You are on page 1of 6

Convergence of the nite volume particle method

for viscous ow
Libor Lobovsk y, Ruairi M. Nestor, Nathan J. Quinlan
Department of Mechanical and Biomedical Engineering
National University of Ireland, Galway
Galway, Ireland
libor.lobovsky@nuigalway.ie
AbstractThe nite volume particle method (FVPM) is a
mesh-free particle method that inherits some of the properties
of the smoothed particle hydrodynamics (SPH) as well as of the
nite volume method (FVM). In this work, a brief description
of the FVPM model for viscous slightly compressible uid is
provided. The presented study addresses convergence and effec-
tivity of the FVPM simulations for a two-dimensional unbounded
viscous ow. The methods sensitivity to some setup parameters
is discussed. The main attention is paid to the role of the particle
kernel function and to the evaluation of interparticle vectors with
respect to further optimisation of the method.
I. INTRODUCTION
The nite volume particle method (FVPM) is a recently
developed computational tool capable of solving complex
problems of uid dynamics, and continuum mechanics in
general, without the need for a computational mesh. The
fundamentals of the FVPM method were derived by Hietel et
al. [3] and extended by several authors during the past decade,
e.g. [5], [6], [10]. It belongs to the family of particle methods,
such as smoothed particle hydrodynamics (SPH), while its
formulation is derived from the integral conservative form
of the governing equations. In FVPM, computational domain
is discretised with a nite number of computational particles
which are represented by smooth overlapping test functions.
The particle interaction is realised through inter-particle uxes
between neighbouring particles. In a special case where neigh-
bouring particles do not overlap but continuously ll in the
computational domain, the method is formally identical to
conventional mesh-based nite volume method.
Unlike the standard SPH method, which is purely La-
grangian, in FVPM the particle velocity does not necessarily
equal the uid velocity, i.e. the FVPM particles can be de-
ned either in Lagrangian, Eulerian or Arbitrary Lagrangian-
Eulerian coordinate frames. The FVPM is conservative and
enables straightforward exact treatment of boundary conditions
via denition of appropriate boundary uxes. Major drawbacks
of the method in comparison to weakly compressible SPH are
signicantly higher computational demands due to numerical
integration of interparticle uxes and difcult implementation
of free surface boundary conditions. An attempt to overcome
both of these issues by applying special smoothing functions
was suggested recently in [9].
The present work is focused on the characterisation of the
error behaviour in FVPM simulations, methods convergence
and overall sensitivity to the simulation setup. Although the
convergence of the method has been investigated in some prior
studies, [7], [11], the current results present a more compre-
hensive evaluation of methods performance for viscous ows.
In the following, the FVPM formulation for weakly com-
pressible viscous uid is compared with the theoretical solu-
tion of Taylor-Green vortex ow in two-dimensional space. A
number of simulation series has been carried out and the error
in numerical solution has been evaluated as a function of both
the initial particle spacing and the particle smoothing length.
II. METHOD
A ow of viscous uid can be described by Navier-Stokes
equations in a conservative form
U
t
+(F G) = 0, (1)
where U is the vector of conserved variables, F is the inviscid
ux, G is the viscous ux and t is time. For a weakly
compressible uid, the system of Navier-Stokes equations
is accompanied by an equation of state, in which the uid
pressure p is an explicit function of density ,
p =

0
c
2
0

__

0
_

1
_
, (2)
where
0
is the reference density, c
0
is the reference value of
speed of sound and equals 7, [1].
The conservative Navier-Stokes equations and the equation
of state are discretised for a nite set of particles which ll in
the computational domain. Each particle is associated with a
compactly supported overlapping test function

i
(x, t) =
W(x x
i
(t), h)

N
j=1
W(x x
j
(t), h)
, (3)
where x is the position vector, N is the number of particles
and W
i
= W(x x
i
(t), h) is a kernel function for particle
at position x
i
and smoothing length h analogous to kernel
functions used in SPH. The radius of kernel support domain
is dened by a multiple of the smoothing length h (e.g.
equals 2 for the quadratic kernel). Similarly to SPH method,
5
th
international SPHERIC workshop Manchester, UK, June, 23-25 2010
Fig. 1. Overlap of FVPM particles with support domain radius of 2h.
the value of smoothing length h is used to dene the distance
at which particles interact with each other. But in FVPM,
the interacting particles are determined by an overlap of
their support domains, i.e. the maximum distance between
interacting particles is 2h, Fig. 1. A schematic illustration of
a distribution of neighbouring particles for several values of
smoothing length is given in Fig. 2. The minimum theoretically
applicable value of h/x is

2/4, where x is initial particle


spacing.
The sum of the test functions
i
over the entire set of
particles equals unity, i.e.
N

i=1

i
(x, t) = 1. (4)
and every particle is associated with a volume V
i
so that
V
i
=
_

i
dx. (5)
Thus a discrete value
i
of any eld variable at particle i
can be given as

i
=
1
V
i
_

i
dx. (6)
The general form of the FVPM governing equations is then
d
dt
(V
i
U
i
) =
N

j=1

ij
(F(U
i
, U
j
) G
ij
) , (7)
where F(U
i
, U
j
) denotes a numerical approximation to the
inviscid ux F
ij
U
ij
x
ij
and G
ij
is viscous ux between
pair of particles i and j. Following the previous work [7],
the viscous ux is evaluated on the basis of the consistency-
corrected SPH gradient approximation of Bonet and Lok [2].
The interaction vector
ij
between particles is given as

ij
=
_

i
W
j

j
W
i

N
k=1
W
k
dx (8)
which can be evaluated using numerical integration over the
overlap region between the particle pair. The vector
ij
is
analagous to the cell interface area vector which weights
intercell uxes in the classical nite volume method.
Fig. 2. Spatial distribution of neighbouring particles (light red) of a dark red
particle for h/x equal to 0.354 (left), 0.501 (center) and 0.708 (right).
Since the particle volume term appears in the time derivative
on the left hand side of equation (7), a separate evaluation
of particle volumes rate of change is necessary. It can be
expressed as
d
dt
V
i
=
N

j=1

ij

x
ij
. (9)
where

x
ij
denotes the average value of velocities at particle i
and j.
In order to complete the set of governing equations, the
pressure p
i
is updated at every particle according to the
equation of state (2), i.e.
p
i
=

0
c
2
0

__

0
_

1
_
. (10)
Further details on the applied FVPM model can be found
in [7].
III. NUMERICAL RESULTS
In the following, the FVPM formulation for weakly com-
pressible viscous uid in Eulerian coordinate frame is consid-
ered. Several sets of numerical experiments are presented and
the error behaviour with respect to the particle spacing x,
the particle smoothing length h, the kernel function denition
and the method for evaluation of interaction vectors
ij
is
investigated.
The test case studied is the two-dimensional Taylor-Green
vortex ow, which represents a laminar incompressible viscous
ow in an unbounded domain, Fig. 3. The ow eld consists
of a decaying vortex described by the exact solution of the
Navier-Stokes equations in the following form
u = u
0
cos(2x/L) sin(2y/L)e

8
2

L
2
t
, (11)
v = u
0
sin(2x/L) cos(2y/L)e

8
2

L
2
t
, (12)
p = u
2
0
/4 [cos(4x/L) + cos(4y/L)] e

16
2

L
2
t
,(13)
where u and v are the x and y components of the velocity, p is
the pressure, u
0
is the maximal initial speed and is kinematic
viscosity. From the equations (11)-(13) it is apparent that the
ow solution is periodic in space. The computational domain
of xy 0, L0, L is considered and periodic boundary
conditions applied.
5
th
international SPHERIC workshop Manchester, UK, June, 23-25 2010
x
y

100.4192
0.78201
101.9833
x
y

0.032811
0.23814
0.44346
Fig. 3. Pressure eld [Pa] (top) and distribution of velocity magnitude
[ms
1
] (bottom) for Taylor-Green ow at time t = 1 s as computed by the
FVPM code for h/x equal to 0.8 and 3600 particles.
The error investigated is the non-dimensional L
2
norm error
in the ow acceleration magnitude evaluated at time t = 0 as
L
2
=

_
1
N
N

i=1
_
a
i
a
th
i
u
2
0
/L
_
2
(14)
where N is the number of particles and a
i
= dv
i
/dt is
the magnitude of velocity time derivative at particle i and a
th
i
is its theoretical value. This measure of error is inuenced
by spatial discretisation only, in isolation from error due to
temporal discretisation and particle motion.
A. Convergence of FVPM simulations
Unless mentioned otherwise, the Taylor-Green ow is
solved for Reynolds number equal to 100, while the computa-
tional domain width L is 1 m and the initial peak velocity
u
0
is 1 ms
1
, Fig. 3. As the weakly compressible model
of the uid is considered, the numerical speed of sound c
0
is set to 10 ms
1
, i.e. resulting Mach number is 0.1. The
numerical experiments are carried out for smoothing length
to particle spacing ratio h/x 0.354, 1.501 and uniform
particle spacing x 0.002, 0.05. Numerical integration
of interparticle uxes
ij
is performed using the Gaussian
quadrature and N
int
N
int
integrating points, where N
int
equals 6.
The rst set of numerical tests is performed for the quadratic
kernel function W(x x
i
(t), h) with support radius of 2h,
Tab. I. The resulting L
2
norm error in dv/dt for selected
10
3
10
2
10
1
10
3
10
2
10
1
10
0
h/L
L
2


h/x = 0.354
h/x = 0.501
h/x = 0.708
h/x = 0.800
h/x = 1.001
Fig. 4. L
2
error norm in dv/dt for simulations using quadratic kernel and
Reynolds number equal to 100; gray dashed lines display functions k(h/L)
3
,
where k is a constant.
values of the smoothing length ratio h/x is plotted as a
function of h/L in Fig. 4. Results of the performed simulations
suggest that the discretisation error is about third-order in
h/L for sufciently high values of h/x and h/L, but is
discretisation-limited at low h/L. The observed levelling off
of the L
2
error norm is similar to the one observed by
Quinlan et al. [8] for SPH viscous ows and is likely to
be related to the application of the SPH approximation of
velocity gradients which are used for evaluation of viscous
uxes between particles. Poor results for smoothing length
ratio lower or equal to 0.501 can be related to insufcient
number of interacting particles, as plotted in Fig. 2. Although
the simulation tests have been run for the smoothing length
ratios up to 1.501, the accuracy of the performed simulations
is not signicantly altered by increasing its value above

2/2.
That suggests that an optimal number of neighbouring particles
can be found for smoothing length ratio close to this value.
The numerical experiments described above have been
repeated for different values of Reynolds numbers. The same
trends in behaviour of the L
2
error norm as in the previous
case have been observed. Selected results are plotted in Fig. 5
showing good agreement between the studied cases.
B. Selected kernel functions
The convergence of the FVPM simulations have been tested
for four selected kernel functions: quadratic, cubic spline, trun-
cated Gaussian and quintic spline function. Details on applied
kernel functions can be found in Tab. I. The quadratic kernel
function has been applied in the series of simulations tests
discussed in the previous section. The cubic spline function
is a commonly used kernel function in SPH simulations. The
same applies also to Gaussian and quintic spline function.
In general, the overall behaviour of L
2
error norm with
respect to h/L is consistent for all four applied kernels.
However, some differences may be observed. A comparison
of the L
2
error norm for smoothing length ratio equal to 0.8
5
th
international SPHERIC workshop Manchester, UK, June, 23-25 2010
10
3
10
2
10
1
10
3
10
2
10
1
10
0
h/L
L
2


h/x=0.354; Re=100
h/x=0.354; Re=1000
h/x=0.354; Re=10000
h/x=0.800; Re=100
h/x=0.800; Re=1000
h/x=0.800; Re=10000
Fig. 5. L
2
error norm in dv/dt for quadratic kernel simulations at several
Reynolds number; gray dashed lines display functions k(h/L)
3
, where k is
a constant.
is provided in Fig. 6. Since the quintic spline function is a
close approximation of the Gaussian function, the results of
numerical simulations for these two kernels show the same
trends and are quantitatively similar. The convergence is about
the same for all kernels at high values of h/L, but levelling
off of the error starts earlier in all cases than for quadratic
kernel.
The FVPM simulations with truncated Gaussian kernel
function have been tested for the values of support domain ra-
dius equal to 3h and 6h while identical results were achieved.
It is obvious that the increase in the radius of particle support
domain requires additional computational costs. The least
computationally expensive simulations have been the FVPM
calculations for quadratic kernel function, which were about
TABLE I
LIST OF APPLIED KERNEL FUNCTIONS, WHERE r
h
= x x
i
/h.
Quadratic kernel function
W =
_
4 r
2
h
0
0 r
h
< 2
2 r
h
Cubic spline kernel function
W =
_
_
_
(1 1.5r
2
h
0.75r
3
h
)/h
2
(2 r
h
)
3
/4h
2
0
0 r
h
< 1
1 r
h
< 2
2 r
h
Gaussian kernel function
W =
_
e
r
2
h/h
2
0
0 r
h
< 3
3 r
h
Quintic spline kernel function
W =
_

_
((3r
h
)
5
6(2r
h
)
5
+ 15(1r
h
)
5
)/h
2
((3r
h
)
5
6(2r
h
)
5
)/h
2
(3r
h
)
5
/h
2
0
0 r
h
< 1
1 r
h
< 2
2 r
h
< 3
3 r
h
10
3
10
2
10
1
10
3
10
2
10
1
10
0
h/L
L
2


Quadratic
Cubic spline
Gaussian
Quintic spline
Fig. 6. L
2
error norm in dv/dt for various kernel function denitions,
Re = 100 and h/x equal to 0.8; gray dashed lines display functions
k(h/L)
3
, where k is a constant.
40% faster than simulations for the Gaussian kernel with
support radius of 3h.
C. Exact evaluation of interparticle vectors
The kernel function W(x x
i
(t), h) may be dened as a
top-hat function, i.e. assigned a value of 1 inside the particle
support domain and 0 elsewhere. It has been shown [9] that
this choice allows the integral of equation (8) to be evaluated
analytically and exactly. This integration procedure is faster by
a factor of approximately 3 than numerical integration using
Gaussian quadrature with 6 6 integrating points, leading to
a signicant reduction of overall computational costs.
Tests for h/x 0.354, 1.501, x 0.002, 0.05 and
a uniform particle distribution have exhibited less sensitivity
of the FVPM simulation with exact
ij
evaluation to value of
smoothing length ratio, Fig. 7. Even for h/x approaching
its limit value

2/4, the convergence of the method remains
consistent with results for higher values of smoothing length
ratio.
When the interaction vectors
ij
are computed analytically,
the required computational effort is about 50% of that required
for numerical evaluation of
ij
for quadratic kernel function.
Decreasing the smoothing length ratio h/x, i.e. decreasing
the number of interacting particle neighbours, brings further
signicant reduction in computational costs. In particular,
decreasing h/x from 0.708 to 0.501 results in another
50% decrease in elapsed time for exact
ij
evaluation for all
interacting particles. This corresponds to 50% decrease in a
number of neighbouring particles as shown in Fig. 2.
D. Number of Integrating Points
In general, the numerical integration of interaction vectors

ij
is the most computationally demanding part of the FVPM
simulations. As pointed out in [7], it can require more than
80% of all computational costs for simulations in Lagrangian
frame. As discussed above, the expenses for this procedure
5
th
international SPHERIC workshop Manchester, UK, June, 23-25 2010
10
3
10
2
10
1
10
3
10
2
10
1
10
0
h/L
L
2


h/x = 0.354
h/x = 0.501
h/x = 0.708
h/x = 0.800
h/x = 1.001
Fig. 7. L
2
error norm in dv/dt for simulations using exact
ij
evaluation
and Reynolds number equal to 100; gray dashed lines display functions
k(h/L)
3
, where k is a constant.
may be decreased by exact
ij
evaluation utilising the top-hat
function. On the other hand, when vectors
ij
are integrated
using the Gaussian quadrature and a kernel function as in
Tab. I, the number of integrating points N
int
becomes another
important factor inuencing not only the computational costs
but also the accuracy of FVPM simulations.
All simulations discussed in sections III-A and III-B have
been performed for N
int
N
int
equal to 6 6 integrating
points. Herewith, the inuence of N
int
{1, 2, 4, 6, 8} on
convergence of the FVPM simulations is examined for a set of
tests outlined in section III-A, Reynolds number equal to 100
and the quadratic kernel function. Selected results are plotted
in Fig. 8.
From the results in Fig. 8, it may be concluded that the
accuracy of the FVPM simulations is not signicantly affected
by varying the number of integrating points for values of
N
int
4. However, values of N
int
2 are shown to be
insufcient. The computational costs for N
int
set to 4 are
about half of the costs for N
int
equal to 6.
In order to improve the results for small number of in-
tegrating points, several corrections to numerical evaluation
of vectors
ij
have been introduced in the past. The two
correction techniques that have been adopted in this work are
the pair-wise correction of Keck and Hietel [4] and Teleagas
correction [10]. However, the results of the performed simula-
tions utilising either Kecks or Teleagas correction technique
have shown a negligible inuence on the studied L
2
error norm
in dv/dt.
IV. CONCLUSION
The results presented in this work suggest that when the
smoothing length ratio h/x is large enough, i.e. there is a
sufcient number of neighbouring particles, the accuracy of
the FVPM calculations is not notably improved by further
increasing the value of smoothing length ratio. The empiri-
cally suggested value for h/x is about

2/2. Third-order
10
3
10
2
10
1
10
3
10
2
10
1
10
0
h/L
L
2


N
int
= 1
N
int
= 2
N
int
= 4
N
int
= 6
N
int
= 8
Fig. 8. L
2
error norm in dv/dt for various number of integrating points
N
int
, quadratic kernel and h/x equal to 0.8; gray dashed lines display
functions k(h/L)
3
, where k is a constant.
convergence can be observed for high values of h/L, but the
error does not fall below a nite minimum level.
Such a behaviour can be observed for all tested kernel
functions as well as for the exact
ij
evaluation technique.
The latter technique represents a signicantly faster approach
to
ij
calculations and enables application of smaller values
for the smoothing length ratio h/x, thus further decreasing
the computational costs. Nevertheless, further verication of
this approach especially for small values of the smoothing
length ratio is required.
When the interaction vectors
ij
are computed using the
Gaussian quadrature and the quadratic kernel function, the
results show low sensitivity to choice of number of integrating
points, provided that at least 4 4 points are used. The
adopted corrections of
ij
from the literature did not show
any signicant inuence on the methods convergence for the
studied case.
Although further work is required to conrm whether the
residual error is dominated by the SPH-based viscous stress
calculation, the presented results already provide some of the
knowledge required for further extension and optimisation of
the method in order to increase its capability to handle large
and computationally demanding problems.
ACKNOWLEDGMENT
The research leading to these results has received funding
from the European Communitys Seventh Framework Pro-
gramme (FP7/2007-2013) under grant agreement no. 225967.
REFERENCES
[1] G.K. Batchelor, An Introduction to Fluid Dynamics, Cambridge Univer-
sity Press, Cambridge, 1967.
[2] J. Bonet and T.-S. L. Lok, Variational and momentum preservation aspects
of smooth particle hydrodynamic formulations, Computer Methods in
Applied Mechanics and Engineering 180 (1999), pp. 97115.
[3] D. Hietel, K. Steiner and J. Struckmeier, A nite volume particle method
for compressible ows, Mathematical Models and Methods in Applied
Science 10 (2000) pp. 13631382.
5
th
international SPHERIC workshop Manchester, UK, June, 23-25 2010
[4] D. Hietel and R. Keck. Consistency by coefcient correction in the
nite volume particle method. In Proc. Meshfree Methods for Partial
Differential Equations, Lecture Notes in Computational Science and
Engineering, pp. 211221. Springer, Berlin, 2003.
[5] R. Keck and D. Hietel, A projection technique for incompressible ow in
the meshless nite volume particle method, Advances in Computational
Mathematics 23 (2005) pp. 143169.
[6] M. Junk and J. Struckmeier, Consistency analysis of meshfree methods
for conservation laws, Mitteilungen der Gesellschaft fur Angewandte
Mathematik und Mechanik 24 (2001), pp. 99126.
[7] R.M. Nestor, M. Basa, M. Lastiwka and N.J. Quinlan, Extension of the
nite volume particle method to viscous ow, Journal of Computational
Physics 228 (2009) pp. 17331749.
[8] N.J. Quinlan, M. Basa and M. Lastiwka, Truncation error in mesh-
free particle methods, International Journal for Numerical Methods in
Engineering 66 (2006) pp. 20642085.
[9] N.J. Quinlan and R.M. Nestor, Fast exact evaluation of particle interaction
vectors in the nite volume particle method, In Proc. Fifth International
Workshop on Meshfree Methods for Partial Differential Equations, Bonn
2009.
[10] D. Teleaga, A nite volume particle method for conservation laws, PhD
Thesis, University of Kaiserslauten, Kaiserslauten 2005.
[11] D. Teleaga and J. Struckmeier, A nite-volume particle method for con-
servation laws on moving domains, International Journal for Numerical
Methods in Fluids 58 (2008), pp. 945967.

You might also like