Vous êtes sur la page 1sur 24

Numerical Simulation of Stokes Flow Down an Inclined Plane

Author(s): D. Rh. Gwynllyw and D. H. Peregrine


Source: Proceedings: Mathematical, Physical and Engineering Sciences, Vol. 452, No. 1946
(Mar. 8, 1996), pp. 543-565
Published by: Royal Society
Stable URL: http://www.jstor.org/stable/52837
Accessed: 28-06-2016 08:17 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact support@jstor.org.

Royal Society is collaborating with JSTOR to digitize, preserve and extend access to Proceedings:
Mathematical, Physical and Engineering Sciences

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow
down an inclined plane
BY D. RH. GWYNLLYWt AND D. H. PEREGRINE
School of Mathematics, University of Bristol, Bristol BS8 1 TW, UK

A boundary-integral equation method is applied to the problem of evolving two-


dimensional flow of a viscous liquid with a free surface down an inclined plane wall.
The flow is assumed to have a sufficiently low Reynolds number that Stokes flow is
a good approximation; the stream function then satisfies the biharmonic equation.
Numerical solutions are found by using a boundary integral equation applied to both
harmonic functions that appear in the Almansi biharmonic representation. Different
approaches to the linear algebraic scheme and different time-stepping routines are
discussed. The motion of the free surface with, or without, surface tension is then
modelled by a simple time-stepping routine.
The numerical scheme is shown to give accurate representations of the free surface
up to and beyond the point of overturning. One example is illustrated in which the
overturning evolves until the free surface is close to forming a cusp. Solutions are
also illustrated by two spatially periodic wave profiles: one is close to a time periodic
flow and the other becomes a steep nearly steady wave.

1. Introduction

Flow of a thin liquid film driven either by gravity, centrifugal force, or surface stresses
caused by the flow of another fluid, is often important in industrial processes which
involve heat and mass transfer or coating flows. In many cases the Reynolds num-
ber of the flow is appreciably less than unity and the Stokes flow approximation is
applicable. For finite Reynolds number, the governing equation of the flow is the
Navier-Stokes equation; solving such an initial boundary value problem must in-
volve the setting up of internal grids at each time step as in the boundary-fitted
coordinates method used in Ohring & Lugt (1991) for viscous flow. For Stokes flow
however, the governing biharmonic equation can be represented by a boundary in-
tegral equation which can significantly improve the computational efficiency of the
numerical simulation of the motion of the free surface.
There are many cases of the boundary integral equation (BIE) method being used
to compute the free surface of steady Stokes flow by combining the method with an
iterative scheme to search for the a priori unknown free-surface solution (see, for
example, Pozrikidis 1988; Ingham & Kelmanson 1984). The use of BIE methods to
solve for evolving Stokes flow is applied mainly to the deformation of drops (see, for
example, Stone & Leal 1989; Kuiken 1990). Boundary integrals for evolving flows
are more common for irrotational inviscid flow, where the governing equation is

t Department of Mathematics, The University of Wales, Penglais, Aberystwyth SY23 3BZ, UK.

Proc. R. Soc. Lond. A (1996) 452, 543-565 () 1996 The Royal Society
Printed in Great Britain 543 TjFX Paper

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
544 D. Rh. Gwynllyw and D. H. Peregrine

the Laplace's equation. A very efficient example is the model developed by Dold &
Peregrine (1986), Tanaka et al. (1987), Cooker et al. (1990) and Cooker & Peregrine
(1990), which has applications to large sea waves.
In the case of Stokes flow, the solution of the biharmonic equation is usually
expressed either as a coupling of Green's integral for the harmonic vorticity and the
biharmonic stream function (Kuiken 1990) or as a single equation based upon the
fundamental solution of the biharmonic equation (Pozrikidis 1988). Either way the
complexity of the integrals requires that evaluation of their numerical counterparts
takes much longer than for Laplace's equation. One motivation for the present work
is to examine if the ideas used in the efficient Dold & Peregrine (1986) scheme could
be transferred to the biharmonic equation.
Dold & Peregrine (1986) use a BIE based upon the Cauchy integral to solve
Laplace's equation for the velocity potential. In calculating a numerical represen-
tation of the integral equation, a point discretization of the free surface is made,
which is in contrast to the segment discretization used by the large majority of other
BIE users. Due to the different nature of the two problems, we were not able to use
any of the special features used by Dold & Peregrine (1986) in the time stepping of
the irrotational flow. However, enough of the experience of the inviscid flow calcula-
tion can be applied to Stokes flow that we think it has an advantage over previous
approaches. We make use of the Almansi representation for the biharmonic function.
This gives two harmonic functions to be found, leading to a combination of two BIES,
each with the same matrix of coefficients in the numerical approximation.
It proved possible to test the accuracy of the numerical results by comparing
numerical solutions with the recent analytical results of Hopper (1990) for the motion
of surface tension driven flow from certain special initial geometries. For the case of
flow down the inclined plane, we compare with the approximate results of linear
theory and finite-amplitude long-wave analysis. This latter analysis has been used in
the search for forms of steady and time-periodic flows, which in the case of Stokes
flow has been unsuccessful. Our numerical results suggest that such flows do exist
and our illustrations include a wave form which is close to being steady and another
example where a travelling wave's profile varies in an almost periodic manner. A
final example of flow down a vertical wall has waves overturning until they are close
to forming a cusp and look like two-dimensional drops running down the wall.

2. The governing equations


The geometry of the flow is illustrated in figure 1. By taking all lengths to be
measured in units of h0, velocities in units of pgh/lft, and pressure in units of pgho,
the dimensionless variables become x' = x/ho, p' = p/pgho, u' = u,l/pgho and
t' - tpgho/l, where g is the acceleration of gravity, ft is viscosity, p is the density, p
is the pressure of the fluid, and ho the mean fluid thickness.
The dimensionless Navier-Stokes equation, having dropped the primes, becomes
Du
Re
Dt
= k-Vp +Au, (2.1)

where Re = p2gh3//t2 is the Reynolds number, k = (sin 3, cos 13) is the vertical unit
vector, and A is the Laplacian operator. Re is assumed to be sufficiently small such
that we can neglect the inertial term on the left-hand side of (2.1) and thus have
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 545
n

A ^^1~ --

Figure 1. Schematic diagram of the geometry of the flow.

Stokes flow. The two components of (2.1) are now


Op o_ A&p A &A'o
+sin3, =- -cosP. (2.2)
Ox Oy Oy Ox
The dimensionless form of the normal and tangential components of the free-
surface boundary conditions are

p2 2 + r, = O, on y = h(x, t), (2.3)

92v-1a = 0, on y = h(x, t), (2.4)


ON2 OT2
where a is the inverse of the Bond number given by a = /t/(pgh2), y being the surface
tension. The unit tangent and outward normal vectors are t and n, respectively, and
define the directional derivatives O/OT = t ' V and 0/ON = n V. n is the surface
curvature, positive for a surface concave away from the liquid, and h(x, t) is the
dimensionless height of the free surface and can be multivalued in x.
Differentiating (2.3) by T, and using the tangential component of equations (2.2)

=- aN +cosO, (2.5)
where 0 is the angle the free-surface tangent makes with the vertical
0 = arctan(-ny/nx),
where n-, ny are the x and y components of the unit normal, we obtain the following
boundary condition:
Od 0930 O
- ON
N + cos O + 2 -+ =- o, on y = h(x, t), (2.6)
ODT2dN OT
where w is the vorticity given by
A = --. (2.7)
Converting from Cartesian to curvilinear coordinates, as described in Kuiken (per-
sonal communication, 1989), equations (2.4) and (2.6) lead to the following dynamic
boundary conditions:
2/nss +- 2(/s)s - wn = -is - COS 0, on y = h(x(s), t), (2.8)
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
546 D. Rh. Gwynllyw and D. H. Peregrine

w + 2s,, - 2k,,n - 0, on y = h(x(s), t), (2.9)


where the subscripts 'n' and 's' refer to derivatives normal and tangential to the
surface, respectively. From (2.2) it is easily shown that the stream function is bihar-
monic, namely,
A2 = 0, Aw = 0. (2.10)
An initial velocity field is not specified in this problem since it is computed from the
initial free surface as part of the problem. Once the initial free surface is given, the
stream function follows from (2.10) and the boundary conditions (2.8), (2.9). The
kinematic boundary condition is given later in equation (4.9).

3. Method of solution

For convenience we consider spatially periodic flow and, as a result of the no-slip
condition at the wall, the stream function is also periodic.
We choose the Almansi representation for the biharmonic solution written in the
form

- = Y + X, (3.1)
giving

-2 , p -2 x, (3.2)
where X and X are two independent harmonic functions. The more popular Goursat
representation (Langlois 1964) gives rise to aperiodic harmonic functions which re-
quire some straightforward but inconvenient arithmetic for application to a periodic
problem. The Almansi functions X and X are also not necessarily periodic. If we con-
sider the perturbed stream function = ?- , where X is the unperturbed stream
function given by
= 1 (y2 _ - y3) sin I,
then the Almansi function ), corresponding to the biharmonic ', is periodic up to a
constant multiple of x. The Almansi representation is not unique since 9 is invariant
under the transformation

-- 0 + ni + n2x, X -- X - y(n + n2x), (3.3)


and hence forcing periodicity in the Almansi functions eliminates the free constant
n2. This is in addition to the stream function being non-unique in the sense that the
transformation

) -) + no (3.4)
does not change the physics of the fluid and hence would still satisfy the governing
equations. Using the stream function, b, requires slightly more arithmetic to ensure
periodicity than using the perturbed stream function, b. Thus, we shall only refer
to the perturbed flow and shall drop the hat notation.
Following the arguments of Jaswon & Symm (1977), it can easily be shown that
the Almansi functions in our model, for constant nl, are single valued, although this
is not the case for some other flow geometries.
By applying Cauchy's integral theorem on a smooth contour and rearranging, Dold
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 547

& Peregrine (1986) obtain the following BIE for a harmonic function 0:

7rn J- Im (R R) ' ds(x') Re e (R s) 's ds(x'), (3.5)

where the primes and non-primes refer to the variables being calculated at the surface
point corresponding to x' and x, respectively, F is a closed contour of the fluid and
R = X+iY is the complex position of the surface profile. The no-slip condition for the
stream function prevents us from following Dold & Peregrine (1986) in considering
the reflection of the governing functions in the plane wall. As in Dold & Peregrine
(1986) and Longuet-Higgins & Cokelet (1976), the surface and the plane wall are
mapped into an annulus by the conformal mapping

Qf2(, t) exp(-iR(c, t)), (3.6)


where ~ is both a point discretization label and a fluid particle label, t is time and
R((, t) = X({, t) + iY(S, t). This leads to the following BIE:

+ Im 2 ( - i' ) d' + Re ( 7 f d2', (3.7)


where ' is any harmonic function and ~ is the boundary point label parameter.
Here, the primes and absence of primes refer to the variables being calculated at
the boundary point labels J' and ~, respectively. The factor IRaj in the change of
integration variable, where

IRE = - (X(j, t) + iY(S, t))/0l,


cancels with its inverse in the transformation of the partial derivatives. The domain
F is the mapping of the free surface and domain B is the mapping of the wall. By
mapping the boundaries of one spatial period on to each other their contributions to
the Cauchy integral cancel.
In addition to the Cauchy integral and the dynamic boundary conditions, the
no-slip and impermeability condition at the plane wall must be considered. These
are

+ Xn =- b-b + Xb , olb - + Xb = 0,= (3.8)


respectively, where the plane wall is taken to be at y = b, and these superscripts
refer to values at the wall, for example, Ob(x) _ (x, b). The unit normal, n, and
unit tangent, t, on the plane wall are given by (0,-1) and (-1, 0), respectively, in
Cartesian coordinates.

4. Numerical representation
Considering (3.1), (3.2), (2.8), (2.9), (3.7) and (3.8), we take our vector of unknown
variables to be

S = (Of f Xf Xf b Xb)T (4.1)


where, for example, Of is the vector of numerical values of s(x, h) with the superscript
indicating values at points on the free surface.
The arc length derivatives are calculated using the 11-point scheme described in
Dold (1992). This gives a straightforward representation of the boundary conditions
as a matrix product with s. The Cauchy integral (3.7) is represented numerically as in
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
548 D. Rh. G'yidrllytll and D. H. Peregrine

Dold & Peregrine (1986). Since periodic flow is being considered and the integration
in (3.7) involves ~', which is discretized at equal intervals around the contour, the
best quadrature formula for integration is the simple trapezium approximation. Note
that S' has integer values but this does not mean that the discretization points
are equidistant in either the physical or transformed plane. Dold & Peregrine deal
with the singularities in the integrand by subtracting them out, which results in the
discrete representation of (3.7) becoming
Nf +Nb Nf +Nb

7r,(-) = E A(E, ')Ot( (') -O(o) + E B( ,$')t(j'), (4.2)


C'=1 C'=1

where

A(S, t) + iB(, A-~,


t) { 2(, t)/(
t) (,+(t)iB(~,
- 2(', t)),
t)if _
S ) (4.3)
'/ (4 3)
(, t)/2f2Q(, t), if C - ',
where Nf and Nb are the number of discretization points per wavelength at the free
surface and the wall, respectively. In our numerical scheme, these two parameters are
equal and denoted by N. VO, Ot and 't9 are the vectors of numerical values at the
discrete points of t9, i9 and 'd, respectively. Equation (4.2) numerically relates the
normal derivative of a generic harmonic function to its surface derivatives at the free
surface and the plane wall. Using t = IRS/st, s = \R 2,n and incorporating surface
differential operators as matrices, equation (4.2) is converted to a matrix equation
for (0df I,, ^ b,0b).
The flat plane conditions can be used to express the values of qb, Xb in terms of
0b, Xb. Hence, applying the converted version of (4.2) twice (d0 _ and 9 X),
together with the two dynamic boundary conditions, gives a linear algebraic scheme
(LAS) for solving for s, namely,
Gs = b. (4.4)
Matrix G is of size L x L where L _ 6N. The rank of G is L - 3 due to the
two free variables, n1 and no, from (3.3) and (3.4), respectively, and also from the
zero eigenvalue corresponding to the eigenvector e3 = (1,0,0,0,0,0), where 1 =
(1, 1,..., 1, 1). This latter zero eigenvalue is due to the fact that the Cauchy equations
(4.2) only relate to the values of the derivatives of the Almansi functions and the
relation between the values of the harmonic functions at the wall and the free surface
is ignored. Fixing an arbitrary value at an arbitrary free-surface discretization point,
the resulting set of numerical values for q at the free surface, q*, is related to the
required numerical values, Of, by

of - * +-n31. (4.5)
n3 is an unknown constant for each flow profile and is given analytically by

3 =- (X, b) - o(X,h) +j O(X, y) dy ], h h(X). (4.6)


Since 4b is calculated directly from the linear system, that is it is not affected by
this arbitrariness, n3 can be evaluated numerically. Similar difficulties do not occur
in finite-depth water waves, such as those studied in Dold & Peregrine, since the
corresponding unknown would be an arbitrary constant in the velocity potential and
thus have no effect upon the fluid field; the result is that the governing equations are
solved in terms of the derivatives of the velocity potential.
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 549

The calculation of n3 is not as straightforward as it might seem. Theoretically, the


calculation of q)(X, y), for use in the quadrature required for (4.6), can be done by
simply applying the Cauchy integral equation. It is documented, as in Cokelet (1978),
that the Cauchy BIE does not work very well for internal points near the surface
(free surface or wall). The reason for this is that for interior points that are near the
boundary, the kernel in the Cauchy integral becomes large, thus making it difficult for
any quadrature to accurately represent the integral. Seemingly the best approach to
eliminate this problem is that proposed by Cokelet (1978) and involves removing the
singularity at the surface points and then using an iterative scheme to converge onto
the required value. This method has theoretically guaranteed convergence, although
in practice this is not so for internal points extremely close to the boundary due to
the accumulation of round-off errors; at such points a Taylor series expansion from
the surface values is sufficient. The integral is evaluated using the trapezium rule with
Richardson extrapolation. Implementing this iterative scheme dramatically improved
our calculation of n3 in test cases over examples which simply applied the integral
equation directly. Due to its accuracy and speed of computation, the implementation
of this vertical integration is not a handicap of the numerical method.
The expected rank deficiency in G is accounted for by appointing arbitrary values
to X(Xi, h) and Xn(X2, b), where (X1, h) and (X2, b) correspond to discretization
points on the free surface and wall, respectively. The effect of imposing global con-
straints instead of appointing arbitrary values is discussed later. For convenience
we set X = X1 = X2 and our choice of X is defaulted to the x coordinate of the
first surface particle point. Thus, the vertical integration required in calculating n3
is from the points (X2, b) and (X1, h). We are free to choose at each time step the
point distribution at the wall and, hence, we can always ensure that there is a wall
discretization point at (X2, b). If our default choice of X is such that h(X) has more
than one value, as in an overturning wave, then X is altered such that this is not
the case. Such alterations involve negligible extra computing overheads.

(a) Reducing computation time


The computing time required to solve (4.4) can be reduced by taking advantage
of the sparsity of G and the fact that the matrices A({, ~') + iB(~, <') in (4.3) are
fixed in time if both ~, I' represent points on the plane wall. The details of how this
is done is given in appendix A. The result is a reduced LAS, given by
Grsr = b, (4.7)
where Gr C ReMXM, M - 4N+2. The rank deficiency of Gr is the same as that of G
and hence it is accounted for in the same way as described earlier. The reduced LAS
is solved using the NAG routine F04ATF, which uses Grout's factorization method.
This factorization for Gr is expected to take less time than for G by a multiplicative
factor of (2)3 for reasonably large N. However, there are additional overheads for
the setting up of the matrix Gr, which include matrix multiplications of matrices of
size (2N)2. The computation of (b, xb), having solved for Sr in (A 4), involves only
four matrix-vector computations where the matrices have (2N)2 elements and can
thus be considered negligible when compared with the factorization of Gr. Based
upon the computation times of using both G and Gr, the latter is chosen as being
a superior method. However, all the basic diagnostic studies of this section were
completed before the improvement with Gr was derived. The two methods of solving
the LAS give identical results.
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
550 D. Rh. Gwynllyw and D. H. Peregrine

(b) Diagnostics of the scheme


We have assumed that the theoretical non-uniqueness of the Almansi functions
(no, n1) is reflected in the numerical scheme. There is no guarantee that this is the
case. The reason for this can be seen by taking the known eigenvector

e-=(1, O,-Y,-Yn O,0-1),


which corresponds to the free variable n1. The residual vector
r= Ge

has a contribution from (2.8), ignoring round-off errors, given by

r(i) = (kD + kDss)y + DDx, i -1,...,Nf, (4.8)


where D , DS are the matrices representing the first and second-order surface deriva-
tives of the 11-point scheme. Since Dsss DsDs, with the Dsss being an 11-point
scheme and DssDs effectively a 21-point scheme, there results a non-zero residual
which is particularly noticeable in regions of high curvature. The residual is a result
of the numerical approximation and is not due to round-off error. As a guide to how
close the numerical scheme reflects the non-uniqueness in the Almansi functions, we
define two residual values Pbc and Pch which are a norm of the contributions to the
residual vector r from the free-surface boundary conditions and the Cauchy integrals,
respectively, and give us a diagnostic test of the accuracy of the discretization. Fur-
ther, we compare our direct solver with that of using singular value decomposition
(SVD) for the LAS which removes the degeneracy from the LAS by effectively removing
the linear combination of rows that cause the degeneracy. Thus, in this case, a global
constraint is applied to remove the causes of degeneracy.

(c) Numerical time-stepping


Having computed the Almansi functions and its derivatives for a given free surface
and thus calculated the surface velocities, the motion of the free surface is calculated
using the kinematic boundary condition that surface particle points remain on the
surface, namely,

dx _ &0 dy _49
dt Oy ' dt Ox (')
Both the Euler scheme and an adaptive fourth-order Runge-Kutta scheme are used
to calculate this motion. The adaptive Runge-Kutta scheme comprises two Runge-
Kutta schemes, one for half the time-step of the other. Thus, in comparing the two
results, an estimate of the time-step error is made leading to an estimate in the size of
the next time step which would lead to a calculation of the free surface's position to
within a given fixed relative tolerance, e. This scheme is described in detail in Press et
al. (1986). This tolerance was also used in determining the accuracy of calculations
of the aforementioned vertical integration. As pointed out in Baker et al. (1982),
there is no necessity to follow the particle points in the numerical time-stepping.
However, with the behaviour of the particle points in our results, we see no reason
to investigate that extra freedom.

(d) Computation times


The relative running times on a single scalar processor are listed below for one
Euler time step out of seven in each RK time step. In all the examples presented
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 551

in this paper, the adaptive fourth-order Runge-Kutta scheme was more efficient
than the Euler scheme due to the fact that the higher order scheme could perform
with larger time-steps and that no a priori knowledge of the optimal time-step was
required. The adaptive time steps were particularly advantageous for the case of a
steepening wave when progressively smaller time steps were required.
We also compare here the running time of setting up and solving the linear al-
gebraic schemes Gs = b and Gs,r = br. We label these two schemes as scheme
(A) and scheme (B), respectively. The LAS solver used here is the Crout's factoriza-
tion since SVD, as reported later, is far more expensive and does not improve the
accuracy of the model. Both Crout's factorization and SVD take O(L3) operations
for an Re LL matrix, but the inclusion of the Householder QU factorization in the
SVD scheme increases the constant of proportionality in the expected time taken.
The time taken to calculate n3 using the vertical integration and iterative scheme is
negligible compared to the time taken to set up and solve the LAS.
The inverting of the LAS dominated the computation time and, hence, in both
schemes A and B, the total run times were approximately proportional to L3. We
denote the (approximate) constant of proportionality by k giving us a computation
time in seconds, per time step, of kL3. For scheme A, k ~ 3.3 x 10-5 while for
scheme B, including computations that need only be done at the initial time stage,
k 1.6 x 10-5. Excluding the 'once-only' computations, the value of k, for scheme
B, reduces to less than 1.0 x 10-5. It is clear that scheme B is significantly faster
than scheme A. Clearly, k is machine dependent and the above values were quoted
purely for comparison of the two schemes.

(e) General comments and observations


Owing to the fact that two harmonic functions are used, only one BIE is set up
and applied to both these functions. Although no direct comparisons were made,
we believe this to have reduced the setting up time as compared with the usual
method of coupling a biharmonic BIE for the stream function with a harmonic BIE
for the vorticity. The difference in overall time, for the same number of discretization
points, will be most significant for low to moderate values of N since at large values
the direct solver dominates the computing time. We believe the accuracy of the 11-
point scheme, which represents use of a tenth-order Taylor polynomial to describe
the surface, is significantly better than the segment discretization schemes used by
most authors. Justification for this belief is, however, based mainly upon the results
for inviscid water waves.
All of the above computations were carried out using both the direct solver and
SVD, the schemes agreeing very well with each other. This was despite the fact that,
for the case of overturning waves, the residual values, Pbc, Pch defined in ? 4 b, were
both of order unity. Furthermore, varying the number of discretization points, and
hence varying the residual values, brought about insignificant changes to our results.
The SVD scheme has the considerable disadvantage of being five times as expensive
to run as the direct solver. The value of Pbc could have been reduced dramatically by
setting up )nss) for example, in the LAS as Dsssg, where p is the complex conjugate of
). This would have required the inverse of Ds and preliminary results show that this
is not an improvement on our original scheme despite the fact that the SVD condition
numbers are significantly lower. The adaptive time stepping also suffered when used
with the SVD solver. Not only were the time steps smaller, but their magnitude could
vary significantly for profiles that showed no such problems in the direct solver case.
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
552 D. Rh. Gwynllyw and D. H. Peregrine

The reason for these variations is not known, although we ascertained that it is not
due to the point of vertical integration being underneath possible problem areas,
such as the trough and peak. We did not pursue this matter further since the SVD
scheme is far too expensive to be used as a non-diagnostic solver.
For evaluation of quantities within the liquid there is a correlation between large
values of Pch and the ability of the Cauchy integral to converge for points near the
surface. However, the effectiveness of the Taylor scheme at such points eliminates
any problems with the vertical integrator.
In our numerical method we found that as long as both Nf and Nb were not too
small, Nf is always the dominating factor in improving the accuracy of the scheme.
The value of Nb could readily be varied at each time step although its only influence
was to vary Pch. Further, this change does not affect the Cauchy iteration scheme
since any convergence problems occur almost always near the free surface and not
the wall.
In every run that was terminated due to the reduction in adaptive time steps,
the residual values would increase dramatically prior to such an end. An increase
in the SVD condition number would be significantly less. Such an increase can be
the result of a decrease in numerical smoothness of the free surface as well as the
development of high curvature and other such phenomena. This is also indicated by
the fact that in certain cases the residual values would increase in transient flow
as the profile tended towards the unperturbed state. No smoothing is performed in
these computations, but might well extend some of the computations.
Due to its effectiveness and accuracy, the following results are calculated using the
21-11-point scheme with a direct matrix solver.
With a dimensionless fluid depth of 1, Ym denotes the y-coordinate of the un-
perturbed free surface and thus the plane wall lies at y = m- 1. Setting Ym to
zero reduces the instabilities which may be aggravated by rounding errors. At large
times, the two time stepping schemes for flow down the inclined plane disagreed and
the results shown are from using a Runge-Kutta scheme with e set to a level which
resulted in no visual difference with lower tolerances.

5. Validation of the numerical method

Before dealing with the flow down the vertical wall, we compare our method with
the analytical results of Hopper (1990), which are for Stokes flow within simply
connected domains, symmetric about the x- and y-axes, driven purely by surface
tension. We thus need only consider a quadrant of the domain and we make use of
the arbitrariness of the Almansi functions to force symmetric and anti-symmetric
properties on these functions. Since the boundary is simply connected (unlike the
film flow case), solving the LAS is sufficient to solve for the unknown functions. In
terms of the governing equations, the only change required for this problem is the
removal of the body force term in (2.8). Again, for this particular geometry, the
Almansi functions are single valued.
Hopper finds a time-dependent conformal mapping,

z = F((,t), (1(1 < Aj = -1), (5.1)


that maps the free surface of the fluid onto the unit circle. In order to find r,
Hopper uses the Goursat representation for a biharmonic function and the boundary
conditions written in terms of the Goursat functions as in Langlois (1964). These
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 553

boundary conditions and the surface velocity are then expressed in terms of the
Goursat functions, r and their derivatives with respect to time and A. Analytic and
single-valued conditions on the Goursat functions limit the existence of solutions
such as (5.1), although Hopper lists several classes of mappings which are acceptable
and satisfy the boundary conditions. One such solution is that of the Nephroid given
by Hopper as
((, v) = (1+ i32)-1/2(( - IV3), (5.2)
where v is a time parameter with time given by

t(v)= 17 / (1 - k)(1 + k2 )2[kK(k)]- dk, (5.3)


where
rT/2
K(k) = ( - k2 sin2 q)-1/2 do,
Jo

is the complete elliptic integral of the first kind.


We compared Hopper's results for the Nephroid with an initial value of v taken to
be 0.9. This initial shape contains large variations in curvature. Two forms of com-
parisons were made. For each profile the velocities were compared at the symmetric
points and the profiles compared graphically. To compare the coordinate values is
extremely tedious, since fixed points on the unit circle do not map onto the same
surface points for different times using Hopper's conformal map.
The profile comparison is illustrated in figure 2, where the numerical time stepping
used the Runge-Kutta scheme with e = 10-6. There is no visible difference between
the numerical results and Hopper's analytical results. This is supported by the fact
that the maximum relative error in the velocities of the symmetry points is 10-9 over
the 20 time steps that the model ran. The Euler scheme also performed well with
this initial profile except for the onset of numerical instabilities as the shape became
very close to the circle.

6. Results for gravity driven flow down a wall


The main physical aim of our study is to examine waves on the gravity-driven flow
of thin viscous films.
For all our examples, the initial free-surface profile is sinusoidal with dimensionless
amplitude am, wavenumber a and free-surface point distribution which is evenly
distributed along x E [0, 2r]. In order to have the strongest effects in wave evolution
the plane wall is vertical, unless otherwise stated.

(a) Comparison with analytical approximations


Most analytical work has used long-wave approximations. To compare with weakly
nonlinear long-wave analysis we would like to consider our model for much longer
waves than those with unit wavenumber (a = 1). Unfortunately, for large spatial
periods, a difficulty arises. As mentioned before, the discretized Cauchy integrals do
not represent the harmonic function very well if any of the discretization points lie
near another point which belongs to a non-adjacent section of the free surface. This
is the case for the points on the wall, as well as for internal points in relation to the
surface points for large spatial periods. The distance within which problems occur
seems to be a function of the distance between neighbouring points, d. Keeping the
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
554 D. Rh. Gwynllyw and D. H. Peregrine

1.0 -

0.75 -

0.5 -

0.25 -

0.0 , i 1 1 1 , i
0.0 0.25 0.5 0.75 1.0

Figure 2. Comparison between our Runge-Kutta results (line) and Hopper's (dots); starting
at v = 0.9 with t = 0.1085,0.2445. These represent the 10th and 20th time stage in the
Runge-Kutta scheme.

number of points per wavelength constant means that d increases for larger periods.
Summarizing, a larger number of discretizing points is required for longer waves if we
are to keep the same accuracy in our numerical scheme. For waves of dimensionless
period 107 or more with dimensionless depth 1 and initial amplitude, am0.1, the
number of discretization points required becomes prohibitively large. This problem,
to a large extent, was alleviated in Dold & Peregrine (1986) by considering the free
surface and its reflection in the plane wall effectively doubling the distance between
near but non-neighbouring points. This could not be done here because there are two
boundary conditions at the wall to be satisfied as opposed to one in the water-wave
case.

For the above reason we deal with comparatively short waves and hence have poor
agreement with the nonlinear long-wave analysis of Benney (1966), Gjevik (1970) and
Nakaya (1975) in the limit of zero Reynolds number. The disagreement is especially
unsatisfactory in the presence of surface tension since higher order terms are required
to accurately represent the surface tension for the wavenumbers considered here.
On the other hand, there is good agreement between our results and the linear
theory given by the Orr-Sommerfeld equation. A study of the linear theory of surface
levelling in viscous fluids has been made by Orchard (1962).
Our experience for flows without surface tension on non-vertical walls with an
initially sinusoidal surface profile is summarized as follows. For a horizontal bed,
the surface maintains a sinusoidal profile as the amplitude decreases and the results
agree very well with the linear theory for initial amplitudes up to at least 0.1. For
an inclined bed, /31, refer to figures 3 and 4. With an initial amplitude of 0.01
the numerical profile keeps its sinusoidal profile leading to good agreement with the
linear theory. For the relatively large initial amplitude of 0.3, there is, unsurprisingly,
rapid departure from the linear solution. The numerical profile steepens with time
while the linear profile keeps its sinusoidal profile. As the wave steepens there is
an increase in the concentration of free-surface particles at the front of the wave
and a corresponding decrease at the rear of the wave crest. These calculations were
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 555

0.01

y 0.0

0 1 2 3 4 5 6 7 8 9
x

Figure 3. Comparison between numerical and linear wave profiles for 3 = 1 rad, am = 0.01 and
a = 1 at times t = 0.0, 2.3, 4.2, 6.1, 8.1, 10.0. The numerical results are shown by the dots which
represent actual free-surface particles.

made with N = 40, the usual choice, with no visible difference between the two time
stepping schemes that we consider.

(b) Small amplitude waves down a vertical wall


Linear theory predicts that all Stokes flow disturbances down a vertical wall are
damped except for flow down the vertical wall free of surface tension. Small amplitude
waves changed very slightly in our numerical results. The model was run with the
Runge-Kutta scheme up to time t = 44.27 with a small initial amplitude taken to
be am = 0.01 with unit wavenumber and no surface tension, a = 0. Only a slight
decrease in the amplitude can be detected with an amplitude of 0.009 99 at t = 44.27.
The wave is almost neutrally stable and we calculate its phase speed to be 0.647,
which agrees with linear theory to three decimal places.

(c) Time periodic flow


We now consider the neutrally stable flow with zero surface tension, a = 0, for a
larger initial dimensionless amplitude of 0.1 and unit wavenumber. At large times
the Euler and Runge-Kutta schemes disagreed. The following results are from using
the Runge-Kutta scheme with e set to a level which resulted in no visual difference
with lower tolerances.
Figure 5 gives the 'waterfall' picture of the flow, where the profiles at each time
step are shifted back by the distance vt where v = 0.652. The comparison of the
profile at time t = 41.0 with the profiles at times t = 54.7, 68.2, 95.4,121 is shown,
magnified, in figure 6. The profile at t = 54.7, shown in figure 6i, has a minimum
height and is included as a contrast to the other profiles which are at approximately
equal phases. This example suggests the possibility of a time periodic propagating
wave form, with period approximately 26.

(d) Large amplitude waves, short waves and steady waves


For large-amplitude waves with no surface tension, the initial profile soon steep-
ened with the face of the wave overhanging the trough, as illustrated in figure 7
(am = 0.3,c = 1). This clearly differs from the overturning experienced with in-
viscid water waves. Beyond the last profile in the figure, the computations become
inaccurate with the time steps becoming too small to be of any use.
Short waves also have a propensity to steepen, as is illustrated in figure 8 (am = 0.1,
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
556 D. Rh. Gwynllyw and D. H. Peregrine
0.3-

-0y 3
0 1 2 3 4 5 6 7 8 9

0.3-

y 0-

- 0 .3 l l l W l l l l T , I i I I I I I j I I i i i, I i i i
0 1 2 3 4 5 6 7 8 9

0.3-

0 1 2 3 4 5 6 7 8 9

0.3-

0 1 2 3 4 5 6 7 8 9

1 at times t 0.0 1.3, 2.9, 4.5 6.1 7.8,6 7 8 9.7. The numerical results are shown by the dots
which represent actual free-surface particles.

c = 3), although the shape of the resulting 'bulb' is significantly different from that
seen in the large-amplitude example with the rear of the wave crest being almost
straight. In the large-amplitude wave the maximum steepness of the rear, throughout
the transient, has kept reasonably constant, leading us to believe that, unlike the
short wave, the rear will not straighten.
The larger amplitude wave steepens more quickly than the shorter wave. Again
this is seen in the comparison between figures 7 (am = 0.3, ca = 1) and 8 (am = 0.1,
a = 3). These waves have the same initial steepness. Figure 7 displays an overhanging
wave at time t = 7.81, while this occurs in the short wave of figure 8 at the later
time of t = 15.9. We suspect that proximity to the wall plays a significant part in
the steepening of the wave. As before, the program is stopped when the RK scheme
returns too small a time-step.
As an additional example of a short wave steepening, we have the wave illustrated
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 557

-0.1 x

Figure 5. The Runge-Kutta scheme from t = 0.0 to t = 150.0 at intervals of dt = 3.0 with
v = 0. 654 for the flow with a = 1r, am = 0.1 and a = 1. The axis is drawn for the t = 0.0 case.

in figure 9 of initial sinusoidal profile am = 0.1 and wavenumber a = 5. In this


figure the discretization point distribution is shown on the wave. For the flow down
the inclined wall, this motion of the surface points is quite typical with the points
concentrating on areas of high curvature variation such as the peak and, especially,
the trough. This is in contrast to the surface tension driven flow mentioned earlier,
where an initially high concentration at the 'kink' of the Nephroid would remain
concentrated up to its steady state as a circle. For this inclined wall flow, the wave
becomes overhanging at time t = 12.35.
The changes in wave profile of these short waves decrease as the wave steepens, as
illustrated in figures 8 and 9. In both cases, at the program's termination, the face
of the wave is overturning in the sense that the wave is becoming pendant and a
portion of the wave has overtaken the trough. In an attempt to further decrease the
rate of steepening of the wave, surface tension was applied to the wave illustrated
in figure 8 (a = 3) at the stage when the wave front was vertical (t = 3.73). The
result was that for a surface tension y corresponding to a = 0.05, the overturning
was soon arrested; for a larger surface tension the amplitude decreases with time and
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
558 D. Rh. Gwynllyw and D. H. Peregrine

0.1 -(
(b)
? (c)
Y 0.0-

-.] " " I 'I 3I I"", II Ig I,


0 1 2 3 4 5 6 7 8 9

0.1 - . (a)
( d)

Y 0.0 -

- .1 I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I
0 1 2 3 4 5 6 7 8 9

0.1-\ /- . . . (a)
(e)

Y 0.0 -

-0.1 II.IIIIIII. IIIIIIIIII IIIIIIIIIIIIIII.III.I I I i


0 1 2 3 4 5 6 7 8 9
x

Figure 6. Comparison of profiles at (a) t = 41.0; (b) t = 68.2; (c) t = 54.7; (d) t = 95.4; (e)
t = 121; assuming a phase velocity v = 0.652 for the flow with 3 = 7r, am = 0.1 and a = 1.

for smaller surface tension the wave continues to overturn with the trough becoming
progressively sharper. Figure 10 is a magnified figure of the aforementioned surface
tension flow with a = 0.05, where the two profiles are taken at a time interval of 1
assuming a phase velocity of 0.508.
The profiles show the close resemblance to a steady wave. The contours of the
stream function travelling with the wave at the phase velocity are shown in fig-
ure 11. The stream function is calculated at the nodes of a 80 x 40 grid by vertical
integration identical to that mentioned earlier. The contours were drawn using the
UNIRAS package GCNW2V which uses bilinear interpolation between the grid nodes.
The machine time used to calculate the internal stream function was comparable
with the time to calculate the surface velocities. We would expect, for a steady wave,
the surface profile to be a streamline. This is not the case here since we only have an
approximation to a steady wave. The contours are, however, drawn for very small in-
crements in the stream function (do = 2.0 x 10-4). Despite this, the streamlines meet
the surface approximately tangentially along most of the surface. Figure 11 clearly
shows that there exists a stagnation point within the bulbous shape. The fluid par-
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 559
0.3

Y 0

-0.3 , ....
0 1 2 3 4 5 6 7 8 9

0.3 -

Y 0-

-0 .3 i i i i i i i i i i i i i i i i i I i i i i i i i i i i i i i i i i i i i i i i i i
0 1 2 3 4 5 6 7 8 9

0.3 -

Y 0-

-0.3
-0.3 I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I
0 1 2 3 4 5 6 7 8 9

0.3 -

Y 0-

-0.3 ,,,, I ,,,, I I I I I j I I I I I I I I I I I I I I I I


0 1 2 3 4 5 6 7 8 9

0.3 -

-0.3 1 I 1 I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I.
0 1 2 3 4 5 6 7 8 9
x

Figure 7. The Runge-Kutta scheme for the 'large' amplitude flow (am = 0.3) down the vertical
wall with a = 1 at t = 0.0, 2.01, 3.51, 4.98, 6.33, 7.08, 7.46, 7.86, 8.21, 8.47 using N = 40.

tides rotate clockwise around this point. This is a feature that can be important in
assessing heat or mass transfer.
These 'drop-like' examples are unlikely to be physically realistic because of their
short wavelength. However, it is likely that such drop-like behaviour can develop for
longer, more realistic, waves which would require more substantial computations.
This behaviour is very similar to the 'rolling' type of motion that occurs as a wetting
front descends down a dry non-wetting surface.

(e) Long waves


Finally we consider the motion of a longer wave of period 107r. By showing one
example of the development of a long wave we do not claim it to be typical of long
wave flow for the particular amplitude chosen. The intention is simply to illustrate
how the wave number can affect the development of waves of the same amplitude.
Figure 12 shows the development of a wave of wave number 0.2 with initial sinusoidal
profile with amplitude 0.1. As explained before, the Cauchy integral representation
is less accurate for long waves. To counteract this we use N = 80, which clearly
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
560 D. Rh. Gwynllyw and D. H. Peregrine

y 0-

-0.1 . . . . . . . . . . . . . . . . . . . . . . . . . . .

II 0.5
0 ).0 0.5 1.0 1.5 2.0 2.5 3.0

0.1

Y 0

-0.I1 I I I I I 1 1 1
0.0 0.5 1.0 1.5 2.0 2.5 3.0

=I
y

-0.1 , J---l

0.0 0.5 5 1.0 1.5 2.0 2.5 3.0

0.1 -

y 0-

-0.1 I I ~ l II I I I I I I I I I I I I I I i
).O 0.5 1.0 1.5 2.0 2.5 3.0

0.1-

y 0-
t,. I
-U. I I I I i I I - I I I
0.0 0.5 1.0 1.5 . ' ' '02.5
2.0 ' I '3.0
' -2
x

Figure 8. Development of a short wave (a = 3) down a vertical wall with am = 0.1 at


times t = 0.0, 0.42, 2.55, 3.55, 4.35, 4.98, 5.16, 5.31, 5.41, 5.49, 5.54, 5.59, 5.63, 5.67, 5.70, 5.74 using
a Runge-Kutta scheme with N= 80.

makes the computation more expensive. For this reason we do not follow the long
wave development any further although we believe our model to be accurate enough
to cope.
Unlike the flow of spatial period 27r, there is no tendency towards a time-periodic
solution. Each wave seems to be developing a solitary nature.
At t = 40 the peak of the wave has travelled 40.15 compared to the linear long-wave
theory value of 40 (phase speed is twice the unperturbed surface velocity).

7. Conclusion

An accurate numerical method of modelling unsteady free-surface Stokes flow using


a BIE method is described. Its major feature is the use of a Cauchy integral for two
harmonic functions, and the point discretization to enable high-order approximations
involving many points rather than discrete elements of more limited geometry. The
method is found to be stable past the point of overturning for steep waves. The effect
of surface tension on such waves is found to generate almost steady waves in some
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 561

O.1 .. ......................
-0.1.. ?.....
0.0 0.25 0.5
- 0.75 1.0 1.25
-~~~- 1.5 1.75
-..I

0.1 I - ""-".........

y C '???.~ltf? ????
.'~. ....????....
... . ...
..'.,,???ts????
n .....
-0.1
0.0 0.25 0.5 0.75 1.0 1.25 1.5 1.75

0.1-
:: .. ...... .?. . *... ..-"'::::'" ... .......
y 0-

-0.1
.:: - . .,?? ?? ~~~~~~~~~~~~~~~~~~~~~~I t
C .0
>.o 0.25 0.5 0.75 1.0 1.5
0.25 0.5 0.75 1.0 1.25 1.25 1.751.5 . .1.75
.. . .

0.1-
: ..

y 0- ??? - t i . . - 1. :." ,???1?? t~~??

0.0 0.25 0.5 0.75 1.0 1.25 1.5 1.75

0.1 .
y 0

?-0.1/ l.iii,,,,i,, , '',,,,,,,, a i.... ,,


0.0 0.25 0.5 0.75 1.0 1.25 1.5 1.75
x

Figure 9. Development of a short wave (a = 5) down a vertical wall with am = 0.1


at times t = 2.10,5.77,8.69,11.94,12.49, 12.76,12.83, 13.02,13.10,13.30,13.37,13.47 using a
Runge-Kutta scheme. Each profile is given by the particle points as in the discretization of
the profile (N = 80). This figure is with natural scaling and the dots represent actual fluid
surface particles.

0.1
? .. (a)
(b)

y 0.0 -

-0.1 III
0.0 k .5 6fs
1.0 ~ 2.0
1.5 ~ o ; ;
2 3.0
x

Figure 10. The comparison of two profiles taken from a continuation of a steepening profile in
figure 8 but now under the effect of a surface tension, 3y, giving a = 0.05. These profiles are
at a time interval of 1 from each other and assuming a phase velocity of 0.508. The unbroken
line, (b), represents the later profile while the dots in the former profile, (a), represent actual
free-surface discretization points.

examples. Further to this, a travelling wave with a time-periodic profile of moderate


amplitude and spatial period is found.
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
562 D. Rh. Gwynllyw and D. H. Peregrine
0.0-

--V. ,I I I I I I I I I
0.0 0.5 1.0 1.5 2.0 2.5 3.0

Figure 11. The streamlines of the flow in a frame of reference with velocity of 0.508. The flow
profile is the later profile in figure 10. At the streamline nearest the base =- 0.174, at stagnation
point ) = -3.9 x 10-4, and contour step db = 2.0 x 10-4. The displayed streamlines correspond
to the lowest 5% values of the streamfunction in the flow. The arrows represent the direction of
flow. This figure is with natural scaling.

0.1

y 0.0
-0.1

0.1

y 0.0
-0.1

0.1.

y 0.0

-0.1
0 5 10 15 20 25 30 35 40 45

0.1

y 0.0

-0.1

0.1 I

y 0.0

-0.1

Figure 12. The development of the long wave of period 107r, down the vertical wall, with
am = 0.1 using Runge-Kutta up to t = 40.02 at equal time intervals of 1.38 (N = 80).

The scheme is accurate despite the large curvature variations found in some of the
Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 563

profiles. After the point of overturning, errors in the calculation of iK, which involves
third-order spatial derivatives, might cause the adaptive Runge-Kutta scheme to
terminate the program. It would be of interest to find whether the trough develops
into an air bubble or if points at the trough can be absorbed into the interior by
cusp development in the absence of surface tension. An analytical study of whether
surface points are absorbed into the interior of a viscous fluid through a free-surface
cusp has been made by Jeong & Moffatt (1992), although in their case only steady
flow is considered. In the case described here, there is an asymmetrical 'rolling over'
of the free surface which seems even more likely to lead to surface absorption.
This numerical method is designed to follow the unsteady motion of Stokes flow.
Precise steady wave forms and time-periodic wave forms might be found explicitly
by using iterative methods for the free surface, as in Pozrikidis (1988), using our
near solutions as initial guesses.
A representation of the biharmonic function, using only one harmonic function,
that satisfies the no-slip boundary conditions is possible by an approach similar to
that described by Zeng & Weinbaum (1994). This method, however, is only applicable
to flows over plane walls. Although we only present results for flows over plane
walls, our numerical method, on the other hand, can be applied to non-plane wall
geometries.

The authors are grateful for the assistance of the referees. This work was supported by SERC
and private funding.

Appendix A. Reducing the linear algebraic scheme

The sparsity of G is due to the fact that the plane wall variables, (bn and Xb, are not
included in the free boundary condition and, further, the relation (4.3) is applied with
_ = X and -- X separately. However, we note that, due to the boundary conditions
at the wall (3.8), both the plane wall variables appear in the relations due to the
Cauchy integral. The form of the relationships are

( M M ) ((Q
M3M4) b R
(~S g)(~)~(~
Of ) ( A 1
) (A1)
^ 3 M.){ -{b0 0 OfR S) f
(3 M4 X(n n ( ? ? ) ( (A 2)
(^ )(^ -(^ .V ^ + (A2)
P3
P3 P P4
a b 0OnTn
0 Of +T Xn
U f
Relations (A 1) and (A 2) are derived from (4.3) by taking ~ to be a wall point and
free-surface point, respectively. The matrices Mi E ReNbxNb for i = 1,... .,4 are fixed
in time since they correspond to both I and ~' representing discretization points on
the plane wall in (4.3). None of the other matrices in (A 1) and (A2) are invariant
in time since they are derived from either or both of J, *' being free-surface points.
Matrix M, given by

\ M1 M2)
M3 M4

Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
564 D. Rh. Gwynllyw and D. H. Peregrine

is singular; however, knowledge of its eigenvector corresponding to its zero eigenvalue


enables us to generate a matrix, M* C Re2Nbx2Nb such that

IX n +M*
- ) ] = M -( )' +
( R nM( ?X? n
) j

+-Fb(1) 0 )+ (o)l
Xn( 7\7:
b) \(A3)
(A3)

where M* is the inverse of a modified non-singular form of M and is independent of


the free-surface profile; Xb(Nb), for example, is the value of Xb at the Nbth point on
the wall and the vector 1 is the Nb-vector of ones.
Inserting (A 3) into (A 2) and combining the result with the free-surface boundary
conditions gives us the reduced LAS

Grsr br, (A 4)
where Gr C ReMX, M _ 4N + 2 and
S - (f, fXf Xf x b(I), X(N))T, (A 5)
and br has the same non-zero elements as b but is of reduced size after having L - M
zero elements removed.

References

Baker, G. R., Meiron, D. I. & Orszag, S. A. 1982 Generalized vortex methods for free-surface
flow problems. J. Fluid Mech. 123, 477-501.
Benney, D. J. 1966 Long waves on liquid films. J. Math. Phys. 45, 150-155.
Cokelet, E. D. 1978 Mechanics of wave induced forces on cylinders (ed. T. L. Shaw), pp. 287-301.
London: Pitman.

Cooker, M. J., Peregrine, D. H., Vidal, C. & Dold, J. W. 1990 The interaction between a solitary
wave and a submerged circular cylinder. J. Fluid Mech. 215, 1-22.
Cooker, M. J. & Peregrine, D. H. 1990 Computations of violent motion due to waves breaking
against a wall. In Proc. 22nd Int. Conf. Coastal Engng. Delft. A.S.C.E., vol. 1, pp. 164-176.
Dold, J. W. 1992 An efficient surface-integral algorithm applied to unsteady gravity waves. J.
Comp. Phys. 103, 90-115.
Dold, J. W. & Peregrine, D. H. 1986 An efficient boundary-integral method for steep unsteady
water waves. In Numerical methods for fluid dynamics (ed. K. W. Morton & M. J. Baines),
vol. II, pp. 671-679. Oxford: Clarendon.
Gjevik, B. 1970 Occurrence of finite-amplitude surface waves on falling liquid films. Physics
Fluids 13, 1918-1925.
Hopper, R. W. 1990 Plane Stokes flow driven by capillarity on a free surface J. Fluid Mech.
213, 349-375.
Ingham, D. B. & Kelmanson, M. A. 1984 Boundary integral equation analyses of singular,
potential and biharmonic problems, Springer Lecture Notes in Engineering 7 (ed. C. A. Brebbia
& S. A. Orszag). Berlin: Springer.
Jaswon, M. A. & Symm, G. T. 1977 Integral equation methods in potential theory and electro-
statics. New York: Academic.

Jeong, J.-T. & Moffatt, H. K. 1992 Free-surface cusps associated with flow at low Reynolds
number. J. Fluid Mech. 241, 1-22.
Kuiken, H. K. 1990 Viscous sintering: the surface-tension-driven flow of a liquid form under the
influence of curvature gradients at its surface. J. Fluid Mech. 214, 503-515.

Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms
Numerical simulation of Stokes flow down an inclined plane 565

Langlois, W. E. 1964 Slow viscous flow. New York: Macmillan.


Longuet-Higgins, M. S. & Cokelet, E. D. 1976 The deformation of steep surface waves on water.
I. A numerical method of computation. Proc. R. Soc. Lond. A 350, 1-26.
Nakaya, C. 1975 Long Waves on thin fluid layer flowing down an inclined plane. Physics Fluids
18, 1407-1412.
Ohring, S. & Lugt, H. J. 1991 Interaction of a viscous vortex pair with a free surface. J. Fluid
Mech. 227, 47-70.
Orchard, S. E. 1962 On surface levelling in viscous liquids and gels. Appl. Sci. Res. A 11, 451-
464.

Pozrikidis, C. 1988 Flow of a liquid film along a periodic wall. J. Fluid Mech. 188, 275-300.
Press, W. H., Flannery, B. P., Teukolsky, S. A. & Vetterling, W. T. 1986 Numerical recipes.
Cambridge University Press.
Stone, H. A. & Leal, L. G. 1989 Relaxation and breakup of an initially extended drop in an
otherwise quiescent fluid. J. Fluid Mech. 198, 399-427.
Tanaka, M., Dold, J. W., Lewy, M. & Peregrine, D. H. 1987 Instability and breaking of a solitary
wave. J. Fluid Mech. 185, 235-248.
Zeng, Y. & Weinbaum, S. 1994 Stokes flow through periodic orifices in a channel. J. Fluid Mech.
263, 207-226.

Received 26 October 1993; revised 26 January 1995; accepted 13 February 1995

Proc. R. Soc. Lond. A (1996)

This content downloaded from 193.194.76.5 on Tue, 28 Jun 2016 08:17:56 UTC
All use subject to http://about.jstor.org/terms

Vous aimerez peut-être aussi