability derivatives makes it prohibitively expensive to compute the gradients of these stability
derivatives for use in a gradient based optimization context. The second approach was developed
using simple linear regression techniques together with a time-spectral CFD solution. Because
this technique does not require an adjoint method to compute the stability derivative values, it
is possible to apply an adjoint method to the combined stability derivative/time-spectral flow
solution process, facilitating efficient gradient calculations for optimization. Further, because
the time-spectral CFD solution includes unsteady flow information, it is possible to estimate
the transient, or “dot”, stability derivatives, which can play an important role in assessing dy-
namic stability. Based on these last two qualities, the time-spectral stability derivative method
was selected for further development and was integrated into an aerodynamic optimization
framework.
To facilitate efficient optimization with this framework, a time-spectral adjoint method for
sensitivity analysis was developed and verified using the complex-step method. The computed
derivatives match reference values to between 7 and 12 digits, giving more than enough accuracy
for the optimizations performed in this work. The efficiency of the sensitivity analysis method
was demonstrated on test cases up to roughly 40 million flow states. The largest adjoint test
case analyzed for the scaling study was solved in under 30 minutes on 224 processors.
Using these techniques, a study was conducted to explore the effects of a series of static and
116
Chapter 7. Conclusions 117
dynamic stability constraints on the optimal design of a flying wing configuration. The study
showed that the various constraints can have a significant effect on the optimal solution. Most
optimizations conducted without the stability constraints resulted in statically unstable designs
that would incur significant drag penalties in order to be made statically stable. However, in the
cases with static stability constraints, the addition of shape variables to the optimization was
able to counteract the adverse effects of the stability constraints by generating reflex airfoils,
virtually eliminating the drag penalty associated with the static-stability constraints in most
cases. The same can not be said for the dynamic stability constrained cases. Increases in sweep
— and the corresponding increases in drag — were required in most of these cases to satisfy
the damping ratio constraint associated with dynamic stability.
Given the relatively simplistic nature of the center of gravity and moment of inertia calcu-
lations used in this work, the specific results shown here must be taken with a certain degree of
caution. However, the fact that optimal wing designs tend to violate the stability constraints if
they are not enforced clearly shows that static and dynamic stability constraints have an impor-
tant impact on aerodynamic performance. Thus, while the work presented here is by no means
the final answer to the problem, the results shown demonstrate the importance of including
stability constraints in aerodynamic optimization, specifically when looking at stability-critical
aircraft such as flying wings.
Chapter 8
Future Work
The results of the work presented in this thesis suggest several areas for future exploration. An
overview of some of these topics is included below.
Viscous Stability Derivative Computations: The work in this thesis has been conducted
entirely with an inviscid Euler flow solver. The predictive capabilities of the stability
derivative methods presented in this work would be significantly improved if combined
with an efficient RANS flow solver. The use of a RANS solver would allow more complete
validation against experimental data, which would allow for more concrete assessment
of the accuracy and validity of the methods. By using a RANS solver, the range of
flow conditions which could be accurately modeled would be significantly increased. This
would help in evaluating the extreme flow conditions which are often critical in the design
process. In addition to considering the analysis with a RANS flow solver, an efficient
adjoint will be needed in order to optimize flows in the RANS regime.
Mass and Moment of Inertia Models: As discussed in the optimization study, there are
several cases where the estimated values of CG location and moment of inertia affect the
optimal result. This is particularly true when the dynamic stability constraints are con-
sidered. Therefore, it is important that further exploration in this area include improved
estimates for these quantities.
Dynamic Stability Constraints: While this thesis has shown that it is possible to optimize
an aircraft using high-fidelity CFD and dynamic stability constraints, this work has really
only scratched the surface of possible formulations for these constraints. An interesting
avenue of exploration would be to investigate the effects of various stability formulations
on both the optimization process and its results. This would likely also necessitate the
consideration of lateral stability derivatives and the impact that they have on design.
118
Chapter 8. Future Work 119
Extended applications: The study presented here is applied to a relatively simple case.
While the broad strokes of the conclusions developed here are applicable to a variety
of cases, it would be interesting to explore the effects of stability-constraints on a more
realistic design, where more is known about the aircraft.
References
[1] Daniël Sarel Agenbag. Longitudinal handing characteristics of a tailless gull-wing aircraft.
Master’s thesis, University of Pretoria, Pretoria, South Africa, July 2008.
[2] Airbus. Global market forecast: 2000 - 2019. Technical report, Airbus Industrie, France,
2000.
[3] W. Kyle Anderson and Daryl L. Bonhaus. Airfoil design on unstructured grids for tur-
bulent flows. AIAA Journal, 37(2):185–191, 1999.
[4] F. Angrand. Optimum design for potential flows. International Journal for Numerical
Methods in Fluids, 3(3):265–282, 1983.
[5] Deric A. Babcock and Dr. Andrew S. Arena Jr. Estimating aircraft stability derivatives
through finite element analysis. AIAA Paper 2004-5174, August 2004.
[6] Satish Balay, Kris Buschelman, Victor Eijkhout, William D. Gropp, Dinesh Kaushik,
Matthew G. Knepley, Lois Curfman McInnes, Barry F. Smith, and Hong Zhang. PETSc
users manual. Technical Report ANL-95/11 - Revision 2.1.5, Argonne National Labora-
tory, 2004.
[7] William Bihrle. A handling qualities theory for precise flight path control. AFFDL
Technical Report AFFDL-TR-65-198, June 1966.
[9] Christian H. Bischof, H. Martin Bücker, Bruno Lang, A. Rasch, and Andre Vehreschild.
Combining source transformation and operator overloading techniques to compute deriva-
tives for MATLAB programs. In Proceedings of the Second IEEE International Workshop
on Source Code Analysis and Manipulation (SCAM 2002), pages 65–72, Los Alamitos,
CA, USA, 2002. IEEE Computer Society.
120
References 121
[10] Christian H. Bischof, Lucas Roh, and Andrew Mauer. ADIC — An extensible automatic
differentiation tool for ANSI-C. Software–Practice and Experience, 27(12):1427–1456,
1997.
[12] Boeing. Current market outlook: 2009 - 2028. Technical report, Boeing Commercial
Airplanes, Seattle, WA, 2009.
[14] Robert D. Braun and Ilan M. Kroo. Development and application of the collaborative
optimization architecture in a multidisciplinary design environment. In N. Alexandrov
and M. Y. Hussaini, editors, Multidisciplinary Design Optimization: State-of-the-Art,
pages 98–116. SIAM, 1997.
[15] Patrick Brezillon, Jean-François Staub, Anne-Marie Perault-Staub, and Gérard Milhaud.
Numerical estimation of the first order derivative: approximate evaluation of an optimal
step. Computers and Mathematics with Applications, 7(4):333–347, 1981.
[16] Howard P. Buckley, Beckett Y. Zhou, and David. W. Zingg. Airfoil optimization using
practical aerodynamic design requirements. Journal of Aircraft, 47(5):1707–1719, 2010.
[17] Alan Carle and Mike Fagan. ADIFOR 3.0 overview. Technical Report CAAM-TR-00-02,
Rice University, 2000.
[18] S. Chai, P. Crisafulli, and W.H.Mason. Aircraft center of gravity estimation in concep-
tual/preliminary design. In Proceedings of the 1st AIAA Aircraft Engineering, Technology,
and Operations Congress, Los Angeles, CA, 1995. AIAA 95-3882.
[19] Eric F. Charlton. Numerical stability and control analysis towards falling-leaf prediction
capabilities of SPLITFLOW for two generic high-performance aircraft models. NASA
Contractor Report CR-1998-208730, September 1998.
[20] Seongim Choi, Mark Potsdam, Kihwan Lee, Gianluca Iaccarino, and Juan J. Alonso.
Helicopter rotor design using a time-spectral and adjoint-based method. AIAA Paper
2008-5810, September 2008. Proceedings of the 12th AIAA/ISSMO Multidisciplinary
Analysis and Optimization Conference, British Columbia, CA.
References 122
[22] Russell M. Cummings, Adam Jirásek, Kristian Petterson, and Stefan Schmidt. SACCON
static and dynamic motion flow physics simulation using cobalt. In Proceedings of the
28th AIAA Applied Aerodynamics Conference, Chicago, IL, 2010. AIAA 2010-4691.
[24] Kivanc Ekici and Kenneth C. Hall. Nonlinear analysis of unsteady flows in multistage
turbomachines using harmonic balance. AIAA Journal, 45(5):1047–1057, 2007.
[25] ESDU. Geometrical properties of cranked and straight tapered wing planforms. Technical
Report Item No. 76003, ESDU International plc, London, UK, January 1976.
[26] ESDU. ESDU dynamics series. Technical Report Various Items, ESDU International plc,
London, UK, 2010.
[27] Bernard Etkin. Dynamics of Atmospheric Flight. Dover Publications, Mineola, New York,
2000.
[28] John W. Gallman, Stephen C. Smith, and Ilan M. Kroo. Optimization of joined-wing
aircraft. Journal of Aircraft, 30:897–905, 1993.
[29] Frank H. Gern, Amir H. Naghshineh-Pour, Erwin Sulaeman, Rakesh K. Kapania, and
Raphael T. Haftka. Structural wing sizing for multidisiplinary design optimization of a
strut-braced wing. Journal of Aircraft, 38:154–163, 2001.
[30] Davendra P. Ghate and Michael B. Giles. Effcient Hessian calculation using automatic
differentiation. AIAA Paper 2007-4059, June 2007. Proceedings of the 18th AIAA Applied
Aerodynamics Conference, Miami, FL.
[32] Philip E. Gill, Walter Murray, and Michael A. Saunders. Snopt: An sqp algorithm for
large-scale constrained optimization. SIAM Review, 47(1):99–131, 2005.
[33] Andrew G. Godfrey and Eugene M. Cliff. Direct calculation of aerodynamic force deriva-
tives: A sensitivity-equation approach. AIAA Paper 98-0393, AIAA, 1998.
References 123
[34] Arathi K. Gopinath and Antony Jameson. Time spectral method for periodic unsteady
computations over two- three dimensional bodies. AIAA Paper 2005-1220, January 2005.
[35] Arathi Kamath Gopinath. Efficient Fourier-Based Algorithms for Time-Periodic Un-
steady Problems. PhD thesis, Stanford University, Stanford, CA 94305, April 2007.
[38] Andreas Griewank, David Juedes, and Jean Utke. Algorithm 755: ADOL-C: A package
for the automatic differentiation of algorithms written in C/C++. ACM Transactions on
Mathematical Software, 22(2):131–167, 1996.
[39] Raphael T. Haftka. Simultaneous analysis and design. AIAA Journal, 23(7):1099–1103,
July 1985.
[41] Kenneth C. Hall, Jeffery P. Thomas, and W.S.Clark. Computation of unsteady nonlinear
flows in cascades using a harmonic balance technique. AIAA Journal, 40(5):879–886,
2002.
[42] L. Hascoët and V Pascual. Tapenade 2.1 user’s guide. Technical report 300, INRIA, 2004.
[43] Jason E. Hicken and David. W. Zingg. Parallel Newton–Krylov solver for the Euler equa-
tions discretized using simultaneous-approximation terms. AIAA Journal, 46(11):2773–
2786, 2008.
[44] Jason E. Hicken and David. W. Zingg. Induced-drag minimization of nonplanar geometries
based on the Euler equations. AIAA Journal, 48(11):2564–2575, 2010.
[45] ICAO. Aviation emissions in context. Technical report, International Civil Aviation
Organization (ICAO), Montreal, CA, 2009.
[46] A. Jameson. Aerodynamic design via control theory. Journal of Scientific Computing,
3(3):233–260, September 1988.
[47] A. Jameson, L. Martinelli, and N. A. Pierce. Optimum aerodynamic design using the
Navier-Stokes equations. Theoretical and Computational Fluid Dynamics, 10:213–237,
1998.
References 124
[48] Antony Jameson. Optimum aerodynamic design using control theory. Computational
Fluid Dynamics Review, pages 495–528, 1995.
[49] Gaetan K.W. Kenway, Graeme J. Kennedy, and Joaquim R. R. A. Martins. A CAD-
free approach to high-fidelity aerostructural optimization. In Proceedings of the 13th
AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, Fort Worth, TX,
2010. AIAA 2010-9231.
[50] Hyung Min Kim, Nestor F. Michelena, Panos Y. Papalambros, and Tao Jian. Target cas-
cading in optimal system design. Journal of Mechanical Design, 125(3):474–480, Septem-
ber 2003.
[51] Andrew B. Lambe and Joaquim R. R. A. Martins. Extensions to the design structure ma-
trix for the description of multidisciplinary design, analysis, and optimization processes.
Structural and Multidisciplinary Optimization, 2011.
[52] Joosung Joseph Lee. Historical and future trends in aircraft perfomance, cost, and emis-
sions. Master’s thesis, Massachusetts Institute of Technology, Cambridge, MA, September
2000.
[53] R. H. Liebeck. Design of the blended wing body subsonic transport. Journal of Aircraft,
41:10–25, 2004.
[54] A.C. Limache and E.M. Cliff. Aerodynamic sensitivity theory for rotary stability deriva-
tives. Journal of Aircraft, 37:676 – 683, 2000.
[55] Alejandro Cesar Limache. Aerodynamic Modeling Using Computational Fluid Dynamics
and Sensitivity Equations. PhD thesis, Virginia Polytechnic Institute and State University,
Blacksburg, VA 24061, April 2000.
[56] Thomas D. Loeser, Dan D. Vicroy, and Andreas Schütte. SACCON static wind tunnel
tests at DNW-NWB and 14’x22’ NASA LaRC. In Proceedings of the 28th AIAA Applied
Aerodynamics Conference, Chicago, IL, 2010. AIAA 2010-4393.
[59] Charles A. Mader and Joaquim R. R. A. Martins. Computation of aircraft stability deriva-
tives using an automatic differentiation adjoint approach. AIAA Journal, 49(12):2737–
2750, December 2011.
[60] Charles A. Mader, Joaquim R. R. A. Martins, Juan J. Alonso, and Edwin van der Weide.
ADjoint: An approach for the rapid development of discrete adjoint solvers. AIAA Jour-
nal, 46(4):863–873, 2008.
[61] Karthik Mani and Dimitri J. Mavriplis. Unsteady discrete adjoint formulation for two-
dimensional flow problems with deforming meshes. AIAA Journal, 46(6):1351–1364, 2008.
[62] Andre C. Marta and Juan J. Alonso. High-speed MHD flow control using adjoint-based
sensitivities. In Proceedings of the 14th AIAA/AHI Space Planes and Hypersonic Systems
and Technologies Conference, Canberra, Australia, 2006. AIAA 2006-8009.
[63] Andre C. Marta, Charles A. Mader, Joaquim R. R. A. Martins, Edwin van der Weide,
and Juan J. Alonso. A methodology for the development of discrete adjoint solvers using
automatic differentiation tools. International Journal of Computational Fluid Dynamics,
21(9):307–327, October 2007.
[64] Joaquim R. R. A. Martins, Juan J. Alonso, and James J. Reuther. High-fidelity aerostruc-
tural design optimization of a supersonic business jet. Journal of Aircraft, 41(3):523–530,
2004.
[65] Joaquim R. R. A. Martins, Peter Sturdza, and Juan J. Alonso. The complex-step deriva-
tive approximation. ACM Transactions on Mathematical Software, 29(3):245–262, 2003.
[66] Joaquim R.R.A. Martins, Juan J. Alonso, and James J. Reuther. A coupled-adjoint
sensitivity analysis method for high-fidelity aero-structural design. Optimization and
Engineering, 6(1):33–62, 2005.
[67] Dimitri J. Mavriplis. Solution of the unsteady discrete adjoint for three-dimensional
problems on dynamically deforming unstructured meshes. In Proceedings of the 46th
AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV, January 2008. AIAA 2008-
727.
[68] Matthew McMullen, Antony Jameson, and Juan Alonso. Demonstration of nonlinear
frequency domain methods. AIAA Journal, 44(7):1428–1435, 2006.
[69] Matthew McMullen, Antony Jameson, and Juan J. Alonso. Application of a nonlinear
frequency domain solver to the Euler and Navier-Stokes equations. In Proceedings of the
References 126
40th AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV, January 2002. AIAA
2002-0120.
[70] M.S. McMullen and A. Jameson. The computational efficiency of non-linear frequency
domain methods. Journal of Computational Physics, 212(2):637–661, 2006.
[71] Duane McRuer, Irving Ashkenas, and Dunstan Graham. Aircraft Dynamics and Auto-
matic Control. Princeton University Press, Princeton, New Jersey, 1973.
[72] Bruno Mialon, Alex Khrabov, Andrea Da Ronch, Luca Cavagna, Mengmeng Zhang, and
Sergio Ricci. Benchmarking the prediction of dynamic derivatives: Wind tunnel tests,
validation, acceleration methods. In Proceedings of the 28th AIAA Applied Aerodynamics
Conference, Chicago, IL, 2010. AIAA 2010-8244.
[73] United States Military. Flying qualities of piloted airplanes. Military Specification MIL-
F-8785c, November 1980.
[75] Siva K. Nadarajah and Antony Jameson. Optimum shape design for unsteady flows with
time-accurate continuous and discrete adjoint methods. AIAA Journal, 45(7):1478–1491,
2007.
[76] Siva K. Nadarajah and Antony Jameson. Optimum shape design for unsteady three-
dimensional viscous flows using a nonlinear frequency-domain method. Journal of Air-
craft, 44(5):1513–1527, 2007.
[77] Siva K. Nadarajah, Antony Jameson, and Juan J. Alonso. Sonic boom reduction using an
adjoint method for wing-body configurations in supersonic flow. AIAA Paper 2002-5547,
2002.
[78] Marian Nemec and David W. Zingg. Newton–Krylov algorithm for aerodynamic design
using the Navier–Stokes equations. AIAA Journal, 40(6):1146–1154, June 2002.
[79] Marian Nemec, David. W. Zingg, and Thomas H. Pulliam. Multipoint and multi-objective
aerodynamic shape optimization. AIAA Journal, 42(6):1057–1065, 2004.
[81] Eric J. Nielsen and W. Kyle Anderson. Aerodynamic design optimization on unstructured
meshes using the Navier–Stokes equations. AIAA Journal, 37(11):1411–1419, 1999.
[82] Michael A. Park and Lawrence L. Green. Steady-state computation of constant rotational
rate dynamic stability derivatives. AIAA Paper 2000-4321, August 2000.
[83] Michael A. Park, Lawrence L. Green, Raymond C. Montgomery, and David L. Raney.
Determination of stability and control derivatives using computational fluid dynamics and
automatic differentiation. AIAA Paper 1999-3136, June 1999.
[85] Sergey Peigin and Boris Epstein. Computational fluid dynamics driven optimization of
blended wing body aircraft. AIAA Journal, 44(11):2736–2745, 2006.
[86] R. Perez and J. Martins. pyACDT: An object-oriented framework for aircraft design
modelling and multidisciplinary optimization. In Proceedings of the 12th AIAA/ISSMO
Multidisciplinary Analysis and Optimization Conference, Victoria, BC, September 10 -
12, 2008.
[87] Ruben E. Perez, Peter W. Jansen, and Joaquim R. R. A. Martins. pyOpt: A Python-
based object-oriented framework for nonlinear constrained optimization. Structures and
Multidisciplinary Optimization, 2011.
[88] Ruben E. Perez, Hugh H.T. Liu, and Kamran Behdinan. Multidisciplinary optimization
framework for control-configuration integration in aircraft conceptual design. Journal of
Aircraft, 43(6):1937–1948, 2006.
[91] James J. Reuther, Antony Jameson, Juan J. Alonso, Mark J. Rimlinger, and David
Saunders. Constrained multipoint aerodynamic shape optimization using an adjoint for-
mulation and parallel computers, part 1. Journal of Aircraft, 36(1):51–60, 1999.
References 128
[92] James J. Reuther, Antony Jameson, Juan J. Alonso, Mark J. Rimlinger, and David
Saunders. Constrained multipoint aerodynamic shape optimization using an adjoint for-
mulation and parallel computers, part 2. Journal of Aircraft, 36(1):61–74, 1999.
[93] Detlef Rohlf, Stefan Schmidt, and Jonathan Irving. SACCON stability and control anal-
ysis applying system identification techniques. In Proceedings of the 28th AIAA Applied
Aerodynamics Conference, Chicago, IL, 2010. AIAA 2010-4399.
[94] A. Da Ronch, M. Ghoreyshi, K.J. Badcock, S. Görtz, M. Widhalm, R.P. Dwight, and
M.S. Campobasso. Linear frequency domain and harmonic balance predictions of dy-
namic derivatives. In Proceedings of the 28th AIAA Applied Aerodynamics Conference,
Chicago, IL, 2010. AIAA 2010-4699.
[96] J.-F. Le Roy and S. Morgand. SACCON CFD static and dynamic derivatives using elsA.
In Proceedings of the 28th AIAA Applied Aerodynamics Conference, Chicago, IL, 2010.
AIAA 2010-4562.
[97] Markus P. Rumpfkeil and Dimitri J. Mavriplis. Efficient Hessian calculations using
automatic differentiation and the adjoint method with applications. AIAA Journal,
48(10):2406–2417, October 2010.
[98] Markus P. Rumpfkeil and David. W. Zingg. The optimal control of unsteady flows with
a discrete adjoint method. Optimization and Engineering, 11(1):5–22, 2010.
[99] Andreas Schütte, Dietrich Hummel, and Stephan M. Hitzel. Numerical and experimental
analyses of the vortical flow around the SACCON configuration. In Proceedings of the
28th AIAA Applied Aerodynamics Conference, Chicago, IL, 2010. AIAA 2010-4690.
[101] B. I. Soemarwoto and Th. E. Labrujére. Airfoil design and optimization methods: Recent
progress at NLR. International Journal for Numerical Methods in Fluids, 30(2):217–228,
1999.
[102] William Squire and George Trapp. Using complex variables to estimate derivatives of real
functions. SIAM Review, 40(1):110–112, 1998.
References 129
[103] Martin Stettner and Ralph Voss. Aeroelastics, flight mechanics and handling qualities of
the MOB BWB configuration. In Proceedings of the 9th AIAA/ISSMO Multidisciplinary
Analysis and Optimization Conference, Atlanta, Georgia, September 4 - 6, 2002.
[104] Jeffery P. Thomas, Kenneth C. Hall, and Earl H. Dowell. Discrete adjoint approach for
modeling unsteady aerodynamic design sensitivities. AIAA Journal, 43(9):1931–1936,
2005.
[105] Magnus Tormalm and Stefan Schmidt. Computational study of static and dynamic vor-
tical flow over the delta wing SACCON configuration using the FOI flow solver Edge.
In Proceedings of the 28th AIAA Applied Aerodynamics Conference, Chicago, IL, 2010.
AIAA 2010-4561.
[106] USAF. The USAF stability and control digital datcom. AFFDL Technical Report
AFFDL-TR-79-3032, February 2003.
[107] E. van der Weide, G. Kalitzin, J. Schluter, and J. J. Alonso. Unsteady turbomachin-
ery computations using massively parallel platforms. In Proceedings of the 44th AIAA
Aerospace Sciences Meeting and Exhibit, Reno, NV, 2006. AIAA 2006-0421.
[108] Edwin van der Weide, Arathi K. Gopinath, and Antony Jameson. Turbomachinery ap-
plications with the time spectral method. AIAA Paper 2005-4905, June 2005.
[109] Dan D. Vicroy. Blended-wing-body low-speed flight dynamics: Summary of ground tests
and sample results. AIAA Paper 2009-933, January 2009. Proceedings of the 47th AIAA
Aerospace Sciences Meeting Including The New Horizons Forum and Aerospace Exposi-
tion, Orlando, Florida.
[110] Sean Wakayama and Ilan Kroo. The challenge and promise of blended-wing-body opti-
mization. AIAA Paper 1998-4736, 1998.
[111] Li Wang, Dimitri J. Mavriplis, and W. Kyle Anderson. Adjoint sensitivity formulation for
discontinuous galerkin discretizations in unsteady inviscid flow problems. AIAA Journal,
48(12):2867–2883, 2010.
[112] Z.U.A. Warsi. Fluid Dynamics: Theoretical and Computational Approaches. Taylor &
Francis, 3rd edition, 2006.
[113] Julian Wolkovitch. The joined wing: An overiew. Journal of Aircraft, 23:161–178, 1986.
Appendix A
The derivation in the following appendix describes how to obtain the linear flight dynamic
model discussed in Chapter 2.
To start the derivation, consider the equations of motion from Etkin [27].
X − sin θ (u̇ + wq − rv)
Y + mg sin φ cos θ = m (v̇ + ru − pw) (A.1)
Z cos φ cos θ (ẇ + pv − qu)
L Ix −Ixy −Izx ṗ 0 −r q Ix −Ixy −Izx p
M = −Ixy Iy −Iyz q̇ + r
0 −p −Ixy
Iy −Iyz q
(A.2)
N −Izx −Iyz Iz ṙ −q p 0 −Izx −Iyz Iz r
φ̇ 1 sin φ tan θ cos φ tan θ p
θ̇ = 0 cos φ − sin φ q
(A.3)
ψ̇ 0 sin φ sec θ cos φ sec θ r
These equations assume that the aircraft is rigid, that the effect of spinning rotors is negligible,
that the mass of the aircraft is constant and that the aircraft is flying through a still air mass.
First, using small disturbance theory, Equations (A.1) through (A.3) can be rewritten as:
X0 + ∆X − sin θ0 − ∆θ cos θ0 ∆u̇ + w0 ∆q − v0 ∆r
Y0 + ∆Y + mg ∆φ cos θ 0
= m ∆v̇ + u0 ∆r − w0 ∆p (A.4)
Z0 + ∆Z cos θ0 − ∆θ0 ∆ẇ + v0 ∆p − u0 ∆q
130
Appendix A. Linear Flight Dynamics 131
L0 + ∆L Ix −Ixy −Izx ∆ṗ
M0 + ∆M = −Ixy Iy −Iyz (A.5)
∆q̇
N0 + ∆N −Izx −Iyz Iz ∆ṙ
∆φ̇ 1 0 tan θ0 ∆p
∆θ̇ = 0 1 0 ∆q . (A.6)
∆ψ̇ 0 0 sec θ0 ∆r
• There are no initial rotations or accelerations - i.e. p0 = q0 = r0 = u̇0 = v̇0 = ẇ0 = ṗ0 =
q̇0 = ṙ0 = 0
Looking in more detail at the condition that there are no initial rotations or accelerations, one
can determine the values of the initial forces (X0 ,Y0 ,Z0 ) and moments (L0 ,M0 ,N0 ). Substituting
zero accelerations and rotations into Equations (A.1) and (A.2) yields the following:
X0 − sin θ0 0
Y0 + mg 0 = 0 (A.7)
Z0 cos θ0 0
L0 0
M0 = 0 (A.8)
N0 0
Substituting these values back into Equations (A.4) and (A.5) yields:
∆X −∆θ cos θ0 ∆u̇ + w0 ∆q − v0 ∆r
∆Y + mg ∆φ cos θ0 = m ∆v̇ + u0 ∆r − w0 ∆p (A.9)
∆Z −∆θ0 ∆ẇ + v0 ∆p − u0 ∆q
∆L Ix −Ixy −Izx ∆ṗ
∆M = −Ixy Iy −Iyz ∆q̇
(A.10)
∆N −Izx −Iyz Iz ∆ṙ
Appendix A. Linear Flight Dynamics 132
Given the preceding equations, one will have the following state variables: u, v, w, p, q, r, φ, θ, ψ
and their time derivatives, u̇, v̇, ẇ, ṗ, q̇, ṙ, φ̇, θ̇, ψ̇. On this basis, the linear aerodynamic reactions
can be expressed as follows:
∆X Xi ∆i Xj ∆j XControl ∆Control
∆Y Yi ∆i Yj ∆j YControl ∆Control
∆Z Zi ∆i Zj ∆j ZControl ∆Control
= + + (A.11)
∆L Li ∆i Lj ∆j LControl ∆Control
∆M M ∆i M ∆j M ∆Control
i j Control
∆N Ni ∆i Nj ∆j NControl ∆Control
where i = u, v, w, p, q, r, φ, θ, ψ and j = di/dt. The subscribed quantities are derivatives of
each force or moment quantity with respect to the corresponding subscript. Note that since
the atmosphere is assumed to be constant this allows the location and altitude variables —
xE , yE , zE — to be ignored in the force equations. Now, by combining Equation (A.12) with
Equations (A.6) (A.9) and (A.10) we can get a matrix version of the equations as follows:
(Xu̇ − m) Xv̇ Xẇ Xṗ Xq̇ Xṙ Xφ̇ Xθ̇ Xψ̇ ∆u̇
Yu̇ (Yv̇ − m) Yẇ Yṗ Yq̇ Yṙ Yφ̇ Yθ̇ Yψ̇ ∆v̇
Zu̇ Zv̇ (Zẇ − m) Zṗ Zq̇ Zṙ Zφ̇ Zθ̇ Zψ̇ ∆ẇ
Lu̇ Lv̇ Lẇ (Lṗ − Ix ) (Ixy + Lq̇ ) (−Ixz + Lṙ ) Lφ̇ Lθ̇ Lψ̇ ∆ṗ
∆q̇ +
Mu̇ Mv̇ Mẇ (Ixy + Mṗ ) (−Iy + Mq̇ ) (Izy + Mṙ ) Mφ̇ Mθ̇ Mψ̇
∆ṙ
Nu̇ Nv̇ Nẇ (Ixz + Nṗ ) (Izy + Nq̇ ) (−Iz + Nṙ ) Nφ̇ Nθ̇ Nψ̇
∆φ̇
0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 1 0
∆θ̇
0 0 0 0 0 0 0 0 1 ∆ψ̇
Xu Xv Xw Xp (Xq − mw0 ) Xr + mv0 Xφ (Xθ − mg cos θ0 ) Xψ ∆u
Yu Yv Yw Yp + mw0 Yq Yr − muo Yφ + mg cos θ0 Yθ Yψ ∆v
Zu Zv Zw Zp − mv0 (muo + Zq ) Zr Zφ Zθ − mg sin θ0 Zψ ∆w
Lu Lv Lw Lp Lq Lr Lφ Lθ Lψ ∆p
∆q +
Mu Mv Mw Mp Mq Mr Mφ Mθ Mψ
Nu Nv Nw Np Nq Nr Nφ Nθ Nψ
∆r
0 0 0 −1 0 − tan θ0 0 0 0
∆φ
0 0 0 0 −1 0 0 0 0 ∆θ
0 0 0 0 0 sec θo 0 0 0 ∆ψ
XControl
YControl
ZControl
LControl
MControl ∆Control = 0 (A.12)
N
Control
0
0
0
Appendix A. Linear Flight Dynamics 133
Further, because we have assumed a constant atmosphere, the orientation of the aircraft
has no impact on the forces, therefore the derivatives of the forces with respect to the change
in the Euler angles —∆φ, ∆θ, ∆ψ — are zero, which simplifies the equations to:
(Xu̇ − m) Xv̇ Xẇ Xṗ Xq̇ Xṙ 0 0 0 ∆u̇
Yu̇ (Yv̇ − m) Yẇ Yṗ Yq̇ Yṙ 0 0 0 ∆v̇
Zu̇ Zv̇ (Zẇ − m) Zṗ Zq̇ Zṙ 0 0 0 ∆ẇ
Lu̇ Lv̇ Lẇ (Lṗ − Ix ) (Ixy + Lq̇ ) (−Ixz + Lṙ ) 0 0 0 ∆ṗ
∆q̇ +
Mu̇ Mv̇ Mẇ (Ixy + Mṗ ) (−Iy + Mq̇ ) (Izy + Mṙ ) 0 0 0
Nu̇ Nv̇ Nẇ (Ixz + Nṗ ) (Izy + Nq̇ ) (−Iz + Nṙ ) 0 0 0
∆ṙ
0 0 0 0 0 0 1 0 0
∆φ̇
0 0 0 0 0 0 0 1 0 ∆θ̇
0 0 0 0 0 0 0 0 1 ∆ψ̇
Xu Xv Xw Xp (Xq − mw0 ) Xr + mv0 0 −mg cos θ0 0 ∆u
Yu Yv Yw Yp + mw0 Yq Yr − muo mg cos θ0 0 0 ∆v
Zu Zv Zw Zp − mv0 (muo + Zq ) Zr 0 −mg sin θ0 0 ∆w
Lu Lv Lw Lp Lq Lr 0 0 0 ∆p
∆q +
Mu Mv Mw Mp Mq Mr 0 0 0
Nu Nv Nw Np Nq Nr 0 0 0
∆r
0 0 0 −1 0 − tan θ0 0 0 0
∆φ
0 0 0 0 −1 0 0 0 0 ∆θ
0 0 0 0 0 sec θo 0 0 0 ∆ψ
XControl
YControl
ZControl
LControl
MControl ∆Control = 0 (A.13)
N
Control
0
0
0
Solving for the time derivative terms yields a differential equation of the form:
h i h ih i h ih i
ẋ = A x + B u (A.14)
which is in a standard state-space matrix equation and can be solved using a normal differential
equation solvers. The inversion of the matrix analytically yields a very messy matrix, so for
actual computation it is probably better to invert the matrix on the fly.
Appendix B
Table B.1: NACA 0012 wing: planform only optimization results summary: 1107k cells, M =
0.5
134
Appendix B. Results Data Tables 135
Table B.2: NACA 0012 wing: shape optimization results summary: 1107k cells, M = 0.5
Appendix B. Results Data Tables 136
Table B.3: NACA 0012 wing: planform only optimization results summary: 1107k cells, M =
0.7
Appendix B. Results Data Tables 137
Table B.4: NACA 0012 wing: shape optimization results summary: 1107k cells, M = 0.7
Appendix B. Results Data Tables 138
Table B.5: NACA 0012 wing: shape optimization results summary: 1107k cells, M = 0.85
Appendix C
139
Appendix C. Relevant Optimization Parameters 140
CFD Parameters
Parameter Value units
Time-spectral ω 6.28 rad./s
Time-spectral amplitude 0.034 rad.
k2 dissipation 0
k4 dissipation 0.006
Time-spectral instances 3
Dissipation lumping parameter 6
Multigrid restriction relaxation 0.8
Dissipation scaling exponent 0.1
Startup CFL number 0.9
Multigrid cycle 4w
CFD Convergence tolerance 10−10
Runge-Kutta to Newton-Krylove switch tolerance 0.001
Newton-Krylov global preconditioner Additive Schwarz
Newton-Krylov local preconditioner ILU
ILU fill level 2
Preconditioner reordering RCM
ASM overlap 2
GMRES subspace size 100
Newton-Krylov Jacobian reset lag 10
ADjoint Parameters
ADjoint convergence tolerance 1e-8
ADjoint global preconditioner Additive Schwarz
ADjoint local preconditioner ILU
ILU fill level 2
Preconditioner reordering RCM
GMRES subspace size 100
ASM overlap 2
SNOPT Parameters
Major feasibility tolerance 10−6
Major optimality tolerance 10−6
Major step limit 0.1
Linesearch Non-derivative
CFD Parameters
Parameter Value units
Time-spectral ω 6.28 rad./s
Time-spectral amplitude 0.034 rad.
k2 dissipation 0.25
k4 dissipation 0.01562
Time-spectral instances 3
Dissipation lumping parameter 6
Multigrid restriction relaxation 0.8
Dissipation scaling exponent 0.1
Startup CFL number 0.9
Multigrid cycle 4w
CFD Convergence tolerance 10−10
Runge-Kutta to Newton-Krylove switch tolerance 0.001
Newton-Krylov global preconditioner Additive Schwarz
Newton-Krylov local preconditioner ILU
ILU fill level 2
Preconditioner reordering RCM
ASM overlap 2
GMRES subspace size 100
Newton-Krylov Jacobian reset lag 10
ADjoint Parameters
ADjoint convergence tolerance 1e-8
ADjoint global preconditioner Additive Schwarz
ADjoint local preconditioner ILU
ILU fill level 2
Preconditioner reordering RCM
GMRES subspace size 100
ASM overlap 2
SNOPT Parameters
Major feasibility tolerance 10−6
Major optimality tolerance 5 × 10−5
Major step limit 0.1
Linesearch Non-derivative
CFD Parameters
Parameter Value units
Time-spectral ω 6.28 rad./s
Time-spectral amplitude 0.0017 rad.
k2 dissipation 0.25
k4 dissipation 0.01562
Time-spectral instances 3
Dissipation lumping parameter 6
Multigrid restriction relaxation 0.8
Dissipation scaling exponent 0.1
Startup CFL number 0.9
Multigrid cycle 4w
CFD Convergence tolerance 10−10
Runge-Kutta to Newton-Krylove switch tolerance 0.001
Newton-Krylov global preconditioner Additive Schwarz
Newton-Krylov local preconditioner ILU
ILU fill level 2
Preconditioner reordering RCM
ASM overlap 2
GMRES subspace size 100
Newton-Krylov Jacobian reset lag 10
ADjoint Parameters
ADjoint convergence tolerance 1e-8
ADjoint global preconditioner Additive Schwarz
ADjoint local preconditioner ILU
ILU fill level 2
Preconditioner reordering RCM
GMRES subspace size 100
ASM overlap 2
SNOPT Parameters
Major feasibility tolerance 10−6
Major optimality tolerance 10−6
Major step limit 0.5
Linesearch Non-derivative
CFD Parameters
Parameter Value units
Time-spectral ω 6.28 rad./s
Time-spectral amplitude 0.034 rad.
k2 dissipation 0.25
k4 dissipation 0.01562
Time-spectral instances 3
Dissipation lumping parameter 6
Multigrid restriction relaxation 0.8
Dissipation scaling exponent 0.1
Startup CFL number 0.9
Multigrid cycle 4w
CFD Convergence tolerance 10−10
Runge-Kutta to Newton-Krylove switch tolerance 0.001
Newton-Krylov global preconditioner Additive Schwarz
Newton-Krylov local preconditioner ILU
ILU fill level 2
Preconditioner reordering RCM
ASM overlap 2
GMRES subspace size 100
Newton-Krylov Jacobian reset lag 10
ADjoint Parameters
ADjoint convergence tolerance 1e-8
ADjoint global preconditioner Additive Schwarz
ADjoint local preconditioner ILU
ILU fill level 2
Preconditioner reordering RCM
GMRES subspace size 100
ASM overlap 2
SNOPT Parameters
Major feasibility tolerance 10−6
Major optimality tolerance 5 × 10−5
Major step limit 0.1
Linesearch Non-derivative
CFD Parameters
Parameter Value units
Time-spectral ω 6.28 rad./s
Time-spectral amplitude 0.0017 rad.
k2 dissipation 0.5
k4 dissipation 0.01562
Time-spectral instances 3
Dissipation lumping parameter 4
Multigrid restriction relaxation 0.8
Dissipation scaling exponent 0.1
Startup CFL number 0.9
Multigrid cycle 4w
CFD Convergence tolerance 10−10
Runge-Kutta to Newton-Krylove switch tolerance 0.001
Newton-Krylov global preconditioner Additive Schwarz
Newton-Krylov local preconditioner ILU
ILU fill level 2
Preconditioner reordering RCM
ASM overlap 2
GMRES subspace size 150
Newton-Krylov Jacobian reset lag 10
ADjoint Parameters
ADjoint convergence tolerance 1e-8
ADjoint global preconditioner Additive Schwarz
ADjoint local preconditioner ILU
ILU fill level 2
Preconditioner reordering RCM
GMRES subspace size 150
ASM overlap 2
SNOPT Parameters
Major feasibility tolerance 10−6
Major optimality tolerance 5 × 10−5
Major step limit 0.5
Linesearch Non-derivative
Note that the baseline FFD is the undeformed FFD that is the starting point for all designs. To
acheive the final design, embed the geometry in the baseline FFD and move its control points
to the values for the desired case. Note that the FFD volumes used in this thesis are cubic
spline volumes.
145
Appendix D. FFD Coordinate Tables 146
Table D.3: FFD coordinates for bending moment constrained case, M = 0.5
Appendix D. FFD Coordinate Tables 149
Table D.7: FFD coordinates for baseline case with shape variables, M = 0.5, part 1
Appendix D. FFD Coordinate Tables 153
Table D.8: FFD coordinates for baseline case with shape variables, M = 0.5, part 2
Appendix D. FFD Coordinate Tables 154
Table D.9: FFD coordinates for Cmα constrained case with shape variables, M = 0.5, part 1
Appendix D. FFD Coordinate Tables 155
Table D.10: FFD coordinates for Cmα constrained case with shape variables, M = 0.5, part 2
Appendix D. FFD Coordinate Tables 156
Table D.11: FFD coordinates for Kn constrained case with shape variables, M = 0.5, part 1
Appendix D. FFD Coordinate Tables 157
Table D.12: FFD coordinates for Kn constrained case with shape variables, M = 0.5, part 2
Appendix D. FFD Coordinate Tables 158
Table D.13: FFD coordinates for CAP constrained case with shape variables, M = 0.5, part 1
Appendix D. FFD Coordinate Tables 159
Table D.14: FFD coordinates for CAP constrained case with shape variables, M = 0.5, part 2
Appendix D. FFD Coordinate Tables 160
Table D.15: FFD coordinates for bending moment constrained case, M = 0.7
Appendix D. FFD Coordinate Tables 161
Table D.19: FFD coordinates for bending moment constrained case with shape variables, M =
0.7, part 1
Appendix D. FFD Coordinate Tables 165
Table D.20: FFD coordinates for bending moment constrained case with shape variables, M =
0.7, part 2
Appendix D. FFD Coordinate Tables 166
Table D.21: FFD coordinates for Cmα constrained case with shape variables, M = 0.7, part 1
Appendix D. FFD Coordinate Tables 167
Table D.22: FFD coordinates for Cmα constrained case with shape variables, M = 0.7, part 2
Appendix D. FFD Coordinate Tables 168
Table D.23: FFD coordinates for Kn constrained case with shape variables, M = 0.7, part 1
Appendix D. FFD Coordinate Tables 169
Table D.24: FFD coordinates for Kn constrained case with shape variables, M = 0.7, part 2
Appendix D. FFD Coordinate Tables 170
Table D.25: FFD coordinates for CAP constrained case with shape variables, M = 0.7, part 1
Appendix D. FFD Coordinate Tables 171
Table D.26: FFD coordinates for CAP constrained case with shape variables, M = 0.7, part 2
Appendix D. FFD Coordinate Tables 172
Table D.27: FFD coordinates for bending moment constrained case with shape variables, M =
0.85, part 1
Appendix D. FFD Coordinate Tables 173
Table D.28: FFD coordinates for bending moment constrained case with shape variables, M =
0.85, part 2
Appendix D. FFD Coordinate Tables 174
Table D.29: FFD coordinates for Cmα constrained case with shape variables, M = 0.85, part 1
Appendix D. FFD Coordinate Tables 175
Table D.30: FFD coordinates for Cmα constrained case with shape variables, M = 0.85, part 2
Appendix D. FFD Coordinate Tables 176
Table D.31: FFD coordinates for Kn constrained case with shape variables, M = 0.85, part 1
Appendix D. FFD Coordinate Tables 177
Table D.32: FFD coordinates for Kn constrained case with shape variables, M = 0.85, part 2
Appendix D. FFD Coordinate Tables 178
Table D.33: FFD coordinates for CAP constrained case with shape variables, M = 0.85, part 1
Appendix D. FFD Coordinate Tables 179
Table D.34: FFD coordinates for CAP constrained case with shape variables, M = 0.85, part 2