Rijksuniversiteit Groningen
ViscousInviscid Interaction
with the QuasiSimultaneous Method
for 2D and 3D Aerodynamic Flow
Proefschrift
door
Promotor:
Referent:
Beoordelingscommissie:
ISBN 9036714729
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, 18981976), which stands in front of the Miro museum in Barcelona, Spain.
Contents
1 Introduction to viscousinviscid interaction
1.1 Introduction . . . . . . . . . . . . . . . . . .
1.2 Viscousinviscid interaction . . . . . . . . . .
1.3 Historical overview . . . . . . . . . . . . . .
1.4 Outline present research . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
. 5
. 6
. 7
. 10
2 Boundarylayer region
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Integral boundarylayer formulation . . . . . . . . . . . . . . . .
2.3 Coordinate systems . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Twodimensional boundary layer . . . . . . . . . . . . . . . . .
2.4.1 Thwaites' method for laminar boundary layers . . . . . .
2.5 Quasithreedimensional boundary layer . . . . . . . . . . . . . .
2.6 Threedimensional boundary layer . . . . . . . . . . . . . . . . .
2.7 Discretisation boundarylayer equations . . . . . . . . . . . . . .
2.7.1 Discretisation two and quasithreedimensional problems
2.7.2 Discretisation threedimensional 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 Fullpotential 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 twodimensional model problem .
3.4.3 Discretisation threedimensional model problem
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
43
47
48
50
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
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
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
5.8
5.9
5.10
5.11
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6 Results
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Twodimensional aerofoil results . . . . . . . . . . . . . .
6.2.1 NACA0012 aerofoil . . . . . . . . . . . . . . . . .
6.2.2 NACA4412 aerofoil . . . . . . . . . . . . . . . . .
6.3 Swept tapered wing results . . . . . . . . . . . . . . . . .
6.4 Threedimensional dented plate results . . . . . . . . . .
6.4.1 Flat plate . . . . . . . . . . . . . . . . . . . . . .
6.4.2 Threedimensional trough
ow . . . . . . . . . . .
6.4.3 Threedimensional dent
ow . . . . . . . . . . . .
6.5 Threedimensional wing results . . . . . . . . . . . . . .
6.5.1 Quasitwodimensional NACA0012 wing
ow . . .
6.5.2 Threedimensional unswept NACA0012 wing
ow
6.5.3 Threedimensional 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
131
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
133
. 133
. 134
. 135
. 136
141
Samenvatting
145
Summary
149
Acknowledgements
153
CONTENTS
Chapter 1
Introduction to viscousinviscid
interaction
1.1 Introduction
The accurate and fast prediction of the development of viscous
ow over two and especially threedimensional 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 (Reynoldsaveraged) NavierStokes equations,
which represent the
ow of a viscous
uid. Whilst NavierStokes simulation potentially
oers generality, the excessive requirements for computer memory and CPU (CentralProcessingUnit) time currently limit their use for practical application, especially within
a design optimisation environment where a large number of congurations 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 viscousinviscid interaction (VII), whereby an inviscid
ow
solver is coupled to a viscous boundarylayer calculation method to provide a composite
solution of the NavierStokes equations.
For twodimensional
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 NavierStokes simulation and this at much
lower computational cost than required for NavierStokes simulation. The development of
threedimensional viscousinviscid 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 setback of the coupled schemes is that they are applicable to a smaller range of
applications, compared to NavierStokes solvers. Viscousinviscid 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 boundarylayer thickness
and, containing (long)
at separation bubbles. The calculation of larger regions with
separation (marginal separation or even larger, massive separation) is beyond the scope of
viscousinviscid 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 viscousinviscid interaction method.
In this thesis, several computational methods based on the idea of viscousinviscid interaction have been developed and results are presented for the prediction of incompressible twodimensional aerofoil
ow, transonic quasithreedimensional swept tapered wing
ow and fully threedimensional incompressible
ow over dented plates and aircraft wings.
The quasisimultaneous interaction method is used to couple the inviscid
ow solver to
the boundarylayer equations and limited regions with two and threedimensional
ow
separation are permitted.
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
noslip 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.
10
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 multizonalmarching technique Lazare and
Le Balleur [55] and Le Balleur and GirodrouxLavigne [58] computed threedimensional
separated
ows with a semiinverse method. However, when the semiinverse 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 quasisimultaneous 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 congurations has been developed [94, 95].
From the quasisimultaneous method similar advantages as were obtained in two dimensions in terms of convergence speed and robustness are expected in threedimensional
calculations. Although some work has been done on threedimensional viscousinviscid
interaction an ecient and robust quasisimultaneous method is still to be fully developed
and tested.
11
Fundamental mathematical theory for the quasisimultaneous 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
wellknown twodimensional CarterWornom trough problem.
Chapter 5 is an extension of chapter 4. The quasisimultaneous interaction method is
extended to threedimensions and the approximations of the external
ow are determined
for the practical problems of twodimensional aerofoil
ow and threedimensional 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 2
Boundarylayer region
2.1 Introduction
In the present section the mathematical and numerical boundarylayer formulations are
given for the two and (quasi)threedimensional
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.
14
dierential equations loose one dimension and only body surface points are required for
the calculation. Integral methods are therefore computationally more economical than
dierential methods. The accuracy of both methods is generally the same, depending
mainly on the closure relations and the transition modelling applied.
With respect to viscousinviscid interaction, integral methods have shown to be more
robust, due to the builtin 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 boundarylayer formulations have been used for the dierent
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
twodimensional
ow as presented in Green's paper [39]. The description for quasithreedimensional 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 boundarylayer formulations
are presented. However, rst the various coordinate systems in which the boundarylayer
equations can be written, are discussed.
n, u n
yc , v c
external
streamline
s, u s
c
c
z
xc , u c
x, u
15
Myring [73] developed a calculation method which uses the boundarylayer equations
transformed to a general nonorthogonal curvilinear coordinate system. He dened 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 sdirection is dened by c .
Metric coecients, accounting for the curvature of the grid, are required for the
boundarylayer equations in a curvilinear coordinate system, and if the direction of the
external streamline changes they have to be recomputed, which for threedimensional
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 threedimensional method
locally using a Cartesian coordinate system (x; y ) tangent to the surface for the formulation of the integral boundarylayer equations. Via a simple rotation the boundarylayer
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
eect 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 twodimensional
ow problems, with no cross
ow present, the boundary layer
equations are simply solved along the external streamline s. The quasithreedimensional
boundarylayer method on the other hand, developed by Smith who based his work
on Myring's paper, uses a nonorthogonal curvilinear coordinate system. For the fully
threedimensional
ow calculations the paper by Mughal and Drela is followed and the
boundarylayer equations are written in a local Cartesian coordinate system. As mentioned, via a transformation the boundary layer variables for the (quasi)threedimensional
methods can be expressed in the orthogonal curvilinear system.
(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 denitions for the streamwise displacement thickness
s and the streamwise momentum thickness ss can be found in appendix A. The shape
factor H is dened as s =ss and shape factor H1 corresponds with ( s )=ss, where
is the boundarylayer thickness. The coecients Cf s and CE are the skinfriction and
e
16
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 dened 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
17
+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 dened in appendix A and Cf s and Cf n are the skinfriction 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 eect, 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 nmomentum 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 boundarylayer equations can be transformed to the general nonorthogonal curvilinear coordinate system (xc ; yc), where c the angle between the xc  and ycdirections and
c the angle between the xc  and sdirections, as seen before in gure 2.1. The metric
coecients h1 and h2 of the xc  and ycdirections, respectively, are dened 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
wall
streamline
cos(c c ) @us
:
h1 us sin c @xc
The three integral boundarylayer 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 twodimensional relation
for the streamwise skinfriction coecient is used as given in appendix B. With
Cf n
un
=
;
(2.14)
tan = zlim
!1 us
Cf s
the cross
ow skinfriction coecient is determined. The lagentrainment 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 =
19
@
@
@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 twodimensional Von Karman equation (2.1). The entrainment equation
(2.17) can be found by integrating the continuity equation across the boundary layer,
as in twodimensions [39]. A derivation of the threedimensional integral boundarylayer
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 proles, it is necessary to work in a streamline coordinate system. The
boundarylayer 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 sdirections. 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
Taking into account the above mentioned closure relations and prescribing the velocity
components ue and ve , only three unknowns remain for the threedimensional integral
boundarylayer equations. The nature of the set of the three partial dierential 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 xdirection 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 boundarylayer equations is strongly aected by the hyperbolic nature the partial
dierential equations, which should be taken into account in the numerical scheme.
The laminar region of the threedimensional boundarylayer
ow is calculated with
the twodimensional Thwaites' method as given in (2.9). For the cases calculated in this
thesis, the threedimensional eects in the small laminar region are negligible, justifying
this approach.
@s
@s i
1
2
s
s 1
:
si
i
(2.19)
The above central dierencing has been tested for the twodimensional aerofoil problem.
21
Discretisation timederivative
For the unsteady boundarylayer formulation a timederivative is present in the momentum equations, which is to be discretised. In combination with upwind discretisation, the
timederivative of s in (2.4) at viscousinviscid iteration n becomes
@s(
@t
n)
s(
n)
s(
t
1)
(2.20)
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 socalled 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 dened 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
22
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 simplied to
P (xc ; yc)
Z
C
Z
C
ZZ
A
dx dy = 0; (2.23)
where C is the contour (anticlockwise) around the area A and subscript c indicates
that values at the centre of the cell are used. The above equation no longer contains
dierentials, so nite dierencing is no longer required.
The present method for the calculation of the fully threedimensional boundarylayer
equations will make use of a grid constructed by parallelogram panels, as shown in gure
2.4. Parallelogram panel (i; j ) is dened 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 nitevolume 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 xdirection 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
(i1, j1)
(i1, j)
1
0
0
1
(i, j)
11
00
00
11
local
z
x
PANEL
(i, j)
11
00
00
11
(i, j1)
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
@
@u
@u
@
(xx qe2 ) + (xy qe2 ) + qe x e + qe y e
@x
@y
@x
@y
w
= 0;
(xy
ij
q2
1 eij 1
ij
q2
1j ei 1j
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 )
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 nitevolume
discretisation for the xmomentum boundarylayer 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 dene 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
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 threedimensional wing problem the stagnation line does not
have to coincide with one and the same spanwise gridline.
ss =
25
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 OrrSomereld equation [79]. First, the point at which the
ow becomes unstable is calculated
Re
ssuns
52
= exp
H
14:8 :
(2.26)
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 modication of Granville's empirical criterion which ignored
freestream turbulence [38].
26
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
signicantly.
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, modied
by Roberts [77] to include the eects 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
Tripped transition
The most used method to model transition is simply to dene 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.
27
28
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 designoptimisation environment, a surface
singularity method for irrotational subsonic
ow is preferred to a physical more complete
Euler or fullpotential
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 boundarylayer displacement eect is taken into account, by modelling an extra source
singularity distribution over the prole'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 nonlifting thinaerofoil/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/boundarylayer calculations.
30
p(X,Y,Z)
r
Y
q(,,)
Sw
Sb
V,
S
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.
31
E (p) =
>
>
<
>
>
>
>
:
1
2 ; for p on the boundary of V ;
(3.6)
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
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 dierence 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 dierence 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 denitions 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, dened as 1 (p) = U1 X + V1 Y + W1 Z , to satisfy the
condition at innity. 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 satised. 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 dene the
ow near sharp trailing edges) to relate in the
wake to on the body.
b
32
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 full 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)
@
@Y ;
@
@Z
Q1 ;
(3.16)
@
@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)
33
The above dened 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)
(3.20)
(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)
@
@
@
(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
34
= @=@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
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)
Qtot =
r :
(3.31)
35
Kutta condition
In the case of modelling wing or aerofoil
ow the problem is not uniquely specied by
only the boundary condition of no normal velocity through the prole'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 dierence 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 boundaryvalue 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 twodimensional
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
From the freestream the source strength can be determined, using the denition 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 constantpotential Dirichlet
problem is uniquely dened 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 prole's surface. Equations (3.10) and (3.11),
dierentiated 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 threedimensional
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)
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 dened 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 socalled transpiration
velocity, describing the transpiration between the boundary layer and the external
ow.
The denition 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 denitions 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
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
modied, 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 eect is a jump in normal velocity w1 across the wake
centre line, dividing the
ows coming from the upper and lower surfaces of the prole
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 dened
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)
39
with the inviscid and the viscous source strengths on the body dened 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).
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
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.
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 straightline panels each of similar length
x = xk+1 xk for k 2 [1; Mp ]. The collocation points are dened 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 denition 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
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 righthand
side of equation (3.54) describes the disturbance eect 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 (denition D.3 and theorem D.10) and therefore positive denite (theorem D.12).
42
ue
ij
ve
ij
1
=
2
=
1
2
xMp +1 yMs +1
Z
Z
(3.56)
(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 wellbehaved. As in the
twodimensional 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
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
(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:
Lwe
ij
= we + (
ij
@w
xi ) e + (
@ ij
@w
yj ) e ;
@ ij
(3.59)
we =
@
@
(u1Zp + qe x ) + (v1 Zp + qe y ):
@x
@y
(3.60)
Using the denition of Lwe (3.59) and centrally discretising the gradients of qe x , qe y ,
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
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 dened 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 loworder 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 ], dened 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 simplied 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
and a similar system is found for the twodimensional
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
Threedimensional case
Amn
Twodimensional 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 dened in a local panelorientated 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 signicant
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
Qtot = Qs + Q1 + n
= Qs + Q1 (Q1 n)n + w1 n;
e
(3.68)
and fulls 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
( n . Q  w ) n
45
Q tot
n
s
1111111111111111111111111111111
0000000000000000000000000000000
0000000000000000000000000000000
1111111111111111111111111111111
0000000000000000000000000000000
1111111111111111111111111111111
0000000000000000000000000000000
1111111111111111111111111111111
t1 and t2 are dened as two dierent 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 dierencing.
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 ] dene 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
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 simplied, using (3.45) and (3.47) for a twolayered
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, dened (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 threedimensional
ow problem where the velocities in the Cartesian coordinate
system are used, the twodimensional
ow problem will make use of a streamline
p 2coordi2
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 denitions 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
47
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
(3.79)
with Amn and Bmn as dened 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 dierencing, 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
L
CL = 1 2 ;
2 Q1 c
D
CD = 1 2 ;
2 Q1 c
(3.82)
48
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 skinfriction 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 =
(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)
w = te
l
te ;
l
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.
49
ellipse for which the analytical solution is known. The geometry of the ellipse is dened
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 complexfunction 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
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
0.8
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
0.2
0.4
0.6
0.8
50
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 dierent 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 shtail. The analytical Kutta condition, implying equal
upper and lower velocities at the trailing edge, is not satised.
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 shtail behaviour near the trailing edge for larger angles of attack remains. For the
boundarylayer coupling this can cause problems and a modication of the trailing edge
values is required and will be discussed in the next section.
The threedimensional 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.
51
The velocities at the trailing edge coming from the wing and aerofoil Dirichlet panel
method are modied before being used for the boundarylayer calculations. Various possibilities have been tested and in the twodimensional 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 boundarylayer 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 threedimensional wing external velocity distribution a slightly dierent modication 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 4
Viscousinviscid interaction with the
quasisimultaneous method
4.1 Introduction
Viscousinviscid 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 twodimensional
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 quasisimultaneous method. In the present chapter the quasisimultaneous method
is introduced and analysed in detail to determine its basic requirements.
ue =
E [x ];
ue =
V [x ];
(4.1)
which describes a nonlinear system in ue, the velocity at the edge of the boundary layer,
and x , the boundarylayer displacement thickness.
54
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 boundarylayer equations to determine the new boundarylayer
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 viscousinviscid 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 denitions 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 skinfriction becomes zero at H = Hsep and corresponds for twodimensional steady
ow calculations with the separation point. Furthermore, it can be shown that
b
a
8
<
:
> 0
= 0
< 0
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 dierential 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 innite, unless the righthand
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 satised. 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 =
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 innite. The solution for x is irrealistic and violates the boundarylayer 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 nonphysical singularity, is the socalled inverse method [11, 8].
The direction of the iterative process between the boundary layer and the external
ow
is inverted and the boundarylayer equations are solved with a prescribed displacement
thickness
8 ( )
< x
= E 1 [u(en 1) ];
(4.7)
( )
: (n)
ue = V [x ]:
n
Analysing boundarylayer 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 boundarylayer 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 twolayered asymptotic structure is no longer correct.
Instead a socalled tripledeck 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
boundarylayer 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 thinaerofoil theory.
56
From the tripledeck theory it is clear that near a singularity the viscous and inviscid
ow
regions have become equally important. A numerical viscousinviscid interaction method
therefore should locally treat both regions together to model the strong interaction. Lagerstrom described the tripledeck idea in 1975 with the following words [54]: \An important
feature is that the pressure is selfinduced, that is, the pressure due to displacement thickness is determined simultaneously with the revised boundarylayer solution. [ : : : ] this
solution exhibits a denite loss of hierarchy."
The rst numerical solutions for the tripledeck 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 dierent asymptotic
ow eld structure exists compared to laminar
ow. However, it is assumed that conclusions from the tripledeck 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 socalled simultaneous method [25, 27], re
ecting the lack of hierarchy by treating
the boundarylayer 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 twodimensional 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 dierential 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
nonexistence/nonuniqueness of a steady
ow solution. Another mathematical model is
to be used instead. With the unsteady boundarylayer 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
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 boundarylayer equations at each iteration.
Furthermore, the simultaneous method can be dicult to implement when employed
to combine an external
ow solver with the boundarylayer equations in their original
dierential form. These problems are avoided by using adaptations of the direct and
inverse schemes, which take into account the character of tripledeck theory.
Semiinverse method
A scheme based on the inverse method is the semiinverse 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 boundarylayer equations.
In (4.11) it is seen that the boundarylayer 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
eV
eE
u(en) + V [x( ) ] = 0;
(4.12)
u(en) = E [x( ) ]:
n
58
to model regions with strong interaction. It is seen that the interaction law ue I [x ], can
also be interpreted as solving the boundarylayer equations with a natural generalisation
of ue prescribed.
The interaction law equation in (4.12) is employed in a decit 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 boundarylayer 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 twodimensional
ow calculations it has been shown that the quasisimultaneous
method, using a suitable interaction law, compared to the semiinverse method, has a
much higher convergence rate and is much more robust [61].
V
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 simplied 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
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
Mmatrix (denition D.3 and theorem D.10) and hence positive denite (theorem D.12).
Before starting with the construction of an interaction law based on E , the various
iterative schemes of the quasisimultaneous method are investigated to determine which
requirements the interaction law has to satisfy.
p
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
boundarylayer equation is matched with the edge velocity obtained from the external
ow equation, i.e. ue = ue . Therefore, using ue = ue , the system dened by (4.17)
and (4.18) can be rewritten as
V
(E + V ) x = 0:
(4.19)
Assumption 4.1 It is assumed that the positive stable matrix E + V possesses nonpositive odiagonal entries, i.e. E + V is an Mmatrix.
60
From (4.16) it is clear that E is a diagonally dominant matrix with nonpositive odiagonal 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 nonpositive
odiagonal entries and assumption 4.1 therefore is not unreasonable.
Introducing ue = Ix as the interaction law and applying (4.12), the quasisimultaneous
viscousinviscid 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)
x
(n)
(n 1)
(4.21)
x
(n)
= !vi (I + V ) 1 (I
E )x
(n 1)
+ (1 !vi )x
(n 1)
(4.24)
J = 1 + !vi (vi
1):
(4.25)
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 denite and E + V is positive real, i.e.
its symmetric part is positive denite (theorem D.13). Another occasion is where E + V
and I + V are Mmatrices 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 Mmatrices and let I E , then the viscousinviscid iterations (4.21) converge.
Proof: Since M I + V is an Mmatrix, hence (I + V ) 1 0, and since N I E 0,
the splitting E + V = (I + V ) (I E ) is found to be regular (denition D.16). Since
also E + V is an Mmatrix, hence (E + V ) 1 0, theorem D.17 provides convergence.
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 odiagonals equal to zero while
keeping the odiagonals of I + V nonpositive.
62
Theorem 4.5 Let E be an external
ow matrix with nonpositive odiagonal entries. Let
the boundarylayer
ow matrix V be such that assumption 4.1 holds. Let the interaction
law I be derived according to construction 4.4. Then the viscousinviscid iterative process
(4.21) is convergent. Further, the more odiagonals of E are set to zero and/or the
more the diagonal is increased for the construction of I , the slower2 the viscousinviscid
iterations converge.
Proof: Under assumption 4.1 E + V is an Mmatrix. Furthermore, setting nonpositive
odiagonal 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 Mmatrix
(theorem D.6) and theorem 4.3 guarantees convergence of the viscousinviscid iterations.
Next let Ia and Ib be two dierent interaction laws, where Ib contains less odiagonals
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)
full
0 G
0 F
slow'.
speaking, according to the proof of the theorem the phrase `slower' should include `equally
63
(4.28)
As before, it is assumed that the viscousinviscid 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 boundarylayer
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)
(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 denition 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 odiagonal 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 nonpositive odiagonal entries. Let
the boundarylayer
ow matrix V be such that assumption 4.1 holds. Let the interaction
64
(I + V ) = (NIV )
(NEV )
= (E + V );
(4.30)
which proves the rst part of the theorem. Next, let Ia and Ib be two dierent interaction
laws, where Ib contains less odiagonals 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)
Il + V ) x
(~
n)
= Iu x
(~
n 1)
(4.32)
Theorem 4.8 Let E be an external
ow matrix with nonpositive odiagonal entries. Let
the boundarylayer
ow matrix V be such that assumption 4.1 holds. Let the interaction
law I be derived according to construction 4.4. Then the boundarylayer iterative process
(4.32) is convergent. Furthermore, the more odiagonals of E are set to zero and/or the
more the diagonal is increased for the construction of I , the faster3 the boundarylayer
iterations converge.
Proof: Firstly, convergence of the GaussSeidel 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
65
two dierent interaction laws, where Ib contains less odiagonals 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 Mmatrices, theorem D.6 yields
0 ((Id )b (Il )b + V ) 1 ((Id )a (Il )a + V ) 1 . Further, straightforward multiplication
of the nonnegative matrices involved shows that the GaussSeidel 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
GaussSeidel 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 odiagonals equal to zero, and/or by increasing
the diagonal (i.e. construction 4.4), has favourable properties: the convergence of the
viscousinviscid iterations is guaranteed, the convergence of the GaussSeidel 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 viscousinviscid iterative method
the convergence improves when the size of the interaction law increases, whereas for the
boundarylayer (GaussSeidel) 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 quasisimultaneous scheme.
66
with positive diagonal entries and nonpositive odiagonal 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 odiagonal 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 satised
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.
(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 boundarylayer 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 boundarylayer iterations, where the values at n~ are used
for (I + V )l Vd leads to a dierent 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.
67
I + V (
!vi x
n)
I +V
!vi
x
(n 1)
(4.35)
where !vi is the relaxation parameter for the viscousinviscid 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 boundarylayer 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 boundarylayer 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 semiimplicit relaxation by using the values at n for Iul + V , only scales the
main diagonal of I and is therefore more suitable. The viscousinviscid 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 viscousinviscid relaxation procedure (4.37) the Newton iterations
need to be solved for the homogeneous boundarylayer system (Id =!vi Iul + V ) x = 0
for which the GaussSeidel 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 boundarylayer systems for which the GaussSeidel schemes (4.39)
and (4.34) are to be solved are dierent, 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 viscousinviscid relaxation leads
to a more robust Newton iterative process than diagonal boundarylayer relaxation.
68
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 GaussSeidel system has become more diagonally
dominant, which is also favourable for the Newton iterations.
It is also interesting to investigate the unsteady boundarylayer formulation used in a
direct scheme (4.2), where ue is prescribed with
V
Ex
(n 1)
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 xderivative of x in boundarylayer 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 wellknown semiinverse
scheme which is known to require a small t !si , which corresponds with slow converge
[61].
69
ei
ik
ei
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 odiagonals used. It is noted that the twosided
interaction law matrix with q = 0 can also be categorised as a onesided interaction law
matrix.
70
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
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
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
1.5
2.5
3.5
X
4.5
5.5
In section 4.7 various forms of relaxation have been discussed of which seven combinations have been assessed. The results obtained using only viscousinviscid relaxation
are displayed in the tables 4.1 and 4.2. For the results given in table 4.2 viscousinviscid
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 boundarylayer relaxation, as described in (4.34), are shown in table 4.3. In tables 4.4 and 4.5 boundarylayer
relaxation is employed in combination with diagonal or full viscousinviscid relaxation.
Relaxation due to the unsteady boundarylayer formulation has been used in combination with viscousinviscid relaxation. Table 4.6 shows the results when applied together
with diagonal viscousinviscid relaxation and in table 4.7 the unsteady relaxation is combined with full viscousinviscid 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
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
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 twosided interaction laws with matrix I based on E by setting certain
odiagonals 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 odiagonals of
the one and twosided 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 twosided 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: Viscousinviscid and boundarylayer iterations with full viscousinviscid relaxation.
74
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
75
next to be discussed.
When comparing the eects of the full and diagonal viscousinviscid relaxation in tables 4.2 and 4.1, it is clear that of the two forms of viscousinviscid relaxation, diagonal
viscousinviscid relaxation is preferred, as more cases have been able to converge. Theoretically this is justied by the fact that with full viscousinviscid 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 boundarylayer relaxation is applied.
Comparing the results of table 4.3 with the results of table 4.1, it is seen that using only
boundarylayer relaxation compared to diagonal viscousinviscid 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
76
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: Viscousinviscid and boundarylayer iterations with full viscousinviscid relaxation and boundarylayer relaxation.
less robust when solving for the boundarylayer system I + V than for the viscousinviscid
relaxed system Id =! Iul + V . It is noted, however, that with the use of boundarylayer
relaxation the robustness for the full interaction law I2128 is improved. When comparing
table 4.3 with table 4.2 the results obtained with boundarylayer relaxation are an improvement compared to the results obtained with full viscousinviscid 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 boundarylayer relaxation is used in combination with diagonal
or full viscousinviscid relaxation. For both cases an improvement is visible, compared
to the cases where viscousinviscid relaxation is applied without boundarylayer relaxation. For the local tridiagonal interaction law I3 for which the boundarylayer system
I 3 + V is very unstable as (I 3 ) < (E ) (proposition 4.6), the use of boundarylayer
relaxation is a disadvantage. Furthermore, although full viscousinviscid relaxation compared to boundarylayer relaxation is less robust, the two forms of relaxation combined
77
is of the three forms most desirable. Of course, diagonal viscousinviscid relaxation using
boundarylayer relaxation compared to full viscousinviscid relaxation using boundarylayer relaxation is far more robust.
Relaxation due to the use of the unsteady boundarylayer 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 viscousinviscid 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 eects of relaxation due to the use of the unsteady
boundarylayer 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
78
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: Viscousinviscid and boundarylayer iterations with full viscousinviscid relaxation using the unsteady boundary layer formulation with t = 1.
79
inviscid iterations obtained with the simultaneous method. If no relaxation is applied also
2 = E,
the number of boundarylayer iterations for the twosided interaction law with I128
only requiring two or three viscousinviscid 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 dier a factor two, for ! = 1 in table 4.1. As I120 can be
identied as a GaussSeidel method and I20 as a Jacobi method, which for consistently
ordered matrices is known to be twice as slow as GaussSeidel, the factor two between
the two methods is explained.
The convergence speed of the boundarylayer iterations is claried by theorem 4.8. It
states that for increasing number of upper diagonals of the interaction law matrix, the
convergence of the boundarylayer iterations deteriorates. This is seen to correspond with
the numerical results, when comparing the number of boundarylayer iterations for the
twosided interaction laws I220 and I2128 for the same number of viscousinviscid iterations
for instance in table 4.1.
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 odiagonals used for the construction of twosided interaction laws, the basic convergence of the system increases. The
80
IiiA =
= location ue
= location *x
A
B
C
D
i1
i+1
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 dierent 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 quasisimultaneous 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 dened 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 odiagonal are
positive, which implies that I D is not an Mmatrix. Furthermore, compared with I A , it
is seen that the main diagonal entries are very small, IdD 0:3IdA .
Above the eect of changing the location of the unknowns has been discussed, and
it is clear that only dening both unknowns in the panel endpoints leads to a suitable
interaction law matrix.
k
82
+ 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 boundarylayer 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 boundarylayer iterations with relaxation parameter
I3JOR =
!JOR = 1 +
ei
2
:
2 + b=a x
(4.59)
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 (n1)
x i1
xi
x i+1
(4.60)
*x (n)
*x (n1)
x i1
2 (~ 1)
+ u0 ;
x x +1
i
*x (n1)
x i1
*x (n1)
x i1
xi
x i+1
Figure 4.6: Geometric interpretation Jacobi, GaussSeidel, JOR and SOR method.
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 denite 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 5
Applications of the
quasisimultaneous method
5.1 Introduction
Although a lot of work has been done on twodimensional coupling, as discussed in chapter
1, results obtained for the industrially more relevant application of threedimensional
viscousinviscid 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 threedimensional coupled problem a more dicult one to solve.
In the preceding chapter, viscousinviscid interaction with the quasisimultaneous
method for the model problem of twodimensional
ow over a plate has been investigated. In the present chapter the quasisimultaneous technique is extended to three
dimensions and applied to the more practical cases of twodimensional aerofoil
ow and
threedimensional dented plate and wing
ow. Furthermore, the solution procedure of the
coupled system is considered in more detail.
(5.1)
86
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 boundarylayer
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 dierential
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 xaxis. If a
system of partial dierential 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 twodimensional steady
ow case,
this point is generally nonsingular 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 nonphysical singularities can occur from the focusing of characteristic lines.
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
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.
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)
Jkl =
(5.4)
88
or analytically, by making use of the chainrule. 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.
U =
s
H
us
1
A;
U =
B
B
B
B
@
ss
H
c
us
1
C
C
C;
C
A
U =
B
B
B
B
@
ss
H
ue
ve
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
( )
Iik qe(n 1) x :
n
(5.5)
The interaction law formulae being linear, the number of equations can be reduced
by preeliminating the velocity unknowns. This has been done for the twodimensional
and quasithreedimensional
ow case to minimise the computational eort. For the fully
threedimensional
ow calculations, this approach requires more rewriting and the preelimination of the velocity unknowns has not been implemented.
In the two and quasithreedimensional cases, pointwise Newton iterations are performed until convergence per station is obtained, before Newton iterations start at the
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
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 specic
ow
case to be calculated.
90
boundarylayer 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) ];
(simultaneous)
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 boundarylayer 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 boundarylayer 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 viscousinviscid 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 viscousinviscid error. When viscousinviscid relaxation is applied,
vi is to be multiplied by !vi , the viscousinviscid relaxation parameter.
The boundarylayer 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
91
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
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 denite Mmatrices (see theorems D.10
and D.12). It follows that interaction law matrices based on Au and B v by setting certain
odiagonals to zero, have the same characteristics and are therefore suitably constructed
for the quasisimultaneous 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 odiagonals of matrices Au and B v can be varied for the construction of
a suitable interaction law as discussed in chapter 4.
Auii
Auijij
4
x
p
1
3 + 2 2
p
=
ln
x 3 2 2
=
1:27=x;
1:12=x:
! 1 in the Cauchy
p
y ( x2 + y 2 + x)2 @we
ln p
2 ( x2 + y 2 x)2 @x ij
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 twodimensional 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 odiagonals of I and I .
ei
ek
se
i
94
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 nonpositive odiagonal entries, is not satised
and the derived theorems in chapter 4 therefore, can not be applied.
The interaction law is based on AU by setting certain odiagonals 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 Mmatrix (theorem D.10). On an equidistant grid matrix I U is symmetric and with (5.17) and (5.18) therefore positive denite (theorem D.12). Interaction law IU is therefore suitably constructed for the quasisimultaneous interaction scheme. A detailed description is also found in [14].
s
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.
95
Similar to the threedimensional 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 satised and the derived theorems in chapter 4 can
not be applied.
Interaction law matrices are constructed from AU and B V by setting certain odiagonals to zero. As was the case for the threedimensional 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 Mmatrices (theorem D.10). On an equidistant grid the Mmatrices I U and I V are symmetric and thus positive denite (theorem D.12). Interaction
laws IU and IV are therefore suitable for the quasisimultaneous interaction scheme.
b
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
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 eect can be obtained.
L
Ue 1 Ue 2
(X Xt 1 );
(5.30)
Xt 1 Xt 2 t
where index t indicates the trailing edge point. Similarly, the boundarylayer 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
97
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
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
where I U
is an incompressible interaction law matrix as derived before.
The above interaction law matrix, with I U
being based on thinaerofoil 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 dene the solution, but to avoid breakdown at separation.
incmpr
se
incmpr
se
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
99
IUq 1 are used with the number of lower odiagonals equivalent to q = 1 and 20. For
comparison also the diagonal thinaerofoil interaction law I20 for a nonequidistant 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 viscousinviscid relaxation.
As discussed in section 5.6 the external
ow matrix E = AU does have some positive
odiagonal 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 fullled, 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 odiagonals 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 onesided interaction laws
IU1 1 , IU20 1 and IU0 2 , that for an increasing number of lower odiagonals the convergence
improves. This also follows from theorem 4.5. A similar behaviour is seen for the twosided
interaction laws when comparing IU1 2 and IU20 2 .
The diagonal Dirichlet interaction law IU0 2 and the diagonal thinaerofoil 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 thinaerofoil 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 dierence between the two diagonal interaction laws I0U 2 and I02 is very small. The thinaerofoil 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
fullled, 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
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 thinaerofoil interaction law for aerofoil
calculations. It was seen that the thinaerofoil interaction law is equally robust and
almost equally fast as the diagonal Dirichlet aerofoil interaction law. Furthermore, the
construction of the thinaerofoil 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 thinaerofoil interaction law is most suitable
for the modelling of aerofoil
ow.
Chapter 6
Results
6.1 Introduction
The two and (quasi)threedimensional viscous potential methods described in the earlier
chapters, using the quasisimultaneous viscousinviscid 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.
102
Chapter 6 Results
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
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_KUTTAJOUWKOWSKI
CL_FORCES
EXPERIMENT
0.4
0.2
0.005
0
0.005
0
10 12
ALPHA
14
16
18
20
0.2
1.4
1.6
103
1.8
0
2D VII RESULTS
1.6
2
1.4
LOG(ERROR)
CL
1.2
1
0.8
0.6
CL_KUTTAJOUWKOWSKI
CL_FORCES
EXPERIMENT
0.4
4
6
8
10
0.2
0
12
0
10 12
ALPHA
14
16
18
20
In the present method tripped transition occurs at a specic point causing the abrupt
change over in the skinfriction 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
KuttaJoukowski's formula (3.87), whereas the CL values calculated via integration of
the pressure and skinfriction 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 dierent 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 viscousinviscid 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
104
Chapter 6 Results
Experimental data of Hastings and Williams [41], obtained in the RAE lowspeed wind
tunnel, is used for the validation of the twodimensional 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 freeair conditions so the
experimental results have to be corrected for the eect 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
0.2
0.4
0.6
0.8
X/C
1.2
1.4
0.08
11
2D VII RESULTS
EXPERIMENT
0.07
2D VII RESULTS
EXPERIMENT
10
9
0.06
8
0.05
7
H
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
0.2
0.4
0.6
0.8
X/C
1.2
1.4
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
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
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 KuttaJoukowski's formula the lift coecient values are
overpredicted, especially near stall. The lift coecient results calculated by integrating the
pressure and skinfriction distribution around the aerofoil are in much better agreement.
It is noted that for the NACA0012 calculations, KuttaJoukowski'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 dierent from experiment, but is required to
avoid laminar boundarylayer 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 boundarylayer 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 twodimensional 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 dier. 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 skinfriction
coecient becomes zero, corresponding with the measurements.
107
1.4
0.003
QScoupling
SIcoupling
1.2
QScoupling
SIcoupling
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
1
x
1.5
1.4
0.004
QScoupling
SIcoupling
1.2
QScoupling
SIcoupling
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
1
x
1.5
1.5
0.0045
QScoupling
SIcoupling
QScoupling
SIcoupling
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
0.5
1
x
1.5
108
Chapter 6 Results
0.7
0.07
QScoupling
SIcoupling
QScoupling
SIcoupling
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
0.03
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
y/s
109
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
1.5
2.5
x
3.5
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 twodimensional
ow case. The results of a twodimensional
quasisimultaneous viscousinviscid interaction method are shown for comparison. The
twodimensional coupled results compare well with the results from the threedimensional
viscous potential
ow program. Only a slight dierence is visible in the results for the
shape factor, due to the dierent boundary conditions used at x = 1.
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 xdirection
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
1.5
2.5
x
3.5
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
1.1
0.0035
Streamwise skin friction coefficient
1.05
1
Velocity xdirection
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 skinfriction coecient for ow over a dent at Re = 11:5 106 .
112
Chapter 6 Results
20
gridline y= 0.00
gridline y= 0.30
gridline y= 1.35
0.03
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
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 .
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
=
=
=
=
=
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
0.8
Figure 6.34: Comparison streamwise momentum thickness at dierent 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 dierent spanwise stations for = 14o .
0.2
0.4
0.6
0.8
X/C
115
results closer to the tip. In gures 6.35 and 6.36 the shape factor and the streamwise
skinfriction 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 viscousinviscid iterations were required to converge, taking
70 minutes of CPU time.
s
1.5
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
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
0.2
0.4
0.6
0.8
X/C
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 dierent 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
viscousinviscid interactive and experimental results is good. However, towards the tip
problems were encountered with the inviscid
ow model due to inaccurate representation
of tip eects. No coupling therefore takes place for Y=Ys > 2:2, being the last 10 percent
of the span, and instead the boundarylayer values are determined by extrapolation.
Boundarylayer velocity proles have been measured for the same wing case at Re =
2:1 106 by Brebner and Wyatt [7] from which the boundarylayer integral thicknesses
are determined [69].
Figure 6.41 shows the midwing streamwise momentum thickness computed by the
threedimensional 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 boundarylayer growth compared with
experiment, especially near the trailing edge. Milewski achieved a similar overprediction
of the boundarylayer 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 eects 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 threedimensional 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 viscousinviscid 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 threedimensional wing
ow 4 hours of CPU time were required to converge.
117
0.008
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
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 XVELOCITY
ERROR YVELOCITY
ERROR ZVELOCITY
0.5
1
40
LOG(ERROR)
BETA, ALPHA+BETA
0.006
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
20
40
60
80
100
120
NUMBER OF VISCOUSINVISCID ITERATIONS
118
Chapter 6 Results
Chapter 7
Conclusions
In this thesis the quasisimultaneous viscousinviscid interaction method is investigated
both mathematically and numerically for two and threedimensional
ow cases. Moreover, the present research concentrates on gaining full understanding of the basic requirements for the quasisimultaneous method and on successfully developing two and
threedimensional quasisimultaneous viscousinviscid interaction methods. Conclusions
on these themes are presented in the next sections.
Boundarylayer discretisation
In chapter 2 the boundarylayer formulations for the various
ow cases are given. For the
threedimensional
ow case the boundarylayer equations are written in Cartesian coordinates, avoiding the cumbersome metric calculation required when using nonorthogonal
curvilinear coordinates. The threedimensional boundarylayer equations are discretised
using an upwind nitevolume scheme (taking into account the hyperbolicity of the system of equations) which works successfully in combination with the quasisimultaneous
method.
Quasisimultaneous interaction
In chapter 4 it is investigated to what extent the approximation of the external
ow, i.e.
the interaction law, can be simplied. Theorems are given from which it is clear what
the basic requirements are for the quasisimultaneous 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 odiagonals equal to zero and/or by increasing the diagonal, has favourable properties. The
convergence of the viscousinviscid iterations can be guaranteed, the convergence of the
GaussSeidel boundarylayer 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
viscousinviscid iterative method the convergence improves when the size of the interaction law increases. However, for the boundarylayer 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 odiagonals
to zero in order to have only a few viscousinviscid 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 odiagonals to zero in order to have only a few
boundarylayer iterations. In this way a fast procedure for the overall quasisimultaneous
scheme is obtained.
The developed theorems are evaluated for the calculation of twodimensional 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 boundarylayer 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 aected by it. Dening the unknowns in the endpoints of the
panels leads to an interaction law with favourable properties for the quasisimultaneous
interaction method. Other discretisations can result in less suitable interaction laws.
In chapter 5 various interaction laws are constructed for the case of twodimensional
aerofoil and threedimensional plate and wing
ow. For the twodimensional aerofoil
ow
in more detail various dierent 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 fullled 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 thinaerofoil theory interaction law is
very suitable for aerofoil viscousinviscid calculations. It can therefore be concluded that
the interaction law based on thinaerofoil 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 modied DERAVFP code quasithreedimensional results with a transonic
quasisimultaneous method are presented. The results are compared with results obtained
with the original semiinverse DERAVFP 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 quasisimultaneous scheme and with the use of the theoretical
results of chapter 4 the run times and robustness could be further improved.
Threedimensional results are obtained for dented plate and wing
ow calculations.
For the dented plate case a twosided 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 threedimensional method
with results obtained with a twodimensional method for quasitwodimensional
ow cases
good agreement is shown.
For the swept RAE101 wing case with
ow separation present, good convergence is
achieved with the threedimensional quasisimutaneous 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 thinaerofoil 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 threedimensional 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 threedimensional quasisimultaneous viscousinviscid interaction,
including separation. This is a noteworthy result as with the use of only a simple diagonal
interaction law, the quasisimultaneous method is just a small modication 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 ThreeDimensional 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 933462CP, 1993.
[4] Arthur, M.T. A Method for Calculating Subsonic and Transonic Flows over Wings
or WingFuselage Combinations with an Allowance for Viscous Eects. RAE TM
Aero 1996, 1984.
[5] Ashill, P.R. and Smith, P.D. An Integral Method for Calculating the Eects on
Turbulent BoundaryLayer Development of Sweep and Taper. RAE TR 83053,
1983.
[6] Ashill, P.R., Wood, R.F. and Weeks, D.J. An Improved, SemiInverse 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 74583, 1974.
[9] Carter, J.E. ViscousInviscid Interaction Analysis of Turbulent Separated Flow.
AIAA 811241, 1981.
[10] Carter, J.E. and Wornom, S.F. Solutions for Incompressible Separated Boundary
Layers Including ViscousInviscid Interaction, Aerodynamic Analysis Requiring Advanced Computers. NASA SP347, 1975.
[11] Catherall, D. and Mangler, K.W. The Integration of the TwoDimensional Laminar
BoundaryLayer 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 PostStall Behaviour
of Airfoils at Low Mach Numbers. AIAA Journal, 33(4):595{602, 1995.
124
BIBLIOGRAPHY
[13] Coenen, E.G.M. Modications to the DERA VFP Code to Include QuasiSimultaneous Coupling (QSVFP). Technical Report 9801, University of Bristol,
1998.
[14] Coenen, E.G.M. QuasiSimultaneous 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. QuasiSimultaneous ViscousInviscid 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. QuasiSimultaneous
ViscousInviscid Interaction for ThreeDimensional 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. ViscousInviscid Interaction
Method for Wing Calculations. In European Congress on Computational Methods
in Applied Sciences and Engineering, ECCOMAS 2000, Barcelona, Spain, 2000.
[18] Cousteix J. ThreeDimensional and Unsteady BoundaryLayer 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 ThreeDimensional Turbulent
BoundaryLayer Calculations and Separation Phenomena. AIAA 814201, 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 ThreeDimensional Turbulent
Boundary Layers Part II: AttachmentLine Flow on an Innite Wing. Aeronautical
Quarterly, 18:150{164, 1967.
[23] Dargel, G. and Thiede, P. Viscous Transonic Airfoil Flow Simulation by an Ecient
ViscousInviscid Interaction Method. AIAA 870412, 1987.
[24] Delery, J.M. and Formery, M.J. A Finite Dierence Method for Inverse Solutions
of ThreeDimensional Turbulent Boundary Layer Flow. AIAA 830301, 1983.
[25] Drela, M. TwoDimensional 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. ViscousInviscid 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 ThreeDimensional 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 LagEntrainment Method. RAE TR
77046, 1977.
[30] Edwards, D.E. Analysis of ThreeDimensional Flow Using Interacting BoundaryLayer Theory. In Smith, F.T. and Brown, S.N., editor, IUTAM Symposium on
BoundaryLayer Separation, pages 163{178. SpringerVerlag, 1987.
[31] Edwards, D.E. and Carter, J.E. Analysis of ThreeDimensional Separated Flow
with the BoundaryLayer Equations. AIAA 851499, 1985.
[32] Edwards, D.E., Carter, J.E. and Smith, F.T. Analysis of ThreeDimensional Separated Flow with the BoundaryLayer 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 Eects. RAE TM Aero 1861, 1980.
[34] Firmin, M.C.P. Applications of RAE Viscous Flow Methods near Separation Boundaries for ThreeDimensional 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. AGARDAR255, 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 Congurations, Bad Harzburg, DGLR Paper 78064, 1978.
[37] Goldstein, S. On Laminar Boundary LayerFlow 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 LagEntrainment 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 BoundaryLayer 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 SemiEmpirical 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. AGARDCP168, 1975.
[48] Houwink, R. and Veldman, A.E.P. Steady and Unsteady Flow Computations for
Transonic Airfoils. AIAA 841618, 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. LowSpeed Aerodynamics; From Wing Theory to Panel
Method. McGrawHill, 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 MultiElement Airfoil Congurations. AIAA 910177, 1991.
[54] Lagerstrom, P.A. Solutions of the NavierStokes 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 ThreeDimensional Flows by
ViscousInviscid Interaction Using the 'MZM' Method. AGARD CP412, 1986.
[56] Le Balleur, J.C. Couplage VisqueuxNon Visqueux: Analyse du Probleme Incluant
Decollements et Ondes de Choc. La Recherche Aerospatiale, 6:349{358, 1977.
[57] Le Balleur, J.C. Couplage VisqueuxNon Visqueux: Methode Numerique et Ap
plications aux Ecoulements
Bidimensionnels Transsoniques et Supersoniques. La
Recherche Aerospatiale, 2:65{76, 1978.
BIBLIOGRAPHY
127
128
BIBLIOGRAPHY
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. ViscousInviscid Interaction for Transonic WingBody Congurations
Including Wake Eects. 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 ThreeDimensional 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 BoundaryLayer Equations in ThreeDimensional Viscous Flow Simulation. AGARDCP412, 1986.
[94] Van der Wees, A.J. and Van Muijden, J. A Robust QuasiSimultaneous Interaction Method for a Full Potential Flow with a Boundary Layer with Application to
Wing/Body Congurartions. 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
ViscousInviscid Interaction Solver for Transonic Flow about Wing/Body Congurations on the Basis of Full Potential Theory. AIAA 933026, 1993.
[96] Varga, R.S. Matrix Iterative Analysis. PrenticeHall, 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, QuasiSimultaneous Method to Calculate Interacting Boundary Layers. AIAA Journal, 19:79{85, 1981.
[99] Veldman, A.E.P. Strong ViscousInviscid Interaction and the Eects 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, SpringerVerlag,
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 BoundaryLayer 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. McGrawHill, 1991.
[105] Wigton, L. and Yoshihara, H. ViscousInviscid 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 ViscousInviscid Interaction Method. Aeronautical Journal, 89:185{197, 1985.
[107] Williams, B.R. ViscousInviscid 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 840265, 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
boundarylayer integral thicknesses
Streamline coordinate denitions
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 dened 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
1
( u
e qe e e
u) dz;
y =
1
( v
e qe e e
v ) dz;
132
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 denitions 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.
x = s cos n sin ;
y = s sin + n cos ;
xy = (ss
yx
yy
Appendix B
Boundarylayer closure relations
B.1 Streamwise closure
In the paper by Head [42] the following shape factors have been dened:
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 dened as
2 @u
2 ss s :
@s
>
:
(B.1)
(B.2)
(B.3)
If Thwaites' method is not used for the laminar boundarylayer 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
for H 4:02923;
for H > 4:02923:
(B.4)
134
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 shockinduced separation [48]. This same idea can be used
for severe trailingedge 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
which gives the following relationship for H1 , to be used for the closure of integral thicknesses in threedimensional
ow:
2H
H1 =
:
(B.8)
(H 1)
(B.9)
135
and together with (B.8) the boundarylayer 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 prole [83], the above given Mager's prole allows for
zero skinfriction and can be used in the wake region.
A setback of using Mager's cross
ow velocity prole 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 prole [21], which has as yet not been implemented.
8
>
>
>
>
<
2
(0:22 + 1:572
Re
>
>
>
>
:
2
Re
ss
ss
1:822);
0:0182
0:22 + 1:4022 +
;
2 + 0:107
(B.14)
0:1 < 2 < 0;
Turbulent
The expression used for the turbulent streamwise skinfriction coecient, Cf s , used in
both the twodimensional and threedimensional calculations, follows the paper by Green
et al. [40]
Cf s
0:9
= Cf 0
H=H0 0:4
0:5 ;
(B.15)
136
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 skinfriction (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)
Cf n = Cf s tan :
(B.17)
(B.18)
Appendix C
Derivation Cartesian integral
boundarylayer equations
C.1 Boundarylayer equations
The
ow of a viscous
uid is represented by the NavierStokes 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)
@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
1 @p
;
@x
1 @p
;
@y
1 @p
;
@z
(C.4)
(C.5)
(C.6)
138
With the assumptions that the freestream Reynolds number Re is large and the characteristic boundarylayer thickness small compared to the characteristic chord length of the
prole, Prandtl's boundarylayer equations are derived from the NavierStokes 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.
@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)
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 boundarylayer integral thicknesses as dened in appendix A, using
= 1, nally gives the integral xmomentum boundarylayer 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 boundarylayer 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
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 xdirection over distance x therefore becomes
@
y x
@x
udz:
(C.20)
140
z
y
@
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)
CE
1 @
=
qe @x
1 @
udz +
qe @y
vdz;
(C.23)
which with the use of the boundarylayer integral thicknesses, dened in appendix A, can
be rewritten to the integral boundarylayer 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 boundarylayer thickness.
Appendix D
Auxiliary denitions and theorems
Denition D.1 (HornJohnson [45, p. 92]) A matrix A is called positive stable if all the
eigenvalues of A full <() > 0.
Denition 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)
jAiij
n
X
k=1
k =i
jAik j
for i = 1; 2; : : : n;
(D.2)
142
Denition 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 Mmatrix.
Denition D.11 (Young [110, p. 21]) A matrix A is positive denite 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 denite.
Theorem D.13 (HornJohnson [45, p. 95])) Let Q be a positive real matrix (i.e. the symmetric part of Q is positive denite), then for every positive denite 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.
143
and A 1 0 then
(M 1 N ) =
(A 1 N )
< 1:
1 + (A 1 N )
(D.4)
i =
j
=1; j 6=i
jAij j;
i = 1; 2; ; n:
(D.6)
jz Aiij
i ;
i = 1; 2; ; n:
(D.7)
144
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 NavierStokes vergelijkingen die
zijn gebaseerd op de behoudswetten van massa, impuls en energie. De NavierStokes 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 NavierStokes
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 vleugelproel. Een aan populariteit winnende manier is het met brute rekenkracht
oplossen van de NavierStokes vergelijkingen. Hoewel NavierStokes 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/nietviskeuze interactie die kan worden gebruikt voor vloeistofstromingen met lage viscositeit (stroperigheid, diusie). 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 eecten
verwaarloosbaar en kan worden verondersteld dat de stroming niet viskeus is. In tegenstelling tot NavierStokes simulatie benutten viskeuze/nietviskeuze interactiemethodes deze
informatie door gebruik te maken van een zogeheten zonetechniek. Het gehele stromingsgebied is opgedeeld in twee zones: een buitengebied waar de nietviskeuze 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 NavierStokes vergelijkingen. Om
de oplossing te vinden voor het gehele stromingsprobleem moeten de oplossingen eerst
in beide zones worden berekend en dan worden gekoppeld. Viskeuze/nietviskeuze interactietechnieken zijn zeer succesvol gebleken voor praktische tweedimensionale 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/nietviskeuze 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/nietviskeuze interactiemethodes ontwikkeld om het systeem van grenslaagvergelijkingen en nietviskeuze 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 quasisimultane methode de beste
prestaties op het gebied van snelheid en robuustheid in twee dimensies. De quasisimultane
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 nietviskeuze stroming in het buitengebied.
De benadering van de nietviskeuze 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 quasisimultane viskeuze/nietviskeuze interactiemethode.
Voor driedimensionale viskeuze/nietviskeuze interactie worden dezelfde voordelen
verwacht die zijn verkregen in twee dimensies. Verschillende pogingen zijn ondernomen
om te komen tot een viskeuze/nietviskeuze interactiemethode voor driedimensionale 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 driedimensionale viskeuze/nietviskeuze
interactiemethode te ontwikkelen, is in dit proefschrift de quasisimultane methode gebruikt. Van de verschillende interactiemethodes levert de quasisimultane 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 driedimensionale berekeningen moeilijker en vereisen ze meer rekentijd dan tweedimensionale berekeningen. Om een driedimensionale 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 tweedimensionale
quasisimultane 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/nietviskeuze quasisimultane interactiemethode te optimaliseren. De nieuwe wiskundige theorie wordt numeriek geevalueerd voor het geval van
tweedimensionale trogstroming, en de theorie en de numerieke resultaten blijken consistent.
Om de ontwikkelde quasisimultane viskeuze/nietviskeuze interactiemethode verder
te testen, worden realistische twee en driedimensionale aerodynamische stromingsproblemen gemodelleerd onder verschillende stromingscondities. Om de volledige driedimensionale 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 quasisimultane interactiemethode, en
richtlijnen worden gegeven voor de optimalisatie van het koppelingsproces van de grenslaag
met de nietviskeuze buitenstroming. Twee en driedimensionale quasisimultane 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 driedimensionale quasisimultane
viskeuze/nietviskeuze 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 socalled NavierStokes equations which are based on the
conservation laws of mass, momentum and energy. The NavierStokes 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 NavierStokes 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 timeconsuming, as one can imagine in the case of windtunnel tests for aircraft conguration 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 NavierStokes
simulation which is being applied more and more these days. Whilst NavierStokes simulation potentially oers 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 viscousinviscid interaction which can be employed for
uid
ows with small viscosity (syrupiness, diusion). For the
ow around an aeroplane the
viscous forces are small and they are conned to a thin layer close to the surface. Outside
this thin (boundary) layer the viscous eects are negligible and the
ow can assumed to
be inviscid (nonviscous). Unlike NavierStokes simulation, viscousinviscid interaction
methods intelligently take this information into account, and make use of a socalled
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
boundarylayer equations are valid. Both sets of equations in the viscous and inviscid
ow
regions are computationally cheap, yet suciently accurate, simplications 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. Viscousinviscid
interaction techniques have proven to be very successful for practical twodimensional
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. Viscousinviscid
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 boundarylayer concept was rst introduced by Prandtl in 1904, several
viscousinviscid interaction methods have been developed to solve the system of boundarylayer and external inviscid
ow equations. The oldest and bestknown is the socalled
direct method which solves the boundarylayer 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 takeo, the direct boundarylayer
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 quasisimultaneous method gives the best performance in terms of speed and robustness in two dimensions. The quasisimultaneous
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 boundarylayer 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 quasisimultaneous viscousinviscid interaction
method.
Similar advantages as obtained in two dimensions, are expected for threedimensional
viscousinviscid interaction. Several eorts have been directed towards the development
of a viscousinviscid interaction method for threedimensional
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 goodworking threedimensional
viscousinviscid interaction method, the quasisimultaneous technique is employed in this
thesis. Of the various interaction methods the quasisimultaneous 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 threedimensional calculations are dicult and require much more computation time than twodimensional calculations. In order to make a threedimensional
viscousinviscid 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 twodimensional quasisimultaneous 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
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 oered 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 viscousinviscid interaction. I would also like to say thanks to Nigel Taylor for having me as his housemate
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 threedimensional 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 nietkameel rijden in Tunesie. Twee goede
vriendinnen, en tevens mijn paranimfen, studiegenootjes, medekenners 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.