Chapter 6

Finite volume particle methods

(FVPM)
Francis Leboeuf
Ecole Centrale de Lyon - LMFA
36 avenue Guy de Collongue
69134 Ecully Cedex, France
Jean-Christophe Marongiu
ANDRITZ-HYDRO
Rue des Deux-Gares 6
CH 1800 Vevey, Suisse
6.1
6.2
6.3
6.4
6.5
6.6

6.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Partition of unity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Average of a function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Derivatives of . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conservation equation and FVPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

119
120
121
122
126
129

Introduction

The Finite Volume Particle Method (FVPM) is a particular SPH method

that has some nice features for conservation properties; it has been developed
by Hietel (2005), Struckmeier (2002, 2008) and applied to some complex ow
phenomena by Quinlan and his co-workers (2009).
FVPM is based on the use of a sort of Shepards kernel (4.42) but without
a weighting of the kernel W by the j .
W (x (t) xj , h)
W (x (t) xj , h)
(x (t) , xj , h) = 
= 
W (xj x (t) , h)
W (x (t) xj , h)
jNx

jNx

where Nx is the set of particles j which contain the point x in their own
domain of inuence Dj . Let
(x (t) , h) =


jNx

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

W (x (t) xj , h)
(x (t) , h)

119

120

Computational Fluid Dynamics

Downloaded by [Banaras Hindu University] at 19:21 15 September 2016

Note that is dierent from one. In general (see Figure 6.1), this is because
the summation is done on the values at x of the kernel function of the particles
j that belongs to Nx ; it does not represent an integral of the kernel function
associated to the particle at x as in (4.9). As a consequence, this is quite
dierent from the normalization used by Feldman and Bonet (2007) in their
CSPH method (see paragraph 5.4.4).
The disadvantage of using instead of W is that the norm depends on
(x, t) and so also, while W depends only on (x xj , t); as a consequence,
the general shape of as to be recomputed for each position (x, t).

6.2

Partition of unity

The main advantage of the kernel function i is that they form a partition
of unity:

(x(t), xi , h) = 1
(6.1)
jNx

because

iNx


W (x(t) xi , h)
 W (x(t) xi , h)
iNx
= 
=1
(x(t), xi , h) =
(x(t), h)
W (x(t) xj , h)
iNx

jNx

As a consequence, the analogy with Finite Volume is closer because the SPH
box is closed (see paragraph 5.2.5).
Note that condition (6.1) is always fullled (at least theoretically) even close
to the boundary , where the domain D(xi ) of the particles are truncated.
To be more clear, consider the Figures 6.1, 6.2 and 6.3 similar to those found
in Junk and Struckmeier (2002). In Figure 6.1, we get the particle positions

0
-4

-2

FIGURE 6.1: Particles position xi and function Wi .

xi and the associated function Wi . Note the symmetry of each function Wi

around each position xi . The sum of function Wi is given in Figure 6.2 for

Finite volume particle methods (FVPM)

121

0
-4

-2

FIGURE 6.2: The function corresponding to Figure 6.1.

Downloaded by [Banaras Hindu University] at 19:21 15 September 2016

0
-4

-2

FIGURE 6.3: The kernel i .

each position x. The kernel i is given in Figure 6.3. Note that this function
is non-symmetric around xi because depends on (x, t) and not on (x xj , t)
and that the maxima of i vary along x.

6.3

Average of a function

A dierence with the classical SPH method is related to the expression of

the average of a function over the domain D(x):

1
(x, t) =
(x , t) (x (t) , x , h) dxd
(6.2)
V (x, t)
D(x)

This discrete value i is associated with each particle, as well as the volume
V:

V (x, t) = (x (t) , x , h) dxd
(6.3)
Di

Their discrete analogues are:

(xi , t) 


1
(xj , t) (xi (t) , xj ) j
V (xi , t)
jNi

and
V (xi , t) 


jNi

(xi (t) , xj ) j

(6.4)

122

Computational Fluid Dynamics

Note also that

(xi , x , t) dxd =

V (xi , t) =
Di

W (xi x , h) d
1
dx
(xi , t)
(xi , t)

Di

Thismeans also that the density can be computed from i = mi i =

W (xi xj , h). This expression is particularly suited to the computami

Downloaded by [Banaras Hindu University] at 19:21 15 September 2016

jDi

tion of the density closed to a boundary on a liquid-gas interface, provided

that the summation is performed on the particles located on one side of the interface only, thereby allowing for a density gradient almost naturally. This
is the basis of the method developed by Hu and Adams (2006).

6.4
6.4.1

Derivatives of
Lagrangian derivative of


As we have a partition of unity



(x (t) , xi ) = 1, then

iNx

(x (t) , xi )

iNx

(x (t) , xi ) = 0

iNx

We will note thereafter:

Wi = W (xi (t) xj )
and
i = (xi (t) , xj ) =

W (xi (t) xj )
(xi (t))

d
i
Assuming that h is constant, then dt
Wi = 0 W
i Wi = 0. We
t + x
compute the gradient of from the gradient of W . We start from
Wi
Wi
i
Wi
i 
i =
i =
i =

Wj
i
i
i
i
i
jNx

and
i
1 Wi
i i
=

t
i t
i t
i  Wj
1 Wi

=
i t
i
t
jNx

x i
i 
Wi +
x j Wj
i
i
jNx

Finite volume particle methods (FVPM)

123

In order to compute the Lagrangian derivative of i , we need:


 
Wi
Wj
Wi
i
i 
= x i
x j i
+
x j Wj =
x i j
t
i
i
i
i
jNxi

jNxi


 
i
i
x i j i x j i j + i j
(x i x j )
=
t
i

(6.5)

Downloaded by [Banaras Hindu University] at 19:21 15 September 2016

jNx


  Wi
Wj
j
i
i
i
jNxi






i
i
j i + i
i j + j
=
i
i

i =

jNx

i =

(j i i j )

(6.6)

jNx

Assume that

We can write:



Wj
i
= i j + j
ij = i
i
i

(6.7)


i (x, t)
(x i ji x j ij )
=
t

(6.8)

jNx

i (x, t) =

(ji ij )

(6.9)

jNx

The Lagrangian derivative of i is then



i
di
=
+ x i i =
(x i ji x j ij ) + x i
(ji ij )
dt
t
jNx

jNx





i
di
=
(x j x i ) ij = i
(x i x j ) j + j
dt
i
jNx

(6.10)

jNx

Note that the summation in all these equations (6.5) to (6.10) is not related
to integration as in (6.4) but to the normalisation in the kernel i .

6.4.2

Other useful coecients and closed box condition

Dene the following integrated coecients by

ij =
ij dxd

124

Computational Fluid Dynamics

ij = ij ji

(ij ji ) dxd

ij =



i
i
=
j i j i
i j + i j
dxd
i
i

(i j j i ) dxd

Downloaded by [Banaras Hindu University] at 19:21 15 September 2016

ij =

Note that the integration has to be done on a domain that is the intersection of the spheres of inuence of the particles i and j. This is certainly a
key point because these coecients play an important role in the transport
equations. This seems also to be a very time consuming process. For instance,
Nestor (2009) reports 6d Gauss points to get an acceptable accuracy.
Integrating by parts, we get:


d
ij = 2
i j dx
i j ndxd1

ij =

=2

j dxd

i j ndxd1

i j dxd

i n

j dxd1

ij =

i ndxd1

(6.11)

ij
|ij | ,

Setting nij =
ij is the analogue in FVPM of the product of a surface
element |ij | and a normal nij joining two cell centres. So that (6.11) expresses
the closed box condition if the FVPM cell is far from the boundary :

ij = 0. We note also that ij = ji , ii = 0
j

REMARK 6.1 It seems that (6.11) is not numerically fullled despite

of the normalisation of the kernel function through the use of the FVPM
kernel i probably because of the numerical errors on the computation of ij .
For instance Teleaga (2008) reports how to correct the uxes
 n terms in the
n
=
ij . Nothing is
conservation equation through an articial term ii
j=i

said on the impact of this correction for the global conservation of the ow

Finite volume particle methods (FVPM)

125

quantities. A dierent technique of correction of ij is also described by Keck

(2002).

6.4.3

Transport of the volume Vi

Downloaded by [Banaras Hindu University] at 19:21 15 September 2016

We have
d
d
(Vi (t)) =
dt
dt

i (t, x) dx =
Di

i d
dx +
t

Di

d1
i xndx

Di


d
d1
(Vi (t)) =
(x i ji x j ij ) +
i xndx

dt
jNi

(6.12)

Di

This is the equivalent for FVPM of (4.4) used for classical SPH method.

6.4.4

On the computation of the gradient i

Three formulations may be used:

the mixed method of Bonet and Lok (1999) (4.44) that is based on
Shepards kernel.
the method of Nestor (2009)
the method of Keck (2005).
We shall detail the last two methods as the rst one has already been given.

6.4.5

Method of Nestor

Nestor proposes to use a renormalized

 kernel Sj (bi ) based on the barycentre bi of each particles, where bi = V1i xi dx. We have then
Di

i =

j (bi )
Vj (j i ) S

jNi

j (bi ) =
S

1
Vj Sj (bi ) (bj bi )

Sj (bi )

jNi

Sj is a cubic-spline used by Monaghan and Lattanzio (1985).

126

6.4.6

Computational Fluid Dynamics

Method of Keck

The method of Keck is based on consistency points xij , which may be

interpreted as midpoints of general cell faces. The consistency points are
used to modify the discrete gradient operator so that we obtain a consistent
approximation. We have then

1 
1 1
Vi I =
(bj ij ) =
xj dx ij
2
2
Vi

Downloaded by [Banaras Hindu University] at 19:21 15 September 2016

jNi

jNi

Di

In practice, xij are explicitly given by:


xij = bi +

1
ik ik

ij Vi

kNi

Note that in general xij = xji . The consistent form of the gradient is then
given by:

()i = V1i
ij rec (xij )
iNi
 
rec (xij ) = (bi ) + (xij bi ) T (bi ) + o h2
()i = (bi ) + o (h)
where rec is the reconstructed function at the consistency point xij . Linear
functions are then properly reproduced.

6.5

Conservation equation and FVPM

We introduce again a weak formulation of the conservation equation (5.1)




+ .F () i dxd =
S () i dxd
t
D(x)

D(x)

Integrating the rst member by parts, we get





i
i d
+ . (F () i ) dxd =
dx +

F () i dxd
t
t
D(x)

D(x)

D(x)

We use also the Reynolds theorem (5.5) in the form

i d
dW
dx +
i dxd =
. (i W ) dxd
dt
t
D(x)

D(x)

D(x)

Finite volume particle methods (FVPM)

127

Combining the two equations, we get

dW
i dxd +
. ([F () W ] i ) dxd
dt
D(x)

Downloaded by [Banaras Hindu University] at 19:21 15 September 2016

D(x)

D(x)

i d
dx +

F () i dxd +
D(x)

S () i dxd
D(x)

The derivatives of i are expressed with (6.8) and (6.9) which gives:



d
i dxd =

(x i ji x j ij ) dxd
dt
D(x)

jNx

D(x)

F ()

(ji ij ) dxd

jNx

D(x)

(F () x i ) i ndxd1

D(x)

S () i dxd

+
D(x)

d
dt

i dxd =

jNx

D(x)

(F () x j ) ij dxd

D(x)

(F () x i ) ji dxd )
D(x)

(F () x i ) i ndx

d1

S () i dxd

D(x)

D(x)

This equation appears in an ALE form with the transport of the particle at
the velocity v0 = x i
We state Gi = F () x i so that we get:

i dxd =
Gj ij dx3
Gi ji dxd

dt
jNx

D(x)

D(x)

D(x)

(6.13)

D(x)

Gi i ndxd1 +

S () i dxd
D(x)

Two important hypotheses are now used in order to reduce the cost of the
integrations:

128

Computational Fluid Dynamics

1. The variation of are small over the domain of intersection of the

two spheres i and j.

2. Gj ji dx Gj ji

Using these hypotheses, the conservation equation can be written:


d
d
d1
i dx
{Gi ji Gj ij }
Gi i ndx
+
S () i dxd
dt
jNx

D(x)

D(x)

D(x)

Downloaded by [Banaras Hindu University] at 19:21 15 September 2016

(6.14)
We can also write:


{Gi ji Gj ij }

jNx

(Gi Gj )

jNx

 (Gi + Gj )
(ij + ji )

(ij ji )
2
2
jNx

Using the rst hypothesis, we then have:


{Gi ji Gj ij }
jNx

 (Gi + Gj )
(ij ji )

2
jNx

 (Gi + Gj )
ij
2

jNx

So that:
d
dt
=


jNx

|ij |

(Gi + Gj )
nij
2

Gi i ndxd1 +

D(x)

i dxd (6.15)
D(x)

S () i dxd
D(x)

(Gi +Gj )
nij is
2
point xij . We

just a numerical ux computed in the direction nij , at the

shall write it in a more general form Gm
ij (t, i , xi , j , xj , nij ).
Note that i is inuenced by the particles j whose domain has a non-zero
intersection with the domain Di .
Finally, equation (6.15) is then:



d
d1
(i Vi ) =
|ij | Gm

ndx
+
S () i dxd (6.16)
i i
ij
dt
jNi

D(x)

D(x)

This form looks very similar to the weak form of Vila (2005) (5.17) for instance.
It is adapted to the use of Riemann solvers.

Finite volume particle methods (FVPM)

Downloaded by [Banaras Hindu University] at 19:21 15 September 2016

6.6

129

Concluding remarks

FVPM is in practice very close to SPH methods. It is possible to apply

the same techniques of improvement of the computation of the gradient, and
an ALE formulation can also be derived, allowing the use of various Riemann
solvers. The theoretical advantage of local conservation associated with the
partition of unity linked to the FVPM kernel is limited to the accuracy of the
numerical integration of coecient ij . The FVPM method will then have
an interest only if a quick algorithm can be used for the computation of the
integrals in the intersection of two spheres.