Вы находитесь на странице: 1из 65

AERODYNAMIC SHAPE OPTIMIZATION OF

A FINITE WING AT LOW REYNOLDS


NUMBER
A Thesis Submitted
in Partial Fulllment of the Requirements
for the Degree of
Master of Technology
by
Sandeep Kumar Attree
to the
Department of Aerospace Engineering
INDIAN INSTITUTE OF TECHNOLOGY KANPUR
INDIA
August, 2012
To my family

Abstract
A continuous adjoint based method is used for shape optimization of a nite wing
exhibiting maximum aerodynamic performance in steady ows, at Reynolds number
= 10
3
, and angle of attack = 4
o
. Objective functions used is minimization of drag
to lift ratio which is same as maximization of lift to drag ratio. The airfoil shape
is parametrized by a 4
th
order non-uniform rational B-Spline (NURBS) curve with
13 control points. The wing surface is parametrized by 6 such control polygons.
Design variables used correspond to y-coordinates of control points for airfoils and
scaling factor of control polygons as well as allowing for wing sweep. To prevent
the optimizer from getting stuck in a local optima with lower aerodynamic perfor-
mance the number of variables for optimization process is increased progressively.
The process begins with optimization of airfoil in 2D; the solution obtained is used
as initial guess to optimize the a rectangular wing with uniform cross-section. Us-
ing the solution obtained from previous step as initial guess, cross-section of the
wing is optimized. Finally, optimal cross-section is used as initial guess for full-wing
optimization. A stabilized nite element method based on streamline upwind Petro-
v/Galerkin (SUPG) and pressure stabilizing Petrov/Galerkin (PSPG) stabilization
technique is employed to solve both, the ow and the adjoint equation. The limited
memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) algorithm is used to mini-
mize or maximize the objective function. The optimal shape has a bulbous leading
edge resulting in increased suction pressure over the upper surface, therefore increas-
ing the lift. It also has bulged leading and trailing edges leading to an increased
suction pressure zone increasing the lift further. This optimization solution is also
iii
iv
associated with an increase in drag but the relative increase in lift is more than that
in drag, hence an increased aerodynamic performance.
Acknowledgments
I would like to express my sincere gratitude towards my thesis supervisor, Prof.
Sanjay Mittal, who has been an immense source of inspiration for the last two years.
His knowledge, expertise and work ethic has really enlightened and motivated me
to complete my thesis.
I would like to thank Nikhil, who was my mentor and whose research I have carried
forward. He helped me at every step and cleared silliest of my doubts with patience.
I have learnt quite a lot from him, not just about research, but also about handling
things in an organized manner.
My thanks also to all my lab mates in the CFD lab, especially Murali, Navrose,
Mohsin, Deva and Suresh, for their help and making the lab environment enjoyable.
I will always cherish the moments that I have spent with them in the lab.
Finally, I would like to thank my family for believing in me and for their uncondi-
tional love and support.
v
Contents
Certicate i
Abstract ii
Acknowledgment v
Contents vi
List of Figures viii
List of Tables xi
Nomenclature xi
1 Introduction 1
1.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Thesis organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Governing Flow Equations and Finite Element Formulation 5
2.1 The Navier-Stokes equations . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Finite element formulation . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Stabilization parameter . . . . . . . . . . . . . . . . . . . . . . . . . . 9
vi
CONTENTS vii
3 Optimization and the adjoint approach 11
3.1 Problem set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2 The adjoint approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 The continuous adjoint equations . . . . . . . . . . . . . . . . 14
3.2.2 Adjoint boundary and terminal conditions . . . . . . . . . . . 15
3.2.2.1 Adjoint boundary conditions for steady ows . . . . 16
3.2.2.2 Adjoint boundary conditions for unsteady ows . . . 16
3.2.3 Finite element formulation for adjoint equations . . . . . . . . 17
3.2.4 The gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2.4.1 The optimizer . . . . . . . . . . . . . . . . . . . . . . 18
4 Low Re wing design for steady ows 21
4.1 Parametrization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.1.1 Parametrization of airfoil . . . . . . . . . . . . . . . . . . . . . 22
4.1.2 Parametrization of Wing . . . . . . . . . . . . . . . . . . . . . 23
4.2 3D Mesh Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.3 Mesh movement strategy . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.4 Mesh convergence study . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.5 Solution strategy for optimization problem . . . . . . . . . . . . . . . 28
4.6 Progressive approach for wing design . . . . . . . . . . . . . . . . . . 29
5 Computational results 31
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2 Optimization of airfoil for a rectangular wing planform . . . . . . . . 32
5.3 Optimization of cross-section for a rectangular wing planform . . . . 33
5.4 Full-wing optimization . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6 Conclusions 41
6.1 Future directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
References 45
List of Figures
2.1 Schematic of the problem set-up: boundary conditions.
U
and
D
are the upstream and downstream boundaries, respectively.
S
and

W
are the lateral boundaries.
B
represent the surface body. . . . . 7
4.1 Parametrization of a NACA 0012 airfoil using 4
th
order NURBS curve
with 13 controls points. The y-coordinates of the control points 26
and 812 are used as design variables for optimizing the airfoil. The
airfoil surface is shown in broken lines. Upper bounds of the design
variables on the upper surface (y
up
max
) and the lower bounds of the
design variables on the lower surface (y
low
min
) are also shown. . . . . . 23
4.2 Control net for a uniform cross-section rectangular wing. . . . . . . . 24
4.3 The close up view of 2D nite element mesh for around wing at 4
o
angle of attack; (a) before wing tip, (b) after wing tip . . . . . . . . 25
4.4 A 3D nite element mesh for a nite wing. . . . . . . . . . . . . . . 26
4.5 Distribution of 2D slices along the span. . . . . . . . . . . . . . . . . 27
4.6 A ow-chart detailing the steps of the optimization process. . . . . . 28
4.7 A graphical representation of the design process; (a) optimization of
airfoil, (b) optimization of cross-section and (c) full-wing optimization 30
5.1 Initial and optimal shapes for maximum performance at Re = 10
3
and = 4
o
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.2 Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: pressure distribu-
tion on the Upper surface of (a) initial shape (b) optimal shape. . . . 34
ix
x LIST OF FIGURES
5.3 Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: C
p
distribution on
the surface at (a) root, (b) 54.5% span and (c) 97.3% span. . . . . . . 34
5.4 Initial and optimal shapes for maximum performance at Re = 10
3
and = 4
o
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.5 Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: pressure distribu-
tion on the Upper surface of (a) initial shape (b) optimal shape. . . . 35
5.6 Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: C
p
distribution on
the surface at (a) root, (b) 54.5% span and (c) 97.3% span. . . . . . . 36
5.7 Initial and optimal shapes for maximum performance at Re = 10
3
and = 4
o
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.8 Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: pressure distribu-
tion on the Upper surface of (a) initial shape (b) optimal shape. . . . 37
5.9 Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: C
p
distribution on
the surface at (a) root, (b) 54.5% span and (c) 97.3% span. . . . . . . 37
5.10 Initial and optimal shapes for maximum performance at Re = 10
3
and = 4
o
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.11 Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: pressure distribu-
tion on the Upper surface of (a) initial shape (b) optimal shape. . . . 39
5.12 Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: C
p
distribution on
the surface at (a) root, (b) 54.5% span and (c) 97.3% span. . . . . . . 39
List of Tables
4.1 The lift and drag coecients at = 4
o
and Re = 10
3
for dierent
nite element meshes. . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Nomenclature
C
d
2D coecient of drag
C
l
2D coecient of lift
C
D
3D coecient of drag
C
L
3D coecient of lift
C
p
coecient of pressure
f body force
h element length
I identity tensor
I
c
objecive function
I augmented objective function
N B-spline basis functions
n Unit normal vector
P coodinates of control points
p pressure
q weight function
Re Reynolds number
s adjoint stress tensor
T deviatric stress tensor
T extent of time integration
t non-dimensional time
U Velocity eld
U

free stream velocity


u, v inline and transverse components of velocity
w weight function
x spatial coordiantes
xii
NOMENCLATURE xiii
Greek letters
density
stress tensor
spatial domain
boundary of spatial domain

U
,
D
,
S
,
B
upstream, downstream, lateral and surface boundaries
dynamic viscosity
kinematic viscosity

SUPG
coecient of SUPG stabilization terms

PSPG
coecient of PSPG stabilization terms

LSIC
coecient of least square stabilization terms
design parameters
adjoint eld

u1
,
u2
inline and transverse components of adjoint velocity

p
adjoint pressure
Chapter 1
Introduction
The wing is the most important component of an airplane. The wing aects not
only its own performance, but also the performance of the entire airplane. From an
aerodynamic perspective, the main source of the airplane drag is associated with
the wing. Nearly two-thirds of the total drag is related to the wing. Therefore
a large emphasis should be placed on the aerodynamic wing-design. A standard
approach used in the past to design a wing involves two main steps. The rst
step is to establish the general wing shape, airfoil types, approximate thickness
to chord ratio and design lift coecient from a design optimization study. The
second step is to design a minimum weight wing while achieving as much of the
two-dimensional airfoil performance as possible, minimizing any parasite or induced
drag penalties, and satisfying all of the other general design criteria such as good
drag characteristics, sucient fuel volume, and minimum weight. For modern wing
design, the procedure still follows the standard approach, but involves the use of
computational uid dynamics (CFD) codes to systematically study variations of
shape, airfoil type, thickness, sweep, etc. The incompressible Euler and Navier-
Stokes equations are solved using CFD and to provide a better understanding of the
1
2 1 Introduction
ow over the wing, but not to drive the wing design process. The wing renement
is done based on the intuition and experience of the designer.
Recently, CFD has been receiving attention as a design optimization tool. Numer-
ical optimization has been combined with CFD. While there exist numerous well
developed optimization techniques, most techniques are tuned to the CFD-based
optimization problems due to dierent nature of the problems. The standard opti-
mization problems generally assume that the function evaluation of the cost function
can be evaluated easily at low computational cost. The standard gradient-based op-
timization further expect similar simplicity of the gradient evaluation. Moreover,
they are mainly developed for problems of a small to moderate number of design
parameters compared to the number of cost functions and constraints. The CFD-
based optimization problems, on the other hand, require solving a non-linear system
of partial dierential equations to calculate the cost function and usually involve a
large number of design parameters.
Early attempts to combine CFD with numerical optimization made use of nite-
dierence calculations to obtain gradient information. This technique can be used
to obtain the derivatives of all the ow quantities with existing ow solvers. One
problem with this approach is the computational time required. To obtain the design
sensitivities for a system involving n design parameters using one sided dierence
requires a well-converged solution of n + 1 ow analysis problems. This becomes
prohibitive when n becomes large.
A practical approach for CFD-based optimization emerged in 1988 when an adjoint
method was rst applied to the governing equations of transonic ow by Jameson
[6]. The method applied control theory techniques to a system of partial dierential
equation and reduced the cost of gradient calculations to only two ow analysis,
regardless of the number of design parameters n. This breakthrough enabled many
1.1. Objectives 3
more design variables than was previously practical for gradient-based aerodynamic
shape optimization problem. During the last decade, adjoint methods have been
extensively developed and proven to be very eective for improving wing section
shapes for xed wing-planform. The framework has been well developed by Jameson
[6, 7, 8, 9]. He applied the adjoint method to redesign airfoils and wing sections in
transonic ow. By using surface mesh points as the design parameter, the adjoint
method to calculate the gradient information, and an iterative method to improve
the shape in each iteration, he was able to achieve a shock-free-wing.
A continuous adjoint based method for aerodynamic shape optimization was formu-
lated and implemented by Srinath and Mittal [21] for steady low Reynolds number
ows. The method was used for nding optimal airfoil shapes at low Re ows [22].
It was further extended for shape optimization in unsteady viscous ows [24]. The
work in 2D was followed by development and validation of tools for optimization in
3D.
1.1 Objectives
The present study is aimed at implementing the continuous adjoint-based method
to optimize a nite wing. The computations are carried out in steady, low Reynolds
number ow at Re = 10
3
and = 4
o
. The adjoint-based methods are well suited
for problems involving large number of design variables, but they may converge to
a worse optimum. Hence, the number of design variables is to be increased pro-
gressively. To investigate the optimized wing obtained for maximum aerodynamic
performance.
1.2 Thesis organization
Chapter 2 describes the ow equations and the the nite element method used to
solve them. It also describes the boundary conditions used in the computations.
4 1.2. Thesis organization
The problem set-up is presented in Chapter 3. A stabilized nite element method
is presented to solve the adjoint equations. Also presented are the details of the
Limited-memory Bradan-Fletcher-Gilda-Shanno(L-BFGS) algorithm that is used as
the optimizer.
Chapter 4 deals with the shape parametrization and 3D mesh generation. It also
explains the wing design strategy.
Chapter 5 focuses on the optimization of wing for maximum performance at Re =
10
3
for an angle of attack of 4
o
. The design process is initiated with two dierent
initial guesses.
Concluding remarks and possible directions for future research are presented in
Chapter 6.
Chapter 2
Governing Flow Equations and
Finite Element Formulation
In this chapter the governing equations for the viscous, incompressible unsteady
ow is presented. This is followed by the nite element discretization that is used
to solve these equations.
2.1 The Navier-Stokes equations
Let R
n
sd
and (0, T) be the spatial and temporal domains, respectively, where n
sd
is the number of space dimensions, and let denote the boundary of . The spatial
and temporal co-ordinates are denoted by x and t. The Navier-Stokes equations
governing incompressible ow are

_
u
t
+ u.u
_
. = 0 on (0, T) (2.1)
.u = 0 on (0, T) (2.2)
5
6 2.1. The Navier-Stokes equations
Here , u and are the density, velocity and stress tensor, respectively. The stress
tensor is written as the sum of its isotropic and deviatric parts:
= pI + T, T = 2(u), (u) =
1
2
(u + (u)
T
)
where p, I and are the pressure, identity tensor and dynamic viscosity, respectively.
The boundary conditions are either on the ow velocity or stress. Both, Dirichlet
and Neumann type boundary conditions are accounted for:
u = g on
g
(2.3)
n. = h on
h
(2.4)
where, n is the unit normal vector on the boundary . Here,
g
and
h
are subsets
of the boundary . More details on the boundary conditions are given in Figure
2.1.
U
and
D
represent the upstream and downstream boundaries, respectively.

S
and
W
represent the lateral boundaries. The surface of the body is represented
by
B
.
The initial condition on the velocity is specied on :
u(x, 0) = u
0
on (2.5)
where u
0
is divergence free.
The drag and lift force coecients, (C
d
, C
l
), on the body are calculated using the
following expression:
(C
d
, C
l
) =
2
U
2
S
_

B
nd (2.6)
2.2. Finite element formulation 7
Figure 2.1: Schematic of the problem set-up: boundary conditions.
U
and
D
are
the upstream and downstream boundaries, respectively.
S
and
W
are the lateral
boundaries.
B
represent the surface body.
The time-averaged coecients are calculated as follows:
C
d
=
1
T
_
t
0
+T
t
0
C
d
(t)dt (2.7)
C
l
=
1
T
_
t
0
+T
t
0
C
l
(t)dt (2.8)
The time-averaging begins at t = t
0
to leave out the transient eect of the initial
condition on the fully developed ow.
2.2 Finite element formulation
Consider a nite element discretization of into subdomain
e
,e = 1, . . . , n
el
, where
n
el
is the number of elements. We dene:
H
1h
() = {
h
|
h
C
0
(),
h
|

e P
1
, e = 1, 2, . . . , n
el
} (2.9)
8 2.2. Finite element formulation
with P
1
representing rst-order polynomials. The trial and test function spaces are
dened as:
S
h
u
= {u
h
|u
h
(H
1h
)
n
sd
, u
h
= g
h
on
g
} (2.10)
V
h
u
= {w
h
|w
h
(H
1h
)
n
sd
, w
h
= 0 on
g
} (2.11)
S
h
p
= V
h
p
= {q
h
|q
h
H
1h
} (2.12)
where n
sd
is the number of space dimensions.
The stabilized nite element formulation of Equations (2.1) and (2.2) is written as
follows: nd u
h
S
h
u
and p
h
S
h
p
such that w
h
V
h
u
, q
h
V
h
p
_

w
h
.
_
u
h
t
+ u
h
.u
h
_
d +
_

(w
h
) : (p
h
, u
h
) d
+
_

q
h
.u
h
d +
n
el

e=1
_

e
1

(
SUPG
u
h
.w
h
+
PSPG
q
h
).
_

_
u
h
t
+ u.u
_
.
_
d
e
+
n
el

e=1
_

LSIC
.w
h
.u
h
d
e
=
_

h
w
h
.h
h
d. (2.13)
The rst three terms and the right-hand side of Equation (2.13) constitute the
Galerkin formulation of the problem. The terms involving the element level integrals
are the stabilization terms added to the basic Galerkin formulation to enhance its
numerical stability. The term with coecient
LSIC
is also a stabilization term
based on the least squares of the incompressibility constraint and is found to be
useful for large Reynolds number ows. The generalized trapezoidal rule is used for
time discretization of u
h
. This is given by:
u
h
= u
i
+ (1 )u
i+1
, (2.14)
u
h
t
=
u
i+1
u
i
t
(2.15)
2.3. Stabilization parameter 9
where 0 1, and u
i
represents the velocity eld at time step i.
2.3 Stabilization parameter
The coecients
SUPG
and
PSPG
in Equation (2.13) are given by:

SUPG
=
PSPG
=
_
_
2||u
h
||
h
_
2
+
_
12
h
2
_
2
_
(2.16)
where h is the element length. In the present work, h is taken as the length of the
shortest edge in an element. The coecient
LSIC
is dened as

LSIC
=
_
_
2
h||u
h
||
2
_
2
+
_
12
h
2
||u
h
||
2
_
_
1/2
(2.17)
More details on the formulation can be found in Tezduyar et al. [25].
Chapter 3
Optimization and the adjoint
approach
In this chapter the problem set-up and the solution strategy used for optimal shape
design is presented. The adjoint method for design in uid ows is presented in
detail in section 3.2. The adjoint equations along with the boundary conditions are
derived in this section.
3.1 Problem set-up
The objective of optimal shape design is to alter either a part or whole of a surface
to achieve a desired objective function. The objective could be to achieve a shape
having the maximum lift, least drag or highest aerodynamic eciency. The objec-
tives are posed in the form of cost functions. The cost function I
c
is a function of
and U, where is the set of shape parameters that are used to represent the shape
and U represents the ow parameters. An optimal shape problem is formulated as:
Obtain , that minimizes or maximizes the cost function I
c
(U, ) subject to satis-
fying the following constraints:
11
12 3.1. Problem set-up
PDE constraints (U) = 0
Geometric constraints g() = 0
Optimization can be broadly classied as non-gradient based and gradient based
methods. Non-gradient based methods use only the cost function to obtain a op-
tima. Simplex [26], simulated annealing [27] and genetic algorithms [28] are a few
such methods that have been used for engineering applications. These methods are
simple to code and are very robust. They are well suited for problems in which the
gradient information is dicult to obtain. The drawback of these methods is the
high computational cost involved. For N design variables, the number of function
evaluations required is O(N
2
) while using genetic algorithms.
Gradient based methods use both the value of cost function and the gradient infor-
mation while seeking an optimum. From a Taylor series expansion
I
c
+ I
c
I
c
+
_
I
c

_
T
The gradient vector (I
c
/) may be used to obtain a search direction that mini-
mizes or maximizes the cost function. If one chooses
=
_
I
c

_
then, for suciently small , the cost function has to diminish.
I
c
+ I
c
= I
c

_
I
c

__
I
c

_
T
< I
c
Further improvements can be obtained by using quasi-Newton methods which use
the Hessian (
2
I
c
/
2
) to obtain new search direction. The simplest way to ob-
tain gradients (I
c
/
i
) is by using nite dierences. For each design variable, its
3.2. The adjoint approach 13
value is varied by a small quantity
i
, the ow and objective function recomputed
for the modied shape and the derivative of I
c
with respect to
i
calculated. An
optimization problem with N design variable requires N + 1 ow computations if
rst order forward dierences are used. Although the approach is simple, it involves
large computational costs if the number of design variables is large. Another disad-
vantage while using this approach is that the appropriate step size
i
is problem
specic and is not known apriori.
The gradients can be obtained in a fast and more ecient manner if the optimiza-
tion problem is formulated in the framework of theory of optimal control of systems
governed by partial dierential equations [29]. For example, shape parameters repre-
senting an airfoil can be considered as controls and the ow equations as constraints.
This leads to the so called adjoint based methods.
3.2 The adjoint approach
Let
B
be the segment of the boundary whose shape is to be determined. Let =
(
1
, . . . ,
m
) be the set of shape parameters that determine
B
. The optimization
problem involves nding the shape parameters that minimize (or maximize) the
objective function, I
c
(U, ) subject to satisfying (U, ) = 0, where represents
the ow equations. Depending on the accuracy needed various approximations to the
ow equations could be used. The ow equations are treated as constraint conditions
on the objective function. An augmented objective function is constructed to convert
the constrained problem to an unconstrained one. The ow equations are augmented
to the objective function by introducing a set of Lagrange multipliers or adjoint
variables .
I = I
c
+
_

.d (3.1)
14 3.2. The adjoint approach
The augmented objective function (3.1) degenerates to the original one if the ow
variables, U, exactly satisfy the ow equations. The rst variation of the augmented
objective function is given by:
I =
I
U
U +
I

+
I

(3.2)
The optimal solution is achieved when the variation of the augmented objective
function vanishes, i.e., I = 0. This requires that the variation of I with respect to
the ow variables U, design parameters and the adjoint variables should go to
zero, independently. These variations are given as:
I

= (U, ) (3.3)
I
U
=
_
I
c
U
+
_

T

U
d
_
(3.4)
I

=
_
I
c

+
_

d
_
(3.5)
Setting the variation of I with respect to to zero gives back the ow equations.
The variation of I with respect to U, given by Equation (3.4), when set to zero,
leads to a set of equations and boundary conditions that are used to evaluate the
adjoint variables. The gradient,
I

, given by Equation (3.5) is utilized to nd the


optimal shape parameters. When the ow and adjoint equations are satised, the
augmented objective function is seen to be a function only of the design parameters.
This implies that the gradient can be determined without the need for additional
ow computations, i.e. the cost of evaluating the gradients is independent of the
number of design parameters.
3.2.1 The continuous adjoint equations
The equations and boundary conditions for the adjoint variables are obtained by
setting the variation of I with respect to the ow variables U, given in equation
3.2. The adjoint approach 15
(3.4), to zero. The equations governing the adjoint variables are:

u
t
+ (u)
T

u
(u.)
u
_
.

= 0 on (0, T) (3.6)
.
u
= 0 on (0, T) (3.7)
where,

is similar to the stress tensor and is given by

=
p
I + [
u
+
(
u
)
T
].
The adjoint equations (3.6) and (3.7) are a set of coupled linear partial dierential
equations. Unlike the ow equations (2.1) and (2.2), the equations for the adjoint
variables are posed backward in time.
3.2.2 Adjoint boundary and terminal conditions
The boundary conditions on the adjoint variables are:

u
= 0 on
U
(3.8)
s = 0 on
D
(3.9)
s
1
= 0,
u2
= 0 on
S
(3.10)

_
T
0
_

B
(.n).
u
ddt +
I
c
u
u +
I
c
p
p = 0 on
B
(3.11)
where, s = {u
u
I
p
+ [
u
+ (
u
)
T
]}.n.
U
,
D
and
S
represent the
upstream, downstream and lateral boundaries, respectively. The surface of the
body is represented by
B
. The terminal condition on the adjoint velocity is given
by:

u
(u, T) = 0 on (3.12)
The conditions on the boundary, given by equation (3.11), depend on the denition
of the objective function. This is illustrated with a few examples for steady and
unsteady ows.
16 3.2. The adjoint approach
3.2.2.1 Adjoint boundary conditions for steady ows
Let I
c
=
1
2
C
2
d
, where, C
d
is the coecient of drag acting on the body. This objective
function is useful in designing a body that yields minimum drag coecient. In this
case, it can be shown that, the boundary conditions for the adjoint variables on the
body surface given by Equation (3.11) can be simplied to:

u
= (C
d
, 0) on
B
3.2.2.2 Adjoint boundary conditions for unsteady ows
The objective function to minimize the time-averaged drag coecient is given by
I
c
=
1
2
C
d
2
. In this situation the boundary condition is given by

u
=
_

C
d
T
, 0
_
on
B
(0, T)
Here, T is the time duration for which the aerodynamic coecients are averaged
and adjoints computed. Similarly, the objective function to maximize the ratio of
the time-averaged lift coecient to the time-averaged drag coecient is given by
I
c
=
1
2
_
C
l
C
d
_
2
. In this case the boundary condition on
B
is given by

u
=
_

1
T
C
l
2
C
d
3
,
1
T
C
l
C
d
2
_
on
B
(0, T)
3.2. The adjoint approach 17
3.2.3 Finite element formulation for adjoint equations
A stabilized SUPG/PSPG nite element method is proposed to solve the adjoint
equations (3.6) and (3.7). The trial and test function spaces are dened as:
S
h
u
= {
h
u
|
h
u
(H
1h
)
n
sd
,
h
u
.
= g
h
on
g
}
V
h
u
= {w
h
u
|w
h
u
(H
1h
)
n
sd
, w
h
u
.
= 0 on
g
}
S
h
p
= V
h
p
= {q
h
p
|q
h
p
H
1h
}
The stabilized nite element formulation of equations (3.6) and (3.7) is as follows:
given u
h
and p
h
satisfying equations (2.1) and (2.2), nd
h
u
S
h
u
and
h
p
S
h
p
such that w
h
u
V
h
u
and q
h
p
V
h
p
,
_

w
h
u
.
_

h
u
t
+ (u
h
)
T

h
u
u.
u
_
d
+
_

(w
h
u
) :

(
h
p
,
h
u
) d +
_

q
h
p
.
h
u
d
+
n
el

e=1
_

e
1

SUPG
(
_
u
h
)
T
w
h
u
u
h
.w
h
u
_
+
PSPG
q
h
p
_
.
_

h
u
t
+ (u
h
)
T

h
u
u.
u
_
.

(
h
p
,
h
u
)
_
d
e
+
n
el

e=1
_

LSIC
.w
h
u
.
h
u
d
e
= 0 (3.13)
The stabilization coecients
SUPG
,
PSPG
and
LSIC
in the formulation proposed
in equation (3.13) are computed based on the ow variables (u, p) and are given in
equations (2.16) and (2.17), respectively.
18 3.2. The adjoint approach
3.2.4 The gradient
The variation of the augmented objective I function with respect to the design
parameters gives the gradient. The gradient is given by:
I

=
I
c

_
T
0
_

p
.uddt
+

_
T
0
_

u
.
_
(
u
t
+ u.u) .
_
ddt (3.14)
It quanties the sensitivity of the objective function to the design parameters. It is
used by the optimizer to rene the search direction. The vanishing of the gradient
reects the attainment of optimal shape.
3.2.4.1 The optimizer
The optimization algorithm used in the present work is the L-BFGS (Limited
memory-Broyden-Fletcher-Goldfarb-Shanno) procedure. This method is derived
from the Newton-Raphson method that seeks to nd a stationary point of a function
f(x) where x could be a vector. In the Newton-Raphson method, marching from
iterate n to n + 1 is is done by the following:
x
n+1
= x
n

n
[Hf(x
n
)]
1
f(x
n
) (3.15)
where
n
is the step length, f(x
n
) and Hf(x
n
) are the gradient and the Hessian
matrix, respectively. In general, computing the Hessian and its inverse are expensive
operations and need to be approximated. In the BFGS approach, the following is
undertaken to march from iterate n to n + 1:
x
n+1
= x
n

n
B
n
f(x
n
) (3.16)
3.2. The adjoint approach 19
where the inverse Hessian approximation B
n
is updated at every iteration by means
of the formula
B
n+1
= V
T
n
B
n
V
n
+
n
s
n
s
T
n
(3.17)
where

n
=
1
y
T
n
s
n
, V
n
= I
n
y
n
s
T
n
and
s
n
= x
n+1
x
n
, y
n
= f
n+1
f
n
The inverse Hessian approximation B
n
is generally very dense. For a problem with
a large number of design variables the cost of storing B
n
becomes prohibitive. To
circumvent this a certain number of vector pairs {s
i
, y
i
} are stored. The product
B
n
f(x
n
) is then obtained by performing a series of inner products and vector
summations involving f(x
n
) and the pairs {s
i
, y
i
}. More details about this method
can be obtained in the book by Nocedal and Wright [30] and in the article by Byrd
et al. [31]. This algorithm is capable of handling upper and lower bounds on the
design variables.
Chapter 4
Low Re wing design for steady
ows
This chapter deals with the design process of wings at low Reynolds numbers.
Parametrization of wing is explained in Section 4.1. along with details of the
parametrization used to represent the airfoil surface. Section 4.2 deals with the
generation of 3D mesh in the computational domain. Mesh movement strategy to
obtain perturbed mesh for gradient calculation and next optimization cycle is dis-
cussed in Section 4.3. A mesh convergence study to determine a suitable mesh that
provides good estimates of the ow is given in Section 4.4. Section 4.5 explains the
strategy employed to obtain the optimal solution using the optimization algorithm.
Finally, the progressive approach to design high performance wing is presented in
Section 4.6.
A semi-wing of unit root chord length and 4.9 span to chord ratio is placed in a
domain whose outer boundary is a cuboid. The upstream and downstream bound-
aries are located at 20 chord lengths, each, from the trailing edge of the wing. The
lateral boundaries are placed at wing root and 15 chord lengths from the root. The
21
22 4.1. Parametrization
top and bottom boundaries are placed at 10 chord length, each, from the wing.
Close to the wing surface a structured mesh is employed to resolve the boundary
layer ow. The remaining domain is discretized using an unstructured mesh via
Delaunay triangulation.
4.1 Parametrization
Shape Parametrization is one of the most important part of automatic shape op-
timization. Parametrization helps to represent a physical shape or surface in a
mathematical model. The type of parametrization will decide our design variables.
The parametrization must be exible enough to represent a large number of shapes
with few control points. It is desirable that the shapes or surfaces obtained during
the course of optimization, be smooth.
A number of approaches for shape parametrization of airfoils or wings have been
used. Castonguay and Nadarajah [32] talks about the various types of parametriza-
tion techniques and their eect on the optimization process. In the work NURBS(Non
uniform rational B-splines) [33] have been used to parametrize an airfoil. It is further
extended to represent a wing surface.
4.1.1 Parametrization of airfoil
A wide range of curves can be obtained by changing the location and weight of the
control points or by changing the knot sequence. In the present work a 4
th
order
NURBS curve with 13 control points is used to model the surface of the airfoil.
The control points 1 and 13 are identical resulting in a closed curve and an airfoil
with sharp trailing edge. In addition, the leading and trailing edge of the airfoil
are xed to achieve the desired chord length and angle of attack. This allows us
to utilize the remaining 10 control points as design variables. In the present work
4.1. Parametrization 23
low
min
max
up
x
6
7
8
9
10
11
12
y
y
y
1
13
2
3
4
5
Figure 4.1: Parametrization of a NACA 0012 airfoil using 4
th
order NURBS curve
with 13 controls points. The y-coordinates of the control points 2 6 and 8 12
are used as design variables for optimizing the airfoil. The airfoil surface is shown
in broken lines. Upper bounds of the design variables on the upper surface (y
up
max
)
and the lower bounds of the design variables on the lower surface (y
low
min
) are also
shown.
only the y-coordinates of the control points are allowed to move. Each of the design
variable is allowed to change between a lower and a upper bound that is specied a
priori. In this study y
low
min
and y
up
max
are set to 0.2 and 0.4 respectively. The other
bounds(y
low
max
and y
up
min
) are set so that the lower and upper surfaces do not cross
each other.
4.1.2 Parametrization of Wing
A wing surface is generated by using the control polygon generated for NACA0012
airfoil. A control net is made by stacking the control polygon for airfoil in z-direction
i.e. span-wise direction. The knot vector for both s and t parametric direction is
taken to be nonperiodic and uniform. The weight w
i,j
associated with each control
point P
i,j
is xed to 1. The surface is of 4
th
order in both parametric direction.
Figure 4.2 shows a control net of uniform cross-section rectangular wing generated
by it.
24 4.2. 3D Mesh Generation
Figure 4.2: Control net for a uniform cross-section rectangular wing.
Dierent planform shapes can be obtained by suitably scaling the control polygon at
dierent span-wise locations. We call this scaling factor ci. c=1 generates an cross-
section of a unit chord length. A factor
i
is introduced to generate unsymmetrical
planforms.The wing is parametrized using 6 control polygons. An optimization
problem is formulated using the same approach as used for airfoil. c
i
and
i
for
i = 2 to 6 are taken as variables.
i
and c
i
for i = 1 i.e. at root are kept xed.
4.2 3D Mesh Generation
A 3D mesh is generated by stacking slices of 2D mesh along the wing span. A 2D
mesh is generated on a NACA0012 airfoil of chord length 1. Figure 4.3 shows a close
up of the mesh near airfoil. Close to the surface of airfoil, there is structured meshing
in order to resolve the boundary layer. The rest of the domain is discretized using
unstructured meshing via delaunay triangulation. After the wing tip, the complete
domain volume needs to be meshed. Therefore, in the 2D slice which is at the wing
4.3. Mesh movement strategy 25
a) b)
Figure 4.3: The close up view of 2D nite element mesh for around wing at 4
o
angle
of attack; (a) before wing tip, (b) after wing tip
tip, the airfoil is meshed inside. Figure 4.3 shows a close up of the mesh near airfoil
before and after the wing tip. The 2D slice at the wing tip is used in the rest of the
domain. The 2D mesh consists of 21222 nodes and 42140 triangular elements with
200 nodes on the airfoil surface. The 2D mesh outside the wing consists of 22290
nodes and 44473 triangular elements. The complete 3D mesh is shown in Figure
4.4. The concentration of slices near the wing tip is kept high to adequately capture
the wing tip vortices.Figure 4.5 shows the distribution of slices along the span. A
mesh moving scheme is utilized to relocate the nodes of the mesh to accommodate
the modied airfoil geometry.
4.3 Mesh movement strategy
The shape of the wing planform changes after each iteration during the optimization
cycle. Hence, the chord lengths of airfoils at dierent sections of the wing change.
One possible way to handle this situation is to re-mesh after each iteration. Al-
though this type of an implementation works for any arbitrary shaped body and
deformations, it can be computationally quite expensive. In the present work, a
mesh moving scheme is employed. A nite element mesh is generated for the ini-
tial geometry. The nodes are subsequently relocated to conform to the changing
shapes while retaining the connectivity between them. The computational domain
26 4.3. Mesh movement strategy
X
Y
Z
Figure 4.4: A 3D nite element mesh for a nite wing.
is modeled as a linearly elastic solid. To reduce distortion of the mesh the smaller
elements are made stier. The modied equations of linear elasticity are solved
for the internal nodal displacements based on the given shape deformation of the
solid boundary. More details about this procedure can be found in the article by
Tezduyar et al. [34]. If the resulting mesh is associated with unacceptable levels of
element distortion one has the option to generate a new mesh.
4.4. Mesh convergence study 27
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
16.00
0 5 10 15 20 25 30
z
/
c
slice no.
wing tip
Figure 4.5: Distribution of 2D slices along the span.
4.4 Mesh convergence study
A mesh convergence study is carried out to obtain the minimum number of slices
required to capture the data accurately. Meshes with 30, 60, and 120 slices are
used to compute ow over a rectangular wing having NACA 0012 airfoil as its
cross-section at Reynolds number 10
3
and angle of attack of 4
o
. The drag and lift
coecients obtained are given in Table 4.1. Since all the meshes lead to virtually
same values of aerodynamic coecients, mesh with 30 slices have been used for
further computations.
Mesh C
D
C
L
30 0.1238 0.1788
60 0.1238 0.1787
120 0.1236 0.1785
Table 4.1: The lift and drag coecients at = 4
o
and Re = 10
3
for dierent nite
element meshes.
28 4.5. Solution strategy for optimization problem
Figure 4.6: A ow-chart detailing the steps of the optimization process.
4.5 Solution strategy for optimization problem
In the present work an iterative approach is used to solve the optimization problem.
This approach is same as used to solve 2D optimization problems. This approach
has been explained in detail by Srinath and Mittal [21, 22]. It is mentioned here for
completion. This approach is straightforward and simple to code. The major disad-
vantage of this approach is the huge memory it requires specially for 3D problems.
A Flowchart depicting the algorithm for the iterative approach is shown in Figure
4.6 and the steps involved are listed below:
1. Initialize design parameters and generate initial mesh.
2. Compute the unsteady ow, u and p and calculate the objective function I
c
.
3. Find
u
and
p
.
4.6. Progressive approach for wing design 29
4. Compute the gradient given by equation (3.14).
5. If the convergence criteria, either on the objective function or the gradient, is
satised, then stop.
6. Update the shape parameters, , in the direction of the gradient. The L-BFGS
algorithm used to update the parameters in this thesis is detailed in section
3.2.4.1.
7. Modify mesh to accommodate new shape A mesh moving strategy used in the
present work is explained in section 4.3.
Although the solution strategy is same for 2D and 3D problems, there is dierence
in the way in which dierent tasks like mesh generation, solution of ow and adjoint
equation and mesh moving are carried out in 2D and 3D codes.
4.6 Progressive approach for wing design
To design an optimal wing, shape optimization algorithm has to work with relatively
large number of design variables. The variation of the objective function with design
parameters, for a richer design space, is most likely associated with more peaks
and valleys. As a result, gradient-based methods that search for a local optimum
do not necessarily lead to a better design as the design space is enriched. It is,
therefore, proposed that the optimization be carried out with progressive increase in
the number of design variables to exploit the richer design space available. Gradual
increase in the number of design variables ensures that the optimization process does
not get stuck in a local optimum that has worse aerodynamic performance. For each
cycle of optimization, the initial geometry used is the optimal shape obtained from
the previous parametrization. Figure 4.7 shows a graphical representation of the
design process. In the rst step, airfoil in a cross-section is optimized and the same
30 4.6. Progressive approach for wing design
-0.04
-0.02
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
y
/
c
x/c
NURBS
fixed control points
control points
y(1)
y(3)
y(4)
y(6)
y(7)
y(8)
y(9)
y(10)
y(2)
y(5)
-1.00
-0.50
0.00
0.50
1.00
0 1 2 3 4 5 6
x
/
c
z/c
planform
-0.04
-0.02
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
y
/
c
x/c
NURBS
fixed control points
control points
y(1)
y(3)
y(4)
y(6)
y(7)
y(8)
y(9)
y(10)
y(2)
y(5)
-1.00
-0.50
0.00
0.50
1.00
0 1 2 3 4 5 6
x
/
c
z/c
planform
-0.50
0.00
0.50
1.00
0 1 2 3 4 5
x
/
c
z/c
planform
chord and thickness
control polygon center
(1) (2) (3)
(4)
(5)
(6)
c(1) c(2) c(3) c(6) c(4) c(5)
-0.04
-0.02
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
y
/
c
x/c
NURBS
fixed control points
control points
y(1)
y(3)
y(4)
y(6)
y(7)
y(8)
y(9)
y(10)
y(2)
y(5)
b)
c)
a)
Figure 4.7: A graphical representation of the design process; (a) optimization of
airfoil, (b) optimization of cross-section and (c) full-wing optimization
airfoil is used throughout the span of the semi-wing. In the second step, airfoil is
allowed to vary across the span using 6 control polygons i.e. optimization of cross-
section is achieved. In the third and nal step, both cross-section and planform are
optimized.
Chapter 5
Computational results
5.1 Introduction
Shape optimization is carried out to design a wing which exhibits good aerodynamic
performance. The computations are done for Reynolds numbers = 10
3
and an angle
of attack = 4
o
. The cross-sectional airfoil is parametrized by a 4
th
order NURBS
curve with 13 control points. These control points along with the scaling factors
form the design space for the optimization problem. The y coordinate of the control
points is taken as the design variable. The airfoil shape is changed by changing the y
coordinate of the control points. The control points at the leading and trailing edge
are kept xed whereas the remaining are allowed to move between specied upper
and lower bounds. The optimization has been carried out progressively in three
steps; (i) optimization of airfoil for a rectangular wing planform, (ii) optimization
of cross-section for a rectangular wing planform and (iii) full wing optimization.
The objective is to nd the optimal wing having maximum aerodynamic eciency.
The objective is to maximize C
l
/C
d
which is equivalent to minimizing C
d
/C
l
. The
31
32 5.2. Optimization of airfoil for a rectangular wing planform
objective function is dened as:
I
c
=
1
2
_
C
d
C
l
_
2
(5.1)
The ow and the adjoint equations are solved by a stabilized nite element method
based on streamline-upwind Petrov/Galerkin (SUPG) and pressure stabilized Petro-
v/Galerkin (PSPG), as discussed in [24]. The limited memory Broyden-Fletcher-
Goldfarb-Shanno (L-BFGS) algorithm is used to minimize the objective function
[31].
5.2 Optimization of airfoil for a rectangular wing
planform
The objective is to nd the optimal shape airfoil with maximum aerodynamic e-
ciency. The computations begin with a geometry corresponding to a NACA 0012
airfoil and rectangular planform. The optimal shape obtained for Re = 10
3
is shown
along with the initial shape in Figure 5.1. The performance of the wing increases
from 1.449 to 1.987. This corresponds to an increase of 37.06% over the NACA
0012 airfoil. The pressure contours on the upper surface of NACA 0012 wing and
the optimal wing is shown in the Figure 5.2. Figure 5.3 shows C
p
distribution on
the surface of the wing at dierent span-wise locations. The optimal shape has
the following characteristics (a) large bump in the lower surface which account for
large increase in positive pressure near the lower surface of the airfoil (b) bulbous
leading edge which results in increase in suction pressure on the upper surface and
hence the increase in lift increasing drag as well simultaneously. Since the optimizer
tends to converge to a local optima computations are carried out for a second initial
guess, which is an airfoil obtained from optimization of NACA 0012 airfoil using a
5.3. Optimization of cross-section for a rectangular wing planform 33
2D solver at Re = 10
3
and = 4
o
. This airfoil has a performance of 2.65 but in
3D it is diminished to 2.31 due to induced drag and tip votices. The optimal so-
lution obtained in this case gives a performance of 2.875 an improvement of 98.3%
over NACA 0012; hence this solution has been used for further computations.The
optimal shape obtained is shown along with the initial shape in Figure 5.4.The pres-
sure contours on the upper surface of initial and the optimal wing is shown in the
Figure 5.5.Figure 5.6 shows C
p
distribution on the surface of the optimal wing at
dierent span-wise locations. The optimal shape has a bulbous leading but a at
trailing edge. The lift coecient in the two cases are comparable but there is a large
reduction in drag in the latter case.
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
y
/
c
x/c
initial shape
optimal shape
Figure 5.1: Initial and optimal shapes for maximum performance at Re = 10
3
and
= 4
o
.
5.3 Optimization of cross-section for a rectangu-
lar wing planform
The optimal airfoil obtained form the previous computations is used as an initial
guess to optimize the cross-section of the wing. The optimal wing shape obtained
is shown along with the initial shape in Figure 5.7. The performance increases from
2.875 to 3.382; an increase of 17.65% over the initial shape. The optimal shape
34 5.3. Optimization of cross-section for a rectangular wing planform
wing root
wing tip
wing root
wing tip
leading edge
trailing edge
leading edge
trailing edge
Figure 5.2: Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: pressure distribution
on the Upper surface of (a) initial shape (b) optimal shape.
-1.5
-1
-0.5
0
0.5
1
1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
C
p
x/c
initial shape
optimal shape
-1.5
-1
-0.5
0
0.5
1
1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
C
p
x/c
initial shape
optimal shape
-2
-1.5
-1
-0.5
0
0.5
1
1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
C
p
x/c
initial shape
optimal shape
b)
c)
a)
Figure 5.3: Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: C
p
distribution on the
surface at (a) root, (b) 54.5% span and (c) 97.3% span.
5.3. Optimization of cross-section for a rectangular wing planform 35
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
y
/
c
x/c
initial shape
optimal shape
Figure 5.4: Initial and optimal shapes for maximum performance at Re = 10
3
and
= 4
o
.
wing root
wing tip
wing root
wing tip
leading edge
trailing edge
leading edge
trailing edge
Figure 5.5: Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: pressure distribution
on the Upper surface of (a) initial shape (b) optimal shape.
obtained is shown along with the initial shape in Figure 5.7.The pressure contours
on the upper surface of initial and the optimal wing is shown in the Figure 5.8.Figure
5.9 shows C
p
distribution on the surface of the optimal wing at dierent span-wise
locations. The optimal shape has a reduced bulge on the leading edge which leads
to lower lift and drag, but the decrease in drag is relatively more than that in the
36 5.4. Full-wing optimization
-2
-1.5
-1
-0.5
0
0.5
1
1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
C
p
x/c
initial shape
optimal shape
-2
-1.5
-1
-0.5
0
0.5
1
1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
C
p
x/c
initial shape
optimal shape
-2
-1.5
-1
-0.5
0
0.5
1
1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
C
p
x/c
initial shape
optimal shape
b)
c)
a)
Figure 5.6: Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: C
p
distribution on the
surface at (a) root, (b) 54.5% span and (c) 97.3% span.
lift. The bulge on the leading edge also reduces along the span and attens out at
the wing tip to facilitate wing tip vortices.
Figure 5.7: Initial and optimal shapes for maximum performance at Re = 10
3
and
= 4
o
.
5.4 Full-wing optimization
The optimal wing cross-section is used as an initial guess for full-wing optimization.
The nal optimal shape obtained is shown along with the initial shape in Figure 5.10.
5.4. Full-wing optimization 37
wing root
wing tip
wing root
wing tip
leading edge
trailing edge
leading edge
trailing edge
Figure 5.8: Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: pressure distribution
on the Upper surface of (a) initial shape (b) optimal shape.
-2
-1.5
-1
-0.5
0
0.5
1
1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
C
p
x/c
initial shape
optimal shape
-2
-1.5
-1
-0.5
0
0.5
1
1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
C
p
x/c
initial shape
optimal shape
-2
-1.5
-1
-0.5
0
0.5
1
1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
C
p
x/c
initial shape
optimal shape
b)
c)
a)
Figure 5.9: Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: C
p
distribution on the
surface at (a) root, (b) 54.5% span and (c) 97.3% span.
38 5.4. Full-wing optimization
The performance increases further from 3.382 to 3.670; an increase of 8.48% over
the initial shape. The optimal shape obtained is shown along with the initial shape
in Figure 5.10.The pressure contours on the upper surface of initial and the optimal
wing is shown in the Figure 5.11.Figure 5.12 shows C
p
distribution on the surface
of the optimal wing at dierent span-wise locations. There is a negligible change
in the cross-section of the nal shape but it has a curved leading edge resulting in
increased suction pressure over the upper surface and consequently increased lift.
Figure 5.10: Initial and optimal shapes for maximum performance at Re = 10
3
and
= 4
o
.
5.4. Full-wing optimization 39
wing root
wing tip
wing root
wing tip
leading edge
trailing edge
leading edge
trailing edge
Figure 5.11: Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: pressure distribution
on the Upper surface of (a) initial shape (b) optimal shape.
-2
-1.5
-1
-0.5
0
0.5
1
1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
C
p
x/c
initial shape
optimal shape
-2
-1.5
-1
-0.5
0
0.5
1
1.5
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
C
p
x/c
initial shape
optimal shape
-1.5
-1
-0.5
0
0.5
1
1.5
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
C
p
x/c
initial shape
optimal shape
b)
c)
a)
Figure 5.12: Minimization of C
D
/C
L
at Re = 10
3
and = 4
o
: C
p
distribution on
the surface at (a) root, (b) 54.5% span and (c) 97.3% span.
Chapter 6
Conclusions
A continuous adjoint based method is used for shape optimization in 3D. A stabilized
nite element method based on streamline upwind Petrov/Galerkin (SUPG) and
pressure stabilizing Petrov/Galerkin (PSPG) stabilization technique is employed
to solve both, the ow and the adjoint equation. The limited memory Broyden-
Fletcher-Goldfarb-Shanno (L-BFGS) algorithm is used to minimize the objective
function.
The method is used for designing optimal wing for maximum performance in viscous
incompressible steady ow at Reynolds number = 10
3
and = 4
o
. Objective func-
tion used is minimization of drag to lift ratio which is same as maximization of lift to
drag ratio. To prevent the optimizer form getting stuck in a local optima with lower
performance the number of variables for optimization process is increased progres-
sively. The process begins with optimization of airfoil in 3D using 10 variables; the
solution obtained is used as initial guess to optimize cross-section of the win using
60 variables. Finally, optimized cross-section is used as initial guess for full-wing
optimization using 70 variables. The nal optimal wing has 153.11% better per-
foramce compared to the rectangular wing having NACA 0012 as its cross-sectional
41
42 6.1. Future directions
airfoil.Since the optimizer is capable of nding only a nearest local minima, the de-
sign process is initiated with two dierent initial guesses. It is observed that optimal
solution corresponding to the airfoil obtained by 2D optimization of NACA 0012
aifoil as initial guess gives 44.67% better performance than the optimal solution ob-
tained from optimization of NACA 0012 airfoil directly in 3D. The airfoil used has
a performance of 2.65 in 2D but it is diminished to 2.31 in 3D due to induced drag
and tip votices. The optimal shape has a bulbous leading edge resulting in increased
suction pressure over the upper surface, therefore increasing the lift. The optimal
shape obtained after cross-section optimization has a reduced bulge on the leading
edge which leads to lower lift and drag, but the decrease in drag is relatively more
than that in the lift, thereby giving an increased performance of 17.65%. Finally,
the optimal wing obtained gives an improved performance by 8.48%. There is a
negligible change in the cross-section of the nal shape but it has a curved leading
edge resulting in increased suction pressure over the upper surface and consequently
increased lift.
6.1 Future directions
A few directions for future research motivated by the present work are presented
here.
Shape optimization with more number of control variables in cross-section can
be carried out to investigate wings with corrugations.
Extension of the present method to unsteady ows.
A few other desired objective functions can be explored. Example can be max-
imization of performance and minimization of drag as the objective function
used in the present study gives a solution corresponding to high lift but drag
also increases signicantly.
6.1. Future directions 43
Turbulence modelling can be incorporated in the present method.
Shape optimization for wings with geometric constraint can be done.
Bibliography
[1] M.J. Lighthill. A new method of two-dimensional aerodynamic design. Aero-
nautical Research Councils Reports and Memoranda, Number 2112, 1945.
[2] G.B.McFadden. An articial viscosity method for design of supercriticial wings.
PhD dissertation, NewYork University, 1979.
[3] R.M. Hicks and P.A. Henne. Wing design by numerical optimization. Journal
of Aircraft, 15:407412, 1978.
[4] C.B. Allen A.M. Morris and T.C.S. Rendall. Cfd-based optimization of aerofoils
using radial basis functions for domain element parameterization and mesh
deformation. International Journal for Numerical Methods in Fluids, 58:827
860, 2008.
[5] C.B. Allen A.M. Morris and T.C.S. Rendall. High-delity aerodynamic shape
optimization of modern transport wing using ecient hierarchical parametriza-
tion. International Journal for Numerical Methods in Fluids, 63:297312, 2009.
[6] A. Jameson. Aerodynamic design via control theory. Journal of Scientic
Computing, 59:117128, 1988.
[7] A. Jameson. Computational aerodynamics for aircraft design. Science, 245:361
371, 1989.
45
46 BIBLIOGRAPHY
[8] A. Jameson. Automatic design of transonic airfoils to reduce the shock induced
pressure drag. In Proceedings of the 31st Israel Annual Conference on Aviation
and Aeronautics, Tel Aviv, Israel, 1990.
[9] A. Jameson. Computational methods for aerodynamic design. In 14th Interna-
tional Conference on Numerical Methods in Fluid Dynamics, Bangalore, India,
1994.
[10] A. Jameson. Optimum aerodynamic design using cfd and control theory. AIAA
Paper 95-1729, 1995.
[11] A. Jameson Juan J. Alonso James J. Reuther L. Martinelli and J.C. Vassberg.
Aerodynamic shape optimization techniques based on control theory. AIAA
Paper 98-2538, 1998.
[12] Joel Brezillon and Richard P. Dwight. Aerodynamic shape optimization using
the discrete adjoint of navier-stokes equations: Applications towards complex
3d congurations. In Proceedings of the CEAS/KATnet II Conference on Key
Aerodynamic Technologies, 2009.
[13] Timothy M. Leung and David W. Zingg. Single and multi-point aerodynamic
shape optimization using a parallel newton-krylov approach. In AIAA Compu-
tational Fluid Dynamics, San Antonio, Texas, 2009.
[14] O. Soto and R. Lohner. Cfd shape optimization using an incomplete-gradient
adjoint formulation. International Journal for Numerical Methods in Engineer-
ing, 51:735753, 2001.
[15] Takayasu Kumano Shinkyu Jeong Shigeru Obayashi Yasushi Ito Keita
Hatanaka Hiroyuki Morino. Multidisciplinary design optimization of wing with
BIBLIOGRAPHY 47
nacelle and pylon. In European Conference on Computational Fluid Dynamics,
2006.
[16] Tad McGeer. Wing design for minimum drag with practical constraints. J.
Aircraft, 21:879886, 1984.
[17] Sean Wakayama and Ilan Kroo. Subsonic wing planform design using multi-
disciplinary optimization. Journal of Aircraft, 32:746753, 1995.
[18] Shigeru Obayashi. Multidisciplinary design optimization of aircraft wing plan-
form based on evolutionary algorithms. In Proceedings of the 1998 IEEE in-
ternational conference on systems man and cybernetics, La Jolla, California,
1998.
[19] K. Leoviriyakit and A. Jameson. Aero-structural wing planform optimization.
In AIAA paper 2004-0029, 2004.
[20] K. Leoviriyakit and A. Jameson. Aerodynamic shape optimization of wings
including planform variations. In AIAA paper 2003-0210, 2003.
[21] D.N. Srinath and S. Mittal. A stabilized nite element method for shape op-
timization in low reynolds number ows. International Journal for Numerical
Methods in Fluids, 54:14511471, 2007.
[22] D.N. Srinath and S. Mittal. Optimal airfoil shapes for low reynolds number
ows. International Journal for Numerical Methods in Fluids, 61:355381, 2008.
[23] D.N. Srinath, S. Mittal, and V. Manek. Multi-point shape optimization of
airfoils at low reynolds numbers. Computer Modelling in Engineering and Sci-
ences, 51:169190, 2009.
[24] D.N. Srinath and S. Mittal. An adjoint method for shape optimization in un-
steady viscous ows. Journal of Computational Physics, 229:19942008, 2010.
48 BIBLIOGRAPHY
[25] T.E. Tezduyar, S. Mittal, S.E. Ray, and R. Shih. Incompressible ow com-
putations with stabilized bilinear and linear equal-order-interpolation velocity-
pressure elements. Computer methods in applied mechanics and engineering,
95:221242, 1992.
[26] D. Landman and C.P. Britcher. Experimental geometry optimization tech-
niques for multi-element airfoils. Journal of Aircraft, 37:707713, 2000.
[27] X. Wang, M. Damodaran, and S. L. Lee. Inverse transonic airfoil design using
parallel simulated annealing and computational uid dynamics. AIAA Journal,
40:791794, 2002.
[28] S. Obyashi. Aerodynamic inverse optimization with genetic algorithms. Journal
of Engineering and Applied Science, pages 421425, 1996.
[29] J.L. Lions. Optimal control of systems governed by partial dierential equations.
Springer-Verlag, Berlin, 1971.
[30] Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer Series
in Operations Research, 1999.
[31] R.H. Byrd, P. Lu, J. Nocedal, and C. Zhu. A limited memory algorithm
for bound constrained optimization. SIAM Journal of Scientic Computing,
16:11901208, 1995.
[32] Patrice Castonguay and Shiv K. Nadarajah. Eect of shape parameterization
on aerodynamic shape optimisation. In 45th AIAA Aerospace Sciences Meeting
and Exhibit, Reno, Nevada, 2007.
[33] Les Piegl and Wayne Tiller. The NURBS Book. Springer, 1997.
[34] T.E. Tezduyar, M. Behr, S. Mittal, and A.A. Johnson. Computation of un-
steady incompressible ows with the nite element methods Space-Time
BIBLIOGRAPHY 49
Formulations, iterative strategies and massively parallel implementations. In
P. Smolinski, W.K. Liu, G. Hulbert, and K. Tamma, editors, New Methods in
Transient Analysis, AMD-Vol.143, pages 724, New York, 1992. ASME.

Вам также может понравиться