Simulation of Aircraft Aeroelasticity

Thesis submitted in accordance with the requirements of

the University of Liverpool for the degree of Doctor in Philosophy
by
Adam Swift
M.Eng.[Mechanical], Liverpool 2007
April 2011

Declaration
I confirm that the thesis is my own work, that I have not presented anyone elses
work as my own and that full and appropriate acknowledgement has been given
where reference has been made to the work of others.

...............................................................
Adam Swift
April 2011

ii

Abstract
Aeroelastic phenomena such as flutter can have a detrimental effect on aircraft
performance and can lead to severe damage or destruction. Buffet leads to a reduced fatigue life and therefore higher operating costs and a limited performance
envelope. As such the simulation of these aeroelastic phenomena is of utmost
importance. Computational aeroelasticity couples computational fluid dynamics
and computational structural dynamics solvers through the use of a transformation method. There have been interesting developments over the years towards
more efficient methods for predicting the flutter boundaries based upon the stability of the system of equations.
This thesis investigates the influence of transformation methods on the flutter
boundary predition and considers the simulation of shock-induced buffet of a
transport wing. This involves testing a number of transformation methods for
their effect on flutter boundaries for two test cases and verifying the flow solver for
shock-induced buffet over an aerofoil. This will be followed by static aeroelastic
calculations of an aeroelastic wing.
It is shown that the transformation methods have a significant effect on the
predicted flutter boundary. Multiple transformation methods should be used to
build confidence in the results obtained, and extrapolation should be avoided.
CFD predictions are verified for buffet calculations and the mechanism behind
shock-oscillation of the BGK No. 1 aerofoil is investigated. The use of steady
calculations to assess if a case may be unsteady is considered. Finally the static
aeroelastic response of the ARW-2 wing is calculated and compared against experimental results.

iii

iv

Acknowledgement
I would like to thank my supervisors Professor Ken Badcock and Professor George
Barakos and all the members of the CFD lab, both past and present, at the
University of Liverpool. It has been a privilege to work with you all.
I would like to thank the members of the R&T group at BAE Systems Warton
for their support and encouragement that helped make my annual placements so
productive. I am especially grateful to Mark Lucking for his guidance and support
during these placements.
I would like to thank my family and friends for their their patience and support. I am especially grateful to Maryam for her unwavering love and patience.
I am grateful for the financial support of the Engineering and Physical Sciences Research Council (EPSRC) and BAE SYSTEMS through a case award. I
would also like to extend my gratitude to the 2nd UK Applied Aerodynamics
Consortium for providing computing time on HPCx and HECTOR.

vi

Publications
Computational Study of the BGK No. 1 Aerofoil Buffet, A. Swift and K.J.
Badcock, submitted to the Aeronautical Journal.
Investigation into the Effect of Transformation Methods on Flutter Boundary
Predictions, A. Swift and K.J. Badcock, submitted to the Aeronautical
Journal.
Transonic Aeroelastic Simulation for Envelope Searches and Uncertainty Analysis, K.J. Badcock, S. Timme, S. Marques, H. Khodaparast, M. Pradina,
J.E. Mottershead, A. Swift and A. Da Ronch, Progress in Aerospace Sciences (online), June 2011.
Framework for Establishing the Limits of Tabular Aerodynamic Models for
Flight Dynamics Analysis, M. Ghoreyshi, K. Badcock, A. Da Ronch, S.
Marques, A. Swift and N. Ames, Journal of Aircraft, 48(1), 2011.
Inter-grid Transfer Influence on Transonic Flutter Predictions, A. Swift and
K. Badcock, AIAA-2010-3049, April 2010.
Framework for Establishing the Limits of Tabular Aerodynamic Models for
Flight Dynamics Analysis, M. Ghoreyshi, K. Badcock, A. Da Ronch, S.
Marques, A. Swift and N. Ames, AIAA-2009-6273, August 2009.

vii

viii

Contents
Declaration

Abstract

iii

Acknowledgement

Publications

vii

Table of Contents

xii

List of Tables

xiii

List of Figures

xvii

Nomenclature

xxiv

1 Introduction
1.1 Aeroelasticity . . . . . . . . .
1.2 Prediction of Flutter . . . . .
1.3 Computational Aeroelasticity
1.4 Thesis Organisation . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

2 Formulation
2.1 Navier-Stokes Equations . . . . . . . . . . . .
2.2 Turbulence . . . . . . . . . . . . . . . . . . . .
2.3 The Closure Problem . . . . . . . . . . . . . .
2.3.1 Reynolds Averaging . . . . . . . . . . .
2.3.2 Favre Mass Averaging for Compressible
2.3.3 Boussinesqs Approximation . . . . . .
2.3.4 Vector Form . . . . . . . . . . . . . . .
2.3.5 Non-dimensional Form . . . . . . . . .
2.3.6 Curvilinear Form . . . . . . . . . . . .
2.4 Turbulence Modelling . . . . . . . . . . . . . .
ix

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

. . . .
. . . .
. . . .
. . . .
Flow .
. . . .
. . . .
. . . .
. . . .
. . . .

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.

1
1
7
8
10

.
.
.
.
.
.
.
.
.
.

13
13
15
18
19
19
21
23
23
24
24

2.5
2.6

2.7
2.8

2.4.1 Menter SST k- Model . . . . . .

Flow Solver . . . . . . . . . . . . . . . .
Mesh Movement . . . . . . . . . . . . . .
2.6.1 Spring Analogy of Block Vertices
2.6.2 TFI of Block Edges . . . . . . . .
2.6.3 TFI of Block Faces . . . . . . . .
2.6.4 TFI of Block Volumes . . . . . .
Structural Solver . . . . . . . . . . . . .
Schur Code . . . . . . . . . . . . . . . .

3 Transformation Methods
3.1 Introduction . . . . . . . . . . . . . . . .
3.1.1 The Transformation Problem . .
3.1.2 Notation . . . . . . . . . . . . . .
3.1.3 Requirements . . . . . . . . . . .
3.1.4 Review . . . . . . . . . . . . . . .
3.2 Types of Structural Models . . . . . . .
3.3 Existing Transformation Methods . . . .
3.3.1 Multiquadric-Biharmonic . . . . .
3.3.2 Infinite Plate Spline . . . . . . . .
3.3.3 Thin-Plate Splines . . . . . . . .
3.3.4 Inverse Isoparametric Mapping .
3.3.5 Finite Plate Spline . . . . . . . .
3.3.6 Non-Uniform Rational B-Splines .
3.3.7 Boundary Element Method . . .
3.3.8 Radial Basis Functions . . . . . .
3.3.9 Constant Volume Tetrahedron . .
3.4 Rigid Section Method . . . . . . . . . . .

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

4 Evaluation of Transformation Methods

4.1 Introduction . . . . . . . . . . . . . . . . . . . .
4.2 Test Cases . . . . . . . . . . . . . . . . . . . . .
4.2.1 Goland Wing . . . . . . . . . . . . . . .
4.2.2 MDO Wing . . . . . . . . . . . . . . . .
4.3 Evaluation of Transfer - Shape . . . . . . . . . .
4.3.1 Case 1 - In-plane . . . . . . . . . . . . .
4.3.2 Case 2 - Goland Wing In-plane . . . . .
4.3.3 Case 3 - Goland Wing Out-of-plane . . .
4.3.4 Case 4 - Goland Wing Beam Stick Model
4.3.5 Case 5 - MDO Wing . . . . . . . . . . .
x

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

25
27
27
27
28
29
29
30
31

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

35
35
35
36
37
40
41
42
42
44
46
47
48
48
49
50
52
55

.
.
.
.
.
.
.
.
.
.

57
57
57
57
58
61
62
64
65
67
67

4.3.6 Summary . .
4.4 Flutter Evaluation .
4.4.1 Goland Wing
4.4.2 MDO Wing .
4.5 Summary . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

5 Aerofoil Buffet
5.1 Introduction . . . . . . . . . . . . . . . .
5.2 Test Case . . . . . . . . . . . . . . . . .
5.2.1 Other Aerofoil Buffet Cases . . .
5.3 BGK No. 1 Aerofoil Results . . . . . . .
5.3.1 Grids . . . . . . . . . . . . . . . .
5.3.2 Probe Locations . . . . . . . . . .
5.3.3 Reduced Frequency Calculation .
5.3.4 Convergence Studies . . . . . . .
5.3.5 Pseudo Steps Study . . . . . . . .
5.3.6 Time Step Study . . . . . . . . .
5.3.7 Grid Refinement Study . . . . . .
5.3.8 Buffet Boundary Estimate . . . .
5.3.9 Mach 0.71 Angle of Attack Sweep
5.4 Summary . . . . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

6 Static Wing Aeroelasticity

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . .
6.2 Previous Work . . . . . . . . . . . . . . . . . . . .
6.3 Test Case Setup . . . . . . . . . . . . . . . . . . . .
6.3.1 CFD Grid . . . . . . . . . . . . . . . . . . .
6.3.2 Structural Model . . . . . . . . . . . . . . .
6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . .
6.4.1 Case 1: Mach Number 0.80, Angle of Attack
6.4.2 Case 2: Mach Number 0.85, Angle of Attack
6.4.3 Case 3: Mach Number 0.92, Angle of Attack
6.4.4 Case 4: Mach Number 0.92, Angle of Attack
6.4.5 Buffet Search . . . . . . . . . . . . . . . . .
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

. . .
. . .
. . .
. . .
. . .
. . .
0.0
1.0
0.0
2.0
. . .
. . .

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

69
69
69
73
75

.
.
.
.
.
.
.
.
.
.
.
.
.
.

77
77
77
82
86
86
88
88
90
90
92
94
94
97
110

.
.
.
.
.
.
.
.
.
.
.
.

117
117
119
121
121
121
124
125
126
135
135
135
137

7 Conclusions
139
7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
xi

A Infinite Plate Spline

153

B Inverse Isoparametric Mapping

155

C Radial Basis Functions

157

xii

List of Tables
2.1 Summary of Solution Method Costs (from Ref. [1]) . . . . . . . .

19

3.1 Basis Functions (kxk) (taken from [2]) . . . . . . . . . . . . . . .

51

4.1 Flutter altitude for Goland wing (all values in feet) . . . . . . . .

4.2 Flutter altitude for MDO wing (all values in meters) . . . . . . .

71
74

5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
5.9

Xiao et al. [3]: Flow conditions . . . . . . . . .

Grid Sizes . . . . . . . . . . . . . . . . . . . . .
Pressure taps and probe locations . . . . . . . .
Results for the pseudo steps study . . . . . . . .
Results for the time step study . . . . . . . . .
Results for grid refinement study . . . . . . . .
Coefficient of lift comparisons for angle of attack
Reduced frequency results . . . . . . . . . . . .
Reduced frequency results . . . . . . . . . . . .

. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
1.396
. . . .
. . . .

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

. 80
. 86
. 89
. 90
. 92
. 94
. 99
. 107
. 110

6.1 Comparison of the structural model frequencies . . . . . . . . . . 121

6.2 ARW-2 test conditions . . . . . . . . . . . . . . . . . . . . . . . . 125

xiii

xiv

List of Figures
1.1
1.2
1.3
1.4

F-16 ventral fin failure (from ref. [4]) . . . . . . . . . . . . . . . .

Vortex breakdown over the F/A-18 (from ref. [5]) . . . . . . . . .
Buffeting criteria (from ref. [6]) . . . . . . . . . . . . . . . . . . .
Comparison between the flutter boundaries predicted using the
linear method and the wing-only configuration for the weakened
structural model. (from ref. [7]) . . . . . . . . . . . . . . . . . . .

4
5
6

2.1 False-colour image of the far-field of a submerged turbulent jet

using laser induced fluorescence (LIF). (from ref. [8]) . . . . . . . 16
2.2 Energy spectrum for a turbulent flow, log-log scales. (from ref. [9]) 16
2.3 Faces of the block volume (from ref. [10]) . . . . . . . . . . . . . . 30
3.1
3.2
3.3
3.4
3.5
3.6

Comparison of typical CFD/CSD grids .

Structural models . . . . . . . . . . . . .
A circle of points rigidly rotated by a bar
NURBS approach [11] . . . . . . . . . .
Boundary element method [10] . . . . . .
Constant volume tetrahedron [10] . . . .

4.1
4.2
4.3
4.4
4.5

. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
using IPS (from ref. [10])
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .

Goland wing geometry and CFD grid . . . . . . . . . . . . . . . .

Goland wing structural models (wing planform in grey) . . . . . .
MDO wing geometry and CFD grid . . . . . . . . . . . . . . . . .
MDO wing structural models . . . . . . . . . . . . . . . . . . . .
Case 1: Section at 98% span for the third mode shape, rigidly
translated and rotated back to the original orientation. . . . . . .
4.6 Case 2: Section at 98% span for the third mode shape, rigidly
translated and rotated back to the original orientation. . . . . . .
4.7 Case 3: Section at 98% span for the third mode shape, rigidly
translated and rotated back to the original orientation. . . . . . .
4.8 Case 3: Sections at 98% of the span through the mode 3 mode
shape for IIM comparison for cases 2 and 3, rigidly translated and
rotated back to the original orientation. . . . . . . . . . . . . . . .
xv

36
43
46
49
49
53
59
60
61
61
63
64
65

66

4.9
4.10
4.11
4.12
4.13
4.14
4.15
4.16
4.17

5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
5.9
5.10
5.11
5.12
5.13
5.14
5.15
5.16
5.17

Case 3: Sections at 98% of the span for the refined extended plate
model, rigidly translated and rotated back to the original orientation.
Case 4: Sections at 98% span through the mode 3 mode shape
rotated and translated back to the original orientation. . . . . . .
Case 5: Vertical displacement of sections at 98% span for the mode
5 mode shape. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Goland wing mode tracking for all modes using CVT . . . . . . .
Goland wing mode tracking for mode 1 for the different structural
models (real part) . . . . . . . . . . . . . . . . . . . . . . . . . . .
Goland wing mode tracking for mode 1 for the different structural
models (real part) . . . . . . . . . . . . . . . . . . . . . . . . . . .
MDO wing mode tracking for all modes using CVT . . . . . . . .
MDO wing mode tracking for mode 1 for the different structural
models (real part) . . . . . . . . . . . . . . . . . . . . . . . . . . .
MDO wing mode tracking for mode 1 for the different structural
models (real part) . . . . . . . . . . . . . . . . . . . . . . . . . . .

66
67
68
70
71
72
73
74
74

BGK No. 1 buffet boundary with the experiment locations (circles) 78

BGK No. 1 supercritical aerofoil . . . . . . . . . . . . . . . . . . . 79
Model of self-sustained shock oscillation for the BGK No. 1 aerofoil
(from ref. [12]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
18% thick circular-arc aerofoil shock oscillation domains (from ref.
[13]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
BGK No. 1 grid details . . . . . . . . . . . . . . . . . . . . . . . . 87
y + distribution around the BGK No. 1 aerofoil at = 6.97 ,
M=0.71 and Re=20 106 . . . . . . . . . . . . . . . . . . . . . . 88
BGK No. 1 pseudo steps study results . . . . . . . . . . . . . . . 91
BGK No. 1 time step study results . . . . . . . . . . . . . . . . . 93
BGK No. 1 grid refinement study results . . . . . . . . . . . . . . 95
Buffet boundary estimate . . . . . . . . . . . . . . . . . . . . . . 96
Comparison of Mach 0.71 angle of attack sweep for steady calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Lift curve for the Mach 0.71 angle of attack sweep . . . . . . . . . 99
BGK No. 1 angle of attack 1.396 results . . . . . . . . . . . . . . 100
BGK No. 1 angle of attack 3.017 results . . . . . . . . . . . . . . 100
BGK No. 1 angle of attack 4.905 results . . . . . . . . . . . . . . 102
Pressure wave propagation for angle of attack 4.905 . . . . . . . 103
Unsteady shock movement during one period for angle of attack
4.905 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
xvi

5.18 Mach number contours at different instants in time for one period
of oscillation for angle of attack 4.905 , M = 0.095 . . . . . . .
5.19 Streamlines at different instants in time for one period of oscillation
for angle of attack 4.905 . . . . . . . . . . . . . . . . . . . . . . .
5.20 Cross-covariance and wave speed for angle of attack 4.905 . . . .
5.21 BGK No. 1 angle of attack 6.97 results . . . . . . . . . . . . . .
5.22 Pressure wave propagation . . . . . . . . . . . . . . . . . . . . . .
5.23 Unsteady shock movement during one period for angle of attack
6.97 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.24 Mach number contours at different instants in time for one period
of oscillation for angle of attack 6.97 , M = 0.095 . . . . . . . .
5.25 Streamlines at different instants in time for one period of oscillation
for angle of attack 6.97 . . . . . . . . . . . . . . . . . . . . . . .
5.26 Cross-covariance and wave speed for angle of attack 6.97 . . . . .
5.27 BGK No. 1 angle of attack 9.0 results . . . . . . . . . . . . . . .

105
106
108
109
110
111
112
113
114
115

6.1 ARW-2 (from ref. [14]) . . . . . . . . . . . . . . . . . . . . . . . . 118

6.2 Maximum PSD peak response from front spar wing-tip accelerometer (from ref. [15]) . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.3 ARW-2 grid details . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.4 Comparison of the mode shapes. Grey - Undeformed shell model,
Red - Report model mode shapes, Green - Shell model mode shapes123
6.5 Comparison of the mode shapes with the beam model. Grey Undeformed shell model, Red - Report model mode shapes, Green
- Shell model mode shapes, Pale Blue - Undeformed CFD mesh,
Blue - Beam model mode shapes on the CFD mesh . . . . . . . . 124
6.6 Case 1: Sectional results . . . . . . . . . . . . . . . . . . . . . . . 127
6.7 Case 1: Mach number contours, M = 0.045 . . . . . . . . . . . . 128
6.8 Case 1: Streamlines . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.9 Case 1: Spar deflections . . . . . . . . . . . . . . . . . . . . . . . 130
6.10 Case 1: Wing twist . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.11 Case 2: Sectional results . . . . . . . . . . . . . . . . . . . . . . . 131
6.12 Case 2: Mach number contours, M = 0.045 . . . . . . . . . . . . 132
6.13 Case 2: Streamlines . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.14 Case 2: Spar deflections . . . . . . . . . . . . . . . . . . . . . . . 134
6.15 Case 2: Wing twist . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.16 Case 3: Convergence . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.17 Case 4: Convergence . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.18 Buffet search results . . . . . . . . . . . . . . . . . . . . . . . . . 136

xvii

xviii

Nomenclature
Acronyms

Definition

AGARD

The Advisory Group for Aerospace Research and Development.

AMB

Aeroelastic Multi-Block solver.

ARW-2

Aeroelastic Research Wing 2.

BEM

Boundary Element Method.

BGK

Bauer-Garabedian-Korn aerofoil.

BILU

Block Incomplete Lower-Upper factorisation.

BSM

Beam Stick Model.

CAE

Computational Aeroelasticity.

CFD

Computational Fluid Dynamics.

CFL

Courant-Friedrichs-Lewy number.

CPU

Central Processing Unit.

CSD

Computational Structural Dynamics.

CVT

Constant Volume Tetrahedron.

DAST

Drones for Aerodynamic and Structural Testing.

DES

Detached Eddy Simulation.

DNS

Direct Numerical Simulation.

DOF

Degrees of Freedom.

FFT

Fast Fourier Transform.

FPS

Finite Plate Splines.

GVT

Ground Vibration Test.

HIRENASD

High Reynolds Number Aero-Structural Dynamics.

xix

IIM

Inverse Isoparametric Mapping.

IMQ

Inverse Multiquadrics.

IPS

Infinite Plate Splines.

LANTRIN

Low Altitude Navigation and Targeting Infrared for Night.

LCO

Limit Cycle Oscillation.

LES

Large Eddy Simulation.

LIF

Laser Induced Fluorescence.

MDO

Multi-Disciplinary Optimisation.

MQ

Multiquadrics.

MUSCL

Monotone Upstream-centred Schemes for Conservation

Laws.

NAE

National Aeronautical Establishment (Canada).

NASA

National Aeronautics and Space Administration.

NUBS

Non-Uniform B-Spline.

NURBS

Non-Uniform Rational Basis Spline.

ONERA

Office National dEtudes

et de Recherches Aerospatiales
(French aerospace research centre).

RANS

Reynolds Averaged Navier-Stokes.

RBF

Radial Basis Functions.

SDM

Standard Dynamics Model.

TDT

Transonic Dynamics Tunnel.

TFI

Transfinite Interpolation.

TPS

Thin Plate Splines.

URANS

Unsteady Reynolds Averaged Navier-Stokes.

USA

United States of America.

Symbols

Definition

qx , qy , qz

Heat fluxes.

Rate of volumetric heat addition per unit mass.

xx

, u ,

Length, velocity and time scales.

dx

Displacements.

fi

Blending functions.

Forces.

Time dependent external force vector.

F, G, H

Flux vectors in the x, y amd z directions.

Complex eigenvector.

Vector of residuals.

Solution vector of conserved variables.

Vector of unknowns.

Eigenvectors.

Coordinates.

x, x,
x

Time dependent displacement, velocity and acceleration

vectors.

Mean part of a flow variable.

Mass averaged flow variable.

Viscous damping matrix.

Chord.

CN

RMS values of the normal-force fluctuations.

Ck

Kolmogorov constant.

CL

Lift coefficient.

cp

Specific heat at constant pressure.

cv

Specific heat at constant volume.

Total energy.

Specific internal energy.

E()

Spectral distribution of energy.

Arbitrary flow variable.

Frequency.
xxi

Fluctuating part of a flow variable.

Fluctuating part of a flow variable.

fi

Body forces.

Total enthalpy.

Specific internal enthalpy.

Identity matrix.

Jacobian.

Stiffness matrix.

Reduced frequency.

Specific turbulent kinetic energy.

kij

Edge spring stiffness.

kx , ky , kz

Norm biasing coefficients.

Length.

Mass matrix.

M , C , K

Modal mass, damping and stiffness matrices.

Number of degrees of freedom.

Ni (, )

Shape functions.

Pressure.

p(x)

Linear polynomial.

Pr

Prandtl number.

P rT

Turbulent Prandtl number.

Rotation matrix.

Specific gas constant.

r()

Blending function.

r2

Non-zero parameter for MQ.

Re

Reynolds number.

ReT

Turbulent Reynolds number.

Temperature.
xxii

Transformation matrix.

Time.

T0

Reference temperature = 288.16K.

Tp

Period of shock oscillation.

Velocity.

u, v, w

Velocity components in the x, y, z directions.

ud

Speed of downstream pressure wave propagation.

uu

Speed of upstream pressure wave propagation.

Virtual work.

x, y, z

Cartesian coordinates.

xs

Time-mean shock position.

s(x)

Interpolant.

Greek Symbols

Definition

, ,

CVT constants.

Coefficients.

, ,

Modal coordinate.

, u ,

Kolmogorov scales of length, velocity and time.

Thermal conductivity.

Wavenumber.

Eigenvalues.

Dynamic viscosity.

Schur solver parameter (altitude).

Reference viscosity = 1.7894 105 .

Turbulent eddy viscosity.

Kinematic viscosity.

Modal matrix.

Displacement mapping to virtual surface.

, ,

Blending functions.
xxiii

Density.

Support radius.

ij

Shear Stresses.

ijR

Reynolds stresses.

Rate of dissipation.

, ,

Local co-ordinate system.

Rotation about the beam.

Subscripts

Definition

Freestream value.

Fluid.

ff

Fluid-fluid.

fs

Fluid-structure.

Inviscid.

i, j, k

Indices=1,2,3,. . . .

Laminar.

Structure.

sf

Structure-fluid.

ss

Structure-structure.

Turbulent.

Viscous.

Superscripts

Definition

Mapped quantity.

Indicates structural node is a member of that element.

Current time-step.

n+1

Next time-step.

Reynolds stress.

Dimensional variable.

xxiv

Chapter 1
Introduction
1.1

Aeroelasticity

Aeroelasticity involves the interaction of aerodynamic, elastic and inertial forces.

The subject can be traced back to before the first powered flight of the Wright
Brothers. The Langley Monoplane that flew a few days before the Wright Brothers
famous flight is suspected to have crashed due to aeroelastic divergence of the
rear wing, which was insufficiently torsionally stiff [16]. The Wright Brothers
used the flexibility in the wings of their aircraft to control roll and this concept
was revisited for the NASA X-53 Active Aeroelastic Wing in 2004 that was based
on the F/A-18 fighter jet. A brief history of the first fifty years of aeroelasticity
was given by Collar [17]. The main areas of interest to aeroelasticians are the
interaction of aerodynamic and elastic forces to form static aeroelasticity, and
the interaction of all three forces to form dynamic aeroelasticity. This can be
extended to include the effect of control systems, which gives rise to aeroservoelasticity, and the effect of heating (important for hypersonic applications) to give
rise to aerothermoelasticity. Aeroelasticity is not solely concerned with aircraft
(although the focus of this thesis), but effects the design of structures such as
bridges (Tacoma Narrows Bridge for example), powerplant cooling towers, Formula 1 racing cars, wind turbines, turbomachinary etc.
Static aeroelasticity is the coupling between the aerodynamic and elastic forces
and is time independent. Since an aircraft structure is flexible the aerodynamic
loads will deform the aircraft structure, which in turn will change the aerodynamic
loads acting on it. For static aeroelasticity the aerodynamic loads and the elastic
restoring forces will come to an equilibrium and the resulting deformed shape will
have an influence on the generated lift and drag. As such airframe manufacturers
put significant work into designing the jig (or manufactured) shape of an aircraft
so that at cruise the aircraft will deform into an optimised design shape. Avoiding
errors in the jig shape is very important due to the large impact the jig shape
1

has on the drag produced [18].

There are negative static aeroelastic effects that are of concern. These are
divergence and control surface reversal. If a wing is assumed to have a finite
torsional stiffness, a twist along the wing will be induced when an aerodynamic
load is applied. For high torsional stiffness or low airspeed (a small aerodynamic
load) this twist angle will be small. However as the airspeed increases, and
therefore the aerodynamic load, so will the twist angle. At a critical speed the
aerodynamic pitching moment exceeds the elastic restoring moment and the wing
is twisted beyond its ultimate strength, leading to structural failure. This is the
static aeroelastic phenomenon of divergence.
The aim of a control surface is to modify the lift or moment of the lifting
surface. If the lifting surface is flexible then control surface reversal can occur.
An example of this is if an aileron is deflected in order to achieve a certain roll
rate, this will augment the lift on the outboard portion of the wing. If the wing is
sufficiently stiff then this will be successful. However if the airspeed and therefore
the aerodynamic loads are large enough the twisting of the wing will lead to a
reduction of lift. Since this is opposite to the desired command the controls
have effectively reversed. Near this critical speed the effect of the control surface
and the twisting cancel leaving little control authority. Control surface reversal
effected the Supermarine Spitfire during the Second World War, which could
happen when the pilot tried to roll during a dive. The fighter had originally been
designed for a control surface reversal speed of 580 mph, which during a dive
could be exceeded. This resulted in the pilots rolling in the opposite direction
than intended. This was later fixed by stiffening the wings to give it a theoretical
control reversal speed of 825 mph.
Dynamic aeroelasticity is concerned with oscillatory and transient effects.
There are four main dynamic aeroelastic phenomenon, which are flutter, limit
cycle oscillation (LCO), control surface buzz and buffet.
Flutter is where two structural vibration modes couple in the aeroelastic system causing the structure to extract energy from the airflow, possibly leading
to the destruction of the aircraft. The first recorded flutter incident was on a
Handley Page O/400 twin engined biplane bomber in 1916. The mechanism was
the coupling of the fuselage torsion mode and an antisymmetric elevator mode.
Since the elevators were independently connected to the control stick the solution
was to connect the elevators together using a torque tube [16]. Flutter can occur
locally, for example panel flutter where an oscillating shockwave causes the structural panel below it to flutter. Control surface limit cycle oscillation leading to
wing flutter is believed to be responsible for the crash of a F-117 stealth fighter
at an airshow in Maryland, USA in 1997. This was found to be caused by four
2

out of five fasteners missing in the left wing, this reduced the torsional stiffness
of the elevator. The reduced torsional stiffness caused a limit cycle oscillation of
the elevator, which caused the wing to flutter [19]. A history of flutter can be
found in ref. [16].
Limit cycle oscillation is a type of flutter, which is a self sustaining limited
amplitude oscillation with the amplitude limited due to non-linearities. Both the
F-16 and F/A-18 have encountered limit cycle oscillations at high subsonic and
transonic speeds for certain store configurations with AIM-9 missiles at the wing
tips and heavy stores on the outboard pylons [20]. Limit cycle oscillations can
damage stores, reduce their fatigue life and targeting effectiveness [20].
Control surface buzz is similar to limit cycle oscillation in that the amplitude
of the oscillations are limited, however the mechanism behind it is different. It is
caused by the interaction of a shock wave, boundary layer and control surface rotation mode. There are two main types of control surface buzz, the first is caused
by buffet of the control surface by the separated flow due to shock-boundary layer
interaction upstream and the second type is due to shock wave oscillation on the
control surface itself with no separation involved.
Buffeting is defined as the structural response to the aerodynamic excitation
produced by separated flows. Buffet can be caused by massive separation over a
lifting surface (stall buffet), separation due to shock/boundary-layer interaction
(Mach number buffet) or vortex breakdown. Related to buffet is transonic buffet
(also refered to as shock buffet), this is an aerodynamic phenomenon in which a
shockwave is generated over a lifting surface that interacts with the boundarylayer causing partially or fully separated flow. This results in a self-sustained
oscillating flow field and occurs without structural motion, however the frequency
of shock oscillation is usually in the range of the low-frequency modes of an elastic
structure which would lead to buffeting.
One of the first records of buffet dates back to 21st July 1930 where a commercial aeroplane crashed at Meopham, Kent. In this case the aircraft flew into
a cloud and shortly afterwards came down in pieces, the subsequent investigation found that there was a strong updraft in the area. This caused the angle
of attack to increase sharply leading to massive separation over the wing. The
tail was situated in this region of highly unsteady separated flow and therefore
subjected to intense forced vibrations that ultimately caused the tail to fail [21].
This accident was an example of stall buffet. A more modern example was on the
American F-15 fighter jet caused by leading edge separation at angles of attack
of over 22 degrees, which excited the first torsion mode [22]. In 1998 the cost of
repairing and replacing F-15 vertical tails due to fatigue was estimated to be $5-6
million per year [4], in addition to the loss of flight readiness. This type of buffet
3

is caused by the separation over wings, the F-16 had severe damage to ventral
fins positioned at the underside of the rear fuselage caused by the upstream separation of the airflow over the LANTRIN targeting pods [4]. This is shown in the
Figure 1.1 and the extensive damage to the ventral can be seen.

(a) F-16 with ventral fins and LANTRIN pod

(b) Ventral fin failure

Figure 1.1: F-16 ventral fin failure (from ref. [4])

Shock-induced boundary-layer separation causes an aerodynamic unsteadiness. At certain conditions this shockwave can exhibit self-excited oscillations
along with separation and reattachment. This dynamic separation and reattachment excites the structure and in some cases can be severe enough to threaten the
structural integrity of the aircraft. An example of this type of buffet is the Aeroelastic Research Wing (ARW-2) which exhibits shock boundary-layer interaction
that leads to buffeting of the structure over a small region of Mach numbers [23].
During windtunnel tests, certain flow conditions could not be tested since the
buffeting was so severe that the safety of the wing was not guaranteed.
The final type of buffet is caused by vortex breakdown, which has generated a
large amount interest. This interest has been caused by the move to highly agile
aircraft capable of performing high angle of attack maneuvers. At high angles of
attack many modern aircraft feature high energy vortices generated by leading
edge extensions, wing, canards etc. and these vortices can become unstable and
burst. Once burst a wake of highly turbulent flow imparts unsteady loading to
the tail of the aircraft that excites the structural modes of the aircraft component.
This excitation is generally at low frequencies that match or are close to the first
bending and torsion modes of the tail. The high dynamic response impacts the
fatigue life and in rare cases threatens the structural safety of the component.
The most researched aircraft that suffers from this type of buffet is the F/A-18.
The vortices are generated by the leading edge extension and breakdown ahead
of the vertical tails at angles of attack of 25 and higher [24] (see Figure 1.2)
with the buffet loads on the vertical tail reaching 450g during test flights [25].
4

This type of buffet is very much an unsolved problem for modern fighter jets with
the EF-2000 type fighter [26, 27], F-22 [28, 29] and the F-35 [30] suffering from
different levels of buffet.

Figure 1.2: Vortex breakdown over the F/A-18 (from ref. [5])
Buffet effects all aircraft, both civil and military, however the causes and flight
regimes are different. Figure 1.3 gives a comparison of the flight envelope for the
buffet criteria for a civil transport aircraft and a military fighter jet. Buffet is
split into different levels of severity denoted as light, moderate and heavy. Civil
aircraft are designed to cruise below buffet onset and maneuver to at least 1.3g
without any buffeting [6]. It is allowed to enter the buffet region for short periods
during rare severe gusts, but it should not create any structural or control issues.
Military fighters fly much further below the buffet boundary, but for performance
reasons will frequently maneuver deep inside the buffet region. This can lead to
a significant reduction in the fatigue life of the aircraft [6].
Over the years many potential solutions to buffet have been proposed, the
simplest being to limit the flight envelope. On the F/A-18, streamline fences
were introduced on the upper surface of the leading edge extension to reduce the
buffet loads, which although significantly reducing the buffet loads also caused a
slight loss in lift. This solution was less effective at higher angles of attack [24].
Piezoelectric actuators have been fitted to the tail of aircraft to actively control
the tail. There are other solutions such as active control surfaces, sucking and
blowing to control vortices and finally stiffening the structure. There have been
hybrid approaches such as combining active control surfaces and piezoelectric
5

(a) Civil transport aircraft

(b) Military fighter aircraft

Figure 1.3: Buffeting criteria (from ref. [6])

actuators to control vertical tails [25], which resulted in a 60% reduction in the
buffeting for a 8% weight penalty.
Aircraft airworthiness regulations require that an aircraft is demonstrated to
be free from aeroelastic instabilities (especially flutter) under flight conditions
using a margin of safety of 15%. Flutter is an aeroelastic phenomenon that can
be prevented through proper design, such as ensuring the bending and torsion
mode frequencies remain separated [31] and decoupling modes. This should be
backed up by flutter analysis throughout the design process. This is achieved
using windtunnel models and computational analysis and, once a prototype has
been manufactured, ground vibration and flight tests [32]. The number of cases
that need to be analysed for a modern transport aircraft was estimated to be
between five to ten thousand cases [33]. Throughout the life of the aircraft any
changes to the structure that may impact the flutter boundary, require the flight
envelope to be demonstrated safe again. Military combat aircraft are typically
fitted with deployable stores both internally and externally with each combination
requiring the flight envelope to be cleared. An example is the F/A-18E that was
estimated to have 400,000 possible store combinations [34]. As such it is not
feasible to fly this number of flight conditions, so the most critical combinations
must be identified prior to flight testing [35]. Clearance is further complicated by
the transonic dip phenomenon, where the flutter boundary moves closer to the
flight envelope due to the formation of shockwaves. An example is shown in Figure
1.4 for the BAE Systems Hawk advanced jet trainer with the flutter boundary
calculated using a weakened structural model by a linear method (NASTRAN)
and by non-linear CFD (Euler equations) [7].
In summary aeroelastic phenomenon have generally been seen to have a negative impact on aircraft with flutter leading to severe damage or destruction of
the aircraft or a component. Buffet leads to a reduced fatigue life and therefore
6

Figure 1.4: Comparison between the flutter boundaries predicted using the linear
method and the wing-only configuration for the weakened structural model. (from
ref. [7])
higher operating costs. This all adds to the complexity of clearing an aircrafts
flight envelope. However in more recent years aeroelasticity has been used in
multidisciplinary optimisation routines to attempt to take advantage of the reduced structural weight possible when aeroelasticity is taken into account [36].
As mentioned previously, the X-53 Aeroelastic Research Wing has attempted to
make use of aeroelasticity to increase the control authority of the aircraft. Also
by taking aeroelasticity into account the drag produced can be reduced through
the definition of a jig shape.

1.2

Prediction of Flutter

The standard methods for calculating flutter have been the k and p-k methods.
Both of these methods work by coupling linear aerodynamics and linear structural
mechanics and then solving an eigenvalue problem. In the k method the aerodynamics are introduced into a vibration analysis using complex inertial terms,
the p-k method introduces the aerodynamics as frequency dependent stiffness
and damping terms. These methods are explained in reference [37]. The linear
aerodynamics usually come from the doublet lattice method for subsonic flow.
The doublet lattice method is based upon potential flow, which assumes the flow
is incompressible, inviscid, irrotational and that the angle of attack is small with
7

thin lifting surfaces. Therefore this method does not capture non-linearities in
the flow due to shockwaves, boundary-layers and flow separation. This methodology is implemented in commercial packages such as NASTRAN and performs
well inside its region of applicability and as such is the industry standard that all
other methods are compared against.
However in the transonic regime as previously mentioned there is the transonic
dip in the flutter boundary. This cannot be captured using linear aerodynamics
due to the nonlinearities caused by the formation of shockwaves. To overcome
this problem there has been a move towards unsteady nonlinear aerodynamics,
which can be predicted by, for example, the Euler equations. This has focused
on time-marching calculations and the technology which has been developed to
allow these calculations will be covered in more detail in the next section. In
these time-marching calculations the system is given an initial perturbation and
the temporal response is calculated to determine its growth or decay to infer the
system stability. This has been shown to be effective at calculating the flutter
boundary, for the example of a wing [7] and for a full aircraft [38]. However
this approach is unfeasible for the use of clearing a flight envelope due to the
computational expense.
In recent years there has been an effort to reduce the cost of flutter calculations
at the University of Liverpool and this has lead to the development of the Schur
code [39, 40, 41]. This solver computes the onset of flutter as a stability problem
for the steady state of the coupled system. It assumes Hopf Bifurcation, which
is when stability is lost as the coupled systems Jacobian matrix has a pair of
eigenvalues that cross the imaginary axis. This system is solved using the Schur
complement and has been parallelised to allow realistically large problems to
be computed. The method has been found to be about two orders of magnitude
faster than the time domain approaches. For the flutter calculations in this thesis
the Schur solver will be used and the solver is presented in more detail in Section
2.8.

1.3

Computational Aeroelasticity

Computational aeroelasticity (CAE) refers to the coupling of computational fluid

dynamics (CFD) and computational structural dynamics (CSD). Over the past
forty years the aerodynamic models have increased in fidelity from the transonic
small disturbance in the 1970s, through the Euler equations in the 1980s to
the Reynolds-averaged Navier-Stokes equations (RANS) in the 1990s. In more
recent years detached eddy simulation (DES) and large eddy simulation (LES)
are starting to be used. A more through review of CFD can be found in reference
8

[42].
There are two methodologies for computational aeroelasticity, the monolithic
and the partitioned approaches. In the monolithic approach the governing equations are reformulated to combine the fluid and structural equations, which are
then solved and integrated in time simultaneously. The partitioned approach is
more commonly used and uses separate aerodynamic and structural solvers. This
necessitates a method of transferring the aerodynamic loads to the structure and
the displacements to the aerodynamic solver. Since there are issues with solving
the system of equations in the monolithic approach, the partitioned approach has
seen considerable effort in solving the various technological challenges required
to make this approach work. The partitioned approach can be extended to include additional disciplines in a straight-forward manner, whereas the monolithic
approach requires the governing equations to be reformulated to include the additional discipline. The main technological challenges for the partitioned approach
are,
how to deform the CFD mesh
how to sequence the aerodynamic and structural solvers
how to transfer the loads and displacements between solvers

These technological challenges have all been solved to various degrees.

A number of solutions to mesh movement have been proposed including
remeshing, spring analogy, boundary element method (BEM), transfinite interpolation, Delaunay mapping and radial basis functions (RBF). Remeshing is where
the entire mesh is regenerated at each time step and the solution is interpolated
to the new grid, however this is a computationally expensive method. Batina
[43] introduced the concept of the Spring Analogy for unstructured grids. The
spring analogy works by replacing all the element edges with springs that have
a stiffness, which is the inverse of the edge length. When the boundary is deformed the system of springs is iterated over to update the grid deformation.
This was improved by Farhat et al. [44] through the addition of torsional springs
in addition to the linear springs to prevent zero or negative areas/volumes in
the mesh. A detailed analysis of the spring analogy and its relationship to elliptic grid generation is given by Blom [45]. The spring analogy can be used for
structured meshes, but it is relatively costly compared to transfinite interpolation
(TFI). Transfinite interpolation is an algebraic grid deformation technique that
interpolates the grid points in the computational domain using prescribed points
along the boundaries. An example implementation can be found in reference [46].
Chen and Jadic [47] proposed an exterior BEM solver, which assumes that the
9

CFD mesh is modelled as an infinite elastic medium with a deformable hollow

slit that is the CFD surface mesh. This method is significantly more robust than
the spring analogy, but has a higher computational cost. Delaunay mapping was
proposed by Liu et al. [48] and was shown to be an order of magnitude faster
than the spring analogy. The algorithm can be summarised as 1) using the surface mesh and a number of boundary points a Delaunay graph is generated, 2) all
the CFD grid points are assigned an element from the Delaunay graph and the
area/volume ratio coefficients are calculated, 3) as the CFD surface deforms the
Delaunay graph is deformed, 4) the deformed CFD mesh can then be calculated
using the previously computed area/volume ration coefficients. If the Delaunay
graph elements intersect one another then the deformation is split recursively into
equal amounts and applied in steps. This approach can be applied to any grid
topology including hybrid grids. Radial basis functions were proposed for mesh
deformation by de Boor et al. [49]. Radial basis functions are described in Chapter 3 as a transformation method, but can be applied as a mesh movement scheme
by using the deforming surface mesh instead of the structural model to drive the
volume mesh instead of the surface mesh. Rendall and Allen [2] commented that
the quality of RBF was as good as the spring analogy and only required matrix
multiplications at each time step, but was expensive in terms of data required
with the matrix size being Nsurface points Nvolume points . This method was parallelised by Rendall and Allen [50] and improved through the use of a greedy
algorithm to reduce the number of centres used for the mesh movement.
Solver sequencing was largely solved by Farhat and Lesoinne [51] and the
references therein. Presented were a number of approaches to sequencing for
both serial and parallel calculations.
The transfer of data between solvers has been approached by many researchers
over the years and is still an active topic of research. The most commonly used
transfer method is infinite plate splines (IPS) developed by Harder and Desmarais
[52] in 1972 and is used in the commercial package NASTRAN. Since then a
number of methods have been proposed and these are covered in Chapter 3.
However current methods still struggle with extrapolation beyond the planform
of the structural model.

1.4

Thesis Organisation

The objective of this work is to investigate the influence of transformation methods on the flutter boundary and to consider the simulation of shock-induced buffet
of a transport wing. This involves testing a number of transformation methods
for their effect on flutter boundaries, verifying the flow solver for shock-induced
10

buffet over an aerofoil and then for static aeroelastic calculations of an aeroelastic
wing. Reviews of these topics can be found in their respective chapters (Chapter
3 for transformation methods, Chapter 5 for shock-induced buffet of an aerofoil
and Chapter 6 for shock-induced buffet of a transport wing).
Chapter 2 introduces the formulation of the solvers to be used in this work.
This includes the Aeroelastic Multiblock (AMB) flow solver and the Schur solver
for fast flutter boundary calculations.
Chapter 3 introduces transformation methods and a description of their formulation. A different approach to transformation is proposed called the rigid ribs
approach that is used for beam stick models.
Chapter 4 presents the evaluation of the transformation methods, both for
mode shape reconstruction and their effect on flutter boundary predictions. This
is shown using two test case, the Goland and the multi-disciplinary optimisation
(MDO) wings. The effect of changing the direction of extrapolation for a beam
stick model is also presented.
Chapter 5 is devoted to the calculation of buffet over the BGK No. 1 aerofoil
to verify the AMB solver for shock-induced buffet. In this chapter a number of
convergence studies are presented to find the effect of various parameters on the
solver and solution. This is followed by an estimate of the buffet boundary using
steady state calculations. The estimated buffet boundary is then verified using
unsteady calculations and the mechanism for the shock oscillation is investigated.
Chapter 6 presents the static aeroelasticity of the Aeroelastic Research Wing
(ARW-2) to verify that the aeroelastic deformation and aerodynamics can be
captured by the AMB solver. This is followed by a buffet search using steady
state calculations to find possible buffet cases.
Chapter 7 concludes the thesis and offers suggestions for future work. The appendices offer more detailed formulations of some of the transformation methods
presented in Chapter 3.

11

12

Chapter 2
Formulation
2.1

Navier-Stokes Equations

The equations in differential conservative form are presented.

Continuity Equation
This is derived from the principle of conservation of mass.
(u) (v) (w)
+
+
+
=0
t
x
y
z

(2.1)

where is the density, u, v, w are the velocity components in the x, y, z directions

and t is time.
Momentum Equations
These equations are derived from Newtons second law. The equation for the x
component of momentum is
(u) (uu) (uv) (uw)
p xx yx zx
+
+
+
=
+
+
+
+ fx (2.2)
t
x
y
z
x
x
y
z
for the y component
p xy yy zy
(v) (vu) (vv) (vw)
+
+
+
=
+
+
+
+ fy (2.3)
t
x
y
z
y
x
y
z
and for the z component
p xz yz zz
(w) (wu) (wv) (ww)
+
+
+
= +
+
+
+ fz (2.4)
t
x
y
z
z
x
y
z
where ij are the shear stresses and fi are the body forces.
13

Energy Equation
This equation is derived from the principle of conservation of energy.


qx qy qz
(E) (uH) (vH) (wH)
+
+
+
= q
+
+
t
x
y
z
x
y
z
(uxx ) (uyx ) (uzx ) (vxy ) (vyy ) (vzy )
+
+
+
+
+
+
x
y
z
x
y
z
(wxz ) (wyz ) (wzz )
+
+
+
+ (fx u + fy v + fz w)
(2.5)
x
y
z
where q is the rate of volumetric heat addition per unit mass, qi are the heat
2
2
2
fluxes, the total energy is given by E = e + u +v2 +w and the total enthalpy is
given by H = E + p .
Additional Equations
Since air is a Newtonian fluid the shear stresses are given by
xx = 2 u
+ ( u
+
x
x
+ ( u
+
zz = 2 w
z
x
xz = zx =

u
z

v
y
v
y

+
+

w
)
z
w
)
z

w
x

yy = 2 v
+ ( u
+ v +
y
 x y
v
xy = yx = x
+ u
y


w
yz = zy = y + v
z

w
)
z

(2.6)

where = 23 and the laminar viscosity, , is given by Sutherlands law,

= 0

T
T0

 23

T0 + 110
T + 110

(2.7)

where T is temperature, 0 is the reference viscosity at the reference temperature,

T0 and are taken as 0 = 1.7894 105kg/ms and T0 = 288.16K. Fouriers law of
heat conduction states that
qx =

T
x

qy =

T
y

qz =

T
z

(2.8)

where is the thermal conductivity. So for a given time and location, the NavierStokes equation have the variables , u, v, w, p, e and T to solve for and only
five equations. By assuming that air is a perfect gas an additional equation can
be specified
p = RT

(2.9)

where R is the specific gas constant. If the gas is assumed to be a calorically

perfect gas then
e = cv T,

h = cp T
14

(2.10)

where cv is the specific heat at constant volume and cp is the specific heat at
constant pressure. Then the heat fluxes can be rewritten as
qx =

h
P r x

qy =

h
P r y

qz =

h
P r z

(2.11)

where P r = cp is the Prandtl number. For most aerodynamic applications it

is usual to neglect the body forces and the volumetric heating. Body forces are
assumed to be negligible, therefore the terms fi from the momentum equations
and (fx u+fy v+fz w) from the energy equation are neglected. Volumetric heating
is assumed to be negligible, therefore q from the energy equation is neglected.

2.2

Turbulence

When the Reynolds number is sufficiently low the fluid flows with no disruption
between layers. This type of flow is called laminar. As the Reynolds number
is increased the flow becomes unstable and transitions to a turbulent flow. A
turbulent flow has rapid fluctuations in velocity and pressure and is inherently
three-dimensional and unsteady. Turbulence has a range of scales that are linked
to structures in the flow called turbulent eddies. This can be seen in Figure 2.1
of a laser induced florescence image of a submerged turbulent jet. There are
large scale eddies of the order of the jet width as well as much smaller eddies.
Turbulent eddies are local swirling motions where the vorticity is large. The
largest eddies interact with and extract energy from the mean flow by a process
called vortex stretching. The mean velocity gradients distort the eddies if they
are aligned in a direction in which the mean velocity gradients can stretch them.
Each turbulent eddy has a characteristic length, , velocity, u and time, scale
associated to it. For the larger eddies their Reynolds number tends to be large
(on the order of the mean flow) suggesting that they are dominated by inertial
forces. The smaller eddies are also stretched, but mainly by slightly larger eddies
as opposed to the mean flow. This leads to a cascade of energy from the larger
eddies to the smaller eddies. An energy spectrum of turbulence is shown in
Figure 2.2. It can be seen that most of the energy is contained in the larger
eddies and as such are referred to as the energy containing eddies. The largest
eddies are unaffected by viscosity and depend on their velocity and length scales.
Dimensional analysis gives the spectral energy content as
E() u2

(2.12)

where is the wavenumber from a Fourier decomposition. As such the largest

eddies are expected to be highly anisotropic. At the other end of the scale the
15

Figure 2.1: False-colour image of the far-field of a submerged turbulent jet using
laser induced fluorescence (LIF). (from ref. [8])

Figure 2.2: Energy spectrum for a turbulent flow, log-log scales. (from ref. [9])

16

smallest eddies contain the least energy and are dominated by viscous forces.
The smallest of these scales of turbulence are called the Kolmogorov scales of
length, , velocity, u and time, . The Kolmogorov scales can be derived using
dimensional analysis and are only dependent on the rate of dissipation, and the
kinematic viscosity, and are given by
=

1/4

u = ()

1/4

 1/2

(2.13)

The smallest of the turbulent scales is fixed by viscous dissipation of energy, but is
still many times larger than any molecular scale, as such turbulence is a continuum
phenomenon. However the rate of dissipation is controlled by the rate at which
these smallest eddies receive energy from the larger eddies. Through dimensional
analysis the ratio of the scales can be obtained based upon the turbulent Reynolds
1/2
number, ReT = k l ,

3/4
ReT ,

u
1/4
ReT ,
u

1/2
ReT

(2.14)

This shows that as the Reynolds number increases so does the range of scales in
the flow. Kolmogorov argued that the spectral energy of the smallest scales should
be a function of the rate of dissipation and the kinematic viscosity. Through
dimensional analysis
E( = 1/) 5/4 1/4

(2.15)

The diffusive action of the viscous stress tends to smear out any anisotropic
behaviour in the smallest eddies leading to the smallest eddies being isotropic.
In Figure 2.2 the smallest eddies are referred to as the viscous range. In Figure
2.2 there is an intermediate range of eddies referred to as the inertial subrange.
These eddies were assumed by Kolmogorov to be large enough to unaffected by
viscosity, but small enough to be expressed as a function of energy dissipation,
so through dimensional analysis
E() = Ck 2/3 5/3

for

1
1

(2.16)

where Ck is the Kolmogorov constant. This is also known as Kolmogorovs -5/3

law.
Large eddies carry smaller scale eddies with them as they move with the flow
and the lifetime of eddies can be as much as 30 times the characteristic length
of the flow. This means that the current state of a turbulent flow depends upon
the upstream history and cannot be uniquely defined by the local strain rate
tensor. Finally viscous and turbulent flows are always diffusive and the diffusion
17

is dominated by the large eddies. These large eddies cause increased mixing of
mass, momentum and energy and is called turbulent mixing.
The rate of dissipation, has been used throughout this section and its definition is
=

dk
dt

(2.17)

where k is the specific kinetic energy of the fluctuating velocity.

2.3

The Closure Problem

The Navier-Stokes equations can be solved for turbulent flow in a number of

ways. Direct numerical simulation (DNS) solves the equations directly using a
sufficiently fine grid and time step to resolve all the turbulent eddies, however
this is hugely expensive and it is infeasible for realistic geometries and Reynolds
numbers with current computer hardware. Another approach is to use large
eddy simulation (LES), where the large eddies are directly simulated and the
small eddies are approximated using a sub-grid-scale model. This method is still
too expensive to use for realistic geometries and Reynolds numbers with current
computer hardware, but it is being used more and more in research. The next
simplification is Deattached Eddy Simulation (DES). In this method LES is used
for the majority of the flow, but as the wall is approached the LES method requires
a large number of grid points and tends towards DNS levels of refinement. To
overcome this DES applies Reynolds Averaged Navier-Stokes in the boundarylayer region, which reduces the computational cost. The final method of solving
the Navier-Stokes equations for turbulent flows is the Reynolds Averaged NavierStokes (RANS), also called Unsteady Reynolds Averaged Navier-Stokes (URANS)
when applied to unsteady flows. This is the cheapest method and it will be
shown below that the Reynolds averaging procedure causes a closure problem,
which then requires a turbulence model to close the system of equations. The
calculations to be presented in this thesis use the RANS/URANS approach. The
cost in terms of grid size and time steps and level of empiricism was estimated by
Spalart [1] in 2000 and is shown in Table 2.1. This gives an idea of the relative
cost of each method and when it will be applied to a whole aircraft case assuming
an increase in computer power by a factor of five every five years. The time steps
column is the number of time steps each method needs for the flow to travel six
times the span of the aircraft.
18

Table
Method
URANS
DES
LES
DNS

2.3.1

2.1: Summary
Empiricism
Strong
Strong
Weak
None

of Solution
Grid Size
107
108
1011.5
1016

Method Costs (from Ref. [1])

Time Steps Estimated Ready
103.5
1995
104
2000
6.7
10
2045
107.7
2080

Reynolds Averaging

Reynolds averaging can be used to simplify the Navier-Stokes equations to the

RANS equations. This is achieved through Reynolds decomposition. Reynolds
decomposition involves splitting the instantaneous flow variables into a mean and
unsteady fluctuation components. There are three methods of averaging: time,
spatial and ensemble averaging. The method used here is time averaging. This
is illustrated with a generic flow variable f , which is split into a mean, f and an
unsteady fluctuation due to turbulence, f .
f = f + f

(2.18)

The time average f is calculated by

1
f=
t
noting that
f

1
=
t

t+t

f dt

(2.19)

f dt 0

(2.20)

t+t

The time average of the product of two flow variables f g is given by f g + f g , because the average of the product of two fluctuating components is not necessarily
zero.

2.3.2

Favre Mass Averaging for Compressible Flow

The above Reynolds averaging approach can be used for compressible flow, however it leads to a very complex system of equations. To avoid this Favre mass
averaging is used. Favre mass averaging defines mass averaged variables, so for a
generic flow variable f the mass averaged variable, f is
f
f =

(2.21)

where is the time averaged density. In a similar way to Reynolds averaging the
flow variables can be decomposed into mass averaged f and fluctuating f parts.
f = f + f
19

(2.22)

It is important to note that the time averages of the double primed fluctuating
quantities are not equal to zero, however the time average of f is equal to
zero. This procedure is a mathematical simplification that removes the density
fluctuations from the averaged equations, but it does not remove the effect of the
density fluctuations from the turbulence. To get to the Favre averaged NavierStokes equations the following flow variables are substituted into the NavierStokes equations given by Equations 2.1 to 2.5 and then time averaged.
u = u + u
e = e + e
qx = qLx
+ qx
= +

v = v + v
w = w + w

+h
h=h
T = T + T
qy = qLy
+ qy qz = qLz
+ qz
p = p + p

The resulting Favre averaged Navier-Stokes equations are then

Continuity Equation
(
u) (
v ) (w)

+
+
+
=0
t
x
y
z

(2.23)

Momentum Equations
In the x direction
R
R
)
u) (
uu) (
uv) (
uw)

p (xx + xx
) (yx + yx
(
+
+
+
=
+
+
t
x
y
z
x
x
y
R
(zx + zx )
+
(2.24)
z

In the y direction
R
R
) (yy + yy
)
p (xy + xy
(
v ) (
v u) (
v v) (
v w)

+
+
+
=
+
+
t
x
y
z
y
x
y
R
(zy + zy )
+
(2.25)
z

In the z direction
R
R
)
(w
p (xz + xz

u) (w
v ) (w w)

) (yz + yz
(w)
+
+
+
=
+
+
t
x
y
z
z
x
y
R
(zz + zz )
+
(2.26)
z

20

Energy Equation
uH) (
v H) (wH)

(E) (
+
+
+
=
t
x
y
z



[qLx
+ qT x ] +
[qLy
+ qT y ] +
[qLz
+ qT z ]
x
y
z


R
R
R
u(xx + xx
+
) + v(yx + yx
) + w(
zx + zx
)
x


R
R
R
) + v(yy + yy
) + w(
zy + zy
)
+
u(xy + xy
y


R
R
R
+
u(xz + xz
) + v(yz + yz
) + w(
zz + zz
)
z

where E and H are redefined as E = e + u

The mean viscous stresses are given by
xx = 2 xu 23 ( xu +

zz = 2 zw 32 ( xu +
xz = zx =

v
y
v
y

w

+
+

)
z
w

)
z

The Reynolds stresses ijR are given by

2 +
v 2 +w
2

+ u
+ k and H = h

v
v
yy = 2 y
32 ( xu + y
+


v
+ yu
xy = yx = x


yz = zy = yw + zv

R
R
xx
= u u
yy
= v v
R
R
R

zz = w w
xy = yx
= u v
R
R
R
R
xz
= zx
= u w yz
= zy
= v w

(2.27)

(2.28)
2 +
v 2 +w
2

+ k.

)
z

(2.29)

(2.30)

and the heat flux components (subscripts L and T are for laminar and turbulence
respectively) are

h
qLx = Pr x
qT x = PrTT xh

(2.31)
qLy = Pr yh qT y = PrTT yh

T h
h
qLz = P r z qT z = P rT z
where T is the turbulent eddy viscosity and P rT is the turbulent Prandtl number.
This introduces six more unknowns that need to be modeled to close the problem.
For general three-dimensional flows there are 6 unknown mean-flow properties,
density, pressure, enthalpy and the three velocities, as well as six Reynolds-stress
unknowns, which brings the total number of unknowns to twelve. There are only
five equations to solve and so more equations are required to close the problem.
This is referred to as the closure problem.

2.3.3

Boussinesqs Approximation

One common approach to solve the closure problem is using the Boussinesq approximation. This is based on an analogy between the Reynolds stresses and
21

the viscous stresses. This leads to the equations for the Reynolds stresses being
rewritten in the form

R
xx
= u u = 2T

v
y

1
3

u
2
x

u
2
x

v2
x

v2
x

w
2
x




2

23 k

23 k

 2

u

v2
w
2
R
zz
= w w = 2T zw 31 x
+ x
+ x
2 k

 2 3 2

 
v
u

v
w
2
R
R
13 x
32 k
+ x
+ x
xy
= yx
= u v = 2T 12 yu + x





u
2
v2
w
2
R
R
23 k
+ x
+ x
xz
= zx
= u w = 2T 21 zu + xw 31 x
 

 2

u

v2
w
2
R
R
yz
= zy
= v w = 2T 21 zv + yw 31 x
+ x
+ x
32 k
(2.32)
where k is the specific turbulent kinetic energy given by
R
yy
= v v = 2T

1
3

1
k = u u
2

(2.33)

The two main assumptions used in the Boussinesq approximation are that the
Reynolds stresses can be defined at each point in space and time by mean velocity
gradients and that the turbulent eddy viscosity is a scalar property of the flow
with the relationship between the Reynolds stresses and the mean velocity gradients being linear. This approximation reduces the number of additional unknowns
from six to one. The majority of turbulence models calculate this turbulent eddy
viscosity as a function of velocity and length scales. There are a number of limitations of using this approximation. When a turbulent boundary-layer is perturbed
from its equilibrium state a new equilibrium state is not reached for at least ten
boundary-layer thicknesses downstream [9]. This observation means that the
Boussinesq approximation along with the equilibrium approximations implicit in
algebraic models will not provide an accurate description of separated flow. The
Boussinesq approximation is not valid for flows over curved surfaces due to significant streamline curvature, which gives rise to uneven normal Reynolds stresses.
Flows that are three-dimensional can cause the Boussinesq approximation not
to be valid, this can be demonstrated in straight non-circular ducts where the
turbulence model using this approximation fails to predict secondary motions.
Flows which have large changes in mean strain rate such as separation cannot be
described by the Boussinesq approximation. Although the Boussinesq approximation may not be valid in these flow conditions many turbulence models using
this approximation have been successfully applied to these flow conditions.
22

2.3.4

Vector Form

The Navier-Stokes equations (in this case the Favre mass averaged form) can be
written in vector form,
W (Fi Fv ) (Gi Gv ) (Hi Hv )
+
+
+
=0
t
x
y
z
where W is the solution vector of conserved variables given by

u
v
W=

w
E
The inviscid vectors are given by

uv

uu + p

v u Gi =
v v + p
Fi =
w
w
u
v

uH

vH
The viscous vector are given by

Fv =
Re

Gv =
Re

Hv =
Re

Hi =

uw

v w
w w + p

w H

(2.34)

(2.35)

0
R
xx + xx
R
xy + xy
R
xz + xz
R
R
R
u(xx + xx ) + v(xy + xy ) + w(
xz + xz
) (qLx
+ qT x )
0
R
yx + yx
R
yy + yy
R
yz + yz
R
R
R
u(yx + yx
) + v(yy + yy
) + w(
yz + yz
) (qLy
+ qT y )
0
R
zx + zx
R
zy + zy
R
zz + zz
R
R
R
u(zx + zx
) + v(zy + zy
) + w(
zz + zz
) (qLz
+ qT z )

where Re is the Reynolds number.

2.3.5

(2.36)

(2.37)

Non-dimensional Form

The Navier-Stokes equations can be non-dimensionalised using the following relationships

x=
=

x
,
L

y=
p=

y
,
L
p
2 ,
U

z = Lz ,
u = Uu , v =

t
t = L /U
T = TT , e =
,

23

v
,
U
e
2 ,
U

w = Uw ,

(2.38)

where the superscript denotes a dimensional variable and L is usually taken

to be the root chord. This allows the characteristic parameters such as Reynolds
number, Mach number etc. to be varied independently of the flow variables.

2.3.6

Curvilinear Form

The Navier-Stokes equations can be rewritten in curvilinear form [10], this makes
solving the equations on a body-fitted grid easier. If the physical region of interest
is described by a Cartesian co-ordinate system x, y, z then a mapping is defined
onto a computational domain , , (a unit cube), the transformation can be
written as
= (x, y, z),

= (x, y, z),

= (x, y, z),

t=t

(2.39)

and the Jacobian determinant of the transformation can be written as

J=

(, , )
(x, y, z)

(2.40)

Then the Equations 2.34 are given by

i F
v ) (G
i G
v ) (H
i H
v)
W
(F
+
+
+
=0
t

(2.41)

= 1W
W
J
i = 1 (x Fi + y Gi + z Hi )
F
J
i = 1 (x Fi + y Gi + z Hi )
G
J
i = 1 (x Fi + y Gi + z Hi )
H
J
v = 1 (x Fv + y Gv + z Hv )
F
J
v = 1 (x Fv + y Gv + z Hv )
G
J
v = 1 (x Fv + y Gv + z Hv )
H
J

(2.42)

where

2.4

Turbulence Modelling

If the Boussinesq approximation has been used then the eddy viscosity needs to
be calculated using a turbulence model. Wilcox defines an ideal turbulence model
24

as a model that introduces the minimum complexity whilst capturing the essence
of the relevant physics. The models for the Favre mass averaged Navier-Stokes
equations can be split into two categories, those that use the Boussinesq approximation and those that do not. This section will concentrate on a Boussinesq
approximation model. The turbulence model that will be presented here is SST
two equation model.

2.4.1

Menter SST k- Model

The Menter SST k- model [53] is a combination of the k- and k- models. It

was observed that the k- model performed very well in the viscous sublayer of
the boundary-layer and predicts the skin-friction and velocity profiles near the
wall well. The model has good numerical stability and is better than the k-
model in the log layer of the boundary-layer, especially for flows with adverse
pressure gradients and compressible flows. However the k- model is sensitive to
freestream values of and it fails to predict accurately pressure induced separation. The k- model is insensitive to freestream values of and performs well
for predicting free shear flows. The k- model is insensitive to adverse pressure
gradients, but overestimates the wall shear stress delaying the point of separation. The SST model combines the k- and k- models using a blending function
F1 that is 1 in the viscous sub-layer and 0 away from the surface. This allows
the k- model to be used in the near wall region and the k- model in free shear
flows with the advantages of both models whilst avoiding their weaknesses. The
SST model also uses the basic idea behind the Johnson-King model, to enforce
Bradshaws observation that the principal turbulent shear stress is proportional
to the turbulent kinetic energy in the wake region of the boundary-layer. By
enforcing this proportionality introduces a lag effect into the equations that accounts for the transport of principal turbulent shear stress. This is achieved by
defining the eddy viscosity as shown in Equation 2.43 with a limiter and adding
a blending function F2 , which is 1 for boundary layer flows and 0 for free shear
layers. The SST model is a major improvement over the k- and k- models for
adverse pressure gradients and separating flows.

Eddy Viscosity

T =

a1 k
,

= max {a1 ,
|
{z

Stress Limiter

25

F2 }
}

(2.43)

Turbulent Kinetic Energy Equation

(k)
|t{z }

Rate of change of k

ui

+
uj k) = ij
(
k +
| {z }
xj
xj
xj
| {z } | {z } Dissipation |
Convection

Production




k k
+ k
xj
{z
}
Diffusion

(2.44)

Specific Dissipation Rate Equation

()
|t {z }

Rate of change of

ui

2 +
(
+
uj ) = ij
| {z }
xj
k
xj
xj
| {z } | {z } Dissipation |
Convection

Production

k
+ 2(1 F1 )w2
xj xj
|
{z
}




k
+ w
xj
{z
}
Diffusion

(2.45)

Cross-diffusion

Closure Coefficients

9
(2.46)
100
the remaining closure coefficients are a blend of the k- and k- coefficients using
the equation
a1 = 0.31,

= 0.41,

= F1 1 + (1 F1 )2

(2.47)

where is a quantity to be blended and the subscripts 1 and 2 stand for k- and
k- coefficients respectively.
!
!

k
4
k
500
w2
F1 = tanh(arg14 ), arg1 = min max
,
,
y y 2
CDkw y 2


k
20
CDkw = max 2w2
(2.48)
, 10
xj xj
The coefficients for the k- part are
k1 = 0.85,

w1 = 0.5,

1 = 0.075,

1 =

1 w1 2

(2.49)

The coefficients for the k- part are

k2 = 1.0,

w2 = 0.856,

2 = 0.0828,

and finally
F2 = tanh(arg22 ),

arg2 = max
26

2 =

2 w2 2

2 k 500
,
y y 2

(2.50)

(2.51)

Additional improvements have been suggested by Menter et al. [54], they are a
production limiter
Pk = ij

ui
xj

Pk = min (Pk , 10 k)

(2.52)

and vorticity is replaced with strain rate in Equation 2.43. However these improvements are not implemented in the current solver.

2.5

Flow Solver

All the computations to be presented were performed using the Aeroelastic MultiBlock (AMB) in-house flow solver [55]. A wide variety of problems have been
studied using this code including cavity flows, delta-wing aerodynamics, rotorcraft problems, flutter[7] and control surface buzz[10]. The governing equations
are discretised using a cell-centred finite volume solver combined with an implicit
dual-time stepping method. In this manner, the solution marches in pseudotime for each real time-step to achieve fast convergence. The discretisation of
the convective terms uses Oshers upwind scheme. Monotone upstream-centred
schemes for conservation laws (MUSCL) interpolation is used to provide nominally second-order accuracy and the van Albada limiter is also applied to remove
any spurious oscillations across shock waves. Central differencing is used to discretise the viscous terms, with the resulting non-linear system of equations generated being solved by integration in pseudo-time using a second-order backward
difference. A Generalised Conjugate Gradient method is then used in conjunction
with a Block Incomplete Lower-Upper (BILU) factorisation as a preconditioner
to solve the linear system of equations, which is obtained from a linearisation in
pseudo-time. A number of turbulence models are available in the solver as well as
large-eddy simulation (LES) and detached eddy simulation (DES), however for
the calculations presented throughout the RANS equations were solved using a
turbulence model.

2.6

Mesh Movement

The mesh is deformed using a hybrid approach where the block vertices are
deformed using the spring analogy and the mesh is deformed inside each block
using transfinite interpolation (TFI).

2.6.1

Spring Analogy of Block Vertices

The spring analogy has been used by many researchers over the years to deform
unstructured [43, 45] and structured meshes. However for structured grids there
27

are more efficient methods of mesh deformation, such as TFI. With this in mind
the spring analogy was used to deform the block vertices and TFI was used for
the edges, faces and volumes. The spring analogy works by assuming each block
edge is a spring with a stiffness that is the inverse of its length. To ensure that
the blocks do not skew or invert additional springs were added diagonally inside
the blocks. For the ith vertex connected to the jth vertex by the edge ij, the
stiffness is given by
kij = p

(xi xj

)2

1
+ (yi yj )2 + (zi zj )2

The displacement at the ith vertex is calculated iteratively by

Pmi
n
j=1 kij dxj
n+1
dxi = Pmi
j=1 kij

(2.53)

(2.54)

where the ith vertex is connected to m other vertices (denoted by j) using the
current displacements (superscript n) for the update the displacement (superscript n + 1). The Dirichlet boundary condition is used (fixed displacements at
the boundaries). Once the displacements have been converged adequately then
the updated vertex coordinates are given by
converged
xnew
= xold
i
i + dxi

(2.55)

Once the block vertices have been updated the block edges can be updated
using TFI with the block vertices as boundary conditions. Then the block faces
are interpolated using TFI with the updated block edges as boundary conditions.
Finally the block volumes are updated via TFI from the new block faces.

2.6.2

TFI of Block Edges

TFI is an algebraic method of grid deformation that is computationally efficient

and easily implemented [46]. To update the block edges the updated block vertices
are used as boundary conditions. The nodes along a block edge are transformed
into a unit computational space (a line with local coordinate ) and inside this
space the displacements are calculated using linear interpolation. The blending
function r() along the line can be calculated using
r() =

length of edge from the first node to the current node

length of edge

(2.56)

and therefore the displacements along the edge can be calculated from the edge
vertices,
dx() = dx(0)[1 r()] + dx(1)r()
(2.57)
28

Finally the updated coordinates of a given node are obtained by

x() = xinitial () + dx()

2.6.3

(2.58)

TFI of Block Faces

Now the updated positions of the block edges are known, they can now be used
as the boundary conditions for updating the block faces [10]. The faces can be
transformed to a unit square with local coordinates (,) to allow the displacement
to be interpolated using linear blending functions.
dx(, ) = f1 (, ) + f2 (, )

(2.59)

f1 (, ) = [1 r1 ()]dx(0, ) + r3 ()dx(1, )

(2.60)

where
and
f2 (, ) = [1 r4 ()] [dx(, 0) f1 (, 0)] + r2 () [dx(, 1) f1 (, 1)]

(2.61)

The blending functions ri along each edge are calculated in the same manner as
before, using Equation 2.56. Finally the updated coordinates of a given node are
obtained by
x(, ) = xinitial (, ) + dx(, )
(2.62)

2.6.4

TFI of Block Volumes

Each block is bounded by six updated faces with the displacements known at all
their nodes [10]. The face numbering are given in Figure 2.3. Linear blending is
used between pairs of faces.
dx(, , ) = f1 (, , ) + f2 (, , ) + f3 (, , )

(2.63)

f1 (, , ) = (1 )dx(0, , ) + dx(1, , )

(2.64)

where
f2 (, , ) = (1 ) [dx(, 0, ) f1 (, 0, )] + [dx(, 1, ) f1 (, 1, )] (2.65)
f3 (, , ) = (1 ) [dx(, , 0) f2 (, , 0)] + [dx(, , 1) f2 (, , 1)] (2.66)
The blending functions, , and are given by
= (1 r )(1 r )r1 + (1 r )r r2 + r (1 r )r3 + r r r4

(2.67)

= (1 r )(1 r )r1 + (1 r )r r2 + r (1 r )r3 + r r r4

(2.69)

= (1 r )(1 r )r1 + (1 r )r r2 + r (1 r )r3 + r r r4

29

(2.68)
(2.70)

Figure 2.3: Faces of the block volume (from ref. [10])

where
r1 + r2 + r3 + r4
4
r5 + r6 + r7 + r8
r =
4
r9 + r10 + r11 + r12
r =
4
r =

(2.71)
(2.72)
(2.73)
(2.74)

and the blending functions along each edge ri are calculated as before using
Equation 2.56. Finally the update coordinates of a given node is calculated using
x(, , ) = xinitial (, , ) + dx(, , )

2.7

(2.75)

Structural Solver

If a structure is modelled as linear then it is possible to model the deformation

as a summation of the modes of deformation. This approximation allows an
n degrees of freedom structural model to be written as a second order linear
ordinary differential equation
Mx
+ C x + Kx = f

(2.76)

where M is the mass matrix, C is the viscous damping matrix and K is the
stiffness matrix, all of size n n. Also x is the time dependent displacements
and f is the time dependent external force vector both of size n. To calculate the
undamped free vibration characteristics, Equation 2.76 is rewritten as
Mx
+ Kx = 0
30

(2.77)

This equation can be solved subsituting x = Xeit into Equation 2.77. The
solution is
[K 2 M]X = 0
(2.78)
This is the premultiplied by M 1 to get
[M 1 K 2 M 1 M]X = 0

(2.79)

Let A = M 1 K and = 2 . This leads to an eigen problem to be solved.

[A I]X = 0

(2.80)

This can be solved to give the eigenvalues and the eigenvectors X, which
are the mode shapes. The mode shapes are usually put in to a modal matrix
= [X1 , X2 . . . , XN ] and are mass normalised to give T M = I. Because the
system is assumed to be linear, its characteristics can be determined once prior
to any aeroelastic calculations.
Equation 2.76 can be transformed into modal space by using x = where
is the modal coordinate. First by premultiplying by T and then substituting
x = into Equation 2.76 to get
M + C + K = T f

(2.81)

where M = T M is the modal mass matrix, C = T C is the modal

damping matrix and K = T K is the modal stiffness matrix. The modal mass
and stiffness matrices are diagonal because the modes of vibration are orthogonal
with respect to the mass and stiffness matrices. The modal damping matrix is
diagonal if the damping is proportional. By ignoring damping the equation for
the ith mode is
i + i2 i = Ti fs
(2.82)
where fs is the external force vector acting on the structure. This equation can
be solved for i using a Runge-Kutta scheme and the structural deformation at
a given time step with p modes is given by
x=

p
X

i i

(2.83)

i=1

2.8

Schur Code

Aeroelastic calculations can be simulated by coupling the fluid and the structural
solvers together and marching through time. This method is computationally
expensive. The Schur solver views the problem of computing flutter onset as a
31

stability problem for a steady state of the coupled problem [41]. The semi-discrete
form of the coupled CFD-CSD system is written as
dw
= R(w, )
dt

(2.84)

w = [wf , ws ]T

(2.85)

where
is a vector containing the fluid unknowns (wf ) and the structural unknowns (ws ),
and
R = [Rf , Rs ]T
(2.86)
is a vector containing the fluid residual (Rf ) and the structural residual (Rs ).
The residual also depends on a parameter (in this case is altitude) which is
independent of w. An equilibrium w0 of this system satisfies R(w0 , ) = 0.
The linear stability of equilibria of Equation 2.84 is determined by eigenvalues
of the Jacobian matrix J = R/w. The Schur solver does a stability analysis
based on the coupled system Jacobian which includes the Jacobian of the CFD
residual with respect to the CFD and structural unknowns. The calculation of
the Jacobian J is most conveniently done by partitioning the matrix as
" R R # 

f
f
Jf f Jf s
wf
ws
J = Rs Rs =
(2.87)
Jsf Jss
w
w
f

The details of the Jacobian calculation are given in references [39] and [40].
It is conventional in aircraft aeroelasticity for the structure to be modelled
by a small number of modes, which leads to the number of the fluid unknowns
being far greater than the structural unknowns. This means that the Jacobian
matrix has a large sparse block Jf f surrounded by thin strips for Jf s and Jsf . As
described in reference [41] the stability calculation is formulated as an eigenvalue
problem, focussing on eigenvalues of the coupled system that originate from the
uncoupled block Jss .
The coupled system eigenvalue problem is written as


Jf f Jf s
p = p
(2.88)
Jsf Jss
where p = [pf , ps ]T and are the complex eigenvector and eigenvalue respectively. The eigenvalue (assuming it is not an eigenvalue of Jf f ) satisfies [56] the
nonlinear eigenvalue problem
S()ps = ps
where S() = Jss Jsf (Jf f I)1 Jf s .
32

(2.89)

The nonlinear Equation 2.89 is solved using Newtons method. Each iteration
requires the formation of the residual, S()ps ps and its Jacobian matrix.
The calculation of the correction matrix, Jsf (Jf f I)1 Jf s , is required to form
the Jacobian matrix with respect to ps and . This can be achieved through
2n solutions of a linear system against Jf f I, one for each column of Jsf
with n being the number modes retained. These solutions are then multipled
against Jsf . Now, for each value of the bifurcation parameter, there are multiple
solutions of the nonlinear system in Equation 2.89, and so the cost of forming
the correction matrix at each Newton step, for each solution and for a range of
structural parameters becomes too high. To overcome this the expansion
2
2 3
(Jf f I) = Jf1
f + Jf f + Jf f + . . .

(2.90)

is used where must be small for the series to converge. Note that this assumption
is not restrictive since we assume that the eigenvalue we are calculating is a
small change from the eigenvalue 0 of Jss . Then 0 can be used as a shift to
the full system eigenvalue by replacing Jf f by Jf f 0 I and Jss by Jss 0 I.
This modifies the nonlinear eigenvalue problem in Equation 2.89 by redefining
S() = (Jss 0 I) Jsf (Jf f 0 I I)1 Jf s . The series approximation then
becomes
(Jf f 0 I I)1 = (Jf f 0 I)1 +(Jf f 0 I)2 +2 (Jf f 0 I)3 +. . . (2.91)
When the shifted problem is solved for , the eigenvalue of the original system is
then 0 +. The terms (Jf f 0 I)1 Jf s , (Jf f 0 I)2 Jf s can be pre-computed to
yield the series approximation which can then be evaluated for any at virtually
no computational cost.
This method is referred to as the Schur method. The series approximation
is used for approximating the Jacobian matrix of the residual from Equation
2.89. For the residual the evaluation of S()ps p s can be made based on
a series approximation at virtually no additional cost after the series matrices
are formed. This formulation leads to a very efficient method of tracing the
aeroelastic eigenvalues as functions of altitude, which in turn provides stability
boundaries.

33

34

Chapter 3
Transformation Methods
3.1
3.1.1

Introduction
The Transformation Problem

Computational aeroelasticity (CAE) refers to the coupling of computational fluid

dynamics (CFD) and computational structural dynamics (CSD) to perform aeroelastic calculations. To solve the coupled aeroelastic problem two approaches have
emerged.
The first is called the monolithic approach that uses a tailored aero-structural
solver. In this approach the governing equations are reformulated combining the
fluid and structural equations, which are then solved and integrated in time simultaneously. While using a fully coupled procedure, one must deal with fluid
equations in an Eulerian reference system, and structural equations in a Lagrangian system. A problem with this type of scheme is that the matrices to be
solved are several orders of magnitude stiffer for the structural system compared
to the fluid system, which makes them difficult to solve [57]. There are examples
of this approach [58, 59], but they are limited to 2D problems.
The second approach is called the partitioned approach. In the partitioned
approach separate aerodynamic and structural solvers are used, which are coupled
through the wetted surface. This requires the aerodynamic loads to be transferred
to the structure and the structural deflections to be transferred to the CFD grid.
This is complicated by the CFD requiring an accurate description of the surface
geometry, but the structural model is usually defined on a simplified geometry,
such as a plate, wing-box or beam. Figure 3.1 shows a typical CFD surface grid
and a structural model. Despite adequately describing the important structural
dynamics, this structural simplification produces the problem that there is a mismatch between the fluid and structural discretisation of the interface between the
two models. This means that projection and extrapolation are usually required
in addition to interpolation to transfer data. To solve this problem one can either
35

use a high resolution structural model, resulting in longer computation time, or

use a transformation method. Collectively the reconstruction of the fluid surface
grid node locations and velocities from the structural model, and the transfer of
forces from the fluid surface grid to the structural grid, is referred to as transformation. There are several ways in which the shape of the wing can be altered
in a non-physical manner (sectional and planform). Reviews of transformation
methods can be found in references [10, 60, 61, 62]. Farhat et al. [38] simplified
the transfer problem of the F-16 fighter by defining both the CFD and structure on the same surface using an unstructured Euler CFD solver and a detailed
structural model. However detailed structural models are not always available
and can be difficult to tune to ground vibration tests.

(a) CFD surface grid

(b) CSD grid

Figure 3.1: Comparison of typical CFD/CSD grids

Transformation methods can be grouped into two groups, local and global
methods. Each have their own advantages and disadvantages. Local methods
often depend on connectivity between the aerodynamic surface and the structural model and do not always give a smooth surface. However as the name
suggests, they only use local information and have low memory requirements.
Global methods have higher memory requirements and can produce non-local
effects, but always provide a smooth surface.

3.1.2

Notation

The following notation will be used throughout

Structural model

- Contains ns grid points

36

- xs = (xs1 xsns )T is the set of structural grid points (used for either
x, y or z)
- dxs = (dxs1 dxsns )T are the displacements calculated at the grid
points by the structural solver (used for either dx, dy or dz)
- Fs = (Fs1 Fsns )T are the forces calculated during the transformation
process (used for either Fx , Fy or Fz )
CFD model

- Contains na grid points

- xa = (xa1 xana )T is the set of aerodynamic grid points (used for
either x, y or z)
- dxa = (dxa1 dxana )T are the displacements calculated during the
transformation process (used for either dx, dy or dz)
- Fa = (Fa1 Fana )T are the forces calculated at the grid points by the
CFD solver (used for either Fx , Fy or Fz )

3.1.3

Requirements

There are a number of requirements for a transformation method to satisfy given

in the literature [2, 10, 62].
1. Conservation of energy.
2. Conservation of total force and moment.
3. Exact recovery of translation and rotation.
4. Force and displacement association.
5. Smoothness.
6. Complex geometries.
7. Minimal memory requirements.
8. Minimal CPU requirements.
These are now elaborated upon.
37

Conservation of energy
Conservation of energy can be achieved through the use of the principle of virtual
work. The virtual work W is defined by
W = dxTs Fs = dxTa Fa

(3.1)

where dx is displacement vector and F is force vector. The subscript a denotes

aerodynamic nodes and s structural nodes. Defining a transformation matrix T
that transforms the structural displacements to the aerodynamic displacements
in a linear fashion as
dxa = T dxs

(3.2)

Then the requirement for energy conservation can be satisfied by substituting

Equations 3.2 into Equation 3.1 so the relationship between the forces is
Fs = T T Fa

(3.3)

Using this method the global conservation of energy is satisfied regardless of the
method that is used to fill the transformation matrix [62].
Conservation of total force and moment
Conservation of the global forces can be written as
ns
X

Fsi =

na
X

Fai

(3.4)

i=1

i=1

Exact recovery of translation and rotation

If the structural grid is subject to a rigid body motion, the displacement transfer
should result in the same motion for the aerodynamic surface grid. Any transformation method should be able to accurately resolve rigid body motions without
introducing any distortion.
Force and displacement association
If all the fluid and structural nodes coincide exactly then the forces and displacements must be associated exactly. In this case the transformation matrix
produced by the transformation method should be an identity matrix. This will
only hold if both the structural and aerodynamic surface meshes coincide. If any
fluid and structural nodes coincide at the beginning of the simulation, then these
points should remain attached throughout the simulation.
38

Smoothness
If the deformed structural grid is smooth then the aerodynamic surface grid once
deformed should also be smooth. If the transformation method artificially generates surface distortions (i.e. ripples, ridges, spikes etc.) it could lead to unexpected results like premature separation or additional shockwaves. There is also
the possibility of the CFD solver failing, especially if the grid folds inside the
aircraft for example at the wing-fuselage junction. If the load distribution on the
aerodynamic surface grid is smooth then the transformed load distribution on the
structural grid should also be smooth. However if there are discontinuities in the
flow solution (i.e. shock-waves) then these should be accurately transferred. Here
smooth is defined in a general sense, but for the CFD solution the surfaces need
to be C 2 continuous to avoid pressure blips. However the local transformation
methods in general are only C 0 continuous, the effect of this lack of sufficient
continuity will be explored in a later chapter.

Complex geometries
When applied to a complex geometry the scheme must not introduce any holes
at component junctions. The scheme should also be able to cope with control
surfaces and stores.

Memory requirements
The methods should use a minimal amount of memory, since high fidelity simulations can involve large CFD and structural grids. For example a full aircraft CFD
surface grid can have in the region of na = 2 105 and a high fidelity structural
model can have in the region of ns = 1.8 104 . Although the use of beam stick
models for aeroelastics is common, which reduces the number of structural nodes
to the region of ns = 200. Using a beam stick model and a global transformation
method can lead to a matrix of around 6 ns na = 2.4 108 non-zero values.
The larger the memory requirements for the transformation method the smaller
the simulation size is that can be run.

CPU time
The CPU time that the transformation method requires should be small in comparison to the CFD calculation, to make the scheme feasible for use in a closely
coupled system.
39

3.1.4

Review

Transformation started to gain interest in the late 1960s as panel methods and
finite element structural models usage became widespread. The initial methods
used were beam and surface splines such as those published by Done [63].
In 1972 infinite plate splines (IPS) were proposed by Harder and Desmarais
[52] and use the solution of a multiply-supported infinite plate for the data transfer. This is still a popular method and is used in commercial finite element
packages such as NASTRAN [64] and ASTROS [65]. This method was an attempt to step away from previous methods, which required a large amount of
user input to choose the best structural and aerodynamic points to get an accurate solution. Also in the seventies a number of methods were proposed in
the field of mathematics to solve the problem of scattered data interpolation,
including Hardys multiquadrics (MQ) [66] and inverse multiquadrics (IMQ) and
Duchons thin plate splines (TPS) [67]. However these new methods were not
applied to the aeroelastic transformation problem until much later.
Franke published a paper in 1982 [68] that reviewed all the methods known
at the time for scattered data interpolation. From this paper the MQ and TPS
were noted to perform very well. Murti and Valliappan published a paper in 1986
on inverse isoparametric mapping (IIM) [69] for the application of remeshing for
crack propagation and this would later be used as a transformation method. At
the end of the eighties Appa published a paper on finite plate splines (FPS)
[70], which was aimed at improving on the infinite plate splines of Harder and
Desmarais. This paper also introduced the concept of virtual surfaces.
The nineties saw an explosion in publications on the transformation problem.
In 1990 Kansa [71] published a series of papers on MQ for use in CFD. In 1992
Pidaparti [72] published the first application of IIM to an aeroelastic problem. A
significant contribution to the field of transformation methods in the nineties was
a paper by Smith et al. [60]. This paper reviewed the transformation methods
IPS, FPS, MQ, TPS, IIM, and NUBS through the use of 260 test cases. The
main conclusions were that TPS was the most accurate, robust and cost-efficient
of all methods tested and IIM performed very well, but requires an extension
to 3D and the elimination of its dependency on structured grids. It is also recommended that multiple methods should be available in any software package.
In 1996 Samareh [73] presented a method for load transfer based upon NURBS.
Cebral and Lohner [74] published a conservative, monotonic, adaptive Gaussian
quadrature load transfer method. In 1998 Chen [47] presented a new boundary
element method (BEM).
In the next decade the new methods that were introduced were radial basis
functions (RBF) and constant volume tetrahedron (CVT). Radial basis functions
40

have been used for transfer for a long time, since IPS, TPS and MQ can be
written in this form. Beckert and Wendland [75] presented radial basis functions
to aeroelastics in 2001 using a number of different basis functions using compact
support. This methodology has been improved upon in a series of papers [2, 50,
76, 77, 78]. Constant volume tetrahedron was proposed by Goura [79] in 2001 and
has subsequently been improved upon [7, 39, 62]. This is a local method where
each aerodynamic node is attached to a structural triangle forming a tetrahedron.
As the structural nodes move the out-of plane component is scaled to keep the
volume of the tetrahedron constant.

3.2

Types of Structural Models

There are a number of different structural models that can be used in aeroelastic
calculations. The simplest are the beam stick models (BSM), these models use
a number of beam elements with lumped masses connected together using stiff
springs. These models are assumed to be chordwise rigid. An example BSM is
shown in Figure 3.2 (a) for the XML12 aircraft test case. The beam elements
are shown in black, the rigid bars are shown in magenta and the lumped masses
are shown as blue triangles. The rigid bars are used for visualisation and to
define additional points to aid the transformation methods. These BSM are still
widely used in industry due to the small number of degrees of freedom (DOF) and
subsequently the ease of matching the model to ground vibration tests (GVT).
These models are also used in the conceptual phase of design because the more
complex models are not available until further along in the design process. These
models are usually used for transport-type configurations with high aspect ratio
wings. Plate models model the lifting surfaces as a plates. An example is shown
in Figure 3.2 (b) for the Standard Dynamics Model (SDM) fighter test case.
These models are usually used for fighter aircraft with low aspect ratio wings.
Wing-box models represent a wing as a wing-box, which only consists of the
spars, ribs and skins above and below the ribs and spars, the rest of the structure
is treated as dead weight and modeled as lumped masses. These models are
also assumed to be chordwise rigid. An example is shown in Figure 3.2 (c) for
the Goland wing test case. These are typically used for transport-type aircraft
whose wings are constructed in this way. Shell models represent the aircraft using
shell elements and usually model the spars, ribs and the full skin. An example
is shown in Figure 3.2 (d) for the ARW-2 wing test case. Full finite element
models using bricks can be used, but are quite rare due to the large number
of degrees of freedom that they contain. These models accurately capture the
structure, but at a high computational cost and for aeroelastic calculations the
41

other types of structural models can provide the required accuracy at a more
reasonable cost. An example is shown in Figure 3.2 (e) for the High Reynolds
Number Aero-Structural Dynamics (HIRENASD) wing test case.
This thesis concentrates on the beam stick and wing-box models. Most transfer methods cannot handle beam stick models directly, since beam elements provide displacements and rotations at the beam nodes. It is common to use rigid
bars attached to the beam elements to provide additional points with displacements. Depending on the transfer method to be used these additional points can
be triangulated or used directly.

3.3

Existing Transformation Methods

Common transformation methods are presented in their approximate historical

order.

3.3.1

Multiquadric-Biharmonic

The multiquadric (MQ) was first derived by Hardy [66] for the approximation of
geographical surfaces and was applied to CFD by Kansa [71]. Hardy assumed that
any function s(x) could be written as an expansion of n continuously differentiable
basis functions ,
n
X
s(x) =
i (kx xi k)
(3.5)
i=1

1/2

where x = (x, y, z)T in this case, (kxk) = [kxk2 + r 2 ] and r 2 is a non-zero

parameter. The coefficients i can be found by solving a set of linear equations
in terms of the basis functions. This can be written in a matrix form and then
be used to calculate a transformation matrix. The form of this equation is very
similar to radial basis functions (see Section 3.3.8), but without a polynomial. To
generate the transformation matrix a matrix Css requires inverting. The higher
the condition number the more accurate the resulting transformation matrix will
be, but the Css matrix is harder to invert. This means a compromise needs
to be reached between having a condition number small enough to invert the
matrix and big enough to achieve the desire accuracy. The parameter r 2 also has
an effect on the condition number, as this parameter is increased the condition
number increases and therefore accuracy until the matrix becomes singular [71].
Improvements
Vary Parameter r 2
Kansa [71] found that the key factor in obtaining an accurate result from

42

(a) Beam stick model (XML12)

(b) Plate model (SDM)

(c) Wing-box model (Goland wing)

(d) Shell model (ARW-2)

(e) Full finite element model (HIRENASD)

Figure 3.2: Structural models

43

MQ was the conditioning of the coefficient matrix. It was found that by

varying the parameter r 2 monotonically the accuracy could be improved by
reducing the matrix condition number. The value of r 2 for the j th basis
function is given by
rj2

2
rmin

2
rmax
2
rmin

j1
 n1

j = 1, 2, ..., n

(3.6)

2
2
where rmin
and rmax
are input parameters. By carefully choosing the values
2
2
of rmin and rmax the transformation can be made more accurate by changing
the condition number of the Css matrix.

Unit Sub-domain
Kansa [71] also recommended that the data should be mapped onto a unit
sub-domain to reduce the errors. One-dimensional problems are mapped
on to a unit line, two-dimensional problems are mapped on to a unit square
and three-dimensional problems are mapped on to a unit cube. Then if
necessary additional rotation and shear transformations were introduced to
make the distances more distinct.
Domain Decomposition
Kansa [71] also recommended using domain decomposition to split the large
coefficient matrix into smaller quasi-local problems. This increases the accuracy and computational efficiency.

Observations
MQ performs better than inverse multiquadric (IMQ) where the basis function
1/2
is (kxk) = [kxk2 + r 2 ]
[68]. The main advantages of MQ are that it is
infinitely differentiable function and it is accurate in regions where the surfaces
are steep. However in regions where surfaces are relatively flat MQ produces a
surfaces that is not as smooth as the original surface, this can be improved using
a hybrid scheme. Smith et al. [60] found that MQ accuracy is case dependent
and performs poorly interpolating highly oscillatory functions. It was also found
that the parameter r should be kept within certain limits to insure a stable linear
system of equations. Finally it was found that the need for scaling was case
dependent and that MQ was sensitive to grid resolution.

3.3.2

Infinite Plate Spline

Infinite plate spline (IPS) was proposed by Harder and Desmarais [52] and is used
by programs such as NASTRAN [64] and ASTROS [65]. IPS is a special 2D case
44

of TPS [60]. It is based upon the superposition of the solutions for the partial
differential equation of equilibrium of an infinite plate. The deflections normal to
the plate surface due to n point forces Fi at given locations (xi , yi ) on the plate
can be written as
dz(x, y) = a0 + a1 x + a2 y +

n
X

Fi ri2 ln ri2

(3.7)

i=1

where ri2 = x2i + yi2 . The unknowns, ak and Fi are obtained from the equilibrium
conditions
n
n
n
X
X
X
yi Fi = 0
(3.8)
Fi =
xi Fi =
i=1

i=1

i=1

and from the given deflections at the n nodes

dzj = a0 + a1 xj + a2 yj +

N
X

Fi rij2 ln rij2

(3.9)

i=1

where rij2 = (xi xj )2 + (yi yj )2 . The derivation of the transformation matrix

is given in Appendix A. Then equations 3.2 and 3.3 can be used to calculate the
displacements and forces respectively.
Improvements
Distributed Loads
It was suggested in the original paper by Harder and Desmarais [52] that
the solution can be smoothed by using distributed loads. This is achieved
by replacing Equation 3.9 with

dzj = a0 + a1 xj + a2 yj +

N
X

Fi rij2 ln(rij2 + )

(3.10)

i=1

where is an input parameter. A parameter value of = 0 is the same as

using point loads, Sadeghi et al. [62] showed that by using = 20 the result
obtained by IPS was improved.
Observations
IPS requires the aerodynamic points and the structural points to be in the same
plane. If the aerodynamic and structural points do not lie on the same plane then
they can be projected onto a neutral plane [10]. The deflections for the projected
aerodynamic points are then calculated with the original offset being added to the
projected points to generate the deflected aerodynamic points. The interpolation
function can be differentiated everywhere, the nodes do not need to form a mesh
45

and a minimum of three nodes are required [60]. It is a global method and as
such generates a large transformation matrix. Extrapolations to the edges of
the planform from the interior structural grid points do not always appear to be
reliable. Figure 3.3 shows that IPS is unable to recover rigid rotations exactly. A
circle is driven by a rigid bar and as the bar is rotated the circle is skewed. Smith
et al. [60] found IPS to be sensitive to the grid resolution and that the accuracy
was only adequate.

Figure 3.3: A circle of points rigidly rotated by a bar using IPS (from ref. [10])

3.3.3

Thin-Plate Splines

Thin-plate splines (TPS) of Duchon [67] provide a means to characterise an irregular surface by using functions that minimise an energy functional. This method
is very similar to RBF and MQ, however uses a different basis function. IPS is a
2D special case of TPS. The basis function for TPS is
(kxk) = kxk2 ln kxk

(3.11)

where x = (x, y, z)T in this case. This is solved in the same way as RBF with the
linear polynomial (see Section 3.3.8).
Improvements
Sub-domaining
Smith et al. [60] stated that TPS outperformed IPS when the data be
mapped onto a number of sub-domains. Then if necessary additional rotation and shear transformations were introduced to make the distances
more distinct. This mapping on to a number of sub-domains also causes
the transformation to be more localised.

46

Observations
Kamakoti et al [61] observed that since splines are invariant to translation and
rotation, it is a useful tool for moving and flexible surfaces. Smith et al [60]
commented that if TPS is used globally in 2D it has all the same limitations as
IPS, but when used in 3D outperforms IPS.

3.3.4

Inverse Isoparametric Mapping

Inverse Isoparametric Mapping (IIM) was proposed by Murti and Valliappan [69]
and was applied by Pitaparti [72]. The same shape functions Ni are used for both
the aerodynamic grid points and the structural deformation. Each aerodynamic
node is projected to a quadrilateral element. A pair of generalised coordinates
(, ) are defined on the quadrilateral using shape functions so that
xa =

n
X

Ni (, )xesi

(3.12)

Ni (, )ysei

(3.13)

i=1

ya =

n
X
i=1

where n is the number of nodes for the element. The element is assumed to lie in
the x y plane and the superscript e indicates the structural node is a member
of that element. The shape function for a quadrilateral element [80] is given by
N1 (, ) = (1 )(1 )/4
N2 (, ) = (1 + )(1 )/4
N3 (, ) = (1 + )(1 + )/4
N4 (, ) = (1 )(1 + )/4
, [1, 1]

(3.14)

Observations
IIM is the most accurate among all methods for interpolation [61] and can be
derived for different types of elements such as 8-noded quadrilaterals. IIM is only
valid for 2D interpolations and it is commented that no extrapolation is possible
purely using IIM [81]. However Pidaparti [72] showed that it was possible to use
this transformation method with common extrapolation techniques to extrapolate
data to control surfaces. This method is only C 0 continuous in displacement
between quadrilateral elements, but since the interpolation is bilinear it is an
improvement over weighted triangles.
47

3.3.5

Finite Plate Spline

The Finite Plate Spline (FPS) proposed by Appa [70] uses a finite plate instead
of the infinite plate of Harder and Desmarais [52]. This was applied to a fighter
aircraft wing by Guruswamy and Byun [57]. This method introduced the concept
of a virtual surface, which is a surface that passes through the aerodynamic
and structural points with the same planform as the aerodynamic surface. The
virtual surface is discretised into finite bending elements. A set of constraints
are established such that when deformed, the virtual surface passes through the
deformed structural points. Appa [70] suggested that C 1 shape functions should
be used, since the C 0 did not give satisfactory results. As such the following
transformation matrix can be derived [60, 70]

1 T
T = a 1 K + Ts s
s

(3.15)

where K is the stiffness matrix of the virtual surface, s is the displacement

mapping from the virtual surface to the structural grid, a is the displacement
mapping from the virtual surface to the aerodynamic grid and is the penalty
parameter. Then equations 3.2 and 3.3 can be used to calculate the displacements
and forces respectively.
Observations
FPS can accommodate changes in models or meshes, however it is a 2D method
and was shown to be CPU and memory intensive due to the virtual surface [60].

3.3.6

Non-Uniform Rational B-Splines

The method proposed by Samareh [82] uses NURBS to transfer the forces and
displacements between the CFD and structural models. The concept behind this
method is to have one main representation of the aircraft in the form of NURBS
surfaces. This single representation can be discretised into a number of other
domains (CFD,CSD, etc.). Each domain is solved and transfers data to and from
the NURBS representation, shown in Figure 3.4. More details can be found in
the papers by Samareh [11, 82].
Observations
This requires the structural model to be a full finite element model that shares
the same surface as the CFD, which will add to the cost. This method requires
access to a CAD package.
48

Figure 3.4: NURBS approach [11]

3.3.7

Boundary Element Method

The Boundary element method (BEM) was proposed by Chen and Jadic [47].
BEM is based upon the elastostatic boundary element method with a BEM solver
being devised to generate a transformation matrix. In this approach the fluid
surface mesh forms the external boundary of an elastic homogeneous body with
the structural nodes as internal points as shown in Figure 3.5.

Figure 3.5: Boundary element method [10]

49

Observations
Sadeghi et al [62] observed that BEM recovers the exact transformation of rigid
body motion and produces a smooth transformation, because all grid points are
connected within continuous elastic bodies, and because of the conditions of minimum strain applied for the inverse BEM. The linear approach allows for the
formulation of a global transformation matrix, which is used for conservative displacement and force transformations. BEM can also be used for CFD volume
grid deformation. Rampurawala [10] states that BEM requires more memory
than IPS and as such is memory intensive.

3.3.8

Radial Basis Functions

The radial basis function interpolant [75] has the form

s(x) =

N
X
i=1

i (kx xi k) + p(x)

(3.16)

where s(x) is the function to be evaluated at x, (kxk) is the basis function, the
index i identifies the centres for the RBF and xi is the location of that centre
(centres are usually the structural points). The linear polynomial p(x) is used to
ensure that translations and rotations are recovered. The coefficients i are found
by requiring the exact recovery of the structural node positions. The polynomial
in the x-direction is given by
px (x) = x0 + xx x + xy y + xz z

(3.17)

When the polynomials are included in the system, the additional requirement is
N
X

i q(x) = 0

(3.18)

i=1

for all polynomials q(x) with degree less than or equal to p(x). Equation 3.16 can
be recast into a transformation matrix, as shown in Appendix C. Then equations
3.2 and 3.3 can be used to calculate the displacements and forces respectively.
Basis Functions
There is an infinite choice of basis functions that can be used. A selection of basis
functions from the literature are given in Table 3.1, where (.)+ = max(0, .) and
in the Euclids Hat basis function r = 2 [76].
50

Table 3.1: Basis Functions (kxk) (taken from [2])

Basis Function
Definition
Gaussian (G)
Thin Plate Spline (TPS)
Multiquadric (MQ)
Inverse Multiquadric (IMQ)
Wendlands C0 (C0)
Wendlands C2 (C2)
Wendlands C4 (C4)
Wendlands C6 (C6)
Euclids Hat (EH)

expkxk
kxk2 ln kxk
1/2
(c2 + kxk2 )
1/2
(c2 + kxk2 )
(1 kxk)2+
(1 kxk)4+ (4kxk + 1)
(1 kxk)6+ (35kxk2 + 18kxk + 3)
2
+ 8kxk + 1)
(1 kxk)8+
(32kxk3 + 25kxk


1
4 3
3
2
12 kxk r kxk + 3 r

Improvements
Norm Biasing
The definition of the norm has a significant impact upon the interpolation,
which is the input for the basis function. Typically, the Euclidean norm is
used, shown below
p
(3.19)
kx xi k = (x xi )2 + (y yi )2 + (z zi )2

The modification to the norm proposed by Rendall and Allen [2] is

q
(3.20)
kx xi k = kx (x xi )2 + ky (y yi )2 + kz (z zi )2
where the coefficients kx , ky and kz have the effect of deforming the sphere
of influence of each centre. It was shown in [2] that by choosing values for
kx , ky , kz that do not equal one improved the trailing edge interpolation.

Support Radius
The support radius is defined as (kxk/) where is the support radius
and is any basis function. This allows the control of the area of influence
of each centre. A very large support radius smooths the deformation over
a large area, whereas a small support radius makes it much more localised.
The larger the support radius the higher the condition number of the Css
matrix, which leads to higher accuracy when there the polynominals are not
used. However when the polynomials are used the matrix sparsity pattern
is more complex and the accuracy is increased as the support radius is
decreased. From numerical tests the wing used effects the whether the
condition number increases or decreases as the support radius decreases,
even though the error decreases for both wings tested. The reason behind
this need further work.

51

 Data Storage
Rendall and Allen [2] suggest using a threshold value above which an element of the matrix is stored. This reduces the storage cost associated with
this method.
Pointwise Partition of Unity
Rendall and Allen [78] proposed the pointwise partition of unity, which is
an improvement of the partition of unity of Wendland [83] and Ahrem et al
[76]. The global solution is generated from a weighted combination of local
solutions. The local solutions are for each aerodynamic point deformed by
a small number of the closest structural points. This results in a much
faster solution with lower storage requirements and a more physical force
distribution.
Data Reduction
Rendall and Allen [50] developed a method that uses a greedy algorithm to
minimise the number of points used as centres whilst minimising the error.
This works by starting with a small number of centres and then applying
a unit translation in the x, y and z directions. This gives a measure of the
error introduced by using a subset of the centres. If the error is above a
threashold amount then an additional centre is added at the loction of the
maximum error, this is repeated until the error is below the target error.

Observations
The great strength of RBF is the fact that the same code can be used to perform
the coupling and/or the mesh movement on structured and unstructured meshes.
RBF is an interpolation scheme for use on scattered data sets and as such is not
based upon the structural dynamics or connectivity. Unfortunately this method
is computationally and memory intensive, since the matrix Css is n2s in size and if
mesh motion is also required the memory required is ns nv . Another disadvantage is the large number of options that can be changed, such as basis functions,
norm biasing and support radius. There is a problem reported by Rendall and
Allen [2] and Ahrem et al [76] where camber is induced at the trailing edge of
a wing especially near the tip, this is probably due to extrapolation. The improvements given above address most of the issues at the cost of a more complex
implementation.

3.3.9

Constant Volume Tetrahedron

Constant Volume Tetrahedron (CVT) was proposed by Goura [79]. In this

method each aerodynamic node is assigned to a structural element defined by
52

three structural grid points, shown in Figure 3.6. The position of the aerody-

Figure 3.6: Constant volume tetrahedron [10]

namic node xa,l is given by
xa,l xs,i = a + b + d

(3.21)

where a = xs,j xs,i , b = xs,k xs,i and d = a b. The constants are calculated
using
(b.b)(a.c) (a.b)(b.c)
=
(3.22)
(a.a)(b.b) (a.b)(a.b)
(a.a)(b.c) (a.b)(a.c)
(3.23)
=
(a.a)(b.b) (a.b)(a.b)
(c.d)
=
(3.24)
(d.d)
The volume of the tetrahedron is held constant during the calculation by recalculating . The method can be linearised in the structural displacements [79],
giving
dxal = Adxs,i + Bdxs,j + Cdxs,k
(3.25)
where

A= I BC
B = I UV (b)
C = I UV (a)
U =I

2
D(d)S(d)
d2

53

(3.26)

d is the magnitude of d.

0 z3 z2
0 z1
V (z) = z3
z2 z1
0

z1 0

D(z) =
0 z2
0 0

z1 z2
S(z) = z1 z2
z1 z2

0
0
z3

z3
z3
z3

(3.27)

(3.28)

(3.29)

Equation 3.25 can be written as a transformation matrix, which can be recalculated each time step to reduce error. Then equations 3.2 and 3.3 can be used to
calculate the displacements and forces respectively.
Improvements
Selection of the Structural Elements
For this method the selection of the correct structural element to map the
aerodynamic node to is critical. This mapping is done as a preprocessing
step and is provided as an input to the solver. The original method of
selecting the structural element was to minimise the distance between the
projected fluid point and the structural element centroid. This could result
in the fluid point being above a structural triangle that it was not mapped
to, since that neighbouring centroid was closer [7].

To minimise the amount of extrapolation required the following modification was made. Each structural element was split into three subtriangles.
The area of the jth structural element ABC (ABC ) along with the subtriangles APB (AP B ), BPC (BP C ), APC (AP C ) are calculated where P
is the projected aerodynamic node and the following function is minimised
Sj = |ABC AP B BP C AP C |

(3.30)

If Sj = 0, the point is above the assigned element, interpolation will be

used. So by minimising Sj it also results in the minimisation of the number
of displacements calculated using extrapolation.
Integral CVT
Sadeghi et al [62] used a different method to linearise the CVT equations
using an integral approach, which is presented as a more efficient approach.

54

Observations
CVT is fully 3D and accounts for the out-of-plane component. It is easy to
implement and has low memory and CPU requirements. However smoothness
is not guaranteed because the structural triangles used are only C 0 continuous.
CVT is sensitive to the resolution of the structural grid (the structural grid is
required to have a similar resolution to the aerodynamic grid [62]). It is also
highly dependent on the quality of the mapping routine used.

3.4

Rigid Section Method

A different approach is taken for beam models. Since one of the assumptions is
that the chordwise section is rigid perpendicular to the beam, a direction is defined
along which the wing section is assumed to be rigid. These directions would be
defined by the ribs in the wing structure. The beam is defined by the points xi
and along with the rigid direction, the corresponding leading xLEi and trailing
xT Ei edge points can be calculated. Then, the motion of this section is defined
by the translation dxi and rotation di of the beam point. The displacement of
any point y on the section can then be derived as
dy = dxi + R (y xi )
where the rotation matrix R is given by
2

a K + cos di abK c sin di acK + b sin di

R = abK + c sin di b2 K + cos di bcK a sin di
acK b sin di bcK + a sin di c2 K + cos di

(3.31)

where K = 1 cos di , the axis of rotation is v = ai + bj + ck and |v| =

1. The fluid grid is defined at yj , and the transformation problem is to define
the displacements dyj based on the definition of the displacements and rotation
on the structural grid. To do this each fluid point is projected on the beam
elements parallel to the rigid direction, the displacements and rotation are then
interpolated and Equation 3.31 is applied to the fluid point.

55

56

Chapter 4
Evaluation of Transformation
Methods
4.1

Introduction

There are some questions of interest that this chapter aims to answer. First,
what influence does C0 continuity (CVT and IIM) have on the section shapes?
Secondly what is the contribution of the out-of-plane component on the section
shapes? Finally, what is the influence of any shape distortion on the flutter
predictions? Two test cases are used to shed light on these questions and are
described below.

4.2
4.2.1

Test Cases
Goland Wing

The Goland wing, shown Figure 4.1, is a rectangular wing that has a chord of 6
feet and a span of 20 feet. The aerofoil section is a 4% thick parabolic section.
The CFD grid used was a coarse version with 35 thousand points and is blockstructured using an O-O topology. This allows points to be focused in the tip
region, which is the most critical region for the aerodynamic contribution to
the aeroelastic response. This was solved using the Euler equations. The Euler
equations can be used because the Goland wing is a slender wing and therefore has
a thin boundary layer and along with a high Reynolds number the viscous effects
can be safely neglected. Throughout the flow regime being investigated the flow is
expected to be attached also allowing the use of the Euler equations. At the time
of the calculations the Schur solver was only capable of using the Euler equations
to calculate the flutter boundaries, which was still a significant improvement over
linear methods due to the ability to capture shockwaves. For LCO the paper by
Beran et al. [84] found that the effect of viscosity has a minimal effect on the onset
57

of LCO, but did limit the amplitude of oscillation due to shock-boundary-layer

interaction. This supports the use of the Euler equations for flutter boundary
calculations of the Goland wing.
In this study three structural models were used. The original structural model
is a wing-box that follows the description in reference [84] and includes a lumped
mass tip store. A plate model is calculated from the wing-box by averaging upper
and lower surface values onto a mid-plane using RBE3 elements. This model is
referred to as the original plate.
The second model is an extension of the first. Points are added along the
leading and trailing edges and also at the tip. These points are tied to the first
plate model using RBAR elements. The element properties of the plate model
have no rotational degrees of freedom, which are required by the RBAR elements
for extrapolating the mode shapes. In order to recreate the mode shapes, the
points on the leading and trailing edges as well as the tip have their displacements extrapolated linearly using a Matlab script after being extracted from the
NASTRAN output. Due to the positioning of the structural model inside the
wing, the only areas of extrapolation are at the trailing edge and a small area at
the tip. This model is referred to as the extended plate.
The final model was derived from the first model and is a beam stick model.
The mid-plane points for the centre spar were chosen to be the beam nodes.
The rotations at the beam nodes were calculated from the first and third spars
using a Matlab script. The displacements of the leading and trailing edges were
calculated in Matlab using Equation 3.31 and a simple rotation matrix about the
z-axis. This model is referred to as the beam stick model (BSM).
In addition to the three structural models refined versions of the original and
extended plates were generated. Each element was split into four elements and the
displacements were linearly interpolated using a Matlab script. All the structural
models generated are shown in Figure 4.2.
The sections used in the shape comparisons are at 98% of the span and the
position as well as the original section at this position is shown in Figure 4.1.
This section is translated and rotated to the original orientation to allow for
easier evaluation of the shape change introduced by the transformation methods.

4.2.2

MDO Wing

The second test case is the multi-disciplinary optimisation (MDO) wing [85],
shown in Figure 4.3, this is a commercial transport wing with a span of 36m.
The wing was optimised to fly at a certain altitude and Mach number and has
a thick supercritical section. The CFD grid used was a coarse grid with 81
thousand points and was solved using the Euler equations. The use of the Euler
58

(a) Geometry

(b) CFD surface grid

Figure 4.1: Goland wing geometry and CFD grid

equations is harder to justify than for the Goland wing, because of the thick
aerofoil section. However the Reynolds number is high, which implies that the
viscous effects are low. The paper on the MDO [85], shows that the shock position
is different when including viscosity, but there is a significant degree of spread in
the results between solvers. The Schur solver was only capable of calculating the
flutter boundary using the Euler equations, but one should be mindful that when
calculated with viscous effect that the flutter boundaries may change slightly. For
this work the in-plane mode is neglected.
In this study two structural models were constructed. The original structural
model is a wing-box model and this was converted into two beam stick models.
The MDO wingbox model has a set of points that resemble a beam that acts
as the connection points for the lumped masses. The points are attached to the
wing-box ribs though RBE3 elements, with one point with a mass for each rib.
These points were chosen to be the beam nodes in the new beam model. The
displacements and rotations were extracted directly from NASTRAN at these
nodes.
This serves as the basis for the two models used. For the first model the
points on the leading and trailing edges are chosen so that the rigid ribs are
approximately perpendicular to the beam, matching the ribs on the wing-box
model. The second model has the points on the leading and trailing edges so
that the rigid ribs are parallel to the x-axis (i.e. fuselage). These models are
referred to as the perpendicular and parallel rib models respectively, and are
shown in Figure 4.4.
The section used in the shape comparisons is taken at 98% of the span per59

(a) Original
model

NASTRAN

(d) Beam stick model

(b) Original plate

(e) Refined original plate

(c) Extended plate

(f) Refined extended plate

Figure 4.2: Goland wing structural models (wing planform in grey)

60

pendicular to the beam and the position as well as the original section is shown
in Figure 4.3.

(a) Geometry

(b) CFD surface grid

Figure 4.3: MDO wing geometry and CFD grid

(a) Original NASTRAN model

(b) Perpendicular ribs

(c) Parallel ribs

Figure 4.4: MDO wing structural models

4.3

Evaluation of Transfer - Shape

The transfer methods were implemented and are tested in this section in terms of
the mapped mode shapes. To test the transformation methods shape comparisons
are shown for several cases in increasing complexity. The test section in each case
is translated and rotated back to the original orientation to allow evaluation of
the shape distortion introduced by the transformation.
61

4.3.1

Case 1 - In-plane

For this case the fluid grid is defined on a plane which is on the same surface
as the structural grid and with the same planform as the Goland wing. The
fluid grid is chosen to be finer than the structural grid and this case tests the
in-plane treatment of the transformation methods only. Both the original plate
and extended plate are compared in this case. Figure 4.5 shows the slice at 98%
of the span for the third mode shape, which is the second torsion mode.
The CVT section exhibits a saw-tooth shape for both structural models, but
passes through all of the NASTRAN points. This can be attributed to the lack
of derivative continuity between structural elements. For the original plate, once
the trailing edge of the structure has been reached, CVT extrapolates linearly.
The IIM section passes through all of the NASTRAN points and provides linear extrapolation parallel to the element edges. The fact that IIM is only C0
continuous does not show in the sections. IPS has trouble in the region of extrapolation for the original plate model where it flicks up to an unrealistically large
deflection. Also throughout both structural models IPS oscillates between the
structural model points. There is a well known problem with IPS[10, 60, 62] that
it cannot recreate rigid rotations. IPS performs much better when there is no extrapolation, but still has minor oscillations. The RBF support radius was chosen
to be 1.0 in order for the first mode shape to pass through the NASTRAN points,
and no norm-biasing was applied. The RBF result shows the same problem with
extrapolation as IPS, however it is much more severe. The observed additional
camber has been seen in other papers [76] (incorrectly attributed to fuselage
interference) and Rendall et al. [2] (where it was reduced by using norm-biasing).
This case highlights the differences between the local (CVT and IIM) and
global (IPS and RBF) methods. The local methods exhibit a problem with slope
discontinuity that the global methods do not suffer from, although IIM did not
suffer from this visibly. IIM uses bilinear elements, so although they are only
C0 continuous this higher order interpolation appears to avoid the slope discontinuities experienced by CVT. The global methods however fail to extrapolate
realistically and have oscillations that are reduced when there are no extrapolation regions. All the transformation methods performed better on the extended
plate than the original plate due to the lack of extrapolation.

62

(a) Original plate

(b) Extended plate

Figure 4.5: Case 1: Section at 98% span for the third mode shape, rigidly translated and rotated back to the original orientation.

63

4.3.2

Case 2 - Goland Wing In-plane

For this case the fluid grid is now defined on the correct wing profile, but the
fluid points are projected onto the structural plane. The transformation methods
are used to define the displacements at the projected points and then these displacements are applied without modification to the wing points. This tests the
discrepancy introduced by failing to calculate the out-of-plane component. Again
only the original and extended plate models are used.
Figure 4.6 shows a slice through the two plate models at 98% of the span for
the third mode shape. All of the methods show the same behaviour as seen for
case 1, but now on the wing section. CVT shows the C0 continuity effect which
leaves the aerofoil nonsmooth. Interestingly IIM shows no sign of this problem
and produces an excellent section shape. RBF and IPS both display the same
behaviour as case 1 with the trailing edge failing to be extrapolated correctly, as
well as section oscillations.

(a) Original plate

(b) Extended plate

Figure 4.6: Case 2: Section at 98% span for the third mode shape, rigidly translated and rotated back to the original orientation.

64

4.3.3

Case 3 - Goland Wing Out-of-plane

Next, the CVT out-of-plane component is added to the in-plane components

calculated from each method. Figure 4.7 shows the transformed slice. There is
no significant change to the mode shapes between cases 2 and 3. It does seem
to have the effect of thickening the section slightly. This is confirmed when the
IIM result for case 2 is cross plotted with the IIM result for case 3, shown in
Figure 4.8. Since the out-of-plane components come from CVT there is a slight
discontinuity introduced in the slope. It can be seen that the effect of the out-ofplane component is small.
To complete the picture of the performance of the transformation methods
on a plate a refinement of the structural grid was undertaken, and is shown in
Figure 4.9 for the extended plate model. For all cases except IIM a significant
improvement can be observed. For IIM there was no improvement shown, which
implies that the IIM solution is already grid independent. For CVT the slope
discontinuity problem persists, but the severity is less for both of the refined
models. IPS is improved in the regions of grid refinement, it is expected that
another level of grid refinement would converge the solution. RBF also benefits
in the region of refinement and the extrapolation is also effected by the refinement.
However although the improvement is significant for the RBF the sections are still
the worst of the methods used.

(a) Original plate

(b) Extended plate

Figure 4.7: Case 3: Section at 98% span for the third mode shape, rigidly translated and rotated back to the original orientation.

65

(a) Original plate

(b) Extended plate

Figure 4.8: Case 3: Sections at 98% of the span through the mode 3 mode shape
for IIM comparison for cases 2 and 3, rigidly translated and rotated back to the
original orientation.

(a) CVT

(b) IIM

(c) IPS

(d) RBF

Figure 4.9: Case 3: Sections at 98% of the span for the refined extended plate
model, rigidly translated and rotated back to the original orientation.
66

4.3.4

Case 4 - Goland Wing Beam Stick Model

Finally the beam stick model is used and all transfer methods have the out-ofplane contributions from the CVT method applied to them. Figure 4.10 shows
a slice at 98% for the third mode shape. All the methods are effectively using
a coarsened extended plate model, since the BSM has one less row of structural
points in the chord-wise direction that corresponds to the trailing edge of the
wing box in the other structural models. The sections are significantly worse as
the trailing edge is approached for all methods except IIM. CVT has a higher
amplitude saw-tooth in the trailing edge region, but the saw-tooth in the leadingedge region is unchanged from case 3. This underlines the local character of
this method. IIM appears to only be affected by the poor CVT out-of-plane
contributions. IPS has the largest deviation from the section shape in the coarse
region of the grid at the trailing edge, but also see an increase in the amplitude
over case 3 in the leading edge region. RBF has the same trend as IPS, but to a
worse degree.

Figure 4.10: Case 4: Sections at 98% span through the mode 3 mode shape
rotated and translated back to the original orientation.

4.3.5

Case 5 - MDO Wing

Next, results are shown for the two structural models of the MDO wing. The
transformation methods are compared in terms of the vertical displacement calculated in Figure 4.11. The rigid rib results are considered exact in this comparison
since all methods only have access to the beam structural model, and the rigid rib
method exploits this information exactly by design. The displacements obtained
by the different methods for the perpendicular ribs are similar except for those
from IPS, which suffers due to the coarse structural grid used. RBF is not shown
67

because the deflections are very different compared to the other methods. There
is much more variety in the results for the parallel ribs, in this case IIM and the
rigid ribs approach match well. CVT has a slight saw-tooth pattern and IPS deviates slightly due to the coarse mesh. RBF prediction is rather poor for the MDO
wing. The reason that the discrepancies are small compared to the Goland wing
is due to the scaling. A significant difference is seen between the displacements
obtained from the different rib orientations, with larger displacements seen for
this mode from the parallel ribs shown in Figure 4.11(c). For other modes this is
the opposite way around. The reason for this difference is that the values for the
rotation of the section are obtained from different points on the structural beam.

(a) Perpendicular rib model

(b) Parallel rib model

(c) Rigid rib comparison of parallel and perpendicular ribs

Figure 4.11: Case 5: Vertical displacement of sections at 98% span for the mode
5 mode shape.

68

4.3.6

Summary

From these test cases it has been shown that allowing the transfer methods to
extrapolate is inadvisable. The in-plane component dominates the final displacements calculated, other factors have a larger effect on the final displacement than
accounting for the out-of-plane affects. Of the methods tested only IIM was unaffected by changing grid density. All the other methods benefited from the structural grid being refined. The local method CVT suffered from a saw-tooth pattern
of discontinuities in the slope, which could be reduced by refining the structural
grid. The other local method IIM did not suffer from this discontinuity in slope,
probably due to the higher order interpolation inside the bilinear elements used.
The RBF result could be improved using the techniques and advanced algorithms
covered in the previous chapter. Also Ahrem et al. [77] published a paper on
using RBF with BSM, however the proposed algorithm was found to be six times
more expensive than the standard algorithm. The proposed rigid ribs method
was shown to perform well producing smooth section shapes similar to IIM. As
expected the orientation of the ribs has an effect on the section shapes, which as
the orientation becomes further from the ideal perpendicular section the section
shapes get worse.

4.4

Flutter Evaluation

Having evaluated the influence of the transformation methods on the sectional

shape, the impact of these distortions is now evaluated for the aeroelastic stability predictions. The Schur eigenvalue method is used to trace the aeroelastic
eigenvalues as a function of altitude (see the Schur solver in Section 2.8).

4.4.1

Goland Wing

The Goland wing mode tracking at a Mach number of 0.80 and zero degrees
incidence for all modes using CVT and the original plate model is shown in Figure
4.12. It can be seen that modes one and two interact, with mode one becoming
undamped at 13,216 feet as the real part of the eigenvalue becomes positive. The
evaluation of the influence of the transformation methods is presented below for
the real part of mode one.
The effect of using different structural models is shown for the Goland wing
in Figure 4.13 and Table 4.1. Consistent with the section shape results presented
above, IIM shows the least spread of results as the structural model is changed.
CVT and IPS show comparable spread in their results, particularly where the
interaction is strong as the mode becomes undamped. RBF shows a large spread
69

and has no real correlation between the different structural model results. The
original plate model did not flutter for mode one, when all the other transfer
methods and structural model combinations did. The extended and BSM models
did flutter in mode one, but their flutter points are almost 9,000 feet different.
Next the results from different methods are compared for each structural model
in turn and this is shown in Figure 4.14. The predictions from the local IIM and
CVT methods are similar for all the structural models. There is considerable
spread in the results for the global methods (IPS and RBF). For the extended
plate where the section shapes were the most reasonable, all the flutter points are
clustered including RBF. The original plate shows the CVT and IIM methods
closely clustered with the IPS flutter point close, RBF however did not flutter
for this mode. The BSM model is effectively a coarsened extended plate and
since the global methods suffered the most in the section shapes their results are
spread out compared to the local methods. If the IIM flutter results are assumed
to be correct due to the excellent performance in the section shapes, then the
other methods results can be compared with them. It would appear that the
discontinuity in slope has a smaller effect on the flutter results than induced
oscillations in the mode shapes.

(a) Real part

(b) Imaginary part

Figure 4.12: Goland wing mode tracking for all modes using CVT

70

(a) CVT

(b) IIM

(c) IPS

(d) RBF

Figure 4.13: Goland wing mode tracking for mode 1 for the different structural
models (real part)

Table 4.1: Flutter altitude for Goland wing

Original plate Extended plate
CVT
13216
13865
IIM
13025
13002
IPS
14480
13890
RBF
N/A
13243
Spread
1455
888

71

(all values in feet)

BSM Spread
12502 1363
12790
235
15373 1483
22175 8932
9673

(a) Original Plate

(b) Extended Plate

(c) BSM

Figure 4.14: Goland wing mode tracking for mode 1 for the different structural
models (real part)

72

4.4.2

MDO Wing

The mode tracking for the MDO wing at a Mach number of 0.85 and an incidence
of one degree based on the beam model with perpendicular ribs using CVT, is
shown in Figure 4.15. Modes one, two and four participate in the instability, with
mode one going undamped first at -2358m.
The effect of the rib orientation for the MDO wing is shown in Figure 4.16.
It can be seen that there are significant differences in damping between the two
rib orientations, even if the crossing at a similar altitude. The comparison of
the predictions from the different transformation methods is shown in Figure
4.17 and Table 4.2. The perpendicular ribs has a tight clustering of the flutter
points for all the methods except RBF which fails to flutter for this mode. This
is not a suprise since its section shape was so different to the other methods it
was omitted. The parallel ribs shows more variation between the methods with
the local methods clustered tightly. The rigid ribs predicts flutter at a slightly
higher altitude compared to the local methods, but the global methods predict
the flutter point at even higher altitudes. It is not obvious from the sections why
the rigid ribs and IIM are not more similar in this case, although it could be due
to the other modes that interacts with mode one having a different section shapes
leading to a different flutter points.

(a) Real part

(b) Imaginary part

Figure 4.15: MDO wing mode tracking for all modes using CVT

73

Figure 4.16: MDO wing mode tracking for mode 1 for the different structural
models (real part)

(a) Rigid ribs perpendicular to the beam

(b) Rigid ribs parallel to the x-axis

Figure 4.17: MDO wing mode tracking for mode 1 for the different structural
models (real part)

Table 4.2: Flutter altitude for MDO wing (all values in meters)
Perpendicular Parallel
CVT
-2358
-2612
IIM
-2175
-2481
IPS
-1848
-915
RBF
N/A
4198
Rigid ribs
-2125
-1743
Spread
510
6810

74

4.5

Summary

The performance of transformation methods was evaluated. Several common

methods were compared for two test cases in terms of their influence on the
aeroelastic damping. It is one thing to have a predictive method for calculating
flutter, however confidence needs to be built up in its ability and robustness.
Through using different transformation methods it has been shown that when
there is no extrapolation there is minimal spread in the resulting flutter point.
When the methods are required to extrapolate there is a large degree of spread in
the flutter point. This clearly shows that to minimise the error associated with
using the transformation methods no extrapolation should be used.
It was also shown that the orientation of the ribs is important in BSM because
most transfer methods do not use rotation data and therefore require additional
displacements to be provided (usually at the leading and trailing edges). Although
it is much simpler and quicker to use the parallel ribs approach, it is more reliable
to use the perpendicular ribs approach that also matches with the assumptions of
BSM. By using the perpendicular ribs the scatter in the results was also reduced.
Whilst it is unlikely to be the norm to have multiple structural models of
different types to test on, it should be possible to have multiple transfer methods
available. Since it has been shown that the choice of transfer method can have
a significant affect on the predicted flutter point then it is strongly advisable to
use more than one transformation method during the verification process.

75

76

Chapter 5
Aerofoil Buffet
5.1

Introduction

Transonic buffet is associated with shock/boundary-layer interaction, although

buffet is a more general term used for any flow unsteadiness inducing structural
vibrations. Consequences of this phenomena can range from reduced fatigue life
to structural damage [4]. Transonic buffeting is interesting because it has more
severe buffet loads than the low-subsonic or supersonic conditions. Transonic
buffet occurs when a shock is strong enough to induce boundary-layer separation
at its foot. Under certain conditions a large amplitude periodic shock motion
can occur. The prediction of this behaviour is challenging for CFD. The success
of the prediction depends on the turbulence model used. This chapter reviews
previous work on the prediction of flow over the Bauer-Garabedian-Korn (BGK)
No. 1 aerofoil. Then the unsteady flow over the BGK No. 1 aerofoil is predicted.

5.2

Test Case

Previous investigations at the National Aeronautical Establishment (NAE) in

Canada have shown that supercritical aerofoils have more favourable buffet characteristics compared to conventional aerofoils. In the 1980s and 1990s the BauerGarabedian-Korn (BGK) No. 1 supercritical aerofoil was investigated experimentally at NAE. The BGK was designed for a Mach number of 0.75 to have a lift
coefficient of 0.63 with a thickness-to-chord ratio of 11.8%. From the results of
these experiments a number of papers were published [86]-[89]. Later this aerofoil
was selected as an AGARD test case and the experimental results were published
in detail by Huang [90]. Finally CFD predictions were published by Xiao et al.
[3, 91].
It was found that the buffet excitation frequency was 70-80Hz for Mach numbers 0.688 to 0.796. The buffet boundary for this aerofoil, along with the ex77

perimental data points is shown in Figure 5.1. The shock oscillation region was
obtained by fixing the Mach number and varying the angle of attack. At each
angle a power spectra for the normal force was computed and the presence of
shock oscillation was deduced from whether a peak at 70-80Hz was present. The
buffet boundary was calculated from a plot of rms values of the normal-force
fluctuations, CN vs. CL and was determined as the point that has a slope of
dCN /dCL = 0.1 [87]. This was an arbitrarily chosen value that matched the
buffet boundary calculated by the trailing-edge pressure divergence criterion.

Figure 5.1: BGK No. 1 buffet boundary with the experiment locations (circles)
The experiments covered the Mach number range of 0.501 to 0.805, angle
of attack range of -0.36 to 11.74 and Reynolds numbers between 15 million
and 21 million. The wind tunnel had a floor and ceiling porosity of 19.3% or
20.5% depending on the reference. The aerofoil used in the experiments had a
chord of 10in. and a span of 15in., and had 50 pressure orifices on the upper
surface and 20 on the lower surface. It also had 16 miniature fast transducers
on the upper surface for unsteady measurements. The locations of these pressure
orifices and fast transducers are given in ref. [90] and are shown in Figure 5.2.
For experiments at Mach number 0.688 skin friction was also measured. It was
estimated that transition occurred at 10% of the chord. The results published
focused on three test cases; case 1: Mach number 0.688 with varying angle of
attack; case 2: Mach number 0.710 with varying angle of attack; case 3: angle of
attack 6 with varying Mach number.
78

(a) Pressure orifice locations (from ref. [88])

(b) Fast transducer locations (from ref. [88])

Figure 5.2: BGK No. 1 supercritical aerofoil

Case 1 also had skin friction measurements taken. The shock boundarylayer interaction was classified into three types and the various angle of attack
values were characterised. The first class of shock boundary-layer interaction is
characterised by a weak shock that results in a low level shock excitation. The
experiment at a Mach number of 0.668 and an angle of attack 3.99 was an
example and showed a large peak in pressure fluctuations at the shock and the
remainder of the aerofoil had very low levels. The second class is characterised
by a strong shock that causes local separation and reattachment, for example in
the experiment at Mach 0.668 and angle of attack 4.95 . The skin friction showed
that there was a region of separation behind the shock followed by reattachment.
The pressure fluctuations showed a large peak at the shock and second small
peak in the separation bubble with the remainder of the aerofoil having very
low levels. The third class is characterised by a strong shock followed by fully
separated flow, but with some of the characteristics of the separation bubble
and trailing-edge separation i.e. there is a peak in the pressure fluctuations that
corresponds to the separation bubble even though this no longer exists. The
experiment at Mach number of 0.688 and an angle of attack 6.94 showed this
behaviour. The pressure fluctuations again showed a large peak at the shock and
a second moderate peak for the separation bubble followed by a moderate level
for the remainder of the aerofoil. When the angle of attack was increased to 9
the flow became very unsteady with very large shock motions. It was also found
that as the angle of attack increased the trailing edge separation expanded faster
than the separation bubble behind the shock. The switch from class two to class
three shock boundary-layer interaction happened with a small increase in angle of
attack. It was found that below angle of attack 4.67 that the flow was attached,
at angle of attack 4.67 a separation bubble is detected. At angle of attack 6.13
trailing edge separation had reached x/c =0.89 and above angle of attack 6.94
the aerofoil became fully separated. The maximum fluctuation in normal force
CN was found to be at an angle of attack 8 .
Case 2 was at Mach 0.710 with an angle of attack range of -0.316 to 6.97.
79

The results were very similar to those of case 1. The flow at angle of attack
-0.316 was sub-critical with fully attached flow and at angle of attack 1.396
there was a weak shock formed with no separation. For angle of attack 3.017
there was a stronger shock with small oscillations which had decayed by the next
transducer 5% of the chord downstream. A separation bubble was formed behind
the strong shock at angle of attack 4.905 and then at angle of attack 6.970 the
flow has become fully separated.
For case 3 the experiments used an angle of attack of approximately 6 whilst
varying the Mach number from 0.597 to 0.772. It was found that at M=0.597
the flow was attached due to the absence of pressure fluctuations. As the Mach
number was increased to 0.688 the flow gained a stronger shock that leads to
a separation bubble. Above Mach 0.722 the flow is fully separated from the
shock to the trailing edge and there is no evidence of the separation bubble. The
maximum CN was found to be at Mach 0.733. The frequency of oscillation was
found to be a linear function of Mach number.
For the BGK No. 1 aerofoil experiments, boundary-layer suction was used and
the ceiling and floor had 19.3% or 20.5% porosity depending on the reference. Also
the span of the aerofoil was only 1.5 chords potentially effecting the accuracy of
the CFD predictions due to the uncertainty caused by the wind tunnel walls. It
is likely that the two-dimensional character of these experiments is not beyond
question.
There has been a CFD study using the BGK No. 1 aerofoil by Xiao et al.
[3]. This CFD study focused on case two, the flow conditions used are shown in
Table 5.1. The grid used had a C-topology and was 640 64 with 512 points on
the aerofoil and 128 in the wake. The initial wall spacing used was 1 106 of
the chord. A grid refinement study was carried out and it was found that this
grid was adequate for their study. The grid refinement study also found that the
dominant frequency was insensitive to the grid refinement.
Table 5.1: Xiao et al. [3]: Flow conditions
Reynolds Number
20 106
Mach Number
0.710
Angles of Attack
1.396 , 6.970 , 9.000
Non-dimensional Time Step
0.05
Non-dimensional Duration
150-200
The first case was at angle of attack 1.396 , where a weak shock is expected
with no separation. The CFD predicted the shock too far downstream, however
when the angle of attack was changed to account for wind tunnel corrections
the shock was predicted closer to the experiment. The second case was at an
80

angle of attack of 6.970 , the experimental result was a strong oscillating shock
causing full separation. The mean lift coefficient from the CFD was 1.030 and the
experimental mean lift coefficient was 1.016. For the dominant reduced frequency
the CFD prediction was 36% lower than the experimental result. For this case
no wind tunnel correction was used. The CFD failed to capture the magnitude
of the unsteady pressure, however there was better agreement for the phase. The
final case at angle of attack 9.000 showed a steady shock at the leading edge
that caused separation to the trailing edge.
An attempt has been made to understand the mechanism behind self-sustained
shock oscillations and shock-induced buffet [12, 92, 93]. This was also studied
computationally by Xiao et al. [3]. The mechanism proposed by Lee [12] is shown
in Figure 5.3. The mechanism is for self-sustained shock oscillations with fully

Figure 5.3: Model of self-sustained shock oscillation for the BGK No. 1 aerofoil
(from ref. [12])
separated flow after the shock wave. Due to the shock motion pressure waves
are formed that propagate downstream in the separated flow region at velocity
ud . When these waves reach the trailing edge they generate new waves either
from the wake fluctuation or from the trailing edge boundary layer, which travel
upstream outside the region of separated flow at velocity uu . These new waves
interact with the shock and impart energy to it to maintain the oscillation. This
loop then repeats, the period of the shock oscillation Tp should agree with the
time it takes for the waves to propagate to the trailing edge and the time it takes
the new wave to propagate upstream to the shock. This can be given by the
following equation
Z c
Z xs
1
1
dx
dx
(5.1)
Tp =
uu
c
xs u d
where ud is the speed of downstream pressure wave propagation, uu is the speed
81

of upstream pressure wave propagation and xs is the time-mean shock position.

It has also been shown experimentally that during buffeting the wake oscillates
and the periodic shock motion is coupled with the disturbances generated at the
trailing edge. It was also shown experimentally that the unsteady pressures in the
wake are random until a periodic shock oscillation occurs. Then the wake shows
a distinct frequency peak which is the same as the shock frequency. Lee et al.
[92] used the non-linear transonic small disturbance equation to investigate the
propagation of the Kutta waves over a series of aerofoils with shock waves. The
empirical formulation that had previously been used to compute ud and uu was
shown to compare favourably with the numerical computations. Xiao et al. [3]
used cross-covariance of signals in various locations in the CFD flowfield to show
that a pressure wave was propagating downstream in the separated region and
that there was a pressure wave propagating upstream outside this region. This
agrees well with Lees model. They also shows that by using the cross-covariance
they could calculate ud and uu and by using Equation 5.1 calculate the time
period of the shock oscillation. This can be converted into a reduced frequency
of 0.175 which compared favourably with the Fast Fourier Transform (FFT) of
the lift coefficient of 0.16.

5.2.1

Other Aerofoil Buffet Cases

In the 1970s experiments were undertaken on an 18% thick circular-arc aerofoil

at NASA for computer code validation. It was tested in a wind tunnel with
contoured upper and lower walls for a nominal Mach number of 0.775. It has
been studied both experimentally and numerically [13]-[99].
The experiments covered the Mach number range of 0.71 to 0.78 and Reynolds
numbers between 1 million to 17 million. The angle of attack was kept constant
at zero degrees. The aerofoil used in the experiments had a chord of 20.3cm and a
span of 25cm. McDevitt et al. [13] showed experimentally that the experimental
setup and conditions were two-dimensional in nature.
This aerofoil has three distinct flow regimes, the first occurs when below the
critical Mach number of 0.76 . It is characterised by a steady weak shock and
trailing edge separation. The second flow regime occurs between Mach number
0.76 and 0.78. In this regime there is a strong shock on the upper and lower
surfaces that oscillate out of phase along with shock induced boundary-layer
separation. Finally the third flow regime is as the Mach number is increased
beyond 0.78 and the shocks become stronger. The shock becomes steady and
induces boundary-layer separation that extends to the trailing edge. When the
Reynolds number is greater than 5 million the flow exhibits a flow hysteresis. The
flow hysteresis is shown in Figure 5.4.
82

Figure 5.4: 18% thick circular-arc aerofoil shock oscillation domains (from ref.
[13])
The mechanism for the flow over this aerofoil at zero degrees was described
by McDevitt et al. [13]. It was explained as the shock induced boundary-layer
separation on the upper surface causes a thickening of the boundary-layer on the
upper surface. This effectively creates a negative camber which has the effect of
slowing the flow on the upper surface. This tends to suppress the shock-induced
phenomenon, but at the same time induce higher velocities over the lower surface,
causing shock induced boundary-layer separation there and the flow field flips.
This mechanism was verified by McDevitt [95] by using a splitter plate at the
leading and trailing edges.
An alternate mechanism was proposed by Chen et al. [99] similar to that
of Lee for the BGK No. 1 aerofoil. The model proposed by Lee assumed the
shock oscillates above the aerofoil, the model proposed by Chen et al. assumes
the shock propagates upstream and leaves the aerofoil. The flow mechanism is
described as a series of compression waves develop in the region near the trailing
edge and move upstream. As they move upstream they coalesce to form a strong
shockwave, this shockwave continues to move upstream whilst strengthening and
induces boundary-layer separation. Once it has reached the mid-chord the shock
weakens to a weak shock and then compression waves that propagate upstream
leaving the aerofoil.
At ONERA in France wind tunnel tests were done on the OAT15A aerofoil
which is a supercritical profile. The experiments were done by Jacquin et al. [100]
and CFD calculations have also been performed by Deck [101].
The experiments covered the Mach number range of 0.7 to 0.75, angle of
83

attack range 2.5 to 3.91 at a Reynolds number of 3 million. The aerofoil is

12.3% thick with a trailing edge thickness of 0.5% and in the experiment the
chord was 230mm with a span of 780mm. The experiment had 68 static pressure
orifices and 36 unsteady Kulite pressure transducers. Transition was forced at
7% of the chord and buffet onset was found to be at 3.1 for Mach 0.73. When
the angle of attack was increased to 3.5 the shock oscillations traversed 20% of
the chord with a frequency of 70Hz. This frequency was insensitive to angle of
attack, but was sensitive to Mach number.
The transonic buffet on the aerofoil is similar in character to the BGK No. 1
aerofoil. The boundary-layer remains attached as the shock moves downstream,
when it reaches its most downstream location the shock causes the boundary-layer
to separate. The region of separation increases in size until it reaches the trailing
edge as the shock reaches the most upstream position. At this point the boundarylayer becomes attached and the loop repeats. The unsteady pressure transducers
detected a pressure wave travelling downstream over the upper surface at 0.072U
and a pressure wave travelling upstream at 0.27U over the lower surface. This
results appears to be similar to the mechanism proposed by Lee [12] for the BGK
No. 1 aerofoil.
Garnier and Deck [102] also studied the OAT15A aerofoil, but using a zonal
URANS/LES solver with the flowfield calculated using LES except on the pressure side of the aerofoil, which uses 2D URANS. This study showed pressure
waves travelling downstream over the upper surface and upstream over the lower
surface. The mechanism of Lee [12] was evaluated and gave a higher frequency
using Equation 5.1 in comparison to the power spectral density of the signals.
This was contributed to the lower surface contributing to the mechanism, not
just the upper surface pressure waves, however further research was needed.
Raghunathan et al. [103] presented results on the NACA 0012 aerofoil that
was pitched from 5 to 6 . The flow at 5 was steady and the flow at 6 had shockoscillations with boundary-layer interaction. The paper studied the mechanism
of the onset of buffet and proposed a mechanism behind the origin of shock
oscillation. The shock was strong enough to generate a separation bubble that
caused the shock to move upstream, strengthening and increasing the length of
the separation bubble. When the separation bubble reached the trailing edge,
the upper surface became fully separated. The separated flow extended into
the wake leading to a deflection of the wake, effectively changing the camber
of the aerofoil pushing the shock further upstream. When the shock reached
far enough upstream it weakened and the boundary-layer reattached. As the
separation bubble reattached the effective camber decreased causing the shock to
move downstream once more. As the shock moves downstream it strengthened
84

until it causes separation and the process repeats.

Iovnovich and Raveh [104] published a study on the shock-induced boundarylayer separation of the flow past three aerofoils, the subsonic NACA 0012, the
supercritical RA16SC1, and the thin transonic/supersonic NACA 64A204. The
type of buffet studied is associated with upper-surface shock oscillations at an
angle of attack. It was found that buffet onset could be well predicted using
URANS, along with a fair prediction of the buffet frequency. However the buffet
amplitude and offset conditions could not be accurately predicted. For buffet
onset it was found that the shock location was the most important factor in the
onset mechanism, with the shock location slightly aft of the location of maximum
camber for all the aerofoils, where the local camber is near zero. Buffet onset is
not driven by the bursting of the separation bubble behind the shock. Buffet
offset for all the aerofoils had the flow fully separated aft of the shock with the
shock located near the leading edge and the aerofoils were stalled. The mechanism
proposed is generally in agreement with Raghunthan et al. [103]. The mechanism
proposed for shock-buffet is dependant on a combination of factors, which are
Wedge effect in which the separation region acts as a geometrical wedge
increasing shock strength and changing the shock from normal to oblique
Dynamic effect in which, as the shock moves upstream, the Mach number
in front of the shock increases relative to the shock, resulting in a stronger
shock
Aerofoil curvature effect that weakens the shock as it moves upstream
and the aerofoil becomes more curved

Starting at the most downstream location the separation bubble pushes the shock
upstream strengthening it, with the wedge and dynamic effects dominating over
the curvature effect. When the shock reaches the upstream position the aerofoil
curvature effect now dominates over the wedge and dynamic effects. As the shock
weakens the boundary-layer reattaches followed by the rest of the flow field after
a lag. As the flow field reattaches the shock travels downstream strengthening
and when it reaches the downstream position it causes separation initiating a new
cycle.

85

5.3
5.3.1

BGK No. 1 Aerofoil Results

Grids

Three grids were generated and their sizes are given in Table 5.2. The blocking is
of the C-H type of a similar setup to Xiao et al.[3] with 256 points on the aerofoil
upper and lower surfaces, 128 in the wake and 64 in the normal direction. The
medium grid is shown in Figure 5.5. The coarse and fine grids are a coarsening or
a refinement of the medium grid by a factor of two in each direction. To calculate
Table 5.2: Grid Sizes
Grid
Number of nodes
coarse
24,576
medium
98,304
fine
221,184
the initial cell spacing the following equations where used. They are based on
the turbulent flow over a flat plate to give a single point in the laminar sub-layer
of the boundary-layer. The inputs for the equations are the Reynolds number,
characteristic length and desired y + . First the temperature was calculated using
T =
followed by the pressure
p=

288.15
1 + 1
2

101325

1
1 + 1
2

Then using Sutherlands law the viscosity can be calculated

 32

288.15 + 110
T
= 1.7894 105
288.15
T + 110

(5.2)

(5.3)

(5.4)

The density is calculated using the equation for a perfect gas

p
=
(5.5)
RT
from the equation for the Reynolds number the freestream velocity can be calculated
Re
U =
(5.6)
l
The Schlichting skin-friction correlation for a turbulent boundary-layer is
cf = (2 log10 (Re) 0.65)2.3

(5.7)

then wall shear stress is calculated using

w =

2
cf U
2

86

(5.8)

(a) Blocking

(b) Near aerofoil blocking

(c) Near aerofoil grid (every other line)

Figure 5.5: BGK No. 1 grid details

87

and the friction velocity

Finally the first cell spacing can be calculate using

u =

y=

(5.9)

y+
u

(5.10)

Using a Reynolds number 20 106 , a characteristic length of 1 and a target y +

of 1 gave a initial cell spacing of 1.46 106 . To ensure that the y + < 1 the first
cell space was chosen to be 1 106 which gives a y + 0.71. Figure 5.6 shows
the y + distribution for a calculation at Mach 0.71, Reynold number 20 106 and
6.97 angle of attack.
Upper Surface
Lower Surface
1

0.8

y+

0.6

0.4

0.2

0.2

0.4

0.6

0.8

x/c

Figure 5.6: y + distribution around the BGK No. 1 aerofoil at = 6.97, M=0.71
and Re=20 106

5.3.2

Probe Locations

Table 5.3 gives the location of the pressure taps and the probe locations used in
the analysis of the calculations.

5.3.3

Reduced Frequency Calculation

Reduced frequency k is defined [3] as

k=

f c
U

88

(5.11)

Table 5.3: Pressure taps

Probe
x/c
E
0.200000
F
0.250000
G
0.300000
H
0.350000
I
0.400000
J
0.450000
K
0.500000
L
0.550000
M
0.591000
N
0.632000
O
0.673000
P
0.714000
Q
0.755000
R
0.796000
S
0.837000
T
0.870000
A
1.000000
B1
0.795000
B2
0.795000
B3
0.795000
C1
0.591000
C2
0.591000
C3
0.591000
D1
0.500000
D2
0.500000
D3
0.500000

89

and probe locations

y/c
0.058386542
0.061595008
0.063864008
0.065301239
0.066004328
0.065981880
0.065223873
0.063659132
0.061656490
0.058887668
0.055237915
0.050626419
0.045050010
0.038576253
0.031455696
0.024903920
0.000000000
0.087300000
0.112000000
0.126000000
0.118000000
0.136000000
0.157000000
0.125000000
0.138000000
0.157000000

where f is the frequency, c is the chord and U is the free-stream velocity.

5.3.4

Convergence Studies

A number of convergence studies where completed to find the effect of various

solver parameters. The parameters studied where the number of pseudo steps,
the time step size and the grid density. For all of these studies the test conditions
used were Mach 0.710, angle of attack 6.97 and Reynolds number 20 106 using
the medium grid. The number of explicit steps used to start the calculation was
set to 500 with a CFL of 0.4 followed by 10,000 implicit steps using a CFL of 2.0.
The steady convergence was set to 1.0 105 using the SST turbulence model.

5.3.5

Pseudo Steps Study

Pseudo steps are used to iterate to the real time solution. The number of these
steps can be changed to allow different levels of convergence. The unsteady
parameters used were 6,000 time steps using a time step of 0.05 non-dimensional
time with an unsteady tolerance of 0.001. The number of pseudo steps allowed for
each unsteady time step was varied in this study, the values used were 100, 500,
1,000 and 1,500. Table 5.4 gives the mean lift coefficient, the reduced frequency of
Table 5.4: Results for the pseudo steps study
Pseudo
Mean Cl Reduced
Percentage of
steps
Frequency Unconverged steps
100
0.967
0.161
100%
500
0.994
0.230
28%
1,000
0.994
0.230
12%
1,500
0.993
0.230
7%
Experiment
1.016
0.25
Xiao et al. [3]
1.03
0.16
oscillation and the percentage of time steps that failed to converge to the desired
unsteady tolerance. By increasing the number of pseudo steps the percentage
of unconverged time steps reduced and the mean lift coefficient and reduced
frequency converged, as shown in Figure 5.7. Figure 5.7(a) shows the lift history
and that the 100 pseudo steps caused the lift oscillations to decay to a steady
solution, whereas increasing the number of pseudo steps gives almost identical
unsteady oscillating lift histories. The time-averaged pressure distribution shown
in Figure 5.7(c) is a reasonable match to the experimental results for all numbers
of pseudo steps except for 100 pseudo steps. The skin-friction shown in Figure
5.7(d) for all numbers of pseudo steps the flow separates after the shock and on
average fails to reattach. Figure 5.7(e) shows the unsteady pressure distribution
90

(a) Time history

(b) Lift power spectral density

(c) Time-averaged pressure distribution

(d) Time-averaged skin-friction distribution

(e) Unsteady pressure distribution

(f) Pseudo steps study convergence

Figure 5.7: BGK No. 1 pseudo steps study results

91

at the experimental pressure taps, this again shows that 100 pseudo steps results
in a steady solution. The CFD results show a higher unsteadiness than the
experiment and point towards the shock oscillating further back on the aerofoil.
However this is in-line with the other CFD study by Xiao et al. [3]. For all the
calculations presented from now on 1,000 pseudo steps will be used, since this is a
good compromise between the potential benefit of a greater number of converged
time steps and the cost of the calculation. Figure 5.7(f) shows this compromise
graphically.

5.3.6

Time Step Study

The non-dimensional time steps used were 0.05 (6,000 steps), 0.025 (12,000 steps),
0.0125 (24,000 steps) and 0.00625 (48,000 steps). All the calculations were ran
to 300 non-dimensional time. The unsteady parameters used were an unsteady
tolerance of 0.001 and 1000 pseudo steps.
A time step study was not performed by Xiao et al. [3], but previous studies
on the 18% thick circular-arc aerofoil by Rumsey et al. [96] showed that the
reducing the time step converges the reduced frequency. This trend can also be
seen in the results presented in Table 5.5, with both the reduced frequency and
the mean coefficient of lift converging.

Table 5.5: Results for the time step study

Time Step
Mean Cl Reduced Frequency
0.05
0.994
0.230
0.025
0.998
0.230
0.0125
1.002
0.238
0.00625
1.001
0.238
Experiment
1.016
0.250
Xiao et al. [3]
1.03
0.16

The results of the present study are given Figure 5.8. The time-averaged
pressure distribution shows all the tested time-steps are in good agreement with
the experimental data. The skin-friction shows that all the time steps are on
average fully separated after the shock with no reattachment. The unsteady
pressure distribution shows that as the time step increases the unsteadiness also
increases with reducing time step. Since the cost of the calculation approximately
doubles when the time step is halved a time step of 0.0125 was chosen for the
remaining calculations as a compromise between cost and accuracy.
92

(a) Time history

(b) Lift power spectral density

(c) Time-averaged pressure distribution

(d) Time-averaged skin-friction distribution

(e) Unsteady pressure distribution

Figure 5.8: BGK No. 1 time step study results

93

5.3.7

Grid Refinement Study

For the grid refinement study the three grids given in Table 5.2 where used. From
the previous studies the unsteady parameters used were 12,000 time steps using
a time step of 0.0125 non-dimensional time leading to a total run length of 150
non-dimensional time with an unsteady tolerance of 0.001 and 1000 pseudo steps.
Xiao et al.[3] found the dominant frequency to be independent of the grid
used. Other CFD studies on the 18% thick circular-arc aerofoil found that the
whilst the reduced frequency was relatively insensitive to grid density it still varied
by 2-5%. Changing the grid in the study by Xiao et al. changed the pressure
distribution and the lift coefficient. Table 5.6 shows the results from the grid
study. The mean lift is reasonably well converged, but the reduced frequency is
still converging towards the experimental result.
Table 5.6: Results for grid refinement study
Grid
Mean Cl Reduced frequency
Coarse
1.042
Medium
1.002
0.238
Fine
0.997
0.245
Experiment
1.016
0.250
Xiao et al. [3]
1.03
0.16
More results are given in Figure 5.9, it can be seen that the coarse grid was
unable to capture the unsteady flow. The medium and fine grids both produced
the expected unsteady flow with the amplitude of the lift oscillations being very
similar. The difference in the mean pressure and skin-friction distributions are
minimal between the medium and fine grids, whereas the coarse grid shows a
strong steady shock. The unsteady pressure distributions are similar for the
medium and fine grids and have a higher intensity than the experimental results.
The coarse grid has very little unsteadiness as expected from the lift history. All
the remaining calculations use the medium grid.

5.3.8

Buffet Boundary Estimate

To calculate the buffet boundary using unsteady CFD calculations takes significant computational time with at least three calculations required for each point
along the buffet boundary. In order to increase the speed of the buffet boundary
prediction using steady computations would be advantageous. If the lift is plotted against iteration number useful patterns emerge. Even though these results
are not time accurate and represent no real time history, if the lift coefficient
stabilises after an initial oscillation the flow is steady. If these oscillations per94

(a) Time history

(b) Lift power spectral density

(c) Time-averaged pressure distribution

(d) Time-averaged skin-friction distribution

(e) Unsteady pressure distribution

Figure 5.9: BGK No. 1 grid refinement study results

95

sist then the flow is possibly unsteady. This was shown by Singh [105] using a
NACA0012 aerofoil.
A large number of steady state calculations were ran using the medium grid
to estimate the buffet boundary. The cases were ran to 40,000 implicit iterations
using a CFL of 2.0 and with the lift coefficient output at each iteration. The
lift coefficient history and the residual of the steady calculation can be used
to estimate whether the the flow is steady or unsteady and therefore estimate
the buffet boundary. The test region is shown graphically in Figure 5.10 for
both using the lift coefficient and residual as the estimator. In the figure the
regions of experimental shock oscillation and the experimental buffet boundary
are highlighted. Using the lift coefficient as the estimator the region of shock
oscillation is well captured, but the experimental buffet boundary is not. This is
not surprising as the most likely feature to cause the lift coefficient to oscillate in
this flow is an oscillating shock wave. Using the residual as the estimator results
in significantly more unsteady results. These cover the entire buffet region, but
also extend outside the experimentally estimated region.

(a) Based upon the Cl history

(b) Based upon the residual

Figure 5.10: Buffet boundary estimate

Figure 5.11 shows a comparison of a number of different angles of attack for
Mach number 0.71. From the lift coefficient history it can be seen that the angles
1.396 , 3.017 and 9.0 are steady. The angle of attack 6.97 is oscillating, but
converging to a steady state and as such was defined as a vague result. The
angle of attack of 4.905 was clearly unsteady. The residual plot contains the
data for the Navier-Stokes residual (dot-dash line) and the turbulence model
residual (solid line). For all angles of attack the Navier-Stokes residual converged
to the desired convergence. However after achieving this level of convergence
the angles of attack 1.396 and 3.017 both rapidly converged and the angle
96

4.905 oscillated. The remaining two angles of attack oscillated, but gradually
damped towards further convergence. The turbulence model residual however
gave a slightly different story. The lowest two angles of attack gave a steady
convergence, 9.0 oscillated before damping to a steady convergence. The angle
of attack of 4.905 failed to converge and its turbulence model residual oscillated.
At angle of attack of 6.97 the turbulence model residual oscillated as it converged,
but never settled down to a steady convergence. To give an indiction that the
reason that the calculations are do not converge is not a numerical issue, this
angle of attack sweep was calculated using CFL number of 5.0 and is also shown
in Figure 5.11. The results at the higher CFL number confirm that the angles of
attack of 1.396 and 3.017 are both steady. The results also confirm that angle
of attack of 4.905 fails to converge and therefore is likely to be unsteady. The
higher angles of attack of 6.97 and 9.0 are less clear, as both converge to the
desired level. The 6.97 angle of attack the turbulence model residual oscillated
as it converged and both angles of attack had oscillations as the Navier-Stokes
residual converged. Since their convergence was quite different from the steady
cases these would also be worth looking into. This method has some promise for
a fast initial estimate of whether a calculation is likely to be unsteady, but as the
results look different for each CFL number care must be taken to only compare
iteration histories of calculations with the same CFL number. This leaves some
calculations with convergence histories that are different from the rapid smooth
convergence of steady calculations giving an indication of which calculations are
worth running unsteady.
Figure 5.12 shows the lift curve for the steady calculations. It can be observed
that the CFD lift coefficient values are higher for all angles of attack tested below
6.97 . Stall is at the same point as the experiment at about 5 .

5.3.9

Mach 0.71 Angle of Attack Sweep

An angle of attack sweep for the Mach number 0.71 ran as unsteady calculations
will be presented here and will cover the region before shock-induced oscillation,
shock-induced oscillation and the region after the shock-induced oscillation. For
the unsteady runs only data from 50 to 150 non-dimensional time will be used
for statistical analysis to remove transient effects at the start of the calculations.
Following the convergence studies each angle of attack will be ran for 150 nondimensional time using a time step of 0.0125 and 1000 pseudo steps on the medium
grid.
97

(a) Lift history at CFL of 2.0

Angle of Attack 1.396
Angle of Attack 3.017
Angle of Attack 4.905
Angle of Attack 6.970
Angle of Attack 9.000

2.5

-1

1.5

-2

-3

0.5

-4

-5

-0.5

10000

20000

30000

Angle of Attack 1.396

Angle of Attack 1.396
Angle of Attack 3.017
Angle of Attack 3.017
Angle of Attack 4.905
Angle of Attack 4.905
Angle of Attack 6.970
Angle of Attack 6.970
Angle of Attack 9.000
Angle of Attack 9.000

log(Residual)

CL

(b) Residual history of CFL of 2.0

-6

40000

Desired
Convergence

10000

20000

30000

40000

Iteration Number

Iteration Number

(c) Lift history at CFL of 5.0

(d) Residual history at CFL of 5.0

Figure 5.11: Comparison of Mach 0.71 angle of attack sweep for steady calculations

98

Figure 5.12: Lift curve for the Mach 0.71 angle of attack sweep
1.396 Angle of Attack
This case is the lowest angle of attack ran and converged to a steady solution.
This agrees with the estimate in Figure 5.10 using both the lift coefficient and
residual, this also agrees with the experimental estimate. The CFD study by
Xiao et al. used an angle of attack of 1.3 to account for wind tunnel correction.
The comparison of the pressure distributions is shown in Figure 5.13 and the lift
coefficient is shown in Table 5.7. The pressure distributions for both the original
and the corrected angles of attack fail to match the experimental results. The
lower angle of attack does shift the shock towards the experimental location, but
does not predict its location correctly. This is in-line with the results by Xiao
et al. who used the lagged k- turbulence model and also failed to predict the
experimental shock location. There is no separation in this case from the skin
friction distribution as expected.

Table 5.7: Coefficient of lift comparisons for angle of attack 1.396

Angle of Attack CFD Cl Experimental Cl
1.396
0.701
0.610
1.300
0.683
-

99

(a) Pressure distribution

(b) Skin-friction distribution

Figure 5.13: BGK No. 1 angle of attack 1.396 results

3.017 Angle of Attack
This angle of attack also converged to a steady solution. This matches both
the experimental and the steady calculation estimates in Figure 5.10. Figure
5.14 presents the pressure and skin-friction distributions. The pressure distribution does not match the experiment so well, as the shock is predicted too far
downstream. There is also a small region of separation directly after the shock
which reattaches, the pressure distribution implies that predicted the region of
separation is smaller than in the experiment, due to the late shock.

(a) Pressure distribution

(b) Skin-friction distribution

Figure 5.14: BGK No. 1 angle of attack 3.017 results

100

4.905 Angle of Attack

This angle of attack is unsteady as shown in Figure 5.10 as shown by both the
experimental and the steady calculation estimates. The time history shown in
Figure 5.15(a) is unsteady, due to shock-induced oscillation. The pressure distribution matches the experimental results reasonably well. The time-averaged
skin-friction plot shows a region of separation after the shock and another at the
trailing edge. The unsteady pressure distribution shows a region 0.45 < x/c < 0.7
where the CFD predicts a much larger amount of unsteadiness. This region corresponds to the time-average region of separation in the skin friction plot. It may
be that in the experiment this region of separation was much smaller than that
predicted by the CFD. The magnitude and phase angle (obtained from an FFT
of each probe at the shock oscillation frequency) of the unsteady pressure at the
pressure taps are shown in Figure 5.16. As expected the magnitude has a very
similar profile to the unsteady pressure in Figure 5.15(e). The phase angle has
three regions, for x/c < 0.35 the phase angle is constant. For 0.4 < x/c < 0.6 the
phase angle is constant at a different phase angle to the first region. This region
corresponds to the time-average separation bubble. The final region 0.6 < x/c
is linear with an increasing phase angle and corresponds to the trailing edge
separation. This could indicate a pressure wave travelling with a constant speed.
Figure 5.17 shows the pressure distribution at different time instants during
one period, (a) is for the downstream shock movement and (b) is for the upstream
shock movement. The time instant = 0.0 corresponds to the time at which the
lift is at a maximum and time instant = 1.0 represents a complete cycle. Shock
motion is clearly visible and covers the range 0.34 < x/c < 0.46. The Mach
number contours and streamline plots are shown in Figures 5.18 and 5.19. At
= 0.0 the shock is at x/c = 0.46 with a separation bubble directly behind
the shock and trailing edge separation. Between = 0.1 and = 0.4 the shock
moves upstream and between = 0.6 and = 0.9 the shock moves downstream.
As the shock moves upstream the two regions of separation merge into one fully
separated region and moves upstream with the shock movement. When the shock
has reached furthest upstream point (x/c = 0.34 at = 0.5) the two regions of
separation split to become distinct again. As the shock moves downstream the two
regions of separation become smaller and by = 0.8 the trailing edge separation
has disappeared with only a small separation bubble directly behind the shock
left. At = 1.0 the shock is at its most downstream position, the shock-induced
separation region has grown and a region of trailing edge separation has also
reappeared. The cycle of separated flow and shock movement then repeats. This
pattern of separation can also be seen in the skin-friction plots in Figure 5.17 (c)
and (d).
101

(a) Time history

(b) Lift power spectral density

(c) Time-averaged pressure distribution

(d) Time-averaged skin-friction distribution

(e) Unsteady pressure distribution

Figure 5.15: BGK No. 1 angle of attack 4.905 results

102

(a) Magnitude

(b) Phase angle

Figure 5.16: Pressure wave propagation for angle of attack 4.905

A similar analysis of the calculation was performed to that of Xiao et al.

[3]. As explained in the introduction the mechanism for a self-sustained shock
oscillation expects a pressure wave to travel downstream from the shock in the
region of separation to the trailing edge. Upon arrival another pressure wave
travels upstream outside the separated region. This is shown in Figure 5.3. The
period of the shock oscillation is given by Equation 5.1. By using two-point
cross-covariance the direction and speed of the pressure waves can be calculated
avoiding any empirical formula. This was achieved using the MATLAB command
xcov(x, y). Figure 5.20 presents the cross-covariance of the surface pressure taps
as well as a number of probes that create three paths outside the separated region.
The location of the probes and the pressure taps are given in Table 5.3. A positive
time delay in Figure 5.20(a) indicates that a pressure wave within the separated
region behind the shockwave travels downstream towards the trailing edge. Since
the distance and time delays between the pressure taps are known, then the
wave speed can be calculated. The Figures 5.20 (b), (c) and (d) are for three
different wave paths outside the separated region and all show a negative time
lag indicating that a pressure wave is moving upstream. Again the speeds can be
calculated in the same manner. The wave speeds are given in the final plot in this
figure. The upstream speeds are all almost identical and linear. The downstream
speed is near constant downstream of the mean shock location, but upstream of
the mean shock location anomalous results are obtained. These velocities can be
used in Equation 5.1 to calculate the period of shock oscillation and therefore the
reduced frequency. The reduced frequency for the shock oscillation calculated
from the lift history using a FFT and from Equation 5.1 using the three different
paths are given in Table 5.8. The different reduced frequencies are of a similar
103

(a) Shock movement downstream: Pressure

distribution

(b) Shock movement upstream: Pressure distribution

(c) Shock movement downstream:

friction

(d) Shock movement upstream: Skin-friction

Skin-

Figure 5.17: Unsteady shock movement during one period for angle of attack
4.905

104

(a)

(b) = 0.0

(c) = 0.1

(d) = 0.2

(e) = 0.3

(f) = 0.4

(g) = 0.5

(h) = 0.6

(i) = 0.7

(j) = 0.8

(k) = 0.9

(l) = 1.0

Figure 5.18: Mach number contours at different instants in time for one period
of oscillation for angle of attack 4.905, M = 0.095

105

(a)

(b) = 0.0

(c) = 0.1

(d) = 0.2

(e) = 0.3

(f) = 0.4

(g) = 0.5

(h) = 0.6

(i) = 0.7

(j) = 0.8

(k) = 0.9

(l) = 1.0

Figure 5.19: Streamlines at different instants in time for one period of oscillation
for angle of attack 4.905

106

magnitude, but are not as conclusive as the results presented by Xiao et al. [3].
Table 5.8: Reduced frequency results
Lift history Path 1 Path 2 Path 3
0.215
0.273
0.250
0.239

6.97 Angle of Attack

This angle of attack is estimated to be unsteady in Figure 5.10, both experimentally and from the steady calculations. This is confirmed by the time history
in Figure 5.21(a). The pressure distribution shown in Figure 5.21(c) has good
agreement with the experimental results and the skin-friction indicates that the
flow is on average separated. The unsteady pressure distribution points towards
the shock oscillating further back than the experiment due to higher unsteadiness between x/c = 0.4 and x/c = 0.5. The magnitude and phase angle of the
unsteady pressure taps is given in Figure 5.22. The magnitude has the same
profile as the unsteady pressure distribution, but has a different magnitude to
the experiment and CFD results published by Xiao et al. [3]. The phase angle
has a similar profile to the experiment, but is offset by an average of 92 . This
implies that there are different starting points in the shock oscillation cycle for
the respective results.
Figure 5.23 shows the pressure and skin-friction distributions at different time
instants throughout one cycle. The shock can be seen to oscillate between x/c =
0.1 and x/c = 0.45. The Mach number contours and streamlines throughout a
single cycle are presented in Figures 5.24 and 5.25. At time instant = 0.0 the
shock is at x/c = 0.4 and this shock induces full separation to the trailing edge.
The shock travels upstream for 0.1 < < 0.4 and travels downstream during
0.5 < < 0.9. As the shock travels downstream there is a separation bubble
behind the shock and a small region of separation at the trailing edge. As the
shock travels upstream the flow fully separates behind the shock and does not
split into distinct regions of separation until the shock starts moving downstream
again. This behaviour can be more easily seen in the skin-friction plots. This
cycle of shock motion and separation then repeats.
The cross-covariance of the probes are given in Figure 5.26. A positive time
delay in Figure 5.26(a) indicates a pressure wave travelling downstream in the
separated region and the negative time delays in the Figures 5.26(b)-(d) indicate
that a pressure wave is travelling upstream outside the separated region. As
with the previous angle of attack the wave speeds can be calculated and are
shown in Figure 5.26(e). The upstream wave speeds are all very similar and are
107

(a) Cross-covariance pressure taps

(b) Cross-covariance upstream path 1

(c) Cross-covariance upstream path 2

(d) Cross-covariance upstream path 3

(e) Wave speeds

Figure 5.20: Cross-covariance and wave speed for angle of attack 4.905

108

(a) Time history

(b) Lift power spectral density

(c) Time-averaged pressure distribution

(d) Time-averaged skin-friction distribution

(e) Unsteady pressure distribution

Figure 5.21: BGK No. 1 angle of attack 6.97 results

109

(a) Magnitude

(b) Phase angle

Figure 5.22: Pressure wave propagation

near linear. The downstream wave speeds are linear until x/c 0.65 at which
point the wave speed rapidly increases. Using Equation 5.1 these speeds can
be integrated to calculate the time period of shock oscillation and therefore the
reduced frequencies, which are given in Table 5.9. Using this method the reduced
frequencies obtained are much higher than the one obtained by a FFT of the lift
history that is close to the experimental value of 0.250.
Table 5.9: Reduced frequency results
Lift History Path 1 Path 2 Path 3
0.245
0.360
0.375
0.378

9.0 Angle of Attack

This angle of attack converged to a steady solution. This matches the steady
calculation estimates in Figure 5.10. Figure 5.27 presents the pressure and skinfriction distributions, but there are no experimental results with which to compare. However the results are in-line with experimental observations that the
flow becomes steady due to massive separation downstream of the shock with no
reattachment. This can be clearly seen in the skin-friction plot.

5.4

Summary

It was shown that steady calculations can be used to estimate if a calculation is

likely to be unsteady. This was confirmed by unsteady calculations at a number of
different angles of attack that represent pre-, post- and shock oscillation regimes.
110

(a) Shock movement downstream: Pressure

distribution

(b) Shock movement upstream: Pressure distribution

(c) Shock movement downstream:

friction

(d) Shock movement upstream: Skin-friction

Skin-

Figure 5.23: Unsteady shock movement during one period for angle of attack
6.97

111

(a)

(b) = 0.0

(c) = 0.1

(d) = 0.2

(e) = 0.3

(f) = 0.4

(g) = 0.5

(h) = 0.6

(i) = 0.7

(j) = 0.8

(k) = 0.9

(l) = 1.0

Figure 5.24: Mach number contours at different instants in time for one period
of oscillation for angle of attack 6.97 , M = 0.095

112

(a)

(b) = 0.0

(c) = 0.1

(d) = 0.2

(e) = 0.3

(f) = 0.4

(g) = 0.5

(h) = 0.6

(i) = 0.7

(j) = 0.8

(k) = 0.9

(l) = 1.0

Figure 5.25: Streamlines at different instants in time for one period of oscillation
for angle of attack 6.97

113

(a) Cross-covariance pressure taps

(b) Cross-covariance upstream path 1

(c) Cross-covariance upstream path 2

(d) Cross-covariance upstream path 3

(e) Wave speeds

Figure 5.26: Cross-covariance and wave speed for angle of attack 6.97

114

(a) Pressure distribution

(b) Skin-friction distribution

Figure 5.27: BGK No. 1 angle of attack 9.0 results

Grid refinement, time step and pseudo step studies were completed with a time
step of 120 steps per cycle using 1,000 pseudo steps on the medium grid proving
to be adequate. The unsteady calculations gave more evidence of pressure waves
travelling in the separated region and returning outside the separated region for
the shock oscillation mechanism. This was achieved through the use of two point
cross-covariances at various points in the flow field. The results obtained for the
4.905 angle of attack results showed reasonable agreement, but the results at
6.97 did not agree very well.

115

116

Chapter 6
Static Wing Aeroelasticity
6.1

Introduction

The aeroelastic research wing (ARW-2) was a full-scale right semispan wing,
which was from part of the NASA Drones for Aerodynamic and Structural Testing (DAST) program [106]. The goals of DAST were to validate, 1) system synthesis and analysis developments for active control of aeroelastic response and 2)
analysis techniques for aerodynamic loads prediction. The wing was designed for
flight testing to investigate the use of active control systems to alleviate manoeuvre and gust loads as well as for flutter suppression. A series of windtunnel tests
were carried out at NASA Langley Research Centre using the Langley Transonic
Dynamics Tunnel (TDT) using heavy gas R-12 (also known as Dichlorodifluoromethane [107]) in the late 1980s and early 1990s. After the first set of tests
were completed the planned flight test program was cancelled. As a result a
second windtunnel test was performed to explore the region of large-amplitude
response that had been uncovered by the first set of windtunnel tests.
The wing planform and dimensional data are presented in Figure 6.1 (a). The
wing has an aspect ratio of 10.3 and a leading edge sweep angle of 28.8. The
wing was designed with supercritical aerofoil sections for a lift coefficient of 0.53
at Mach 0.8 with a cruise angle of attack of 1.53 at 46,800ft and dynamic pressure of 126.4psf. The wing has hydraulically driven trailing edge control surfaces,
two inboard and one outboard, their locations are given in Figure 6.1 (b). For
the windtunnel test the wing was instrumented with 182 pressure transducers, 10
accelerometers and strain gauges at the wing root to measure bending moments.
The instrumentation locations are also shown in Figure 6.1 (b). The wing structure was designed iteratively taking into account the load and stiffness reduction
benefits associated with the use of an active control system. This led to a more
flexible wing than traditional designs. Due to this flexibility a jig shape was defined, which is the shape that the wing is manufactured to, so that at cruise the
117

wing will deform into the desired aerodynamic shape. The first four modes in
still air were first-bending (8.1Hz), second-bending (29.7Hz), in-plane (39.9Hz)
and first-torsion (62.6Hz). The wing geometry is given in Sandford et al. [14]
with the coordinates for the jig shape. This report also included data for the
structural model that was generated with the mode shapes given on the upper
surface of the wing.

(a) Sketch of complete wind tunnel model (dimensions are in

inches)

(b) Wing instrumentation (dimensions are in inches)

Figure 6.1: ARW-2 (from ref. [14])

The ARW-2 was found to have an instability in the Mach number range of
0.90 to 0.92 with a frequency that is very close to that of the first-bending mode.
Buffet for the ARW-2 was found to be at its worst at Mach 0.92 with an angle
of attack of 1 and a dynamic pressure of 325.6psf. At the test point the wing
118

tip acceleration reached a peak of 32g. The remainder of the chapter discusses
the previous work, the aerodynamic grid and structural model used in this work,
followed by results of aeroelastic CFD calculations.

6.2

Previous Work

In 1987 Seidel et al. [108] presented experimental results from the first windtunnel
test of the ARW-2 with the objective of providing an early assessment of the wing
aeroelastic stability over a wide range of angles of attack. The experiment varied
the Mach number between 0.6 to 0.9 at dynamic pressures of 50 to 300psf. The
angle of attack was also varied in the range 0 to 2 . They found an unusual
wing instability at a Mach number of 0.9 with a dynamic pressure of 100psf.
Zero damping points where avoided to prevent damage to the wing, since it was
expected to be flight tested. The instability had a frequency of 8.6Hz, which is
very close to the first-bending mode frequency. The frequency of the instability
was found to increase with dynamic pressure to 13Hz at the highest dynamic
pressure tested and was found to be sensitive to angle of attack with minimum
damping near zero degrees. A sustained limit amplitude oscillation at Mach 0.895
with the lowest dynamic pressure was discovered, however the wing became stable
again at Mach 0.9.
Following the first windtunnel test the flight test program was cancelled, so
a second windtunnel test was completed to firmly establish the existence of the
instability that had been found and to gather data to help understand the mechanism forcing the wing oscillations. The results from this second test were published in a number of papers [23, 15, 109]. The second test did not find the
instability predicted in the first windtunnel test, however a region of high dynamic wing response was found instead between Mach numbers 0.92 and 0.94.
Eckstrom et al. [15] published a large amount of data from the second windtunnel
test. This is a brief summary of the findings of the report. At Mach 0.80 there
was little to no response, however there was a high frequency (40Hz) unsteady
flow on the upper surface. At Mach 0.85 there was little response, but a strong
shock was starting to form on the upper surface. There was also flow unsteadiness on both the upper and lower surfaces. At Mach 0.88 there was a moderate
increase in motion for the medium dynamic pressure and a significant increase
in motion for the high dynamic pressure. At this Mach number there was now
a strong oscillating shock with a frequency of about 15Hz. At Mach 0.90 there
was a significant increase in wing motion for all dynamic pressures. There was a
strong shock on both the upper and lower surfaces and for some of the dynamic
pressures the lower surface pressure was lower than the upper surface. This in119

dicated that there was a strong upwards bending and, due to the wing sweep, a
nose-down twist. There were also regions of separation on the lower surface. At
Mach 0.92 the near maximum motion was achieved for all dynamic pressures and
for the high dynamic pressure, at an angle of attack of 1 , wing-tip acceleration
peaked at 32g. Some combinations of angle of attack and dynamic pressure were
not tested due to concerns for the structural safety of the wing. There was a large
region of separation on the lower surface and another region at the trailing edge
of the upper surface. At Mach 0.94 there was a significant decrease in motion for
the high dynamic pressure, whereas for the other dynamic pressures the motion
remained at the same level. At Mach 0.96 for all dynamic pressures the wing
motion decreased substantially. This is summarised in Figure 6.2, which shows
the relative peak wing-tip response as the Mach number is increased. The mech-

Figure 6.2: Maximum PSD peak response from front spar wing-tip accelerometer
(from ref. [15])
anism for the response is related to the chordwise shock movement in conjunction
with flow separation and reattachment on both the upper and lower surfaces of
the wing. After a Mach number of 0.94 the wing response reduced, which was
attributed to full flow separation downstream of the shockwaves with no reattachment. The wing-tip accelerometers show that the frequency of the response
between 8 to 10Hz, which is near the first-bending mode. The amplitude of the
dynamic wing response was found to increase with increasing dynamic pressure.
It was also found to be angle of attack dependant with the maximum motion occurring near an angle of attack of zero degrees and as the angle of attack moved
120

from zero degrees, the maximum motion was shifted to Mach 0.94 at angles of
attack of 2 . Seidel et al. [23] tested two methods to reduce the buffet, the first
was a mean control surface deflection of 6 and the second was a lower surface
spanwise fence. Both methods were found to significantly reduce the maximum
wing response and reduce the Mach number of the peak wing response slightly.
In 1996 Farhangnia et al. [110] published CFD results and compared them
with the experimental data. The calculations were done using ENSAERO which
used the strong conservation-law form of the thin-layer Reynolds-averaged NavierStokes equations with the Baldwin-Lomax turbulence model and a modal approach for the structural solver. Pressure coefficient comparisons were presented
for rigid, static and dynamic cases. An improvement in the results was shown
as the aeroelasticity was taken into account. This paper says that the structural
response is due to the upper shock oscillation.

6.3
6.3.1

Test Case Setup

CFD Grid

The geometry was built from the design coordinates from Sandford et al. [14]
and a C-H type blocking was used to block the wing. The grid size is 2,333,940
nodes and is shown in Figure 6.3.

6.3.2

Structural Model

The structure and the mode shapes are given in Sandford et al. [14] in the form
of coordinates and displacements at 100 points on the upper surface of their finite
element model. A shell model (referred to as the shell model) was generated to
replicate the shell model presented in Sandford et al. [14] (referred to as the
report model). The resulting model is compared with the model [14] in both
frequencies and mode shapes, with the frequencies given in Table 6.1.
Table 6.1: Comparison of the structural model frequencies
Mode Model [14] (Hz) Shell Model (Hz)
1
8.0935
8.0272
2
29.2299
29.0527
3
32.7980
32.6560
4
60.8920
61.4565
The frequencies are all within 1% of the model [14]. The mode shape comparison of the upper surfaces is shown in Figure 6.4. It can be seen that the shell
121

(a) Near wing blocking

(b) Surface blocking

(c) Surface grid

Figure 6.3: ARW-2 grid details

122

model mode 4 has less torsion compared to the report model, however all the
other modes are captured well.

(a) Mode 1

(b) Mode 2

(c) Mode 3

(d) Mode 4

Figure 6.4: Comparison of the mode shapes. Grey - Undeformed shell model,
Red - Report model mode shapes, Green - Shell model mode shapes
The shell model was converted into a beam model using the leading and
trailing edge deflections. The mode shapes are almost exactly the same as the
original data, except for the torsion mode. The torsion mode from the shell
model has the section stretching which cannot be recreated using a beam model,
however the wing is unlikely to stretch to the degree of the shell model, so this
is believed to be an artifact of the finite element model that is presented in the
reference and recreated by the shell model. This is backed up by the paper by
Byrdsong et al. [111] that states that there was sufficient chordwise rigidity to
prevent chordwise bending. The beam model along with a rigid direction that
is parallel with the majority of the ribs was used to generate the mode shapes
123

on the aerodynamic surface using the method discussed in Chapter 3. The mode
shapes defined on the surface mesh are shown in Figure 6.5.

(a) Mode 1

(b) Mode 2

(c) Mode 3

(d) Mode 4

Figure 6.5: Comparison of the mode shapes with the beam model. Grey - Undeformed shell model, Red - Report model mode shapes, Green - Shell model mode
shapes, Pale Blue - Undeformed CFD mesh, Blue - Beam model mode shapes on
the CFD mesh

6.4

Results

A number of static calculations were performed to compare the aerodynamics

and the static deflections. The flow conditions for the test cases are given in
Table 6.2. The Reynolds number is based upon the mean aerodynamic chord.
Each calculation was run for 500 explicit steps at a CFL of 0.4 followed by 40,000
implicit steps at a CFL of 2.0. The desired convergence is set to 1 105 using
the SST turbulence model. Since the windtunnel data was collected using heavy
124

gas R-12, the ratio of specific heats was set to 1.14 as per the data given in
Cole and Rivera [107].

Case
1
2
3
4

6.4.1

Mach
number
0.80
0.85
0.92
0.92

Table 6.2: ARW-2 test conditions

Angle of Reynolds Dynamic
Eckstrom
Byrdsong
attack
number
pressure et al. [15] tab et al. [111, 112] tab

6
0.0
2.41 10
123.6psf
91
82
1.0
2.41 106 134.6psf
243
914
0.0
2.60 106 152.5psf
98

6
2.0
2.60 10
154.8psf
111
-

Case 1: Mach Number 0.80, Angle of Attack 0.0

The experiments by Eckstrom et al. [15] showed that this test condition was
free from buffet and there is experimental data for the static deflections from
Byrdsong et al. [111, 112]. Byrdsong et al. did not record any results if the
deflections were not static or quasi-static, implying that this test condition is
static.
The pressure distributions are shown in Figure 6.6 for both the static and
rigid cases. The effect of not taking the structural flexibility into account is considerable with a strong shockwave generated in the outboard sections. By taking
into account the structural flexibility the pressure distributions match the experimental data better. The noticeable discrepancy is a large suction peak in the
static results that are not shown in the experimental results. This improvement
appears to be due a nose-down twist, however the lower surface is not matched
as well as expected. This is likely to be due to an insufficient amount of twist
being generated.
The Mach number contours for the rigid and static solutions are shown in
Figure 6.7 side by side. The sections at 10% span are very similar, since the
effect of the static deflections will have a minimal effect so close to the root of
the wing. The section at 56% span has the static effects starting to show as
the static section has risen slightly. There are also small differences appearing
in the Mach contours. At 80% span the deflection of the static section is more
pronounced and the shockwave seen in the rigid section has weakened significantly
and shifted to the leading edge in the static solution. Finally in the 95% sections,
again there is significant deflection in the static case. The Mach contours have a
strong shock in the rigid case that has shifted forwards to the leading edge and
weakens considerably in the static case. This implies that the large suction peak
is possibly a weak shockwave.
125

The streamlines are shown in Figure 6.8 for both the rigid and static cases.
For both the rigid and static cases, the flow over the entire span is attached. This
is backed up by the experimental observations by Eckstrom et al. [15].
Figure 6.9 shows the 3D representation of the wing in its rigid and deflected
states. The plots showing the front and rear spar deflections show a very good
match with the experimental data of Byrdsong et al. [111, 112]. The twist angle
is shown in Figure 6.10 along with the change in twist between the rigid and
static cases. Although the deflections match the experimental values well, the
twist does not match as well. The change in twist for the CFD follows a different
profile to the experimental result as seen in Figure 6.10(b). This observation
matches with the pressure distribution results, which appears to have insufficient
nose down twist. This could be due to the errors introduced due to the modelling
of the structure, both by Sandford et al. [14] and subsequently by the author.
In the paper by Sandford et al. the frequencies of the structural model did not
match those obtained by the experiment exactly. Then the shell model generated
by the author was within 1% of the Sandford et al. model and mode 4 did not
recreate the same level of twist as the Sandford et al. model. This could be the
reason for the lack of agreement for the twist.

6.4.2

Case 2: Mach Number 0.85, Angle of Attack 1.0

The experimental results of Eckstrom et al. [15] report no large scale motion
at these test conditions. There are also static deflection results available from
Byrdsong et al. [111, 112], implying it is a static case.
The sectional results are shown in Figure 6.11. As seen before the rigid results
produce a strong shockwave and much higher pre-shock levels. The static results
are a significant improvement over the rigid results and match the experimental
results well. Again this improvement is indicative of a nose-down twist.
The Mach number contours are shown in Figure 6.12 and a strong shockwave
can be seen to have formed on the upper surface throughout the entire span for
both rigid and static cases. However the shock weakens as the tip is approached
for the static case, due to the deflection of the wing and the slight nose down
twist.
The streamline plots are given in Figure 6.13. The rigid case does have regions
of separated flow caused by the shock wave especially at the midspan. The static
case shows no regions of separation for any spanwise station. The experiments
by Eckstrom et al. [15] did not report any regions of separated flow, but did
comment that there were regions of unsteady flow on both the upper and lower
surfaces. It is not known whether this unsteadiness is due to shock oscillation or
separated flow.
126

(a) 70.7% Section pressure distribution

(b) 87.1% Section pressure distribution

(c) 97.2% Section pressure distribution

Figure 6.6: Case 1: Sectional results

127

(a) 10% Section rigid solution

(b) 10% Section static solution

(c) 56% Section rigid solution

(d) 56% Section static solution

(e) 80% Section rigid solution

(f) 80% Section static solution

(g) 95% Section rigid solution

(h) 95% Section static solution

Figure 6.7: Case 1: Mach number contours, M = 0.045

128

(a) 10% Section rigid solution

(b) 10% Section static solution

(c) 56% Section rigid solution

(d) 56% Section static solution

(e) 80% Section rigid solution

(f) 80% Section static solution

(g) 95% Section rigid solution

(h) 95% Section static solution

Figure 6.8: Case 1: Streamlines

129

(a) Wing deflection comparison

(b) Front spar

(c) Rear spar

Figure 6.9: Case 1: Spar deflections

(a) Twist

(b) Change in twist between rigid and

static cases

Figure 6.10: Case 1: Wing twist

130

The spar deflections in Figure 6.14 do not match the experimental data of
Byrdsong et al. [111, 112] as well as the previous case. This is mainly due to the
dynamic pressure used that matches the experimental data by Eckstrom et al.
[15], but is about halfway between two of the dynamic pressures used to record
the static deflections by Byrdsong et al.. As such it would be expected that the
static deflections do not match.
The twist results are shown in Figure 6.15 again showing a discrepancy between the CFD and the experimental results. This will be due to the dynamic
pressure used and the failure of the structural model to fully capture the amount
of twist in mode 4.

(a) 70.7% Section pressure distribution

(b) 87.1% Section pressure distribution

(c) 97.2% Section pressure distribution

Figure 6.11: Case 2: Sectional results

131

(a) 10% Section rigid solution

(b) 10% Section static solution

(c) 56% Section rigid solution

(d) 56% Section static solution

(e) 80% Section rigid solution

(f) 80% Section static solution

(g) 95% Section rigid solution

(h) 95% Section static solution

Figure 6.12: Case 2: Mach number contours, M = 0.045

132

(a) 10% Section rigid solution

(b) 10% Section static solution

(c) 56% Section rigid solution

(d) 56% Section static solution

(e) 80% Section rigid solution

(f) 80% Section static solution

(g) 95% Section rigid solution

(h) 95% Section static solution

Figure 6.13: Case 2: Streamlines

133

(a) Wing deflection

(b) Front spar

(c) Rear spar

Figure 6.14: Case 2: Spar deflections

(a) Twist

(b) Change in twist between rigid and

static cases

Figure 6.15: Case 2: Wing twist

134

6.4.3

Case 3: Mach Number 0.92, Angle of Attack 0.0

This case did not produce static deflections in the experiments and had near
maximum dynamic wingtip motions [15]. As such there are no static results to
compare against by Byrdsong et al. [111, 112], since they did not record any
dynamic deflections. The static calculation lift and residual histories are shown
in Figure 6.16. These show that the calculation has not converged to a steady
solution and is therefore is likely to be an unsteady case. The lift history shows
oscillations that indicate a possible shock-oscillation.

(a) Lift History

(b) Residual History

Figure 6.16: Case 3: Convergence

6.4.4

Case 4: Mach Number 0.92, Angle of Attack 2.0

This case like the last also has large dynamic wingtip motion in the experiment
[15] and therefore has no static deflection results to compare with from Byrdsong
et al. [111, 112]. This experimental observation is backed up by the calculation
lift and residual histories in Figure 6.17. The lift history has an oscillation,
which indicates a possible shock-oscillation. The residual history shows that the
static calculation converges to about -4 and then converges no further, which also
indicates that this case is unsteady.

6.4.5

Buffet Search

A similar search to the one performed for the BGK number one aerofoil was
performed for the ARW-2 to find the most likely buffet cases using steady static
calculations. The calculations were performed at angle of attack 0.0 , Reynolds
number 2.5 106 , dynamic pressure 150 psf and the Mach number was varied
135

(a) Lift History

(b) Residual History

Figure 6.17: Case 4: Convergence

between 0.80 and 0.94. The results are shown in Figure 6.18. Mach 0.85 has a
sudden jump due to the calculation being restarted, but it otherwise converged
well. However if that case did not require restarting it is likely it would have
converged before the Mach 0.88 case. The time history for the lift shows little
variation unlike the BGK No. 1 aerofoil. However of the unsteady cases tested
in the BGK No. 1 aerofoil case, one was obviously unsteady and the other was
a more subtle case in the buffet search. As such the most likely candidates for
unsteady cases are Mach 0.92 and Mach 0.94, since Mach 0.94 did not converge
and Mach 0.92 converged much slower than the lower Mach numbers and the lift
history shows oscillations.

(a) Convergence history

(b) Lift history

Figure 6.18: Buffet search results

136

6.5

Summary

It was shown that for the static case the AMB solver matched the experiments
well. However when the flow conditions where expected to lead to an unsteady
result the solver was unable to match the experiment using steady static calculations. This is unsurprising and can be rectified by running computationally
expensive unsteady calculations in the future. A buffet search was completed to
find the most promising case for buffet calculations, leading to Mach 0.92 and 0.94
being the most likely candidates. Cases 3 and 4 are also promising candidates for
buffet calculations as these also failed to converge and hint at shock-oscillation.
Although all the preparation was completed for buffet calculations on the ARW
wing, the actual unsteady calculation was not ran. This was because the calculation was unable to be ran in a reasonable time period on the computational
resources available. To run this calculation in a reasonable time period more computational nodes are required and the grid needs to be split into more domains
to allow it to be well load balanced on the increased number of nodes.

137

138

Chapter 7
Conclusions
7.1

Conclusions

The objective of this work was to investigate the influence of transformation

methods on the flutter boundary and to consider the simulation of shock-induced
buffet of a transport wing.
The formulation of the CFD, structural and Schur solvers was presented in
Chapter 2.
The transformation problem was presented in Chapter 3, along with the different types of structural models and a number of transformation methods. A
new method based on an assumption of beam stick models that the section perpendicular to the beam is rigid was presented. In this method the direction that
is assumed to be rigid is defined. These rigid ribs should ideally be the same
direction as the wing ribs and perpendicular to the beam, but can be varied to
assess the effect of the transformation scheme on the predictions.
The transfer methods IPS, IIM, CVT and RBF were tested on the Goland
and MDO wings in Chapter 4. Through the use of different structural models it
was found that the in-plane component of a transformation method dominated
the calculated displacements. It was also shown that the transformation methods
are sensitive to the structural grid density, with only IIM relatively unaffected.
The local method CVT suffers from discontinuities in slope at the structural
element edges, due to the formulation only being C 0 continuous. The other
local method tested, IIM, is also only C 0 continuous, but did not show any
significant discontinuities. This is probably due to the higher order interpolation
used. Both global methods tested suffered due to the coarse grids used and tended
to introduce oscillations to the section shape. Another important consideration
is to avoid transformation methods extrapolating, and wherever possible to do
the extrapolation as a pre-processing step. This can be achieved in NASTRAN
through the use of rigid bars (RBAR). A method was proposed for extrapolating
139

beam stick models by specifying a direction for extrapolation and it was shown
to perform as well as IIM. It was also shown that when the rigid direction was
not close to perpendicular to the beam, the section shapes were worse.
The effect of the transformation methods on the flutter boundary of the
Goland and MDO wings was also investigated in Chapter 4. It was shown that
the local methods had less spread in their prediction of the flutter point as the
structural model was varied than the global methods. More importantly there
was more spread in the flutter point when the transformation methods extrapolated. This adds weight to the previous conclusion that the transfer methods
should not extrapolate and that extrapolation should be done as a pre-processing
step. It is important to stick to the structural model assumptions and extrapolate perpendicular to the beam, as this was shown to reduce the spread in the
predicted flutter point. Since it was shown that the choice of transfer method can
have a significant effect on the predicted flutter point, it is therefore advisable to
use multiple transfer methods to build confidence in the results obtained.
To facilitate buffet calculations the AMB solver was verified for shock-induced
buffet. This was completed using the BGK No. 1 aerofoil and a number of convergence studies in Chapter 5. A buffet search using steady state calculations
was performed to estimate the buffet boundary. Two methods were used, the
first was to use the lift history and the second was to use the calculation residual.
The lift history method matched the experimental shock oscillation region very
well, but performed less well against the experimental buffet onset. The residual
method performed better at predicting buffet onset, but predicted a slightly larger
region than the experiment. This did provide evidence that steady state calculations could be used to predict shock-oscillation cases and potentially buffet onset
boundary. A number of unsteady calculations were used to check the results
of the steady state predictions and covered the pre, post and shock-oscillation
regimes. The results from these calculations matched the steady state predictions and the experimental data. For the two cases that proved to be unsteady,
the mechanism behind the shock-oscillation was investigated. Through the use of
cross-covariance and probes inside and outside the boundary-layer, pressure waves
could be detected travelling downstream inside the boundary-layer and upstream
outside the boundary-layer. The time lags of these pressure waves could be calculated and they correlated to the shock-oscillation period for the 4.905 angle of
attack case.
Finally the static aeroelasticity of the ARW-2 wing was calculated to verify
the AMB solver for static aeroelasticity in Chapter 6. For cases that had been
shown to be static experimentally the solver produced good results providing reasonable agreement with experimental pressure distributions and wing deflections.
140

The results did not match as well as expected for wing twist, this is thought to
be due to the structural model failing to capture the torsion mode well. The
experimentally unsteady cases failed to converge when computed using steady
state calculations. This led to a similar buffet search to the BGK No. 1 aerofoil,
which predicted the Mach numbers 0.92 and 0.94 as the most likely candidates for
unsteady cases for an angle of attack of zero degrees, Reynolds number 2.5 106
and dynamic pressure 150psf. This was consistent with the experimental data.

7.2

Future Work

Buffet calculations can now be performed on the ARW-2 wing. This would be a
combination of the experience gained from the 2D aerofoil buffet calculations and
the static aeroelastic work already preformed on the ARW-2 wing. It has already
been commented that the literature indicates the buffet is due to shock-oscillation
on the upper surface, but the mechanism behind it has not been studied. As such
it would be interesting to investigate the buffet mechanism on the wing and see
if it is the same as the aerofoil case with propagating pressure waves.
There has been a lot of work in recent years on the RBF transfer method
for aeroelasticity and grid deformation. The most interesting development is
the pointwise partition of unity and has been shown to localise the method. It
would be an interesting extension of the current work to see how this would effect
the extrapolation characteristics of RBF and if it prevents the oscillation in the
section shapes that were be generated. The effect of the pointwise partition of
unity on flutter prediction variability would also be an interesting study.
Although the rigid ribs transformation model was useful in testing how varying
the direction of the ribs affects the flutter prediction, it requires extending for
use with more complex geometries. This can be achieved in a similar way to that
used for CVT, by using a hierarchy of components and blending functions.
An extension of the aerofoil buffet work would be to model the turbulence
using DES or LES to investigate whether significant improvements could be made
by using these more expensive options over URANS.

141

142

Bibliography
[1] Spalart P.R. Strategies for Turbulence Modelling and Simulations. International Journal of Heat and Fluid Flow, 21:252263, 2000.
[2] Rendall T.C.S. and Allen C.B. Unified Fluid-Structure Interpolation and
Mesh Motion using Radial Basis Functions. International Journal for Numerical Methods in Engineering, 74:15191559, 2008.
[3] Xiao Q., Tsai H.M., and Liu F. Numerical Study of Transonic Buffet on a
Supercritical Airfoil. AIAA Journal, 44(3):620628, March 2006.
[4] Canfield R.A., Morgenstern S.D., and Kunz D.L. Alleviation of BuffetInduced Vibration using Piezoelectric Actuators. Computers and Structures, 86:281291, 2008.

[5] NASA. Website, 1989. http://www.dfrc.nasa.gov/Gallery/Photo/F-18HARV/HTML/EC89[6] ESDU International plc. An Introduction to Aircraft Buffet and Buffeting.
Technical Report 87012, ESDU International plc, 2006.
[7] Woodgate M.A., Badcock K.J., Rampurawala A.M., Richards B.E., Nardini
D., and Henshaw M.J. de C. Aeroelastic Calculations for the Hawk Aircraft
using the Euler Equations. Journal of Aircraft, 42(4):10051012, 2005.
[8] Fukushima
C.
and
Westerweel
J.
http://en.wikipedia.org/wiki/File:Jet.jpg.

Website,

2007.

[9] Wilcox D.C. Turbulence Modelling for CFD. DCW Industries, Inc., 3
edition, 2006.
[10] Rampurawala A.M. Aeroelastic Analysis of Aircraft with Control Surfaces
using CFD. PhD thesis, University of Glasgow, 2006.
[11] Samareh J.A. and Bhatia K.G. A Unified Approach to Modelling Multidisciplinary Interactions. AIAA Paper AIAA-2000-4704, 2000.
143

[12] Lee B.H.K.

Oscillatory Shock Motion Caused by Transonic Shock
Boundary-Layer Interaction. AIAA Journal, 28(5):942944, 1990.
[13] McDevitt J.B., Levy Jr. L.L., and Deiwert G.S. Transonic flow about a
thick circular-arc airfoil. AIAA Journal, 14(5):606613, 1976.
[14] Sandford M.C., Seidel D.A., and Spain C.V. Geometrical and structural
properties of an aeroelastic research wing. Technical Report NASA TM
4110, NASA, Langley Research Center, Hampton, Virginia, 1989.
[15] Eckstrom C.V., Seidel D.A., and Sandford M.C. Measurements of unsteady
pressure and structural response for an elastic supercritical wing. Technical Report NASA TP 3443, NASA, Langley Research Center, Hampton,
Virginia, November 1994.
[16] Garrick I.E. and Reed III W.H. Historical Development of Aircraft Flutter.
Journal of Aircraft, 18(11):897912, 1981.
[17] Collar A.R. The First Fifty Years of Aeroelasticity. Aerospace, 5(2):1220,
February 1978.
[18] Wright J.R. and Cooper J.E. Introduction to Aircraft Aeroelasticity and
Loads. Wiley, 2007.
[19] Walters J.M. and Sumwalt III R.L. Aircraft Accident Analysis: Final Reports. McGraw-Hill, 2000.
[20] Bunton R.W. and Denegri Jr. C.M. Limit Cycle Oscillation Characteristics
of Fighter Aircraft. Journal of Aircraft, 37(5):916918, 2000.
[21] Fung Y.C. An Introduction to the Theory of Aeroelasticity. Dover Publications Inc., 2002.
[22] Triplett W.E. Pressure Measurements in Twin Vertical Tails in Buffeting
Flow. Journal of Aircraft, 20(11):920925, 1983.
[23] Seidel D.A., Eckstrom C.V., and Sandford M.C. Transonic Region of High
Dynamic Response Encountered on an Elastic Supercritical Wing. Journal
of Aircraft, 26(9):870875, 1989.
[24] Sheta EF. Alleviation of vertical tail buffeting of f/a-18 aircraft. Journal
of Aircraft, 42(2):322330, 2004.
[25] Wickramasinghe V.K., Chen Y., and Zimcik D.G. Experimental Evaluation of an Advanced Buffet System on Full-Scale F/A-18 Fin. Journal of
Aircraft, 44(3):733740, 2007.
144

[26] Breitsamter C. and Laschka B. Fin Buffet Pressure Evaluation Based on

Measured Flowfield Velocities. Journal of Aircraft, 35(5):806808, 1998.
[27] Breitsamter C. and Schmid A. Airbrake-Induced Fin-Buffet Loads on
Fighter Aircraft. Journal of Aircraft, 45(5):16191630, 2008.
[28] Moses R.W. and Huttsell L. Fin Buffeting Features of an Early F-22 Model.
AIAA Paper AIAA 200-1695, 2000.
[29] Anderson W.D., Patel S.R., and Black C.L. Low-Speed Wind Tunnel Buffet
Testing on the F-22. Journal of Aircraft, 43(4):879885, 2006.
[30] Pototzky A.S. and Moses R.W. A Tail Buffet Loads Prediction Method for
Aircraft at High Angles of Attack. Technical Report 20050217424, NASA
Langley Research Center, 2005.
[31] ESDU International plc. Fluctuating Loads and Dynamic Response of Bodies and Structures in Fluid Flows - Background Information. Technical
Report 77032, ESDU International plc, 2007.
[32] Henshaw M.J de C., Badcock K.J, Vio G.A., Allen C.B., Chamberlain J.,
Kaynes I., Dimitriadis G., Cooper J.E., Woodgate M.A., Rampurawala
A.M., Jones D., Fenwick C., Gaitonde A.L., Taylor N.V., Amor D.S., Eccles T.A., and Denley C.J. Nonlinear aeroelastic prediction for aircraft
applications. Progress in Aerospace Sciences, 43:65137, 2007.
[33] Bhatia K.G. Airplane Aeroelasticity: Practice and Potential. Journal of
Aircraft, 40(6):10101018, 2003.
[34] Yurkovich R. Status of Unsteady Aerodynamic Prediction for Flutter of
High-Performance Aircraft. Journal of Aircraft, 40(5):832842, 2003.
[35] Dowel E.H., Thomas J.P., Hall K.C., and Denegri Jr. C.M. Theoretical Predictions of F-16 Fighter Limit Cycle Oscillations for Flight Flutter Testing.
Journal of Aircraft, 46(5):16671672, 2009.
[36] Sobester A. and Keane A.J. Multidisciplinary design optimisation of UAV
airframes. In 47th AIAA/ASME/ASCE/AHS/ASC Structures, Structural
Dynamics and Materials Conference, number 47, page 13, Newport, Rhode
Island, May 2006. AIAA.
[37] MSC Software. Aeroelastic Analysis Users Guide, 2004.
145

[38] Farhat C., Geuzaine P., and Brown G. Application of a Three-Field Nonlinear Fluid-Structure Formulation to the Prediction of the Aeroelastic Parameters of an F-16 Fighter. Computers and Fluids, 32:329, 2003.
[39] Badcock K.J., Rampurawala A.M., and Richards B.E. Intergrid Transformation for Aircraft Aeroelastic Simulations. AIAA Journal, 42(9):1936
1939, 2004.
[40] Badcock K. J., Woodgate M. A., and Richards B. E. Direct Aeroelastic
Bifurcation Analysis of a Symmetric Wing Based on the Euler Equations.
Journal of Aircraft, 42(3):731737, 2005.
[41] Badcock K.J. and Woodgate M.A. Prediction of Bifurcation Onset of Large
Order Aeroelastic Models. In AIAA-2008-1820 in the proceedings of the
49th Structural Dynamics and Materials Conference, AIAA, Schaumberg,
Illinois, 2008.
[42] Shang J.S. Computational Fluid Dynamics Application to Aerospace Science. The Aeronautical Journal, 113(1148):619632, 2009.
[43] Batina J.T. Unsteady Euler Airfoil Solutions Using Unstructured Dynamic
Meshes. AIAA Journal, 28(8):13811388, August 1990.
[44] Farhat C., Degand C., Koobus B., and Lesoinne M. Torsional Springs
for Two-Dimensional Dynamic Unstructured Fluid Meshes. Computational
Methods in Applied Mechanics and Engineering, 163:231245, 1998.
[45] Blom F.J. Considerations on the Spring Analogy. International Journal
for Numerical Methods in Fluids, 32:647668, 2000.
[46] Dubuc L., Cantariti F., Woodgate M., Gribben B., Badcock K.J., and
Richards B.E. A Grid Deformation Technique for Unsteady Flow Computations. International Journal for Numerical Methods in Fluids, 38:285311,
2000.
[47] Chen P.C. and Jadic I. Interfacing of Fluid and Structural Models via Innovative Structural Boundary Element Method. AIAA Journal, 36(2):282
287, 1998.
[48] Liu X., Qin N., and Xia H. Fast Dynamic Grid Deformation based on
Delaunay Graph Mapping. Journal of Computational Physics, 211:405
423, 2006.
146

[49] de Boer A., van der Schoot M.S., and Bijl H. Mesh Deformation based on
Radial Basis Function Interpolation. Computers and Structures, 85:784
795, 2007.
[50] Rendall T.C.S. and Allen C.B. Parallel Efficient Mesh Motion using Radial
Basis Functions with Application to Multi-Bladed Rotors. International
Journal for Numerical Methods in Engineering, 81:89105, 2009.
[51] Farhat C. and Lesoinne M. Two Efficient Staggered Algorithms for the Serial and Parallel Solution of Three-Dimensional Nonlinear Transient Aeroelastic Problems. Computer Methods in Applied Mechanics and Engineering,
182:499515, 2000.
[52] Harder R.L. and Desmarais R.N. Interpolation Using Surface Splines. Engineering Notes, pages 189191, February 1972.
[53] Menter F.R. Two-Equation Eddy-Viscosity Turbulence Models for Engineering Applications. AIAA Journal, 32(8):15981605, 1994.
[54] Menter F.R., Kuntz M., and Langtry R. Ten Years of Industrial Experience
with the SST Turbulence Model. In Hanjalic K., Nagano Y., and Tummers
M., editors, Turbulence, Heat and Mass Transfer 4. Begell House Inc., 2003.
[55] Badcock K.J., Richards B.E., and Woodgate M.A. Elements of Computational Fluid Dynamics on Block Structured Grids using Implicit Solvers.
Progress in Aerospace Sciences, 36:351392, 2000.
[56] Bekas C. and Saad Y. Computation of Smallest Eigenvalues using Spectral
Schur Complements. SIAM Journal on Scientific Computing, 27(2):458
481, 2005.
[57] Guruswamy G.P. and Byun C. Direct Coupling of Euler Flow Equations
with Plate Finite Element Structures. AIAA Journal, 33(2):375377, 1995.
[58] Felker F.F. Direct Solution of Two-Dimensional Navier-Stokes Equations
for Static Aeroelasticity Problems. AIAA Journal, 31(1):148153, 1993.
[59] H
ubner B., Walhorn E., and Dinkler D. A Monolithic Approach to FluidStructure Interaction using Space-Time Finite Elements. Computer Methods in Applied Mechanics and Engineering, 193:20872104, 2004.
[60] Smith M.J., Hodges D.H., and Cesnik C.E.S. An Evaluation Of Computational Algorithms to Interface Between CFD and CSD Methodologies.
Technical Report WL-TR-96-3055, U.S. Air Force Research Lab., November 1995.
147

[61] Kamakoti R. and Shyy W. Fluid-Structure Interaction for Aeroelastic Applications. Progress in Aerospace Science, 40(8):535558, 2004.
[62] Sadeghi M., Lui F., Lai K.L., and Tsai H.M. Application of ThreeDimensional Interfaces for Data Transfer in Aeroelastic Computations. In
22nd Applied Aerodynamics Conference and Exhibit, number 2004-5376,
August 2004.
[63] Done G.T.S. Interpolation of Mode Shapes: A Matrix Scheme using TwoWay Spline Curves. Aeronautical Quarterly, XVI:333349, 1965.

[64] McNeal. NATRAN. Website. http://www.mscsoftware.com/products/msc_nastran.cfm?Q=131&Z

[65] ZONA
Technology
Inc.
ASTROS.
http://www.zonatech.com/ASTROS.htm.

Website.

[66] Hardy R.L. Multiquadric Equations of Topography and Other Irregular

Surfaces. Journal of Geophysical Research, 76(8):19051915, March 1971.
[67] Duchon J. Splines Minimising Rotational-Invariant Semi-Norms in Sobolev
Spaces. In Schempp W. and Zeller K., editors, Constructive Theory of
Functions of Several Variables, Oberwolfach, 1976, pages 85100. SpringerVerlag, 1977.
[68] Franke R. Scattered Data Interpolation: Tests of Some Methods. Mathematics of Computation, 38(157), 1982.
[69] Murti V. and Valliappan S. Numerical Inverse Isoparametric Mapping in
Remeshing and Nodal Quantity Contouring. Computers and Structures,
22(6):10111021, 1986.
[70] Appa K. Finite-Surface Spline. Journal of Aircraft, 26(5):495496, 1989.
[71] Kansa E.J. Multiquadrics - A Scattered Data Approximation Scheme with
Applications to Computational Fluid-Dynamics - I: Surface Approximations and Partial Derivative Estimates. Computers and Mathematics with
Applications, 19(8/9):127145, 1990.
[72] Pidaparti R.M.V. Structural and Aerodynamic Data Transformation Using
Inverse Isoparametric Mapping. Journal of Aircraft, 29(3):507509, 1992.
[73] Samareh J.A. Use of CAD Geometry in MDO. AIAA Paper AIAA-96-3991,
1996.
148

[74] Cebral J.R. and Lohner R. Conservative Load Projection and Tracking for
Fluid-Structure Problems. AIAA Journal, 35(4):687692, April 1997.
[75] Beckert A. and Wendland H. Multivariate Interpolation for Fluid-StructureInteraction Problems using Radial Basis Functions. Aerospace Science and
Technology, 5:125131, 2001.
[76] Ahrem R., Beckert A., and Wendland H. A Meshless Spatial Coupling
Scheme for Large-Scale Fluid-Structure-Interaction Problems. Computer
Modelling in Engineering and Sciences, 12:121136, 2006.
[77] Ahrem R., Beckert A., and Wendland H. Recovering Rotations in Aeroelasticity. Journal of Fluids and Structures, 23:874884, 2007.
[78] Rendall T.C.S. and Allen C.B. Improved Radial Basis Function FluidStructure Coupling via Efficient Localised Implementation. International
Journal for Numerical Methods in Engineering, 78:11881208, 2009.
[79] Goura G.S.L. Time Marching Analysis of Flutter using Computational
Fluid Dynamics. PhD thesis, University of Glasgow, 2001.
[80] Burnett D.S. Finite Element Analysis from Concepts to Applications.
Addison-Wesley Publishing Company, 1987.
[81] Smith M.J., Hodges D.H., and Cesnik C.E.S. Evaluation of Computational
Algorithms Suitable for Fluid-Structure Interactions. Journal of Aircraft,
37(2):282294, 2000.
[82] Samareh J.A. Aeroelastic Deflection of NURBS Geometry. In 6th International Conference on Numerical Grid Generation in Computational Field
Simulation, July 1998.
[83] Wendland H. Fast Evaluation of Radial Basis Functions: Method Based
on Partition of Unity. Approximation Theory X: Wavelets, Splines and
Applications, pages 473483, 2002.
[84] Beran P.S., Khot N.S., Eastep F.E., Snyder R.D., and Zweber J.V. Numerical Analysis of Store-Induced Limit-Cycle Oscillation. Journal of Aircraft,
41(6):13151326, 2004.
[85] Girodroux-Lavigne P., Grisval J.P., Guillemot S., Henshaw M., Karlsson
A., Selmin V., Smith J., Teupootahiti E., and Winzell B. Comparison of
Static and Dynamic Fluid-Structure Interaction Solutions in the Case of a
Highly Flexible Modern Transport Aircraft Wing. Aerospace Science and
Technology, 7:121133, 2003.
149

[86] Lee B.H.K. and Ohman L.H. Unsteady Pressure and Forces during Transonic Buffeting of a Supercritical Airfoil. Journal of Aircraft, 21(6):439441,
1984.
[87] Lee B.H.K., Ellis F.A., and Bureau J. Investigation of the Buffet Characteristics of Two Supercritical Airfoils. Journal of Aircraft, 26(8):731736,
1989.
[88] Lee B.H.K. Investigation of Flow Separation on a Supercritical Airfoil.
Journal of Aircraft, 26(11):10321037, 1989.
[89] Lee B.H.K. Transonic Buffet on a Supercritical Aerofoil. Aeronautical
Journal, pages 143152, May 1990.
[90] Huang R.F. and Lee H.W. Turbulence Effect on Frequency Characteristics
of Unsteady Motions in Wake of Wing. AIAA Journal, 38(1):8794, 2000.
[91] Xiao Q., Tsai H.M., and Liu F. A Numerical Study of Transonic Buffet
on a Supercritical Airfoil. In 42nd AIAA Aerospace Sciences Meeting and
Exhibit, 5-8 January 2004, Reno, Nevada, AIAA 2004-1056.
[92] Lee B.H.K. and Tang F.C. Characteristics of the Surface Pressures on a
F/A-18 Vertical Fin due to Buffet. Journal of Aircraft, 31(1):228235, 1994.
[93] Lee B.H.K. Self-Sustained Shock Oscillations on Airfoil at Transonic
Speeds. Progress in Aerospace Sciences, 37:147196, 2001.
[94] Levy Jr. L.L. Experimental and Computational Steady and Unsteady Transonic Flows about a Thick Airfoil. AIAA Journal, 16(6):564572, 1978.
[95] McDevitt J.B. Supercritical Flow about a Thick Circular-Arc Airfoil. Technical Report 78549, NASA, 1979.
[96] Rumsey C.L., Sanetrik M.D., Biedrom R.T., Melson N.D., and Parlette
E.B. Efficient and Accuracy of Time-Accurate Turbulent Navier-Stokes
Computations. Computations and Fluids, 25(2):217236, 1996.
[97] Prananta B.B. Two-Dimensional Transonic Aeroelastic Analysis using
Thin-Layer Navier-Stokes Method. Journal of Fluids and Structures,
12:665676, 1998.
[98] Raghunathan S., Gillan M.A., Cooper R.K., Mitchell R.D., and Cole J.S.
Shock Oscillations on Biconvex Aerofoils. Aerospace Science and Technology, 1:19, 1999.
150

[99] Chen L.W., Xu C.Y, and Lu X.Y. Numerical Investigation of the Compressible Flow Past an Aerofoil. Journal of Fluid Mechanics, 643:97126,
2010.
[100] Jacquin L., Molton P., Deck S., Maury B., and Soulevant D. Experimental
Study of Shock Oscillation over a Transonic Supercritical Profile. AIAA
Journal, 47(9):19851994, September 2009.
[101] Deck S. Zonal-Detached-Eddy Simulation of the Flow Around a High-Lift
Configuration. AIAA Journal, 43(11):23722384, 2005.
[102] Garnier E. and Deck S. Large-Eddy Simulation of Transonic Buffet over a
Supercritical Airfoil. In Direct and Large-Eddy Simulation VII, volume 13
of ERCOFTAC, pages 549554. Springer Netherlands.
[103] Raghunathan S., Mitchell R.D., and Gillan M.A. Transonic shock oscillations on naca 0012 aerofoil. Shock Waves, 8:191202, 1998.
[104] Iovnovich M. and Raveh D.E.
Reynolds-Averaged Navier-Stokes
Study of the Shock-Buffet Instability Mechanism.
In 52nd
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and
Materials Conference, number AIAA 2011-2078, 2011.
[105] Singh J.P. Estimation of Shock-Induced Buffet Onset. Acta Mechanica,
151:245253, 2001.
[106] Murrow H.N. and Eckstrom C.V. Drones for Aerodynamic and Structural
Testing (DAST) - A Status Report. Journal of Aircraft, 16(8):521526,
1979.
[107] Cole S.R. and Rivera Jr. J.A. The New Heavy Gas Testing Capability in
the NASA Langley Transonic Dynamics Tunnel. Technical Report NASATM-112702, NASA Langley Research Center, 1997.
[108] Seidel D.A., Sandford M.C., and Eckstrom C.V. Measured Unsteady Transonic Aerodynamic Characteristics of an Elastic Supercritical Wing. Journal of Aircraft, 24(4):225230, April 1987.
[109] Eckstrom C.V., Seidel D.A., and Sandford M.C. Unsteady Pressure and
Structural Response Measurements on an Elastic Supercritical Wing. Journal of Aircraft, 27(1):7580, 1990.
[110] Farhangnia M., Guruswamy G., and Biringen S., editors. Transonic-Buffet
Associated Aeroelasticity of a Supercritical Wing, number AIAA 96-0286 in
34th Aerospace Sciences Meeting and Exhibit, Reno, NV, January 1996.
151

[111] Byrdsong T.A., Adams R.R., and Sandford M.C. Close-Range Photogrammetric Measurement of Static Deflections for an Aeroelastic Supercritical
Wing. Technical Report NASA TM 4194, NASA, Langley Research Center,
Hampton, Virginia, 1990.
[112] Byrdsong T.A., Adams R.R., and Sandford M.C. Close-Range Photogrammetric Measurement of Static Deflections for an Aeroelastic Supercritical
Wing. Technical Report Supplement to NASA TM 4194, NASA, Langley
Research Center, Hampton, Virginia, 1990.

152

Appendix A
Infinite Plate Spline
This appendix will provide the derivation of the transformation matrix for infinite
plate splines. The deflections normal to the plate surface due to N point forces
Fi at given locations (xi , yi ) on the plate can be written as
dz(x, y) = a0 + a1 x + a2 y +

N
X

Fi ri2 ln ri2

(A.1)

i=1

where ri2 = x2i + yi2 . The unknowns, ak and Fi are obtained from the equilibrium
conditions
N
N
N
X
X
X
Fi =
xi Fi =
yi Fi = 0
(A.2)
i=1

i=1

i=1

and from the given deflections at the N nodes

dzj = a0 + a1 xj + a2 yj +

N
X

Fi rij2 ln rij2

(A.3)

i=1

where rij2 = (xi xj )2 +(yi yj )2 . By applying Equations A.2 and A.3 to both the
structural and aerodynamic points the following equations are obtained, expressed
in matrix form as
dxs = Rs a + Zs Fs
(A.4)
dxa = Ra a + Zas Fs

(A.5)

and Equation A.4 is subject to

RsT Fs = 0
where

Rs =

1 xs1 ys1
1 xs2 ys2
..
..
..
.
.
.
1 xsN ysN
153

(A.6)

and

Ra =

1 xa1 ya1
1 xa2 ya2
..
..
..
.
.
.
1 xaN yaN

and where Zs and Zas contain the elements

h

2

2 i h

2

2 i
Zsij = xsi xsj + ysi ysj
ln xsi xsj + ysi ysj
Zasij =

xai xsj

2

+ yai ysj

2 i

ln

xai xsj

2

+ yai ysj

2 i

which are effectively matrices of the basis function (rij2 ln rij2 ). Given that
dxa = T dxs

(A.7)

where T is the transformation matrix and from the principle of virtual work
Fs = T T Fa

(A.8)

The transformation matrix T can be derived as




T = Ra Ys1 RsT + Zas I Ys1 RsT Zs1 Rs Zs1

where Ys = RsT Zs1 Rs and I is an identity matrix.

154

(A.9)

Appendix B
Inverse Isoparametric Mapping
This appendix will provide the derivation of the transformation matrix for Inverse
Isoparametric Mapping. The same shape functions Ni are used for both the
aerodynamic grid points and the structural deformation
xa =

n
X

Ni (, )xesi

(B.1)

Ni (, )ysei

(B.2)

i=1

ya =

n
X
i=1

where n is the number of nodes for the element. The element is assumed to lie in
the x y plane and the superscript e indicates the structural node is a member
of that element. The shape function for a quadrilateral element[80] is given by
N1 (, ) = (1 )(1 )/4
N2 (, ) = (1 + )(1 )/4
N3 (, ) = (1 + )(1 + )/4
N4 (, ) = (1 )(1 + )/4
, [1, 1]

(B.3)

The implementation shown here is based upon the implementation by Smith[60].

First Equation B.1 is expanded and using Equation B.3 one can get

where

xa = ax1 + ax2 + ax3 + ax4

(B.4)

ax1 = (xes1 xes2 + xes3 xes4 )/4

ax2 = (xes1 + xes2 + xes3 xes4 )/4
ax3 = (xes1 xes2 + xes3 + xes4 )/4
ax4 = (xes1 + xes2 + xes3 + xes4 )/4

(B.5)

ya = ay1 + ay2 + ay3 + ay4

(B.6)

The same is true for y

155

where

ay1 = (yse1 yse2 + yse3 yse4 )/4

ay2 = (yse1 + yse2 + yse3 yse4 )/4
ay3 = (yse1 yse2 + yse3 + yse4 )/4
ay4 = (yse1 + yse2 + yse3 + yse4 )/4

(B.7)

Rearranging Equation B.6 for

=

ya ay4 ay2
ay1 + ay3

(B.8)

This can be substituted into Equation B.4 and rearranged into the form

where

A 2 + B + C = 0

(B.9)

A = ax2 ay1 ax1 ay2

B = ay1 (ax4 xa ) + ax1 (ya ay4 ) + ax2 ay3 ax3 ay2
C = ay3 (ax4 xa ) + ax3 (ya ay4 )

(B.10)

If ay1 + ay3 = 0 then use

xa ax4 ax2
ax1 + ax3
to substitute into Equation B.6, this leads to
=

A = ay2 ax1 ay1 ax2

B = ax1 (ay4 ya ) + ay1 (xa ax4 ) + ay2 ax3 ay3 ax2
C = ax3 (ay4 ya ) + ay3 (xa ax4 )

(B.11)

(B.12)

Equation B.9 can be solved as a normal quadratic, there will be one solution
of the equation inside the bounds [1, 1]. Then can be solved for using either
Equation B.8 or B.11. Once the local coordinates for the aerodynamic grid points
have been calculated the shape functions can be calculated using Equation B.3.
Then the displacements can be calculated using
dxa =

n
X

Ni (, )dxesi

(B.13)

i=1

This can be expanded assuming n = 4

dxa = N1 dxes1 + N2 dxes2 + N3 dxes3 + N4 dxes4

(B.14)

and then written in matrix format

dxa = T dxes

(B.15)

where T is the local transformation matrix. This then allows the forces to be
calculated using
Fes = T T Fa
(B.16)
from the principle of virtual work and therefore satisfies conservation of energy.
156

Appendix C
Radial Basis Functions
This appendix will provide the derivation of the transformation matrix for radial
basis functions. The radial basis function interpolant [2] has the form

s(x) =

N
X
i=1

i (kx xi k) + p(x)

(C.1)

where s(x) is the function to be evaluated at x, is the basis function, the index i
identifies the centres for the RBF and xi is the location of that centre. The linear
polynomial p(x) is used to ensure that translations and rotations are recovered.
The coefficients i are found by requiring the exact recovery of the structural
node positions. The polynomial in the x-direction is given by
px (x) = x0 + xx x + xy y + xz z

(C.2)

When the polynomials are included in the system, the additional requirement is
also required
N
X

i q(x) = 0

(C.3)

i=1

for all polynomials q(x) with degree less than or equal to p(x). For ease of
computation the exact recovery of centres gives using up to linear polynomial
terms
Xs = Css ax

(C.4)

Ys = Css ay

(C.5)

Zs = Css az

(C.6)

157

where

Xs =

0
0
0
0
xs

xs1

.
, xs =
..

xsns

, ax =

There are analogous definitions for Ys , Zs , ay and az .

0 0
0
0
1
1
0 0
0
0
xs1
xs2

0 0
0
0
y
ys2
s1

0 0
0
0
zs1
zs2
Css =
1 xs
ys1 zs1 s1 s1 s1 s2
1

..
..
..
..
..
..
.
.
.
.
.
.
1 xsns ysns zsns sns s1 sns s2

x0
xx
xy
xz
xs1
..
.
xsns

...
...
...
...
...
..
.

xsns
ysns
zsns
s1 sns
..
.

. . . sns sns

(C.7)

(C.8)

where s1 s2 = (kxs1 xs2 k) and s indicating a structural node. An analogous

matrix Caa where a indicates an aerodynamic node. To locate the aerodynamic
surface points the following matrix must be formed

1 xa1 ya1 za1 a1 s1 a1 s2 . . . a1 sns

..
..
..
..
..
..
..
Aas = ...
(C.9)

.
.
.
.
.
.
.
1 xana yana zana ana s1 ana s2 . . . ana sns

The positions of the aerodynamic surface points given by the the vectors xa ,
ya and za analogous to xs in Equation C.7 is given by

xa
T 0 0
Xs
ya = 0 T 0 = Ys
(C.10)
za
0 0 T
Zs

1
where T = Aas Css
. Since the matrix Css has a zero block, the matrix can be
subdivided as shown below


0 P
(C.11)
Css =
PT M

The vectors ax , ay and az are split into the parts which multiply the RBF (superscript RBF) and polynomial (superscript PLY)terms,
aPx LY = Mp P M 1 xs

(C.12)

= M 1 M 1 P T Mp P M 1 xs
(C.13)
aRBF
x


1
where Mp = P M 1 P T
. Forming the transformation matrix is now straight
forward


Mp P M 1
(C.14)
T = Aas
M 1 M 1 P T Mp P M 1
158