Journal of Computational Physics 172, 119–165 (2001)

doi:10.1006/jcph.2001.6817, available online at http://www.idealibrary.com on

A Novel Velocity–Vorticity Formulation of the

Navier–Stokes Equations with Applications to
Boundary Layer Disturbance Evolution
Christopher Davies∗ and Peter W. Carpenter†
∗ School of Mathematical and Information Sciences, Coventry University, Priory Street, Coventry CV1 5FB,
United Kingdom; and †School of Engineering, University of Warwick, Coventry CV4 7AL, United Kingdom
E-mail: [email protected]

Received November 27, 2000; revised April 23, 2001

A novel velocity–vorticity formulation of the unsteady, three-dimensional, Navier–

Stokes equations is presented. The formulation is particularly suitable for simulating
the evolution of three-dimensional disturbances in boundary layers. A key advantage
is that there are only three governing equations for three primary dependent variables.
Another advantage is that no wall boundary conditions are needed for the vorticity.
Instead the conditions placed on the velocity are linked to the vorticity field through
integral constraints based on the definition of vorticity.
Numerical methods are presented in the context of application to the three-
dimensional boundary layer over a rotating disc. The discretization scheme uses
spectral expansions in the wall-normal and azimuthal (or spanwise) directions and
compact finite differences in the radial (or streamwise) direction. The scheme is
implemented so that the advantages of spectral convergence can be combined with
the use of an efficient line-iteration solution procedure. The linearized form of the
new velocity–vorticity method is validated for the case of convective instabilities
evolving over both rigid and compliant discs. °c 2001 Academic Press

1. INTRODUCTION

Velocity–vorticity methods have many attractive features as a basis for calculating un-
steady flow fields. Accordingly, it is all the more unfortunate that the methods currently
in use generally suffer from two major drawbacks. Firstly, for unsteady three-dimensional
problems they require six dependent variables which compares poorly with the four required
in formulations using primitive variables. Secondly, the lack of a boundary condition on
vorticity can make it necessary to adopt relatively elaborate schemes in order to ensure that
the computed velocity and vorticity fields are divergence-free. We will show that both of
these drawbacks can be overcome for an important class of problems.
119
0021-9991/01 $35.00
Copyright ° c 2001 by Academic Press
All rights of reproduction in any form reserved.
120 DAVIES AND CARPENTER

Our novel formulation stems from a desire to develop an efficient, yet versatile, method
for simulating the evolution of three-dimensional perturbations in boundary layers. We
are particularly interested in simulating passive and active flow control; for example, the
use of finite-length compliant panels [14] and MEMS and micro actuators in the form of
deformable bumps or jets issuing from orifices in the wall [8, 9, 11, 32, 33, 50]. These
are highly elliptic problems featuring rapid change in the streamwise direction and also
local upstream propagation of disturbances. This latter feature is also found when there are
absolute instabilities. We wished to develop a computational method that could be used for
full direct numerical simulations but, at the same time, would also be suitable for integrating
the linearized Navier–Stokes equations. In this respect it would emulate the PSE (Parabolic
Stability Equation) approach developed by Bertolotti et al. [4] (see, also, [31]). However,
unlike the PSE approach, our method is fully elliptic and no restrictions are placed on the
form of the disturbance. We do not, however, expect our method to compete in terms of
computational efficiency for problems which are suitable for using the PSE approach. That
said, we can see no reason why our velocity–vorticity formulation could not be used as the
basis of a three-dimensional PSE scheme.
Fasel [22], Speziale [57], and Gatski [25] review the advantages of velocity–vorticity
methods. One of the main advantages is the elimination of the pressure by using the Navier–
Stokes equations in the form of a vorticity transport equation

DΩ
= Ω · ∇U + ν∇ 2 Ω, (1)
Dt

where D/Dt denotes the material derivative, Ω and U denote, respectively, the three-
dimensional vorticity and velocity vector fields, and ν is the kinematic viscosity. Thus,
only three dependent variables need to be updated in time, as compared with four when
using primitive variables. Equation (1) is supplemented by the continuity equation and the
definition of vorticity, i.e.,

∇ · U = 0, Ω = ∇ × U. (2a,b)

Here Eq. (2a) is written in the form used for incompresssible flow. In principle, Eqs. (1),
(2a) and (2b) are sufficient for determining the velocity and vorticity fields. However, many
velocity–vorticity methods use a Poisson equation, obtained by taking the curl of Eq. (2b)
and making use of (2a) to determine the velocity from the vorticity. This Poisson equation
takes the form:

∇ 2 U = −∇ × Ω. (3)

Gatski [25] classified velocity–vorticity methods proper into two basic categories, both
of which use the vorticity transport equation (1) to update the vorticity but differ in how
the velocity is then subsequently determined from the vorticity field: (1) Method A solves
Eqs. (2a) and (2b) sequentially or simultaneously for U—examples include the studies
made by Gatski et al. [26, 27], Koh and Bradshaw [41], and Bertagnolio and Daube [3]; and
(2) Method B solves the Poisson equation (3) for U—examples include the works of Fasel
[21, 22], Dennis and Quartapelle [20], Fasel and Konzelmann [23], Fasel et al. [24], Kral
and Fasel [42], Kloker et al. [39], Rempfer and Fasel [52, 53], Rist and Fasel [54], Wu, Wu,
and Wu [61], Trujillo and Karniadakis [59], and Meitz and Fasel [49].
 VELOCITY–VORTICITY FORMULATION 121

As mentioned above, difficulties can arise with velocity–vorticity methods, in ensuring

that both Eqs. (2a) and (2b) are properly satisfied by the numerical solution. For example, in
the case of Method B, although Eq. (2a) is used to obtain Eq. (3), a divergence-free velocity
field is not necessarily guaranteed. However, it can be shown [22] that provided Eq. (2a)
is satisfied on the boundary of the computational domain, it will be satisfied in the interior
(in practice this implies that the error is maximal on the boundary). It can also be shown
that provided the computed vorticity field is solenoidal, the velocity field computed using
Eq. (3) will satisfy Eq. (2b) everywhere in the domain, provided that Eq. (2b) is satisfied
on the boundary. These considerations led Fasel and his coworkers to formulate numerical
boundary conditions in terms of the first and second derivatives of the perturbation vorticity
components. This approach usually works well in practice, but, as will be discussed in more
detail below, it can lead to difficulties in the case of parallel or near-parallel undisturbed
flows.
Gresho [28, 29] and more recently Weinan and Liu [60] critically review and analyze
the various forms of the vorticity boundary condition used by previous authors. A common
view has developed over the past ten years or so that any vorticity boundary condition has
to be global in the sense that the values at all points on the boundary should be linked.
However, Weinan and Liu show that in the context of finite-difference schemes, at least,
many of the newly developed global boundary conditions are actually equivalent to older
local formulae. They thereby cast serious doubt on the usefulness of these global boundary
conditions since they are much more complicated to implement than local ones. Our view,
as will be seen, is that an integral constraint is a far more satisfactory and simpler, truly
global, alternative to a boundary condition.
Many authors (e.g., Gatski et al. [26, 27], Wu, Wu, and Wu [61] and Bertagnolio and
Daube [3]) decompose their computed vorticity field Ωc , which is not divergence-free in
general, into a solenoidal ∇φ and rotational part Ω. It therefore follows that

∇ 2 φ = ∇ · Ωc . (4)

Equation (4) can then be solved for ∇φ which is used to project the computed Ωc field onto
a divergence-free field,

Ω = Ωc − ∇φ. (5)

Wu, Wu, and Wu [61] established the conditions for optimum projection and also showed
that it is possible to solve Eq. (3) for a divergence-free velocity field using an arbitrary non-
solenoidal Ωc in place of Ω. They showed that the projection can be carried out relatively
cheaply without the need for a time-consuming inversion of a Poisson equation. Recently,
Bertagnolio and Daube [3] generalized this approach of a Helmholtz decomposition com-
bined with projection onto a space of divergence-free vectors to generalized curvilinear
coordinates.
In any discussion of vorticity boundary conditions, it is important to appreciate that
the only fundamental boundary conditions are those that must be imposed on the velocity
field. (See [28, 29] for a detailed discussion of this point.) When boundary conditions on
the vorticity are derived and utilized within a particular formulation of the Navier–Stokes
equations, they cannot provide any constraints that are genuinely additional to those that are
required for the velocity. For instance, in the velocity–vorticity formulation adopted by Fasel
122 DAVIES AND CARPENTER

et al. [24] the conditions imposed on the velocity field provide the boundary conditions for
Poisson-type equations that relate the velocity field to the vorticity field. Exactly the same
information is employed, a second time, to obtain boundary conditions for the vorticity.
These latter conditions are derived by simply substituting the boundary conditions imposed
on the velocity field into algebraic relations that follow from the Poisson equations, the
vorticity definition, and the continuity equation.
Because there is no general requirement to impose any constraint on the vorticity that
cannot be traced back to the conditions imposed on the velocity field, the apparent necessity
for vorticity boundary conditions is often an artifice of the chosen numerical method. Indeed,
as pointed out by Gresho, nearly thirty years ago Davis [17] formulated a streamfunction–
vorticity method based on finite differences that implicitly discarded the need for any
boundary conditions on vorticity. A little later, Barrett [2] and Campion–Renson and Crochet
[7] independently also proposed streamfunction–vorticity methods, this time based on finite
elements, in which no boundary condition is imposed on vorticity. In effect, in such schemes
the vorticity is constrained or determined by some sort of integral relation, or its equivalent,
over the computational domain which relates the unknown value of the vorticity at the
boundary to its values in the interior without any need to specify the former. An approach
of this kind has been proposed by Dennis and Quartapelle [20] who use Green’s theorem
to generate a set of integral conditions of this sort. Dennis and Hudson [18, 19] propose
a hybrid of Gatski’s Methods A and B which for two-dimensional flows solves one of the
Poisson equations (3) for one of the velocity components and then integrate the continuity
equation to solve for the other. Shen and Loc [56] have proposed a three-dimensional version
of this approach.
Perhaps the most straightforward example of the use of an integral constraint on vorticity
is provided by Guevremont et al. [30]. They impose no boundary conditions on vorticity
in their two-dimensional finite-element method. Instead, a weighted integral relation is im-
posed which is based on the definition of vorticity and evaluated over the entire domain.
Our approach bears some similarity to that of Guevremont et al., although it was devel-
oped in ignorance of their work. It is a hybrid of Methods A and B. We solve the Poisson
equation (3) for the normal velocity component and use the definition of vorticity (2b) to
link the surface values of the other two velocity components to integrals of the vorticity
components. The two-dimensional version was given by Davies and Carpenter [14]. Thus,
this method replaces boundary conditions with integral conditions on vorticity which link
it to the boundary conditions on velocity on the solid surface in a natural and very straight-
forward way. In many cases, of course, it is simply the no-slip condition that is imposed
on the solid surface. This approach provides a relatively simple, but rigorous, method for
ensuring a divergence-free velocity field and finding a solution to the vorticity field that
satisfies Eq. (2b).
We will now discuss the principal disadvantage of the current three-dimensional velocity–
vorticity methods, namely that it appears to be necessary to solve the three transport equa-
tions (1) for the vorticity components, as well as the three Poisson equations (3), or their
equivalents, for the velocity components. In contrast, for two dimensions the velocity–
vorticity formulation can readily be reduced to two governing equations. An obvious way
to achieve this is to use a streamfunction–vorticity formulation such as, for example, Dennis
and Quartapelle [20]. It is also possible to retain the velocity–vorticity formulation and elim-
inate the streamwise velocity component between the continuity equation (2a) and the def-
inition of vorticity (2b). This approach was followed independently by Koh and Bradshaw
 VELOCITY–VORTICITY FORMULATION 123

[41] and Davies and Carpenter [14]. Wu, Wu and Wu [61] point out that in three dimensions
the continuity equation (2a) can be used to determine the third component of velocity rather
than solving a third Poisson equation. Adopting this procedure can lead to considerable
savings in CPU time, especially if projection onto a divergence-free space is required. Nev-
ertheless, our intuition suggested that it should be possible to carry out three-dimensional
calculations with only three (primary) governing equations. And, indeed, in the present
paper we show how it is possible for a wide class of problems to derive a velocity–vorticity
formulation which is based on only three primary dependent variables.
In fact, Kim, Moin, and Moser [37] have already shown that for homogeneous flow
in a planar channel, at least, it is feasible to employ a velocity–vorticity formulation that
involves only two primary dependent variables and two evolution equations. One evolution
equation determines the Laplacian of the normal velocity component, while the other is just
the usual transport equation for the normal component of the vorticity. The remaining two
velocity components are obtained by applying the continuity equation and the definition
of the normal component of the vorticity. For the doubly spatially periodic computational
domain appropriate to homogeneous channel flow, the construction of the spanwise and
streamwise components of the velocity from the normal components of the velocity and
the vorticity may be accomplished in a direct and efficient manner, using relatively simple
manipulations of Fourier series representations for the flow-variables. However, it is not
obvious that a similarly efficient numerical method could be implemented for more general
flow configurations that are not periodic along two spatial directions, or geometrically
simple in some other fashion.
A distinctive, but potentially disadvantageous, feature of the formulation adopted by
Kim, Moin, and Moser is that it involves a fourth-order evolution equation for the normal
velocity component. This apparent drawback was successfully dealt with by introducing an
intermediate variable within a numerical scheme that split the fourth-order equation into a
pair of coupled second-order equations. The first of these split equations may be classified
as a vorticity-type transport equation, while the second takes the simpler form of a Poisson
equation. Thus, the system of governing equations used by Kim, Moin, and Moser consisted,
in practice, of two vorticity-type transport equations and a single Poisson equation. We
shall develop a velocity–vorticity formulation that involves a similar system of governing
equations. However, our formulation would appear to be less restrictive with regard to the
flow and the bounding surface geometry. Moreover, the ease of its numerical implementation
is not fine-tuned to any presumed streamwise periodicity of the computational domain. We
shall show, in particular, that our velocity–vorticity formulation readily lends itself to the
efficient simulation of disturbances in spatially inhomogeneous flows.
In our standard formulation, the dependent variables are perturbations, which are finite in
general, to a known undisturbed flow field. The geometry is relatively simple, as illustrated
in Fig. 1. We envisage a three-dimensional flow field over a flat or slightly curved solid
surface. The solid surface need not be completely rigid. Part of it may be replaced by a
compliant surface or some other sort of interactive device, such as a suction hole or slot, or
a microjet-type actuator.
The three primary dependent variables are the two perturbation vorticity components
in the plane of the solid surface and the perturbation normal velocity component. These
are governed by two vorticity transport equations and a Poisson equation. The secondary
dependent variables are the remaining vorticity and velocity components and pressure.
The distinction between primary and secondary dependent variables is that the latter can
124 DAVIES AND CARPENTER

FIG. 1. Arbitary three-dimensional disturbance in a three-dimensional flow over a solid surface.

be determined explicitly entirely from the current values of the former. Accordingly, the
secondary variables can, in principle, be eliminated from consideration and storage is not
required for them. We will show that, provided the dependent variables can be forced to
satisfy certain rather general conditions at infinity, this formulation is fully equivalent to
the Navier–Stokes equations in primitive-variable form.
Owing to our interest in simulating the evolution of perturbations in boundary layers,
the velocity–vorticity approach of Fasel and his coworkers, was an attractive starting point.
In many respects, despite the evident differences, our method can be regarded as a natu-
ral development from theirs. It is obviously computationally more advantageous to work
with only three primary governing equations. A further, less-evident advantage of our for-
mulation is that it allows spectral discretization in the normal, as well as the spanwise,
directions. Finite differences are used only for the streamwise direction. Furthermore, it
is possible to retain simple pentadiagonal and tridiagonal matrix schemes even with our
spectral discretization. This feature leads to additional computational efficiency.
The present paper is mostly devoted to describing the formulation and other theoretical
and numerical aspects of our velocity–vorticity method. But we also provide evidence of
its practical utility by describing a numerical study of the convective instabilities of the
three-dimensional boundary layer over a rotating disc. The study includes cases where a
compliant annulus is inset into the disc’s surface. For this particular application, we consider
small-amplitude disturbances and employ a linearized form of the governing equations. An
overview of the literature on the rotating disc boundary layer is given in the reviews by
Reed and Saric [51] and Cooper and Carpenter [12]. Here we will only provide a minimal
account that is sufficient to motivate our numerical study.
The standard linear stability analysis for the rotating disc boundary layer is based on a
sixth-order system of ordinary differential equations [46, 48]. The results of this analysis are
well-documented and can be used to provide validation for our computer code. However,
the standard stability analysis neglects the slow variation of the flow variables in the radial
direction. Our method, being based on the complete linearized governing equations, does
not need to make any such assumption. Accordingly, we can study the effects of including
the so-called “nonparallel terms.” It turns out that they have only a slight effect on the
convective instabilities. This is in accord with the results of an earlier study by Malik and
Balakumar [47] using the PSE approach.
Three distinct families of eigenmodes have been identified for the rotating-disc boundary
layer. Recently Lingwood [43, 44] showed that the Type I mode (which is essentially an
 VELOCITY–VORTICITY FORMULATION 125

inviscid inflexion-point instability) coalesces with the, hitherto obscure, Type III mode to
form an absolute instability. She also noted that the Types I and II modes could coalesce.
(The Type II is a viscous mode and is destabilized by the effects of Coriolis acceleration.)
In this case, the coalescence first occurs when the individual eigenmodes are convectively
stable. Koch [40] suggested that, since such a coalescence, which he termed direct resonance,
corresponds to a double root, the Cauchy residue theorem could be used to evaluate the
form of the coalesced mode. He thereby showed that algebraic growth would occur. For
sufficiently high levels of background noise, algebraic growth could provide a route to
transition. Our numerical simulations were successful in confirming the existence of the
algebraic growth, thus providing a particularly stringent validation of the computer code.
However, in this instance the effects of the so-called “nonparallel” terms were found to be
much more significant. The “nonparallel” simulations suggest that, in practice, for a rigid
disc the algebraic growth would be dominated by the onset of convective instability slightly
further outboard.
A separate paper will be devoted to a study of the implications of the absolute instability
for global behavior. It suffices here to note that such a study has, in fact, been successfully
undertaken [16] using our new velocity–vorticity formulation and discretization scheme
to conduct an appropriate set of numerical simulations in an efficient manner. The results
of these simulations appear to suggest that the slow radial flow variation does have an
important effect, which leads to a qualitative change in the global behavior of the rotating-
disc boundary layer, compared with Lingwood’s [43] local “parallel-flow” theory.
The remainder of the present paper is set out as follows. Section 2 describes our velocity–
vorticity formulation of the Navier–Stokes equations for the general case. In Section 3, we
adapt the formulation for the rotating-disc flow. The numerical methods are described in
Section 4. The results of a study of the convective instabilities of the rotating-disc boundary
layer obtained with our velocity–vorticity method are presented and discussed in Section 5.
Finally, the conclusions are given in Section 6.

2. REFORMULATION OF THE NAVIER–STOKES EQUATIONS

In this section we describe the new velocity–vorticity formulation of the Navier–Stokes

equations and demonstrate its equivalence to the usual formulation in terms of primitive
variables. We suppose that there is an incompressible fluid flow in the semi-infinite domain
z ∈ [η, ∞), where z = η(x, y, t) defines the location of a solid, but possibly nonstationary
or porous boundary to the flow domain. For the case where the boundary is a flat rigid wall,
we will set η = 0. Consequently, z will be referred to as the wall-normal coordinate. It is
assumed that there is a known undisturbed-flow solution represented by the velocity and
vorticity fields
¡ ¢
U B = (U B , V B , W B ), Ω B = ΩxB , Ω yB , ΩzB .

The total velocity and vorticity fields are then decomposed as

U = U B + u, Ω = Ω B + ω, (6a,b)

where

u = (u, v, w), ω = (ωx , ω y , ωz )

126 DAVIES AND CARPENTER

represent perturbations from the prescribed undisturbed flow. We will consider the governing
equations for perturbations rather than the total flow-variables, since this yields a more
convenient formulation for present purposes. (If so desired, the undisturbed flow can be
determined by making slight modifications to the governing equations for the perturbations.
Details are included Appendix A, together with some comments about how the formulation
could be amended when there is an upper physical boundary located at finite z.) The
assumption of a Cartesian coordinate system provides for an economic exposition, but is
not essential. Cylindrical polar coordinates will be employed for the rotating-disc problem.
We first divide the components of the perturbation flow-fields u, ω into two sets. The
components {ωx , ω y , w} will be referred to as primary variables while the remaining com-
ponents {u, v, ωz } will be identified as secondary variables. The primary variables are
obtained by projecting the velocity field along the wall-normal direction and projecting
the vorticity field on to the orthogonal plane, which is parallel to the wall. The order of
combination of these projections with the velocity and vorticity fields is reversed in the
case of the secondary variables. The distinction between primary and secondary variables
is made because the latter may be defined explicitly in terms of the former, and thus, in
principle, eliminated completely from consideration.
From this point on, unless explicitly stated otherwise, all variables will be dimensionless.
We will show that the evolution of the three dimensionless primary variables {ωx , ω y , w}
may be determined using just three equations, namely

∂ωx ∂ Nz ∂ Ny 1
+ − = ∇ 2 ωx (7)
∂t ∂y ∂z R
∂ω y ∂ Nx ∂ Nz 1
+ − = ∇ 2ωy (8)
∂t ∂z ∂x R
∂ωx ∂ω y
∇ 2w = − , (9)
∂y ∂x

where N = (N x , N y , Nz ) is defined as

N = Ω B × u + ω × U B + ω × u. (10)

The Reynolds number R is defined in the usual manner, using the appropriate velocity and
length scales chosen for the nondimensionalization. The two equations (7) and (8) are the
transport equations for the x and y components of the vorticity. The form for N shows that the
full nonlinearity of the transport equations has been retained. For the rotating-disc numerical
simulations, a linearization was performed by dropping the final term in Eq. (10). It should
be stressed that the linearization, useful though it is for studies of the evolution of small-
amplitude disturbances, is not an inherent feature of our velocity–vorticity formulation.
Equation (9) can be obtained by taking the wall-normal component of the relationship

∇ 2 u = −∇ × ω,

which is the perturbation form of Eq. (3).

Clearly, the convective quantity N appearing in the transport equations can only be fully
evaluated if the secondary variables {u, v, ωz } are known. Accordingly we make the explicit
 VELOCITY–VORTICITY FORMULATION 127

definitions
Z ∞µ ¶
∂w
u=− ωy +
dz (12)
z ∂x
Z ∞µ ¶
∂w
v= ωx − dz (13)
z ∂y
Z ∞µ ¶
∂ωx ∂ω y
ωz = + dz. (14)
z ∂x ∂y

The first two of these definitions can be obtained by integrating the appropriate components
of the definition (2b) of vorticity with respect to z. It is assumed that both u and v vanish at
infinity. The definition of ωz follows from integration of the condition that the vorticity be
solenoidal. With the definitions (12)–(14) of the secondary variables, the quantity N can be
written explicitly in terms of the primary variables. The two transport equations (7) and (8),
together with the Poisson equation (9), can thus be viewed as a system of three governing
equations for the three unknown primary variables; all reference to the secondary variables
may be eliminated.
Thus far, we have indicated how three differential equations and three explicit definitions
can be derived for the six components of the velocity and vorticity perturbation fields. We
now show that, from such a limited basis, it is possible to recover the full Navier–Stokes
equations in primitive-variables form, provided two further conditions are satisfied for the
behavior of the perturbations as z → ∞.

2.1. Equivalence to the Primitive Variables Formulation

Differentiating Eqs. (12) and (13) with respect to z we obtain the relations:

∂w ∂v ∂u ∂w
ωx = − , ωy = − . (15a,b)
∂y ∂z ∂z ∂x

Substituting Eqs. (15a) and (15b) into Eq. (14) and carrying out the integration gives

∂v ∂u
ωz = − . (16)
∂x ∂y

Thus, the secondary variables {u, v, ωz } are defined so that the vorticity does, in fact, satisfy
the usual definition (2b) of vorticity. The solenoidal property of the vorticity perturbation
field

∇ · ω = 0.

follows immediately. (It also follows automatically from the definition of the secondary
variable ωz given in Eq. (14).)
If we differentiate the definition (12) of u with respect to x, the definition (13) of v with
respect to y and then add we obtain
Z ∞µ ¶
∂u ∂v ∂ω y ∂ωx ∂ 2w ∂ 2w
+ =− − + + dz. (17)
∂x ∂y z ∂x ∂y ∂x2 ∂ y2
128 DAVIES AND CARPENTER

Substituting for the integrand using the Poisson equation (9) then gives

∂u ∂v ∂w ∂w
+ + = lim . (18)
∂x ∂y ∂z z→∞ ∂z

It follows therefore that we can recover the incompressibility condition (2a), provided that

∂w
lim = 0. (19)
z→∞ ∂z

This does not represent an additional constraint so far as the primitive-variables formulation
of the Navier–Stokes equations is concerned. It is satisfied automatically if incompressibility
is enforced and the u, v components of the perturbation velocity vanish as z → ∞. However,
in our formulation, it would appear that the condition (19) needs to be imposed directly. We
will show that, in practice, this presents no difficulty, provided an appropriate coordinate
transformation is utilized to map the physically semi-infinite domain in the z-direction on
to a finite computational domain.
Differentiating the vorticity transport equations (7) and (8) with respect to x, y, respec-
tively, adding the results, and then using the solenoidal property of the perturbation vorticity
field gives
µ ¶ µ ¶
∂ ∂ωz ∂ Ny ∂ Nx ∂ 1 2
+ − = ∇ ωz . (20)
∂z ∂t ∂x ∂y ∂z R

By integrating this with respect to z it may be concluded that the transport equation for ωz
will hold at all z-locations if it can be satisfied in the limit as z → ∞. Thus, we can recover
the transport equation

∂ωz ∂ Ny ∂ Nx 1
+ − = ∇ 2 ωz , (21)
∂t ∂x ∂y R

provided the perturbation flow-fields can be determined in a manner that ensures

½ ¾
∂ωz ∂ Ny ∂ Nx 1
lim + − − ∇ 2 ωz = 0. (22)
z→∞ ∂t ∂x ∂y R

The secondary variables {u, v, ωz } are all defined so as to vanish in the limit as the wall-
normal coordinate approaches infinity. We will also assume that the velocity perturbation
component w vanishes and that the corresponding component W B of the undisturbed flow
tends to a constant. (This constant will be zero for many boundary-layer flows, for example
the flow over a flat plate, but is nonzero for the von-Kármán flow over a rotating disc.) The
condition (22) may then be simplified to
µ ¶µ ¶
1 ∂ ∂ωx ∂ω y
lim −W B + + = 0, (23)
z→∞ R ∂z ∂x ∂y

where the solenoidal property of the vorticity has been reapplied to obtain a condition in
terms of primary variables only. Such a condition would automatically be satisfied for the
vorticity components ωx , ω y that could be derived from solutions of the Navier–Stokes
equations in the usual primitive variables formulation. In our formulation it would appear
 VELOCITY–VORTICITY FORMULATION 129

to be necessary to impose (23) directly. However, as with the previous condition (19)
required for ensuring incompressibility, the need for any explicit enforcement of (23) can
be circumvented by making a judicious choice of coordinate mapping in the z-direction.
More details will be given later.
Assuming that some means is found for satisfying condition (23), the vorticity transport
equation (21) for the component ωz can be combined with the vorticity transport equations
(7) and (8) for the other two components ωx , ω y to give
½ ¾
∂u 1 2
∇× + U · ∇u + u · ∇U + u · ∇u − ∇ u = 0,
B B
(24)
∂t R
where the definition (2b) of vorticity has been used to eliminate any direct reference to
the perturbation vorticity field and there has been some straightforward manipulation using
vector identities. (Note that definition 2b is not assumed a priori, but is shown above to follow
from our formulation.) Since we are dealing with the simply connected domain defined by
z ≥ η(x, y, t), the irrotational vector quantity which appears enclosed in brackets in (24)
may be written as the gradient of some scalar field. This scalar field can be identified as the
negative of the perturbation pressure field. Hence, we can recover the momentum equations
for the perturbation velocity components in the usual primitive-variables form.
We have now shown that the primitive-variables formulation of the Navier–Stokes equa-
tions can be obtained from our compact velocity–vorticity formulation. It should be recalled
that the latter formulation comprises the three governing equations (7), (8), and (9) for the
primary variables {ωx , ω y , w} together with the relations (12)–(14) which define the sec-
ondary variables {u, v, ωz }. Full equivalence can only be ensured if the primary variables are
constrained to satisfy the limiting conditions (19) and (23). The assumption was also made
that all three components of the perturbation velocity field vanish for z → ∞. Such behavior
was built into the definitions of the secondary variables u, v. By contrast, the vanishing of
w may be used to provide a boundary condition for the solution of the Poisson equation (9).
We will now discuss what other boundary conditions must be imposed, and how they can
be associated, individually, with each of the governing equations for the primary variables.

2.2. Velocity Boundary Conditions at the Wall

The motion of the fluid must be matched to that of the solid wall which is located at
z = η(x, y, t). This gives boundary conditions of the general form

u(x, y, η, t) = u w (x, y, t), (25)

v(x, y, η, t) = vw (x, y, t), (26)
w(x, y, η, t) = ww (x, y, t), (27)

where u w , vw , ww are functions determined by the wall motion or otherwise. For example,
if the fluid was bounded by a flexible wall which could only move in the z-direction we
would need to set

u w (x, y, t) = −U B (x, y, η, t), (28)

vw (x, y, t) = −V (x, y, η, t),
B
(29)
∂η
ww (x, y, t) = (x, y, t) − W B (x, y, η, t). (30)
∂t
130 DAVIES AND CARPENTER

The case where there is a rigid wall at z = 0 can be treated by taking η = 0 and then simply
setting

u w = vw = ww = 0. (31)

It is a straightforward matter to associate the boundary condition (27) on w with one of

the governing equations; the prescribed wall-motion function ww can be used to define a
Dirichlet condition for the Poisson equation (9).
As noted in Section 1, the derivation of vorticity boundary conditions has been a source
of difficulty for most previous velocity–vorticity methods. When transport equations for
all three components of the vorticity are used in conjunction with all three Poisson equa-
tions (3), it would appear that a total of six boundary conditions must be imposed at the
wall. The Poisson equations can be solved by using the boundary conditions on the ve-
locity perturbations in a obvious manner, but the conditions required for the solution of
the vorticity transport equations are not so readily supplied. Fasel and his collaborators
[21, 23, 39, 54] made use of boundary conditions involving first and second derivatives
of the perturbation velocity and vorticity components. This approach usually works well
in practice but its limitations can be illustrated by considering an almost-parallel, two-
dimensional, undisturbed flow along the x-direction. (This, of course, includes the case of
Blasius flow over a flat plate studied by Fasel and his coworkers.) If attention is restricted
to two-dimensional disturbances, there is only a single nonvanishing component of the
perturbation vorticity, namely ω y . Its evolution is governed by a transport equation of the
form

∂ω y ∂ω y ∂ 2U B 1
+UB +w + non-parallel terms + nonlinear terms = ∇ 2 ω y . (32)
∂t ∂x ∂z 2 R

It can be seen that, when nonparallel and nonlinear terms are neglected, there is no direct
coupling between the perturbation vorticity ω y and the streamwise perturbation veloc-
ity component u. For the transport equation (32) Fasel et al. used the vorticity boundary
condition

∂ω y ∂ 2 w ∂ 2 ww
=− 2 − for z = η (33)
∂x ∂z ∂x2

obtained by substituting the boundary condition (27) for w into the Poisson equation (9).
(Fasel et al. set η = 0 but allowed ww 6= 0 in order to generate disturbances by suction and
blowing at the wall.) It may be noted that no use has been made of the no-slip boundary
condition. Thus, should an approximation of parallel undisturbed flow be applied and a
linearization performed, the evolution of the vorticity would not be constrained, either
directly or indirectly, by the no-slip condition on u. It follows that nonparallel or nonlinear
terms must be retained in the vorticity transport equation (32) in order for there to be any
prospect of obtaining a well-posed problem. Even then, the coupling between ω y and u can
be expected to remain weak if the effects of nonparallelism and nonlinearity are small. In
fact, it can be formally argued that the scheme adopted by Fasel et al. remains ill-posed
even if such effects are taken fully into account [13]. In practice, of course, this apparent
difficulty has not prevented the nonparallel and nonlinear version of the scheme being used
to conduct viable numerical simulations.
 VELOCITY–VORTICITY FORMULATION 131

A major advantage of our formulation is that there is natural means of associating the
wall-motion boundary conditions on u and v with the vorticity transport equations. We
avoid the difficulties mentioned above by replacing wall boundary conditions for the vor-
ticity with integral constraints obtained directly from the definitions (12) and (13) for
the secondary variables with account taken of the conditions (25) and (26). Thus, we
obtain

Z ∞ Z ∞
∂w
ω y dz = −u w − dz (34)
η η ∂x
Z ∞ Z ∞
∂w
ωx dz = vw + dz. (35)
η η ∂y

Since the substitution can be reversed, it is clear that these two relations are fully equivalent
to the wall-motion boundary conditions for u and v. The relations (34) and (35) may thus
be viewed as constraints on the evolution of the primary variables ω y , ωx , respectively.
Each relation can be associated with one of the vorticity transport equations (7) and (8)
to provide a means of satisfying the no-slip conditions, or their equivalents, for the fluid
perturbation velocity. This ensures that the problem remains well-posed even when non-
parallel and nonlinear terms are neglected or absent from the vorticity transport equations.
Moreover, there is an attractive economy in the usage of the wall boundary conditions
on the fluid perturbation velocities; each of the conditions for u, v, w is applied once,
and once only, to provide conditions that constrain the development of ω y , ωx , and w,
respectively.

2.3. Wall-Normal Coordinate Mapping and the Conditions for z → ∞

As suggested earlier, there is no difficulty in implementing a condition on w in the limit
z → ∞. The vanishing of w provides a boundary condition for the solution of the Poisson
equation (9). It is not so obvious how, in the same limit, the behavior of the other two
primary variables, ωx and ω y , can be constrained. The secondary variables u and v are
defined in such a way that they must both approach zero as z → ∞. This implies that the
limiting behavior of u and v cannot be used to obtain any additional constraint on ωx and
ω y . Accordingly, such constraints must be obtained by some other means. In fact, we have
already derived two further conditions that must be satisfied for z → ∞, namely (19) and
(23). Unfortunately, the first condition only constrains w, while the second one involves both
components of the perturbation vorticity in a symmetric manner. As a consequence, there
is no unambiguous way in which the two conditions can be associated with the individual
transport equations for the vorticity components. In the present study, we will sidestep the
issue by simply replacing the conditions (19) and (23) with the more easily implemented and
stronger conditions that both ωx and ω y vanish for z → ∞. We thus impose the same form of
limiting behavior for all three of the primary variables. The validity of such an approach can
be monitored, retrospectively, by checking that the primary variables computed using these
replacement limiting conditions do, in fact, satisfy the conditions (19) and (23). Clearly,
this consistency requirement would be met if it was found that the z-derivatives of all three
of the computed primary variables vanished for z → ∞. In practice, it is possible to obtain
132 DAVIES AND CARPENTER

satisfactory solutions by making use of the coordinate transformation

l
ζ = , (36)
z +l

which maps the semi-infinite physical domain z ∈ [0, ∞) on to the finite computational
domain ζ ∈ (0, 1]. The parameter l is a stretching factor. It can be seen that the limit
z → ∞ corresponds to the limit ζ → 0. Derivatives with respect to the physical coordinate
z are related to those with respect to the transformed co-ordinate ζ according to

∂f ζ2 ∂f
=− , (37)
∂z l ∂ζ

so that it may be inferred that the z-derivative will vanish for z → ∞ provided the
ζ -derivative remains bounded as ζ approaches zero. Thus, if the primary variables are
determined as functions of ζ , rather than z, it becomes a straightforward matter to de-
termine whether the z-derivatives appearing in the conditions (19) and (23) vanish. We
just need to check that the ζ -derivatives are bounded as ζ → 0. Consequently, provided
our numerical simulations yield solutions for the primary variables which remain smooth
functions of ζ when ζ → 0, the incompressibility condition and the transport equation for
the wall-normal vorticity will automatically be satisfied. Further discussion concerning the
usage of the mapping (36) is included in the section on numerical methods. In addition, a
brief account is given in Appendix B of how the mapping could be modified to deal with
a nonplanar flow boundary located at z = η(x, y, t), for configurations where it is not ap-
propriate to perform any linearization to obtain conditions to be applied at z = 0. For such
cases, it is possible to perform a simple change of coordinates to obtain a modified set of
governing equations for which the boundary conditions required at the physical boundary
z = η can be imposed along a planar computational boundary.

3. FORMULATION FOR THE ROTATING-DISC FLOW

3.1. Transformation to Cylindrical Polar Coordinates

Thus far, for expositional convenience, we have used a Cartesian coordinate system
to present our velocity–vorticity formulation. However, it is not difficult to see that the
formulation only relies upon a decomposition of the perturbation velocity and vorticity
fields that combines projections in the z-direction with projections on to the orthogonal plane
parallel to the wall. Hence, there is no problem changing to a cylindrical polar coordinate
system; all that is involved is a transformation of coordinates in planes defined by constant
values of z. It is also straightforward to change to a noninertial frame of reference rotating
at a constant angular velocity about the z-axis. The only significant difference from the
formulation for a nonrotating frame is that the Poisson equation (9) for w needs to be
employed in order to recover the Coriolis term in the transport equation for ωz . (There is
no additional term to be included in the condition on the limiting behavior that corresponds
to condition (23). The Coriolis term in the transport equation for ωz is proportional to the
derivative ∂w/∂z, and so must vanish for z → ∞ in accordance with (19), if the continuity
equation is to be satisfied.)
 VELOCITY–VORTICITY FORMULATION 133

For the rotating disc the velocity and vorticity fields may be represented as

u = (u r , u θ , w), ω = (ωr , ωθ , ωz ),

where the subscripts r and θ refer, respectively, to the radial and azimuthal directions defined
by the cylindrical polar coordinate system. The components {ωr , ωθ , w} can be taken as the
primary variables, while {u r , u θ , ωz } form the set of secondary variables. The governing
equations for the primary variables, obtained by transforming equations (7)–(9), take the
form
µ ¶ ½µ ¶ ¾
∂ωr 1 ∂ Nz ∂ Nθ ∂w 1 1 2 ∂ωθ
+ − − 23 ωθ + = ∇ 2 − 2 ωr − 2 (38)
∂t r ∂θ ∂z ∂r R r r ∂θ
µ ¶ ½µ ¶ ¾
∂ωθ ∂ Nr ∂ Nz 1 ∂w 1 1 2 ∂ωr
+ − + 23 ωr − = ∇ − 2 ωθ + 2
2
(39)
∂t ∂z ∂r r ∂θ R r r ∂θ
µ ¶
1 ∂ωr ∂(r ωθ )
∇ 2w = − , (40)
r ∂θ ∂r

where N = (Nr , Nθ , Nz ) can be defined as previously by Eq. (10). There are additional
terms on the left-hand sides of Eqs. (38) and (39) arising from the Coriolis acceleration.
They have been written entirely in terms of the primary variables. The factor 3 represents the
nondimensionalized angular velocity. In a similar manner to before, the secondary variables
may be determined from the primary variables, using the explicit definitions
Z ∞µ ¶
∂w
ur = − ωθ + dz, (41)
z ∂r
Z ∞µ ¶
1 ∂w
uθ = ωr − dz, (42)
z r ∂θ
Z µ ¶
1 ∞ ∂(r ωr ) ∂ωθ
ωz = + dz. (43)
r z ∂r ∂θ

3.2. Undisturbed Flow

The undisturbed flow is taken to be the laminar similarity solution of von Kármán [36]
which is an exact solution to the Navier–Stokes equations for the flow field above a rotating
disc in an unbounded fluid [55]. It is assumed that the disc rotates at a constant angular
velocity 3∗ about the vertical z ∗ -axis, where asterisks denote dimensional quantities. The
undisturbed base flow in a frame of reference rotating with the disc is given by
¡ ¢
U B∗ = UrB∗ , UθB∗ , W B∗ . (44)

These velocity components are related to the Kármán similarity variables as

1 1 1
F(z) = U B∗ , G(z) = ∗ ∗ UθB∗ , H (z) = 1 W
B∗
, (45)
r ∗ 3∗ r r 3 ∗
(ν 3 )∗ 2

where ν ∗ is the kinematic viscosity and z = z ∗ (3∗ /ν ∗ )1/2 .

134 DAVIES AND CARPENTER

FIG. 2. Radial (F), azimuthal (G) and axial (H) velocity components for boundary layer flow over a rotating
disc.

Thus, the nondimensional undisturbed flow is given by

µ ¶
r r 1
U =B
F, G, H , (46)
R R R

where the Reynolds number R is defined as R = ra∗ (3∗ /ν ∗ )1/2 for some selected radial posi-
tion ra∗ . Since lengths are nondimensionalized using the fixed lengthscale δ ∗ = (ν ∗ /3∗ )1/2 ,
we have R = ra . It may also be noted that the nondimensionalized rotation rate is equal to
1/R. Thus, when we choose to work in a frame of reference that rotates with the disc, it is
necessary to set 3 = 1/R for the Coriolis terms included in the vorticity transport equations
(38) and (39). The velocity-profile functions F, G, and H are plotted in Fig. 2.

3.3. Linearization and the Excitation of Azimuthal Modes

In the present study we are interested in disturbances with vanishingly small amplitude,
so we will use linearized governing equations. These are obtained by simply neglecting the
nonlinear term ω × u in the definition (10) of the quantity N. The linearization makes the
problem separable in the azimuthal direction which allows us to consider individual modes
of the form

u = (û r , û θ , ŵ) einθ , ω = (ω̂r , ω̂θ , ω̂ z ) einθ , (47)

where n is the integer-valued, azimuthal mode number. In order to excite disturbances in the
boundary-layer flow, we will prescribe a radially localized motion of the disc surface with
the same form of azimuthal variation. Because attention is restricted to small-amplitude
disturbances, the boundary conditions may be linearized about the undisturbed wall, which
is taken to be at z = 0. If it is supposed that the disc surface is only allowed to move in the
vertical direction, the linearized boundary conditions may be written as

r 0 r ∂η
ur = − F (0)η, u θ = − G 0 (0)η, w = at z = 0, (48a,b,c)
R R ∂t
 VELOCITY–VORTICITY FORMULATION 135

where z = η(r, θ, t) is the perturbed location of the disc surface. Substituting (48a) and
(48b) into the definitions (41) and (42) of the secondary variables yields the following
integral constraints on the primary variables
Z ∞ Z ∞
r 0 ∂w
ωθ dz = F (0)η − dz (49)
0 R 0 ∂r
Z ∞ Z ∞
r inw
ωr dz = − G 0 (0) η + dz, (50)
0 R 0 r

where, as will frequently be the case in what follows, the assumed azimuthal mode structure
has been used to replace partial derivatives with respect to θ by the multiplicative factor in.
In order to obtain a radially localized excitation of the fluid the wall displacement takes
the form

η(r, θ, t) = a(r − re ) b(t) einθ , (51)

where the function a specifies the radial variation centred on the radius r = re and the
function b defines the time-dependent amplitude. Typically, we have employed radial dis-
tributions of the form

a(r ) = e−λr eiαr ,

2
(52)

where λ is a scaling factor which fixes the radial extent of the forcing and α is a wavenumber
which, when not set equal to zero, may be chosen with the intention of enhancing the
response for disturbances with a particular radial wavelength. The time-dependent amplitude
may be chosen so as to give a time-periodic excitation, for instance by setting
¡ 2¢
b(t) = 1 − e−σ t e−iβt , (53)

where β is the prescribed temporal frequency and σ is a parameter chosen to scale the forcing
up from zero amplitude at t = 0. Stationary disturbances, generated by stationary “bumps”
on the disc surface distributed in accordance with the imposed azimuthal periodicity, can
be considered by setting β equal to zero. Similarly, disturbances can be excited impulsively
by employing
¡ 2¢
b(t) = 1 − e−σ t e−σ t
2
(54)

or some other smoothed approximation to an impulse function.

As an alternative to exciting the flow by introducing a vertical wall motion, it is possible
to specify a suction/blowing distribution. This can be achieved by dropping the terms
involving η in the integral constraints (49) and (50) and replacing the condition on w by
the requirement that w = ww at z = 0 where ww is a prescribed function describing the
character of the suction/blowing.

3.4. Dynamics of the Compliant Wall

The effects of wall compliance can be considered by embedding an annulus of compliant
material within the otherwise rigid disc (see Fig. 3). The dynamics of the compliant wall may
136 DAVIES AND CARPENTER

FIG. 3. Compliant annulus embedded within an otherwise rigid rotating disc.

be simulated using a spring-backed plate model, as in a previous study that was conducted
for a different flow-configuration by Davies and Carpenter [14]. It is assumed that the
compliant surface only moves in the z-direction. Denoting, as before, the wall displacement
by η(r, θ, t), the equation of motion for the wall may be written as

∂ 2η ∂η
m +d + B ∇ˆ 2 ∇ˆ 2 η + K η = − pw , (55)
∂t 2 ∂t

where pw is the fluid perturbation pressure acting at the wall, and ∇ˆ 2 is the two-dimensional
Laplacian operator defined by

∂2 f 1 ∂f 1 ∂2 f
∇ˆ 2 f = + + .
∂r 2 r ∂r r 2 ∂θ 2

The nondimensional parameters m, d, B, K are respectively the areal density, damping

coefficient, flexural rigidity, and spring stiffness per unit area of the compliant wall. The
perturbation pressure distribution can be found by integrating the linearized z-momentum
component of the Navier–Stokes equations to obtain
Z ∞½ µ ¶ µ ¶¾
∂w 1 ∂w ∂w 1 ∂(r ωθ ) ∂ωr
pw = + rF + (G + 1) + − dz. (56)
0 ∂t R ∂r ∂θ rR ∂r ∂θ

In practice, the terms involving the vorticity may be expected to be negligible at radial
locations of interest. They will give contributions that are of the same order as the wall-
normal viscous stresses that were ignored in Eq. (55). Assuming that some suitable means
can be found for solving the equation of motion for the compliant wall, values for the
displacement η and the wall velocity ∂η/∂t may be substituted into the linearized boundary
conditions and vorticity integral constraints given in Eqs. (48)–(50). The only difference
from before is one of interpretation. Near the forcing location r = re , the displacement η is
simply prescribed, while within the compliant portion of the disc surface the displacement
is interactively coupled to the fluid disturbances via the fluid perturbation pressure.

3.5. Inflow and Outflow Conditions

In many simulations of transitional boundary layers, the perturbation flow-fields are
assumed to be periodic in two directions parallel to the wall [38]. This periodicity can
 VELOCITY–VORTICITY FORMULATION 137

be built into the numerical methods, through the use of Fourier expansions, leading to
highly efficient spectral codes. There is no guarantee that such an approach will lead to
a satisfactory representation of spatially evolving disturbances, particularly when strong
upstream influence is anticipated, or the wall boundary conditions are nonuniform. If the
assumption of streamwise periodicity is dropped, it becomes necessary to address the issue
of inflow and outflow conditions. The formulation of such conditions for the rotating-disc
flow introduces some additional complications that are not encountered for two-dimensional
boundary layers. If the computational domain is defined as ri ≤ r ≤ ro , where ri and r0
denote the radius of the inner and outer boundaries, respectively, then it is not at all obvious
that it remains appropriate to use the terms “inflow” and “outflow” conditions. There is the
distinct possibility that there will be significant propagation and growth of disturbances in
the radially inward direction, as well as in the radially outward direction.
In practice, the inner radial boundary does not appear to cause any serious difficulties,
provided that it is located at a sufficiently small radius. Disturbances can be expected to be
increasingly damped as the disc centre is approached. Linear stability results [1, 43, 46] can
be used as a guide. Assuming that a suitable selection of the inner radius ri can be made, the
simplest means of constraining the disturbance evolution at the boundary is to impose null
values on an appropriate set of perturbation flow variables. For instance, all components of
the perturbation velocity could be set to zero, giving the conditions that

ur = u θ = w = 0 at r = ri . (57)

Using the definitions (41) and (42) of the secondary variables, u r and u θ , these conditions
can be translated into the following equivalent conditions on the primary variables:
∂w
ωθ = − , ωr = w = 0 at r = ri . (58)
∂r
Alternatively, the condition on ωθ may be replaced by a straightforward null-value condition,
giving the requirement that all of the primary variables vanish at the inner computational
boundary, that is

ωθ = ωr = w = 0 at r = ri . (59)

In the numerical simulations, it was found that the results obtained using either conditions
(58) or (59) were virtually identical, provided the inner radial boundary remained sufficiently
removed from any source of disturbances.
The choice of boundary conditions at the outer radius of the computational domain must
allow for the fact that in cases of interest, the boundary will lie in a region where disturbances
are unstable. The difficulties this causes are exacerbated when there is strong radially inward
influence and temporal disturbance growth. For our simulations, we considered three distinct
strategies for dealing with the outer radial boundary. The simplest is to ensure that the outer
radial boundary is always kept some distance away from locations where disturbances have
evolved to appreciable amplitudes. Such an approach makes it unnecessary to model the
behavior of disturbances at the outer boundary, but is computationally expensive.
The second strategy depends upon an assumption that the disturbances are wavelike at
the outer radius, and involves the imposition of conditions of the form

∂ 2 ωr ∂ 2 ωθ ∂ 2w
= −α 2 ωr , = −α 2 ωθ , = −α 2 w (60)
∂r 2 ∂r 2 ∂r 2
138 DAVIES AND CARPENTER

at r = ro . The wavenumber α, which we can allow to be complex-valued to take account

of spatial growth, can be set equal to the corresponding eigenvalue determined from linear
stability theory. Since these boundary conditions can be imposed at a radial location where
the disturbance amplitude is not negligible, the length of the computational domain may be
significantly reduced. A similar approach has previously been applied, with some success,
in studies of disturbance evolution in plane channel flow and Blasius flow (see for instance
the works by Fasel [21], Fasel and Konzelmann [23] and Davies and Carpenter [14]).
The validity of such wavelike boundary conditions usually depends upon the disturbance
evolution remaining both linear and convective in nature. The first of these requirements is
always met in the present numerical studies, since we only use linearized equations. The
requirement that the disturbance convect through the outer radial boundary cannot, however,
be so easily satisfied. When nonconvective behavior is anticipated, the simple wavelike
boundary conditions (60) need to be applied with extreme caution if the disturbance is
allowed to reach an appreciable amplitude at the outer radial computational boundary. (A
more detailed discussion of the precautions that were necessary for rotating disc simulations
involving absolute instability can be found in reference [16].) However, for the simulations
that we have selected to report in detail in the present paper, the disturbance behavior
remained purely convective. Under such circumstances, the wavelike conditions (60) were
found to be perfectly adequate. Moreover, they proved to be particularly useful during code
validation, when they facilated the use of relatively short computational domains.
The third strategy for dealing with the outer radial boundary is to add an outflow buffer
region to the computational domain. A variety of different implementations of the buffer-
region strategy have been applied in previous studies, for example as described by Bertolotti
et al. [4], Kloker et al. [39] and Liu and Liu [45]. However, the success of such an approach,
no matter how sophisticated in its implementation, can still be expected to depend upon
whether or not the disturbances display a convective nature up to the beginning of the buffer
region. Thus, in a sense, the buffer-region strategy involves little more than an elaboration
of simple wavelike boundary conditions (60). It does, however, have the advantage that it
can more readily cope with nonlinear disturbances and multimodal forms of disturbance
whose behavior cannot be adequately modelled by simply prescribing a single outflow
wavenumber. Nevertheless, the buffer region is only designed to allow disturbances to
be convected out of the computational domain without spurious reflections. We do not
claim to have undertaken extensive experimentation with the use of buffer regions for cases
where nonconvective behaviour could be discerned, but our few tests suggest that, at least
for linearized simulations involving only a single mode of disturbance, there was little
advantage to be gained over the use of the simpler wavelike boundary conditions.

4. NUMERICAL METHODS

We will describe the numerical methods in the context of the rotating-disc problem, but the
corresponding formulae for Cartesian coordinates should, for the most part, be fairly evident.
We employ a mixed finite-difference/spectral method for the numerical discretization. The
time-derivatives in the two vorticity transport equations (38) and (39) are discretized using
a second-order scheme. For the sake of being definite, we will describe a particular three-
point backward-difference scheme found to be robust for the most numerically challenging
configuration that we have considered to date. Namely, the case where part of the disc
 VELOCITY–VORTICITY FORMULATION 139

surface is compliant. For less demanding cases, where the disc’s surface motion could
be simply prescribed, a Crank–Nicholson/Adams–Bashforth type of time-stepping method
was also implemented. This was more efficient than the method based on a three-point
backward-difference scheme, when interactive calculations of the wall motion were not
required. Further details are given in Appendix C.
Variations in the radial direction are discretized using a fourth-order, centered, compact
finite-difference scheme. The wall-normal discretization is based upon a Chebyshev series
representation using the mapped variable introduced in Eq. (36). We will focus our account
of the numerical methods on features of the Chebyshev discretization that are of particular
interest. It will be shown that is possible to obtain a discretization which is spectrally
accurate in the z direction, but nevertheless involves only pentadiagonal operators acting
on the Chebyshev expansion coefficients.

4.1. Chebyshev Discretization for the Wall-Normal Variation

Restricting attention to an individual azimuthal mode, the primary variables are expanded
in terms of odd Chebyshev polynomials as follows

( M )
X
ωr (r, θ, z, t) = ωrk (r, t) T2k−1 (ζ ) einθ (61)
k=1
( M )
X
ωθ (r, θ, z, t) = ωθk (r, t) T2k−1 (ζ ) einθ (62)
k=1
( )
X
M
w(r, θ, z, t) = wk (r, t) T2k−1 (ζ ) einθ , (63)
k=1

where Tk is the kth Chebyshev polynomial and ζ ∈ (0, 1] is the mapped wall-normal co-
ordinate defined in Eq. (36). It is acceptable to use only the odd Chebyshev polynomials
because our semi-infinite physical domain is mapped on to half of the usual Chebyshev inter-
val. We assume, implicitly, that the primary variables and all of their even-order ζ -derivatives
vanish for ζ → 0. This will certainly be the case if the disturbances decay exponentially for
z → ∞. (If for some flow variable f we have f ∼ e−sz for z → ∞, where s > 0, then we
must also have ∂ n f /∂ζ n → 0 when ζ → 0 for all values of n.) The requirement that there
is exponential decay of the disturbances should also be sufficient to ensure the coefficients
in the Chebyshev expansion converge faster than any inverse power of their order [6].
The secondary variables may be expanded in terms of even Chebyshev polynomials
( )
1 XM
u r (r, θ, z, t) = u r (r, t) + u rk (r, t) T2k (ζ ) einθ (64)
2 0 k=1
( )
1 XM
u θ (r, θ, z, t) = u θ (r, t) + u θk (r, t) T2k (ζ ) einθ (65)
2 0 k=1
( )
1 XM
ωz (r, θ, z, t) = ωz (r, t) + ωzk (r, t) T2k (ζ ) einθ , (66)
2 0 k=1
140 DAVIES AND CARPENTER

where we set the zeroth-order expansion coefficient for u r to be given by

X
M
u r0 (r, t) = −2 u rk (r, t) (−1)k , (67)
k=1

and similarly for u θ0 (r, t) and ωz0 (r, t). This ensures that the secondary variables all vanish
for ζ = 0, i.e. for z → ∞. Fixing the zeroth-order expansion coefficients in such a fashion is
equivalent to expanding the secondary variables using a series of modified even Chebyshev
polynomials T̄2k = T2k − T2k (0).
The Chebyshev expansions for the primary and secondary variables can be substituted
into the defining Eqs. (41)–(43) for the secondary variables to yield the following tridiagonal
relations between the Chebyshev coefficients,
µ ¶
∂
(k − 1)u rk−1 + 2ku rk + (k + 1)u rk+1 = −l ωθk − ωθk+1 + (wk − wk+1 ) (68)
∂r
µ ¶
in
(k − 1)u θk−1 + 2ku θk + (k + 1)u θk+1 = l ωrk − ωrk+1 − (wk − wk+1 ) (69)
r
µ ¶
l ∂ ¡ ¡ ¢¢ ¡ ¢
(k − 1)ωzk−1 + 2kωzk + (k + 1)ωzk+1 = r ωrk − ωrk+1 + in ωθk − ωθk+1 ,
r ∂r
(70)

for k = 1, . . . , M, where l is the stretching factor in Eq. (36) defining ζ . Details of the
derivation are included in Appendix D. These relations can readily be used to determine the
Chebyshev coefficients for the secondary variables from given values of those for the primary
variables. All that is involved is a simple application of the Thomas algorithm. It should be
remarked that the Chebyshev coefficients of the secondary variables are only required for
the purposes of computing the convective terms in the vorticity transport equations. These
involve the linearized quantity

N = ΩB × u + ω × U B , (71)

which is treated explicitly, via a predictor-corrector method, in the time-marching procedure.

Thus, there is no need to allocate any permanent storage for the secondary variables; they
only appear at an intermediate stage in the calculation of the quantity N from specified
values of the primary variables. In fact, the representation of the secondary variables using
Chebyshev coefficients can immediately be exchanged, via an application of a Fast Fourier
Transform, for a representation in terms of collocation values. These collocation values can
then be utilised in the pseudospectral evaluation of the mean-flow products in N.
Before discussing the discretization of the governing equations we note that the integral
conditions, (49) and (50), placed on the primary variables can be cast in the forms

X
M
r 0 X ∂wkM
pk ωθk = F (0) η − pk (72)
k=1
R k=1
∂r

X
M
r X inwk M
pk ωrk = − G 0 (0) η + pk (73)
k=1
R k=1
r
 VELOCITY–VORTICITY FORMULATION 141

for fixed constants pk . These constraints on the primary-variable Chebyshev coefficients

may be interpreted as discretized versions of the linearized no-slip conditions for the dis-
turbances. The constants pk can be computed by extending the argument used to derive
the tridiagonal relations (68)–(70). More details are given in Appendix D. The boundary
condition (48c) on w can be discretized in a more straightforward manner to yield

X
M
∂ η̂
q k wk = , (74)
k=1
∂t

where qk = T2k−1 (1) = 1 and the wall-displacement amplitude η̂ is defined by writing

η(r, θ, z, t) = η̂(r, z, t) einθ .
The governing equations (38)–(40) are integrated twice with respect to the mapped vari-
able ζ . This removes the z-derivative operators and replaces them with ζ -integral operators.
Such a procedure is adopted because the integral operators take a very convenient form
which facilitates the employment of an efficient line iteration along the radial direction.
The integrated equations may be written as
· µ ¶¸ ¸
∂ωr in ∂w 1
I + Nz − 23 ωθ + + JNθ
∂t r ∂r l
½ ·µ ¶ ¸ ¾
1 1 2in 1
= I ∇ − 2 ωr − 2 ωθ + 2 K ωr
ˆ 2
(75)
R r r l
· µ ¶¸
∂ωθ ∂ Nz in 1
I − + 23 ωr − w − JNr
∂t ∂r r l
½ ·µ ¶ ¸ ¾
1 1 2in 1
= I ∇ − 2 ωθ + 2 ωr + 2 K ωθ
ˆ 2
(76)
R r r l
· ¸
1 in 1 ∂(r ωθ )
I ∇ w + 2 Kw = I
ˆ 2
ωr − , (77)
l r r ∂r

where the integral operators I, J, and K are defined by

Z ζZ ζ0
I f (ζ ) = f (ζ 00 ) dζ 0 dζ 00
Z ζ Z ζZ ζ0
J f (ζ ) = ζ 02 f (ζ 0 ) dζ 0 − 2 ζ 00 f (ζ 00 ) dζ 0 dζ 00
Z ζ Z ζZ ζ0
03 0 0
K f (ζ ) = ζ f (ζ ) − 6
4
ζ f (ζ ) dζ + 6 ζ 002 f (ζ 00 ) dζ 0 dζ 00 ,

and, using the assumed form of the azimuthal variation,

∂2 f 1 ∂f n2
∇ˆ 2 f = + − f
∂r 2 r ∂r r2

for any function f . When they are applied to an appropriate Chebyshev series, the integral
operators can be represented by banded matrices that are, at most, pentadiagonal, except for
the lowest-order terms. For instance, the integral operator I acts on series of odd Chebyshev
142 DAVIES AND CARPENTER

polynomials in the following tridiagonal manner

X
M
I f k T2k−1 = a T0 + b T1
k=1

X
M+1 µ ¶
1 f k−1 fk f k+1
+ − + T2k−1 , (78)
k=2
8 (2k − 1)(k − 1) (k − 1)k (2k − 1)k

where the arbitrary integration constants a, b arise from the double indefinite integration. A
similar, but pentadiagonal, representation can be obtained for the operator K. The integral
operator J, which acts on the series of even Chebyshev polynomials used to approximate
the quantities Nθ and Nr , can be represented by a matrix with a bandwidth of four.
Substituting the Chebyshev series for the primary variables into the integrated governing
equations and then matching the coefficients of T2k−1 for k = 2, . . . , M leads to a system
of 3(M − 1) partial differential equations for the 3M unknowns {ωrk , ωθk , wk }. The three
equations that would have been obtained by matching the coefficients of the lowest-order
odd polynomial T1 can be dispensed with because they only serve to introduce additional
unknowns in the form of integration functions. They are replaced by the conditions given
in Eqs. (72)–(74) corresponding to the wall boundary conditions for the fluid disturbance
velocity components, the no-slip conditions having first been formulated as the equivalent
integral constraints on the vorticity. We then obtain a set of 3M partial differential equations
for the 3M unknown Chebyshev coefficients of the primary variables. This procedure may
be classified as a form of the tau-method.

4.2. Temporal and Radial Discretization

Using the notation

f l = f |t=l1t

the three-point, backward-difference, time discretization takes the form

µ ¶l
∂f 1
= (3 f l − 4 f l−1 + f l−2 ). (79)
∂t 21t

The time-stepping is performed using a predictor-corrector scheme for the convective terms
in the integrated vorticity transport equations. The Coriolis terms in the transport equations
may be dealt with either implicitly or by applying the same predictor-corrector that is used for
the convective terms. The predictor-corrector method can also be chosen for some selected
viscous terms arising from the use of a non-Cartesian coordinate system. The remaining
viscous terms, which involve double indefinite ζ -integrals of second-order derivatives along
each of the r , θ, z coordinate directions, can all be treated implicitly. The Poisson equation
for w and all of the boundary conditions, including the integral constraints on the vorticity,
are also applied in an implicit manner. (In what follows we will assume, for the sake of being
definite, that the Coriolis terms and selected viscous terms are computed using a predictor-
corrector. The modifications required for an implicit treatment of all but the convective
terms would introduce only marginal differences in the computational cost.)
 VELOCITY–VORTICITY FORMULATION 143

The predictor step for the convective terms, which requires the evaluation of the quantity
N, is specified by setting

(Nl ) p = 2 Nl−1 − Nl−2

= 2(Ω B × ul−1 + ωl−1 × U B ) − (Ω B × ul−2 + ωl−2 × U B ), (80)

while the corrector step is defined by

(Nl )c = N((ul ) p , (ωl ) p ) = Ω B × (ul ) p + (ωl ) p × U B , (81)

where (ul ) p , (ωl ) p are the disturbance velocity and vorticity fields determined from the
predictor stage. This time stepping process involves two evaluations of the undisturbed-
flow product terms per time step. In addition, because of the implicit treatment of some of
the remaining terms in the governing equations, there is a large system of linear equations
that must be solved twice per time step. As mentioned previously, the products involving
undisturbed-flow quantities are calculated in a pseudospectral manner using a Fast Fourier
Transform to convert between Chebyshev series coefficients and collocation values and vice
versa. When the predictor-corrector employed for the convective terms is also applied to
the Coriolis terms and appropriately selected viscous terms, the relevant Chebyshev series
coefficients can be evaluated directly without resort to any transformation.
The Crank–Nicholson method, used in conjunction with an Adams–Bashforth treatment
of the convective terms, might appear at first sight to be a more obvious, less computationally
expensive, choice for the time-stepping procedure. However, on the basis of our previous
experience [14] we expected a backward-difference method to perform more robustly, par-
ticularly when there is interactive wall-motion. The advantages of such time-stepping pro-
cedures have also been documented in other circumstances where there is motion at a fluid
boundary (see, for example, [5]). Nevertheless, we make no claim that the particular time-
stepping procedure that we have elected to describe here in detail has been optimized. For
simulations that did not involve any interactively coupled wall motion, improvements to the
efficiency of the numerical scheme were readily made by treating more terms in an explicit
manner and adopting a modified predictor-corrector method. (As was mentioned earlier,
further details of the alternative time-stepping procedure can be found in Appendix C.)
Introducing the notation

f k,l j = f k |r = j1r,t=l1t ,

where, as before, the suffix k is used to denote a Chebyshev series coefficient, the dis-
cretization of the governing equations can be completed by approximating the radial
derivatives using compact, fourth-order, centered, finite-difference schemes of the general
form:
µ ¶ µ ¶ µ ¶
∂fl ∂fl ∂fl
α + +α
∂r k, j−1 ∂r k, j ∂r k, j+1
à ! à !
2(α + 2) f k,l j+1 − f k,l j−1 4α − 1 f k,l j+2 − f k,l j−2
= + , (82)
3 21r 3 41r
144 DAVIES AND CARPENTER

µ ¶ µ ¶ µ ¶
∂2 f l ∂2 f l ∂2 f l
β + +β
∂r 2 k, j−1 ∂r 2 k, j ∂r 2 k, j+1
à ! à !
4(1 − β) f k,l j+1 − 2 f k,l j + f k,l j−1 10β − 1 f k,l j+2 − 2 f k,l j + f k,l j−2
= + .
3 (1r )2 3 4(1r )2
(83)

In principle, the parameters α, β can be freely chosen. The choices α = 1/4, β = 1/10
minimize the stencil of discrete radial locations involved in the computation of the first and
second derivatives. Most of our computations were performed using these values. Setting
α = β allows both first- and second-order radial derivatives to be completely eliminated
from the fully discretized governing equations, which is more convenient when computer
storage is at a premium. We conducted a few simulations with α = β = 1/10 and obtained
results that were virtually identical to those obtained with α = 1/4, β = 1/10.
The finite-difference approximations for the radial derivatives may be employed together
with the temporal discretization to obtain discretized versions of the partial differential
equations for the primary-variable Chebyshev coefficients discussed in Section 4.1. The
full set of 3M partial differential equations contains three subsets of M equations that are
associated, in turn, with the two vorticity transport equations and the Poisson equation for
w. The first equation in each M-equation subset is chosen to be the condition that enforces
the appropriate boundary condition on the fluid perturbation velocity at the wall, while the
remaining M − 1 equations are obtained directly from the integrated governing equation.
Using the notation
¡ ¢T
Flj = f 1,l j , f 2,l j , . . . , f M,
l
j ,

for vectors of Chebyshev coefficients, the fully discretized governing equations can be
written in the forms

S j Ωlr j = Blr j + Clr j , S j Ωlθ j = Blθ j + Clθ j , T j W lj = Blwj , (84)

where {Ωlr j , Ωlθ j , W lj } represent the Chebyshev vectors of the primary variables {ωr , ωθ , w}
and S j , T j are M × M matrices. The vector Blr j is a simple linear function of vectors Ωlri for
i 6= j and the vector W lj , the latter being involved because the integral constraint coupling
ωr and w is incorporated as the first component of the vector equation. The vectors Blθ j ,
Blw j are specified in a similar manner; the former can be computed using vectors other
than Ωlθ j and the latter computed from vectors other than W lj . In the two fully discretized
vorticity transport equations, the vectors Clr j and Clθ j collect together quantities which
can be calculated explicitly using the primary-variable vector values from the previous
two time steps and, in the corrector stage of the time stepping, from the primary-variable
vectors computed during the predictor stage. There is no corresponding vector in the fully
discretized Poisson equation for w because the equation is treated in a fully implicit fashion.
The matrices S j , T j are pentadiagonal except for their first rows. This pentadiagonal
structure is the consequence of our choosing to replace z-derivative operators by ζ -integral
operators for the purposes of the Chebyshev discretization. The first rows of the matrices
are used to incorporate the perturbation velocity boundary conditions at the wall. It should
be recalled that the conditions to be imposed on the primary variables {ωr , ωθ , w} were
stated in Eqs. (72)–(74). Because the coefficients pk , qk appearing in these condition are
nonzero for k = 1, . . . M, the first rows of the matrice S j , T j are full.
 VELOCITY–VORTICITY FORMULATION 145

It may be seen that the fully-discretized equations (84), have been cast in a form that
is well-suited for efficient solution by means of a line iteration along the radial direction.
The inversions required to obtain estimates for each of the primary-variable Chebyshev
vectors {Ωlr j , Ωlθ j , W lj } at a given radial location, using previous estimates for the vectors
at locations that are radially inward and outward from the given location, can be performed
by employing a pentadiagonal equation solver that is modified to take account of the full
first rows of the matrices S j , T j .

4.3. Numerical Solution for the Compliant-Wall Motion

When an annular region of the wall is compliant (see Fig. 3), values of the wall displace-
ment must be supplied to specify the boundary conditions for the fluid motion. In order to
obtain a stable numerical scheme, the equation of motion (55) for the wall must be recast
in form
µ Z ∞ ¶
∂
mws + w dz + d ws + B ∇ˆ 2 ∇ˆ 2 η + K η
∂t 0
Z ∞½ µ ¶¾
1 ∂w ∂w 1 ∂(r ωθ ) ∂ωr
=− rF + (G + 1) + − dz, (85)
R 0 ∂r ∂θ r ∂r ∂θ

where boundary condition (48c) has been used and we have substituted the expression (56)
for the fluid perturbation pressure at the wall. Equation (85) can be viewed as an evolution
equation for the total wall-normal momentum defined by
Z ∞
µ = mws + w dz. (86)
0

The compliant-wall and hydrodynamic momenta are combined in order to avoid the domi-
nation of one over the other when they are included in the radial line iteration outlined in
Section 4.2. When this is not done, the iteration can fail to converge. The details of how
the reformulated equation of motion for the compliant wall may be employed to derive a
numerically stable coupling between the wall and the fluid are similar to those that were
presented in [14], to which the interested reader should refer. For present purposes it suffices
to note two important features. First, the fully-discretized boundary condition for w may
be implemented in the modified form

X
M
(γ j pk + qk ) wk,
l
j = w̃ j ,
l
(87)
k=1

where γ j is defined in terms of the wall parameters and the discretization constants. (The
coefficients pk , qk which are used, respectively, to evaluate w at the wall and its integral
across the boundary layer, were defined in Section 4.1.) The variable w̃lj appearing on
the right-hand side of (87) can be calculated using vectors other than W lj . Thus, it may
determined within the radial line-iteration procedure. Secondly, we note that the fully dis-
cretized version of the unmodified boundary condition on w, which was stated in a partially
discretized form in Eq. (74), may be written as

1 ¡ l−1 ¢ 21t X
M
η̂lj = 4 η̂ j − η̂l−2
j + qk wlj,k (88)
3 3 k=1
146 DAVIES AND CARPENTER

where

η̂lj = η̂|r = j1r,t=l1t .

This can be used to determine the wall displacement amplitude η̂, that must be known
in order to impose the integral constraints on the vorticity, when the boundary condition
applied directly to w is cast in the form (87).

5. SIMULATION RESULTS FOR CONVECTIVE DISTURBANCES

IN THE ROTATING-DISC BOUNDARY LAYER

5.1. Stationary Disturbances and Code Validation

We undertook extensive validation of our computer code through simulations of the
disturbances in the rotating-disc boundary layer. Initially, the governing equations were
“parallelized” (i.e., slow radial change was neglected) so that direct comparisons could be
made with results from previous linear-stability studies [12, 46–48]. The “parallelization”
was achieved by neglecting the radial dependence of the coefficients which multiply the
differential operators appearing in the governing equations (38)–(40) and the definitions
(41)–(43) of the secondary variables. All the coefficients were set to constant values by
simply substituting the fixed Reynolds number R for the variable radius r .
Radial wavenumbers and growth rates determined from the “parallelized” numerical
simulations were compared with corresponding eigenvalues determined from the linear
stability theory of previous investigators. (The specific form of the sixth-order system
of ordinary differential equations, and the numerical method used for their solution, are
described in reference [12]). Good agreement was found over a range of Reynolds numbers
for a variety of stationary and time-periodic forms of disturbance, including both Type I
and Type II instabilities.
Comparisons were also made for numerical simulations based on the complete linearized
Navier–Stokes equations. It was found that locally defined radial wavenumbers and growth
rates computed from the numerical simulations could be well-predicted using the eigenval-
ues obtained from standard, (i.e., quasi-parallel), linear stability theory. This is illustrated
in Fig. 4, which shows radial wavenumbers and growth rates for the case of a stationary
disturbance with azimuthal mode number n = 32. (This is the most strongly growing eigen-
mode.) In the numerical simulations, the disturbance was excited by introducing bumps on
the wall in the manner described in Section 3.3. The maximum wall displacement was
located at r = 350. Locally defined, complex, radial wavenumbers α = αr + iαi could be
computed using the simulation data by setting

1 ∂2 A
α2 = − , (89)
A ∂r 2
where A is the complex amplitude of some selected flow-field variable. For the results
presented in Fig. 4, the complex radial wavenumber was calculated using the amplitude of
the integral of ωθ across the boundary layer. It can be seen that, away from radial locations
close to where the wall deformation is greatest, the radial wavenumbers and growth rates
agree well with standard linear stability theory. The oscillations in the curves near r = 350
are a reflection of near-field effects in the simulations; the disturbance only has a well-
defined complex radial wavenumber at a sufficient radial distance from its source.
 VELOCITY–VORTICITY FORMULATION 147

FIG. 4. Radial wavenumbers αr and radial growth rates αi for a stationary disturbance with azimuthal
wavenumber n = 32. The solid curves were determind by solving, for each of the Reynolds number over the
range of radii, the eigenvalue problem derived by applying the parallel-flow approximation. The dotted curves
correspond to locally calculated complex radial wavenumbers obtained directly from numerical simulation data.

The numerical-resolution requirements for the simulations were established by undertak-

ing extensive and detailed, grid-refinement studies. It was found that, typically, the variation
of the primary perturbation variables ωr , ωθ , w across the boundary layer in the z-direction,
could be fully resolved using a Chebyshev expansion involving 48 polynomials. The radial
resolution was selected to give at least 24 points per wavelength, which usually translated to
taking a radial increment 1r ' 1. The temporal resolution was dictated by the convergence
demands of the line-iteration scheme used to solve the set of implicit equations obtained
at each time step, rather than by accuracy requirements. Typically, the need to obtain a
converged solution led to a choice of time increment 1t ' 1r that was sufficient for full
resolution of the temporal evolution when the Reynolds number R was O(103 ) or less. (It
should be noted that, with the specified nondimensionalization, the Reynolds number is
equal to the value of the dimensionless radial coordinate at the inner radius of the com-
putational domain.) The nondimensional rotation period T for the disc is equal to 2π R.
Thus, it may be seen that our typical choice of time-increment was such that 1t/T ∼ 10−3 .
In addition, as has mentioned above, the effects of the positions of the inner and outer
computational radial boundaries were carefully monitored. For most of the simulations,
the computational domain extended well beyond the limits suggested by the figures used
to illustrate the behavior of physical interest. For example, the perturbation velocity fields
for the time-periodic Type I disturbances considered at beginning of the next section are
displayed for 375 ≤ r ≤ 525, whereas the simulations were actually conducted using a
computational domain with 250 ≤ r ≤ 950.

5.2. Nonstationary Type I and II Disturbances

In order to further illustrate the utility of our numerical simulation method, we provide
a brief description of results for the radial evolution of Type I and Type II travelling-wave
disturbances. Both types of disturbance were excited by prescribing a radially localized,
time-periodic, wall displacement, with a nondimensional frequency β R. The form of the
wall motion used for the excitation was described in Section 3.3.
148 DAVIES AND CARPENTER

FIG. 5. Perturbation velocity contours for a Type I disturbance with n = 43 and β R = −8, excited by localized
disk surface motion centred at re = 400. The contours are drawn at levels ±A2m for m = 0, 1 . . . , N , where A is an
arbitary normalization factor and N defines the contour level corresponding to the largest disturbance magnitudes.
The normalization remains the same for each of the plotted perturbation velocity components but the value of N
varies: (a) u r ; N = 11, (b) u θ ; N = 12, (c) w; N = 8. Thus, the z-component of the perturbation velocity can be
seen to be an order of magnitude smaller than the other two components.

Figure 5 displays perturbation velocity contours, at a particular instant of time and a

fixed value of the azimuthal angle, for a Type I disturbance with azimuthal mode number
n = 43. The disturbance was excited by setting the nondimensional wall-motion frequency
β R = −8. The forcing was centred on re = 400. Both the n and β R were chosen so that,
according to standard linear stability theory, the radial growth rate would be as large as
possible at radial locations close to excitation. The pronounced radial growth experienced
by the disturbance can be inferred from the contour plots upon noting the logarithmic scale
 VELOCITY–VORTICITY FORMULATION 149

FIG. 6. Perturbation azimuthal vorticity ωθ at the disk surface (solid lines), together with corresponding
envelopes ± |ωθ | (broken lines), for Type I and Type II disturbances exicted at re = 400. (a) n = 43 and β R = −8
(b) n = 5 and β R = 7.9 The amplitudes are normalized so that, in both cases, ωθ is O(1) at locations near to
where the disturbances are generated.

used to define the contour levels. This radial growth can also be discerned, more readily,
from Fig. 6a, which displays the radial variation of ωθ at the disc surface. The time instant
and the azimuthal orientation are the same as for Fig. 5. Figure 6b shows the corresponding
result for the Type II disturbance, which, according to the standard linear stability theory,
should undergo maximal radial growth at the chosen radial location re = 400. In this case,
n = 5 and β R = 7.9. It can be seen that, as expected from previous studies, the radial growth
of the Type II disturbance is much weaker than that exhibited by the Type I disturbance.
In particular, it should be noted that the amplitude of the localized wall motion, used as
excitation, has been normalized in a commensurable fashion for both types of disturbance.
Thus, the large differences between the magnitudes of the perturbation vorticities, at radial
locations away from the source of excitation, reflect the disparity between the radial growth
rates of the two disturbance types. The difference in the radial growth rate is also apparent
from the contour plots given in Fig. 7. By making a comparison with Fig. 5, it can be seen
that the two types of disturbance also differ in other respects. For instance, there is a marked
contrast between the radial wavelengths. There are also differences in the distance above
the disc surface where the perturbation velocities achieve their maxima.

5.3. The Effects of Wall Compliance

The different characters of the Type I and Type II disturbances remained evident when
simulations were conducted for configurations where a compliant annulus was inserted
in the rotating-disc surface, as shown in Fig. 3. Figure 8 illustrates the stabilizing effect
on Type I disturbances of having such a compliant insert between r = 450 and 500. The
figure shows the instantaneous radial variation of the wall pressure, at a fixed azimuthal
angle for a case where the disc surface was taken to be entirely rigid, compared with
a corresponding case where there was a compliant insert. The wall displacement is also
plotted. As before, the disturbance was excited by time-periodic wall motion centered on
r = 400. It is clear that wall compliance leads to a significant reduction in the disturbance
growth in the radial direction, even though the compliant annulus extends for less than two
disturbance wavelengths. This is in accordance with the results of the standard linear stability
of Cooper and Carpenter [12] for compliant walls in the form of a single layer of elastomeric
material, the dynamics of which were modelled using modified Navier equations.
The wall parameters m, B, K defining our simpler spring-backed-plate model were
selected so that no flow-induced surface instabilities would be expected over the range of
radii considered. (It should be recalled from Section 3.4 that m, B, K are, respectively, the
150 DAVIES AND CARPENTER

FIG. 7. Perurbation velocity contours for a Type II disturbance with n = 5 and β R = 7.9, exited by localised
disk surface motion centered at re = 400. The contours are drawn at levels ±A2m for m = 0, 1 . . . , N , in the
same manner as in Fig. 5. (a) u r ; N = 6, (b) u θ ; N = 7, (c) w; N = 5.

nondimensional areal density, flexural rigidity, and spring stiffness for the compliant wall.)
Crude estimates for the onset velocities of flow-induced surface instabilities were derived by
modelling the undisturbed von Kármán flow as a uniform rotational flow and then assuming
that established results for the case of uniform uni-directional flow over a compliant plate
[10] could be applied in a localized manner. Such a procedure leads to two restrictions on
the choice of the nondimensional wall parameters, each of which gives a recipe for the
avoidance of a distinct mode of flow-induced surface instability. These restrictions can then
be applied to choose the wall parameters to make the compliant annulus as soft as possible
without introducing any flow-induced surface instability. When such a choice is made, the
wall parameters m, B, K can all be specified in terms of a single critical wavenumber
 VELOCITY–VORTICITY FORMULATION 151

FIG. 8. Instantaneous surface displacements and perturbation pressures at the disk surface for a Type I
disturbance with n = 43 and β R = −8 exicted at r = 400. (i) Pressure for an entirely rigid disk, (ii) pressure at
the same time instant in the periodic forcing cycle when there is a compliant annulus which extends from r = 450
to r = 500, (iii) corresponding compliant surface displacements. The surface compliance parameters are such that,
notionally, there is marginal stability with respect to divergence and travelling wave flutter forms of disturbance
at the outermost radius of the annulus. The critical wavenumber for the compliant surface is αc = 0.2.

defined by
µ ¶ 14
K
αc = . (90)
3B

The compliant part of the disc surface would only be expected to give rise to significant
stabilizing effects on Type I disturbances when αc is selected so as to be comparable with
the wavenumbers of the most unstable Type I disturbances. Accordingly, we set αc = 0.2
for the simulation presented here. (For a more detailed discussion of these matters, in the
context of plane channel flow between compliant walls, the interested reader should refer
to [15].)
Simulations were also conducted to investigate the effect of wall compliance on Type II
disturbances. The changes identified in the radial growth rates were negligible when the
wall parameters were selected in the above manner. This null result concurs with the work
of Cooper and Carpenter [12], who found that the modification of Type II disturbances
resulting from wall compliance was rather weak, compared with Type I disturbances.

5.4. Algebraic Radial Growth

Modal coalescence between Type I and Type II disturbances has been studied, indepen-
dently, by Lingwood [43], Cooper and Carpenter [12], and Turkyilmazoglu and Gajjar [58].
152 DAVIES AND CARPENTER

The standard linear stability analysis carried out by these previous investigators predicted
that modal coalescence would first occur at R = 437 for a temporal frequency β = 0.0061
and an azimuthal wavenumber given by n/R = 0.041 (which, it should be noted, actually
corresponds to an unphysical noninteger value of n = 17.9). They found that the coa-
lesced eigenmode was convectively stable at its onset Reynolds number. Examination of
the solution branches for the spatio-temporal eigenvalue problem also established that the
modal coalescence did not give rise to an absolute instability. There remained the interest-
ing possibility that the coalesced eigenmode could exhibit prolonged algebraic growth [40].
However, upon closer scrutiny, the physical significance of any such algebraic growth might
appear to be rather limited. Quasi-parallel linear stability theory predicts that, at above a
Reynolds number R ' 439, which is only very slightly higher than the Reynolds number
of R = 437 corresponding to the onset of modal coalescence, the coalesced mode becomes
convectively unstable. Thus, when the slow radial variation is properly accounted for, it
is likely that any algebraic radial growth would be eclipsed by exponential radial growth
over a relatively short radial distance. Notwithstanding this, there is still some purpose to
be achieved from a brief study of disturbance development when modal coalescence is
expected. Numerical simulations can be used to examine the effects of “nonparallelism”
and to verify that algebraic radial growth does, in fact, occur within the framework of the
“parallel-flow” approximation. Moreover, the successful identification of algebraic radial
growth in simulations conducted with “parallelized” governing equations can be viewed as
providing an additional stringent validation for the computer code.
Figure 9 displays “parallelized” numerical simulation results for the radial variation
of ωθ at the wall for the case of a time-periodic disturbance at the predicted onset of
modal coalescence R = 437. A cursory inspection suggests that the disturbance is subject
to algebraic growth, rather than the exponential decay expected in the absence of modal
coalescence. Stronger confirmation of the predictions of the quasi-parallel linear stability
theory of Cooper and Carpenter can be inferred from the excellent fit to the simulation data

FIG. 9. Instantaneous azimuthal vorticity ωθ at the surface of a rigid disk for the modal coalescence involving
the Type I and Type II forms of disturbance. R = 437, n/R = 0.041 and β = 0.0061. (i) Numerical simulation
results obtained with the parallel-flow approximation applied to the fluid governing equations, (ii) data-fit using
a function Ä(r ) = (r − re ) <{Aeiα(r −re ) }, where re locates the time-periodic forcing, A is a fixed complex con-
stant and α = 0.185 + 0.0015i is the complex radial wavenumber. (Note that, owing to use of the parallel-flow
approximation, the origin of the radial co-ordinate system is arbitrary.)
 VELOCITY–VORTICITY FORMULATION 153

that is shown for a function of the form

© ª
Ä(r ) = (r − re ) < A eiα(r −re ) for r ≥ re , (91)

where re defines the innermost radial location at which there is time-periodic forcing. The
term A is a complex constant that fixes the arbitrary phase and amplitude and α is the complex
radial wavenumber. The value α = 0.185 + 0.0015 i selected for the wavenumber was
determined, independently of the results of the numerical simulation, using the quasi-parallel
linear stability theory for the specified azimuthal wavenumber n/R = 0.041 and temporal
frequency β = 0.0061. It can be seen from Fig. 9 that deviations between the fitted function
and the simulation data are only found in the immediate vicinity of the locations where
time-periodic wall motion is used to generate the disturbance. They can thus be discounted
as being attributable to near-field effects. It should be noted that the disturbance would have
slowly decayed, over a radial lengthscale 1/αi ∼ 103 if algebraic growth were absent.
The effects on the modal coalescence of using the complete linearized Navier–Stokes
equations are illustrated in Fig. 10. Radial distributions of the ωθ amplitudes are plotted for
numerical simulations conducted with and without the use of the “parallel flow” approxi-
mation. The time-periodic forcing used to excite the disturbance was applied in the same
manner with the same normalization, in both cases. An amplitude distribution fitted using
the eigenvalue α determined from the standard linear stability theory is also displayed. In
order to facilitate comparisons, a slight geometric inconsistency is accommodated for the
“nonparallel” case. The azimuthal mode number n is permitted to be noninteger valued, just
as before. (Similar results were obtained from other simulations for which n was, more cor-
rectly, taken to be an integer.) Inspection of the plotted simulation data suggests that there
is still a region of algebraic radial growth when the complete linearized Navier–Stokes
equations are used. However, in so far as it can be identified with any precision, the region
of algebraic growth does not appear to extend very far beyond the source of the disturbance.
It can also be seen that, close to the source of the disturbance, the radial growth appears to
be slightly weaker than for standard linear theory. As expected, the algebraic growth gives
way to exponential radial growth, when convective instability sets in further outboard.

FIG. 10. Azimuthal vorticity amplitude |ωθ | at the surface of a rigid disk for the modal coalescence involving
the Type I and Type II forms of disturbance. R = 437, n/R = 0.041 and β = 0.0061. Simulation results obtained
with (i) the parallel-flow approximation, (ii) nonparallel mean flow effects included. (iii) Data-fit using a func-
tion |Ä(r )| = (r − re )|A|e−αi (r −re ) , where αi = 0.0015 is the imaginary part of the complex radial wavenumber
obtained from the solution of the parallel-flow linear stability eigenvalue problem, re locates the time-periodic
forcing and A is a fixed constant.
154 DAVIES AND CARPENTER

6. CONCLUSIONS

We have described a new velocity–vorticity formulation of the unsteady three-

dimensional Navier–Stokes equations. The approach is designed to be suitable for simulat-
ing the evolution of three-dimensional disturbances in boundary-layer flows, particularly
for flow-control applications. It is assumed that the undisturbed flow field is known. (As
shown in Appendix A, this is not an essential assumption.) Accordingly, only perturbations
to this known flow field are calculated. In general, the perturbations are of finite amplitude.
A key advantage of the formulation is that only three governing equations for three primary
dependent variables are required. The three primary variables are the perturbations to the
wall-normal velocity component and to the two vorticity components in the plane of the
wall. The secondary variables are the perturbations to the remaining velocity and vorticity
components and, for some applications, the pressure. These are defined explicitly in terms
of the primary variables. We show that, providing the primary variables can be made to
satisfy certain fairly weak conditions as z → ∞, our formulation is fully equivalent to the
Navier–Stokes equations in primitive-variable form.
For our formulation, only three dependent variables need to be computed and stored. This
contrasts favourably to previous three-dimensional, velocity–vorticity formulations (e.g.,
Fasel and his coworkers) which require all six velocity and vorticity components to be pri-
mary variables. Our approach even offers advantage compared with the primitive-variable
formulation with its four dependent variables. Thus, since our formulation only employs
three flow-field variables and three governing equations, it offers considerable scope for
improvements in algorithm design and computational times in applications involving di-
rect numerical simulations of boundary-layer disturbances. Another important advantage
offered by our approach lies in the treatment of the constraints placed on vorticity. The ap-
parent need for vorticity boundary conditions has been problematic with previous velocity–
vorticity methods. We show that the wall boundary conditions placed on velocity may be
linked to vorticity through integral constraints. Such an approach ensures solenoidal veloc-
ity and vorticity fields and provides a simple and rigorous alternative to vorticity boundary
conditions.
The numerical methods were presented in the context of application to the three-
dimensional boundary layer over a rotating disc. However, the corresponding methodology
for a Cartesian coordinate system should, for the most part, be fairly evident. Our scheme of
discretization employs spectral expansions in the wall-normal and azithmuthal (or spanwise)
directions and compact finite differences in the radial (or streamwise) direction. Discretiza-
tion in time takes the form of a three-point, backward-difference scheme. Time stepping is
based on a predictor-corrector scheme for the convective terms. For the application to the
rotating disc, but not for those using Cartesian coordinates, the predictor-corrector scheme is
also used for the Coriolis and some viscous terms. The remaining viscous terms (all viscous
terms in applications based on Cartesian coordinates) can be treated implicitly. It is shown
that we can retain an efficient line-iteration scheme, similar to that used by Fasel [21] and
in his later papers, but with the added advantage of spectral convergence in the wall-normal
direction. (For simulations that do not involve any interactive wall motion, the line-iteration
scheme can be simplified. The radial line iteration used to determine the vorticity may be
replaced by a more direct, radial line-marching procedure as described in Appendix C.)
For some of the applications reported above we have used a compliant wall. Wall com-
pliance is, of course, a passive method of flow control. Nevertheless, computationally it is
 VELOCITY–VORTICITY FORMULATION 155

much more challenging than active boundary-layer control methods such as suction and
blowing or the use of MEMS- or micro-actuators. In such cases, the wall boundary con-
ditions are normally specified a priori. Whereas for a compliant wall, the calculation of
wall and flow dynamics must be carried out interactively. In such circumstances, it can be
difficult to achieve a stable numerical scheme. In our previous work [14] on the simulation
of the development of two-dimensional boundary-layer disturbances over compliant walls,
we overcame the problem of achieving a stable scheme by combining the flow and wall
momenta as, effectively, a single variable. In the present paper, we show how this approach
can be extended to the three-dimensional case.
The applications presented in the present paper concern convective disturbances devel-
oping in the boundary layer over a rotating disc. (Investigations of absolute instabilities in
the same system will be reported in a separate paper. A preliminary account is given in
reference [16].) The simulations are restricted to small-amplitude disturbances for which
the governing equations can be linearized. This permits us to decouple the azimuthal modes
and compute them separately, thereby making the simulations much less computationally
expensive. It also allows us to validate the methods by comparison with known results from
standard linear stability analysis. We excite the most unstable stationary Type I instability
by introducing small-amplitude bumps on the disc surface. Good agreement is found with
previous linear-stability results. In our standard formulation, we use the full linearized gov-
erning equations without suppressing the slow radial variation of the disturbances. In fact,
we find that the results are not greatly different from the standard, “parallel-flow,” linear
stability theory which omits the slow radial variation. This is in accord with the PSE re-
sults of Malik and Balakumar [47]. We have also studied the travelling forms of the Type I
and Type II instabilities over rigid and compliant walls. Finally, we simulated the case
where the standard, linear, stability theory predicts a coalescence between Types I and II
eigenmodes. Theory suggests that, even though the disturbances are convectively stable,
coalescence implies localized algebraic growth. This is confirmed by the “parallelized”
simulations. When the full linearized governing equations were used, localized algebraic
growth was still observed, but exponentially growing convective instability set in at slightly
more outboard locations and soon dominated the algebraic growth.
The present application for the rotating disc has been restricted to small-amplitude dis-
turbances and only linearized governing equations were used. It is not too difficult to see
how the scheme could be generalized to allow the simulation of nonlinear finite-amplitude
disturbances. In this case, the nonlinear terms introduce coupling between the various
azimuthal modes and we would be unable to compute them separately. However, if, like the
convective terms in the linearized case, the nonlinear terms were treated explicitly within
the time-marching procedure, essentially the same line-iteration process could be used for
the nonlinear computations. The main difference would be the need to repeat the radial line-
iteration solution procedure for every resolved azimuthal mode, possibly using some form
of parallel processing. The nonlinear terms would only generate quantities which remained
fixed within the radial line iteration undertaken for each time step, in much the same manner
as was described for the convective terms in the linearized case.
We end by briefly mentioning some of the further applications of the new velocity–
vorticity formulation that have been made to date. This serves to emphasize the fact that the
utility of the formulation is by no means restricted to the rotating disc flow that we considered
in detail in the present paper. For example, the linearized version of the governing equations
has been used to simulate interactions between MEMS actuators and various forms of
156 DAVIES AND CARPENTER

small-amplitude three-dimensional disturbances developing in the Blasius boundary layer

and other laminar boundary layers. These simulations were conducted using essentially the
same numerical scheme as described in the present paper. The only significant change, aside
from the incorporation of MEMS devices and some new methods of disturbance generation,
was that a Cartesian coordinate system was used, instead of the cylindrical polar coordinates
that were more appropriate for the rotating disc. Careful code validation was undertaken,
in the same manner as for the simulations reported in the present paper. In particular, we
found that it was possible to reproduce standard linear stability results for both two- and
three-dimensional disturbances developing within the boundary layer. We were also able
to undertake more demanding simulations that involved the development of disturbances
that extended out beyond the boundary layer and into the free-stream. More details can be
found in reference [11].
In addition to studies conducted for three-dimensional linear disturbances, we have un-
dertaken some simulations of two-dimensional, fully nonlinear, disturbance development
in Blasius flow [34]. (A short account of the two-dimensional restriction of the velocity–
vorticity formulation is given in Appendix E.) The disturbance evolution was investigated
both with and without the use of the parallel flow approximation. Good agreement was
found with the behavior predicted by large Reynolds number asymptotic theories. More-
over, the numerical simulation code remained robust even when the Reynolds number, based
on displacement thickness, was taken to be in excess of 105 .
A two-dimensional version of our velocity–vorticity formulation has also been used
to compute the steady wake flow near the trailing edge of an aligned flat plate. For this
inherently nonlinear computation, the vorticity integral constraint used to impose no-slip
along the surface of the plate was replaced by a center-line symmetry condition within the
wake itself. The computed steady flow was found to develop, along the streamwise direction,
in precisely the manner predicted by triple-deck asymptotic theory (see, for example, [35]).

APPENDIX A

Velocity–Vorticity Formulation for the Total Flow and for Finite Domains
For the present study of the rotating-disc boundary layer, the undisturbed flow can be
determined directly from the solution of a set of ordinary differential equations [55]. Thus,
there would be little interest in using our formulation to determine the undisturbed flow.
This may not be the case for other potential applications. We will now describe how our
formulation may be modified to deal with the undisturbed flow or, more generally, with the
total flow. If we set U B = Ω B = 0 in the decomposition defined in Eqs. (6a) and (6b) then
the total velocity and vorticity field U, Ω are identical with the fields previously taken to be
perturbations, and the definition of the convective quantity N becomes

N = Ω × U.

If, in order to conform with the notation used in the main text, we take the components
of U = (u, v, w) and Ω = (ωx , ω y , ωz ), the three governing equations (7)–(9) can now be
interpreted as equations for the total-flow primary variables. It remains to specify the total-
flow secondary variables {u, v, ωz } in terms of the primary variables. Allowing for the fact
 VELOCITY–VORTICITY FORMULATION 157

that U may no longer vanish for z → ∞, we set

Z ∞µ ¶
∂w
u=− ωy + dz + ū ∞
z ∂x
Z ∞µ ¶
∂w
v= ωx − dz + v̄∞
z ∂y
Z ∞µ ¶
∂ωx ∂ω y
ωz = + dz + ω̄z∞ ,
z ∂x ∂y

where ū ∞ , v̄∞ are the prescribed velocity components for z → ∞ and

∂ v̄∞ ∂ ū
ω̄z∞ = − ∞.
∂x ∂y

For example, for a zero-pressure gradient flow over a flat plate we would simply have

ū ∞ = 1, v̄∞ = ω̄z∞ = 0.

The argument used to recover the primitive-variables form of the Navier–Stokes equations
for the perturbation flow fields can be reapplied to the case of the total flow fields. The only
difference is that the conditions which must be imposed on the primary variables for z → ∞,
in order to recover the incompressibility constraint and the z-component of the vorticity
transport equation, both take a slightly more complicated form. For instance, it is necessary
to ensure that
µ ¶
∂w ∂ ū ∞ ∂ v̄∞
→− + for z → ∞
∂z ∂x ∂y

in order to obtain the incompressibility constraint.

It is of interest to note that the formulation could, in principle, be extended to cases where
there is an upper boundary at z = L. This would involve replacing the z → ∞ upper limit by
z = L in the integrals defining the secondary variables. It would also be necessary to impose
the incompressibility constraint and the z-component of the vorticity transport equation at
the location z = L. For example, if there were a rigid stationary boundary located at z = L
(where L is constant), then the definitions of secondary variables would be given by
Z Lµ ¶
∂w
u=− ωy + dz,
z ∂x
Z Lµ ¶
∂w
v= ωx − dz,
z ∂y
Z Lµ ¶
∂ωx ∂ω y
ωz = + dz,
z ∂x ∂y

while the conditions to be imposed at z = L would take the simple form

µ ¶
∂w 1 ∂ ∂ωx ∂ω y
= + = 0.
∂z R ∂z ∂x ∂y
158 DAVIES AND CARPENTER

APPENDIX B

Coordinate Mapping for the Case of a Nonplanar Boundary

The system of equations that comprises the new velocity–vorticity formulation is first
described in the main part of the text for a general configuration where there is a bounding
surface located at z = η(x, y, t). However, for the specific application that is considered in
detail, namely the evolution of small-amplitude disturbances in the rotating disk boundary
layer, a linearization is performed, giving boundary conditions that are applied at z = 0. In
this section we provide a brief account of how the velocity–vorticity formulation, including
the boundary conditions, can be implemented in a computationally convenient fashion when
the flow boundary is taken to be more genuinely nonplanar and when it is not appropriate
to apply any form of linearization.
We suppose, as before, that the flow occupies the semi-infinite region region η ≤ z < ∞
and introduce the simple change of coordinates Z = z − η(x, y, t), which maps the bound-
ary at z = η(x, y, t) on to the plane defined by Z = 0. (More generally we could also
incorporate a transformation of the form X = f (x, y, t), Y = g(x, y, t) in the x-y plane.
Such a transformation was used, implicitly, when cylindrical polar coordinates were ap-
plied in a rotating frame for our study of the rotating disk boundary layer flow. How-
ever, we omit any further details here in the interests of brevity.) By making the substi-
tutions

ω̃x (x, y, Z , t) = ωx (x, y, z, t),

ω̃ y (x, y, Z , t) = ω y (x, y, z, t),
w̃(x, y, Z , t) = w(x, y, z, t),

for the primary variables {ωx , ω y , w}, and similarly for the secondary variables {u, v, ωz },
the two vorticity transport equations, the Poisson equation and the secondary variable def-
initions can all be transformed in a straightforward manner to yield a system in which
both the primary variables and the secondary variables, as well as the differential and
integral operators that act upon them, are expressed in terms of Z , rather than z. Of par-
ticular interest are the transformed definitions of the secondary variables. These take the
form
Z ∞µ ¶
∂ w̃ ∂η
ũ = − ω̃ y + dZ − w̃,
Z ∂x ∂x
Z ∞µ ¶
∂ w̃ ∂η
ṽ = ω̃x − dZ − w̃,
Z ∂y ∂y
Z ∞µ ¶
∂ ω̃x ∂ ω̃ y ∂η ∂η
ω̃z = + dZ + ω̃x + ω̃ y .
Z ∂x ∂y ∂x ∂y

As would be expected, the secondary variables are still defined explicitly in terms of the
primary variables. The only change is that there are some additional terms that involve the
slope of the nonplanar physical boundary. Using the above definitions of the secondary
variables, it may be readily shown that the no-slip conditions stated in Section 2.2 can be
 VELOCITY–VORTICITY FORMULATION 159

cast as vorticity integral conditions that take the form

Z ∞ Z ∞
∂ w̃ ∂η
ω̃ y d Z = −u w − dZ − ww ,
0 0 ∂x ∂x
Z ∞ Z ∞
∂ w̃ ∂η
ω̃x d Z = vw + dZ + ww ,
0 0 ∂y ∂y

where u w , vw , ww are the prescribed velocities for z = η, as in Eqs. (25)–(27). Thus, the
no-slip conditions can be imposed using constraints that only involve integration over the
fixed semi-infinite interval 0 ≤ Z < ∞. Finally, we note that the wall-normal coordinate
mapping considered in Section 2.3 can be applied to Z , rather than z, to map the semi-infinite
domain Z ∈ [0, ∞) on to a finite computational domain. Thus, we set

l
ζ = ,
Z +l
where l is a mapping constant, to obtain a computational domain such that ζ ∈ (0, 1]. A
discretization based upon Chebyshev expansions can then be applied for the wall-normal
variation, as described in Section 4.

APPENDIX C

An Alternative Time-Stepping Method

In Section 4.2 we described a time-stepping method that employs a second-order, three-
point, backward-difference formula for the time derivatives. This formula was applied in
conjunction with a predictor-corrector method for the convective terms in the vorticity
transport equations. The decision to implement the time integration in such a fashion was
motivated by our previous experience in simulating fluid-structure interactions for a two-
dimensional flow [14]. We had found that the use of a fully implicit time-stepping procedure
could provide a robust means of overcoming numerical stability problems. However, in the
absence of any interactive coupling between the motion of the fluid and its bounding walls,
the conditions required for numerical stability can be relaxed. This allows more explicit
time-stepping methods to be adopted.
Accordingly, for cases without any interactive wall motion coupling, we have developed
and implemented a Crank–Nicholson/Adams–Bashforth type of method for the time step-
ping. In this alternative numerical scheme, all the terms appearing in the vorticity transport
equations, with the exception of those viscous terms involving z-derivatives, can be treated
in an explicit fashion by means of a predictor-corrector scheme. It is also possible to treat
the vorticity integral constraints in an explicit manner. The advantage of the resulting dis-
cretization scheme is that the two primary-variable vorticity components can be determined
directly, at every radial location, prior to the solution of the Poisson equation for w. The
modified scheme for obtaining the each vorticity component thus involves only a single
radial line march, rather than the more expensive radial line iteration described previously.
Line iteration is still needed to solve the Poisson equation, but w can now be constructed
using already determined values of the vorticity.
The improvements to the time-stepping procedure can be illustrated by re-employing the
notation that was first introduced in Section 4.2. In the same manner as for Eq. (84), the
160 DAVIES AND CARPENTER

discretized equations to be solved at each time step may be cast in the form,

SΩrl j = Crl j ,
SΩθl j = Cθl j ,
T j Wlj = Bwl j + Dwl j ,

where the matrix S is now independent of the radial location. The right-hand sides of
the two discretized vorticity transport equations can be computed explicitly, using known
values of the primary variables. Thus, the terms Brl j and Bθl j appearing in Eq. (84) are no
longer present. Because the matrix S remains pentadiagonal, apart from its first row, the
Chebyshev vectors Ωrl j and Ωθl j can be determined from Crl j and Cθl j in a very efficient
manner. The right-hand side of the discretized Poisson equation contains an additional term
Dwl j , denoting the quantities that can be computed directly, without any need for iteration,
from the previously determined values of Ωrl j and Ωθl j . The vector Bwl j is now defined so
that it depends only on the values of w at locations that are radially inward or outward from
the selected radial position. Plainly, then, radial line iteration is still required in order to
obtain the Chebyshev vectors Wlj .
To summarize, it may be stated that in the alternative time-stepping scheme, radial line
iteration is replaced by a single radial line march for each of the vorticity transport equations;
but a slightly modified, radial line iteration is retained for the Poisson equation. In the main
body of the text there are a number of references to radial line iterations. When interactive
wall motion is not involved, most of these references should, strictly speaking, be replaced by
a reference to the radial line marching/radial line iteration combination outlined immediately
above.

APPENDIX D

Chebyshev Discretization for the Secondary Variables

The definitions of the secondary variables take the general form
Z ∞
f = g dz,
z

where f is the secondary variable and g is specified in terms of the primary variables
and their derivatives in the planes orthogonal to the z-direction. Introducing the coordinate
mapping ζ = l/(z + l), which maps the semi-infinite interval [0, ∞) to the finite interval
(0, ζ ], yields the relationship
Z ζ
g
f =l dζ.
0 ζ2

Setting

∞
X X ∞
g
g= gk T2k−1 (ζ ), h = = h k T2k−1 (ζ ),
k=1
ζ 2
k=1
 VELOCITY–VORTICITY FORMULATION 161

and using the fact that

1
ζ 2 T2k−1 = (T2k+1 + 2T2k−1 + T2k−3 ),
4

we can conclude that

1
gk = (h k−1 + 2h k + h k+1 ).
4

We also have
Z ∞
ζ
1X1
h dζ = (h k − h k+1 )(T2k (ζ ) − T2k (0)),
0 4 k=1 k

where we have employed the identity

1 0 1
4T2k−1 = T2k − T0
k k − 1 2(k−1)

for k > 1. Thus, we can set

∞
X
f = f k (T2k (ζ ) − T2k (0)) ,
k=1

where

l
fk = (h k − h k+1 ).
4k

Hence, as may easily be checked,

l
(k + 1) f k+1 + 2k f k + (k − 1) f k−1 = (h k−1 + 2h k + h k+1 − (h k + 2h k+1 + h k+2 )).
4

Using the relation between the Chebyshev coefficients of g and h given earlier, we can
remove all reference to the intermediate variable h. We thus obtain

(k + 1) f k+1 + 2k f k + (k − 1) f k−1 = l(gk − gk+1 ),

which, when truncated to finite order, is the relationship that was used to derive the tridiag-
onal schemes for the secondary-variable Chebyshev coefficients given in Eqs. (68)–(70).
Finally, we outline the method used to compute the constants pk which are required for
the numerical evaluation of integrals across the boundary layer. Such integrals appear in the
integral constraints imposed on the vorticity and also in the expression for the perturbation
fluid pressure at the solid surface. For the case of untruncated Chebyshev expansions, the
constants pk may be defined by setting
Z ∞ ∞
X
f |z=0 = g dz = pk gk .
0 k=1
162 DAVIES AND CARPENTER

For the finite-order truncation, the constants may be determined by first computing the
matrix P jk which satisfies

(k − 1)P j,k−1 + 2k P j,k + (k + 1)P j,k+1 = δ j,k − δ j,k+1

for j, k = 1, . . . , M, and then setting

X
M/2
pk = 2l Pk,2 j−1 ,
j=1

as may be demonstrated by applying the previously obtained relationship between the

Chebyshev coefficients f k , gk to evaluate f |z=0 = f |ζ =1 when gk = δ j,k .

APPENDIX E

Two-Dimensional Version of the Velocity–Vorticity Formulation

With a view to the applications mentioned in the concluding section of the main part of the
paper, we present the two-dimensional restriction of the velocity–vorticity formulation for
the case of deviations from a prescribed parallel flow. If we let u, w denote streamwise and
wall-normal velocity components corresponding to a two-dimensional perturbation from
the parallel flow U = (U p (z), 0, 0), and let ω denote the associated vorticity perturbation,
then the Navier–Stokes equations can be cast in the form

∂ω ∂ω ∂(uω) ∂(wω) 1 1
+ Up + + + wU p00 = ∇ 2 ω + U p000 ,
∂t ∂x ∂x ∂z R R
∂ω
∇ 2w = − ,
∂x
Z ∞µ ¶
∂w
u=− ω+ dz,
z ∂x

where ∇ 2 now denotes the two-dimensional Laplacian.

It may be seen that, just as for the three-dimensional formulation, the streamwise pertur-
bation velocity component u is defined explicitly in terms of the wall-normal component
w and the vorticity perturbation ω. Thus, u may be eliminated from the vorticity transport
equation, leaving a system of two partial differential equations for the two unknowns ω,
w. It should also be noted that the profile U p (z) of the parallel flow may be chosen to take
any form that is convenient for the problem at hand: it certainly does not need to be an
exact solution of the Navier–Stokes equations. Parallelized simulations can be conducted
by simply omitting the final term 1/RU p000 from the right-hand side of the transport equation.
This term would otherwise act as a source for the generation of a nonparallel base flow.
We suppose that the boundary conditions at the location z = η(x, t) of the solid surface
bounding the flow take the form

u(x, η, t) = u w (x, t), w(x, η, t) = ww (x, t).

The integral condition that constrains the evolution of the vorticity is then given by
Z ∞ Z ∞
∂w
ω dz = −u w − dz.
η η ∂x
 VELOCITY–VORTICITY FORMULATION 163

As before, this is fully equivalent to the no-slip condition. Continuity holds at all locations
provided only that the normal-velocity satisfies the condition

∂w
→0 for z → ∞.
∂z
As was mentioned briefly in the main text, the two-dimensional formulation has also
been employed to compute the steady flow near the trailing edge of a flat plate. Essentially,
all that was required was that the vorticity integral condition be exchanged for the simple
center-line symmetry condition

ω(x, 0, t) = −U p0 (0),

for locations in the wake beyond the end of the plate.

Finally, we remark that some of the ideas that we developed within our velocity–vorticity
formulation could also find an application in a streamfunction–vorticity formulation for the
case of two-dimensional flow. In particular, if we introduce a streamfunction ψ, then the
no-slip condition can be very simply translated into a vorticity integral constraint of the form
Z ∞ Z ∞
∂ 2ψ
ω dz = −u w + dz.
η η ∂x2

Coupling the enforcement of this condition to the solution of the vorticity transport equation
provides a rigorous alternative to the employment of an artificial wall-vorticity boundary
condition.

ACKNOWLEDGMENTS

The research described in this paper was carried out in part with the support of the Engineering and Physical
Sciences Research Council.

REFERENCES

1. P. Balakumar and M. R. Malik, Traveling disturbances in rotating-disk flow, Theoret. Comput. Fluid Dyn. 2,
125 (1990).
2. K. E. Barrett, A variational principle for the stream function-vorticity formulation of the Navier–Stokes
equations incorporating no-slip conditions, J. Comput. Phys. 26, 153 (1978).
3. F. Bertagnolio and O. Daube, Solution of the div-curl problem in generalized curvilinear coordinates,
J. Comput. Phys. 138, 121 (1997).
4. F. P. Bertolotti, Th. Herbert, and P. R. Spalart, Linear and nonlinear stability of the Blasius boundary layer,
J. Fluid Mech. 242, 441 (1992).
5. V. Bojarevics, G. Tinios, K. Pericleous, and M. Cross, CFD modelling of magnetic levitation casting. In
Proc. Int. Cong. Electromagnetic Processing of Materials. (Centre Français de l’Electricité, 1997), Vol. 2,
pp. 211–216.
6. C. Canuto, M. Y. Hussaini, A. Quarteroni, and T. A. Zang, Spectral Methods in Fluid Dynamics (Springer-
Verlag, Berlin/NewYork, 1988).
7. A. Campion-Renson and M. J. Crochet, On the stream function-vorticity finite element solution of Navier–
Stokes equations, Int. J. Num. Meth. Eng. 12, 1809 (1978).
8. H. A. Carlson and J. L. Lumley, Active control in the turbulent wall layer of a minimal flow unit, J. Fluid
Mech. 329, 341 (1996).
9. H. A. Carlson and J. L. Lumley, Flow over an obstacle emerging from the wall of a channel, AIAA J. 34, 924
(1996).
164 DAVIES AND CARPENTER

10. P. W. Carpenter, Status of transition delay using compliant walls, in Viscous Drag Reduction in Boundary
Layers edited by D. M. Bushnell and J. N. Hefner, (AIAA Press, Washington, DC, 1990), pp. 79–113.
11. P. W. Carpenter, D. A. Lockerby, and C. Davies, Numerical Simulation of the Interaction of MEMS Actuators
and Boundary layers, Technical Paper 2000–4330 (AIAA Press, Washington, DC, 2000).
12. A. J. Cooper and P. W. Carpenter, The stability of rotating-disc boundary-layer flow over a compliant wall,
Part 1. Type I and II instabilities, J. Fluid Mech. 350, 231 (1997).
13. O. Daube, Resolution of the 2D Navier–Stokes equations in velocity–vortcity form by means of an influence
matrix technique, J. Comput. Phys. 103, 402 (1992).
14. C. Davies and P. W. Carpenter, Numerical simulation of the evolution of Tollmien–Schlichting waves over
finite compliant panels, J. Fluid Mech. 335, 361 (1997).
15. C. Davies and P. W. Carpenter, Instabilities in a plane channel flow between compliant walls, J. Fluid Mech.
352, 205 (1997).
16. C. Davies and P. W. Carpenter, A Novel Velocity–Vorticity Formulation of the Navier–Stokes Equations
Part II: Global Behaviour Corresponding to the Absolute Instability of the Rotating-Disk Boundary Layer,
Technical Report 2000/1 (Fluid Dynamics Research Centre, Univ. Warwick, 2000).
17. R. T. Davis, Numerical solution of the Navier–Stokes equations for symmetric laminar incompressible flow
past a parabola, J. Fluid Mech. 51, 417 (1972).
18. S. C. R. Dennis and J. D. Hudson, An h4 accurate vorticity-velocity formulation for calculating flow past a
cylinder, Int. J. Num. Meth. in Fluids 21, 489 (1995).
19. S. C. R. Dennis and J. D. Hudson, Methods of solution of the velocity-vorticity formulation of the Navier–
Stokes equations, J. Comput. Phys. 122, 300 (1995).
20. D. C. R Dennis and L. Quartapelle, Some uses of Green’s theorem in solving the Navier–Stokes equations,
Int. J. Numer. Meth. Fluids 9, 871 (1989).
21. H. Fasel, Investigation of the stability of boundary layers by a finite difference model of the Navier–Stokes
equations, J. Fluid Mech. 78, 355 (1976).
22. H. Fasel, Recent developments in the numerical solution to the Navier–Stokes and hydrodynamic stability
problems, in Computational Fluid Dynamics (Hemisphere, 1980), pp. 167–280.
23. H. Fasel and U. Konzelmann, Non-parallel stability of a flat-plate boundary layers using the complete Navier–
Stokes equations, J. Fluid Mech. 221, 311 (1990).
24. H. Fasel, U. Rist, and U. Konzelmann, Numerical investigation of the three-dimensional development in
boundary-layer transition, AIAA J. 28, 29 (1990).
25. T. B. Gatski, Review of incompressible fluid flow computations using the vorticity–velocity formulation,
Applied Numer. Meth. 7, 227 (1991).
26. T. B. Gatski, C. E. Grosch, and M. E. Rose, A numerical study of the 2-dimensional Navier–Stokes equations
in vorticity–velocity variables, J. Comput. Phys. 48, 1 (1982).
27. T. B. Gatski, C. E. Grosch, and M. E. Rose, A numerical-solution of the Navier–Stokes equations for 3-
dimensional, unsteady, incompressible flows by compact schemes, J. Comput. Phys. 82, 298 (1989).
28. P. M. Gresho, Some interesting issues in incompressible fluid dynamics, both in the continuum and in numerical
simulation, Adv. Applied Mech. 23, 45 (1991).
29. P. M. Gresho, Incompressible fluid dynamics: some fundamental formulation issues, Ann. Rev. Fluid Mech.
23, 413 (1991).
30. G. Guevremont, W. G. Habashi, and M. M. Hafez, Finite element solution of the Navier–Stokes equations by
a velocity–vorticity method, Int. J. Num. Meth. Fluids 10, 461 (1990).
31. Th. Herbert, Parabolized stability equations, Ann. Rev. Fluid Mech. 29, 245 (1997).
32. L. M. Hofmann and Th. Herbert, Disturbances produced by motion of an actuator, Phys. Fluids 9, 3727 (1997).
33. L. M. Hofmann and Th. Herbert, Reproducing the flow response to actuator motion, J. Comput. Phys. 142,
264 (1998).
34. S. Houten, J. J. Healey, and C. Davies, Nonlinear evolution of Tollmien-Schlichting waves at finite Reynolds
numbers, in Proc. IUTAM Symp. on Laminar-Turbulent Transition, Sedona, USA, edited by H. Fasel and
W. S. Saric (Springer-Verlag, Berlin/New York, 1999).
35. C. E. Jobe and O. R. Burggraf, The numerical solution of the asymptotic equations of trailing edge flow, Proc.
Roy. Soc. London A 340, 91 (1974).
 VELOCITY–VORTICITY FORMULATION 165

36. Th. von Kármán, Über laminare und turbulente Reibung. Z. Angew. Math. Mech. 1, 233 (1921).
37. J. Kim, P. Moin, and R. Moser, Turbulence statistics in fully developed channel flow at low Reynolds number,
J. Fluid Mech. 177, 133 (1987).
38. L. Kleiser and T. Zang, Numerical simulation of transition in wall-bounded shear flows, Ann. Rev. Fluid Mech.
23, 495 (1991).
39. M. Kloker, U. Konzelmann, and H. Fasel, Outflow boundary conditions for spatial Navier–Stokes simulations
of transition boundary-layers, AIAA J. 31, 620 (1993).
40. W. Koch, Direct resonance in Orr–Sommerfeld problems, Acta Mechanica 58, 11 (1986).
41. Y. M. Koh and P. Bradshaw, Numerical solution of two-dimensional or axisymmetric incompressible flow
using the vorticity equations, KSME Journal 8, 264 (1994).
42. L. D. Kral and H. F. Fasel, Direct numerical-simulation of passive control of 3-dimensional phenomena in
boundary-layer-transition using wall heating, J. Fluid Mech. 264, 213 (1994).
43. R. J. Lingwood, Absolute instability of the boundary layer on a rotating disc, J. Fluid Mech. 299, 17 (1995).
44. R. J. Lingwood, An experimental study of absolute instability of the rotating-disk boundary-layer flow,
J. Fluid Mech. 314, 373 (1996).
45. C. Liu and Z. Liu, High order finite difference and multigrid methods for spatially evolving instability in a
planar channel, J. Comput. Phys. 106, 92 (1993).
46. M. R. Malik, The neutral curve for stationary disturbances in rotating-disk flow, J. Fluid Mech. 164, 275
(1986).
47. M. R. Malik and P. Balakumar, Nonparallel stability of rotating disk flow using PSE, in Instability, Transition
and Turbulence, edited by M. Y. Hussaini, A. Kumar, and C. L. Streett (Springer-Verlag, Berlin/New York,
1992), pp. 168–180.
48. M. R. Malik, S. P. Wilkinson and S. A. Orszag, Instability and transition in rotating disk flow, AIAA J. 19,
1131 (1981).
49. H. L. Meitz and H. F. Fasel, A compact-difference scheme for the Navier–Stokes equations in vorticity-velocity
formulation, J. Comput. Phys. 157, 371 (2000).
50. R. Rathnasingham and K. S. Breuer, Coupled fluid-structural characteristics of actuators for flow control,
AIAA J. 35, 832 (1997).
51. H. L. Reed and W. Saric, Stability of three-dimensional boundary layers, Ann. Rev. Fluid Mech. 21, 235
(1989).
52. D. Rempfer and H. F. Fasel, Dynamics of 3-dimensional coherent structures in a flat-plate boundary-layer,
J. Fluid Mech. 275, 257 (1994).
53. D. Rempfer and H. F. Fasel, Evolution of 3-dimensional coherent structures in a flat-plate boundary-layer,
J. Fluid Mech. 260, 351 (1994).
54. U. Rist and H. Fasel, Direct numerical simulation of controlled transition in a flat-plate boundary layer,
J. Fluid Mech. 298, 211 (1995).
55. H. Schlichting, Boundary-Layer Theory, 7th ed. (McGraw-Hill, NewYork, 1979), pp. 102–104.
56. W. Z. Shen and T. Ph. Loc, Numerical method for unsteady 3D Navier–Stokes equations in velocity–vorticity
form, Comput. Fluids 26, 193 (1997).
57. C. G. Speziale, On the advantages of the vorticity velocity fomulation of the equations of fluid-dynamics,
J. Comput. Phys. 73, 476 (1987).
58. M. Turkyilmazoglu and J. S. B. Gajjar, Absolute and Convective Instabilities in Incompressible Boundary
Layer on a Rotating Disc, Internal Report CLSCM-1998-002 (University of Manchester, 1998).
59. J. T. Trujillo and G. E. Kaniadakis, A penalty method for the vorticity-velocity formulation, J. Comput. Phys.
149, 32 (1999).
60. E. Weinan and J. G. Liu, Vorticity boundary condition and related issues for finite difference schemes,
J. Comput. Phys. 124, 368 (1996).
61. X. H. Wu, J. Z. Wu, and J. M. Wu, Effective vorticity-velocity formulations for 3D incompressible viscous
flows, J. Comput. Phys. 122, 68 (1995).