You are on page 1of 160

Viscous-Inviscid Interaction

with the Quasi-Simultaneous Method


for 2D and 3D Aerodynamic Flow

Rijksuniversiteit Groningen

Viscous-Inviscid Interaction
with the Quasi-Simultaneous Method
for 2D and 3D Aerodynamic Flow
Proefschrift

ter verkrijging van het doctoraat in de


Wiskunde en Natuurwetenschappen
aan de Rijksuniversiteit Groningen
op gezag van de
Rector Magni cus, dr. D.F.J. Bosscher,
in het openbaar te verdedigen op
vrijdag 12 oktober 2001
om 16.00 uur

door

Edith Gerda Maria Coenen


geboren op 2 maart 1973
te Zwolle

Promotor:

Prof.dr. A.E.P. Veldman

Referent:

Prof. S.P. Fiddes

Beoordelingscommissie:

Prof.dr.ir. P.G. Bakker


Prof.dr.ir. H.W.M. Hoeijmakers
Prof.dr.ir. H.W. Hoogstraten

ISBN 90-367-1472-9

The research described in this thesis is funded by the University of Bristol, U.K., the
Engineering and Physical Sciences Research Council (EPSRC), U.K., the Defence Evaluation and Research Agency (DERA), U.K., and the University of Groningen.
The cover is based on the statue `Four Wings' (1972) by Alexander Calder (American
sculptor, 1898-1976), which stands in front of the Miro museum in Barcelona, Spain.

Contents
1 Introduction to viscous-inviscid interaction
1.1 Introduction . . . . . . . . . . . . . . . . . .
1.2 Viscous-inviscid interaction . . . . . . . . . .
1.3 Historical overview . . . . . . . . . . . . . .
1.4 Outline present research . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

5
. 5
. 6
. 7
. 10

2 Boundary-layer region
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Integral boundary-layer formulation . . . . . . . . . . . . . . . .
2.3 Coordinate systems . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Two-dimensional boundary layer . . . . . . . . . . . . . . . . .
2.4.1 Thwaites' method for laminar boundary layers . . . . . .
2.5 Quasi-three-dimensional boundary layer . . . . . . . . . . . . . .
2.6 Three-dimensional boundary layer . . . . . . . . . . . . . . . . .
2.7 Discretisation boundary-layer equations . . . . . . . . . . . . . .
2.7.1 Discretisation two- and quasi-three-dimensional problems
2.7.2 Discretisation three-dimensional problem . . . . . . . . .
2.8 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . .
2.8.1 Attachment line/stagnation point . . . . . . . . . . . . .
2.8.2 Transition line . . . . . . . . . . . . . . . . . . . . . . . .
2.8.3 Root boundary condition . . . . . . . . . . . . . . . . . .
2.8.4 Tip boundary condition . . . . . . . . . . . . . . . . . .

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

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

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

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

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

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

13
13
13
14
15
16
17
19
20
20
21
24
24
25
27
27

.
.
.
.
.
.
.
.
.
.
.

29
29
29
30
32
34
36
36
39
39
40
41

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

3 External ow region
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Inviscid steady potential- ow problem . . . . . . . . .
3.2.1 Formulation of the integral equation . . . . . .
3.2.2 Model problem . . . . . . . . . . . . . . . . . .
3.2.3 Wing/aerofoil problem . . . . . . . . . . . . . .
3.2.4 Full-potential problem . . . . . . . . . . . . . .
3.3 Viscous potential- ow problem . . . . . . . . . . . . . .
3.4 Discretisation viscous potential- ow problem . . . . . .
3.4.1 Global and local coordinate systems . . . . . . .
3.4.2 Discretisation two-dimensional model problem .
3.4.3 Discretisation three-dimensional model problem

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

CONTENTS

3.4.4 Discretisation wing/aerofoil problem .


3.5 Aerodynamic coecients . . . . . . . . . . . .
3.6 Veri cation two-dimensional Dirichlet method
3.7 Adjustment trailing edge velocity distribution

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

43
47
48
50

4 Viscous-inviscid interaction with the quasi-simultaneous method


4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Interaction algorithms . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Quasi-simultaneous method . . . . . . . . . . . . . . . . . . . . . . .
4.4 Analysis of interaction laws . . . . . . . . . . . . . . . . . . . . . . .
4.5 Viscous-inviscid iterations . . . . . . . . . . . . . . . . . . . . . . . .
4.5.1 Basic conditions for viscous-inviscid convergence . . . . . . . .
4.5.2 Convergence rate viscous-inviscid iterations . . . . . . . . . .
4.6 Boundary-layer iterations . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.1 Convergence rate boundary-layer iterations . . . . . . . . . . .
4.6.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.7 Basic convergence pointwise iterations . . . . . . . . . . . . . . . . .
4.7.1 Relaxation boundary-layer iterations . . . . . . . . . . . . . .
4.7.2 Relaxation viscous-inviscid iterations . . . . . . . . . . . . . .
4.7.3 Relaxation due to unsteady viscous-inviscid iterations . . . . .
4.8 Interaction law variations . . . . . . . . . . . . . . . . . . . . . . . . .
4.8.1 Two-sided interaction laws . . . . . . . . . . . . . . . . . . . .
4.8.2 One-sided interaction laws . . . . . . . . . . . . . . . . . . . .
4.8.3 Local tridiagonal interaction law . . . . . . . . . . . . . . . . .
4.9 Numerical evaluation of interaction laws . . . . . . . . . . . . . . . .
4.9.1 Basic convergence a ected by the choice of interaction law . .
4.9.2 Basic convergence a ected by the choice of relaxation . . . . .
4.9.3 Convergence rate a ected by the choice of interaction law . . .
4.9.4 Convergence rate a ected by the choice of relaxation . . . . .
4.9.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10 E ect discretisation on interaction law . . . . . . . . . . . . . . . . .
4.11 Geometric interpretation iterative variations . . . . . . . . . . . . . .

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

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

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

53
53
53
57
58
59
60
61
63
64
65
65
66
66
68
69
69
69
69
70
72
74
78
79
79
80
82

.
.
.
.
.
.
.
.
.
.
.

85
85
85
85
87
87
89
89
91
92
93
94

5 Applications of the quasi-simultaneous method


5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Interaction three-dimensional systems . . . . . . . . . . . . . . .
5.2.1 Two-dimensional unsteady interaction . . . . . . . . . .
5.2.2 Three-dimensional interaction . . . . . . . . . . . . . . .
5.3 Newton solution procedure . . . . . . . . . . . . . . . . . . . . .
5.3.1 Initial conditions . . . . . . . . . . . . . . . . . . . . . .
5.4 Three-dimensional quasi-simultaneous scheme . . . . . . . . . .
5.5 Three-dimensional plate interaction law . . . . . . . . . . . . . .
5.5.1 Comparison two- and three-dimensional interaction laws
5.6 Dirichlet aerofoil interaction law . . . . . . . . . . . . . . . . . .
5.7 Dirichlet wing interaction law . . . . . . . . . . . . . . . . . . .

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

CONTENTS

5.8
5.9
5.10
5.11

5.7.1 Lumping Dirichlet wing interaction law


Interaction at the trailing edge . . . . . . . . .
Wake interaction law . . . . . . . . . . . . . .
Inclusion compressibility into interaction law .
Numerical evaluation aerofoil interaction law .
5.11.1 Conclusion . . . . . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

6 Results
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Two-dimensional aerofoil results . . . . . . . . . . . . . .
6.2.1 NACA0012 aerofoil . . . . . . . . . . . . . . . . .
6.2.2 NACA4412 aerofoil . . . . . . . . . . . . . . . . .
6.3 Swept tapered wing results . . . . . . . . . . . . . . . . .
6.4 Three-dimensional dented plate results . . . . . . . . . .
6.4.1 Flat plate . . . . . . . . . . . . . . . . . . . . . .
6.4.2 Three-dimensional trough ow . . . . . . . . . . .
6.4.3 Three-dimensional dent ow . . . . . . . . . . . .
6.5 Three-dimensional wing results . . . . . . . . . . . . . .
6.5.1 Quasi-two-dimensional NACA0012 wing ow . . .
6.5.2 Three-dimensional unswept NACA0012 wing ow
6.5.3 Three-dimensional 45o swept RAE101 wing ow .

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

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

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

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

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

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

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

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

.
.
.
.
.
.

.
.
.
.
.
.

95
96
96
97
98
99

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

101
. 101
. 101
. 102
. 103
. 106
. 109
. 109
. 110
. 111
. 112
. 113
. 114
. 115

7 Conclusions

119

Bibliography

123

A Streamline and Cartesian boundary-layer integral thicknesses

131

B Boundary-layer closure relations


B.1 Streamwise closure . . . . . . .
B.2 Cross ow closure . . . . . . . .
B.3 Skin-friction closure . . . . . . .
B.4 Entrainment closure . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

133
. 133
. 134
. 135
. 136

C Derivation Cartesian integral boundary-layer equations


137
C.1 Boundary-layer equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
C.2 Integral momentum equations . . . . . . . . . . . . . . . . . . . . . . . . . 138
C.3 Entrainment equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
D Auxiliary de nitions and theorems

141

Samenvatting

145

Summary

149

Acknowledgements

153

CONTENTS

Chapter 1
Introduction to viscous-inviscid
interaction
1.1 Introduction
The accurate and fast prediction of the development of viscous ow over two- and especially three-dimensional surfaces is an important problem in both aerodynamics and
hydrodynamics.
The continuing advances in eciency and accuracy of numerical algorithms, together
with the increasing speed and memory size of computers, are enabling viscous ows to be
calculated by methods that solve the full (Reynolds-averaged) Navier-Stokes equations,
which represent the ow of a viscous uid. Whilst Navier-Stokes simulation potentially
o ers generality, the excessive requirements for computer memory and CPU (CentralProcessing-Unit) time currently limit their use for practical application, especially within
a design optimisation environment where a large number of con gurations is to be considered.
A possible alternative for the prediction of viscous ow over aerofoil sections and wings,
is to use the older technique of viscous-inviscid interaction (VII), whereby an inviscid ow
solver is coupled to a viscous boundary-layer calculation method to provide a composite
solution of the Navier-Stokes equations.
For two-dimensional ow problems extensive research has been done into viscousinviscid interaction. Various coupling algorithms have been developed and have shown to
be very ecient and robust [61]. For many cases of aerodynamic interest the coupled solution matches the experimental data as well as Navier-Stokes simulation and this at much
lower computational cost than required for Navier-Stokes simulation. The development of
three-dimensional viscous-inviscid interaction methods has so far been limited but similar
advantages, in terms of speed and robustness as were obtained in two dimensions, are to
be expected.
A set-back of the coupled schemes is that they are applicable to a smaller range of
applications, compared to Navier-Stokes solvers. Viscous-inviscid interaction methods are
restricted to the calculation of ows with only limited amounts of separation, i.e. ows
having a thickness in the separated region comparable to the boundary-layer thickness
and, containing (long) at separation bubbles. The calculation of larger regions with

Chapter 1 Introduction to viscous-inviscid interaction

separation (marginal separation or even larger, massive separation) is beyond the scope of
viscous-inviscid interaction. However, viscous ow problems relevant for the civil aircraft
industry contain mostly only limited regions with ow separation, and are suitable to be
treated with a viscous-inviscid interaction method.
In this thesis, several computational methods based on the idea of viscous-inviscid interaction have been developed and results are presented for the prediction of incompressible two-dimensional aerofoil ow, transonic quasi-three-dimensional swept tapered wing
ow and fully three-dimensional incompressible ow over dented plates and aircraft wings.
The quasi-simultaneous interaction method is used to couple the inviscid ow solver to
the boundary-layer equations and limited regions with two- and three-dimensional ow
separation are permitted.

1.2 Viscous-inviscid interaction


Viscous-inviscid interaction techniques divide the ow domain in two regions: an external
region, where the inviscid ow equations hold, and a thin shear layer region, where the
boundary-layer equations are valid. By combining the two regions, the solution for the
whole ow is found.
It all started in 1904, when Prandtl presented his paper with the title \Uber Flussigkeitsbewegung bei sehr kleiner Reibung" at a congress in Heidelberg [76]. He explained
that viscous e ects only play an important role in a very small region near the surface,
which he termed the boundary layer. Outside this region, the ow can be assumed to be
inviscid.
From the Navier-Stokes equations, which describe completely all possible uid ow
situations, the so-called ( rst-order) boundary-layer equations to model the thin shear
layer ow can be derived, assuming that the thickness of the shear layer is small compared
to the characteristic overall length and that the freestream Reynolds number is large. In
the inviscid ow region the Navier-Stokes equations simplify to the Euler equations and
an inviscid ow calculation technique can be applied.
The e ect of the boundary layer is to displace the external inviscid ow away from
the physical body over a certain distance, termed displacement thickness, to create an
e ective displacement body. The external inviscid ow therefore does not encounter the
original body but a thickened one. Alternatively, the e ect of the presence of the boundary
layer on the inviscid ow is explained as uid injected normal to the surface of the body
to displace the dividing streamline from the surface. The required transpiration velocity
of this uid is related to the rate of change of mass de cit in the boundary layer.
The inviscid ow region and the boundary layer have to be combined to provide a
composite solution, representing the genuine approximation of the physical ow problem,
which is the basic idea behind all viscous-inviscid interaction techniques.
The iterative approach employed to couple the two regions together is through the
boundary conditions of each region. To include the displacement e ect of the boundary
layer, the inviscid ow method uses instead of the zero normal velocity boundary condition, the transpiration velocity boundary condition at the body's surface. The alternative
of changing the shape of the body for each iteration to include the boundary-layer growth,

1.3 Historical overview

is usually computationally inecient because of the required regridding. In the boundarylayer region use is made of the calculated inviscid velocity distribution together with a
no-slip boundary condition at the surface, to determine a new solution of the boundarylayer equations. The iteration process is repeated until the velocities at the edge of the
boundary layer, as computed from the external inviscid ow and those computed from
the boundary layer, match.

1.3 Historical overview


Several viscous-inviscid interaction schemes to solve the system of boundary-layer and
external inviscid ow equations have been developed since Prandtl rst introduced his
boundary-layer concept in 1904. A good overview can be found in the paper by Lock and
Williams [61] in which the relative merits of each scheme are described.
The inclusion of the viscous e ects in the external inviscid ow was originally organised via an iterative calculation process based on the mathematical theory of matched
asymptotic expansions [87, 66, 63]: (1) assuming that no boundary layer is present, the
inviscid velocity (or pressure) distribution along the surface of the body is calculated, (2)
with this velocity distribution, the solution to the boundary-layer equations is found, (3)
via the transpiration velocity boundary condition the presence of the boundary layer is
then included in the external inviscid ow and a new (disturbed) inviscid ow velocity
distribution is determined. Step two and three are repeated until convergence in terms of
boundary-layer solution and inviscid ow solution is reached. This method is referred to
as the direct method as it makes use of a sequential way of calculating based on a direct
hierarchy between the outer ow and the boundary layer.
Diculties arise when the boundary layer no longer stays attached but separates
from the surface. This usually happens in the trailing edge region or at the foot of a
strong shock, when a severe positive pressure gradient is present. The direct scheme
with prescribed velocity distribution for the boundary-layer equations proves to be illconditioned, which can be demonstrated analytically as the `Goldstein singularity' [37]
for two-dimensional laminar ow, leading to failure of the iteration procedure.
After many fruitless attempts to resolve the diculty with the Goldstein singularity,
Catherall and Mangler [11] came around the problem by solving the set of equations with
an inverse approach. They organised their iteration process the other way around. Instead
of using a prescribed velocity distribution, the boundary-layer equations are solved with a
prescribed displacement thickness which indicates the increased importance of the viscous
layer in separated ow. The external inviscid ow makes use of the newly calculated
velocity distribution from the boundary-layer equations to predict a new displacement
thickness. In this way the hierarchy as was used in the direct scheme is reversed and
problems that occurred at separation are avoided. This interaction scheme is known as
the inverse method.
Another successful inverse approach has been demonstrated by Carter [8], specifying
either the displacement thickness or the skin-friction for the boundary-layer calculations.
However, a general setback of the inverse techniques is that severe underrelaxation is
required, making the method very slow to converge [61].

Chapter 1 Introduction to viscous-inviscid interaction

Simultaneously to the numerical progress, advances were made in the mathematical


theory of matched asymptotic expansions. Independently Stewartson and Williams [88],
Neiland [74] and Messiter [68] developed the so-called triple-deck theory, which shows
that for ow near a singularity point, the thin-shear layer approximations are still valid
and that the boundary layer is equally important as the external inviscid ow. No longer
does the boundary layer merely provide small corrections to the external ow as was the
case for the classical direct method.
Following the triple-deck theory numerical methods were developed to represent the
interaction correctly for separated ows. As the hierarchy between the inviscid and viscous regions near a singularity disappears, both regions should be treated together. The
methods taking this into account and still used today, are the semi-inverse method, the
quasi-simultaneous method and the simultaneous method.
The semi-inverse scheme, closely related to the inverse scheme, was developed by Le
Balleur [57] and Carter [9] independently. The semi-inverse method calculates the outer
ow in a direct way and the boundary layer in an inverse way, both regions thus having a
prescribed displacement thickness. With this approach the advantages of both the direct
and inverse method are incorporated. A new displacement thickness is determined by
a special relaxation formula for which the most optimal relaxation factor is determined
with the use of Fourier-stability analysis [57].
The semi-inverse method is widely applied for two-dimensional ow calculations and
is quite successful in dealing with separation as can be seen in the papers by Ashill, Wood
and Weeks [6] who developed the Viscous Garabedian and Korn (VGK) method, Dargel
and Thiede [23] who developed the VILMA aerofoil code, and Williams [106].
To further improve the convergence rate, Veldman developed the quasi-simultaneous
coupling method. Instead of solving the boundary-layer equations with a prescribed displacement thickness, the boundary layer is treated together with a simple approximation
of the inviscid ow model. The quasi-simultaneous method can be seen to resemble the
classical direct method, as the external inviscid ow is calculated rst producing a velocity distribution which is next used for the boundary-layer calculations. The important
di erence with the direct scheme is, however, that the boundary-layer calculations are performed simultaneously with a simpli cation for the inviscid external ow. Compared to
the semi-inverse method, the quasi-simultaneous method has shown for two-dimensional
ow calculations to have a much improved convergence rate and to be more robust [61].
Veldman developed the quasi-simultaneous coupling originally for two-dimensional
laminar ow calculations over a at plate using the di erential boundary-layer formulation [97, 98]. In later work he applied the method to steady and unsteady transonic
aerofoil calculations [48, 101]. Cebeci combined the quasi-simultaneous method with his
di erential boundary-layer equations and applied it to various test cases [12, 102]. Others
that used the quasi-simultaneous method for two-dimensional ow calculations are Arnold
and Thiele [3], Kusunose, Wigton and Meredith [53], and King and Williams [52] who
applied the quasi-simultaneous scheme for their multi-element aerofoil computations.
A few years later Drela adopted a fully simultaneous approach, rst introduced by
Lees and Reeves [59] for the calculation of supersonic ow. Drela coupled an Euler
method with the integral boundary-layer equations for the calculation of two-dimensional
transonic aerofoil ow [25, 27]. In a second simultaneous method Drela combines the

1.3 Historical overview

integral boundary-layer equations with a vorticity streamfunction panel method, to reduce


the computational e ort of the external ow calculations [26].
The fully simultaneous method has shown to be very robust, however, within a design optimisation method it is quite time-consuming to solve the full inviscid outer ow
equations together with the boundary-layer equations for each iteration. For an ecient
and fast calculation of ows with limited amounts of separation the quasi-simultaneous
approach is therefore to be preferred.

Three-dimensional viscous-inviscid interaction history


Two-dimensional viscous-inviscid interaction methods are nowadays quite advanced, however, although several e orts have been directed towards the extension to three dimensions, more research is still needed. The development of a three-dimensional viscousinviscid interaction method proves dicult due to the physical and numerical complexity
of the boundary-layer equations.
Weak coupling schemes have typically been used for three-dimensional interacting
ows. Firmin [33, 34] and Arthur [4] have calculated transonic ows over wings and
wing-fuselage combinations by combining a potential- ow solver together with Smith's
turbulent integral boundary-layer formulation [83] using the direct method. Streett [90]
developed a similar method for solving the potential ow and the integral boundary layer,
where the wake is modelled with simple strip theory.
As in two dimensions, singularities appear in the direct mode for three-dimensional
ow calculations. Cousteix and Houdeville [20] examined the occurrence of these singularities near the separation line when solving the three-dimensional boundary-layer equations
in integral form with prescribed pressure. They found that as in two dimensions a local
singularity can be identi ed by a zero streamwise component of the skin-friction. However, in general there is no reason why the point of zero streamwise skin-friction should be
a singularity. It moreover indicates a change in orientation of the limiting wall streamline.
Cousteix and Houdeville concluded that the singularities in three-dimensional ow can
be avoided if the equations are solved in an inverse mode.
The application of the inverse method is described in a paper by Smith [84] and was
shown to be successful for the calculation of separated ow by Wigton and Yoshihara [105],
combining an integral boundary layer and the full-potential equation, and by Delery and
Formery [24], using the boundary-layer equations in di erential form.
In later work by Wai and Yoshihara [108] some questions were raised on the applicability of the inverse method in three dimensions as they encountered numerical problems.
The work by Smith [80] and Sykes [92] on laminar triple-deck theory suggests that the
inverse form of the three-dimensional boundary-layer equations can lead to an elliptic
system. Solving the three-dimensional boundary-layer equations with a forward-marching
technique would hence lead to exponential departure solutions.
The pseudospectral approach by Duck and Burggraf [28] avoids this problem and
as mentioned Delery and Formery [24] did not encounter diculties with their inverse
scheme. Van Dalsem and Steger [93] successfully employed a time-relaxation algorithm
to obtain an inverse solution.
To clarify the situation, several inverse schemes were investigated by Edwards, Carter

10

Chapter 1 Introduction to viscous-inviscid interaction

and Smith [31, 32]. It was concluded that only the inverse scheme with prescribed normal
vorticity results in an elliptic system. Specifying one or no edge velocity component and
one or two integral thicknesses will result in a stable inverse scheme, capable of calculating
within the region of separation.
More recently, with the use of a special multi-zonal-marching technique Lazare and
Le Balleur [55] and Le Balleur and Girodroux-Lavigne [58] computed three-dimensional
separated ows with a semi-inverse method. However, when the semi-inverse method
is applied to test cases with reasonable amounts of separation present, the inviscid and
viscous solution are mismatched and an update procedure for rectifying this has yet to be
developed [35]. To avoid the use of an update procedure and to obtain a faster convergence rate, for ow calculations with strong interaction, a more simultaneous approach is
preferred.
The simultaneous method, developed by Drela [27], has been extended to three dimensions by Nishida and Drela [75] and Milewski [69]. They obtained good results, however,
especially in three dimensions the simultaneous method is computationally expensive.
The ecient quasi-simultaneous method has a straightforward extension to threedimensions and was shown to work successfully by Edwards [30], Smith [81], and Roget
et al. [78] for the calculation of laminar ow over a at plate with a dent or a hump. At
the Dutch National Aerospace Laboratory NLR the work of Veldman has been continued
and a code for transonic ow over wing/body con gurations has been developed [94, 95].
From the quasi-simultaneous method similar advantages as were obtained in two dimensions in terms of convergence speed and robustness are expected in three-dimensional
calculations. Although some work has been done on three-dimensional viscous-inviscid
interaction an ecient and robust quasi-simultaneous method is still to be fully developed
and tested.

1.4 Outline present research


The work described in this thesis is an investigation into quasi-simultaneous viscousinviscid interaction when applied to two- and especially three-dimensional ow problems.
The quasi-simultaneous interaction process is analysed and its requirements and implementation are discussed. The outline of this thesis is as follows.
In chapter 2 the mathematical and numerical boundary-layer formulations are presented for the various ow cases to be examined. Both laminar and turbulent ow models
are given and various mechanisms are discussed to predict transition to turbulence. Furthermore, the required boundary conditions are given.
The mathematical and numerical external inviscid ow modelling with the inclusion
of viscous e ects through the transpiration sources is explained in chapter 3. Thinwing/aerofoil theory is applied for the two- and three-dimensional dented plate calculations, whereas the constant potential Dirichlet method is used for the wing and aerofoil
cases. Required for the coupled potential- ow/boundary-layer computations, the edge
velocity equations are determined for each ow case.
In chapter 4 the various techniques to couple the boundary layer to the external inviscid ow region are discussed of which the quasi-simultaneous method in great detail.

1.4 Outline present research

11

Fundamental mathematical theory for the quasi-simultaneous method is presented providing guidelines for the optimisation of the coupling process between the boundary layer
and the external inviscid ow. The developed theory is evaluated numerically for the
well-known two-dimensional Carter-Wornom trough problem.
Chapter 5 is an extension of chapter 4. The quasi-simultaneous interaction method is
extended to three-dimensions and the approximations of the external ow are determined
for the practical problems of two-dimensional aerofoil ow and three-dimensional dented
plate and wing ow. Furthermore, the Newton solution procedure for the two- and threedimensional coupled system is studied in greater detail.
In chapter 6 the results of the various ow cases are presented and nally, the conclusions are given in chapter 7.

12

Chapter 1 Introduction to viscous-inviscid interaction

Chapter 2
Boundary-layer region
2.1 Introduction
In the present section the mathematical and numerical boundary-layer formulations are
given for the two- and (quasi-)three-dimensional ow cases to be examined in this thesis. Both laminar and turbulent ow models are presented and various mechanisms are
discussed to predict transition to turbulence.
For the practical wing and aerofoil problems a wake is included, which is modelled as
a turbulent boundary layer without the wall layer. Boundary conditions are used at the
edges of the ow domain and conditions are supplied at the attachment line/stagnation
point.

2.2 Integral boundary-layer formulation


The Navier-Stokes equations, based on the conservation laws for mass, momentum and
energy, together with the equations of state, describe all ow situations completely in a
Newtonian medium. The ratio of the inertia (convective) terms to the viscous (di usive)
terms in these equations is de ned as the dimensionless Reynolds number Re = Q1 L= ,
where  is the kinematic viscosity, and Q1 and L are the characteristic scales of the
velocity of the ow and the length of the body, respectively.
With the assumptions that the Reynolds number is large and the characteristic boundary-layer thickness small compared to the characteristic length L of the body, the rstorder boundary-layer equations can be derived from the Reynolds-averaged Navier-Stokes
equations. As a consequence of the assumptions, the pressure does not vary across the
boundary layer. It is noted that for situations where the streamlines are highly curved, as
is the case immediately behind the trailing edge for rear-loaded aerofoils, the assumption
of constant pressure across the boundary layer is no longer justi ed [99].
In order to solve the boundary-layer equations two approaches can be followed. One
approach is to solve the boundary-layer equations in its original di erential form. Alternatively, by introducing suitable boundary-layer integral thicknesses, the boundary-layer
equations can be integrated across the characteristic thickness of the boundary layer, to
the integral boundary-layer equations. With the latter approach the governing partial

14

Chapter 2 Boundary-layer region

di erential equations loose one dimension and only body surface points are required for
the calculation. Integral methods are therefore computationally more economical than
di erential methods. The accuracy of both methods is generally the same, depending
mainly on the closure relations and the transition modelling applied.
With respect to viscous-inviscid interaction, integral methods have shown to be more
robust, due to the built-in empiricism in the closure relations. Furthermore, as the displacement thickness, which is a variable in the interaction procedure, is also a variable in
the integral equations, the coupling of the inviscid ow region to the boundary layer is
straightforward.
The objective of the present research is to develop an ecient and robust computational method and the integral approach has therefore been adopted.
Various integral boundary-layer formulations have been used for the di erent ow
cases. In the laminar ow regime up to transition Thwaites' method [70] is employed.
In the turbulent ow regime Head's entrainment idea is followed, which is applied to
two-dimensional ow as presented in Green's paper [39]. The description for quasi-threedimensional swept tapered wing ow is as given by Smith [83] and for the fully threedimensional ow case the formulation is based on the paper by Mughal and Drela [72].
In the remainder of this chapter the above mentioned boundary-layer formulations
are presented. However, rst the various coordinate systems in which the boundary-layer
equations can be written, are discussed.

2.3 Coordinate systems


As experimental results are most readily obtained along the external streamlines of
the ow, providing closure for the boundary-layer method, the integral boundary-layer
equations have traditionally been written in an orthogonal curvilinear coordinate system
(s; n; z ), based on the local external streamline direction s. The direction normal to the
external streamline and tangent to the surface is de ned by n, see gure 2.1, and is more
frequently indicated as the cross ow direction. The z -direction is for each of the discussed
coordinate systems taken to be outward and normal to the surface of the geometry.
For general applicability it is more ecient to use a grid suitable for the inviscid ow
calculations, which will give a uniform coverage of the geometry's surface. For this reason
y, v

n, u n

yc , v c
external
streamline
s, u s

c
c
z

xc , u c
x, u

Figure 2.1: Local coordinate systems.

2.4 Two-dimensional boundary layer

15

Myring [73] developed a calculation method which uses the boundary-layer equations
transformed to a general non-orthogonal curvilinear coordinate system. He de ned the
directions, xc and yc, in the body surface, which are at an angle c. The relationship
between the streamline coordinate system (s; n) and (xc ; yc) can be seen in gure 2.1,
where the angle between the xc -direction and the s-direction is de ned by c .
Metric coecients, accounting for the curvature of the grid, are required for the
boundary-layer equations in a curvilinear coordinate system, and if the direction of the
external streamline changes they have to be recomputed, which for three-dimensional ow
calculations is a cumbersome process.
In order to obtain a grid suitable for inviscid ow calculations and which avoids the calculation of metric coecients, Mughal and Drela developed a three-dimensional method
locally using a Cartesian coordinate system (x; y ) tangent to the surface for the formulation of the integral boundary-layer equations. Via a simple rotation the boundary-layer
variables in the Cartesian coordinate system can be expressed in terms of the streamline
coordinate system to make use of the empirical closure relations. The rotational terms in
e ect replace the metric coecients of the curvilinear coordinate system.
The above mentioned coordinate systems have all been used for the various ow cases.
For the two-dimensional ow problems, with no cross ow present, the boundary layer
equations are simply solved along the external streamline s. The quasi-three-dimensional
boundary-layer method on the other hand, developed by Smith who based his work
on Myring's paper, uses a non-orthogonal curvilinear coordinate system. For the fully
three-dimensional ow calculations the paper by Mughal and Drela is followed and the
boundary-layer equations are written in a local Cartesian coordinate system. As mentioned, via a transformation the boundary layer variables for the (quasi-)three-dimensional
methods can be expressed in the orthogonal curvilinear system.

2.4 Two-dimensional boundary layer


The two-dimensional unsteady incompressible integral boundary layer is described by
the following two equations, obtained by integrating the boundary-layer momentum and
continuity equations in the direction normal to the wall [104]:
1 @
@
 @u
C
(us s ) + ss + (H + 2) ss s = f s ;
2
us @t
@s
us @s
2
@
 @u
@H
H1 ss + H1 ss s + ss 1 = CE :
@s
us @s
@s
e

(2.1)

(2.2)

The above equations are given in a streamline coordinate system and are valid for both
laminar and turbulent ow.
The term us represents the streamwise edge velocity component, which corresponds
with the total edge velocity qe . The de nitions for the streamwise displacement thickness
s and the streamwise momentum thickness ss can be found in appendix A. The shape
factor H is de ned as s =ss and shape factor H1 corresponds with ( s )=ss, where
is the boundary-layer thickness. The coecients Cf s and CE are the skin-friction and
e

16

Chapter 2 Boundary-layer region

entrainment coecient, respectively. Coecient CE describes the normalised rate at


which the ow enters the boundary layer through its external edge.
The steady boundary-layer equations are a special case of the above equations and it is
seen that the rst equation (2.1) without time-derivative is the well-known Von Karman
momentum integral equation. The second equation (2.2) is known as Head's entrainment
equation, based on the concept of entrainment velocity.
Equation (2.2) can be rewritten, using that shape factor H1 is a function of H
 @u
0 @
0 @ 
(H1 HH1 ) ss + H1 ss s + H1 s = CE ;
(2.3)
@s
us @s
@s
e

where gradient H1 corresponds with dH1 =dH . Equation (2.3) combined with (2.1) gives
@ 
@ 
@u
(2.4)
a s + b s + c s = d;
@s
@s
@t
having assumed for clarity of the presentation of the theory that us is constant and
us  1 hence, @us s =@s  @s =@s. The terms a, b, c and d are de ned as

0
a =
[HH1(H + 2) H1 (H + 1)];
(2.5)
us
0
b = H1 ;
(2.6)
1
0
c = 2 (H1 HH1 );
(2.7)
us
Cf s
0
(H1 HH1):
(2.8)
d = CE
2
Suitable closure relations are required to uniquely determine the system of equations.
In appendix B algebraic expressions are given for the shape parameter H1 , and the coecients CE and Cf s in both laminar and turbulent ow. The unknowns remaining are
ss , H and us .
e

2.4.1 Thwaites' method for laminar boundary layers


For the calculation of the steady two-dimensional incompressible boundary layer in the
laminar region, up to transition, it is much easier to use, instead of (2.1) and (2.2),
Thwaites' method [70]
1d 2 6
(ss us ) = 0:45u5s ;
(2.9)
 ds
which relies on the momentum integral equation.
For any given velocity distribution us and initial guess of ss , a new solution for ss
can be determined by simple integration of the rst-order ordinary di erential equation
(2.9). Once the streamwise momentum thickness is known, closure relations de ned in
appendix B can be applied to nd information on the skin-friction distribution Cf s and
shape factor H .
Most boundary-layer prediction methods described in this thesis will make use of
Thwaites' method in the laminar region.
e

2.5 Quasi-three-dimensional boundary layer

17

2.5 Quasi-three-dimensional boundary layer


The steady compressible integral boundary layer, incorporated in the DERA-VFP code
for the calculation of quasi-three-dimensional ow over swept tapered wings, is based on
Smith's boundary-layer method [83]




@ss @sn
H + 2 Me2 @us
2 Me2 @us
+
+ ss
Ks + sn
2Kn
@s
@n
us
@s
us @n
e

+s

1 @us
us @n

Kn + Ksnn =


@ns @nn
2 Me2 @us
+
+ ns
@s
@n
us
@s

2Ks + nn

Cf s
;
2

(2.10)


2 Me2 @us
us @n

Kn Kn ss (H + 1)

Cf n
;
2

1 Me2 @us

s )
us
@s

+Kn ss (H + 1) =

@ (

@s

s )

@n
+ (
@n

(2.11)
e

Ks

Me2 @us
Kn = CE ;
(2.12)
n
us @n
described in a local streamline coordinate system (s; n). The integral thicknesses are de ned in appendix A and Cf s and Cf n are the skin-friction coecients in the streamwise
and crosswise directions, respectively. The Mach number at the edge of the boundary
layer is denoted by Me . The terms Ks and Kn are the geodesic curvatures of the curves
s=constant and n=constant, respectively. In e ect, Ks is a measure of the rate of external streamline divergence. Based on the equation for the component of vorticity of
the external ow in the direction normal to the surface, Kn can be derived, which for
irrotational ow results in Kn = 1=us @us =@n [84].
Equations (2.10) and (2.11) are the s- and n-momentum integral equations, and (2.12)
is the entrainment equation. With the operators
@
sin(c c ) 1 @
sin c 1 @
=
+
;
@s
sin c h1 @xc sin c h2 @yc
1
+ 

cos(c c) 1 @
cos c 1 @
@
=
+
;
@n
sin c h1 @xc sin c h2 @yc
the boundary-layer equations can be transformed to the general non-orthogonal curvilinear coordinate system (xc ; yc), where c the angle between the xc - and yc-directions and
c the angle between the xc - and s-directions, as seen before in gure 2.1. The metric
coecients h1 and h2 of the xc - and yc-directions, respectively, are de ned in such a way,
that a length s on the surface is given by
s2 = h21 xc 2 + h22 yc2 + 2h1 h2 cos cxc yc:

18

Chapter 2 Boundary-layer region


z
external
streamline

wall
streamline

Figure 2.2: Three-dimensional boundary-layer pro le.


For the special case of a swept tapered wing the integral equations (2.10), (2.11)
and (2.12) can be simpli ed. With the two assumptions, that the boundary-layer integral
thicknesses are proportional to the local chord and that the isobars lie along the generators
of the wing, implying us and c to be invariant with yc, the three-dimensional boundarylayer equations reduce to a system of ordinary di erential equations. The calculations
can be performed separately along each chordwise strip on the wing, similar to the twodimensional ow problem. For this particular case the geodesic curvatures are given by
@us
cos2 (c c)
;
Ks =
h1 us sin c sin(c c) @xc
e

cos(c c ) @us
:
h1 us sin c @xc
The three integral boundary-layer equations are not sucient to arrive at a uniquely
determined system. An extra equation is required for c . With the assumptions that
the ow is irrotational and that us and c are invariant with yc , the following equation
can be derived from the equation for the z -component of vorticity of the external ow,
relating c to us :
1 @ cos(c c )
sin(c c ) @h1
cos(c c ) @us
+
=
:
(2.13)
us h1 sin c @xc h1 sin c
@xc
h1 h2
@yc
The streamwise and cross ow turbulent closure relations are presented in appendix B,
reducing the seven unknowns , s , n , ss , sn , ns and nn to only three unknowns, being
ss , H and , where is the angle between the external streamline and the limiting wall
streamline (streamline at the surface), as seen in gure 2.2. The two-dimensional relation
for the streamwise skin-friction coecient is used as given in appendix B. With
Cf n
un
=
;
(2.14)
tan = zlim
!1 us
Cf s
the cross ow skin-friction coecient is determined. The lag-entrainment method of Green,
Weeks and Brooman [40] is used to provide the closure for the entrainment coecient.
The ve unknowns remaining are ss , H , , c and us .

Kn =

2.6 Three-dimensional boundary layer

19

The laminar boundary-layer region is calculated via an extension of Thwaites' method


for compressible ow [85].

2.6 Three-dimensional boundary layer


The fully three-dimensional steady incompressible boundary-layer equations in integral
formulation are written in a local Cartesian coordinate system (x; y; z ), following the
approach of Mughal and Drela. The x and y directions lie in the plane tangent to the
surface of the boundary layer and z is normal to the plane. The governing three integral
boundary-layer equations are

@
@
@u
@u
(xx qe2 ) + (xy qe2 ) + qe x e + qe y e = w ;
@x
@y
@x
@y

(2.15)

@
@v
@v
@
(yx qe2 ) + (yy qe2 ) + qe x e + qe y e = w ;
@x
@y
@x
@y

(2.16)

1 @
(u
qe @x e

qe x ) +

1 @
(v
qe @y e

qe y ) = CE :

(2.17)

Equations (2.15) and (2.16) are the x- and y -momentum integral equations, which are the
extension of the two-dimensional Von Karman equation (2.1). The entrainment equation
(2.17) can be found by integrating the continuity equation across the boundary layer,
as in two-dimensions [39]. A derivation of the three-dimensional integral boundary-layer
equations in a Cartesian system is given in appendix C.
In the above equations (2.15), (2.16) and (2.17) ue and ve are the velocity edge components in the x- and y -direction, with qe2 = u2e + ve2 the square of the total edge velocity.
The wall shear stress components are given by w and w . The integral thickness definitions for , x , y , xx , xy , yx and yy are presented in appendix A for a Cartesian
coordinate system.
In order to make use of the established closure relations, based on streamwise and
cross ow velocity pro les, it is necessary to work in a streamline coordinate system. The
boundary-layer quantities in the (x; y ) system need to be transformed to the (s; n) system.
Looking again at gure 2.1, it is seen that via vector rotation the (x; y ) quantities can be
expressed in (s; n) quantities. The relationship between the two systems can be given in
terms of the velocities
x

u = us cos un sin ;
v = us sin + un cos ;
where is the angle between the x- and s-directions. With the above relationship the
Cartesian integral thicknesses can easily be determined. In appendix B the streamwise
and cross ow closure relations are given, reducing the number of unknowns to ve, being
ss , H , , ue and ve . Following Smith [82] it is assumed that the empirical relationship
for CE in two dimensions also holds in three dimensions.

20

Chapter 2 Boundary-layer region

Taking into account the above mentioned closure relations and prescribing the velocity
components ue and ve , only three unknowns remain for the three-dimensional integral
boundary-layer equations. The nature of the set of the three partial di erential equations
with prescribed velocities is found to be totally hyperbolic, having three real and distinct
characteristic directions [73, 20]. The characteristics correspond to the tangent of the
angle they make with the x-direction and are bounded by the direction of the external
streamline and the limiting wall streamline. One characteristic almost corresponds with
the limiting wall streamline as was described in [20]. The behaviour of the solutions of
the boundary-layer equations is strongly a ected by the hyperbolic nature the partial
di erential equations, which should be taken into account in the numerical scheme.
The laminar region of the three-dimensional boundary-layer ow is calculated with
the two-dimensional Thwaites' method as given in (2.9). For the cases calculated in this
thesis, the three-dimensional e ects in the small laminar region are negligible, justifying
this approach.

2.7 Discretisation boundary-layer equations


To nd the numerical solution, the integral boundary-layer equations have to be discretised. As discussed, the two- and quasi-three-dimensional formulations are described by
a set of ordinary di erential equations and use can be made of two-dimensional upwind
or central discretisation schemes.
The fully three-dimensional problem is described by a set of partial di erential equations which are totally hyperbolic. The hyperbolicity of the system is to be taken into
account in the discretisation which should be compatible with the characteristic directions.

2.7.1 Discretisation two- and quasi-three-dimensional problems


The two-dimensional and quasi-three-dimensional integral boundary-layer equations both
lead to a system of ordinary di erential equations with dependence solely on upstream
conditions. The discretisation of the streamwise derivative terms in the boundary-layer
equations are therefore to be computed with an upwind scheme, which is rst-order accurate. For example the upwind discretisation for the gradient of s in point i is written

s s 1
@s
=
;
(2.18)
@s i
si
where si = si si i is the distance in s-direction between two adjacent node points.
In order to obtain second-order accuracy, central two-point di erencing (Keller-line)
can be applied. In this case the equations are to be evaluated halfway two adjacent node
points. The values in point i 21 , lying in between i 1 and i, are determined by averaging
between the two stations, thus s 1 = (s 1 + s )=2, and the central derivative becomes
i

@s
@s i

1
2

s

s 1
:
si
i

(2.19)

The above central di erencing has been tested for the two-dimensional aerofoil problem.

2.7 Discretisation boundary-layer equations

21

Discretisation time-derivative
For the unsteady boundary-layer formulation a time-derivative is present in the momentum equations, which is to be discretised. In combination with upwind discretisation, the
time-derivative of s in (2.4) at viscous-inviscid iteration n becomes
@s(
@t

n)

s(

n)

s(
t

1)

(2.20)

with t being the timestep.

2.7.2 Discretisation three-dimensional problem


The system of integral boundary-layer equations that is to be solved for the fully threedimensional ow case is discretised following a nite-volume approach, as used by Mughal
and Drela for their calculation method for three-dimensional integral boundary-layer equations [72]. The use of nite-volume discretisation eliminates the calculation of cumbersome
metric gradients, which have to be evaluated for the traditional boundary-layer formulation in a general curvilinear coordinate system.
The equations in a local Cartesian system (x; y ) to be solved are of the form

P (x; y )

@F (x; y )
@G(x; y )
+ Q(x; y )
@x
@y

R(x; y ) = 0:

(2.21)

If the ow domain is discretised by a grid of general quadrilateral cells, the so-called nite
volumes, then by integrating the above equation over the area A of a cell in the (x; y )
plane gives
ZZ

@F (x; y )
@G(x; y )
P (x; y )
+ Q(x; y )
@x
@y

R(x; y ) dx dy = 0:
(2.22)
A
With a, b, d, e, f and g as de ned in gure 2.3, area A of a general quadrilateral cell is
calculated with
1p 2 2
A =
4f g (b2 + e2 a2 d2 ):
4
d

A
f

g
b

Figure 2.3: Area quadrilateral cell ( + + + = 2 ).

22

Chapter 2 Boundary-layer region

If it is assumed that the terms P (x; y ), Q(x; y ) and R(x; y ) do not vary much over the
region of the cell, then these terms can be taken outside of the integral. Their values are
instead determined at the centre of the cell. With the use of Green's theorem equation
(2.22) can be simpli ed to

P (xc ; yc)

Z
C

F (x; y ) dy Q(xc ; yc)

Z
C

G(x; y ) dx R(xc ; yc)

ZZ

A

dx dy = 0; (2.23)

where C is the contour (anti-clockwise) around the area A and subscript c indicates
that values at the centre of the cell are used. The above equation no longer contains
di erentials, so nite di erencing is no longer required.
The present method for the calculation of the fully three-dimensional boundary-layer
equations will make use of a grid constructed by parallelogram panels, as shown in gure
2.4. Parallelogram panel (i; j ) is de ned by the four edge points (i 1; j 1), (i; j 1),
(i; j ) and (i 1; j ). The angle made by the panel edge (i; j 1) to (i; j ) and the local
y -axis is termed  and corresponds for the present grid with the sweep angle of the wing.
To be able to use the nite-volume discretisation in node point (i; j ) of the parallelogram panel, a cell is chosen with point (i; j ) as its centre. The new cell, in gure 2.4
drawn with the dashed lines, has points n(orth), e(ast), s(outh) and w(est) lying halfway
the cell edges.
The hyperbolic nature of the system of equations, leading to the concept of domains of
dependence and in uence, is to be taken into account in the discretisation. The domains,
seen in gure 2.5, are constructed by the outmost characteristics going through point P .
For node point (i; j ) this implies that when the domain of dependence corresponds
with panel (i; j ) upwinding can be applied in both the x- and y -direction. For this case
the angle , which corresponds with the angle the most outbound characteristic makes
with the x-direction in point (i; j ), is greater than zero and smaller than 90o .
The x- and y -upwind schemes are applied by transporting the values of panel (i; j )
to the dashed panel. Values at point (i; j ) go to n and e and values at points (i; j 1)
and (i 1; j ) go to s and w, respectively. The shape of the dashed cell is chosen to be
equivalent to panel (i; j ) as it provides the values in the points n, e, s and w.
n

tip

1111111111
0000000000
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
(i,j+1) (i+1,j+1)
0000000000
Uoo 1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
(i,j)
0000000000
1111111111
(i+1,j)
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
root

11
00
00
11

(i-1, j-1)

(i-1, j)
1
0
0
1

(i, j)

11
00
00
11

local
z
x
PANEL
(i, j)

11
00
00
11

(i, j-1)

Figure 2.4: Parallelogram panels.

2.7 Discretisation boundary-layer equations

23

Domain of
1111111111111
0000000000000
influence
0000000000000
1111111111111
0000000000000
1111111111111
0000000000000
1111111111111
0000000000000
1111111111111
0000000000000
1111111111111
0000000000000
1111111111111
0000000000
1111111111
0000000000
1111111111
0000000000 P
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
Domain1111111111
of
0000000000
0000000000
1111111111
dependence
0000000000
1111111111

Figure 2.5: Domains of dependence and in uence.


The application of the described x- and y -upwind nite-volume approach to the xmomentum boundary-layer equation (2.15)

@
@u
@u
@
(xx qe2 ) + (xy qe2 ) + qe x e + qe y e
@x
@y
@x
@y

w

= 0;

in point (i; j ) becomes

xx qe2 yijij 1 + xx


ij

(xy

ij

q2

1 eij 1

ij

q2

1j ei 1j

xij 1i 1j 1 + xy qe2 (xijij 1 + xi 1jij ) + xy


ij

ij

+x qe (ue yijij 1 + ue


ij

y qe (ue
ij

ij

yi 1j 1i 1j

ij

ij

ij

1j

q2

1j ei 1j

xi 1j 1i 1j )

yi 1j 1i 1j )

xij 1i 1j 1 + ue (xijij 1 + xi 1jij ) + ue


ij

1j

xi 1j 1i 1j )

w A = 0;
xij

where xklmn = xkl xmn and yklmn = ykl ymn . The discretisation of the y -momentum
equation and entrainment equation is analogous to the above.
For the special case of a rectangular panel, the above x- and y -upwind nite-volume
discretisation for the x-momentum boundary-layer equation will lead to
xx qe2 xx 1 qe2 1 xy qe2 xy 1 qe2 1
+
+
x
y
ij

ij

ue
qe x

ij

ij

ij

ij

ue
x

1j

ij

ij

ue
+ qe y

ij

ij

ij

ij

ue
y

ij

= w ;
xij

where x and y de ne the lengths of the cell edges in the x- and y -directions.
When the domain of dependence corresponds with panel (i; j + 1), 90o  <  < 0
in point (i; j ), downwinding should be applied in the y -direction. The values of panel
(i; j + 1) are then transported to a dashed cell of equivalent shape as panel (i; j + 1).

24

Chapter 2 Boundary-layer region

Values at point (i; j ) go to e and s and values at points (i; j + 1) and (i 1; j ) go to n


and w, respectively.
When the domain of dependence corresponds with panel (i +1; j ) or panel (i +1; j +1),
implying   90o  or   90o , downwinding should be applied in the xdirection. However, for most cases with  not too large, the two previously discussed
x-upwind schemes are sucient. A scheme dealing with characteristics coming from
upstream therefore has not been implemented.

2.8 Boundary conditions


Boundary conditions need to be prescribed at the attachment line for the three-dimensional
ow case and at the stagnation point for the two-dimensional ow case to start the calculations. Conditions are also required at the transition line, when transition from laminar
ow to turbulent ow takes place.
For the fully three-dimensional integral boundary-layer equations boundary conditions
are required at the wing root and tip, taking into account the ow of information.

2.8.1 Attachment line/stagnation point


The ow along the attachment line is assumed to correspond with the ow along the
leading edge of an in nite-span wing. If the y -direction is taken parallel to the leading
edge, the three-dimensional boundary-layer equations for an in nite-span wing reduce to
a decoupled system (@=@y = 0). The continuity equation and the x-momentum equation
are independent of v and equivalent to the two-dimensional boundary-layer equations
[104]. From [22] it is known that in the immediate vicinity of the attachment line ue = kx
and that ve is constant. Flow along the attachment line can therefore be modelled as
two-dimensional stagnation ow with ve = v1 .
For the various boundary-layer prediction methods the initialisation at the threedimensional attachment line or two-dimensional stagnation point is now as follows. The
velocity components at the edge of the boundary layer are initialised with the external
purely inviscid velocity and the angles and c are set to zero. For the laminar ow case
the two-dimensional stagnation ow is assumed to correspond with the stagnation ow
around a small circle and

s = 0:6478

r
;
2Re

H = 2:216;

(2.24)

where r is the radius of the small circle taken to be 0:02 [19]. The momentum thickness
is determined via ss = s =H .
For turbulent ow the following is prescribed:
0:005
;
H = 1:35;
= 0:
(2.25)
Re0:2
It is noted that for the three-dimensional wing problem the stagnation line does not
have to coincide with one and the same spanwise gridline.

ss =

2.8 Boundary conditions

25

2.8.2 Transition line


Transition to turbulence is to be included in the various boundary-layer prediction methods. The most common method to model transition is to set speci c transition points
in correspondence with the determined experimental locations. In the experiment wire
strips are used to enforce the transition from laminar to turbulent ow, also called tripping. Transition modelled with xed locations is therefore termed tripped transition.
Other two-dimensional transition mechanisms have also been used and are next to be
discussed.
For the three-dimensional ow problems, no special three-dimensional transition mechanisms have been incorporated. It is assumed that the ow in the small laminar region is
two-dimensional and again the two-dimensional mechanisms are applied.

Natural transition
At high Reynolds numbers natural transition occurs before laminar separation, after which
the ow almost immediately turns turbulent. The empirical method that can be used to
determine the point of transition is based on a paper of Arnal et al. [2]. This general
method, written in integral parameters, is derived from theoretical results for the OrrSomer eld equation [79]. First, the point at which the ow becomes unstable is calculated
Re

ssuns

52
= exp
H

14:8 :

(2.26)

The transition is said to be complete when


206 exp[25:7  2t ln(16:8T F ) 2:77 2t ];
(2.27)
with T F the freestream turbulence and  2t the average value of the Pohlhausen parameter
between the instability point xuns and the transition point xt

Re

sst

Re

ssuns

1
 2t =
xt xuns

Zxt

2 dus
ss
ds:
 ds
e

xuns

(2.28)

The above formula (2.27) is a modi cation of Granville's empirical criterion which ignored
freestream turbulence [38].

Laminar separation bubble


For low Reynolds number ow, the suction peak near the leading edge can make the laminar boundary layer want to separate before natural transition has taken place. The ow
has become unstable (Tollmien-Schlichting waves), due to surface roughness or freestream
turbulence. Separation followed by transition to turbulence takes place and, if the pressure recovery is sucient, the free shear layer may reattach. If the ow reattaches we
speak of a short laminar separation bubble [79].

26

Chapter 2 Boundary-layer region

To detect leading edge separation, a criterion based on Thwaites' method is used. The
ow is said to be separated when
2 dus
ss
 0:09:
(2.29)
 ds
Alternatively, the ow is assumed to be separated when Cf s has reached a very small
positive value, close to zero. Both criteria have been used.
At high angles of attack, or at very low Reynolds numbers, the free shear layer might
fail to reattach. This is called bubble bursting. After bursting the ow might reattach
as a long bubble or stay separated, which reduces the pressure gradient over the bubble
signi cantly.
A short bubble, on the other hand, only has a small in uence on the pressure distribution. Outside the bubble region the pressure distribution resembles an inviscid solution.
Inside the bubble the pressure will remain almost constant over the laminar part and will
then decrease linearly to approximately the inviscid solution at reattachment.
To be able to model a laminar separation bubble, Horton's model [46] is used, modi ed
by Roberts [77] to include the e ects of the freestream turbulence T F .
The bubble is divided into two parts: a laminar part of length l1 and a turbulent part
of length l2 . The length of the laminar part of the bubble is given by
e

l1
K
=
;
ss
Re
s

c
K = 25000 1 log10 (100 T F )
Ls

sss

1 
5

with K an approximation of Roberts result. The subscript s indicates that the value is
taken at separation. Furthermore, c is the local chord length and Ls is a macroscale of
the ow turbulence, which is conventionally taken here to be equal to the chord length.
Because the pressure gradient is zero between separation and transition there is no
growth of ss and ss = ss , subscript t indicating transition.
All the possible loci for l2 are given with the relation below. The bubble is presumed
to be reattached when the velocity of the free shear layer compared with the velocity of
the inviscid ow is the same. Hence, knowing the inviscid ow solution, the correct l2 is
determined. When l2 is known the momentum thickness at reattachment, ss , is found
as well
l2
85:23(1 ue )
=
;
(2.30)
ss
4ue 0:497
4
ss = 1 + 0:0058 l2 1 ue :
(2.31)
u3e
ss u3e (1 ue )
t

The bar indicates normalisation with respect to the values at separation.

Tripped transition
The most used method to model transition is simply to de ne the transition points which
is also known as tripping the boundary layer. It allows good comparison with experimental
results in which wire transition trips have been applied.

2.8 Boundary conditions

27

Furthermore, if the empirical methods give problems with transition prediction at


higher angles of attack, tripping is applied. On the other hand, when the trip locations
are not suitable for the calculation, one of the previously mentioned empirical transition
mechanisms is automatically employed.

2.8.3 Root boundary condition


The wing root corresponds with the symmetry plane of the aeroplane (without fuselage)
and the ow along the root should be parallel to this plane.
The values of the velocity in the Y -direction, normal to the symmetry plane, and
the limiting wall streamline angle are set to zero at the root. The values of the other
variables are prescribed by the values at the gridline next and parallel to the root, in
order to have the derivatives in the direction normal to the symmetry plane at the root
equal to zero.

2.8.4 Tip boundary condition


As the boundary-layer equations are to be solved with the quasi-simultaneous viscousinviscid interaction method, requiring downstream in uence for particular cases (see chapter 4), boundary conditions need to be provided at the tip.
For the three-dimensional wing ow calculations to be presented in this thesis the
ow of information at the tip is going out of the calculation domain, and the values of
the variables at the tip therefore are determined via extrapolation from the values at the
adjacent gridline on the wing.

28

Chapter 2 Boundary-layer region

Chapter 3
External ow region
3.1 Introduction
For the modelling of the ow outside the boundary layer an inviscid ow solver is needed.
As it is of primary importance in the work described in this thesis to develop a fast and
ecient computational method, suitable for a design-optimisation environment, a surface
singularity method for irrotational subsonic ow is preferred to a physical more complete
Euler or full-potential- ow solver.
In the next sections the potential- ow problem is discussed for which solutions can
be obtained by distributing singularity elements on the boundaries. First, the purely
inviscid potential- ow problem with no boundary layer present is examined. Later on,
the boundary-layer displacement e ect is taken into account, by modelling an extra source
singularity distribution over the pro le's surface and the wake. This is termed the viscous
formulation of the potential- ow problem.
The unknown strengths of the singularities are to be determined numerically. This
is done straightforwardly with a panel method, which requires the discretisation of the
geometry into a number of panel segments.
Various problems are discussed, of which the model problem of the ow over at plates,
which corresponds with non-lifting thin-aerofoil/wing theory, is described in greater detail.
The constant potential Dirichlet method is applied for the more practical problems of
modelling wing and aerofoil ow. For each of these problems the edge velocity equations
are determined, required for the coupled potential- ow/boundary-layer calculations.

3.2 Inviscid steady potential- ow problem


Consider an arbitrary body within an enclosed three-dimensional volume V , with the
contours of the body surface and wake, Sb and Sw , respectively, and the outer control
surface S1 , surrounding the body and the wake surface, as shown in gure 3.1.
The unit normal n is the vector pointing towards the inside of volume V and towards
the outside of Sb and Sw . The body, which has a body- xed Cartesian coordinate system
(X; Y; Z ), is subject to the in ow velocity Q1 = (U1 ; V1; W1 )T .
The inviscid ow inside domain V over the solid body when assumed incompressible

30

Chapter 3 External ow region

p(X,Y,Z)
r
Y

q(,,)

Sw

Sb

V,
S

Figure 3.1: Nomenclature potential- ow problem.


and irrotational, can be treated as potential ow [71]. This is governed by the Laplace
equation
r2  = 0;
(3.1)
for the velocity potential (X; Y; Z ).
On the solid body surface Sb in the ow regime there is zero normal velocity which
can be expressed as the Neumann boundary condition

r  n

= 0:

(3.2)

On the outer control surface S1 the disturbance due to the presence of the body must
vanish in the limit and the potential must be equal to the undisturbed freestream potential
1
lim (r

S1

r1)

!1

= 0:

(3.3)

Once  has been obtained, the pressure follows from Bernoulli's equation.

3.2.1 Formulation of the integral equation


Using Green's identity the above constructed boundary-value problem can be transformed
into an integral equation for the potential  [50]


ZZ
1 1
r  n dS;
(3.4)
4E (p)(p) =
r
r r
S +S +S1
with r(p; q ) the distance between the point of integration q (; ;  ) and a xed eld point
p(X; Y; Z ) at which the potential is calculated
p
(3.5)
r(p; q ) = (X  )2 + (Y  )2 + (Z  )2 :
b

3.2 Inviscid steady potential- ow problem

31

The function E (p) in (3.4) is given by


8
>
> 0; for p outside the ow eld V ;

E (p) =

>
>
<
>
>
>
>
:

1
2 ; for p on the boundary of V ;

(3.6)

1; for p inside the ow eld V :

If the ow eld inside the body is considered, an equation similar to (3.4) can be derived
for the internal potential in [50]


ZZ
1
1
4 (1 E (p))in (p) =
r in r r  n dS:
(3.7)
r in
S +S
Equations (3.4) and (3.7) can be combined to a form relating  and in
w

4 (E (p)(p) + (1 E (p))in (p)) =




ZZ

(3.8)


ZZ

1 1
1 1
r
( in )  n dS +
r
r  n dS:
( in )r
r r
r r
S1
S +S
The wake surface Sw is assumed to have zero thickness, such that the normal velocity
jump across the wake dividing streamline Sw is zero, while a jump in the potential is
allowed. With this condition the problem reduces to determining the values of  in
and r( in )  n on the boundaries. The di erence in external and internal potential
can be provided by a continuous doublet distribution of strength  over the surface of the
body and the wake. The di erence between the normal derivatives of the external and
internal potential can be induced by a continuous source distribution of strength  on the
body
b

 =  in ;
 = r( in )  n:
(3.9)
Using the above de nitions for  and , the potential for points p inside the ow regime
V is given by
ZZ
ZZ
1

1
1
(p) =
dS +
 n  r dS + 1 (p);
(3.10)
4 S r
4 S +S
r
with the freestream potential, de ned as 1 (p) = U1 X + V1 Y + W1 Z , to satisfy the
condition at in nity. In two dimensions equation (3.10) is written as [50]
Z
Z
1
1
(p) =
 ln rdS
 n  r(ln r) dS + 1 (p):
(3.11)
2 S
2 S +S
Solving the Laplace equation (3.1) has now been reduced to nding an appropriate
singularity distribution over the boundaries so that the boundary conditions (3.2) and
(3.3) are satis ed. It is noted, however, that equations (3.10) and (3.11) do not immediately specify a unique combination of sources and doublets for a particular problem.
Additional considerations are usually required which depend on the physics of the problem (i.e. Kutta condition to de ne the ow near sharp trailing edges) to relate  in the
wake to  on the body.
b

32

Chapter 3 External ow region

3.2.2 Model problem


From the previous section it is clear that a solution for the ow over arbitrary bodies can
be obtained by a singularity distribution on the modelled surface. Before applying the
described method to practical problems, the method is rst investigated for the model
problem of two- and three-dimensional non-lifting small disturbance ow over dented
plates, which corresponds with non-lifting thin-aerofoil/wing theory.
Let the surface of the dented plate be given by Z = Zp(X; Y ), with Zp assumed to be
small compared to the length and the width of the plate. In order to nd the normal to
the surface, the function F (X; Y; Z ) is de ned

F (X; Y; Z )

 Z Zp(X; Y )

= 0:

(3.12)

The unit normal vector n, pointing outward on the upper surface of the dented plate is
determined by

n =

rF
jrF j

1
=
jrF j

T

@Zp @Zp
;
;1
@X @Y

(3.13)

The velocity potential due to a freestream Q1 = (U1 ; V1; 0)T , corresponds with 1 =
U1 X + V1 Y , and the potential can be constructed to be
 =  + 1 ;

(3.14)

which has to ful l the boundary condition (3.2) of zero normal velocity at the surface of
the plate

r  n = jrF1 j

@
@Zp
@
@Zp
+ U1
+
+ V1
@X
@X
@Y
@Y

@
@Z

= 0:

(3.15)

Consequently, the Laplace problem r2  = 0 is to be solved for the perturbation potential


 with boundary condition (3.15) for @=@Z on Z = Zp. Assuming the geometry of the
dent to be shallow, the perturbation velocity can be taken to be small compared to the
freestream


@


@X ;



@


@Y ;



@


@Z

 Q1 ;

(3.16)

where Q1 = jQ1 j and boundary condition (3.15) reduces to

@
@Z
@Z
(X; Y; Z = Zp) = U1 p + V1 p :
(3.17)
@Z
@X
@Y
With the use of a Taylor expansion, boundary condition (3.17) at the plate's surface is
transferred to the (X; Y ) plane (Z = 0)
@Z
@Z
@
(X; Y; 0) = U1 p + V1 p :
@Z
@X
@Y

(3.18)

3.2 Inviscid steady potential- ow problem

33

The above de ned potential- ow problem for the perturbation potential  and boundary
condition (3.18) can be solved by a source distribution on the (X; Y ) plane. Assuming
a source distribution of strength  on the (X; Y ) plane over the region X 2 [0; 1] and
Y 2 [0; 1], the velocity potential and velocity eld are given by
1
(X; Y; Z ) =
4

@
1
=
@X
4
@
1
=
@Y
4
@
1
=
@Z
4

Z1 Z1
0

Z1 Z1
0

Z1 Z1
0

Z1 Z1
0

 (;  ) dd
1;
 )2 + (Y  )2 + Z 2 ) 2

(3.19)

 (;  )(X  ) dd


3 ;
((X  )2 + (Y  )2 + Z 2 ) 2

(3.20)

 (;  )(Y  ) dd


3 ;
((X  )2 + (Y  )2 + Z 2 ) 2

(3.21)

 (;  )Z dd
3 :
 )2 + (Y  )2 + Z 2 ) 2

(3.22)

((X

((X

The normal velocity @=@Z (X; Y; 0) at the surface is related to the source strength,
see gure 3.2. This can be shown either by a limit process for Z ! 0 in equation (3.22),
0+ and 0 representing the upper and lower surface, respectively, or by observing the
ux over the area X by Y . The ux due to a source of strength  (X; Y ) over the area
X by Y is given by

 (X; Y ) X Y:

(3.23)

At the same time, the ux due to the ow rate can be found to be

@
@
@
(X; Y; 0+) X Y
(X; Y; 0 ) X Y = 2 (X; Y; 0+) X Y; (3.24)
@Z
@Z
@Z
the uxes from the side becoming negligible when Z ! 0. Modelling the ow over the
upper surface of the plate, only 0+ is of interest and @=@Z (X; Y; 0 )= @=@Z (X; Y; 0+)
Z

(X,Y)
Zp

d
(X,Y,0+)
dZ

V
U

X
d
(X,Y,0-)
dZ

Figure 3.2: Source distribution for a dented plate.

34

Chapter 3 External ow region

= @=@Z (X; Y; 0). The two uxes (3.23) and (3.24) should be the same and

@
(X; Y; 0):
(3.25)
@Z
Comparing (3.25) with boundary condition (3.18), it is seen that modelling the perturbation potential over a dented plate corresponds with a source distribution over the (X; Y )
plane when
 (X; Y ) = 2

@Z
@Z
 (X; Y ) = 2 U1 p + V1 p :
@X
@Y

(3.26)

The velocity eld at the surface of the dented plate, Zp being small, is now given by

U (X; Y; Z = Zp )

V (X; Y; Z = Zp )

Z1 Z1

W (; ; 0)(X  ) dd


2 + (Y  )2 ) 23 + U1 ; (3.27)
((
X

)
0 0
Z1 Z1
W (; ; 0)(Y  ) dd
1
V (X; Y; 0) =
2 + (Y  )2 ) 23 + V1 ; (3.28)
2
((
X

)
0 0
1
U (X; Y; 0) =
2

where W (X; Y; 0) = @=@Z (X; Y; 0) as de ned in (3.18). In two dimensions a similar


result is found [50]

U (X; Z = Zp )

1
U (X; 0) =


Z1

W (; 0)d
+ U1 ;
X 

(3.29)

where

W (X; 0) =

@
@Z
(X; 0) = U1 p :
@Z
@X

(3.30)

3.2.3 Wing/aerofoil problem


In the previous section 3.2.2 the modelling of the inviscid ow over dented plates was
discussed with the boundary condition applied on the projection of the dent in the (X; Y )
plane (Z = 0). In order to more realistically model the geometry of a wing or aerofoil the
boundary conditions should be applied at the actual surface. The singularity elements
are to be distributed over the real surface to model the whole ow eld, and the problem
is reduced to nding the source and doublet strengths so that the boundary conditions
(3.2) and (3.3) are satis ed.
Boundary condition (3.3) is automatically ful lled as the singularity solutions have a
velocity eld that decays for r ! 1. For boundary conditions (3.2) two formulations
exist. If the boundary-value problem is solved explicitly using boundary condition (3.2)
we speak of a Neumann problem, and equations (3.10) and (3.11) are di erentiated in
order to solve for the total velocity eld Qtot

Qtot =

r :

(3.31)

3.2 Inviscid steady potential- ow problem

35

Instead of applying the direct boundary condition (3.2), it is also possible to de ne  as


a constant. This is termed the Dirichlet formulation of the problem.
For the work described in this thesis the Dirichlet method is applied on the wing's
surface for the modelling of the really inviscid wing ow. The solution in the wake
region follows, once the Dirichlet problem on the body is solved. The modelling of both
regions is discussed next in more detail. However, rst the Kutta condition is de ned
to determine the problem uniquely by relating the doublet strengths in the wake to the
doublet strengths at the trailing edge.

Kutta condition
In the case of modelling wing or aerofoil ow the problem is not uniquely speci ed by
only the boundary condition of no normal velocity through the pro le's surface.
A condition is required at the trailing edge to relate the doublets in the wake to the
body doublets in order to uniquely determine the solution. The Kutta condition is used
for this and in its most general form requires that the velocity at the trailing edge is
bounded

jrjte < 1:

(3.32)

For the present formulation an implicit Kutta condition is used. The potential jump
across the wake is set equal to the di erence between the potential values at the upper
and lower surface at the trailing edge. This is the same as relating the doublet strength in
the wake w to the unknown doublets te and te at the upper and lower trailing edge
of the wing's body as
u

w = te

te :

(3.33)

Dirichlet method
Specifying on the boundary the potential inside the body surface in , instead of using
(3.2), is called the Dirichlet boundary condition and can be employed to determine the
solution of the boundary-value problem (3.1). For eld points p inside the body, the
internal potential is given by (3.10)
1
4

in =

ZZ

Sb


1
dS +
r
4

ZZ

Sb Sw

@ 1
dS + 1 :
@n r

(3.34)

With the Dirichlet boundary condition, in = in + 1 = 1 , where the internal
perturbation potential in is set to zero, equation (3.34) simply becomes
ZZ

ZZ

@ 1
1


dS =
dS;
@n r
4 S r
S +S
which for the two-dimensional ow case corresponds with
1
4

1
2

@
1
 (ln r) dS =
@n
2
S +S
b

(3.35)

Sb

 ln r dS:

(3.36)

36

Chapter 3 External ow region

From the freestream the source strength  can be determined, using the de nition of
source strength (3.9)

@ 1
= Q1  n;
(3.37)
@n
with in = 1 and the normal velocity @ =@n at the surface being zero.
When applying the implicit Kutta condition (3.33), linking the doublet strengths in the
wake to the upper and lower trailing edge doublets, the above constant-potential Dirichlet
problem is uniquely de ned and the unknown doublet distribution can be determined.
Assuming the wake to have zero thickness, the velocity distribution everywhere within
the ow regime V is automatically determined by the source and doublet distribution
provided by the Dirichlet method on the pro le's surface. Equations (3.10) and (3.11),
di erentiated in the normal direction with respect to the eld point p, form the integral
equations for the velocity eld in the wake and for the three-dimensional ow problem
given by equation (3.10) this results in


Qtot (p) =
=

in
 @@n @@n

= 0

r(
p)

ZZ
ZZ
1

1
r 4
dS +
r
4
Sb

@ 1

dS + 1 (p) :
@n r
S +S
b

(3.38)

3.2.4 Full-potential problem


As part of this thesis, contract work has been carried out for the Defence Evaluation and
Research Agency (DERA) in Farnborough, United Kingdom, and use has been made of
their existing Viscous Full-Potential (VFP) code for transonic ow calculations.
The inviscid ow part of the VFP program is based on the method developed by Forsey
and Carr [36] for calculating potential ow over wing-fuselage combinations. It is valid
for non-zero, subsonic Mach numbers and the scheme can be applied to con gurations
with swept, cranked and tapered wings. As the full-potential (FP) part of the VFP code
is treated as a blackbox it will not be discussed in this thesis. More information can be
found in the papers [33, 34, 4].

3.3 Viscous potential- ow problem


In the previous sections the modelling of the inviscid external ow has been discussed,
ignoring the presence of the boundary layer. However, as the boundary layer displaces
the outer potential ow away from the surface over a distance, termed the displacement
thickness, this boundary-layer displacement e ect should be taken into account into the
external ow modelling.
The potential ow should be modelled over the resulting e ective body shape, constructed from the actual physical body plus the displacement thickness. This can be done
via the distribution of singularities over the resulting e ective body shape which forms a
streamsurface [60]. A simpler approach is to change the zero normal velocity boundary
condition at the original surface in such a way that the normal velocity at the real surface

3.3 Viscous potential- ow problem

37

in the potential ow corresponds with the normal velocity at the edge of the boundary
layer.
Before continuing, a local Cartesian coordinate system (x; y; z ) is de ned where the
x- and y -axis are tangent to the surface and the z -axis is normal to the surface. The
local total velocity vector is given by qtot = (u; v; w; )T . The velocity at the edge of the
boundary layer is indicated with the subscript e, thus ue and ve are the x- and y -velocity
components at the edge of the boundary layer in the local Cartesian system.
Using the continuity equation and assuming that w grows linearly with (@ue =@x+
@ve =@y ) z within the boundary layer, the following approximation for the normal velocity
in the boundary layer can be obtained:

w(x; y; z )

@ue @ve
+
z + w1 (x; y ); z ! 1;
@x @y

(3.39)

where z ! 1 indicates the edge of the boundary layer. The disturbance velocity w1 6= 0 is
the next term of the expansion of w for z ! 1 and represents the so-called transpiration
velocity, describing the transpiration between the boundary layer and the external ow.
The de nition for the transpiration velocity w1 is determined by

@ue
@ve
z
+
z]
w1 (x; y ) = zlim
[
w
(
x;
y;
z
)
+
!1
@x
@y
Zz
@w
@ue
@ve
= zlim
[
dz
+
z
+
z]
!1 @z
@x
@y
0

Zz 
@ue
@ve
@u @v
dz
+
z
+
z]
= zlim
[
!1
@x @y
@x
@y
0


Zz 
Zz 
@ue @u
@ve @v
= zlim
[
dz +
dz ]
!1
@x @x
@y @y
0
0
Z1
Z1
@
@
=
(ue u)dz +
(ve v )dz
@x
@y
0
0
@
@
=
(qe x ) + (qe y );
@x
@y

(3.40)

p
where qe = u2e + ve2 + we2 is the velocity magnitude at the edge of the boundary layer.
The integral thickness de nitions in the local Cartesian coordinate system for the displacement thicknesses x and y in the x- and y -directions can be found in appendix
A.
Physically, the above transpiration velocity represents the rate of change of mass
defect, being mx = qe x and my = qe y in the local Cartesian coordinate system.
In two dimensions equation (3.40) reduces to

w1 (x) =

@
(q  ):
@x e x

(3.41)

38

Chapter 3 External ow region

To include the viscous in uence in the potential- ow problem described in the previous sections, the `zero' normal velocity boundary condition at the body surface (3.2) is
modi ed, giving

r  n

= w1 (x; y ) =
6 0:

(3.42)

For the discussed potential- ow problems the change in boundary condition, due to the
presence of the boundary layer, results in modifying the source strengths. For the threedimensional dented plate problem, where the global coordinate system corresponds with
the local coordinate system, the source strength (3.26) becomes


@
@
@
@
3D = 2
(U1 Zp) + (V1 Zp) + 2
(qe x ) + (qe y ) = inv + vis ; (3.43)
@x
@y
@x
@y
or in two dimensions

@
@
(U1 Zp) + 2 (qe x ) = inv + vis ;
(3.44)
@x
@x
where inv represents the inviscid source strength based on the freestream and vis represents the viscous source strength due to the presence of the boundary layer.
For the wing and aerofoil problem the viscous sources should be included on the body
and in the wake, and equation (3.37) for the source distribution on the body surface
changes to
2D = 2

Q1  n + w1 = inv + vis :

 =

(3.45)

In the wake, the displacement e ect is a jump in normal velocity w1 across the wake
centre line, dividing the ows coming from the upper and lower surfaces of the pro le
w1 =

@
@
(qe x + qe x ) + (qe y + qe y );
@x
@y
u

(3.46)

where subscripts u and l indicate upper and lower surface values, respectively.
Dependent on the modelling of the boundary layer for the various ow cases, the
wake's upper and lower layers will be treated separately or together as just one layer. In
the rst case the above boundary condition (3.46) on the wake dividing streamsurface
leads to the following viscous source strength in the wake:

 =

@
@
(qe x + qe x ) + (qe y + qe y ) = vis :
@x
@y
u

(3.47)

For the case where the wake is modelled as just one layer with the wake thicknesses de ned
with x = x + x and y = y + y , and the velocity qe = qe = qe , the viscous source
strength becomes
w

vis =

@
@
(qe x ) + (qe y ):
@x
@y
w

(3.48)

3.4 Discretisation viscous potential- ow problem

39

The general formulation of the two-dimensional potential- ow problem in (3.11) with


the inclusion of viscous e ects can be written
Z
Z
1
1
 ln r dS
( n  r(ln r) vis ln r) dS + 1 (p); (3.49)
(p) =
2 S inv
2 S +S
b

with the inviscid and the viscous source strengths on the body de ned in (3.45). The
wake is modelled as just one layer, implying that vis in the wake is given by the twodimensional version of (3.48).

3.4 Discretisation viscous potential- ow problem


For the discretisation of the potential-based equations use is made of a panel method.
The body surface, and (if present) the wake surface, are divided into a number of panel
elements as seen in gure 3.3. For the three-dimensional case planar panels are used,
whereas in two dimensions the panels simply reduce to straight lines.
The discretisation for the viscous potential problem for indented plate ow and wing/aerofoil ow is examined, and the velocity equations at the edge of the boundary layer
required for the coupled viscous/inviscid problem are determined.
For clarity, rst the used local and global Cartesian coordinate systems are re-introduced.

3.4.1 Global and local coordinate systems


In section 3.2 the general formulation of the potential- ow problem has been described
in a global body- xed Cartesian coordinate system (X; Y; Z ) with velocity vector Qtot =
(U; V; W )T . When the geometry is discretised, it is useful to de ne for each panel segment
a local panel Cartesian coordinate system (x; y; z ) with velocity vector qtot = (u; v; w)T .
For the model problem of ow over a dented plate, the local and global Cartesian
coordinate systems are chosen to correspond with each other, with the origin de ned
at a corner point of the plate. However, for the panel method applied for a wing or
aerofoil section the panels no longer lie in the same plane. Each panel has a di erent local
individual
panel

local
y
x

Z
global

Y
tip

leading
edge
Q

trailing edge
WING

X
root

Figure 3.3: Global (X; Y; Z ) and local (x; y; z ) Cartesian coordinate systems.

40

Chapter 3 External ow region

coordinate system, with the origin chosen in the midpoint of each panel, see gure 3.3.
The z -axis is taken to be normal to the panel and the x- and y -axis are tangent to the
panel.

3.4.2 Discretisation two-dimensional model problem


A two-dimensional plate is de ned between x1 and xM +1 along the x-axis and integral
(3.29), which represents the x-velocity component at the edge of the boundary layer,
becomes
p

1
ue (x) =


xMp +1

x1

we( )d
+ u1(x);
x 

(3.50)

where 2we = inv + vis as given in equation (3.44). For the discretisation of the above
integral, segment [x1 ; xM +1 ] is divided into Mp straight-line panels each of similar length
x = xk+1 xk for k 2 [1; Mp ]. The collocation points are de ned at the panel endpoints
x = xi for i 2 [2; Mp] and
p

ue

1
=


xMp +1

we( )d
+ u1 :
xi 

(3.51)

x1

It is seen that for  = xi the above integral is singular and in the region surrounding the
singularity [xi 1 ; xi+1 ] a linear expansion Lwe of we is to be used. The Cauchy principal
part (C:p:p:) of the singular integral can be calculated and (3.51) is

ue

1
=

1
=


Mp
X
k=1
k=i 1;i

Mp
X
k=1
k=i 1;i

where

Lwe = we + (
i

x
Zk+1

we + 1 d 1
2
+
xi 


x
Zi+1

xk

we + 1
k


x
ln i
x
i

Lwe d
+ u1
xi 
i

xi 1

xk 2x @we
+u :
xk+1  @x i 1
{z
}
|

(3.52)

C:p:p:

@w
xi ) e
@ i

@
@2
(u1Zp + qe x )ji + ( xi ) 2 (u1Zp + qe x )ji :
(3.53)
@
@
Using the above de nition of Lwe and centrally discretising the gradients of u1Zp and
qe x , equation (3.52) becomes
=

ue =

Mp
X

k=1
k=i 1;i



qe +1 x +1 qe x i k
qe +1 x +1 2qe x + qe 1 x
ln
2
 x
i k 1
 x
k

+ u0 ;
ei

3.4 Discretisation viscous potential- ow problem

which can be rewritten to the form


M +1
X
ue =
Auik qe x + u0 ;
k =1
p

ei

41

i = 2; : : : ; Mp ;

(3.54)

where u0 is the undisturbed external inviscid ow at the edge of the boundary layer due
to the freestream and the geometry of the dent. The summation part on the right-hand
side of equation (3.54) describes the disturbance e ect due to the presence of the boundary
layer, with matrix Au constructed as follows:
e

4
;
i = 2; : : : ; Mp ;
 x
1
(2 ln 2);
i 6= 2;
Auii 1
=
 x
1
Auii+1
=
(2 ln 2);
i 6= Mp ;
 x


1
1
u

Aik
=
ln 1
;
k 6= 1; i 1; i; i + 1; Mp;
 x (i k)2
2
;
Au21
=
 x
1
i
ln
;
i 6= 2;
Aui1
=
 x i 1
1
M +1 i
AuiM +1 =
ln p
;
i 6= Mp ;
 x
Mp i
2
AuM M +1 =
:
 x
The constructed in uence matrix Au has the following characteristics:

Auii

Auii > 0;
Auik = Auki < 0; (i 6= k),
M +1
X
Auik  0; with at least for one i the inequality sign.
k =1

(3.55)

From (3.55) it follows, that matrix Au is a symmetric weakly diagonally dominant Mmatrix (de nition D.3 and theorem D.10) and therefore positive de nite (theorem D.12).

3.4.3 Discretisation three-dimensional model problem


In the Cartesian coordinate system a rectangular plate is de ned between x1 and xM +1
and between y1 and yM +1 in the (x; y ) plane. The segments along the x-axis are de ned
by the intervals [xk ; xk+1] of length x for k 2 [1; Mp] and along the y -axis by [yl ; yl+1] of
length y for l 2 [1; Ms ]. The grid is constructed by rectangulars, resulting in panels of
x by y . The double integrals are evaluated in the node points of the panels (xi ; yj ),
p

42

Chapter 3 External ow region

for i = 2; : : : ; Mp and j = 2; : : : ; Ms and the velocity components (3.27) and (3.28) at


the edge of the boundary layer become

ue

ij

ve

ij

1
=
2
=

1
2

xMp +1 yMs +1
Z
Z

we (;  )(xi  )dd


3 + u1 ;
((xi  )2 + (yj  )2 ) 2

(3.56)

we (;  )(yj  )dd


3 + v1 :
((xi  )2 + (yj  )2 ) 2

(3.57)

ij

x1
y1
xMp +1 yMs +1
Z
Z

ij

x1

y1

It is seen that the above integrals, when evaluated in  = xi and  = yj are singular.
Outside a region surrounding the singularity, the integrals are well-behaved. As in the
two-dimensional case it is assumed that a linear source distribution exists over the region
xi 1 < x < xi+1 and yj 1 < y < yj +1, including the singularity. The Cauchy principal
part of the singular integral can be calculated and equation (3.56) is rewritten

ue

ij

1
=
2

Mp
X

1
2

Mp

yl

Lwe (xi  )dd


3 + u1
((xi  )2 + (yj  )2 ) 2
ij

ij

xi 1 yj 1

k=1
l=1
k=i 1;i l=j 1;j

xk

xi+1 y
Z
Zj +1

Ms
X

we + 1 + 1 (xi  )dd
2 2
3
((xi  )2 + (yj  )2 ) 2
k

k=1
l=1
k=i 1;i l=j 1;j

+
1
=
2

x
yl+1
Zk+1 Z

Ms
X

we + 1 + 1
k

2l 2

p

(xi
ln p

xk+1 )2 + (yj yl )2 + (yj


xk+1 )2 + (yj yl+1 )2 + (yj

(xi

p
!
(xi xk )2 + (yj yl+1 )2 + (yj yl+1 )

+ ln p
(xi xk )2 + (yj yl )2 + (yj yl )

p

y ( x2 + y 2 + x)2 @we
+u ;
ln p
2 ( x2 + y 2 x)2 @x ij 1ij
{z
}
|

yl )
yl+1 )

(3.58)

C:p:p:

with Lwe the linear expansion of we

Lwe

ij

= we + (
ij

@w
xi ) e + (
@ ij

@w
yj ) e ;
@ ij

(3.59)

and 2we = inv + vis given by (3.43) is

we =

@
@
(u1Zp + qe x ) + (v1 Zp + qe y ):
@x
@y

(3.60)

Using the de nition of Lwe (3.59) and centrally discretising the gradients of qe x , qe y ,

3.4 Discretisation viscous potential- ow problem

43

u1Zp and v1 Zp, the equations for ue and ve are found to be of the form
ij

ue

ij

+1 MX
s +1
X

Mp

=1 l=1
M +1 M +1
X X

ij

u
qe Auijkl x + Bijkl
y + u0 ;
kl

kl

(3.61)

eij

kl

ve

ij

=1 l=1

v
qe Avijkl x + Bijkl
y + v0 :
kl

kl

(3.62)

eij

kl

The double summation part contains the disturbance due to the boundary layer, and
as before, u0 and v0 are the undisturbed inviscid velocity components, containing the
freestream and the discretisation of the geometry.
e

3.4.4 Discretisation wing/aerofoil problem


The Dirichlet approach of the potential- ow problem is to be applied for the wing/aerofoil
problem. With the inclusion of the boundary-layer displacement e ect in the Dirichlet
formulation of the potential- ow problem on the wing's surface, equation (3.35) becomes
1
4

ZZ

@ 1
1

dS =
@n r
4
S +S
w

ZZ
Sb

inv
1
dS +
r
4

ZZ

vis
dS;
r
S +S
b

(3.63)

with the inviscid source strength inv and the viscous source strength vis as de ned in
(3.45) on the body and (3.47) in the wake.
In order to calculate the above integrals the wing surface and wake are discretised.
Let the number of segments along the contour of the body surface be denoted by Mb , and
spanwise along the Y -axis by Ms , resulting in Nb = Mb  Ms panels on the body surface.
The wake centre line is divided into Mw segments, leading to Nw = Mw  Ms panels in
the wake.
A low-order panel method is used, assuming that the singularity strengths are constant
and equal to the value at the centroid of each panel. Each panel will have a constant
doublet strength n and a constant source strength n , for n 2 [1; Nb + Nw ]. The inviscid
source strengths are determined from the freestream and the viscous source strengths are
known from the viscous quantities. The discretisation of equation (3.63) then leads to a
linear system of equations to be solved for the unknowns n , n 2 [1; Nb + Nw ], de ned in
the panel centres
ZZ
N
N +N
N +N
X
X 1 Z Z vis
X 1 ZZ
@ 1
1
inv
n
dS =
dS +
dS:
4

@n
r
4

r
4

r
S
S
S
n=1
n=1
n=1
b

A more simpli ed way of writing the above system for body collocation points m 2 [1; Nb ],
located in the midpoint of the panel, is
N +N
N
N +N
X
X
X
Bmn vis ;
(3.64)
Bmn inv +
Amn n =
n=1
n=1
n=1
b

44

Chapter 3 External ow region

and a similar system is found for the two-dimensional ow problem. For the construction
of (3.64) the source and doublet strengths, assumed constant per panel, have been taken
outside the double integral. In uence matrices Amn and Bmn , describing the in uence at
collocation point m due to a constant doublet strength distribution and source strength
distribution, respectively, placed at panel n, are given by
Three-dimensional case

Amn

Two-dimensional case

ZZ

1
@ 1
=
dS;
4 S @n rmn
ZZ
1
1
=
dS;
4 S rmn

Amn

Bmn

1
@
=
(ln rmn ) dS;
2 S @n
Z
1
=
ln rmn dS:
2 S

(3.65)

Bmn

(3.66)

with rmn the distance between collocation point m and the point of integration n. The
matrices Amn and Bmn are de ned in a local panel-orientated coordinate system, and the
evaluation of the above integrals can be found in the paper of Hess and Smith [44].
The above system (3.64) consists of Nb equations, however, there are Nb + Nw unknown
doublet strengths to be determined. For the problem to be unique the implicit Kutta
condition is used, to prescribe the unknown doublet strength in the wake. Providing
there is no signi cant ow along the trailing edge, the doublet strength along a wake strip
is taken to be constant, and its value, w , is related to the upper and lower trailing edge
doublets as

w = b

b ;

(3.67)

where b is the doublet strength of the upper body trailing edge panel and b is the
doublet strength of the lower body trailing edge panel. With the above implicit Kutta
condition, the in uence matrix Amn can be reduced to a Nb  Nb matrix. The Nb unknown
doublet strengths on the body surface are to be calculated from the resulting system of
Nb equations.
u

Edge velocity equations body


For the interaction with the viscous layer, a relation between the velocity from the external
ow at the edge of the boundary layer, and the boundary-layer displacement thicknesses,
is required. For the Dirichlet method this means that equation (3.64) has to be further
manipulated.
The total edge velocity vector Qtot , located at the endpoint of a panel, instead of at
the centroid, can be decomposed as given in the paper by Kerwin et al. [51] and shown
in gure 3.4
e

Qtot = Qs + Q1 +  n
= Qs + Q1 (Q1  n)n + w1 n;
e

(3.68)

and ful ls the boundary condition Qtot  n = w1 as given in (3.42). The term Qs = rs 
corresponds to the component of perturbation velocity tangential to the body surface. If
e

3.4 Discretisation viscous potential- ow problem

-( n . Q - w ) n

45

Q tot

n
s
1111111111111111111111111111111
0000000000000000000000000000000
0000000000000000000000000000000
1111111111111111111111111111111
0000000000000000000000000000000
1111111111111111111111111111111
0000000000000000000000000000000
1111111111111111111111111111111

Figure 3.4: Composition velocity vectors.

t1 and t2 are de ned as two di erent unit vectors tangential to the surface, the following
holds:
@
@
= t1  rs ;
= t2  rs :
@t1
@t2
When using the above, the tangential velocity Qs can be expressed
@
@
[ t1 (t1  t2 ) t2 ] +
[ t2 (t1  t2 ) t1 ]
@t
@t
1
2
Qs =
:
(3.69)
k t1  t2 k2
The two perturbation velocity components in the t1 and t2 directions are determined by
the two components of the doublet gradient in the two corresponding tangential directions
@
@
@
@
=
;
=
;
@t1
@t1
@t2
@t2
which can be discretised using central di erencing.
After some manipulations, combining (3.64) with the Kutta condition (3.67) and
(3.69), the equation for Qs in the panel endpoints indicated by (i; j ), with i 2 [2; Mb ]
and j 2 [2; Ms ], can be written
M M
M +M M
X
X
X X
1
Qs =
(CA B )ijkl inv +
(CA 1B )ijkl vis ;
(3.70)
k =1 l=1
k =1
l=1
where indices k 2 [1; Mb ] and l 2 [1; Ms ] de ne the panel number. Furthermore, matrix
(CA 1B ) is constructed from three matrices A 1 , B and C . Matrix A 1 is the inverse of
A which represents the doublet in uence and matrix B represents the source in uence,
both given in (3.65) and (3.66). Matrix C contains the discretisation of expression (3.69).
It follows that equation (3.68) for the total edge velocity Qtot , with Qs given by (3.70),
can be rewritten as:
M +M M
X X
Qtot = vis n +
Dijkl vis + Q0 ;
(3.71)
k =1
l=1
e

eij

kl

kl

eij

ij

kl

eij

46

Chapter 3 External ow region

with matrix D being (CA 1B ). In the above expression, Q0 is the undisturbed inviscid
velocity, due to the freestream and the inviscid sources.
Equation (3.71) can be further simpli ed, using (3.45) and (3.47) for a two-layered
wake and centrally discretising the gradients of qe x and qe y , giving nally for the global
edge velocity components Qtot = (Ue ; Ve; We )T in the body collocation points i 2 [2; Mb ]
and j 2 [2; Ms ]
M +2M +1 M +1
X
X
U
Ue =
qe [AUijkl x + Bijkl
y ] + U0 ;
(3.72)
k =1
l=1
M +2M +1 M +1
X
X
V
Ve =
qe [AVijkl x + Bijkl
y ] + V0 ;
(3.73)
k =1
l=1
M +2M +1 M +1
X
X

W

We =
qe [AW
(3.74)
ijkl x + Bijkl y ] + W0 :
k =1
l=1
It is noted that the rst wake node point corresponds with the upper and lower trailing
edge points, de ned (Mw + 1; l) and (Mb + Mw + 1; l), and is therefore omitted. Indices
(Mw ; l) and (Mb + Mw + 2; l) thus represent the upper and lower `second' wake point just
after the trailing edge for spanwise station l.
In two dimensions a similar manipulation can be followed and relations for the velocity
components U and W in the Cartesian coordinate system can be found. However, unlike the three-dimensional ow problem where the velocities in the Cartesian coordinate
system are used, the two-dimensional ow problem will make use of a streamline
p 2coordi-2
nate system (s; n). The tangential velocity in the streamline direction Us = U + W
becomes
Us = Qtot  sji
@ @ 1
+
;
(3.75)
=
@s
@s
having used (3.68) with Qs  s = @=@s = @=@s,  n  s = 0 and Q1  s = @ 1 =@s. By
manipulating (3.64) and using the Kutta condition (3.67) it is then found that
M
M +M
X
X
1
Us =
(CA B )ik inv +
(CA 1B )ik vis + Us1 ;
(3.76)
k =1
k =1
with matrix C containing the discretisation of (3.75). Modelling the boundary layer in
the wake as just one layer with (3.48) and using the integral thickness de nitions in a
streamline coordinate system, as given in appendix A, Us becomes
M +M +1
X
Us =
AUik Us s + U0 :
(3.77)
k =1
From equations (3.72) and (3.74) it follows that:
AUik = cos k (cos i AUik + sin i AW
(3.78)
ik );
where k is the angle panel k makes with the X -axis.
e

ij

kl

kl

eij

kl

kl

kl

eij

kl

kl

kl

ij

kl

ij

ei

eij

ei

ei

ei

ek

se
i

3.5 Aerodynamic coecients

47

Edge velocity equations wake


In the wake region, the doublet strengths are known from the Dirichlet method on the
body surface and the implicit Kutta condition. The source strengths are given by the
freestream and viscous quantities.
For the discretisation of the potential- ow method in the wake, the potentials are to
be determined in the midpoints of the panels. By taking the gradient, the edge velocity
equations are calculated back in the node points of the panels. The total potential at the
collocation points m 2 [1; Nw ], located in the midpoints of the wake panels, is given by
+
X

m = 1

Nb Nw
m

1
4
n=1

ZZ

N
X
@ 1
1
n
dS +
@n r
4
n=1
b

Sn

ZZ

N +N
X 1
inv
dS +
r
4
n=1
b

Sn

ZZ

vis
dS:
r
S
n

A more simpli ed way of writing the above system is


N +N
N
N +N
X
X
X
m = 1
Amn n + Bmn inv +
Bmn vis
n=1
n=1
n=1
= 1 + m ;
b

(3.79)

with Amn and Bmn as de ned before in (3.65) and (3.66) and  the perturbation potential.
The doublet strengths  in (3.79) are known from the Dirichlet method on the body and
can be replaced with (3.64), which results in an equation for  only dependent on the
sources strengths inv and vis .
The velocity distribution Qtot can then be determined in the panel endpoints with
index (i; j ), i 2 [1; Mw 1] and j 2 [2; Ms ], by di erencing, using (3.69). Note that
i = 1 represents the `second' wake point after the trailing edge. Relations for the velocity
components in the wake are thus found of a similar format as equations (3.72), (3.73) and
(3.74).
e

3.5 Aerodynamic coecients


In a frame of reference with velocity Q1 , Bernoulli's equation for incompressible and
inviscid steady ow is
p1 + 21 Q21 = p + 21 Q2tot ;
(3.80)
where p is the pressure and  = 1 the density. From (3.80) the pressure coecient follows:
p p
Q2tot
Cp = 1 21 = 1
(3.81)
21 :
Q
Q
1
2
In two dimensions the lift and the drag coecient are de ned by

L
CL = 1 2 ;
2 Q1 c

D
CD = 1 2 ;
2 Q1 c

(3.82)

48

Chapter 3 External ow region

for which the lift L, is the force acting in the direction perpendicular to the freestream
direction and the drag D, the force acting in the direction of the freestream.
Assuming the pressure coecient Cp and the skin-friction coecient Cf to be constant
over each panel, the forces in the X - and Z - direction per wing section are given by

Cf
li cos i );
2
i
X
C
Fz =
( Cp li cos i + sign f li sin i );
2
i
X

Fx =

(Cp li sin i + sign

(3.83)
(3.84)

where i the angle between panel i and the X -axis and li2 = X 2 + Z 2 , the length of
panel i. The term sign is equivalent to +1 or 1 for panel i on the upper or lower surface,
respectively. Then for an incidence the lift and drag are

L = Fz cos Fx sin ;
D = Fx cos + Fz sin :

(3.85)
(3.86)

Alternatively, the lift coecient per wing section can be determined with the KuttaJoukowski's formula [50]. This formula states that the resultant aerodynamic force in
an incompressible, inviscid, steady and irrotational ow is directly proportional to the
circulation and acts normal to the freestream, hence

CL

KJ

= 1 2 :
2 Q1 c

(3.87)

The total circulation is determined by


= w

w = te
l

te ;
l

having used the implicit Kutta condition (3.33).


The integration formula for the drag coecient (3.86) can give quite incorrect results.
A more suitable prediction is given with Young's formula [109]

CD

2ss1
c

= ss Us =Q1 (H +5)=2 ;
=

te

te

ete

(3.88)

which is derived from the Von Karman equation, with ss1 the streamwise momentum
thickness far downstream in the wake.

3.6 Veri cation two-dimensional Dirichlet method


To verify the two-dimensional Dirichlet method for the calculation of inviscid ow, the
described panel method has been tested for the calculation of inviscid ow around an

3.6 Veri cation two-dimensional Dirichlet method

49

ellipse for which the analytical solution is known. The geometry of the ellipse is de ned
by
(X 12 )2 Y 2
+ 2 = 1;
(3.89)

( 12 )2
where 2 is the maximum thickness of the ellipse. The maximum width of the ellipse
corresponds with 1. The exact solution for the ow past an ellipse can be determined
with complex-function theory and the velocity components in the X - and Y -direction on
the ellipse are


(2 8(x 0:5)2 ) cos + (4y 4yx)= sin
0:5 
Uellipse =
+1
;
(3.90)
5  0:5  16(x 0:5)2 + 2
0:5 + 
1 + 00::5+
 0:5+

Vellipse =

0:5 
0:5 + 

(4x

8(x 0:5)2 ) sin + (4y (x 0:5))= cos


;
1 + 00::55+ 00::55+ 16(x 0:5)2 + 2

0.06

4
ELLIPS

EXACT
DIRICHLET
EXACT
DIRICHLET
EXACT
DIRICHLET

3.5

0.04

3
U_s_ellips

0.02
Y

(3.91)

0
-0.02

2.5

alpha
alpha
alpha
alpha
alpha
alpha

=
=
=
=
=
=

0
0
10
10
45
45

2
1.5
1

-0.04

0.5

-0.06

0
0

0.2

0.4

0.6

0.8

0.2

0.4

Figure 3.5: Geometry ellipse for  = 0:06.

0.8

Figure 3.6: Velocity distribution for N = 15.

3
2.5

alpha
alpha
alpha
alpha
alpha
alpha

=
=
=
=
=
=

0
0
10
10
45
45

EXACT
DIRICHLET
EXACT
DIRICHLET
EXACT
DIRICHLET

3.5
3
U_s_ellips

EXACT
DIRICHLET
EXACT
DIRICHLET
EXACT
DIRICHLET

3.5

U_s_ellips

0.6
X

2
1.5

2.5

=
=
=
=
=
=

0
0
10
10
45
45

2
1.5

0.5

0.5

alpha
alpha
alpha
alpha
alpha
alpha

0
0

0.2

0.4

0.6

0.8

Figure 3.7: Velocity distribution for N = 35.

0.2

0.4

0.6

0.8

Figure 3.8: Velocity distribution for N = 61.

50

Chapter 3 External ow region

with being the angle of attack. The total magnitude of the velocity corresponds with
the streamwise velocity at the surface of the ellipse and is given by

Us

ellipse

2
2 :
Uellipse
+ Vellipse

(3.92)

For the present calculations  is set to 0:06 for which the geometry of the ellipse is shown
in gure 3.5. The trailing edge point corresponds with x = 1. The computations are
performed for three di erent incidences, being = 0o , 10o and 45o .
In gures 3.6, 3.7 and 3.8 the velocity distribution obtained with the panel method
for the three incidences is shown, compared to the exact results obtained with (3.92). For
the results in gure 3.6, 15 points are use for the discretisation of the body surface, and
it is seen that the results from the Dirichlet method are not very accurate. Increasing the
number of points to 35 gives much better correspondence with the exact solution, as is
seen in gure 3.7. However, near the trailing edge the velocity distribution determined
by the panel method for the upper surface is seen to cross the velocity distribution for
the lower surface, resembling a sh-tail. The analytical Kutta condition, implying equal
upper and lower velocities at the trailing edge, is not satis ed.
The results obtained using 61 body points, shown in gure 3.8, are nearly similar to
the results for 35 body points in gure 3.7. The accuracy has not really improved and
the sh-tail behaviour near the trailing edge for larger angles of attack remains. For the
boundary-layer coupling this can cause problems and a modi cation of the trailing edge
values is required and will be discussed in the next section.
The three-dimensional Dirichlet method has been programmed by George Patrianakos,
a fellow Ph.D. student at the University of Bristol, and the validation for the threedimensional Dirichlet method will be presented in his thesis.

3.7 Adjustment trailing edge velocity distribution


As is clear from the previous section, the inviscid ow solution obtained at the trailing
edge with the two-dimensional Dirichlet panel method is not very accurate. The threedimensional wing Dirichlet panel method, making use of the same Kutta condition as the
two-dimensional method, displays the same inaccuracies.
With the implemented implicit Kutta condition (3.67) the actual Kutta condition
is not explicitly ful lled as (3.67) apparently allows for the trailing edge velocities to
be di erent from each other on the upper and lower surfaces, whereas analytically they
should be exactly the same. Already a small di erence can have a large in uence on the
overall solution, especially when coupled to a boundary layer. An even larger problem
arises when the inaccurate solution at the trailing induces a steep pressure gradient, which
can cause the boundary-layer calculations to breakdown.
For the work described in this thesis the external ow solver is treated as a blackbox.
To avoid the breakdown of the boundary-layer calculations, and to possibly improve the
overall solution, therefore only some ad hoc modi cations have been implemented. To
come to a more suitable trailing edge treatment the problem is to be examined more
closely.

3.7 Adjustment trailing edge velocity distribution

51

The velocities at the trailing edge coming from the wing and aerofoil Dirichlet panel
method are modi ed before being used for the boundary-layer calculations. Various possibilities have been tested and in the two-dimensional aerofoil code the following simple
procedure is used:

Ue

te

= min(Ue ; Ue ):
teu

(3.93)

tel

The new value at the trailing edge is set to the minimum of the upper and lower trailing
edge velocities.
For a wing section at angle of attack the most critical area is near the trailing edge
on the upper surface, where separation can occur for severe positive pressure (negative
velocity) gradients. Adjustment by using the maximum value of Ue at the trailing
edge, instead of the minimum value as is done in (3.93), would decrease the velocity
gradient near the trailing edge of the upper surface and one would expect this to improve
the boundary-layer calculations. However, using the minimum has shown to work much
better. An explanation for this could be that the pressure gradient just after the trailing
edge in the wake is more problematic than the pressure gradient at the trailing edge on
the upper surface. Taking the maximum value of Ue at the trailing edge could in this
case increase the (positive) pressure gradient just after the trailing edge in the wake,
making the calculations there more dicult, whereas taking the minimum would make
the calculations easier.
For the three-dimensional wing external velocity distribution a slightly di erent modi cation procedure is used. The solution for the three velocity components in the trailing
edge point on the upper and lower surface is determined by averaging the calculated upper
and lower trailing edge velocities.
In the wake the solution for the Z -velocity component in the rst point downstream of
the trailing edge is also redetermined. This is done via interpolation with the new trailing
edge velocity and the downstream wake values.
Other adjustment variants have been tested as well, however, the above ones have
shown to be the most robust and accurate and have been applied.
te

te

52

Chapter 3 External ow region

Chapter 4
Viscous-inviscid interaction with the
quasi-simultaneous method
4.1 Introduction
Viscous-inviscid interaction techniques divide the ow regime in an external ow region,
where the inviscid ow equations hold, and a thin shear layer region, where the boundarylayer equations are valid. By combining the solution in the two regions with a suitable
coupling scheme, the solution for the whole ow is to be found.
Over the years various coupling schemes have been developed, with the aim to arrive
at a robust and fast method, capable of calculating ows with limited amounts of ow
separation. A lot of work has been done for steady two-dimensional ow problems and
for these cases the situation at present is quite satisfactory.
The most robust interaction algorithm, capable of handling the separated ow region,
is the quasi-simultaneous method. In the present chapter the quasi-simultaneous method
is introduced and analysed in detail to determine its basic requirements.

4.2 Interaction algorithms


The aim of a viscous-inviscid interaction algorithm is to combine iteratively the separate
boundary-layer and external ow calculations to provide the overall ow solution.
Simpli ed, if the operators E , for the external inviscid ow equations, and V , for the
viscous boundary-layer equations, are introduced, the formulation in both regions in two
dimensions can be written as
(

ue =

E [x ];

ue =

V [x ];

(4.1)

which describes a non-linear system in ue, the velocity at the edge of the boundary layer,
and x , the boundary-layer displacement thickness.

54

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

Direct method
Following the classical mathematical theory of matched asymptotic expansions, the solution is calculated in a sequential way, based on a direct hierarchy between the inner
boundary layer and the outer inviscid ow. First, the outer ow velocity is calculated,
which is then used in the boundary-layer equations to determine the new boundary-layer
displacement thickness
8
< u(en) =
E [x( 1) ];
(4.2)
: ( )
(
n)
1
x = V [ue ];
n

which is called the direct method. Superscript n indicates the nth viscous-inviscid iteration.
It is interesting to analyse the system of equations when solved with the direct method.
In two dimensions, the steady Von Karman equation and the entrainment equation can
be combined to give

d
due b dx
+
= ;
(4.3)
dx a dx
a
for which the derivation and the de nitions of the coecients a; b and d have been given in
section 2.4. It is noted that coecient b corresponds with dH1 =dH , with function H1 (H )
having a minimum due to which the calculation of H can become impossible when H1 is
given [47, 56]. Using the closure relation for H -H1 as given in appendix B, b equals zero
at H = Hsep, where Hsep = 4:029 for laminar ow and Hsep = 2:73 for turbulent ow.
The skin-friction becomes zero at H = Hsep and corresponds for two-dimensional steady
ow calculations with the separation point. Furthermore, it can be shown that
b
a

8
<
:

> 0
= 0
< 0

for H < Hsep (attached ow),


for H = Hsep (separation point),
for H > Hsep (separated ow),

d
> 0:
a

(4.4)

When treating equation (4.3) with the classical direct method, the velocity ue is known
from the external ow calculation and a simple ordinary di erential equation for x remains to be solved
d due
b dx
=
:
(4.5)
a dx
a dx
At H = Hsep the coecient b=a is zero and dx =dx becomes in nite, unless the right-hand
side of (4.5) cancels against the left

d due
:
(4.6)
a dx
For ue being a converged solution (4.6) is valid. For an arbitrary prescribed ue distribution
the above relation is usually not satis ed. Furthermore, even if (4.6) is true at H = Hsep,
in (4.5) the derivative dx =dx remains undetermined, being of the form 0=0.
0 =

4.2 Interaction algorithms

55

Similarly, from the Von Karman equation (2.1), it follows that at H = Hsep the gradient dxx =dx remains nite and the gradient dx =dx therefore becomes in nite. The solution for x is irrealistic and violates the boundary-layer hypothesis where the boundarylayer thickness is assumed to be small compared to the length scale of the geometry. We
speak of a Goldstein singularity, as Goldstein in 1948 rst wrote a detailed paper about
the above mentioned problem [37].
It is noted that the direct scheme converges to a solution for attached ow up to the
point where H = Hsep, as it then correctly models the global behaviour of the ow.

Inverse method
A technique to remove the non-physical singularity, is the so-called inverse method [11, 8].
The direction of the iterative process between the boundary layer and the external ow
is inverted and the boundary-layer equations are solved with a prescribed displacement
thickness
8 ( )
< x
= E 1 [u(en 1) ];
(4.7)
( )
: (n)

ue = V [x ]:
n

Analysing boundary-layer equation (4.3) with the inverse method, with the displacement thickness prescribed, leads to

due
d b dx
=
:
(4.8)
dx
a a dx
Term d=a remains positive when H = Hsep, which implies that with the inverse scheme
the calculations can be continued beyond the point of separation.
The above discussed direct and inverse schemes are examples of weak interaction between the boundary-layer and external ow regions, as they are based on a hierarchical
iterative treatment between the viscous and inviscid ow regions.
Strong interaction
Following the mathematical theory of matched asymptotic expansions for laminar ow,
developed independently by Stewartson [88, 87], Neiland [74] and Messiter [68], it was seen
that near a singularity the classical two-layered asymptotic structure is no longer correct.
Instead a so-called triple-deck structure should be applied, in which the boundary layer
is subdivided into three regions as follows [97]:
1. a viscous ow sublayer of thickness O(Re 5=8 L) which is governed by the classical
boundary-layer equations;
2. an inviscid ow middle layer of thickness O(Re 1=2 L) which is a continuation of the
oncoming boundary layer;
3. inviscid irrotational ow top layer of thickness O(Re 3=8 L) which can be described
by thin-aerofoil theory.

56

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

From the triple-deck theory it is clear that near a singularity the viscous and inviscid ow
regions have become equally important. A numerical viscous-inviscid interaction method
therefore should locally treat both regions together to model the strong interaction. Lagerstrom described the triple-deck idea in 1975 with the following words [54]: \An important
feature is that the pressure is self-induced, that is, the pressure due to displacement thickness is determined simultaneously with the revised boundary-layer solution. [ : : : ] this
solution exhibits a de nite loss of hierarchy."
The rst numerical solutions for the triple-deck equations were obtained by three independent investigations: by Jobe and Burggraf [49], Melnik and Chow [65] and Veldman
and Van de Vooren [100].
In turbulent ow a di erent asymptotic ow eld structure exists compared to laminar
ow. However, it is assumed that conclusions from the triple-deck theory for laminar ow
with regard to hierarchy and modelling are also valid for turbulent ow. Mathematical
evidence for this is given in the papers by Melnik [64, 66, 67] and Sychev and Sychev [91].

Simultaneous method
The so-called simultaneous method [25, 27], re ecting the lack of hierarchy by treating
the boundary-layer equations and the external inviscid ow equation together, is written
as
8
( )
< u(en) E [x ] = 0;
(4.9)
( )
: (n)

ue + V [x ] = 0:
n

To analyse the above system of equations for the two-dimensional steady ow case, it
is assumed that the external ow is modelled by a simple linear relation between ue and
x

ue

1 x = 2 ;

(4.10)

where the coecients 1 > 0 and 2 are two constants. Combining (4.3) and (4.10)
following the simultaneous scheme leads to the following ordinary di erential equation to
be solved for x :


b
1 +
a

dx
d
= :
dx
a

It is seen that when b=a = 0 at H = Hsep, the solution for x is feasible and no singularity
is encountered.
When the ow becomes too strongly separated, however, the term 1 + b=a reaches
zero, see (4.4), leading to the breakdown of the system. As the boundary layer is no longer
thin, the mathematical model fails to represent the physics, which is characterised by the
non-existence/non-uniqueness of a steady ow solution. Another mathematical model is
to be used instead. With the unsteady boundary-layer formulation the calculations can
be continued, until it looses its validity for unsteady detached ow [43]. To describe the
unsteady detached continuation of the ow (i.e. marginal separation), a new asymptotic

4.3 Quasi-simultaneous method

57

theory is to be applied [89]. Alternatively, the problem is to be solved with the NavierStokes equations, giving the complete mathematical description of the ow eld.
The simultaneous method can be quite computationally expensive, requiring the full
external ow equations to be solved with the boundary-layer equations at each iteration.
Furthermore, the simultaneous method can be dicult to implement when employed
to combine an external ow solver with the boundary-layer equations in their original
di erential form. These problems are avoided by using adaptations of the direct and
inverse schemes, which take into account the character of triple-deck theory.

Semi-inverse method
A scheme based on the inverse method is the semi-inverse method [57, 9], which combines
the viscous and inviscid ow regions via a special matching condition
8
(n) = E [ ( 1) ];
>
> ue
>
>
<

u(en) =

>
>
>
>
: (n)

V [x(

= x(

1)

1)

(4.11)

];

+ !si (u(en)
V

u(en) ):
E

where ue represents the edge velocity obtained from the external ow equations and ue
represents the edge velocity obtained from the viscous boundary-layer equations.
In (4.11) it is seen that the boundary-layer region is calculated in an inverse way,
avoiding breakdown at separation, whereas the external ow is computed in a direct way.
With a suitable update procedure, in which !si is the relaxation parameter of the semiinverse scheme, a new displacement thickness is determined from the calculated viscous
and external velocities. A converged solution is obtained when ue = ue .
E

4.3 Quasi-simultaneous method


In order to more economically solve the coupled system of equations than the fully simultaneous method and following the triple-deck theory, the quasi-simultaneous method was
developed, based on the direct technique [98].
Instead of solving the system of equations (4.1) fully simultaneously, the boundarylayer calculations are performed together with an approximation for the external inviscid
ow. The external outer ow calculations are performed with a prescribed boundary-layer
displacement thickness, as was done in the direct method
8 8 (n)
> < u
I [ ( ) ] = u(n 1) I [ ( 1) ];
>
>
>
<
>
>
>
>
:

eV

eE

u(en) + V [x( ) ] = 0;

(4.12)

u(en) = E [x( ) ]:
n

The approximation derived from the inviscid ow is formulated in terms of an interaction


law I and describes the essential part of the interaction with the boundary layer, in order

58

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

to model regions with strong interaction. It is seen that the interaction law ue  I [x ], can
also be interpreted as solving the boundary-layer equations with a natural generalisation
of ue prescribed.
The interaction law equation in (4.12) is employed in a de cit formula, not in uencing
the nal solution. When a converged solution is obtained, the I terms in (4.12) disappear
and the edge velocity obtained from the boundary layer and the edge velocity obtained
from the external ow are matched, i.e. ue = ue . As the interaction law determines
the convergence speed of the method a good choice of I is very important, which will be
discussed in detail in the following sections.
It is noted that the analysis of the system of equations presented for the fully simultaneous method in the previous section corresponds with the analysis for the quasisimultaneous method, assuming that equation (4.10) corresponds with the interaction
law equation. Solving the boundary-layer equation together with an appropriate interaction law removes the singularity, whereas without the interaction law the above scheme
corresponds with the direct method, which would breakdown at H = Hsep.
For two-dimensional ow calculations it has been shown that the quasi-simultaneous
method, using a suitable interaction law, compared to the semi-inverse method, has a
much higher convergence rate and is much more robust [61].
V

4.4 Analysis of interaction laws


From section 4.3 it is clear that with the use of a suitable representation of the external
inviscid ow, to be solved together with the boundary-layer equations, the singularities
occurring when dH1 =dH passes through zero, are avoided. A very simple interaction law
was used in (4.10). Here, the choice of the interaction law is investigated in detail and
the basic requirements are discussed.
For the model problem of two-dimensional viscous potential ow over a plate, as
discussed in chapter 3, the ow eld at the edge of the boundary layer is described by
thin-aerofoil theory, consistent with property 3 of the triple-deck
1
ue(x) =


xend
Z

d
d
(uex )
+ u0 (x);
d
x 
e

xbegin

(4.13)

with u0 the undisturbed velocity and duex =d the transpiration velocity. For clarity
in the following analysis, the transpiration velocity is simpli ed to dx =d and (4.13)
becomes1
e

1
ue(x) =


xend
Z

dx d
+ u0 (x):
d x 
e

xbegin

(4.14)

retrospect, it would have been more suitable to use the mass defect m  ue x as a variable in the
present analysis.
1 In

4.5 Viscous-inviscid iterations

59

After the discretisation of the plate into Mp segments the above integral (4.14) can be
calculated, as was done in section 3.4.2, and rewritten to the form
M +1
X
ue =
Eik x + u0 ;
(4.15)
k =1
which is similar to equation (3.54) with the external ow in uence matrix Au renamed as
E . It has been established in (3.55) that matrix E has the following properties:
p

ei

Eii > 0;
Eik = Eki  0; (i 6= k),
(4.16)
M +1
X
Eik  0; with at least for one i the inequality sign.
k =1
From appendix D it follows that matrix E is a symmetric weakly diagonally dominant
M-matrix (de nition D.3 and theorem D.10) and hence positive de nite (theorem D.12).
Before starting with the construction of an interaction law based on E , the various
iterative schemes of the quasi-simultaneous method are investigated to determine which
requirements the interaction law has to satisfy.
p

4.5 Viscous-inviscid iterations


In order to keep the following analysis clear, boundary-layer equation (4.3) is simpli ed
to a linear relation between ue and x . The homogeneous system consisting of equation
(4.15) and the simpli ed equation (4.3) is written
V

ue
ue

=
=

E x ;
V x ;

(4.17)
(4.18)

where E is the external ow in uence matrix and V the matrix representing the boundarylayer ow. Moreover, V is a lower triangular matrix, due to the discretisation of the xderivatives (section 2.7.1), with coecient b=a on its diagonal. The solution of the above
system has to satisfy the requirement that the edge velocity obtained from the viscous
boundary-layer equation is matched with the edge velocity obtained from the external
ow equation, i.e. ue = ue . Therefore, using ue = ue , the system de ned by (4.17)
and (4.18) can be rewritten as
V

(E + V ) x = 0:

(4.19)

As our main interest is in stable steady ow solutions of the ow equations, it may be


assumed that E + V is a positive stable matrix (de nition D.1). However, in order to
enable some theory occasionally the following stronger assumption is made.

Assumption 4.1 It is assumed that the positive stable matrix E + V possesses nonpositive o -diagonal entries, i.e. E + V is an M-matrix.

60

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

From (4.16) it is clear that E is a diagonally dominant matrix with non-positive o diagonal entries. The diagonal entries of E have elements  1=x (section 3.4.2), whereas
matrix V has entries O(1) with a diagonal that is positive and/or slightly negative. For a
small stepsize x, matrix E + V hence will also be diagonally dominant with non-positive
o -diagonal entries and assumption 4.1 therefore is not unreasonable.
Introducing ue = Ix as the interaction law and applying (4.12), the quasi-simultaneous
viscous-inviscid iteration process to solve for (4.19) is given by
I

(I + V ) x

= (I

E ) x

= (I + V ) 1 (I

E ) x

(n)

(n 1)

(4.20)

which can be rewritten as

x

(n)

(n 1)

(4.21)

The iteration matrix of the system of viscous-inviscid iterations (4.21) is given by


(I + V ) 1 (I E ) = I (I + V ) 1 (E + V );
(4.22)
where matrix I denotes the identity matrix. An iterative method converges if and only if
the spectral radius  (de nition D.2) of the iteration matrix is smaller than one, and the
present scheme therefore obtains convergence when
((I + V ) 1 (I E )) < 1:
(4.23)
It is preferable to have ((I + V ) 1 (I E )) close to zero, as the smaller , the faster the
iterative method converges. When I = E , which actually corresponds with the simultaneous coupling scheme,  = 0 and only one viscous-inviscid sweep is required. However,
using I = E would require the calculation of the full inverse of E and a simpler I is
preferred. Some guidelines on how to choose I are to be presented next.

4.5.1 Basic conditions for viscous-inviscid convergence


In case (4.23) is not satis ed, the simplest way to try to obtain viscous-inviscid convergence
is by means of Jacobi-like (under)relaxation

x

(n)

= !vi (I + V ) 1 (I

E )x

(n 1)

+ (1 !vi )x

(n 1)

(4.24)

Theorem 4.2 There exists a (suciently small) viscous-inviscid relaxation parameter


!vi > 0 such that Jacobi-like relaxation of the viscous-inviscid iterations as given in (4.24)
converges, if and only if matrix (I + V ) 1 (E + V ) is positive stable.
Proof: Denote the eigenvalues of the iteration matrix (4.22) by vi and the eigenvalues
of the Jacobi-like iteration matrix, implicitly de ned in (4.24), by J . The eigenvalues of
(I + V ) 1 (E + V ) are denoted by . It is then found that

J = 1 + !vi (vi

1):

(4.25)

4.5 Viscous-inviscid iterations

61

It easily follows that one can achieve jJ j < 1 for (suciently small) positive !vi , if
and only if all eigenvalues vi satisfy <(vi ) < 1. Rewriting the latter condition, using
vi = 1  from (4.22), gives <() > 0, i.e. convergence can be achieved if and only if
(I + V ) 1 (E + V ) is positive stable.

The next step is to determine the conditions for which matrix (I + V ) 1 (E + V ) is
positive stable. With E + V positive stable, it is natural to demand that I + V (and hence
(I + V ) 1 ) is positive stable too. However, the product of positive stable matrices does not
have to be positive stable. In order to give theoretical proof to show that (I + V ) 1 (E + V )
is positive stable (slightly) stronger conditions are needed.
One such occasion is where I + V is positive de nite and E + V is positive real, i.e.
its symmetric part is positive de nite (theorem D.13). Another occasion is where E + V
and I + V are M-matrices with I + V  E + V (theorem D.6). It can even be shown that
in the latter case relaxation is not necessary.

Theorem 4.3 Let both E + V and I + V be M-matrices and let I  E , then the viscousinviscid iterations (4.21) converge.
Proof: Since M  I + V is an M-matrix, hence (I + V ) 1  0, and since N  I E  0,
the splitting E + V = (I + V ) (I E ) is found to be regular (de nition D.16). Since
also E + V is an M-matrix, hence (E + V ) 1  0, theorem D.17 provides convergence. 

The above theorem gives mathematical inspiration on how to construct appropriate


interaction laws. The physical idea for constructing an interaction law is that it should
contain the essential part of the interaction with the external ow, i.e. it should contain
the local e ect of the interaction. It then naturally follows to de ne interaction laws in
which some outer o -diagonals of E are omitted. With E having non-positive o -diagonal
entries this will result in I  E , as in theorem 4.3. Hence mathematical and physical
ideas come together and when E + V and I + V are M-matrices (compare assumption
4.1) theory can be developed (this theory will also cover the situation when the diagonal
of E is increased). Thus one is led to study the behaviour of interaction laws constructed
as in the following proposal.

Construction 4.4 Let the interaction law I be constructed from E by increasing the
value of its diagonal, and/or by setting certain outer o -diagonals equal to zero while
keeping the o -diagonals of I + V non-positive.

4.5.2 Convergence rate viscous-inviscid iterations


Thus far, the basic convergence of the viscous-inviscid iterative scheme has been discussed.
Next the requirements for an interaction law to improve the viscous-inviscid convergence
rate are investigated. Theorem D.18 gives a possibility to analyse and compare the convergence rates of iterative methods; here it will be applied to interaction laws as de ned
from construction 4.4.

62

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

Theorem 4.5 Let E be an external ow matrix with non-positive o -diagonal entries. Let
the boundary-layer ow matrix V be such that assumption 4.1 holds. Let the interaction
law I be derived according to construction 4.4. Then the viscous-inviscid iterative process
(4.21) is convergent. Further, the more o -diagonals of E are set to zero and/or the
more the diagonal is increased for the construction of I , the slower2 the viscous-inviscid
iterations converge.
Proof: Under assumption 4.1 E + V is an M-matrix. Furthermore, setting non-positive
o -diagonal elements of E equal to zero and/or increasing the diagonal gives I  E , hence
I + V  E + V . Moreover, by construction I + V 2 Zn , hence also I + V is an M-matrix
(theorem D.6) and theorem 4.3 guarantees convergence of the viscous-inviscid iterations.
Next let Ia and Ib be two di erent interaction laws, where Ib contains less o -diagonals
and/or a larger main diagonal, hence Ib  Ia . Then Nb  Ib E  Ia E  Na  0, and
since the splittings are regular theorem D.18 yields
0  ((Ia + V ) 1 (Ia E ))  ((Ib + V ) 1 (Ib E )) < 1;
(4.26)

which proves the theorem.

A strict inequality in (4.26) can be obtained if matrix E + V is full and diagonally


dominant, which may safely be assumed with E de ned in (4.15) as the thin-aerofoil
Hilbert matrix. Matrix E + V can be written E + V = DEV (I C ), with DEV > 0
(theorem D.5) the diagonal matrix of E + V . Consequently, matrix C has zero diagonal
elements and strictly positive o -diagonal elements, as matrix E + V is full and has
negative o -diagonal elements. As matrix E + V is diagonally dominant also I C is
diagonally dominant and it follows that (C ) < 1 (theorem D.19). Then, (E + V ) 1 =
(I C ) 1 DEV1 = (I + C + C 2 +    )DEV1 . Matrix I + C is a full matrix and as C  0
and DEV1 > 0 it follows that also the inverse of E + V is a full matrix.
Interaction law I is constructed according to assumption 4.1. Via the permutation of
rows and columns, the non-zero o -diagonals of matrix I E are transported to the last
m columns. Then (E + V ) 1 (I E ) corresponds with


full



0 G
0 F

where F is a full square m  m matrix and ((E + V ) 1 (I E )) = (F ). As F is a full


matrix, F is irreducible (theorem D.9) and for increasing F also (F ) increases (theorem
D.14). For interaction laws Ia and Ib , where Ib contains less o -diagonals and/or a larger
main diagonal, hence Ib > Ia , it follows that:
((E + V ) 1 (Ia E )) = (Fa ) < (Fb ) = ((E + V ) 1 (Ib E ));
(4.27)
where Fa and Fb the full F matrices for Ia and Ib , respectively. Next, with theorem D.17
the inequality sign in (4.26) is yielded.
2 Strictly

slow'.

speaking, according to the proof of the theorem the phrase `slower' should include `equally

4.6 Boundary-layer iterations

63

4.6 Boundary-layer iterations


Keeping the right-hand side of the viscous-inviscid iterative scheme (4.20) constant, gives
the following homogeneous boundary-layer system to be solved with a pointwise Newton
method:
(I + V ) x = 0:

(4.28)

As before, it is assumed that the viscous-inviscid system of equations to solve for, i.e.
E + V (4.19), is positive stable. Iterative inversion of E + V is possible and it would
be very unfortunate when iterative problems would occur in the inner boundary-layer
iterations applied to solve I + V (4.28). Therefore, from a robustness point of view it is
natural to demand from I that whenever E + V is positive stable, also I + V is positive
stable. This immediately leads to a requirement on the eigenvalues of I , in particular the
one with the smallest real part, as translated in the following proposition.

Proposition 4.6 Let matrices E and I be such that for all matrices V for which E + V
is positive stable also I + V is positive stable. Then
 (I )   (E );

(4.29)

where  (A) = min <((A)).




Proof: The proof proceeds by constructing a contradiction. Suppose  (I ) <  (E ) and


de ne    (E )  (I ) > 0. De ne V  diag ( 21   (E )), then
 (E + V ) =  (E ) + (V ) =  (E ) + 12   (E ) = 12  > 0;

hence matrix E + V is positive stable. But

 (I + V ) =  (I ) + (V ) =  (I ) + 12   (E ) =

1
2  < 0;

hence, for this particular choice of V , matrix I + V is not positive stable. A contradiction
has been obtained and the assertion  (I )   (E ) follows.


Remark When A 2 Zn (see de nition D.3), the eigenvalue with the minimum real part
 (A) turns out to be real and is called the minimum eigenvalue [45].
From proposition 4.6 it is now clear that for a robust construction of a suitable interaction
law, matrix I should be derived from E in such a way that  (I )   (E ).
Now, recall construction 4.4 in which the interaction law is proposed to be derived
from the external ow matrix E by setting certain o -diagonal elements equal to zero
and/or by increasing its diagonal. This obviously gives I  E , which indeed will lead to
the desired properties, as formulated in the following theorem.

Theorem 4.7 Let E be an external ow matrix with non-positive o -diagonal entries. Let
the boundary-layer ow matrix V be such that assumption 4.1 holds. Let the interaction

64

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

law I be derived from E according to construction 4.4. Then  (I + V )   (E + V ) and


 (I + V ) is a non-decreasing function of the number of outer o -diagonals set to zero and
of the magnitude of the diagonal entries, respectively.
Proof: Assumption 4.1 implies that E + V 2 Zn. Moreover, by construction I + V 2 Zn
and I  E . Now let  = max((E + V )ii ; (I + V )ii ) > 0 (theorem D.5) and de ne NEV 
I (E + V ) and NIV  I (I + V ). By construction it is seen that 0  NIV  NEV .
With theorem D.15 it follows that max (NIV ) = (NIV )  (NEV ) = max (NEV ) and

 (I + V ) =  (NIV )

  (NEV )

=  (E + V );

(4.30)

which proves the rst part of the theorem. Next, let Ia and Ib be two di erent interaction
laws, where Ib contains less o -diagonals and/or a larger main diagonal. This gives Ib  Ia ,
hence Ib + V  Ia + V . As before, it follows that:

 (Ib + V )

  (Ia + V );

(4.31)

which completes the proof.

A strict inequality in (4.30) and (4.31) can be obtained if matrices E + V and I + V


are both irreducible.

4.6.1 Convergence rate boundary-layer iterations


The system of equations (4.28) consisting of an approximate interaction law and the
boundary-layer description will be solved by repeated marching sweeps through the boundary layer in a Gauss-Seidel fashion
(Id

Il + V ) x

(~
n)

= Iu x

(~
n 1)

(4.32)

where I = Id Il Iu , with Id a diagonal matrix, Il a strictly lower triangular matrix


and Iu a strictly upper triangular matrix. Further n~ indicates the n~ th boundary-layer
iteration. Like before, it is possible to formulate a comparison theorem.

Theorem 4.8 Let E be an external ow matrix with non-positive o -diagonal entries. Let
the boundary-layer ow matrix V be such that assumption 4.1 holds. Let the interaction
law I be derived according to construction 4.4. Then the boundary-layer iterative process
(4.32) is convergent. Furthermore, the more o -diagonals of E are set to zero and/or the
more the diagonal is increased for the construction of I , the faster3 the boundary-layer
iterations converge.
Proof: Firstly, convergence of the Gauss-Seidel iterations (4.32) is guaranteed because
assumption 4.1 and construction 4.4 imply that the corresponding splitting is regular
(theorem D.17). Next, let Ia = (Id )a (Il )a (Iu )a and Ib = (Id )b (Il )b (Iu )b be
3 Strictly

fast'.

speaking, according to the proof of the theorem the phrase `faster' should include `equally

4.7 Basic convergence pointwise iterations

65

two di erent interaction laws, where Ib contains less o -diagonals and/or a larger main
diagonal. Hence, (Id )b  (Id )a , (Il )a  (Il )b and (Iu)a  (Iu )b  0. It follows that
(Id )b (Il )b + V  (Id )a (Il )a + V and as both are M-matrices, theorem D.6 yields
0  ((Id )b (Il )b + V ) 1  ((Id )a (Il )a + V ) 1 . Further, straightforward multiplication
of the non-negative matrices involved shows that the Gauss-Seidel iterations matrices for
the two interaction laws satisfy 0  ((Id )b (Il )b + V ) 1 (Iu )b  ((Id )a (Il )a + V ) 1 (Iu )a .
Finally theorem D.15 yields a similar relation for the respective spectral radii, hence the
Gauss-Seidel iterations for interaction law Ib converge at least as fast.


4.6.2 Summary
Summarising the results from section 4.5 and 4.6, it has been found that, under the
reasonable assumption 4.1, the construction of an interaction law from an external ow
matrix, either by setting certain outer o -diagonals equal to zero, and/or by increasing
the diagonal (i.e. construction 4.4), has favourable properties: the convergence of the
viscous-inviscid iterations is guaranteed, the convergence of the Gauss-Seidel boundarylayer iterations is guaranteed, and nally, the robustness of the inner iterations is enhanced.
Predictions on the rate of convergence have been made. However, it is seen that theorem 4.5 is the opposite result from theorem 4.8. For the viscous-inviscid iterative method
the convergence improves when the size of the interaction law increases, whereas for the
boundary-layer (Gauss-Seidel) iterative scheme the convergence deteriorates when the size
of the interaction law increases. The following rule can now be established to come to a
fast procedure for the overall quasi-simultaneous scheme.

Rule of thumb When the external ow calculations E are computationally intensive,


then take I  E , in order to have only a few viscous-inviscid iterations. On the other
hand, when solving E is not computationally intensive, take I with only a few non-zero
diagonals, in order to have a small number of boundary-layer iterations.

4.7 Basic convergence pointwise iterations


The non-linear system of equations described by (4.28) is to be solved with pointwise
Newton iterations. Convergence of the Newton iterations depends on the initialisation: for
an initial guess suciently close to the solution of the non-linear system Newton's method
will converge quadratically to the solution. However, in case the attraction domain for a
non-linear problem is small, intermediate Newton solutions can be quite deviated from the
nal solution. In some cases a bad initial guess can even keep the scheme from converging.
As during the Newton iterations intermediate solutions can be produced that are quite
di erent from the actual solution, it is important for the present problem to investigate
the behaviour of the boundary-layer matrix V . It is clear from (4.4) that the diagonal of
V , b=a, changes sign if separation is predicted, from positive to negative. For too negative
values of b=a the Newton iterations might stop converging.
It is recalled that the external ow matrix E (4.16) is a diagonally dominant matrix

66

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

with positive diagonal entries and non-positive o -diagonal entries. Interaction law I is
constructed as described by construction 4.4. For attached ow (b=a > 0) assumption 4.1
holds and from theorem 4.8 it follows that Newton's method will converge.
For separated ow (b=a < 0) the main diagonal entries (I + V )d = Ed + b=a decrease
whereas the o -diagonal entries increase and ((Id Il + V ) 1 Iu ) advances towards one.
If the ow is not too strongly separated (i.e. the diagonal coecients of V are not too
negative), the spectral radius remains below one and the Newton iterations will converge.
However, when diagonal elements of V become too negative assumption 4.1 is not satis ed
and the method may no longer converge.
It is clear that having strong diagonal dominance for matrix I + V helps the Newton
iterations when the ow is separated. The basic convergence of the pointwise Newton
iterations can thus be improved by increasing the magnitude of the main diagonal Id + b=a,
which can be done either via increasing Id or via increasing the diagonal coecient of
matrix V . With the use of relaxation the diagonals can be scaled, which has been applied
to the various iterative schemes to improve the basic convergence of the Newton iterations.

4.7.1 Relaxation boundary-layer iterations


Relaxation applied to the Gauss-Seidel scheme (4.32), using the values at n~ for the lower
triangular matrix (I + V )l , leads to the well-known SOR iterative method


(I + V )d
!bl

(I + V )l

(~ )
x =

(I + V )d
!bl

(I + V )l

(I + V )

x

(~
n 1)

; (4.33)

to solve for I + V , with !bl the boundary-layer relaxation parameter. If the ow is not
too strongly separated, the splitting of I + V for 0 < !bl  1 is regular and the above
SOR scheme can be made to converge (theorem D.17).
If the ow is strongly separated, the diagonal entries (I + V )d become negative and
as both Vd and Id are scaled by the relaxation parameter !bl , relaxation cannot avoid the
breakdown of the scheme.
Relaxation applied to the boundary-layer iterations, where the values at n~ are used
for (I + V )l Vd leads to a di erent scheme


Id
+ Vd
!bl

(I + V )l

(~ )
x =
n

Id
!bl

Il

I + Vu

x

(~
n 1)

(4.34)

The above method in (4.34) is more robust than the SOR method in (4.33). Method (4.34)
can keep the diagonal matrix (I + V )d diagonally dominant for very negative diagonal
coecients of V by using underrelaxation 0 < !bl  1, as only Id is scaled by the relaxation
parameter.

4.7.2 Relaxation viscous-inviscid iterations


Besides the possibility of applying relaxation to the boundary-layer iterations, relaxation
can be applied to the viscous-inviscid iterations to improve the basic convergence of the

4.7 Basic convergence pointwise iterations

67

Newton iterations. Using explicit Jacobi-like relaxation on (4.20) results in

I + V (

!vi x

n)

I +V
!vi

x

(n 1)

(4.35)

where !vi is the relaxation parameter for the viscous-inviscid iterations. It is seen that
both V and I are scaled by !vi , and as before if Vii < Iii relaxation cannot avoid the
breakdown of the Newton iterations to solve for the homogeneous boundary-layer system
(I + V )=!vi x = 0.
Relaxation should be done in such a way that it only scales matrix I , which can be
achieved by using the values at n for V , then


I
+V
!vi

(


n)

I
!vi

x

(n 1)

(4.36)

and the Newton iterations are to be solved for the homogeneous boundary-layer system
(I=!vi + V ) x = 0.
It is seen that with underrelaxation 0 < !vi  1 the full interaction law matrix is
scaled, which does not improve the overall diagonal dominance but does make the main
diagonal of I + V stronger.
Applying semi-implicit relaxation by using the values at n for Iul + V , only scales the
main diagonal of I and is therefore more suitable. The viscous-inviscid iterations become


Id
!vi

Iul + V

(


n)

Id
!vi

Iul

x

(n 1)

(4.37)

That relaxation procedure (4.37) leads to a more robust Newton scheme than (4.36) can
also be derived from theorem 4.7. For 0 < !vi  1, it is seen that Ib  Id =!vi Iul 
I=!vi  Ia , and with theorem 4.7


I
 d
!vi

Iul + V

I

+V :
!vi

(4.38)

With the diagonal viscous-inviscid relaxation procedure (4.37) the Newton iterations
need to be solved for the homogeneous boundary-layer system (Id =!vi Iul + V ) x = 0
for which the Gauss-Seidel scheme is written as


Id
+ Vd
!vi

(I + V )l x

(~
n)

= (I + V )u x

(~
n 1)

(4.39)

It is noted that the boundary-layer systems for which the Gauss-Seidel schemes (4.39)
and (4.34) are to be solved are di erent, and whereas (4.39) requires Id =!vi Iul + V
to be positive stable, (4.34) requires I + V to be positive stable, which is a stronger
requirement. It can be concluded that applying diagonal viscous-inviscid relaxation leads
to a more robust Newton iterative process than diagonal boundary-layer relaxation.

68

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

4.7.3 Relaxation due to unsteady viscous-inviscid iterations


Another form of relaxation can be obtained by using the unsteady formulation of the Von
Karman equation, given in section 2.4, leading to
b @x c @x
d
@ue
+
+
= ;
(4.40)
@x a @x a @t
a
where for analysis purposes @ue x =@t is replaced by @x =@t (see footnote p. 56). For the
discretisation of @x =@t the values of x at the previous viscous-inviscid iteration are to
be used. Using the values of the previous boundary-layer iteration would lead to a less
robust Newton scheme, similar to boundary-layer relaxation. The homogeneous linearised
boundary-layer equation, following from (4.40), is written as
b ( ) c x  ( ) ( 1) 
(
n)
ue =

x
a x
a t x
b c x ( ) c x ( 1)
+
+

:
(4.41)
=
a a t x
a t x
V

It is seen in that here V corresponds with b=a + c=a x=t, which improves the diagonal
dominance of I + V as c=a > 0. The Gauss-Seidel system has become more diagonally
dominant, which is also favourable for the Newton iterations.
It is also interesting to investigate the unsteady boundary-layer formulation used in a
direct scheme (4.2), where ue is prescribed with
V

Ex

(n 1)

Convergence is obtained when

b (

a x

c x (
(
a t x

n)

b
c
t + x
a
a

1 c

x

n)

x E t

(n 1)

):

(4.42)

< 1:

(4.43)

It is seen that when b = 0 at separation,  remains nite and a numerical singularity does
not occur. The timestep t should be chosen in such a way that jbt=a + cx=aj > 0,
and unlike in the steady case the unsteady direct scheme is now capable of calculating
within the region of separation.
If the x-derivative of x in boundary-layer equation (4.40) is discretised with an explicit
scheme, the direct method in (4.42) becomes
b ( 1) c x ( ) ( 1)
( 1)

(
x );
(4.44)
Ex
=
a x
a t x
which can be rewritten to the form
b=a + E ( 1)
( )
( 1)
x = x
t

:
(4.45)
cx=a x
n

Comparing (4.45) with (4.11) it is seen that (4.45) resembles the well-known semi-inverse
scheme which is known to require a small t  !si , which corresponds with slow converge
[61].

4.8 Interaction law variations

69

4.8 Interaction law variations


In the previous sections the factors dominating the basic convergence and the convergence
rate of the viscous-inviscid, boundary-layer and pointwise Newton iterations have been
discussed in detail. The derived analytical results are to be validated numerically, which
will be carried out for three types of interactions laws I, to be introduced next.

4.8.1 Two-sided interaction laws


The rst type of interaction law to be investigated is the so-called two-sided interaction
law, where the interaction law matrix contains a non-zero main diagonal and a certain
number of non-zero upper and lower diagonals next to the main diagonal. For the present
two-sided interaction laws, the (o -)diagonal entries are based on the external ow matrix
E (4.15) and the number of non-zero o -diagonals above the main diagonal is the same as
the number of non-zero o -diagonals below the main diagonal. The interaction law matrices constructed in this way based on (4.15) are symmetric positive de nite M-matrices,
and the two-sided interaction laws I2 are given by
i+q
i+q
X
X
2

Iq =
Eik x + u0 =
Iq2 x + u0 :
(4.46)
k =i q
k =i q
i

ei

ik

ei

where q is the number of upper and lower o -diagonals used.


It is noted that, as downstream in uence is present in the two-sided interaction laws
for q  1, boundary-layer iterations will be required.

4.8.2 One-sided interaction laws


It is known from theorem 4.8 that by decreasing the number of upper diagonals of the
interaction law matrix, the convergence of the boundary-layer iterations improves. For
one-sided interaction law matrices, with the number of upper diagonals set to zero, only
one boundary-layer sweep is required.
For the present one-sided interaction laws I1 , the main diagonal and the lower o diagonal entries are based on the external ow matrix E (4.15) leading to

I1q =
i

i
X

k i q

Eik x + u0
k

ei

i
X

Iq1 x + u0 ;
ik

ei

(4.47)

k i q

where q indicates the number of lower o -diagonals used. It is noted that the two-sided
interaction law matrix with q = 0 can also be categorised as a one-sided interaction law
matrix.

4.8.3 Local tridiagonal interaction law


The above one- and two-sided interaction laws are constructed from the external ow
in uence matrix E by setting a certain number of o -diagonals of E to zero. Another

70

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

possibility is to use a di erent number of in uencing panels surrounding a collocation


point for the construction of matrix I . Recall that E itself is constructed by taking the
in uence of all panels.
For the choice of using only the two neighbouring panels of collocation point xi , the
local tridiagonal interaction law I3 is obtained
1
I3i =

xi+1
Z

xi 1

dx
+ (
d i


d2 x
d
xi ) 2
+ u0
d i xi 

ei

= +1

kX
i

= 1

Iik3 x + u0 ;
k

ei

(4.48)

k i

where x has been approximated by a second-order polynomial to be able to calculate the


Cauchy integral.
For the interaction law constructed via this second approach it can be shown numerically that  (I 3 ) <  (E ), which is the opposite result from (4.29), leading to a non-robust
system according to proposition 4.6. Furthermore, the above constructed matrix I 3 does
not lead to a regular splitting E + V = (I 3 + V ) (I 3 E ) as Iij3 < Eij < 0 for
j = i 1; i + 1, hence giving I 3 E  0 (de nition D.16). The viscous-inviscid and
boundary-layer convergence rate theorems 4.5 and 4.8 therefore can not be used.

4.9 Numerical evaluation of interaction laws


The one- and two-sided interaction laws I1q and I2q and the local tridiagonal interaction
law I3 are to be compared for the calculation of two-dimensional incompressible boundary
layers, which strongly interact with the external ow. The geometry used is the wellknown Carter-Wornom trough [10] de ned by

y =
;
(4.49)
cosh(4(x 3:5))
where < 0 indicates the depth of the trough which has its lowest point at x = 3:5.
Unlike the Carter and Wornom problem the ow is assumed to be turbulent and the
freestream Reynolds number has been set to Re = 1  107 . The boundary-layer model
used is as described in section 2.4. The total number of points over the interval x 2 [1; 6]
is 128 and as an initial guess Blasius' at plate solution is used. In the stagnation point
x = 1 initial conditions are prescribed for the values of the variables and in the endpoint
x = 6 they are determined by extrapolation.
In gures 4.1, 4.2, 4.3 and 4.4 the edge velocity, displacement thickness, shape factor
and skin-friction results are given for a trough with depths = 2; 8 and 12. In
gure 4.4 a large region with ow separation (Cf < 0) is clearly visible, for which the edge
velocity shows the typical plateau region between the start of separation and reattachment.
Four two-sided interaction laws I2q are tested: q = 0, leading to a diagonal interaction
law matrix; q = 1; q = 20; and q = 128, for which the interaction law matrix is equivalent
with the full external ow in uence matrix E . The one-sided interaction laws are used
with q = 1 and q = 20. The results obtained with the quasi-simultaneous method using
these interaction laws I1q , I2q , and the local tridiagonal interaction law I3 , are compared
with results from a fully simultaneous method.

4.9 Numerical evaluation of interaction laws

71

1.15

12

1.1

gamma = -2
gamma = -8
gamma = -12

10
delta^*_x

1.05

ue

1
0.95
0.9

8
6
4

0.85
2

gamma = -2
gamma = -8
gamma = -12

0.8
0.75
1

1.5

2.5

0
3

3.5
X

4.5

5.5

Figure 4.1: Edge velocity distribution for turbulent ow over a trough at Re = 1  107 .

1.5

2.5

3.5
X

4.5

5.5

Figure 4.2: Displacement thickness for turbulent ow over a trough at Re = 1  107 .

2500

0.007
gamma = -2
gamma = -8
gamma = -12

2000

gamma = -2
gamma = -8
gamma = -12

0.006
0.005
0.004

Cf

1500

0.003
0.002

1000

0.001
0

500

-0.001
0

-0.002
1

1.5

2.5

3.5
X

4.5

5.5

Figure 4.3: Shape factor for turbulent ow


over a trough at Re = 1  107 .

1.5

2.5

3.5
X

4.5

5.5

Figure 4.4: Skin-friction distribution for turbulent ow over a trough at Re = 1  107 .

In section 4.7 various forms of relaxation have been discussed of which seven combinations have been assessed. The results obtained using only viscous-inviscid relaxation
are displayed in the tables 4.1 and 4.2. For the results given in table 4.2 viscous-inviscid
relaxation is applied to the full I (4.36), whereas for table 4.1 only the main diagonal of
I is scaled with ! (4.37). The results obtained with the use of only boundary-layer relaxation, as described in (4.34), are shown in table 4.3. In tables 4.4 and 4.5 boundary-layer
relaxation is employed in combination with diagonal or full viscous-inviscid relaxation.
Relaxation due to the unsteady boundary-layer formulation has been used in combination with viscous-inviscid relaxation. Table 4.6 shows the results when applied together
with diagonal viscous-inviscid relaxation and in table 4.7 the unsteady relaxation is combined with full viscous-inviscid relaxation.
In the rst column of the tables the depth of the trough is indicated. The relaxation
parameter ! is shown in the second column. When ! is set to ! = 1 no relaxation
is employed. The relaxation parameters ! = 0:7, ! = 0:4 and ! = 0:1 correspond

72

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method


I3

-2
-2
-2
-2

1.0
0.7
0.4
0.1

78
263
1486

-4
-4
-4
-4

1.0
0.7
0.4
0.1

90
313
1739

-6
-6
-6
-6

1.0
0.7
0.4
0.1

99
333
1865

-8
-8
-8
-8

1.0
0.7
0.4
0.1

348
1909

-10
-10
-10
-10

1.0
0.7
0.4
0.1

2100

-12
-12
-12
-12

1.0
0.7
0.4
0.1

342
1990

I20

680
1200
3749
812
1474
4588
959
1598
5012

1688
5279

5736

1761
5670

164
236
413
1579

61
132
315
1525

194
280
490
1889

71
157
375
1814

208
298
523
2058

169
400
1884

308
544
2160

175
419
1988

223
2218

2137

543
2162

413
2083

I21

I220

480
705
1182
3618
590
863
1447
4428
946
1578
4766

1085
1759
5090

5422

1721
5397

6
88
271
1467
6
103
323
1742
109
348
1881
115
364
1973
380
2149
375
2040

I2128

243
742
1336
4107
283
903
1627
5032
989
1776
5518

1041
1877
5861

1974
6271

1971
6177

2
88
274
1514
2
104
325
1784
3
112
347
1939
3
115
362
1967
383
2047
3
378
2008

=E

I11

86
747
1349
4241
101
912
1652
5194
106
997
1797
5708
106
1064
1899
5988

2001
6313
147

2001
6285

I120

113
186
361
1542

86
160
340
1534

simul.
85
160
339
1573

133
219
431
1855

101
190
403
1866

100
189
403
1819

142
234
459
1986

106
202
430
1940

105
201
430
1947

242
480
2065

208
442
2073

205
443
1997

2185

2167

2180

464
2130

221
413
2077

228
399
2017

Table 4.1: Viscous-inviscid and boundary-layer iterations with diagonal viscous-inviscid


relaxation.
with underrelaxation. If viscous-inviscid relaxation and boundary-layer relaxation are
combined, for both the same indicated ! is applied.
In the other columns of the tables the number of iterations is indicated, where the
roman number gives the number of viscous-inviscid iterations and the number written in
italics the number of boundary-layer iterations. For the one-sided interaction laws and
the `two-sided' interaction law with q = 0 no boundary-layer iterations are registered as
these interaction laws do not contain any downstream information and hence only require
one sweep through the boundary layer. A small bar in the columns implies that the
calculations have crashed, which usually occurs during the Newton iterations.

4.9.1 Basic convergence a ected by the choice of interaction law


From these tables a large number of conclusions can be drawn. In terms of basic convergence it is immediately clear that interaction law I3 is the least robust. If no relaxation
is applied the scheme with I3 does not converge. The row summation of matrix I 3 being

4.9 Numerical evaluation of interaction laws

73

zero, implies that I3 is not very stable and when the diagonal of matrix V becomes too
negative during the Newton iterations, the solution procedure breaks down. Furthermore,
as was discussed in section 4.8.3, it can be shown numerically that  (I 3 ) <  (E ). From
proposition 4.6 it then follows that the Newton iterations to solve for I + V are not robust.
With the help of suitable underrelaxation the row summation can be led away from zero,
and as is seen in tables 4.1 and 4.6, I3 can be made to converge.
For the one- and two-sided interaction laws with matrix I based on E by setting certain
o -diagonals to zero, it follows from theorem 4.7 that  (Ib + V )   (Ia + V )   (E + V )
for two interaction laws matrices Ia and Ib with 0  (Iul )b  (Iul )a , and Ib should lead to
a more robust I + V iterative scheme than Ia . In other words, the more o -diagonals of
the one- and two-sided interaction law matrices are set to zero, the robuster the Newton
iteration process. As a consequence, the diagonal interaction law matrix should be the
most robust.
Looking at the tables, especially 4.1 and 4.2, and comparing the two-sided interaction laws, it is seen that corresponding with the theory, for increasing q , the robustness

I3

-2
-2
-2
-2

1.0
0.7
0.4
0.1

92

1003

61
88
169
723

-4
-4
-4
-4

1.0
0.7
0.4
0.1

71
105
196
904

-6
-6
-6
-6

1.0
0.7
0.4
0.1

210
950

-8
-8
-8
-8

1.0
0.7
0.4
0.1

215
970

-10
-10
-10
-10

1.0
0.7
0.4
0.1

993

-12
-12
-12
-12

1.0
0.7
0.4
0.1

206
952

I21
480
677
1162
4169
590
839
1431
5177

1575
5732

1658
6082

6397

1779
6485

6
11
28
135
6
10
77
394
12
14
-

I220

I2128

243
376
805
3506
283
430
970
4315

491

2
10
258
2
3
-

3
-

3
-

546

=E

I11

I120

133
192
341
1324

101
147
264
1044

142
206
364
1437

106
156
277
1139

374
1503

148
280
1137

1522

288
1028

209
334
1411

199
340
1260

86
413

3910
101

106

106

147

113
163
289
1153

86
126
222
885

Table 4.2: Viscous-inviscid and boundary-layer iterations with full viscous-inviscid relaxation.

74

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method


I3

-2
-2
-2
-2

1.0
0.7
0.4
0.1

61
66
73
82

-4
-4
-4
-4

1.0
0.7
0.4
0.1

71
77
88
99

-6
-6
-6
-6

1.0
0.7
0.4
0.1

96
95

-8
-8
-8
-8

1.0
0.7
0.4
0.1

121

-10
-10
-10
-10

1.0
0.7
0.4
0.1

95

-12
-12
-12
-12

1.0
0.7
0.4
0.1

109

I21
480
759
1441
4527
590
938
1798
5403

2071
5721

5972

5436

6413

6
7
7
22
6
8
10
44
43
43
37
42

I220

I2128

243
317
289
1634
283
354
354
2037

2115

2238

2258

2298

2
3
7
19
2
3
8
34
3
3
10
32
3
4
11
27
10
25
3
-

=E
86
160
344
1605
101
190
408
1960
106
201
440
2108
106
208
455
2075

481
2263
147

Table 4.3: Viscous-inviscid and boundary-layer iterations with boundary-layer relaxation.


generally decreases. However, for some reason, as seen in table 4.1 and 4.3, interaction
2 is also quite robust. For the one-sided interaction laws this behaviour is
law matrix I128
not visible in the tables. The one-sided interaction laws, containing only zeros above the
diagonal, are very diagonally dominant and the e ect, therefore, of altering the number
of lower o -diagonals used for its construction, is negligible. Comparing the one-sided
interaction laws with the two-sided interaction laws for the same q , it is seen in each table
that the one-sided interaction laws are far more robust, as predicted.
It can be concluded that in terms of robustness the one-sided interaction laws and
especially I20 are the most robust.

4.9.2 Basic convergence a ected by the choice of relaxation


Basic convergence of the boundary-layer iteration process can be improved with relaxation. By decreasing ! more cases might be able to converge to a solution as is seen in
each table. However, for ! too small problems can occur as the scheme might not be able
to converge away from the initial solution. The various forms of relaxation applied are

4.9 Numerical evaluation of interaction laws

75

next to be discussed.
When comparing the e ects of the full and diagonal viscous-inviscid relaxation in tables 4.2 and 4.1, it is clear that of the two forms of viscous-inviscid relaxation, diagonal
viscous-inviscid relaxation is preferred, as more cases have been able to converge. Theoretically this is justi ed by the fact that with full viscous-inviscid relaxation the whole matrix
I is scaled with ! , and if the row summation of I is close to zero, also with relaxation
it remains close to zero, leading to breakdown of the Newton iterations. Applying only
relaxation on the diagonal will lead the row sum away from zero. A similar explanation
can be obtained from theorem D.15 and theorem 4.7: as I=! = (Id Iul )=!  Id =! Iul
for 0 < !  1, hence  (I=! + V )   (Id =! Iul + V ), implying that matrix I using
diagonal relaxation leads to a more robust Newton system.
In table 4.3 the results are shown when only boundary-layer relaxation is applied.
Comparing the results of table 4.3 with the results of table 4.1, it is seen that using only
boundary-layer relaxation compared to diagonal viscous-inviscid relaxation leads to a far
less robust system. As explained in section 4.7 this is due to the Newton iterations being

I3

-2
-2
-2
-2

1.0
0.7
0.4
0.1

61
136
314
1227

-4
-4
-4
-4

1.0
0.7
0.4
0.1

71
160
367
1464

-6
-6
-6
-6

1.0
0.7
0.4
0.1

173
406
1569

-8
-8
-8
-8

1.0
0.7
0.4
0.1

179
414
1597

-10
-10
-10
-10

1.0
0.7
0.4
0.1

182
439
1792

-12
-12
-12
-12

1.0
0.7
0.4
0.1

414
1662

I21
480
1089
3268
54048
590
1344
4042
66893

1480
4565
74462

1583
4865
77808

1742
5178
81632

5194
80672

6
83
258
1235
6
98
306
1434
106
321
1578
109
339
1633
116
368
1665
347
1603

I220

I2128

243
1191
3800
56170
283
1455
4687
68196

1600
5124
74998

1697
5472
79520

1813
5905
82394

5885
82528

2
82
256
1176
2
98
304
1471
3
98
323
1541
3
107
335
1567
114
352
1688
3
1624

=E
86
1203
3867
55460
101
1486
4766
69239
106
1619
5256
75471
106
1714
5620
78863

1894
6014
83631
147

83153

Table 4.4: Viscous-inviscid and boundary-layer iterations with diagonal viscous-inviscid


relaxation and boundary-layer relaxation.

76

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method


I3

-2
-2
-2
-2

1.0
0.7
0.4
0.1

61
91
163
516

-4
-4
-4
-4

1.0
0.7
0.4
0.1

71
106
191
597

-6
-6
-6
-6

1.0
0.7
0.4
0.1

116
201
626

-8
-8
-8
-8

1.0
0.7
0.4
0.1

216
667

-10
-10
-10
-10

1.0
0.7
0.4
0.1

216
661

-12
-12
-12
-12

1.0
0.7
0.4
0.1

718

I21
480
1061
3301
36781
590
1311
4125
44024
1461
4573
46918
4919
49311
5182
51152
51234

6
10
20
201
6
16
25
227
12
26
242
12
27
229
13
29
265
279

I220

I2128

243
522
1546
17217
283
552
1892
19875
669
2050
21084
707
2141
21194
770
2272
23116
23691

2
10
23
196
2
10
26
265
3
257
3
301
281
3
308

=E
86
629
1807
17452
101
685
2058
21378
106
22020
106
24029
24953
147
25510

Table 4.5: Viscous-inviscid and boundary-layer iterations with full viscous-inviscid relaxation and boundary-layer relaxation.
less robust when solving for the boundary-layer system I + V than for the viscous-inviscid
relaxed system Id =! Iul + V . It is noted, however, that with the use of boundary-layer
relaxation the robustness for the full interaction law I2128 is improved. When comparing
table 4.3 with table 4.2 the results obtained with boundary-layer relaxation are an improvement compared to the results obtained with full viscous-inviscid relaxation. Here,
the same reasoning used when comparing tables 4.2 and 4.1 does apply. In table 4.3 only
the diagonal of I is scaled, leading the row sum away from zero, whereas in table 4.2 the
whole matrix I is relaxed.
In tables 4.4 and 4.5 boundary-layer relaxation is used in combination with diagonal
or full viscous-inviscid relaxation. For both cases an improvement is visible, compared
to the cases where viscous-inviscid relaxation is applied without boundary-layer relaxation. For the local tridiagonal interaction law I3 for which the boundary-layer system
I 3 + V is very unstable as  (I 3 ) <  (E ) (proposition 4.6), the use of boundary-layer
relaxation is a disadvantage. Furthermore, although full viscous-inviscid relaxation compared to boundary-layer relaxation is less robust, the two forms of relaxation combined

4.9 Numerical evaluation of interaction laws

77

is of the three forms most desirable. Of course, diagonal viscous-inviscid relaxation using
boundary-layer relaxation compared to full viscous-inviscid relaxation using boundarylayer relaxation is far more robust.
Relaxation due to the use of the unsteady boundary-layer formulation has been used
for calculations presented in the tables 4.6 and 4.7. It is seen, comparing tables 4.6 and
4.7 with tables 4.1 and 4.2, that when relaxation due to the use of the unsteady boundarylayer formulation is used in combination with viscous-inviscid relaxation, the robustness
of the system is generally decreased. The timestep t might have been taken too small,
making it impossible for the system to converge away from the initial solution. Further
research is required to examine the e ects of relaxation due to the use of the unsteady
boundary-layer formulation more closely.

I3

-2
-2
-2
-2

1.0
0.7
0.4
0.1

93
277
1460

-4
-4
-4
-4

1.0
0.7
0.4
0.1

105
326
1736

-6
-6
-6
-6

1.0
0.7
0.4
0.1

111
345
1901

-8
-8
-8
-8

1.0
0.7
0.4
0.1

360
1983

-10
-10
-10
-10

1.0
0.7
0.4
0.1

2075

-12
-12
-12
-12

1.0
0.7
0.4
0.1

372
-

781
1253
3774

865
1517
4639

932
1642
5091

1732
5434

5770

1848

I20

175
247
427
1655

73
146
330
1591

205
290
503
1919

172
388
1847

218
309
539
2057

182
414
1969

320
557
2128

188
424
2033

339
2228

2101

330
573
-

441
-

I21

I220

543
761
1228
3717

913
1486
4500

991
1667
4963

1117
1724
5238

5483

1820

96
288
1499
118
339
1822
125
359
1932
129
374
1998
2028
395
-

800

1391
4200

960
1675
5173

1043
1816
5633

1090
1913
5948

6211

2064

I2128

102
287
1505
119
341
1767
126
360
1878
129
375
2028
137
2103
396
-

=E

817
1403
4281

977
1699
5226

1106
1849
5696

1118
1947
6101

1263

6426

2084

I11

I120

127
198
377
1583

103
174
353
1571

simul.
103
174
353
1582

147
233
444
1912

118
204
417
1853

117
203
418
1826

155
247
473
2009

124
216
442
1992

123
215
440
1967

254
491
2075

221
458
2103

221
456
2071

266
2213

232
2163

229
2141

255
498
-

460
-

449
2097

Table 4.6: Viscous-inviscid and boundary-layer iterations with diagonal viscous-inviscid


relaxation using the unsteady boundary layer formulation with t = 1:

78

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method


I3

-2
-2
-2
-2

1.0
0.7
0.4
0.1

73
102
180
740

-4
-4
-4
-4

1.0
0.7
0.4
0.1

57

1884

119
209
887

-6
-6
-6
-6

1.0
0.7
0.4
0.1

125
224
909

-8
-8
-8
-8

1.0
0.7
0.4
0.1

909

-10
-10
-10
-10

1.0
0.7
0.4
0.1

1004

-12
-12
-12
-12

1.0
0.7
0.4
0.1

235
-

I21
543
739
1216
4262

879
1554
5245

964

1621
5770

6188

6515

1928

36
36
38
37
-

I220

=E
-

I11

127
177
302
1167

I120

147
206
354
1358

118
165
281
1072

155
217
375
1465

124
173
296
1122

227
388
1511

178
304
1178

237
1564

179
1107

384
-

190
315
-

987

1025

1068

1035

I2128

103
142
243
890

Table 4.7: Viscous-inviscid and boundary-layer iterations with full viscous-inviscid relaxation using the unsteady boundary layer formulation with t = 1.

4.9.3 Convergence rate a ected by the choice of interaction law


From theorem 4.5 it follows that the smaller the (number of) o -diagonals of the one- and
two-sided interaction law matrices, the slower the viscous-inviscid iteration process converges. Comparing the number of viscous-inviscid iterations for the one-sided interaction
laws and I20 in each table, it is seen that for decreasing q , in correspondence with theorem 4.5, the number of iterations increases, except for three cases with ! = 0:1. When
comparing the two-sided interaction laws I21 and I220 for approximately the same number
of boundary-layer iterations, a similar behaviour is seen.
It is also interesting to compare the two tridiagonal interaction laws I3 and I21 . As
matrix I 3 has larger o -diagonals than matrix I12 , the number of viscous-inviscid iterations
should according to theorem 4.5 be smaller. This is shown in tables 4.1 and 4.6 comparing
ow cases for which the number of boundary-layer iterations is approximately the same.
It is furthermore clear from tables 4.1 and 4.6 that the number of viscous-inviscid
iterations for the one-sided interaction law I120 corresponds with the number of viscous-

4.9 Numerical evaluation of interaction laws

79

inviscid iterations obtained with the simultaneous method. If no relaxation is applied also
2 = E,
the number of boundary-layer iterations for the two-sided interaction law with I128
only requiring two or three viscous-inviscid iterations, can be seen to correspond (table
4.1), as expected.
The number of iterations of interaction law I120 compared to the number of iterations
of interaction law I20 is seen to di er a factor two, for ! = 1 in table 4.1. As I120 can be
identi ed as a Gauss-Seidel method and I20 as a Jacobi method, which for consistently
ordered matrices is known to be twice as slow as Gauss-Seidel, the factor two between
the two methods is explained.
The convergence speed of the boundary-layer iterations is clari ed by theorem 4.8. It
states that for increasing number of upper diagonals of the interaction law matrix, the
convergence of the boundary-layer iterations deteriorates. This is seen to correspond with
the numerical results, when comparing the number of boundary-layer iterations for the
two-sided interaction laws I220 and I2128 for the same number of viscous-inviscid iterations
for instance in table 4.1.

4.9.4 Convergence rate a ected by the choice of relaxation


With the use of relaxation more cases can be made to converge but a price is usually
paid in terms of convergence rate. This immediately follows from theorem D.18 as for
0  Na  N  N=!  Nb , where 0 < !  1, (Ma 1 Na )  (Mb 1 Nb ).
Comparing the convergence rate of the one-sided interaction laws using full or diagonal
viscous-inviscid relaxation it is seen that with diagonal viscous-inviscid relaxation far
more viscous-inviscid iterations are required. This is explained with theorem D.18 as
Nb  Id =! Iul E  I=! E  Na , which implies that I using the full viscous-inviscid
relaxation converges faster.
Comparing the results for the two-sided interaction laws using boundary layer relaxation in table 4.3 and viscous-inviscid relaxation in tables 4.1 and 4.2, it is clear that
boundary-layer relaxation leads to faster convergence for the viscous-inviscid iterations.
The reason for this is that only relaxation is applied to the boundary-layer iterations and
not to the viscous-inviscid iterations. Viscous-inviscid relaxation a ects both the viscousinviscid iterations and the boundary-layer iterations. boundary-layer relaxation used in
combination with full or diagonal viscous-inviscid relaxation deteriorates the convergence
when compared to using only viscous-inviscid relaxation, as during the boundary-layer
iterations the amount of relaxation has doubled.
Adding relaxation due to the unsteady boundary layer formulation to the viscousinviscid relaxation can even further deteriorate the convergence. However, for the present
timestep t = 1 the e ect is very small as seen in tables 4.6 and 4.7.

4.9.5 Summary
The results obtained by the numerical evaluation of interaction laws are brie y summarised in the present section.
It has been shown that by decreasing the number of o -diagonals used for the construction of two-sided interaction laws, the basic convergence of the system increases. The

80

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

basic convergence is even further enhanced by ignoring the downstream in uence in an


interaction law, and the one-sided interaction laws and especially the diagonal interaction
law therefore lead to the most robust systems, in correspondence with theorem 4.7.
The rate of convergence is clearly a ected by the number of o -diagonals used for
the construction of an interaction law. It was seen that by decreasing the number of
o -diagonals, the viscous-inviscid convergence deteriorates, which is consistent with theorem 4.5. Decreasing the number of o -diagonals also resulted in faster boundary-layer
convergence for the two-sided matrices, consistent with theorem 4.8.
Basic convergence can be improved by relaxation, and of the various forms tested it
is clear that diagonal viscous-inviscid relaxation (possibly in combination with boundarylayer relaxation) leads to the best basic convergence. Relaxation due to the unsteady
boundary-layer formulation seems to decrease the robustness. Further research is required
to investigate the e ects of relaxation due to the use of the unsteady boundary-layer
formulation in greater detail.
The use of relaxation slows down the viscous-inviscid and/or boundary-layer convergence. However, it is noted that applying full viscous-inviscid relaxation resulted in faster
convergence for the viscous-inviscid iterations than diagonal viscous-inviscid relaxation.
It can be concluded that using a diagonal interaction law matrix with the help of
diagonal viscous-inviscid relaxation is the most simple and robust option. Its convergence
is slightly slower compared to the one-sided interaction laws. However, compared to
the two-sided interaction laws, the overall convergence is much faster, as the diagonal
interaction law does not require boundary-layer iterations.

4.10 E ect discretisation on interaction law


The interaction laws, discussed in the previous sections, have been based on the external
ow in uence matrix E , given in (4.15). For the construction of E , the locations of the
unknowns ue and x have both been de ned in the panel endpoints, which corresponds
with discretisation A in gure 4.5. The diagonal and two upper and lower o -diagonal
entries of I A = E are given by
ln 2 2
ln 3=4
4
; IiiA 1 = IiiA+1 =
; IiiA 2 = IiiA+2 =
:
(4.50)
 x
 x
 x
In this section it will be shown that a di erent form of discretisation will lead to a di erent

IiiA =

= location ue
= location *x

A
B
C
D
i-1

i+1

Figure 4.5: Location unknowns.

4.10 E ect discretisation on interaction law

81

sized and structured interaction law matrix, which might be less suitable for the quasisimultaneous interaction.
By taking ue in the panel centres and keeping x in the panel endpoints, shown in gure
4.5 as discretisation B, a di erent matrix structure for I is obtained. Using interaction law
I A (4.50) and determing the velocity by averaging, interaction law matrix I B is created,
for which
2 + ln 2
ln 3=2 2
;
IiB 21 i 2 = IiB 21 i+1 =
:
(4.51)
IiB 21 i = IiB 21 i 1 =
2 x
2 x
The above choice of locations for the unknowns is not very suitable for the quasi-simultaneous coupling scheme. As is seen in (4.51) IiB 1 i = IiB 1 i 1 , implying that both symmetry
2
2
and diagonal dominance have been lost. Furthermore, the diagonal entries are very small,
IdB  0:3IdA , which will lead to problems for more dicult ow cases. Actually, interaction
law I B is constructed from I A by averaging,
Another alternative is to determine the velocity in point xi 12 directly with the potential ow theory integral for the velocity
x
Z +1 
M
M +1
X
dx
d
1X
+
u
=
I~iB 12 k x + u0 1 ; (4.52)
ue 1 =
0

1
2
 k=1
d k+ 21 xi 12 
2
2
k =1
x
p

e
i

where
ln 3
ln 9=5
I~iB 12 i = I~iB 12 i 1 =
;
I~iB 12 i 2 = I~iB 12 i+1 =
:
(4.53)
 x
 x
It is seen that the structure of matrix I~B resembles the structure of I B . However, the
matrix entries of I~B are even smaller than those of I B .
Discretisation C in gure 4.5, where the location of ue is taken in the panel endpoints
and the location of x in the panel centres, leads to a similar interaction law as was
obtained with discretisation B. Replacing x in interaction law I A by (x 1 + x + 1 )=2
2
2
leads to
2 + 2 ln 2
ln 3=2 2
;
IiiC 23 = IiiC+ 32 =
;
(4.54)
IiiC 12 = IiiC+ 21 =
2 x
2 x
and it is seen that interaction law matrix I C has the same unsuitable structure as I B .
When both unknowns are de ned in the panel centres, as seen in discretisation D in
gure 4.5, interaction law matrix I D is constructed, with
2 + ln 2
ln 3
IiD 21 i 12 =
;
IiD 21 i+ 12 = IiD 21 i 23 =
:
(4.55)
2 x
4 x
Matrix I D is symmetric but the values on the rst upper and lower o -diagonal are
positive, which implies that I D is not an M-matrix. Furthermore, compared with I A , it
is seen that the main diagonal entries are very small, IdD  0:3IdA .
Above the e ect of changing the location of the unknowns has been discussed, and
it is clear that only de ning both unknowns in the panel endpoints leads to a suitable
interaction law matrix.
k

82

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

4.11 Geometric interpretation iterative variations


In section 4.9 it has been concluded that interaction law I3 is most unsuitable for the quasisimultaneous method. The discussed one- and two-sided interaction laws perform much
better, however, as they are not easily physically interpreted, the geometric interpretation
of the iterative variations is clari ed with interaction law I3 .
The geometric interpretation of the local growth of the boundary layer described by
an interaction law can be visualised by a small parabolic bump. For the case that the
boundary-layer iteration process corresponds with a Gauss-Seidel scheme, using the local
tridiagonal interaction law (4.48), the interaction law is written as
4 (~ )
2 (~ 1)
2 (~ )
x 1 +
x

+ u0 :
(4.56)
 x
 x
 x x +1
The local growth of the boundary layer, modelled by (4.56), corresponds with the second
bump depicted in gure 4:6. If instead the boundary-layer iterative process follows the
Jacobi method, where

I3GS =

ei

2 (~ 1)
4 (~ )
2 (~ 1)
x 1 +
x

+ u0 ;
(4.57)
 x
 x
 x x +1
the geometric interpretation is given by the rst bump in gure 4:6.
When the gradient over the second part of the parabolic bump is kept constant, the
following interaction law version is produced:

I3JAC =

ei

2 (~ 1)
2 (~ )
2 (~ 1)
2 (~ 1)
x 1 +
x +
x

+ u0 ; (4.58)
 x
 x
 x
 x x +1
which results in a JOR method for the boundary-layer iterations with relaxation parameter

I3JOR =

!JOR = 1 +

ei

2
:
2 + b=a x

(4.59)

The SOR version is given by


2 (~ )
2 (~ )
2 (~
x 1 +
x +

 x
 x
 x x

I3SOR =

bump

JAC

*x (n)

bump

GS

*x (n)

xi

x i+1

1)

bump

JOR

xi

x i+1

*x (n)

ei

bump

SOR

~
~

*x (n-1)

x i-1

xi

x i+1

(4.60)

*x (n)

*x (n-1)

x i-1

2 (~ 1)

+ u0 ;
 x x +1
i

*x (n-1)

x i-1

*x (n-1)

x i-1

xi

x i+1

Figure 4.6: Geometric interpretation Jacobi, Gauss-Seidel, JOR and SOR method.

4.11 Geometric interpretation iterative variations

83

for which the geometric interpretation is displayed by the fourth bump in gure 4:6. For
a symmetric matrix I + V the SOR iterative method, with a constant relaxation factor
!SOR , converges if and only if I + V is positive de nite and 0 < !SOR < 2. As !SOR
is also given by (4.59), it is seen that at separation (b=a = 0) !SOR = 2, whereas inside
the separated ow region !SOR is even larger than 2. As !SOR is not constant the above
statement does not strictly apply, however, it can be considered highly unlikely that SOR
as in (4.60) will converge for separated ow.

84

Chapter 4 Viscous-inviscid interaction with the quasi-simultaneous method

Chapter 5
Applications of the
quasi-simultaneous method
5.1 Introduction
Although a lot of work has been done on two-dimensional coupling, as discussed in chapter
1, results obtained for the industrially more relevant application of three-dimensional
viscous-inviscid interaction are still limited. The numerical modelling of the hyperbolic
nature of the system of equations and the lack of suitable empirical closure relations make
the three-dimensional coupled problem a more dicult one to solve.
In the preceding chapter, viscous-inviscid interaction with the quasi-simultaneous
method for the model problem of two-dimensional ow over a plate has been investigated. In the present chapter the quasi-simultaneous technique is extended to three
dimensions and applied to the more practical cases of two-dimensional aerofoil ow and
three-dimensional dented plate and wing ow. Furthermore, the solution procedure of the
coupled system is considered in more detail.

5.2 Interaction three-dimensional systems


The analysis of interaction algorithms for three-dimensional systems is di erent from the
analysis for the two-dimensional steady ow problems presented in section 4.2. Unlike
the two-dimensional steady ow case, where the system of equations is parabolic, the
three-dimensional ow case leads to a hyperbolic system of equations.
An analogy exists between the three-dimensional and the unsteady two-dimensional
case, and as the unsteady two-dimensional case is easier to understand it is analysed next.

5.2.1 Two-dimensional unsteady interaction


For the two-dimensional unsteady ow problem the boundary-layer equation, given in
section 2.4, is written as

@ue b @x c @x


d
+
+
= ;
@x a @x a @t
a

(5.1)

86

Chapter 5 Applications of the quasi-simultaneous method

with b=a changing sign from positive to negative at H = Hsep (4.4) and c=a > 0 and
d=a > 0.
It is noted that in the preceding chapter the unsteady term in the boundary-layer
equation has been treated numerically as a form of relaxation for the steady parabolic
problem.

Direct method
Applying the classical direct interaction algorithm (4.2) to equation (5.1), with the edge
velocity assumed known from the external inviscid ow, the following partial di erential
equation for x remains to be solved:
d
b @x c @x
+
=
a @x a @t
a
The characteristic of the above equation is given by
 =

dx
dt

=)

 =

@ue
:
@x

(5.2)

b=a
;
c=a

where  represents the tangent of the angle between the characteristic and the x-axis. If a
system of partial di erential equations has complex 's the equations are called elliptic. If,
as is the case here, the 's are real and distinct the equations are called totally hyperbolic.
It is seen that for the above scheme the characteristic  has a positive sign when
H < Hsep (b=a > 0), and a negative sign when H > Hsep (b=a < 0). The characteristic
passes through zero when H = Hsep, and unlike for the two-dimensional steady ow case,
this point is generally non-singular for the hyperbolic problem. The passage through
zero only implies occurrence of in uence from downstream to upstream because the wall
streamline becomes orthogonal to the external streamline, which is associated to ow
reversal. However, Cousteix and Houdeville [20] have shown that for this characteristic
family non-physical singularities can occur from the focusing of characteristic lines.

Strong interaction methods


For comparison, boundary-layer equation (5.1) is solved together with a representation of
the external inviscid ow. Using again the simple linear relation ue 1 x = 2 (4.10),
representing either the external inviscid ow equation (fully simultaneous method) or the
interaction law equation (quasi-simultaneous method), equation (5.1) is reduced to


b
1 +
a

@x c @x
d
+
= ;
@x a @t
a

and  is found to be

 =

1 + b=a
:
c=a

It is seen that with the use of the outer ow representation (4.10) the characteristic  does
not change sign at H = Hsep. Hence, if the ow is not too strongly separated, implying

5.3 Newton solution procedure

87

that the angle between the wall streamline and the external streamline is only slightly
larger than 90o , coecient 1 + b=a remains positive, as the value of b, although negative,
is small.
Cousteix and Houdeville in [20] have shown that with an inverse method the occurrence
of any singularity is avoided. Similarly, the strong coupling techniques most probably
prevent the occurrence of a discontinuity line.

5.2.2 Three-dimensional interaction


The above discussed behaviour of the two-dimensional unsteady hyperbolic problem shows
great resemblance with the behaviour of the steady and unsteady three-dimensional problem, when solved with the direct scheme, as described in [18, 20, 94]. When for the steady
or unsteady three-dimensional boundary-layer equations coecient b passes through zero,
some characteristics change direction from upwind to downwind and convergence of characteristics might lead to the formation of a non-physical discontinuity line. When solved
simultaneously with an interaction law, the characteristics generally originate from the
upwind direction, most probably avoiding the occurrence of any singularity.

5.3 Newton solution procedure


Before continuing with the formulation of the three-dimensional quasi-simultaneous coupling scheme, the procedure to solve for the boundary-layer and external ow equations
which are combined with the quasi-simultaneous method, is explained in more detail.
The coupled system being non-linear, an iterative scheme is to be employed to solve the
set of equations. As a consequence of the viscous-inviscid interaction technique chosen,
the Newton method is to be implemented which is the best-known procedure for solving
non-linear problems. For a system of non-linear equations of the form

R(U) = 0;
where U is the vector of unknowns, Newton's method is written


@R (n) (n+1)
(U
U(n) ) = R(U(n) ):
@U

(5.3)

The term @R=@U corresponds with the Jacobian matrix J of the system of equations
at some Newton iteration level n . Each entry Jkl of the Jacobian represents the partial
derivative of the kth equation with respect to the lth variable of vector U

Jkl =

@Rk (U)
:
@U (l)

The Jacobian matrix J can be calculated numerically via

Jkl =

Rk (U (l) + ") Rk (U (l) ")


;
2"

(5.4)

88

Chapter 5 Applications of the quasi-simultaneous method

or analytically, by making use of the chain-rule. Of the two possibilities the numerical
computation is employed for which it is important that " is suitably chosen for U (l).
A setback of the numerical approach is that for each iteration residuals at positions l,
l " and l + " have to be calculated, and hence by implementing the analytical approach
Newton's solution procedure might be faster.

Present application Newton's method


Newton's method is applied per station for both the two- and three-dimensional ow cases
to solve for the unknowns vector U. For the various cases U, containing the discrete
viscous and inviscid ow variables, is given by
Two-dimensional plate/aerofoil case

Quasi-three-dimensional swept tapered wing case


0

U =

s
H
us

1
A;

U =

B
B
B
B
@

ss
H

c
us

1
C
C
C;
C
A

Three-dimensional dented plate case


0

U =

B
B
B
B
@

ss
H

ue
ve

Three-dimensional wing case


0

1
C
C
C;
C
A

U =

B
B
B
B
B
B
@

ss
H

Ue
Ve
We

1
C
C
C
C;
C
C
A

The coupled system of equations R contains the set of discretised boundary-layer


equations described for each ow case in chapter 2, and the interaction law formulae, as
derived in chapter 4.
For reasons of simplicity the interaction law formulae are linearised by using the velocity of the previous viscous-inviscid iteration, which in two-dimensions results in
n)
Ii = u(
=
e
Ii

( )
Iik qe(n 1) x :
n

(5.5)

The interaction law formulae being linear, the number of equations can be reduced
by pre-eliminating the velocity unknowns. This has been done for the two-dimensional
and quasi-three-dimensional ow case to minimise the computational e ort. For the fully
three-dimensional ow calculations, this approach requires more rewriting and the preelimination of the velocity unknowns has not been implemented.
In the two- and quasi-three-dimensional cases, pointwise Newton iterations are performed until convergence per station is obtained, before Newton iterations start at the

5.4 Three-dimensional quasi-simultaneous scheme

89

next point. The Newton stopcriterion for these cases with i and j xed is
n+1)
n)
X jU (l)(
U (l)(
ij
ij j
< N ;
jU (l)(n+1)j + 1  10 10

(5.6)

ij

where N is the allowed error of the Newton iterations.


For the fully three-dimensional case the scheme is somewhat di erent. Pointwise
Newton iterations are performed per spanwise section until convergence is obtained for
the spanwise section. For only i xed
!
n+1)
n)
X jU (l)(
U (l)(
ij
ij j
< N :
(5.7)
max
j
jU (l)(n+1) j + 1  10 10
ij

5.3.1 Initial conditions


A good initial guess is required for every grid point to obtain convergence with Newton's
method. For the methods presented in this thesis three possibilities are included, being

initialisation with the Blasius at plate solution, which is useful to initialise plate
ow or wing/aerofoil ow at low angles of attack;

initialisation with the solution from a prior calculation, which is useful at higher
angles of attack;

initialisation by setting at the rst iteration for each point (i; j ) to be calculated
the values of all downstream points equal to the values at the previously calculated
upstream point

Ukj = Uk 1j ;

for k > i

1:

The last approach can be applied for each ow case. It might be slower than the previous
two approaches but does on the other hand not require any knowledge of the speci c ow
case to be calculated.

5.4 Three-dimensional quasi-simultaneous scheme


The quasi-simultaneous method has a straightforward extension to three dimensions. For
the present three-dimensional quasi-simultaneous method this is done by constructing
an interaction law consisting of three separate formulae for the global Cartesian edge
velocities Ue , Ve and We , termed I U , I V and I W .
Introducing the operators, Mx, My and En for the boundary-layer x-momentum, y momentum and entrainment equations, the total system of equations to be solved during

90

Chapter 5 Applications of the quasi-simultaneous method

the nth downstream viscous-inviscid march, can be written symbolically as follows:


8
> U (n)
I U [U(n) ]
= U (n 1) I U [U(n 1) ];
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
<

boundary-layer region

>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
:

eV

external ow region

>
>
>
>
>
:

Ve(n)
V

I V [U(Vn) ]

= Ve(n 1)

I V [U(Vn 1) ];

We(n)

I W [U(Vn) ]

= We(n 1)

I W [U(Vn 1) ];

Mx [Ue(n) ; Ve(n) ; We(n) ; U(Vn) ] = 0;


V

My [Ue(n) ; Ve(n) ; We(n) ; U(Vn) ] = 0;


V

(simultaneous)

En [Ue(n) ; Ve(n) ; We(n) ; U(Vn) ] = 0;


V

8
>
>
>
>
>
<

eE

Ue(n)

= E U [U(Vn) ];

Ve(n)
E

= E V [U(Vn) ];

We(n)

= E W [U(Vn) ];

(direct)

where UV represents the vector with the viscous boundary-layer unknowns, which is part
of the unknowns vector U (section 5.3), having excluded the velocity components.
The three interaction law equations are solved simultaneously with the three boundarylayer equations, producing three new global velocity components, Ue , Ve , We and a
new boundary-layer unknowns vector UV of dimension three. With the new boundarylayer unknowns vector UV , a new external inviscid ow calculation can be performed,
using the direct scheme, etc.
V

Stopcriteria
The viscous-inviscid iterations between the boundary layer and external ow are continued
until
max jU (n) U (n 1) j < U ; max jV (n) V (n 1) j < V ; max jW (n) W (n 1) j < W ;
i;j

eVij

eVij

vi

i;j

eVij

eVij

vi

i;j

eVij

eVij

vi

where vi is the allowed viscous-inviscid error. When viscous-inviscid relaxation is applied,
vi is to be multiplied by !vi , the viscous-inviscid relaxation parameter.
The boundary-layer iterations have either been set to the xed number of three or
they are performed until
max jU (~n) U (~n 1) j < U ; max jV (~n) V (~n 1) j < V ; max jW (~n) W (~n 1) j < W ;
i;j

eVij

eVij

bl

i;j

eVij

eVij

bl

i;j

eVij

eVij

bl

where n~ indicates the n~ th boundary-layer iteration. The allowed boundary-layer error is


de ned by bl , and when boundary-layer relaxation is applied, bl should be multiplied by
!bl .

5.5 Three-dimensional plate interaction law

91

5.5 Three-dimensional plate interaction law


In chapter 4 the various requirements for the construction of a suitable interaction law
have been discussed. In particular it was described in section 4.10 that the edge velocity
and the boundary-layer unknowns should be de ned in the node points of the panels in
order to obtain an interaction law matrix with the correct structure.
For the construction of a three-dimensional plate interaction law discretisation A of
gure 4.5 is therefore to be applied. The velocity components are determined from the
potential- ow theory integrals for the velocity components and, following discretisation
A, calculated in node point (xi ; yj ). In other words, the interaction law can be constructed
from the three-dimensional viscous potential- ow model, as derived in section 3.4.3
1
ue (x; y ) =
2

yend 
xend Z
Z

@
@
(x  )dd
(qe x ) + (qe y )
3 + u0 (x; y ); (5.8)
@
@
((x  )2 + (y  )2 ) 2
e

xbegin ybegin

1
ve (x; y ) =
2

x
Zend

yend 
Z

@
@
(y  )dd
(qe x ) + (qe y )
3 + v0 (x; y ); (5.9)
@
@
((x  )2 + (y  )2 ) 2
e

xbegin ybegin

where we(x; y ) = @=@x(qe x ) + @=@y (qe y ). After the discretisation equations (5.8) and
(5.9) become
M +1 M +1
X X


u
ue =
qe Auijkl x + Bijkl
y + u0 ;
(5.10)
k =1 l=1
M +1 M +1
X X


v
ve =
qe Avijkl x + Bijkl
y + v0 :
(5.11)
k =1 l=1
A third interaction law formula for we = We is not required as both ue = Ue and ve = Ve
are tangent to the surface of the plate.
If the grid is constructed by squares, y = x, the following matrix entries of Au are
obtained:
p

ij

eij

kl

kl

kl

eij



1
3 + 2 2
p ;
ln
=
 x 3 2 2

Auiji

kl

kl

ij

u
ijij

kl

(5.12)




2
1
3 + 2 2
u
p
p
p
=
A
=
ln
+ ln

1j
iji+1j
3 2 2
2 x
9+3 5 6 2

!

;
10

(5.13)

which contain the main in uence at node point (xi ; yj ). The other matrix entries of Au
and those of B u are much smaller. For the present square grid the matrix entries of B v
v
u
for the y -velocity component resemble the entries of Au as Bijij
+
k = Aiji+kj and similarly
do the entries of Av resemble those of B u .
The matrix entries for B u and Av being small, it is decided that for the construction of
interaction law formulae for ue and ve matrices B u and Av are to be ignored. The formula

92

Chapter 5 Applications of the quasi-simultaneous method

for ue hence only contains a qe x part and the formula for ve only takes the contribution
of the qe y part into account. The remaining matrices Au and B v have the following
characteristics:
v
Auijij ; Bijij
> 0;
u
u
v
v
Aijkl = Aklij ; Bijkl = Bklij  0; (i 6= k, j 6= l);
M +1 M +1
M +1 M +1
X X
X X
u
v
Aijkl ;
Bijkl
 0; with inequality for at least one index ij;
k =1 l=1
k =1 l=1
p

which means that both Au and B v are positive de nite M-matrices (see theorems D.10
and D.12). It follows that interaction law matrices based on Au and B v by setting certain
o -diagonals to zero, have the same characteristics and are therefore suitably constructed
for the quasi-simultaneous interaction scheme.
Looking in more detail at the matrix entries of Au , it is seen that for the interaction
law formula for ue the entries Auijkl for l 6= j are small and can be omitted. Similarly for
v
ve the entries Bijkl
for k 6= i can be neglected. This results in the following interaction
law formulae for ue and ve , respectively:
ij

ij

Iuij =

kj

Ivij =

Auijkj qe x + u0

v
Bijil
qe y + v0
il

eij

kj

il

eij

(5.14)

v
Iijil
qe y + v0 :

(5.15)

kj

u
Iijkj
qe x + u0 ;

il

eij

kj

il

eij

The number of o -diagonals of matrices Au and B v can be varied for the construction of
a suitable interaction law as discussed in chapter 4.

5.5.1 Comparison two- and three-dimensional interaction laws


The above result for three-dimensional plate ow can be shown to correspond with the
result found earlier in two dimensions. When comparing the diagonal matrix entries of
Au for the two- and three-dimensional problem, (3.54) and (5.10), it is seen that

Auii
Auijij

4
 x
p

1
3 + 2 2
p
=
ln
 x 3 2 2
=

 1:27=x;
 1:12=x:

Even closer correspondence can be shown by taking the limit y


principal part of the singular integral (3.58)
lim
y!1

! 1 in the Cauchy

p


y ( x2 + y 2 + x)2 @we
ln p
2 ( x2 + y 2 x)2 @x ij

5.6 Dirichlet aerofoil interaction law

lim
y!1

y
2

lim
y!1

y
2

93
p



@we
4x x2 + y 2


p
2x2 + y 2 2x x2 + y 2 @x ij
p
!
2
2
4x x + y
@we
p
2x2 + y 2 2x x2 + y 2 @x ij


ln 1 +

1

x22 + 1
4
x

y
@we A
1
q
= lim @
2
2
y!1
2 2 x2 + 1 2 x x2 + 1 @x ij
y
y y

2x @we
=
;
 @x i
0

which equals the Cauchy principal part in the two-dimensional case (3.52).
It follows from the above that the diagonal entries of the interaction law matrices
1
u
I (4.47) and I u (5.14) constructed from Au are comparable: Iii1  Iijij
. A similar
1
u
resemblance exists for the o -diagonals of I and I .

5.6 Dirichlet aerofoil interaction law


Before continuing with the construction of an interaction law for the three-dimensional
wing problem, an interaction law is constructed for the simpler two-dimensional aerofoil
problem.
An often seen choice is the use of a two-dimensional plate interaction law, as constructed in chapter 4, to represent the aerofoil's edge velocity. The alternative is to derive
an interaction law from the actual external ow formulation. The latter option more
accurately represents the inviscid ow region, taking into account the Kutta condition,
and is to be applied for the present aerofoil ow case.
For the calculation of the actual two-dimensional aerofoil velocity two well-known
approaches exist, being the Neumann and the Dirichlet method. The Neumann method,
uses the zero normal velocity boundary condition and determines the velocity integrals in
the panel centres. The boundary layer unknowns being de ned in the panel endpoints, the
Neumann method can be seen to correspond with discretisation B, discussed in section
4.10, leading to an unsuitable interaction law matrix.
The Dirichlet method on the other hand uses a constant potential boundary condition and from the doublet strengths the velocity is determined in the panel endpoints,
corresponding with discretisation A. Unlike the Neumann method, the Dirichlet method
produces a suitable interaction law for the quasi-simultaneous interaction technique and
is to be applied.
In section 3.4.4 the two-dimensional streamwise edge velocity equation for aerofoil
ow, using the Dirichlet approach has been derived
M +M +1
X
(5.16)
Us =
AUik Us s + U0 ;
k =1
b

ei

ek

se
i

94

Chapter 5 Applications of the quasi-simultaneous method

for i 2 [2; Mb ], and as expected matrix AU has the following properties away from the
leading and trailing edge:
s

AUii > 0;
AUki  0; (i 6= k).

(5.17)
(5.18)

Us
ik

Due to the Kutta condition, and its induced strong coupling between the upper and
lower surface near the trailing edge (indicated by large matrix entries), AU is not weakly
diagonally dominant. It follows that assumption 4.1, which assumes that the positive
stable matrix E + V = AU + V possesses non-positive o -diagonal entries, is not satis ed
and the derived theorems in chapter 4 therefore, can not be applied.
The interaction law is based on AU by setting certain o -diagonals to zero, leading to
s

IUi =
s

AUik Us s + U0
s

ek

se
i

IikU Us s + U0 :
s

ek

se
i

(5.19)

Interaction law IU maintains the characteristics described by (5.17) and (5.18) and for
I U 6= AU constructed according to construction 4.4
M +M +1
X
IikU  0; with inequality for at least one i,
(5.20)
k =1
s

which implies that interaction matrix I U is an M-matrix (theorem D.10). On an equidistant grid matrix I U is symmetric and with (5.17) and (5.18) therefore positive de nite (theorem D.12). Interaction law IU is therefore suitably constructed for the quasisimultaneous interaction scheme. A detailed description is also found in [14].
s

5.7 Dirichlet wing interaction law


Like for the two-dimensional aerofoil problem, an interaction law, using the Dirichlet
method, is derived from the edge velocity equations derived in section 3.4.4.
In equations (3.72), (3.73) and (3.74) the discretised equations for the velocity components Ue , Ve and We in the global Cartesian coordinate system have been given
M +2M +1 M +1
X
X
U
Ue =
qe [AUijkl x + Bijkl
y ] + U0 ;
(5.21)
k =1
l=1
M +2M +1 M +1
X
X
V
Ve =
qe [AVijkl x + Bijkl
y ] + V0 ;
(5.22)
k =1
l=1
M +2M +1 M +1
X
X

W

We =
qe [AW
;
(5.23)
ijkl x + Bijkl y ] + W0
k =1
l=1
b

ij

ij

kj

kj

eij

kj

kj

kj

eij

kj

kj

kj

ij

kj

eij

for i 2 [2; Mb ] and j 2 [2; Ms ]. From these equations three approximations are to be
derived for the interaction law formulae for the velocity components.

5.7 Dirichlet wing interaction law

95

Similar to the three-dimensional dented plate case, the in uence on the velocity of
matrices B U  AV , AW and B W is small and they are therefore to be omitted. The
remaining matrices AU and B V have the following properties away from the leading and
trailing edge:

AUijkl

V
AUijij ; Bijij
> 0;
V
V
AUklij ; Bijkl
 Bklij
;  0; (i 6= k, j 6= l):

(5.24)
(5.25)

Due to the Kutta condition, and its induced strong coupling between the upper and lower
surface near the trailing edge, matrices AU and B V are not weakly diagonally dominant.
It follows that assumption 4.1 is not satis ed and the derived theorems in chapter 4 can
not be applied.
Interaction law matrices are constructed from AU and B V by setting certain o diagonals to zero. As was the case for the three-dimensional plate in uence matrices,
U
the entries AUijkl for l 6= j and the entries for Bijkl
for k 6= i are small and are omitted,
leading to the following Dirichlet interaction law matrices:

IUij =

AUijkj qe x + U0
kj

V
ij

V
ijil

kj

qe y + V0
il

il

eij

eij

(5.26)

V
Iijil
qe y + V0 ;

(5.27)

kj

U
Iijkj
qe x + U0 ;

il

eij

kj

il

eij

IW
= W0 :
ij

(5.28)

eij

The interaction law formulae have similar properties as described by (5.24) and (5.25)
and for I U 6= AU and I V 6= B V constructed according to construction 4.4
M +2M +1 M +1
M +2M +1 M +1
X
X
X
X
U
V
Iijkl ;
Iijkl
 0; with inequality for at least one ij;
k =1
l=1
k =1
l=1
implying that I U and I V are M-matrices (theorem D.10). On an equidistant grid the Mmatrices I U and I V are symmetric and thus positive de nite (theorem D.12). Interaction
laws IU and IV are therefore suitable for the quasi-simultaneous interaction scheme.
b

5.7.1 Lumping Dirichlet wing interaction law


For the three-dimensional Dirichlet interaction law it is possible to improve the diagonal
dominance of the interaction law matrix via lumping.
U
As has been discussed, the matrix entries AUijkl for l 6= j and the entries Bijkl
for k 6= i
are small and have previously been omitted for the construction of the interaction law
formulae (5.26) and (5.27). However, via lumping they can have a positive in uence on
the interaction law matrix. The entries of the lumped interaction law matrix I U can be
constructed as follows:
L

U
Iijkj
= AUijkj + (AUijkj 1 + AUijkj +1)=2;
(5.29)
entries AUijkj 1 and AUijkj +1 having the same sign as AUijkj . In a similar way a lumped
interaction law matrix for Ve can be constructed.
L

96

Chapter 5 Applications of the quasi-simultaneous method

It is seen that with (5.29) a stronger diagonal is obtained for the interaction law
U
U
matrices, as Iijij
> Iijij
, which is most suitable for the Newton iteration process in order
to be able to handle more separation. Lumping has been applied for the construction of
the Dirichlet wing interaction law, however, it is noted that with the use of relaxation a
similar e ect can be obtained.
L

5.8 Interaction at the trailing edge


At the trailing edge of an aerofoil or wing thus far no interaction law has been constructed.
The most straightforward approach to determine an approximation for the upper and
lower edge velocities there is by di erentiation of the doublet strengths, as was done in
(3.69) and (3.75), using the known doublet strengths of the upper and lower trailing edge
panels, respectively, and the wake. However, it is known from anti-symmetric thin-aerofoil
theory, corresponding with the Kutta condition, that the analytical expression for the
velocity degenerates at the trailing edge [101, 86], leading to quite complicated interaction
law formula for the quasi-simultaneous interaction method. Two other approaches are
therefore suggested here, to be able to deal with the trailing edge.
One possibility is to determine the values at the trailing edge via extrapolation from
upstream values. For the present method this is done by constructing the X -velocity in
the trailing edge points i = 1 and i = Mb + 1 for every section j as follows:

Ue 1 Ue 2
(X Xt 1 );
(5.30)
Xt 1 Xt 2 t
where index t indicates the trailing edge point. Similarly, the boundary-layer variables
and the velocity components in the Y - and Z -direction can be determined.
The other possibility is to apply a simpler way of coupling in the trailing edge points,
which can be done with the use of the interaction law matrix entries for the upstream
wing body points. This should be a more accurate approach than extrapolation.
For the calculations to be presented in chapter 6 the rst approach has been followed
as the latter is still to be implemented.
U e = Ue
t

5.9 Wake interaction law


In the wake region of a two-dimensional aerofoil or a three-dimensional wing the Dirichlet
panel method can not be applied to determine the interaction law formulae for the edge
velocities. The Dirichlet approach requires a body with a thickness, whereas the wake
geometry is modelled as just a line without thickness.
The edge velocity equations in the wake are therefore to be calculated via another
approach. The potential is to be determined in the panel centres and via di erencing
the edge velocities in the node points can be obtained. As the location of the boundarylayer unknowns and the velocities are taken in the panel endpoints, the constructed wake
in uence matrix will have a structure suitable for the quasi-simultaneous method as is
next to be discussed.

5.10 Inclusion compressibility into interaction law

97

For two-dimensional ow over a at plate, which corresponds with a two-dimensional


wake ow for a non-inclined aerofoil, the perturbation potential , due to the presence of
the boundary layer, in point xi 12 is given by

i

M
1X
=
 k=1
w

1
2

x
Zk+1

xk

dx
ln jxi
d k+ 12

1
2

 j d;

(5.31)

and similarly the perturbation potential at i+ 21 is to be found. The interaction law for
the velocity at point xi , by taking the gradient of  in point xi , becomes

I =
w
i

i+ 21

i
x

1
2

+ u0

ei

Mw
X
k

=1

Iikw x + u0 ;
k

(5.32)

ei

where index w implies wake and


3 ln 3
5 ln 5 9 ln 3
7 ln 7 15 ln 5 + 9 ln 3
Iiiw =
; Iiiw 1 = Iiiw+1 =
; Iiiw 2 = Iiiw+2 =
:
 x
2 x
2 x
It is seen that on an equidistant grid the interaction law matrix I w has the same characteristics as matrix I A in (4.50). Matrix I w is a symmetric positive de nite M-matrix
(theorems D.10 and D.12), suitable for the quasi-simultaneous interaction method. However, a small di erence is that the diagonal entries of I w are slightly smaller than those
of I A .
The procedure described by (5.32) is to be followed to determine a suitable interaction
law for the edge velocities in the wing and aerofoil wake. It is noted that the wake
interaction law for the wing is to be further simpli ed, similarly as was done for the
Dirichlet wing interaction law.
Via rotation the geometry of the wake centre line, aligned with the freestream or
otherwise located, is included in the wake interaction law.

5.10 Inclusion compressibility into interaction law


Thus far, only incompressible ow problems have been discussed for which interaction laws
have been derived from potential- ow theory. However, for the quasi-three-dimensional
swept tapered wing case the ow is assumed to be transonic (0:8 < M < 1:2, say, with
M being the freestream Mach number).
For subsonic ow (M < 0:8) the potential- ow representations are still valid if a
compressibility correction term such as the Prandtl-Glauert rule is used [50]

Uscmpr =
e

U incmpr
:
1 M2

p se

An incompressible interaction law can thus be transformed into a subsonic interaction law

cmpr

Use
i

= U0cmpr +
ei

incmpr
se

IikU
 = U cmpr + X I U
p

0
ik
1 M2 x
k

cmpr
se

ei

x ;
k

(5.33)

98

Chapter 5 Applications of the quasi-simultaneous method

where I U
is an incompressible interaction law matrix as derived before.
The above interaction law matrix, with I U
being based on thin-aerofoil theory,
has been shown to work successfully for the calculation of transonic ows in the paper by
Veldman et al. [101].
It would be more computationally ecient to use an interaction law that takes account
of the local ow conditions, including the local Mach number of the external ow. For
transonic ow it is not easy to nd a more accurate interaction law based on a linear
representation for the external ow from the transonic small perturbation (TSP) equation
[50], not least because the TSP equation is essentially nonlinear. However, the function
of the interaction law is not to de ne the solution, but to avoid breakdown at separation.
incmpr
se

incmpr
se

5.11 Numerical evaluation aerofoil interaction law


The two-dimensional Dirichlet interaction law IU given by (5.19) is to be tested in terms
of basic convergence and convergence rate for various constructions. This is done for the
two-sided Dirichlet interaction laws IUq 2 with the number of upper and lower o -diagonals
q equal to 0, 1, 20 and 146, for which I U = E = AU . The one-sided interaction laws
s

0.0

1.0

5.0

I0 s

I1 s

I20s

s
I146

29

14

159

97

12

1.0

83

30

344

19

274

10.0

1.0

150

54

548

32

11.0

1.0

182

62

608

12.0

1.0

233

96

13.0

1.0

344

126

13.5
13.5

1.0
0.7

534
762

226
424

13.6
13.6

1.0
0.7

638
911

286
505

13.7
13.7

1.0
0.7

825
1207

395
1041

13.8
13.8
13.8
13.8

1.0
0.7
0.4
0.1

=E

I1 s

I20s

162

27

22

342

76

73

101

497

39

590

137

129

183

40

607

48

677

165

155

220

825

50

745

63

815

211

197

281

1065

82

1023

95

1994

311

290

409

440

3008

486
715

456
686

617
881

552

3684

585
859

551
827

727
1038

1707
2160
2102
2453

151
422
202
530

1588
2208
1863
2582

36

4208

837

3565

936

5831

1166

1139

924
1319

2836

25

I20

Table 5.1: Viscous-inviscid and boundary-layer iterations with diagonal viscous-inviscid


relaxation.

5.11 Numerical evaluation aerofoil interaction law

99

IUq 1 are used with the number of lower o -diagonals equivalent to q = 1 and 20. For
comparison also the diagonal thin-aerofoil interaction law I20 for a non-equidistant grid is
tested.
The calculations have been performed for the (symmetric) NACA0012 aerofoil section.
The freestream Reynolds number is set to Re = 3:9  106 and the number of points used
for the discretisation is 121 on the aerofoil and 25 in the wake. Transition to turbulent
ow is tripped at X=C = 0:0006 and 0:65 on the upper and lower surfaces, respectively,
except for the incidence = 0. For = 0 the transition locations are determined with
the natural transition formula given in section 2.8.2. Each calculation is started from
scratch initialising each individual point during the rst iteration with the solution of the
previous upstream point. The results are presented in table 5.1, where indicates the
incidence and ! is the applied diagonal viscous-inviscid relaxation.
As discussed in section 5.6 the external ow matrix E = AU does have some positive
o -diagonal entries and does not satisfy assumption 4.1, which is all due to the Kutta
condition. Although the requirements for the theorems derived in chapter 4 therefore are
not ful lled, the numerical results presented here follow the theorems closely.
In terms of basic convergence it is seen that the scheme using interaction law IU1462 = E
is the least robust. This is consistent with theorem 4.7, which implies that by increasing
the number of lower and/or upper o -diagonals of I , the robustness of the iterative scheme
I + V decreases. Consequently the diagonal interaction laws, in which the diagonal of I
can be increased to make the matrix strongly diagonally dominant, should be the most
robust which can also been seen in the table.
In terms of convergence rate it is clear, when comparing the one-sided interaction laws
IU1 1 , IU20 1 and IU0 2 , that for an increasing number of lower o -diagonals the convergence
improves. This also follows from theorem 4.5. A similar behaviour is seen for the two-sided
interaction laws when comparing IU1 2 and IU20 2 .
The diagonal Dirichlet interaction law IU0 2 and the diagonal thin-aerofoil interaction
law I20 both lead to an equally robust scheme. However, in terms of convergence rate it
is seen that IU0 2 requires fewer iterations. Physically this is explained by the fact that
IU0 2 gives a more accurate description of the inviscid ow, taking into account the Kutta
condition which is not present in the thin-aerofoil formulation. Mathematically, as it can
be shown that I0U 2 < I02 , theorem 4.5 states that interaction law IU0 2 leads to faster
convergence.
From the table it is clear, however, that the di erence between the two diagonal interaction laws I0U 2 and I02 is very small. The thin-aerofoil interaction law I20 is determined
by a simple formula. The calculation of Dirichlet interaction law IU0 2 on the other hand
requires a more expensive calculation to take into account the Kutta condition.
s

5.11.1 Conclusion
From the numerical evaluation of the Dirichlet aerofoil interaction law for the NACA0012
aerofoil calculations two important conclusions can be drawn.
Firstly, although the requirements for the theorems described in chapter 4 are not
ful lled, due to the presence of the Kutta condition in the external ow matrix, the results
obtained with the Dirichlet aerofoil interaction law are very good and in correspondence

100

Chapter 5 Applications of the quasi-simultaneous method

with the theory. It is noted that these calculations are started from scratch for every
incidence.
The other conclusion follows from the use of the thin-aerofoil interaction law for aerofoil
calculations. It was seen that the thin-aerofoil interaction law is equally robust and
almost equally fast as the diagonal Dirichlet aerofoil interaction law. Furthermore, the
construction of the thin-aerofoil interaction law is very simple, whereas the Dirichlet
aerofoil interaction law requires an expensive calculation to include the Kutta condition.
It can therefore be concluded that a simple thin-aerofoil interaction law is most suitable
for the modelling of aerofoil ow.

Chapter 6
Results
6.1 Introduction
The two- and (quasi-)three-dimensional viscous potential methods described in the earlier
chapters, using the quasi-simultaneous viscous-inviscid interaction technique, have been
evaluated for several test cases. The results are presented in the following sections, and
to demonstrate the accuracy of the methods, comparisons are made with experimental
data and other computational results.

6.2 Two-dimensional aerofoil results


Results obtained with the two-dimensional viscous Dirichlet method are presented for the
NACA0012 and NACA4412 aerofoils for which the coordinates can be found in Abbott
and Von Doenhof [1].
For both test cases the calculations including separation are performed from scratch
using during the rst iteration for each point the solution of a previously calculated
upstream point, as discussed in section 5.3.1. For the approximation of the external ow
required for the quasi-simultaneous coupling scheme, the diagonal Dirichlet interaction
law has been taken. For the discretisation 121 points are used for the aerofoil's surface
and 25 points for the wake.
The geometry of the wake dividing streamline is modelled as a straight line, being
half the length of the chord and making an angle  with the X -axis. For an aerofoil at
an angle of attack below 6o , the position of the wake dividing streamline is similar to
the freestream, i.e.  = . For an angle of attack above 6o ,  is set to half the trailing
edge angle of the aerofoil, which for both test cases is equivalent to 6o . In this way the
geometry of the wake dividing streamline is more or less equivalent with the direction in
which the ow leaves the trailing edge, here along the lower surface.
The boundary-layer equations are discretised using the second-order accurate Kellerline central di erencing scheme as described in section 2.7.1. The velocity distribution in
the two trailing edge points is adjusted by setting both values to the minimum of the two
as discussed in section 3.7.

102

Chapter 6 Results

6.2.1 NACA0012 aerofoil


The symmetric NACA0012 is a favourite wing section, despite the fact that it has a fairly
abrupt stall behaviour. The calculations for the NACA0012 are performed at a Reynolds
number of 9  106 .
In gure 6.1 the pressure distribution at an incidence of 6o is shown, compared with
computational results obtained by Verho et al. [102] who used an iterative boundarylayer method coupled with a Smith and Hess panel method [12]. The results are nearly
identical except near the trailing edge, which is possibly due to a di erent trailing edge
treatment. The skin-friction results in gure 6.2 are also in good comparison. Transition
is tripped at X=C = 0:043 on the upper surface and X=C = 0:77 on the lower surface.

-3

0.016
2D VII RESULTS
RESULTS VERHOFF ET AL

-2.5
-2

0.012
0.01
Cf_s

Cp

-1.5
-1

0.008

-0.5

0.006

0.004

0.5

0.002

0
0

0.2

0.4

0.6

0.8
X/C

1.2

1.4

0.2

0.4

0.6

0.8

X/C

Figure 6.1: Comparison pressure distribution


results with calculated results of Verho et al.
[102] at Re = 9  106 and = 6o .

Figure 6.2: Comparison skin-friction results


with calculated results of Verho et al. [102]
at Re = 9  106 and = 6o .

1.8

0.04

1.6

0.035

1.4

0.03

1.2

0.025

CD

CL

2D VII RESULTS
RESULTS VERHOFF ET AL

0.014

0.8

CD_YOUNG
CD_FORCES
EXPERIMENT

0.02
0.015

0.6

0.01

CL_INVISCID
CL_KUTTA-JOUWKOWSKI
CL_FORCES
EXPERIMENT

0.4
0.2

0.005
0

-0.005
0

10 12
ALPHA

14

16

18

20

Figure 6.3: Lift coecient NACA0012 at Re =


9  106 , having used the Keller-line boundarylayer discretisation.

0.2

0.4 0.6 0.8


1
1.2
CL_KUTTA_JOUWKOWSKI

1.4

1.6

Figure 6.4: Drag coecient NACA0012 at


Re = 9  106, having used the Keller-line
boundary-layer discretisation.

6.2 Two-dimensional aerofoil results

103

1.8

0
2D VII RESULTS

1.6
-2
1.4
LOG(ERROR)

CL

1.2
1
0.8
0.6

CL_KUTTA-JOUWKOWSKI
CL_FORCES
EXPERIMENT

0.4

-4
-6
-8
-10

0.2
0

-12
0

10 12
ALPHA

14

16

18

20

Figure 6.5: Lift coecient NACA0012 at Re =


9  106, having used the upwind boundary-layer
discretisation.

100 200 300 400 500 600 700


NUMBER OF VISCOUS-INVISCID ITERATIONS

Figure 6.6: Convergence for the calculation at


Re = 9  106 and = 15o.

In the present method tripped transition occurs at a speci c point causing the abrupt
change over in the skin-friction coecient, whereas the results by Verho et al. show a
more gradual behaviour. At higher angles of attack transition on the upper surface was
assumed to correspond to laminar separation.
Lift and drag results for the NACA0012 aerofoil are compared with experimental data
from Abbott and Von Doenhof [1]. Figure 6.3 shows the lift coecient compared with
experimental and inviscid ow results. For this case the best CL prediction is given by
Kutta-Joukowski's formula (3.87), whereas the CL values calculated via integration of
the pressure and skin-friction coecients around the aerofoil (3.82) are underpredicted.
Maximum lift is obtained at an incidence of 15:5o in correspondence with experiment.
Flow separation starts to occur at an incidence of 13o .
The drag prediction results are plotted in gure 6.4. The results obtained via integrating Cp and Cf are not very accurate. The drag results obtained with Young's formula
(3.88) on the other hand t the experimental data very well. In the low lift region small
discrepancies are visible which are probably due to a di erent transition location.
In gure 6.5 the lift coecient results are shown, when using the upwind boundarylayer discretisation. It is seen that for high angles of attack the obtained results are less
accurate. Separation is predicted earlier and maximum lift is obtained at = 14o .
The number of viscous-inviscid iterations versus the error decay for the calculation of
the NACA0012 aerofoil at an incidence of 15o is displayed in gure 6.6. The convergence is
regular and would decrease nicely to machine accuracy. For this case separation occurred
at approximately 65 percent chord.
s

6.2.2 NACA4412 aerofoil


The NACA4412 is an ideal aerofoil to test the prediction of trailing edge separation as it
has a smooth stall behaviour for high Reynolds number ow.

104

Chapter 6 Results

Experimental data of Hastings and Williams [41], obtained in the RAE low-speed wind
tunnel, is used for the validation of the two-dimensional viscous Dirichlet method. The
experimental measurements were carried out at a Reynolds number of 4:17  106 and
a Mach number of 0:18. In order to have no laminar separation bubble at the leading
edge, transition was tripped at X=C = 0:014 and 0:113 on the upper and lower surfaces,
respectively. Most measurements were done at an incidence of 12:15o, at which nearly
maximum lift is obtained. The calculations are performed in free-air conditions so the
experimental results have to be corrected for the e ect of tunnel constraint. At the
experimental angle of attack of 12:15o, the corrected angle of attack corresponds with
12:49o.

1.6

-5

1.4

-4

1.2

-3

-2

0.8

-1

CL_INVISCID
CL_KUTTA_JOUWKOWSKI
CL_FORCES
EXPERIMENT

0.6

0.4

1
0

8
10
ALPHA

12

14

16

Figure 6.7: Lift coecient NACA4412 at


Re = 4:17  106 , having used the Keller-line
boundary-layer discretisation.

0.2

0.4

0.6

0.8
X/C

1.2

1.4

Figure 6.8: Comparison pressure distribution


results with experiment at Re = 4:17  106 and
= 12:5o.

0.08

11
2D VII RESULTS
EXPERIMENT

0.07

2D VII RESULTS
EXPERIMENT

10
9

0.06
8
0.05

7
H

Streamwise displacement thickness

2D VII RESULTS
EXPERIMENT

Cp

-6

CL

1.8

0.04

6
5

0.03

4
0.02
3
0.01

1
0

0.2

0.4

0.6

0.8
X/C

1.2

1.4

Figure 6.9: Comparison streamwise displacement thickness results with experiment at Re


= 4:17  106 and = 12:5o.

0.2

0.4

0.6

0.8
X/C

1.2

1.4

Figure 6.10: Comparison shape factor results


with experiment at Re = 4:17  106 and =
12:5o .

105

0.009

0.035
2D VII RESULTS
EXPERIMENT

0.008

2D VII RESULTS
EXPERIMENT

0.03

0.007

0.025

0.006

0.02

0.005

Cf_s

Streamwise momentum thickness

6.2 Two-dimensional aerofoil results

0.004

0.015
0.01

0.003

0.005

0.002
0.001

-0.005
0

0.2

0.4

0.6

0.8

X/C

0.2

0.4

0.6

0.8

X/C

Figure 6.11: Comparison streamwise momentum thickness results with experiment at Re =


4:17  106 and = 12:5o.

Figure 6.12: Comparison skin-friction results


with experiment at Re = 4:17  106 and =
12:5o .

The calculations for the NACA4412 wing section are performed for Re = 4:17  106 at
several angles of attack. In gure 6.7 the lift coecient is compared with experimental and
inviscid results. With the use of Kutta-Joukowski's formula the lift coecient values are
overpredicted, especially near stall. The lift coecient results calculated by integrating the
pressure and skin-friction distribution around the aerofoil are in much better agreement.
It is noted that for the NACA0012 calculations, Kutta-Joukowski's formula gave results
closer to experiment.
Compared to experiment maximum lift is obtained slightly earlier at = 12o , whereas
the experimental measurements found Cl = 1:439 at a corrected incidence of 12:49o.
Flow separation starts to occur at an incidence = 8o .
Transition from laminar to turbulent ow is tripped at X=C = 0:110 for the lower
boundary layer and X=C = 0:0035 on the lower surface for the upper boundary layer,
respectivily. The latter trip location is di erent from experiment, but is required to
avoid laminar boundary-layer separation. In the experiment wire transition strips are
used which are not implicitly contained in the theoretical method to enable a change
in the state of the boundary layer in this critical area. To simulate the transition strip
in the present method the displacement thickness is increased at the upper boundarylayer transition location by s = 0:000243, leading to reasonable agreement with the
measured data at X=C = 0:2 and 0:4. No interaction with the external ow takes place
and the shape factor and the velocity in the upper boundary-layer transition point are
determined via averaging the two neighbouring values.
A similar approach was adopted by Williams [107] who increased the momentum
thickness by ss = 0:00018 at the upper surface transition point X=C = 0:014 to have
good comparison with experiment at X=C = 0:2 and 0:4. At transition the shape factor
obtained by Williams is approximately 1.6 and s hence corresponds with H ss =
0:000288 at X=C = 0:014. As the two-dimensional viscous Dirichlet method predicts
transition earlier, the increase in streamwise displacement thickness is required to be
max

106

Chapter 6 Results

slightly smaller, and with s = 0:000243 the measured and computed results correspond
well. On the lower surface no special treatment is applied as the computed s and H
compare well with experiment at X=C = 1.
In gure 6.8 the pressure distribution is compared with experimental results at an
incidence of 12:5o . The results are in good agreement; only in the trailing edge region,
where there is separation, the results di er. The displacement thickness is shown in gure
6.9 and as is seen the results compare well with experiment.
The shape factor results are shown in gure 6.10. In the region of separation, i.e.
H > 2:73, the shape factor is slightly overpredicted. The experimental measurements
found H = 8:7 at the trailing edge, whereas here H = 10:5 is calculated. These incorrect
shape factor predictions are probably caused by the empirical closure relation for H1 . The
small peak seen in H near the leading edge corresponds with the stagnation point where
the values have been kept constant.
The incorrect shape factor results also cause the discrepancies for the momentum
thickness results near the trailing edge as shown in gure 6.11. Separation took place
at X=C = 0:75 which is displayed in gure 6.12 where at X=C = 0:75 the skin-friction
coecient becomes zero, corresponding with the measurements.

6.3 Swept tapered wing results


The DERA Viscous Full-Potential (VFP) code [85] is based on a viscous-coupled method
for computing the compressible ow over wing-body combinations. It combines a fullpotential- ow solver, brie y mentioned in section 3.2.4, and an integral boundary-layer
method, as described in section 2.5. The boundary-layer equations are derived in a local
coordinate system for each chordwise strip on the wing. This derivation assumes that
each section of the wing behaves as if it were part of a larger simply-tapered wing with
the same leading and trailing edge sweeps as the local section. Ashill and Smith [5] show
that results obtained with VFP for conventional wings are similar to those obtained using
a fully three-dimensional version of the boundary-layer equations.
In the original code the interaction between the external inviscid and the viscous ow
regions is governed by the boundary-layer equations through the semi-inverse method.
The aim of the contract work carried out as part of this thesis, was to replace the semiinverse method with a quasi-simultaneous coupling method to improve performance in
terms of convergence rate.
A four-stage Runge-Kutta method is used in the original VFP program to solve the
set of boundary-layer equations. In combination with a quasi-simultaneous scheme the
Runge-Kutta method is not suitable, as it requires the calculation of the extra interaction
law equation, which is to be solved together with the boundary-layer equations, at each
stage of the four-stage Runge-Kutta scheme, i.e. at each of the four stages between grid
points. To avoid the calculation of an interaction law at all the intermediate stages
between grid points, a Newton iterative method is to be applied in combination with the
quasi-simultaneous interaction technique.
Newton's method however, has the disadvantage that the scheme is rst-order accurate, using upwind di erencing, whereas the four-stage Runge-Kutta scheme is fourth-

6.3 Swept tapered wing results

107

eta=0.13191, alpha=1.0, Re=13.3x10^6, M=0.78

eta=0.13191, alpha=1.0, Re=13.3x10^6, M=0.78

-1.4

0.003
QS-coupling
SI-coupling

-1.2

QS-coupling
SI-coupling

-1

0.0025

-0.8
-0.6

0.002

theta

Cp

-0.4
-0.2

0.0015

0
0.2

0.001

0.4
0.6

0.0005

0.8
1

0
0

0.5

1.5

0.5

Figure 6.13: Pressure distribution at  =


0:132, = 1o , Re = 13:3  106 and M = 0:78.

1
x

1.5

Figure 6.14: Momentum thickness at  =


0:132, = 1o , Re = 13:3  106 and M = 0:78.

eta=0.53440, alpha=1.0, Re=13.3x10^6, M=0.78

eta=0.53440, alpha=1.0, Re=13.3x10^6, M=0.78

-1.4

0.004
QS-coupling
SI-coupling

-1.2

QS-coupling
SI-coupling
0.0035

-1
-0.8

0.003

-0.6
0.0025

theta

Cp

-0.4
-0.2

0.002

0
0.0015
0.2
0.4

0.001

0.6
0.0005
0.8
1

0
0

0.5

1.5

0.5

Figure 6.15: Pressure distribution at  =


0:534, = 1o , Re = 13:3  106 and M = 0:78.

1
x

1.5

Figure 6.16: Momentum thickness at  =


0:534, = 1o , Re = 13:3  106 and M = 0:78.

eta=0.98754, alpha=1.0, Re=13.3x10^6, M=0.78

eta=0.98754, alpha=1.0, Re=13.3x10^6, M=0.78

-1.5

0.0045
QS-coupling
SI-coupling

QS-coupling
SI-coupling

0.004

-1

0.0035
0.003

Cp

theta

-0.5
0.0025
0.002

0
0.0015
0.001

0.5

0.0005
1

0
0

0.5

1.5

Figure 6.17: Pressure distribution at  =


0:988, = 1o , Re = 13:3  106 and M = 0:78.

0.5

1
x

1.5

Figure 6.18: Momentum thickness at  =


0:988, = 1o , Re = 13:3  106 and M = 0:78.

108

Chapter 6 Results

order accurate. This consequently leads to a loss of accuracy.


The quasi-simultaneous calculations performed for the W4 wing case use a tridiagonal
subsonic plate interaction law to represent the external inviscid ow as derived in section
5.10 with I U
based on thin-aerofoil theory. Other choices of interaction law can
lead to a more robust system, but have not been tested. The number of boundary-layer
iterations is set to three.
For the present VFP calculations, transition is tripped on the upper and lower surface
at 15 percent chord. For the discretisation on the ne grid 162 points are used around the
wing section and 19 in the wake. In the spanwise direction 30 points are taken. The quasisimultaneous computations are started from a previously calculated solution, a scheme
organising a start from scratch not being implemented. The external velocity distribution
in the last four trailing edge points is modi ed before the boundary-layer calculations are
started and are determined via extrapolation from upstream values.
The pressure distribution and streamwise momentum thickness results at the root,
semi-span and near the tip for the swept tapered W4 wing are shown in gures 6.13, 6.14,
6.15, 6.16, 6.17 and 6.18. The results have been computed for an incidence of = 1o and
Re = 13:3  106 and M = 0:78.
Comparing solutions from the semi-inverse and quasi-simultaneous versions of the VFP
code, the agreement is generally good, particularly in view of the quite di erent integration
procedures being used to compute the boundary layer in each case. In particular the shock
is predicted by the quasi-simultaneous method to be at the same location as predicted
by the semi-inverse method. The solution of the quasi-simultaneous scheme di ers only
slightly at the trailing edge and in the wake. In gures 6.19 and 6.20 the spanwise lift and
drag distributions are shown, being in good comparison with the semi-inverse results.
The calculations for both the semi-inverse scheme and the quasi-simultaneous scheme
took similar amounts of computing time, with the number of viscous-inviscid iterations
being xed. However, little experimentation has been carried out with the numerical
parameters of the quasi-simultaneous scheme. With the use of the theoretical results presented in chapter 4, which have been derived after the completion of the VFP calculations,
the run times can most likely be reduced. For more details see also [15, 13].
incmpr
se

alpha=1.0, Re=13.3x10^6, M=0.78

alpha=1.0, Re=13.3x10^6, M=0.78

0.7

0.07
QS-coupling
SI-coupling

QS-coupling
SI-coupling

0.06

0.65

0.05
0.6
0.04
0.03

Cd

Cl

0.55
0.5

0.02
0.01

0.45

0
0.4
-0.01
0.35
0.3
0.1

-0.02

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

y/s

Figure 6.19: Span-wise lift distribution.

-0.03
0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

y/s

Figure 6.20: Span-wise drag distribution.

6.4 Three-dimensional dented plate results

109

6.4 Three-dimensional dented plate results


The three-dimensional viscous potential- ow method, combining the turbulent integral
boundary-layer equations from section 2.6 and the potential- ow method as described in
section 3.4.3, has been tested for a at plate with several forms of protuberance. Here
the studies of at plate ow, three-dimensional trough ow and three-dimensional dent
ow are presented. The leading edge of the plate is at x = 1, and the total computational
domain is 1 < x < 4 and 1:5 < y < 1:5. The grid is assumed to be uniform, with 41
points in the x-direction and 21 in the y -direction. The oncoming ow is aligned with the
x-axis and the freestream is hence taken to be u1 = 1 and v1 = 0.
At the leading edge of the plate x = 1 boundary conditions are prescribed. At the
trailing edge x = 4 the values are determined by extrapolation in the x-direction. The
values of the variables at the two side edges, y = 1:5 and y = 1:5, are copied from the
values of the gridlines in the domain lying next and parallel to these two side edges.
The interaction law used for the quasi-simultaneous coupling scheme has been derived
in section 5.5. For the present calculations two-sided interaction law matrices for the
u- and v -velocity are taken with the number of upper and lower o -diagonals set to
one. From the theory presented in chapter 4, obtained after the completion of the threedimensional dented plate computations, it follows that a diagonal interaction law most
likely would improve the convergence and lead to a more robust system. As an initial guess
for ow cases with separation the solution of a previous calculation is used, initialisation
option three given in section 5.3.1 not being implemented. The number of boundary-layer
iterations is xed at three. The computations are performed on a 300 MHz Pentium II
and take under a minute per complete solution.

6.4.1 Flat plate


The accuracy of the three-dimensional viscous potential- ow method is tested by assuming
simple two-dimensional ow over a three-dimensional at plate at Re = 11:5  106 .
1.38

Streamwise displacement thickness

1.36
Shape factor H

0.005

gridline y=-0.45
gridline y= 0.00
gridline y= 0.45
2D result

1.34
1.32
1.3
1.28
1.26

0.0045
0.004
0.0035
0.003
0.0025
0.002
0.0015
gridline y=-0.45
gridline y= 0.00
gridline y= 0.45
2D result

0.001
0.0005
0

1.5

2.5
x

3.5

Figure 6.21: Shape factor for ow over a at


plate at Re = 11:5  106 .

1.5

2.5
x

3.5

Figure 6.22: Displacement thickness for ow


over a at plate at Re = 11:5  106 .

110

Chapter 6 Results

The results are shown in gures 6.21 and 6.22, displaying the shape factor and the
streamwise displacement thickness. The displacement thickness and the shape factor,
plotted along the gridlines y = 0:45 and y = 0:0, all give the same result, which is
to be expected as it is a two-dimensional ow case. The results of a two-dimensional
quasi-simultaneous viscous-inviscid interaction method are shown for comparison. The
two-dimensional coupled results compare well with the results from the three-dimensional
viscous potential- ow program. Only a slight di erence is visible in the results for the
shape factor, due to the di erent boundary conditions used at x = 1.

6.4.2 Three-dimensional trough ow


Next, the three-dimensional viscous potential- ow program has been tested for twodimensional ow over a three-dimensional trough with a depth = 0:1

;
(6.1)
zd =
cosh(4x 10)
as shown in gure 6.23.
The Reynolds number is again taken to be 11:5  106 . In gure 6.24 the edge velocity
distribution in the x-direction is displayed. The solutions for gridlines y = 0:0 and
y = 0:45 are again seen to be similar for this two-dimensional ow case. Also, they are in
good comparison with the results from the two-dimensional viscous-inviscid interaction
method. There is, however, a di erence visible for the points in the middle of the trough,
around x = 2:5, where the three-dimensional viscous potential- ow method predicts a
lower velocity distribution. This is due to the fact that in the external inviscid ow
modelling the at plate is not assumed to be in nitely long and wide. The e ect of this
is also visible in gure 6.25 in which the streamwise displacement thickness is presented.
The results compare well until just before x = 2:5, but again, in the middle of the trough,
the three-dimensional code underpredicts the displacement thickness and this e ect is
carried downstream.
1.1
Channel geometry

1.05

0
-0.01
-0.02
-0.03
-0.04
-0.05
-0.06
-0.07
-0.08
-0.09
-0.1

1.5

2
x

2.5

3.5

-1
4 -1.5

0
-0.5

0.5

1.5

Velocity x-direction

z
1
0.95
0.9
0.85
0.8
gridline y= 0.00
gridline y= 0.45
2D result

0.75
0.7
1

Figure 6.23: Geometry of a three-dimensional


trough with depth 0:1.

1.5

2.5
x

3.5

Figure 6.24: Velocity distribution in the xdirection for ow over a three-dimensional


trough at Re = 11:5  106 .

6.4 Three-dimensional dented plate results

Streamwise displacement thickness

0.014

111

gridline y= 0.00
gridline y= 0.45
2D result

0.012

z
0.01

0
-0.02
-0.04
-0.06
-0.08
-0.1
-0.12
-0.14
-0.16
-0.18

0.008
0.006
0.004

0.002

1.5

0
1

1.5

2.5
x

3.5

2.5

3.5

-1
4 -1.5

0
-0.5

1.5

Figure 6.25: Streamwise displacement thickness for ow over a three-dimensional trough


at Re = 11:5  106 .

Figure 6.26: Geometry of a three-dimensional


dent with depth 0:174.

1.1

0.0035
Streamwise skin friction coefficient

1.05
1
Velocity x-direction

0.5

0.95
0.9
0.85
0.8
0.75
gridline y= 0.00
gridline y= 0.30
gridline y= 1.35

0.7
0.65
1

1.5

2.5
x

3.5

0.003
0.0025
0.002
0.0015
0.001
0.0005
gridline y= 0.00
gridline y= 0.30
gridline y= 1.35

0
-0.0005

Figure 6.27: Velocity distribution in the xdirection for ow over a dent at Re = 11:5 
106 .

1.5

2.5
x

3.5

Figure 6.28: Streamwise skin-friction coecient for ow over a dent at Re = 11:5  106 .

6.4.3 Three-dimensional dent ow


The full three-dimensional capability of the viscous potential method has been tested for
three-dimensional ow over a dent with depth = 0:174

zd =
;
(6.2)
cosh(4x 10) cosh(4y )
displayed in gure 6.26.
In gure 6.27 the velocity distribution in x-direction is shown. Results for three
di erent gridlines are plotted: one going through the middle of the dent, y = 0:0, one
slightly away from the middle, y = 0:3, and one near the edge of the plate, y = 1:35,
where the e ect of the dent is nearly negligible. For gridlines y = 0:3 and y = 1:35
the velocity distribution is symmetric around x = 2:5, however, for the gridline through
the middle of the dent, this is no longer the case. The velocity distribution attens o

112

Chapter 6 Results

20

gridline y= 0.00
gridline y= 0.30
gridline y= 1.35

0.03

Limiting streamline angle (deg)

Streamwise displacement thickness

0.035

0.025
0.02
0.015
0.01
0.005
0

gridline y=-1.35
gridline y=-0.30
gridline y= 0.00
gridline y= 0.30
gridline y= 1.35

15
10
5
0
-5
-10
-15
-20

1.5

2.5
x

3.5

Figure 6.29: Streamwise displacement thickness for ow over a dent at Re = 11:5  106 .

1.5

2.5
x

3.5

Figure 6.30: Limiting wall streamline angle


for ow over a dent at Re = 11:5  106 .

before x = 2:5, implying separation and reattachment. The two-dimensional type of ow


reversal can also be clearly seen in gure 6.28 which displays the streamwise skin-friction
coecient. For gridline y = 0:0 near x = 2:5 the skin-friction coecient is below zero.
The other plotted gridlines have a skin-friction coecient well above zero.
In gure 6.29 for the same three gridlines the streamwise displacement thickness is
plotted. For gridline y = 0:0 a steep growth is seen near x = 2:5. Finally in gure 6.30
the limiting wall streamline angle is plotted in degrees. The ow along gridline y = 0:0
is not diverted. For the gridlines slightly away from the midpoint of the dent, the ow
is rst diverted outwards and then inwards again. As is seen for gridlines y = 0:3 and
y = 0:3 the results are symmetric in y = 0:0. The gridlines close to the side edges of
the plate are hardly in uenced by the dent and almost have no diversion.

6.5 Three-dimensional wing results


The three-dimensional viscous Dirichlet method, coupling the three-dimensional integral
boundary-layer equations and the wing potential- ow method as described in section
3.4.4, has been applied to several test cases and some results are presented in this section.
Quasi-two-dimensional ow is modelled over an unswept symmetric NACA0012 wing
of high aspect ratio for comparison with the two-dimensional viscous Dirichlet method. To
demonstrate the three-dimensional capabilities of the method separated ow is simulated
for the unswept NACA0012 wing and calculations are performed for a 45o swept RAE101
wing.
The geometry of the wing's wake is chosen as a at surface, making an angle  with
the X=C -axis in the (X; Z ) plane. For a wing at an angle of attack below 6o , the
position of the wake surface is set to . For an angle of attack above 6o ,  is set to half
the trailing edge angle of the wing, which for the test cases is approximately 6o .
The external velocity distribution in the upper and lower trailing edge points and in
the rst wake point are modi ed before the boundary-layer calculations are started, as

6.5 Three-dimensional wing results

0.004

113

-2

3D VII RESULTS
2D VII RESULTS

0.0035

3D VII RESULTS
2D VII RESULTS

-1.5
-1

0.0025
Cp

Theta_ss

0.003

0.002
0.0015

-0.5
0

0.001
0.5

0.0005
0

1
0

0.2

0.4

0.6

0.8

0.2

0.4

x/c

Figure 6.31: Comparison of three- and twodimensional viscous Dirichlet methods for the
streamwise momentum thickness at = 4o .

0.6

0.8

x/c

Figure 6.32: Comparison of three- and twodimensional viscous Dirichlet methods for the
pressure distribution at = 4o .

discussed in section 3.7.


The quasi-simultaneous computations are started from scratch, using during the rst
iteration for each point the solution of the previous calculated upstream point. The interaction law used for the quasi-simultaneous coupling scheme has been derived in section
5.7. Based on the ndings in chapter 4 for the present calculations diagonal interaction
law matrices are taken for the U - and V -velocity, i.e. the number of upper and lower
o -diagonals is set to zero.

6.5.1 Quasi-two-dimensional NACA0012 wing ow


For the simulation of quasi-two-dimensional ow an unswept NACA0012 wing con guration is used with an aspect ratio of 14. The freestream Reynolds number is set to 9  106
and results are obtained at an incidence = 4o . For the discretisation a rectangular
grid is used with 71 points around the wing section, parallel to the freestream direction,
and 9 points in the spanwise direction. It is noted that fewer points are used for the
discretisation of the wing section than for the two-dimensional ow case, which is done to
keep the computation time acceptable. Transition was tripped on the upper surface close
to the attachment line and on the lower surface at X=C = 0:35.
Midwing the ow has the least in uence of the tip and root e ects, and these results
are therefore shown. In gure 6.31 the streamwise momentum thickness is displayed
for the upper and lower surfaces and the results compare well with the two-dimensional
results from the two-dimensional viscous Dirichlet program. In gure 6.32 the pressure
distributions of both viscous Dirichlet methods are seen to be in good agreement. Close
to the leading edge there are, however, some small di erences. For the present calculation
10 viscous-inviscid iterations were required. Other early results can be found in [17].

114

Chapter 6 Results

3D
3D
3D
3D
3D

-6
-5

VII
VII
VII
VII
VII

Y/Ys
Y/Ys
Y/Ys
Y/Ys
Y/Ys

=
=
=
=
=

Streamwise momentum thickness

-7
0.88
1.97
3.50
5.03
6.12

Cp

-4
-3
-2
-1
0
1
0

0.2

0.4

0.6

0.8

0.008
3D
3D
3D
3D
3D

0.007
0.006
0.005

VII
VII
VII
VII
VII

Y/Ys
Y/Ys
Y/Ys
Y/Ys
Y/Ys

=
=
=
=
=

0.88
1.97
3.50
5.03
6.12

0.004
0.003
0.002
0.001
0

0.2

0.4

X/C

Figure 6.33: Comparison pressure distribution


at di erent spanwise stations for = 14o .

0.8

Figure 6.34: Comparison streamwise momentum thickness at di erent spanwise stations for
= 14o.

0.012
3D
3D
3D
3D
3D

3.5

VII
VII
VII
VII
VII

Y/Ys
Y/Ys
Y/Ys
Y/Ys
Y/Ys

=
=
=
=
=

0.88
1.97
3.50
5.03
6.12

3D
3D
3D
3D
3D

0.01
0.008

CF_s

0.6
X/C

2.5

VII
VII
VII
VII
VII

Y/Ys
Y/Ys
Y/Ys
Y/Ys
Y/Ys

=
=
=
=
=

0.88
1.97
3.50
5.03
6.12

0.006
0.004

2
0.002
1.5

-0.002
0

0.2

0.4

0.6

0.8

X/C

Figure 6.35: Comparison shape factor at di erent spanwise stations for = 14o .

0.2

0.4

0.6

0.8

X/C

Figure 6.36: Comparison streamwise skinfriction at di erent spanwise stations for =


14o .

6.5.2 Three-dimensional unswept NACA0012 wing ow


As for the quasi-two-dimensional unswept NACA0012 calculations, the three-dimensional
unswept NACA0012 calculations are performed at a Reynolds number of 9  106 . The
same rectangular grid is used and transition is tripped on the upper surface close to the
attachment line and on the lower surface at X=C = 0:35.
By increasing the incidence beyond 10o , small three-dimensional e ects become present.
Results in gures 6.33 to 6.36 are shown for an incidence of 14o for which ow case there
is a small region with (two-dimensional) separation. Relaxation is applied using ! = 0:5.
In gure 6.33 the pressure distribution is shown for ve di erent chordwise stations,
where Y=Ys = 0:88 lies close to the root, Y=Ys = 3:5 lies midwing and Y=Ys = 6:12 is
close to the tip of the wing. Away from the root the pressure distribution decreases.
Figure 6.34 shows the streamwise momentum thickness for the various gridlines. The
results near the root show a more severe growth of ss near the trailing edge than the

6.5 Three-dimensional wing results

115

results closer to the tip. In gures 6.35 and 6.36 the shape factor and the streamwise
skin-friction coecient are shown. From these graphs it is clear that the region of ow
separation near the trailing edge increases when going towards the root. It is seen that,
apart from gridline Y=Ys = 6:12, Cf < 0 and H > 2:73 in the trailing edge region.
The calculations have been performed on a 300 MHz Pentium II. For the above ow
case of separated wing ow, 90 viscous-inviscid iterations were required to converge, taking
70 minutes of CPU time.
s

6.5.3 Three-dimensional 45o swept RAE101 wing ow


To simulate fully three-dimensional wing ow, the three-dimensional viscous Dirichlet
method is applied to a swept RAE101 wing of aspect ratio 5 and with a sweep angle
 = 45o . The grid is constructed using 71 points around the wing section, parallel to

-1.5

GEOMETRY 45 DEG SWEPT WING RAE101

3D VII RESULTS
3D INVISCID RESULTS
EXPERIMENT

Z
-1
0.06
0.04
0.02
0
-0.02
-0.04
-0.06

Cp

-0.5

0
2.5
2

0.5

0.5

1.5
1

1.5
X

1
2

2.5

0.5

Y
1

3.5 0

0.2

0.4

0.6

0.8

X/C

Figure 6.37: Geometry RAE101 45o swept


wing.

Figure 6.38: Comparison inviscid, interactive


and experimental Cp at Y=Ys = 0:22 for =
6:3o .

-1.5

-1.5
3D VII RESULTS
3D INVISCID RESULTS
EXPERIMENT

-1

3D VII RESULTS
3D INVISCID RESULTS
EXPERIMENT

-1

Cp

-0.5

Cp

-0.5

0.5

0.5

1
0

0.2

0.4

0.6

0.8

X/C

Figure 6.39: Comparison inviscid, interactive


and experimental Cp at Y=Ys = 1:25 for =
6:3o .

0.2

0.4

0.6

0.8

X/C

Figure 6.40: Comparison inviscid, interactive


and experimental Cp at Y=Ys = 1:69 for =
6:3o .

116

Chapter 6 Results

the freestream direction, and 11 points in the spanwise direction, as seen in gure 6.37.
Initial results are given in [16].
Pressure measurements for this wing have been done in the 11 12 ft.  8 12 ft. wind tunnel
at the RAE in 1950 and are reported in [103]. The tests were performed at a Reynolds
number of 1:68  106 and at a wind speed low enough to treat the ow as incompressible.
In gures 6.38, 6.39 and 6.40 the measured and computed interactive and inviscid
pressure distributions are shown for an incidence of 6:3o at three di erent spanwise stations, Y=Ys = 0:22; Y=Ys = 1:25 and Y=Ys = 1:69. These computational results have
been obtained at a Reynolds number of 2:1  106 and transition is tripped close to experimentally observed locations, being X=C = 0:08 on the upper surface and X=C = 0:5
on the lower surface. Relaxation is applied using ! = 0:2. The agreement between the
viscous-inviscid interactive and experimental results is good. However, towards the tip
problems were encountered with the inviscid ow model due to inaccurate representation
of tip e ects. No coupling therefore takes place for Y=Ys > 2:2, being the last 10 percent
of the span, and instead the boundary-layer values are determined by extrapolation.
Boundary-layer velocity pro les have been measured for the same wing case at Re =
2:1  106 by Brebner and Wyatt [7] from which the boundary-layer integral thicknesses
are determined [69].
Figure 6.41 shows the midwing streamwise momentum thickness computed by the
three-dimensional viscous Dirichlet program at = 6:3o and Re = 2:1  106 . The
experimental values of the streamwise momentum thickness are shown by the squares. It is
seen that the computational results overpredict the boundary-layer growth compared with
experiment, especially near the trailing edge. Milewski achieved a similar overprediction
of the boundary-layer growth with his interactive viscous potential- ow method [69].
The discrepancies between experiment and the computational methods are in all
probability due to the inaccurate modelling of the Kutta condition in the potential- ow
method. Furthermore, it should be noted that the experimental results have not been
corrected for the e ects of tunnel constraints.
In gure 6.42 the streamwise momentum thickness at various spanwise stations is
compared, showing the steepest growth of ss near the trailing edge midwing. In gures
6.43 and 6.44 the interactive and experimental shape factor H and cross ow displacement
thickness n results are shown. As discussed before the computed results are overpredicted,
but again quite similar to Milewski's results.
From gure 6.45 it immediately follows that three-dimensional separation has occurred. It is seen that towards the trailing edge + , being the angle between the
X -axis and the limiting wall streamline, is larger than 90o  = 45o , which corresponds
to separation. Close to the stagnation point, , the angle between the X -axis and the
streamwise direction, is quite large as the streamwise direction is perpendicular to the
X -axis.
In gure 6.46 the history of the viscous-inviscid iteration process is plotted for the
three global velocity components.
The calculations have been performed on a 300 MHz Pentium II and for the above
ow case of three-dimensional wing ow 4 hours of CPU time were required to converge.

117

0.008

Streamwise momentum thickness

Streamwise momentum thickness

6.5 Three-dimensional wing results

3D VII RESULTS
EXPERIMENT
MILEWSKI RESULTS

0.007
0.006
0.005
0.004
0.003
0.002
0.001
0
0

0.2

0.4

0.6

0.8
X/C

1.2

3D VII RESULTS
EXPERIMENT
MILEWSKI RESULTS

3
2.5
2
1.5
1
0

0.2

0.4

0.6

0.8
X/C

1.2

VII
VII
VII
VII

Y/Ys
Y/Ys
Y/Ys
Y/Ys

=
=
=
=

0.46
1.25
1.69
2.28

0.005
0.004
0.003
0.002
0.001
0
0.2

0.4

0.6

0.8
X/C

1.2

1.4

Figure 6.42: Comparison ss at di erent spanwise stations for = 6:3o .

0
-0.001
-0.002
-0.003
-0.004
-0.005
-0.006
3D VII RESULTS
EXPERIMENT
MILEWSKI RESULTS

-0.007
-0.008

1.4

Figure 6.43: Comparison interactive and experimental H at Y=Ys = 1:25 for = 6:3o .

0.2

0.4

0.6

0.8
X/C

1.2

1.4

Figure 6.44: Comparison interactive and experimental n at Y=Ys = 1:25 for = 6:3o .

60

0
ANGLE BETA
ANGLE ALPHA+BETA

50

ERROR X-VELOCITY
ERROR Y-VELOCITY
ERROR Z-VELOCITY

-0.5
-1

40
LOG(ERROR)

BETA, ALPHA+BETA

0.006

Normal displacement thickness

3D
3D
3D
3D

0.007

1.4

Figure 6.41: Comparison interactive and experimental ss at Y=Ys = 1:25 for = 6:3o .

3.5

0.008

30
20

-1.5
-2
-2.5
-3

10
-3.5
0

-4

-10

-4.5
0

0.2

0.4

0.6

0.8
X/C

1.2

1.4

Figure 6.45: Comparison and + at


Y=Ys = 1:25 for = 6:3o.

20
40
60
80
100
120
NUMBER OF VISCOUS-INVISCID ITERATIONS

Figure 6.46: Error decay versus the number


of iterations.

118

Chapter 6 Results

Chapter 7
Conclusions
In this thesis the quasi-simultaneous viscous-inviscid interaction method is investigated
both mathematically and numerically for two- and three-dimensional ow cases. Moreover, the present research concentrates on gaining full understanding of the basic requirements for the quasi-simultaneous method and on successfully developing two- and
three-dimensional quasi-simultaneous viscous-inviscid interaction methods. Conclusions
on these themes are presented in the next sections.

Boundary-layer discretisation
In chapter 2 the boundary-layer formulations for the various ow cases are given. For the
three-dimensional ow case the boundary-layer equations are written in Cartesian coordinates, avoiding the cumbersome metric calculation required when using non-orthogonal
curvilinear coordinates. The three-dimensional boundary-layer equations are discretised
using an upwind nite-volume scheme (taking into account the hyperbolicity of the system of equations) which works successfully in combination with the quasi-simultaneous
method.
Quasi-simultaneous interaction
In chapter 4 it is investigated to what extent the approximation of the external ow, i.e.
the interaction law, can be simpli ed. Theorems are given from which it is clear what
the basic requirements are for the quasi-simultaneous method in order to optimise its
robustness and convergence rate.
The most important conclusions are that subject to certain conditions the construction
of an interaction law from an external ow matrix, either by setting certain outer o diagonals equal to zero and/or by increasing the diagonal, has favourable properties. The
convergence of the viscous-inviscid iterations can be guaranteed, the convergence of the
Gauss-Seidel boundary-layer iterations can be ensured and the robustness of the inner
Newton iterations can be enhanced.
Predictions on the rate of convergence lead to the following conclusions. For the
viscous-inviscid iterative method the convergence improves when the size of the interaction law increases. However, for the boundary-layer iterative method the opposite result

120

Chapter 7 Conclusions

is obtained: its convergence deteriorates when the size of the upper interaction law matrix increases. Concluding, when the external calculations are expensive, construct an
interaction law from the external ow matrix by only setting a few outer o -diagonals
to zero in order to have only a few viscous-inviscid iterations. On the other hand, when
the external calculations are not expensive, construct an interaction law from the external ow matrix by setting many outer o -diagonals to zero in order to have only a few
boundary-layer iterations. In this way a fast procedure for the overall quasi-simultaneous
scheme is obtained.
The developed theorems are evaluated for the calculation of two-dimensional trough
ow using various choices of interaction law. The numerical results are in correspondence
with the theory and it is shown that a diagonal interaction law is the most simple and
robust option. Furthermore, not requiring any boundary-layer iterations it is also one of
the fastest options.
Another aspect relevant for the construction of a suitable interaction law, is the choice
of the location for the unknowns (i.e. ue and x ), as both the magnitude and the structure
of the interaction law are a ected by it. De ning the unknowns in the endpoints of the
panels leads to an interaction law with favourable properties for the quasi-simultaneous
interaction method. Other discretisations can result in less suitable interaction laws.
In chapter 5 various interaction laws are constructed for the case of two-dimensional
aerofoil and three-dimensional plate and wing ow. For the two-dimensional aerofoil ow
in more detail various di erent interactions laws are tested to evaluate again the above
mentioned theory. It is found that although the requirements for the theorems described
in chapter 4 are not ful lled due to the presence of the Kutta condition in the external
ow matrix, the convergence behaviour obtained with the Dirichlet aerofoil interaction
law is in good correspondence with the theory. Furthermore, it is noted that besides the
diagonal Dirichlet interaction law also the diagonal thin-aerofoil theory interaction law is
very suitable for aerofoil viscous-inviscid calculations. It can therefore be concluded that
the interaction law based on thin-aerofoil theory, being the most simple to implement,
should be the one to be preferred.

Results
In chapter 6 results are presented for various ow calculations. The calculations do not
require an initial solution from a prior calculation and only small amounts of relaxation
are used for the ow cases with regions with ow reversal present.
Two symmetric aerofoil models are tested for ow cases including separation, and the
results show to be in good agreement with experimental and other computational results.
With the modi ed DERA-VFP code quasi-three-dimensional results with a transonic
quasi-simultaneous method are presented. The results are compared with results obtained
with the original semi-inverse DERA-VFP code. The results are in good agreement and
require similar amounts of computing time. Little experimentation is carried out with the
numerical parameters of the quasi-simultaneous scheme and with the use of the theoretical
results of chapter 4 the run times and robustness could be further improved.
Three-dimensional results are obtained for dented plate and wing ow calculations.
For the dented plate case a two-sided interaction law is used with the number of o -

121

diagonals set to one, whereas for the latter wing case a diagonal Dirichlet interaction law
is taken. When comparing the results obtained with the fully three-dimensional method
with results obtained with a two-dimensional method for quasi-two-dimensional ow cases
good agreement is shown.
For the swept RAE101 wing case with ow separation present, good convergence is
achieved with the three-dimensional quasi-simutaneous method. The obtained results are
compared with experimental and other computational data, and are in reasonably good
correspondence. However, improvement of the trailing edge modelling in the used external
inviscid ow solver is required, as discussed in section 3.7.

Conclusion
It is demonstrated that a diagonal interaction law derived from thin-aerofoil theory can be
used for aerofoil calculations, despite the fact that it does not include the Kutta condition
with its global in uence. Similarly, it is shown that with a diagonal interaction law based
on three-dimensional Dirichlet potential ow theory good results are obtained for wing
calculations. These observations are supported by the mathematical theory presented
in chapter 4 and it is concluded that a diagonal interaction law is sucient for fast
and robust two- and three-dimensional quasi-simultaneous viscous-inviscid interaction,
including separation. This is a noteworthy result as with the use of only a simple diagonal
interaction law, the quasi-simultaneous method is just a small modi cation of the classical
direct method, which is unable to deal with ow separation.

122

Chapter 7 Conclusions

Bibliography
[1] Abbott, I.H. and Von Doenho , A.E. Theory of Wing Sections. Dover, New York,
1959.
[2] Arnal, D., Habiballah, M. and Coustols, E. Laminar Instability Theory and Transition Criteria in Two and Three-Dimensional Flow. La Recherche Aerospatiale,
2:45{63, 1984.
[3] Arnold, F. and Thiele, F. A Laplace Interaction Law for the Computation of Viscous
Airfoil Flow in Low and High Speed Aerodynamics. AIAA 93-3462-CP, 1993.
[4] Arthur, M.T. A Method for Calculating Subsonic and Transonic Flows over Wings
or Wing-Fuselage Combinations with an Allowance for Viscous E ects. RAE TM
Aero 1996, 1984.
[5] Ashill, P.R. and Smith, P.D. An Integral Method for Calculating the E ects on
Turbulent Boundary-Layer Development of Sweep and Taper. RAE TR 83053,
1983.
[6] Ashill, P.R., Wood, R.F. and Weeks, D.J. An Improved, Semi-Inverse Version of
the Viscous Garabedian and Korn Method (VGK). RAE TR 87002, 1987.
[7] Brebner, G.G. and Wyatt, L.A. Boundary Layer Measurements at Low Speed on
Two Wings of 45o and 55o Sweep. ARC CP 554, 1961.
[8] Carter, J.E. Solutions for Laminar Boundary Layers with Separation and Reattachment. AIAA 74-583, 1974.
[9] Carter, J.E. Viscous-Inviscid Interaction Analysis of Turbulent Separated Flow.
AIAA 81-1241, 1981.
[10] Carter, J.E. and Wornom, S.F. Solutions for Incompressible Separated Boundary
Layers Including Viscous-Inviscid Interaction, Aerodynamic Analysis Requiring Advanced Computers. NASA SP-347, 1975.
[11] Catherall, D. and Mangler, K.W. The Integration of the Two-Dimensional Laminar
Boundary-Layer Equations past the Point of Vanishing Skin Friction. Journal of
Fluid Mechanics, 26(1):163{182, 1966.
[12] Cebeci, T., Hefazi, H., and Roknaldin, F. Predicting Stall and Post-Stall Behaviour
of Airfoils at Low Mach Numbers. AIAA Journal, 33(4):595{602, 1995.

124

BIBLIOGRAPHY

[13] Coenen, E.G.M. Modi cations to the DERA VFP Code to Include QuasiSimultaneous Coupling (QS-VFP). Technical Report 98-01, University of Bristol,
1998.
[14] Coenen, E.G.M. Quasi-Simultaneous Coupling for Wing and Aerofoil Flow. In Lai,
C.-H., Bjorstad, P.E., Cross, M. and Widlund, O.B., editor, Domain Decomposition
Methods in Sciences and Engineering, pages 197{205, Bergen, Domain Decomposition Press, 1999.
[15] Coenen, E.G.M. Quasi-Simultaneous Viscous-Inviscid Interaction for Swept Tapered Wings. In 7th Annual Conference of the Association for Computational Mechanics in Engineering, ACME 1999, Durham, United Kingdom, 1999.
[16] Coenen, E.G.M., Veldman, A.E.P. and Patrianakos, G. Quasi-Simultaneous
Viscous-Inviscid Interaction for Three-Dimensional Turbulent Wing Flow. In 22nd
International Congress of Aeronautical Sciences, ICAS 2000, Harrogate, United
Kingdom, 2000.
[17] Coenen, E.G.M., Veldman, A.E.P. and Patrianakos, G. Viscous-Inviscid Interaction
Method for Wing Calculations. In European Congress on Computational Methods
in Applied Sciences and Engineering, ECCOMAS 2000, Barcelona, Spain, 2000.
[18] Cousteix J. Three-Dimensional and Unsteady Boundary-Layer Computations. Annual Review Fluid Mechanics, 18:173{196, 1986.
[19] Cousteix, J. Couche Limite Laminaire. Cepadeus Editions, 1988.
[20] Cousteix, J. and Houdeville, R. Singularities in Three-Dimensional Turbulent
Boundary-Layer Calculations and Separation Phenomena. AIAA 81-4201, 1981.
[21] Cross, A.G.T. Calculation of Compressible Three Dimensional Turbulent Boundary
Layers with Particular Reference to Wings and Bodies. BAe Brough, YAD Note
3379, 1979.
[22] Cumpsty, N.A. and Mead, M.R. The Calculation of Three-Dimensional Turbulent
Boundary Layers Part II: Attachment-Line Flow on an In nite Wing. Aeronautical
Quarterly, 18:150{164, 1967.
[23] Dargel, G. and Thiede, P. Viscous Transonic Airfoil Flow Simulation by an Ecient
Viscous-Inviscid Interaction Method. AIAA 87-0412, 1987.
[24] Delery, J.M. and Formery, M.J. A Finite Di erence Method for Inverse Solutions
of Three-Dimensional Turbulent Boundary Layer Flow. AIAA 83-0301, 1983.
[25] Drela, M. Two-Dimensional Aerodynamic Design and Analysis Using the Euler
Equations. PhD thesis, MIT, 1985.
[26] Drela, M. XFOIL: An Analysis and Design System for Low Reynolds Number
Airfoils. Low Reynolds Number Aerodynamics, 1989.

BIBLIOGRAPHY

125

[27] Drela, M. and Giles, M.B. Viscous-Inviscid Analysis of Transonic and Low Reynolds
Number Airfoils. AIAA Journal, 25:1347{1355, 1987.
[28] Duck, P.W. and Burggraf, O.R. Spectral Solutions for Three-Dimensional TripleDeck Flow over Surface Topography. Journal of Fluid Mechanics, 162:1{22, 1986.
[29] East, L.F., Smith, P.D. and Merryman, P.J. Prediction of the Development of
Separated Turbulent Boundary Layers by the Lag-Entrainment Method. RAE TR
77046, 1977.
[30] Edwards, D.E. Analysis of Three-Dimensional Flow Using Interacting BoundaryLayer Theory. In Smith, F.T. and Brown, S.N., editor, IUTAM Symposium on
Boundary-Layer Separation, pages 163{178. Springer-Verlag, 1987.
[31] Edwards, D.E. and Carter, J.E. Analysis of Three-Dimensional Separated Flow
with the Boundary-Layer Equations. AIAA 85-1499, 1985.
[32] Edwards, D.E., Carter, J.E. and Smith, F.T. Analysis of Three-Dimensional Separated Flow with the Boundary-Layer Equations. AIAA Journal, 25(3):380{387,
1987.
[33] Firmin, M.C.P. Calculations of Transonic Flow over Wing/Body Combinations with
an Allowance for Viscous E ects. RAE TM Aero 1861, 1980.
[34] Firmin, M.C.P. Applications of RAE Viscous Flow Methods near Separation Boundaries for Three-Dimensional Wings in Transonic Flow. RAE TM Aero 2065, 1986.
[35] Fluid Dynamics Panel Working Group 10. Calculation of 3D Separated Turbulent
Flows in Boundary Layer Limit. AGARD-AR-255, 1990.
[36] Forsey, C.R. and Carr, M.P. The Calculation of Transonic Flow over ThreeDimensional Swept Wings Using the Exact Potential Equation. DGLR Symposium
Transonic Con gurations, Bad Harzburg, DGLR Paper 78-064, 1978.
[37] Goldstein, S. On Laminar Boundary Layer-Flow near a Position of Separation.
Quarterly Journal of Mechanics and Applied Mathematics, 1:43{69, 1948.
[38] Granvill, P.S. The Calculation of the Viscous Drag on Bodies of Revolution. David
Taylor Model Basin Report 849, 1953.
[39] Green, J.E. Application of Head's Entrainment Method to the Prediction of Turbulent Boundary Layers and Wakes in Compressible Flow. RAE TR 72079, 1972.
[40] Green, J.E., Weeks, D.J. and Brooman, J.W.F. Prediction of Turbulent Boundary
Layers and Wakes in Compressible Flow by a Lag-Entrainment Method. RAE TR
72231, 1973.
[41] Hastings, R.C. and Williams, B.R. Studies of the Flow Field near a NACA 4412
Aerofoil at Nearly Maximum Lift. Aeronautical Journal, 91(901):29{44, 1987.

126

BIBLIOGRAPHY

[42] Head, M.R. Entrainment in the Turbulent Boundary Layer. ARC R&M 3152, 1960.
[43] Henkes, R.A.W.M. and Veldman, A.E.P. On the Breakdown of the Steady and
Unsteady Interacting Boundary-Layer Description. Journal of Fluid Mechanics,
179:513{529, 1987.
[44] Hess, J.L. and Smith, A.M.O. Calculation of Potential Flow about Arbitrary Bodies.
Progress in Aeronautical Sciences, 8:1{138, 1966.
[45] Horn, R.A. and Johnson C.R. Topics in Matrix Analysis. Cambridge University
Press, 1991.
[46] Horton, H.P. A Semi-Empirical Theory for the Growth and Bursting of Laminar
Separation Bubbles. ARC CP 1073, 1967.
[47] Horton, H.P. Numerical Investigation of Regular Laminar Boundary Layer Separation. AGARD-CP-168, 1975.
[48] Houwink, R. and Veldman, A.E.P. Steady and Unsteady Flow Computations for
Transonic Airfoils. AIAA 84-1618, 1984.
[49] Jobe, C.E. and Burggraf, O.R. The Numerical Solution of the Asymptotic Equations
of Trailing Edge Flow. Proceedings Royal Society, A 340:91{111, 1974.
[50] Katz, J. and Plotkin, A. Low-Speed Aerodynamics; From Wing Theory to Panel
Method. McGraw-Hill, 1991.
[51] Kerwin, J.E., Kinnas, S.A., Lee, J.-T. and Shih, W.-Z. A Surface Panel Method for
the Hydrodynamic Analysis of Ducted Propellers. SNAME Transactions, 95:93{122,
1987.
[52] King, D.A. and Williams, B.R. Developments in Computational Methods for HighLift Aerodynamics. Aeronautical Journal, pages 265{288, 1988.
[53] Kusunose, K., Wigton, L. and Meredith, P. A Rapidly Converging Viscous/Inviscid
Coupling Code for Multi-Element Airfoil Con gurations. AIAA 91-0177, 1991.
[54] Lagerstrom, P.A. Solutions of the Navier-Stokes Equations at Large Reynolds Number. SIAM Journal of Applied Mathemtics, 28(1):202{214, 1975.
[55] Lazare , M. and Le Balleur, J.C. Computation of Three-Dimensional Flows by
Viscous-Inviscid Interaction Using the 'MZM' Method. AGARD CP-412, 1986.
[56] Le Balleur, J.C. Couplage Visqueux-Non Visqueux: Analyse du Probleme Incluant
Decollements et Ondes de Choc. La Recherche Aerospatiale, 6:349{358, 1977.
[57] Le Balleur, J.C. Couplage Visqueux-Non Visqueux: Methode Numerique et Ap
plications aux Ecoulements
Bidimensionnels Transsoniques et Supersoniques. La
Recherche Aerospatiale, 2:65{76, 1978.

BIBLIOGRAPHY

127

[58] Le Balleur, J.C.and Girodroux-Lavigne, P. Calculation of Fully Three-Dimensional


Separated Flows with an Unsteady Viscous-Inviscid Interaction Method. In Cebeci,
T., editor, Numerical and Physical Aspects of Aerodynamic Flows V, 1992.
[59] Lees, L. and Reeves, B. L. Supersonic Separated and Reattaching Laminar Flows:
I. General Theory and Application to Adiabatic Boundary-Layer Shock-Wave Interactions. AIAA Journal, 2:1907{1920, 1964.
[60] Lighthill, M.J. On Displacement Thickness. Journal of Fluid Mechanics, 4:383{392,
1958.
[61] Lock, R.C. and Williams, B.R. Viscous-Inviscid Interactions in External Aerodynamics. Progress Aerospace Sciences, 24:51{171, 1987.
[62] Mager, A. Generalization of Boundary-Layer Momentum-Integral Equations to
Three-Dimensional Flows Including those of Rotating System. NACA R 1067, 1952.
[63] Mellor, G.L. The Large Reynolds Number Asymptotic Theory of Turbulent Boundary Layers. International Journal of Engineering Science, 10:851{873, 1972.
[64] Melnik, R.E. Some Applications of Asymptotic Theory to Turbulent Flow. AIAA
91-0220, 1991.
[65] Melnik, R.E. and Chow, R. Asymptotic Theory of Two Dimensional Trailing Edge
Flows. NASA SP-347 (also Grumman Research Dept. Report RE-510J), 1975.
[66] Melnik, R.E. and Grossman, B. On the Turbulent Viscid-Inviscid Interaction at a
Wedge-Shaped Trailing Edge. In Cebeci, T., editor, Numerical and Physical Aspects
of Aerodynamic Flows I, pages 211{235, 1982.
[67] Melnik, R.E., Chow, R. and Mead, H.R. Theory of Viscous-Inviscid Transonic Flow
over Airfoils at High Reynolds Number. AIAA 77-680, 1977.
[68] Messiter, A.F. Boundary Layer Flow near the Trailing Edge of a Flat Plate. SIAM
Journal on Applied Mathematics, 18:241{257, 1970.
[69] Milewski, W.M. Three-Dimensional Viscous Flow Computations Using the Integral
Boundary Layer Equations Simultaneously Coupled with a Low Order Panel Method.
PhD thesis, Department of Ocean Engineering, MIT, 1997.
[70] Moran, J. An Introduction to Theoretical and Computational Aerodynamics. John
Wiley and Sons, 1984.
[71] Morino, L. and Kuo, C.-C. Subsonic Potential Aerodynamics for Complex Con gurations: A General Theory. AIAA Journal, 12(2):191{197, 1974.
[72] Mughal, B. and Drela, M. A Calculation Method for the Three-Dimensional
Boundary-Layer Equations in Integral Form. AIAA 93-0786, 1993.

128

BIBLIOGRAPHY

[73] Myring, D.F. An Integral Prediction Method for Three-Dimensional Turbulent


Boundary-Layers in Incompressible Flow. RAE TR 70147, 1970.
[74] Neiland, V.Ya. Theory of Laminar Boundary Layer Separation in Supersonic Flow.
Fluid Dynamics, 4:33{35, 1970.
[75] Nishida, B. and Drela, M. Fully Simultaneous Coupling for Three-Dimensional
Viscous/Inviscid Flows. AIAA 95-1806-C, 1995.
[76] Prandtl, L. Uber Flussigkeitsbewegung bei sehr kleiner Reibung. Proceedings International Mathematical Congress, pages 484{491, 1904.
[77] Roberts, W.B. Calculation of Laminar Separation Bubbles and their E ect on
Airfoil Performance. AIAA Journal, 18(1):25{31, 1980.
[78] Roget, C., Brazier, J.Ph., Cousteix, J. and Mauss, J. A Contribution to the Physical
Analysis of Separated Flows past Three-Dimensional Humps. European Journal of
Mechanics B/Fluids, 17(3):307{329, 1998.
[79] Schlichting, H. Boundary Layer Theory. McGraw-Hill, 1979.
[80] Smith, F.T. Theoretical Aspects of Steady and Unsteady Laminar Separation.
AIAA 84-1582, 1984.
[81] Smith, F.T. Steady and Unsteady 3-D Interactive Boundary Layers. Computers
and Fluids, 20(3):243{268, 1991.
[82] Smith, P.D. Calculation Methods for Three-Dimensional Turbulent Boundary Layers. ARC R&M 3523, 1966.
[83] Smith, P.D. An Integral Method for Three-Dimensional Compressible Turbulent
Boundary Layer. ARC R&M 3739, 1972.
[84] Smith, P.D. The Numerical Computation of Three-Dimensional Boundary Layers.
In Fernhoh, H.M. and Krause, E., editor, IUTAM Symposium on Three-Dimensional
Turbulent Boundary Layers, pages 266{285. Springer-Verlag, 1982.
[85] Smith, P.D. A Viscous Package for Attached and Separated Flows on Swept and
Tapered Wings. RAE TR 89027, 1989.
[86] Somers, M.A.M. and Veldman, A.E.P. The Inclusion of Streamline Curvature in a
Quasi-Simultaneous Viscous-Inviscid Interaction Method for Transonic Airfoil Flow.
Preprint available at http://www.math.rug.nl/~veldman/preprints.html, 1999.
[87] Stewartson, K. Multistructured Boundary Layers on Flat Plates and Related Bodies.
Advances in Applied Mechanics, 14:145{239, 1974.
[88] Stewartson, K., F.R.S. and Williams, P.G. Self-Induced Separation. Proceedings
Royal Society, A 312:181{206, 1969.

BIBLIOGRAPHY

129

[89] Stewartson, K., Smith, F.T. and Kaups, K. Marginal Separation. Studies in Applied
Mathematics, 67:45{61, 1982.
[90] Streett, C.L. Viscous-Inviscid Interaction for Transonic Wing-Body Con gurations
Including Wake E ects. AIAA Journal, 20(7):915{923, 1982.
[91] Sychev, V.V. and Sychev, V.V. Turbulent Separation (Russian translation). Zhurnal
Vychislitelnoi Matematiki i Matematicheskoi, 20(6):1500{1512, 1980.
[92] Sykes, R.J. On Three-Dimensional Boundary Layer Flow over Surface Irregularities.
Proceedings Royal Society, A 373:311{329, 1980.
[93] Van Dalsem, W.R. and Steger, J.L. Using the Boundary-Layer Equations in ThreeDimensional Viscous Flow Simulation. AGARD-CP-412, 1986.
[94] Van der Wees, A.J. and Van Muijden, J. A Robust Quasi-Simultaneous Interaction Method for a Full Potential Flow with a Boundary Layer with Application to
Wing/Body Con gurartions. In Cebeci, T., editor, Numerical and Physical Aspects
of Aerodynamic Flows V, 1992.
[95] Van der Wees, A.J., Van Muijden, J. and Van der Vooren, J. A Fast and Robust
Viscous-Inviscid Interaction Solver for Transonic Flow about Wing/Body Con gurations on the Basis of Full Potential Theory. AIAA 93-3026, 1993.
[96] Varga, R.S. Matrix Iterative Analysis. Prentice-Hall, 1962.
[97] Veldman, A.E.P. Boundary Layers with Strong Interaction from Asymptotic Theory
to Calculation Method. Proceedings BAIL 1 Conference, Dublin, 1980.
[98] Veldman, A.E.P. New, Quasi-Simultaneous Method to Calculate Interacting Boundary Layers. AIAA Journal, 19:79{85, 1981.
[99] Veldman, A.E.P. Strong Viscous-Inviscid Interaction and the E ects of Streamline
Curvature. CWI Quarterly, 10(3&4):353{359, 1997.
[100] Veldman, A.E.P. and Van de Vooren, A.I. Drag of a Finite Plate. In R.D. Richtmyer,
editor, Proceedings 4th International Conference on Numerical Methods in Fluid
Dynamics, Lecture Notes in Physics, volume 35, pages 422{430, Springer-Verlag,
Berlin, 1975.
[101] Veldman, A.E.P., Lindhout, J.P.F., De Boer, E. and Somers, M.A.M. VISTRAFS:
A Simulation Method for Strongly Interacting Viscous Transonic Flow. In Cebeci,
T., editor, Numerical and Physical Aspects of Aerodynamic Flows IV, pages 37{51,
1990.
[102] Verho , A., Chen, H.H., Cebeci, T. and Michal, T. An Accurate and Ecient
Interactive Boundary-Layer Method for Analysis and Design of Airfoils. AIAA 960328, 1996.

130

BIBLIOGRAPHY

[103] Weber, J. and Brebner, G.G. Low Speed Tests on 45o Sweptback Wings. Part I:
Pressure Measurements on Wings of Aspect Ratio 5. RAE R Aero 2374, 1950.
[104] White, F.M. Viscous Fluid Flow. McGraw-Hill, 1991.
[105] Wigton, L. and Yoshihara, H. Viscous-Inviscid Interactions with a Three Dimensional Inverse Boundary Layer Code. Presented at Symposium on Numerical and
Physical Aspects of Aerodynamic Flows II, 1983.
[106] Williams, B.R. The Prediction of Separated Flow Using a Viscous-Inviscid Interaction Method. Aeronautical Journal, 89:185{197, 1985.
[107] Williams, B.R. Viscous-Inviscid Interaction Schemes for External Aerodynamics.
RAE TM Aero 2152, 1989.
[108] Yoshihara, H. and Wai, J.C. Transonic Turbulent Separation on Swept Wings - A
Return to the Direct Formulation. AIAA 84-0265, 1984.
[109] Young, A.D. Boundary Layers. BSP Professional Books, 1989.
[110] Young, D.M. Iterative Solution of Large Linear Systems. Academic Press, 1971.

Appendix A
Streamline and Cartesian
boundary-layer integral thicknesses
Streamline coordinate de nitions
In a streamline coordinate system system (s; n; z ), where s is the streamline and n the
cross ow direction in the plane tangent to the surface of the boundary layer and z normal
to the plane, the integral thicknesses are de ned as follows:
s

ss =
ns =

(1

us
) dz;
e qe

us
(1
e qe
Z

us
) dz;
qe

us un
dz;
e qe2

n

sn =
nn =

un
dz;
e qe

un
(1
e qe

us
) dz;
qe

u2n
dz;
e qe2

where represents the real boundary-layer thickness. In the above de nitions


us and un
p
are the velocity components in the s and n direction. Velocity qe = u2e + u2e = ue
is the total velocity, subscript e indicating that the value is taken at the edge of the
boundary layer.
When the ow is assumed to be incompressible the density  = e = constant.
s

Cartesian coordinate de nitions


In a Cartesian coordinate system system (x; y; z ), where x and y lie in the plane tangent
to the surface of the boundary layer and z is normal to the plane, the integral thicknesses
are de ned as follows, representing the boundary-layer thickness:
x =

1
( u
e qe e e

u) dz;

y =

1
( v
e qe e e

v ) dz;

132

Appendix A Streamline and Cartesian boundary-layer integral thicknesses

xx =
yx =

u
(u
e qe2 e
u
(v
e qe2 e

u) dz;

xy =

v ) dz;

yy =

v
(u
e qe2 e

u) dz;

v
(v
e qe2 e

v ) dz:

In thepabove de nitions u and v are the velocity components in the x and y direction and
qe = u2e + ve2 is the total velocity at the edge of the boundary layer. When the ow is
incompressible the density  = e = constant.

Relationship streamline and Cartesian systems


Through rotation over the angle , a relationship exists between the Cartesian coordinate
system and the streamline coordinate system, where the angle between the x and s
directions, de ned by
u
v
cos = e ;
sin = e :
qe
qe
With this transformation the integral thicknesses in the Cartesian coordinate system can
be expressed in terms of the integral thicknesses in the streamline system. Hence, the
displacement thicknesses can be written

x = s cos n sin ;
y = s sin + n cos ;

and the momentum thicknesses can be given by


xx = ss cos2 (2sn +  ) sin cos + nn sin2 ;
n

nn ) sin cos + sn (cos2 sin2 ) n sin2 ;


= (ss nn ) sin cos + sn (cos2 sin2 ) + n cos2 ;
= ss sin2 + (2sn + n ) sin cos + nn cos2 :

xy = (ss
yx
yy

Appendix B
Boundary-layer closure relations
B.1 Streamwise closure
In the paper by Head [42] the following shape factors have been de ned:

H =

s
;
ss

H1 =

s )

ss

Laminar
Laminar closure for H to be used by Thwaites' method is given by [70]
8
2
> 2:61 3:752 + 5:242 ; 0 < 2 < 0:1;
H =

<

0:0731
;
0:1 < 2 < 0;
2:088 +
2 + 0:14
with 2 the dimensionless pressure gradient parameter de ned as
2 @u
2  ss s :
 @s
>
:

(B.1)

(B.2)

(B.3)

If Thwaites' method is not used for the laminar boundary-layer calculations, but instead
the Von Karman equation together with the entrainment equation, a laminar closure
relation for H1 is required. In a book by Cousteix [19] the following relation is given:


H
1
H
H1
+ a 10 = b
+
+ c;
H10
H1
H (4:02923)2
where H10 = 12:37 and

a = 1:2706; b = 1:5022; c = 3:1924;


a = 0:33044; b = 0:31993; c = 1:03094;

for H  4:02923;
for H > 4:02923:

(B.4)

134

Appendix B Boundary-layer closure relations

Turbulent
Turbulent closure required for the shape factor H1 , can be given by the expression in Lock
and Williams [61], being
8




H 1 1:093
1:12 1:093
>
>
< 2 + 1:5
+ 0:5
; H < 4;
H 1
1:12
H1 =
(B.5)
>
>
:
4 + 13 (H 4);
H  4;
which has a minimum at H = 2:7, corresponding with separation.
In a paper by Hastings and Williams [41] it was remarked that if separation occurs as
a result of strong adverse pressure gradients, H is overpredicted and the H -H1 relation is
incorrect. A unique relationship between H and H1 for all practical cases still has to be
found.
By reducing the slope of @H1 =@H in separated ow by a factor two, Houwink was able
to suppress the instabilities at shock-induced separation [48]. This same idea can be used
for severe trailing-edge separation. Houwink's H -H1 relationship is given by
8
(0:5H + 1)H
>
>
;
H < 2:732;
>
>
>
H
1
>
>
H1 =

with

>
>
<
>
>
>
>
>
>
>
>
>
:

(0:5ht + 1)ht
;
ht 1

1:75 +

ht  4 and H  2:732;

(B.6)

5:2273ht
; ht > 4 and H  2:732;
ht + 5:818181

ht = 0:5(H 2:732) + 2:732:


Both H -H1 relations, (B.5) and (B.6) have been incorporated in the prediction methods.
A third relation for H -H1 relation can be found by using the simple power law streamwise velocity pro le of the form [83]
us  z n
=
;
(B.7)
qe

which gives the following relationship for H1 , to be used for the closure of integral thicknesses in three-dimensional ow:
2H
H1 =
:
(B.8)
(H 1)

B.2 Cross ow closure


Turbulent
The empirical cross ow velocity pro le proposed by Mager, given in [62] is
un  z  2
= 1
tan ;
qe

(B.9)

B.3 Skin-friction closure

135

and together with (B.8) the boundary-layer integral thicknesses become expressions in
terms of only ss , H and [83]

ns =
nn =
n =

2ss tan
;
(H 1)(H + 2)
(H
(H

sn = ns

(B.10)

24ss tan2
;
1)(H + 2)(H + 3)(H + 4)
16Hss tan
;
1)(H + 3)(H + 5)

s :

(B.11)
(B.12)
(B.13)

Unlike Johnston's cross ow velocity pro le [83], the above given Mager's pro le allows for
zero skin-friction and can be used in the wake region.
A set-back of using Mager's cross ow velocity pro le is that it has been established only
for fairly small values of . For close to 45o it gives implausible values for the integral
thicknesses [61]. It would therefore be more accurate to use the more sophisticated Cross
velocity pro le [21], which has as yet not been implemented.

B.3 Skin-friction closure


Laminar
In the book by Moran [70] a laminar closure relation for Cf s is given for Thwaites' method,
being
Cf s =

8
>
>
>
>
<

2
(0:22 + 1:572
Re

>
>
>
>
:

2
Re

ss


ss

1:822);

0 < 2 < 0:1;




0:0182
0:22 + 1:4022 +
;
2 + 0:107

(B.14)
0:1 < 2 < 0;

with 2 de ned in (B.3).

Turbulent
The expression used for the turbulent streamwise skin-friction coecient, Cf s , used in
both the two-dimensional and three-dimensional calculations, follows the paper by Green
et al. [40]
Cf s

0:9
= Cf 0
H=H0 0:4

0:5 ;

(B.15)

136

Appendix B Boundary-layer closure relations

with the following empirical correlations for Cf 0 and H0 :


0:01013
0:00075;
log Re
1:02
1
p
;
H0 =
1 6:55 Cf 0 =2

Cf 0 =

ss

in which Re is the Reynolds number based on ss . It has been mentioned in [61] that
the above expression for the streamwise skin-friction (B.15) bears no relation with reality
for separated ow. It is sucient then to set Cf s to a small negative value
ss

Cf s =

0:00001;

(B.16)

as in separated ow the boundary-layer development is not strongly in uenced by Cf s .


The wake layer resembles a turbulent boundary layer but is without the wall layer and
Cf s and Cf 0 are set to zero.
The cross ow component of the skin-friction coecient, Cf n , is given by

Cf n = Cf s tan :

(B.17)

B.4 Entrainment closure


The empirical relationship used for CE , found by Head in [42] is written
CE = 0:0306(H1 3) 0:6169 ;

(B.18)

which is used in both two- and three-dimensional programs.


In the wake region, where there is more vigorous turbulent mixing, the above formula
might need to be changed. For symmetric ow the wake can be seen as one viscous layer,
being twice as thick as the viscous layers on the wing. Hence, the largest eddies in the
wake can be twice the size as well and the entrainment coecient should be multiplied
by two. For non-symmetric ows there can be a signi cant di erence at the trailing edge
in displacement thickness, upper and lower. The wake layer is, in that case, not twice as
thick and the closure relation for the entrainment coecient should remain unchanged.
Other closure for CE can be provided by the lag-entrainment method, which uses an
extra partial di erential equation, the so-called lag-equation, instead of (B.18). For more
details of the lag-entrainment method see papers [29, 40].

Appendix C
Derivation Cartesian integral
boundary-layer equations
C.1 Boundary-layer equations
The ow of a viscous uid is represented by the Navier-Stokes equations [104]
Conservation of mass

@
+ r  q = 0;
@t

Conservation of momentum

@q
+ (q  r)q =
@t

(C.1)

rp + r  ;

(C.2)

where  is the density and q = (u; v; w)T is the velocity vector. The term  represents
the stress tensor which for a Newtonian uid, in Cartesian tensor notation is given by

ij =

@q
@qi @qj
2
ij k + 
+
;
3 @xk
@xj @xi

(C.3)

in which  is the dynamic viscosity and ij the Kronecker symbol.


When the dynamic viscosity equals zero the Navier-Stokes equations reduce to the
Euler equations, which for incompressible steady ow in three dimensions are written as

@u
@u
@u
+v +w
=
@x
@y
@z
@v
@v
@v
u +v +w
=
@x
@y
@z
@w
@w
@w
=
u +v +w
@x
@y
@z
u

in a Cartesian coordinate system.

1 @p
;
 @x
1 @p
;
 @y
1 @p
;
 @z

(C.4)
(C.5)
(C.6)

138

Appendix C Derivation Cartesian integral boundary-layer equations

With the assumptions that the freestream Reynolds number Re is large and the characteristic boundary-layer thickness small compared to the characteristic chord length of the
pro le, Prandtl's boundary-layer equations are derived from the Navier-Stokes equations
@u @v @w
+ +
= 0;
(C.7)
@x @y @z
@u
@u
@u
1 @p 1 @x
u +v +w
=
+
;
(C.8)
@x
@y
@z
 @x  @z
@v
@v
@v
1 @p 1 @y
u +v +w
=
+
;
(C.9)
@x
@y
@z
 @y  @z
1 @p
;
(C.10)
0 =
 @z
where the plane z = 0 is the wall.

C.2 Integral momentum equations


Subtracting from the x-momentum boundary-layer equation (C.8) the Euler equation
(C.4), at the edge of the boundary layer (with we = 0) gives
@u
@u
@u
@u
@u 1 @x
u +v +w
= ue e + ve e +
:
(C.11)
@x
@y
@z
@x
@y  @z
The above equation is to be integrated over the boundary-layer thickness . First this is
done for the term involving w
Z

@u
w dz = [ wu ]0
@z
= we ue

@w
dz
@z
Z

w0 u0

= ue @w0 +
= ue w0

ue

(C.12)

@w
dz
@z

@w A
dz
w0 u0
@z

Z 

@u @v
+
dz
@x @y

@w
dz
@z

w0 u0 +

@u @v
u
+
dz;
@x @y

0
0
where in the last step use is made of the continuity equation. Replacing the integral over
the z derivative (C.12), equation (C.11) becomes
Z

@u
u dz +
@x
+

@u
v dz + uew0
@y


ue


Z 

@u @v
+
dz
@x @y

@u @v
@u
@u
u
+
dz = ue e + ve e
@x @y
@x
@y

w0 u0
w
;

x

(C.13)

C.3 Entrainment equation

139

with w0 being zero at the surface and taking the gradients outside of the integrals gives

@
ue
@x

@
ue
@y

udz

ue

@ue
@x

@
vdz +
@x
ve

@ue
=
@y

@
u2 dz +
@y
w
:


uvdz

(C.14)

Combining terms

@
@x

@ue
@x

u(u ue)dz
@ue
@y

@
u)dz +
@y

(ue

(ve

v )dz =

v (u ue)dz

(C.15)

w
;

x

0
and applying the boundary-layer integral thicknesses as de ned in appendix A, using
 = 1, nally gives the integral x-momentum boundary-layer equation

@
@
@u
@u
(xx qe2 ) + (xy qe2 ) = qe x e qe y e + w :
(C.16)
@x
@y
@x
@y
Similarly the y -momentum integral boundary-layer equation can be found to be
@
@
@v
@v
(yx qe2 ) + (yy qe2 ) = qe x e qe y e + w :
(C.17)
@x
@y
@x
@y
x

C.3 Entrainment equation


The entrainment equation in three dimensions can be found using Head's concept of the
entrainment velocity. From the physics of a control volume, sketched in gure C.1, with
(x; y ) the boundary-layer thickness, the volume ow rate within the boundary layer at
x is
y

udz:

0
The volume ow rate out of the control volume at x + x is

y

@
udz + y x
@x

udz:

(C.18)

(C.19)

0
0
The increase in volume ow in the x-direction over distance x therefore becomes

@
y x
@x

udz:

(C.20)

140

Appendix C Derivation Cartesian integral boundary-layer equations

z
y

Figure C.1: Control volume.


In a similar way, the increase in volume ow in the y -direction can be found to be

@
y x
@y

vdz:

(C.21)

Per unit base area the entrainment velocity, E , describing the rate at which the external
velocity enters the boundary layer through its external edge is

@
E =
@x

@
udz +
@y

vdz:

(C.22)

Introducing the dimensionless entrainment velocity, CE = E=qe then

CE

1 @
=
qe @x

1 @
udz +
qe @y

vdz;

(C.23)

which with the use of the boundary-layer integral thicknesses, de ned in appendix A, can
be rewritten to the integral boundary-layer entrainment equation

CE =

1 @
1 @
( qe x + ue ) +
qe @x
qe @y

qe y + ve :

(C.24)

The above derived entrainment equation (C.24) can also be found in a more theoretical
way by integrating the continuity equation (C.7) over the boundary-layer thickness.

Appendix D
Auxiliary de nitions and theorems
De nition D.1 (Horn-Johnson [45, p. 92]) A matrix A is called positive stable if all the
eigenvalues  of A ful l <() > 0.
De nition D.2 (Varga [96, p. 9]) Let A be an n  n (complex) matrix with eigenvalues
i , for 1  i  n. Then
(A) = 1max
j j;
in i

(D.1)

is the spectral radius of the matrix A.

De nition D.3 (Horn-Johnson [45, p. 113]) A real n  n matrix A is called an M-matrix


if A is positive stable and A 2 Zn , where the set Zn is de ned by
Zn = fA = [Aij ] : Aij  0 if i 6= j (i; j = 1;    ; n)g.
Theorem D.4 (Horn-Johnson [45, p. 114]) If A is an M-matrix, then A is nonsingular
and A 1  0.
Theorem D.5 (Horn-Johnson [45, p. 114]) If A is an M-matrix, then A has positive
diagonal entries.
Theorem D.6 (Horn-Johnson [45, p. 117 & p. 127]) Let A; B 2 Zn , suppose A is an
M-matrix, and assume B  A. Then B and B 1 A are both M-matrices. Furthermore,
0  B 1  A 1.
De nition D.7 (Varga [96, p. 37]) A n  n matrix A is weakly diagonally dominant if

jAiij 

n
X
k=1
k =i

jAik j

and with at least for one i the inequality sign.

for i = 1; 2; : : : n;

(D.2)

142

Appendix D Auxiliary de nitions and theorems

De nition D.8 (Young [110, p. 37]) A matrix A is irreducible if and only if there does
not exist a permutation matrix P such that P 1 AP has the form
P 1AP =

F 0
G H

(D.3)

where F and H are square matrices and where all elements of 0 vanish.

Theorem D.9 (Young [110, p. 38]) A matrix A of order n is irreducible if and only if
n = 1 or, given any two distinct integers i and j with 1  i  n, 1  j  n, then Aij 6= 0
or there exist i1 ; i2 ;    ; is such that Aii1 Ai1 i2    Ai j 6= 0.
s

Theorem D.10 (Varga [96, p. 85]) An irreducible, weakly diagonally dominant matrix A
with positive diagonal elements and with Aik  0 (8 i 6= k) is an M-matrix.
De nition D.11 (Young [110, p. 21]) A matrix A is positive de nite if A is Hermitian
(i:e: A = A) and (x; Ax) > 0, for all x 6= 0.
t

Theorem D.12 (Varga [96, p. 85]) If matrix A has positive diagonal elements and is
irreducible, symmetric and weakly diagonally dominant, then matrix A is positive de nite.
Theorem D.13 (Horn-Johnson [45, p. 95])) Let Q be a positive real matrix (i.e. the symmetric part of Q is positive de nite), then for every positive de nite matrix A the product
AQ is positive stable.
Theorem D.14: Perron(1907)-Frobenius(1912) (Varga [96, p. 30]) Let A  0 be an
irreducible n  n matrix. Then,
(1) A has a positive real eigenvalue equal to its spectral radius,
(2) to (A) there corresponds an eigenvector x > 0,
(3) (A) increases when any entry of A increases,
(4) (A) is a simple eigenvalue of A.
Historically, Perron proved theorem D.8 assuming A > 0. Later, Frobenius extended
Perron's results for nonnegative (A  0) and irreducible matrices. If the entries of matrix
A are at the very least nonnegative, the following theorem can be applied.

Theorem D.15 (Young [110, p. 18]) Let A  0 be an n  n matrix. Then,


(1) to (A) there corresponds an eigenvector x  0,
(2) (A) does not decrease when any entry of A increases.
De nition D.16 (Varga [96, p. 88]) For n  n real matrices A, M and N , A = M
is a regular splitting of the matrix A if M is nonsingular with M 1  0 and N  0.
Theorem D.17 (Varga [96, p. 89]) If A = M

N is a regular splitting of the matrix A

143
and A 1  0 then

(M 1 N ) =

(A 1 N )
< 1:
1 + (A 1 N )

(D.4)

Theorem D.18 (Varga [96, p. 90]) Let A = Ma Na = Mb Nb be two regular splittings


of A, where A 1  0. If Nb  Na  0, then
0  (Ma 1 Na )  (Mb 1 Nb ) < 1:
(D.5)
Theorem D.19: Gerschgorin (Varga [96, p. 16]) Let A be an arbitrary n  n (complex)
matrix, and let
n
X

i =
j

=1; j 6=i

jAij j;

i = 1; 2;    ; n:

(D.6)

Then, all the eigenvalues  of A lie in the union of the disks

jz Aiij 

i ;

i = 1; 2;    ; n:

(D.7)

144

Appendix D Auxiliary de nitions and theorems

Samenvatting
Stromingsverschijnselen zijn overal om ons heen. Te denken valt aan bijvoorbeeld de
golven in de zee of de wind die de bladeren kan doen dansen op straat. De meeste stromingsverschijnselen worden beschreven door de zogeheten Navier-Stokes vergelijkingen die
zijn gebaseerd op de behoudswetten van massa, impuls en energie. De Navier-Stokes vergelijkingen zijn complex en moeilijk op te lossen. Alhoewel ze meer dan anderhalve eeuw
geleden zijn geformuleerd, zijn er tot op heden alleen exacte oplossingen gevonden voor
heel eenvoudige stromingsproblemen. Recentelijk is het oplossen van de Navier-Stokes
vergelijkingen uitgeroepen tot een van de zeven `Millennium Prijs Problemen' door het
Clay Wiskunde Instituut in Cambridge (Massachusetts), waaraan een beloning van een
miljoen dollar is verbonden (zie www.claymath.org of NRC Handelsblad 5/8/2000).
Doordat analytische oplossingen voor problemen die relevant zijn voor de industrie
niet bestaan, zijn andere manieren nodig om stromingverschijnselen te bestuderen. Een
manier om inzicht te krijgen in deze moeilijkere stromingsverschijnselen is via experimenteel onderzoek. Experimenten hebben echter als nadeel dat ze erg duur zijn en veel tijd
vergen, zoals in het geval van windtunneltests voor het ontwerpen van vliegtuigen. Een
modern alternatief is stromingsverschijnselen met behulp van een computer te bestuderen.
Dankzij de voortdurende ontwikkelingen in ecientie en nauwkeurigheid van numerieke
algoritmes, samen met de toenemende snelheid en geheugenruimte van computers, beginnen numerieke methodes daarom een steeds grotere rol te vervullen in de aero- en
hydrodynamica.
In dit proefschrift wordt een eciente numerieke methode ontwikkeld voor het analyseren en ontwerpen van vliegtuigvleugels. Twee werkwijzes kunnen worden gevolgd om
een numerieke methode te ontwikkelen voor het modelleren van de luchtstroming langs
een vleugelpro el. Een aan populariteit winnende manier is het met brute rekenkracht
oplossen van de Navier-Stokes vergelijkingen. Hoewel Navier-Stokes simulatie in principe
algemeen kan worden toegepast, verhinderen momenteel de hoge kosten van rekentijd en
de complexe turbulentiemodellering het gebruik van deze aanpak.
De tweede manier, die in dit proefschrift onderzocht wordt, is de oudere en ecientere
methode van viskeuze/niet-viskeuze interactie die kan worden gebruikt voor vloeistofstromingen met lage viscositeit (stroperigheid, di usie). In het geval van stroming rond een
vliegtuig, bijvoorbeeld, zijn de viskeuze krachten klein en slechts merkbaar in een dunne
laag langs het oppervlak. Buiten deze dunne laag (grenslaag) zijn de viskeuze e ecten
verwaarloosbaar en kan worden verondersteld dat de stroming niet viskeus is. In tegenstelling tot Navier-Stokes simulatie benutten viskeuze/niet-viskeuze interactiemethodes deze
informatie door gebruik te maken van een zogeheten zone-techniek. Het gehele stromingsgebied is opgedeeld in twee zones: een buitengebied waar de niet-viskeuze vergelijkingen

146

Samenvatting

gelden, en een dunne grenslaag langs het oppervlak, waar de viskeuze grenslaagvergelijkingen geldig zijn. Beide systemen van vergelijkingen zijn numeriek goedkope, echter
toch voldoende nauwkeurige, vereenvoudigingen van de Navier-Stokes vergelijkingen. Om
de oplossing te vinden voor het gehele stromingsprobleem moeten de oplossingen eerst
in beide zones worden berekend en dan worden gekoppeld. Viskeuze/niet-viskeuze interactietechnieken zijn zeer succesvol gebleken voor praktische twee-dimensionale stromingsproblemen. Ze zijn eenvoudig te implementeren en numeriek erg economisch. De
berekende resultaten komen voor een groot aantal toepassingen, zoals bijvoorbeeld voor
de stroming rond een vliegtuig in kruisvlucht, goed met experimentele resultaten overeen.
Kortom, viskeuze/niet-viskeuze interactietechnieken zijn erg geschikt voor het modelleren
van aerodynamische stromingen en worden derhalve in dit proefschrift geanalyseerd en
gebruikt voor het onderzoek van verscheidene stromingsproblemen.
Sinds de introductie van het grenslaagconcept door Prandtl in 1904 zijn verschillende
viskeuze/niet-viskeuze interactiemethodes ontwikkeld om het systeem van grenslaagvergelijkingen en niet-viskeuze vergelijkingen op te lossen. De oudste en meest bekende
methode is de zogeheten directe methode, die de grenslaagvergelijkingen oplost met een
voorgeschreven snelheidsverdeling (of drukverdeling). Voor situaties met gladde (aanliggende) stroming werkt de directe methode goed. Echter, voor realistische situaties met
gebieden waar terugstroming plaatsvindt, zoals bij de achterrand van een vliegtuigvleugel
tijdens het opstijgen of landen, laten de directe grenslaagberekeningen het volledig afweten. Andere interactiemethodes zijn ontwikkeld die wel in staat zijn de berekeningen voort
te zetten in deze gebieden. Dit zijn de (semi-)inverse en (quasi-)simultane methodes.
Van de verschillende interactiemethodes levert de quasi-simultane methode de beste
prestaties op het gebied van snelheid en robuustheid in twee dimensies. De quasi-simultane
methode lijkt op de klassieke directe methode. Echter, in plaats van de grenslaagvergelijkingen op te lossen met een voorgeschreven snelheidsverdeling, worden de grenslaagvergelijkingen opgelost met een benadering voor de niet-viskeuze stroming in het buitengebied.
De benadering van de niet-viskeuze buitenstroming wordt `interactiewet' genoemd, en
beschrijft het essentiele deel van de fysische interactie tussen de grenslaag en het nietviskeuze buitengebied. Opgemerkt moet worden dat de interactiewet de uiteindelijke
oplossing niet benvloedt. Echter de interactiewet heeft wel belangrijke gevolgen voor de
snelheid en de robuustheid van de quasi-simultane viskeuze/niet-viskeuze interactiemethode.
Voor drie-dimensionale viskeuze/niet-viskeuze interactie worden dezelfde voordelen
verwacht die zijn verkregen in twee dimensies. Verschillende pogingen zijn ondernomen
om te komen tot een viskeuze/niet-viskeuze interactiemethode voor drie-dimensionale problemen. De extra dimensie maakt de keuze van een numeriek schema en het modelleren
van terugstroming een ingewikkelder probleem dan in twee dimensies. Om deze problemen te onderzoeken en om een goedwerkende drie-dimensionale viskeuze/niet-viskeuze
interactiemethode te ontwikkelen, is in dit proefschrift de quasi-simultane methode gebruikt. Van de verschillende interactiemethodes levert de quasi-simultane methode niet
alleen de beste prestaties in twee dimensies, maar hij bezit ook de extra eigenschap dat
hij gemakkelijk naar drie dimensies kan worden uitgebreid.
Zoals gezegd zijn drie-dimensionale berekeningen moeilijker en vereisen ze meer rekentijd dan twee-dimensionale berekeningen. Om een drie-dimensionale viskeuze/niet-

Samenvatting

147

viskeuze numerieke methode interessant te maken voor de industrie, is het essentieel dat
de methode zo robuust mogelijk is en dat de rekentijden acceptabel zijn. Ter voorbereiding op deze bovengenoemde punten wordt in dit proefschrift het twee-dimensionale
quasi-simultane interactieproces grondig onderzocht. Dit leidt tot de ontwikkeling van
fundamentele wiskundige theorie over de keuze van een interactiewet om de robuustheid
en de snelheid van de viskeuze/niet-viskeuze quasi-simultane interactiemethode te optimaliseren. De nieuwe wiskundige theorie wordt numeriek geevalueerd voor het geval van
twee-dimensionale trogstroming, en de theorie en de numerieke resultaten blijken consistent.
Om de ontwikkelde quasi-simultane viskeuze/niet-viskeuze interactiemethode verder
te testen, worden realistische twee- en drie-dimensionale aerodynamische stromingsproblemen gemodelleerd onder verschillende stromingscondities. Om de volledige drie-dimensionale capaciteiten van de methode te tonen is terugstroming gesimuleerd voor een vleugel
met een pijlhoek van 45 graden. De voorspelde krachten, drukverdelingen en grenslaagvariabelen zijn in goede overeenstemming met de experimentele data en andere numerieke
resultaten.
Samenvattend, in dit proefschrift wordt nieuwe fundamentele wiskundige theorie gepresenteerd die een duidelijk inzicht geeft in de quasi-simultane interactiemethode, en
richtlijnen worden gegeven voor de optimalisatie van het koppelingsproces van de grenslaag
met de niet-viskeuze buitenstroming. Twee- en drie-dimensionale quasi-simultane methodes zijn ontwikkeld, die de geformuleerde richtlijnen volgen en die goede resultaten
geven voor de verschillende gesimuleerde stromingsproblemen. Het onderzoek beschreven in dit proefschrift leidt tot een beter begrip van drie-dimensionale quasi-simultane
viskeuze/niet-viskeuze interactie, hetgeen in de toekomst de mogelijkheid biedt vliegtuigvleugels ecienter te ontwerpen.

148

Samenvatting

Summary
Flow phenomena are around us everywhere. One can think of the waves in the sea or
the wind playing with the leaves fallen from the trees. Most of these uid ow situations are described by the so-called Navier-Stokes equations which are based on the
conservation laws of mass, momentum and energy. The Navier-Stokes equations, though
fundamental, are complex and dicult to solve. To this day exact solutions have only
been found for very simple ow cases. Recently solving the Navier-Stokes equations was
named one of the seven `Millennium Prize Problems' by the Clay Mathematics Institute
in Cambridge (Massachusetts), with a prize fund of one million dollars allocated to it (see
www.claymath.org or NRC Handelsblad 5/8/2000).
Meanwhile, as analytical solutions for industrially relevant problems are still undetermined, other approaches need to be followed to study uid ow phenomena. One
approach to obtain knowledge about these more dicult ow cases is by experimentation. Experiments however, have the major disadvantage that they are very expensive
and time-consuming, as one can imagine in the case of wind-tunnel tests for aircraft con guration design. A modern alternative is to study uid ow problems with the help of a
computer. Due to the continuing advances in eciency and accuracy of numerical algorithms, together with the growing speed and memory size of computers, ow calculation
methods therefore start playing an increasing role in uid dynamics.
In this thesis an ecient computational method is developed to be used for the design
and analysis of aircraft wings. Two approaches can be followed to come to a numerical method for the modelling of wing ow. The rst approach is to use Navier-Stokes
simulation which is being applied more and more these days. Whilst Navier-Stokes simulation potentially o ers generality, the high computational cost involved and the complex
turbulence modelling currently limits its use.
The second approach, which is investigated in this thesis, is to use the older and
more ecient method of viscous-inviscid interaction which can be employed for uid
ows with small viscosity (syrupiness, di usion). For the ow around an aeroplane the
viscous forces are small and they are con ned to a thin layer close to the surface. Outside
this thin (boundary) layer the viscous e ects are negligible and the ow can assumed to
be inviscid (non-viscous). Unlike Navier-Stokes simulation, viscous-inviscid interaction
methods intelligently take this information into account, and make use of a so-called
zonal technique. The whole ow eld is divided into two zones: an external region where
the inviscid equations hold, and a boundary layer close to the surface, where the viscous
boundary-layer equations are valid. Both sets of equations in the viscous and inviscid ow
regions are computationally cheap, yet suciently accurate, simpli cations of the NavierStokes equations. In order to nd the composite solution for the whole ow problem,

150

Summary

both zones are rst calculated separately and then coupled repeatedly. Viscous-inviscid
interaction techniques have proven to be very successful for practical two-dimensional
ow cases. They are straightforward to implement and computationally very economical.
Furthermore, for most applications the computed results compare well with experiment,
as for example for ow around an aeroplane under cruise conditions. Viscous-inviscid
interaction techniques therefore are most suitable for the modelling of aerodynamic ows
and are analysed and applied for the various ow scenarios investigated in this thesis.
Since the boundary-layer concept was rst introduced by Prandtl in 1904, several
viscous-inviscid interaction methods have been developed to solve the system of boundarylayer and external inviscid ow equations. The oldest and best-known is the so-called
direct method which solves the boundary-layer equations with a prescribed velocity (or
pressure) distribution. For situations with smooth (attached) ow the direct method
works well. However, for situations where regions with ow reversal are present, such as
near the trailing edge of a wing during landing and take-o , the direct boundary-layer
calculations breakdown. Other interaction methods have been developed that are able
to continue the calculations into the region with ow reversal. They are known as the
(semi-)inverse and (quasi-)simultaneous methods.
Of the various interaction methods the quasi-simultaneous method gives the best performance in terms of speed and robustness in two dimensions. The quasi-simultaneous
method resembles the classical direct method. However, instead of solving the boundarylayer equations with a prescribed external velocity distribution as is the case with the
direct method, the boundary-layer equations are solved together with an approximation
of the external inviscid ow. The approximation of the external inviscid ow is termed
`interaction law', and describes the essential part of the physical interaction between the
boundary layer and the external inviscid ow. It is noted that the interaction law does not
in uence the nal solution. However, the interaction law plays an important role into determining the speed and robustness of the quasi-simultaneous viscous-inviscid interaction
method.
Similar advantages as obtained in two dimensions, are expected for three-dimensional
viscous-inviscid interaction. Several e orts have been directed towards the development
of a viscous-inviscid interaction method for three-dimensional ow problems, however,
more research is still needed. The added dimension makes the choice of numerical scheme
and the modelling of ow reversal a more complex problem than it is in two dimensions.
In order to deal with these problems and to develop a good-working three-dimensional
viscous-inviscid interaction method, the quasi-simultaneous technique is employed in this
thesis. Of the various interaction methods the quasi-simultaneous method not only gives
the best performance in two dimensions, however, it also has the additional quality of a
straightforward extension to three dimensions.
As mentioned three-dimensional calculations are dicult and require much more computation time than two-dimensional calculations. In order to make a three-dimensional
viscous-inviscid computational method interesting to industry, it is therefore essential to
make the method as robust as possible and to optimise the run times. To prepare for
these issues, the two-dimensional quasi-simultaneous interaction process is thoroughly investigated in this thesis. This leads to the development of fundamental mathematical
theory and clear guidelines on how to chose the interaction law in order to optimise the

Summary

151

robustness and speed of the viscous-inviscid quasi-simultaneous interaction method. The


new mathematical theory is evaluated numerically for the test case of two-dimensional
trough ow, and theory and numerics show to be consistent.
To further validate the developed quasi-simultaneous viscous-inviscid interaction method realistic two- and three-dimensional aerodynamic ow cases are modelled under di erent ow conditions. To show the full three-dimensional capabilities of the method, ow
reversal is simulated for a 45 degree swept wing. The predicted forces, pressure distributions and boundary-layer variables are in good comparison with experimental data and
other computational results.
Concluding, in this thesis new basic mathematical theory is presented which gives
clear insight into the quasi-simultaneous interaction method and which provides guidelines for the optimisation of the coupling process between the boundary layer and the
external inviscid ow. A two- and three-dimensional quasi-simultaneous method is developed, following the formulated guidelines and producing good results for the di erent ow
scenarios calculated. With the research described in the present thesis the understanding
of three-dimensional quasi-simultaneous viscous-inviscid interaction is enhanced, which
will be useful for future industrial design and analysis of aircraft wings.

152

Summary

Acknowledgements
Dit proefschrift is het resultaat van onderzoek begonnen als Ph.D. student aan de University of Bristol in Engeland en na drie jaar voortgezet als assistent in opleiding (AiO)
aan de Rijksuniversiteit Groningen. Veel mensen hebben me geholpen en gesteund tijdens
mijn promotieonderzoek en een aantal van hen wil ik graag hier bedanken.
In de eerste plaats ben ik heel veel dank verschuldigd aan Arthur Veldman, mijn promotor. Arthur, je hebt met mij je kennis en grote enthousiasme voor grenslaagstromingen
gedeeld. Je had op ieder moment tijd om naar me te luisteren en met me te praten over
allerhande zaken, waardoor ik het in Groningen heel erg naar mijn zin heb gehad. Ook
wist je me altijd te steunen en te motiveren als ik het moeilijk had. Ik kan je niet genoeg
bedanken!
Secondly, I would like to thank Steve Fiddes, my supervisor at the University of Bristol,
who o ered me the opportunity to start a Ph.D. in aerodynamics and helped me through
the rst years.
Verder gaat mijn dank uit naar de beoordelingscommissie, bestaande uit prof.dr.ir.
P.G. Bakker, prof.dr.ir. H.W.M. Hoeijmakers en prof.dr.ir. Hendrik Hoogstraten, voor
het nauwkeurig lezen van mijn manuscript en het geven van kritisch commentaar.
Eugen Botta wil ik graag bedanken voor het controleren van de stellingen en bewijzen in hoofdstuk 4, en voor het geven van mooie aanvullingen op enkele stellingen die
daardoor nog scherper geformuleerd konden worden.
I would like to thank John Doherty and Kevin Hackett from DERA, Farnborough,
for giving me the opportunity to work on an industrial wing code in a very stimulating
environment and for introducing me to various people working on viscous-inviscid interaction. I would also like to say thanks to Nigel Taylor for having me as his house-mate
during my time at DERA and always making me laugh with his stories.
Many people have helped to make my stay in England a very pleasant one. I would like
to mention my colleagues Chris, Dorian, Graham, John, Max, Pablo and his girlfriend
Hestia, Paul, Rob, Taqi, Yacine and especially George, Nick and Jothi. George, I am
really grateful for all your support and the many `discussions' about potential ow theory
and life in general. Also when I had just returned to Groningen you always made time to
answer my emails or chat with me on the phone which I really appreciated. Furthermore,
I would like to thank you for supplying me with your three-dimensional potential ow
code required for my interactive calculations. Nick, I would like thank you very much
for your patience explaining me panel methods and helping me with all my computer
problems. Jothi, thanks for all the female chatting and laughter we had and the fun times
shopping in Bristol. Also I would like to thank prof. David Birdsall for always looking
out for me and supporting me during my last months in England.

154

Acknowledgements

Na mijn periode in Bristol werd ik in Groningen al snel opgenomen door de AiO's van
de afdeling. Graag wil ik Ena, Erwin, Evgeny, Geert, Lucian, Marc, Robert, Roland en
Theresa bedanken voor hun bijdrage aan de goede sfeer op de afdeling. Met name wil
ik echter graag Jeroen bedanken bij wie ik altijd terecht kon met een vraag of voor een
praatje, en Barteld met wie ik behalve de kamer ook al mijn verhalen deelde. Barteld, je
hebt me ontzettend gesteund tijdens mijn ups en downs en ook met het editen van mijn
sollicitatiebrieven en samenvattingen heb je me erg geholpen. Natuurlijk ook bedankt
voor alle gezelligheid en leuke dingen die we gedaan hebben, zoals het naar de lm gaan,
zwemmen, en dansen in De Kar.
Ook wil ik graag mijn woongroepgenoten bedanken voor alle heerlijke etentjes en
de gezelligheid. Bedankt Anniek, Ester, Javier en de Baulig super kok Coert, die ik
ook zeer dankbaar ben voor zijn hulp met het maken van het plaatje op de voorkant.
In het bijzonder wil ik Wendy bedanken voor het zijn van niet alleen een geweldige
huisgenoot maar ook een goede vriendin. We hebben heel wat afgepraat en ontzettend
veel lol gemaakt, gillend in het huis, op de tennisbaan en natuurlijk niet te vergeten in
het zwembad.
Er zijn een aantal mensen die ik al lang ken en wiens goede vriendschap heel erg
belangrijk voor me is en hen wil ik daarvoor hier ook graag zeer bedanken. Kris, thanks
for the good times we had, going for pub lunches and high teas, or driving through the
English countryside, Scotland and Sweden in the mini, me `reading' the map. Also I
would like to thank you immensely for all your support and encouraging telephone calls
during the dicult times. Marco, heel erg bedankt dat je er altijd voor me bent geweest.
Ook wil ik je heel erg bedanken voor al de gekke leuke (in)spannende dingen die we
hebben ondernomen, zoals het hardlopen, etsen en restaurantjes verkennen in Groningen,
wijn drinken in Assen en toch ook voor het niet-kameel rijden in Tunesie. Twee goede
vriendinnen, en tevens mijn paranimfen, studiegenootjes, mede-kenners van het AiObestaan, zijn Krista en Sacha, die me al tien jaar bijstaan met hart en ziel, waarvoor ik ze
ontzettend dankbaar ben. We hebben veel met elkaar gedeeld en meegemaakt en hopelijk
blijven we dat doen.
Als laatste bedank ik mijn zus en mijn ouders voor alle steun en hulp altijd vanaf het
begin.