Meshless methods for conservation laws

D. Hietel1 , M. Junk2 , J. Kuhnert1 , and S. Tiwari1

1
Fraunhofer-Institut für Techno-und Wirtschaftsmathematik,
Gottlieb-Daimler-Straße, Geb. 49, 67663 Kaiserslautern, Germany
2
Fachbereich Mathematik, Universität Kaiserslautern,
Erwin-Schrödinger-Straße, 67663 Kaiserslautern, Germany

Summary. In this article, two meshfree methods for the numerical solution of con-
servation laws are considered. The Finite Volume Particle Method (FVPM) gen-
eralizes the Finite Volume approach and the Finite Pointset Method (FPM) is a
Finite Difference scheme which can work on unstructured and moving point clouds.
Details of the derivation and numerical examples are presented for the case of in-
compressible, viscous, two-phase flow. In the case of FVPM, our main focus lies on
the derivation of stability estimates.

1 Introduction
Meshfree techniques play an increasing role as solution methods for conser-
vation laws. Practically, all meshfree methods are based on clouds of points,
where each point carries the relevant information for the problem to be solved.
One of the biggest advantages of these methods is that no mesh has to be
established. Generating a grid can be very costly, sometimes it is even the
dominating part in the problem. Compared to that, it is relatively simple to
establish a point cloud, even within very complex geometries. Moreover, the
cloud is easy to maintain or to adapt locally. Due to the free movement of the
points, an optimal adaptivity of the cloud is provided towards changes in the
geometry or towards movement of free surfaces as well as phase boundaries.
Among the pioneering meshfree methods, Smoothed Particle Hydrody-
namics (SPH) is certainly the most famous [10]. SPH is a Lagrangian idea,
which is based on the movement of finite mass points. However, for a long
time, SPH was suffering from several problems, among them stability and
consistency. Facing these problems, the development of meshfree methods
went into various directions, starting in the early nineties. On one hand, peo-
ple tried to improve SPH. One idea was to avoid inconsistency problems of
SPH by reproducing kernel methods [30], another idea was to improve the ap-
proximation properties of SPH by the introduction of so called Moving Least
Squares (MLS) ideas [5, 15]. On the other hand, many new types of meshfree

preprint -- preprint -- preprint -- preprint -- preprint

 2 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari

methods were developed. Widely used methods are the Element Free Galerkin
(EFG) idea [23] or the Partition of Unity Method (PUM) [8]. As the EFG and
PUM ideas provide the possibility to carry out Finite Element computations
on gridfree structures, the present paper introduces meshfree Finite Volume
(FV) and Finite Difference (FD) concepts.
In section 2, we present the Finite Volume Particle Method (FVPM), which
incorporates FV ideas into a meshfree framework [7]. In particular, the ap-
proach uses the concept of numerical flux functions and guarantees conserva-
tion on a discrete level. As we show in section 2, the similarity to classical
Finite Volume schemes allows us to derive stability results using standard ar-
guments. For numerical results obtained with FVPM, we refer to [7, 12, 24, 31].
In section 3, we concentrate on the Finite Pointset Method (FPM) which
is a general Finite Difference Method for conservation laws on a meshfree basis
[15, 16]. The latest FPM development which covers the discretization of the
incompressible Navier-Stokes equations with multiple phases is presented.

2 The Finite Volume Particle Method

2.1 Derivation

The Finite Volume Particle Method (FVPM) has been developed in an at-
tempt to combine features of SPH (Smoothed Particle Hydrodynamics) with
Finite Volume Methods (FVM) [7]. To explain the idea, we consider the prob-
lem to find a function u : [0, T ] × Rd → R which satisfies
∂u
+ divx F (u) = 0, u(0, x) = u0 (x). (1)
∂t
In the classical Finite Volume Method, conservation laws like (1) are dis-
cretized by introducing a Finite Volume mesh on Rd and integrating (1) over
each volume element (see, for example, [6, 14]). To reformulate this procedure
on a more abstract level, we note that a mesh gives rise to a particular par-
tition of unity which is generated by the indicator functions of the elements.
More specifically, if Rd is the disjoint union of mesh d
P cells Ci ⊂ R , then the
functions χi = 1Ci have the property χi ≥ 0, and i χi = 1 where the sum is
locally finite because at every x ∈ Rd , at most one function χi is nonzero (the
one for which x ∈ Ci ). The Finite Volume scheme is then obtained by taking
χi as test functions in a weak formulation of (1). In this context, the basic
idea in FVPM is to choose a smooth partition of unity instead of a mesh-based
partition. Let us therefore assume that a smooth, locally finite partition of
unity {ψi }i∈I is given on Rd , i.e.Psmooth and compactly supported functions
ψi : Rd → R with ψi ≥ 0 and i ψi = 1 where ψi (x) 6= 0 for only finitely
many indices at every x ∈ Rd (the functions ψi are called particles). Such
a partition can be constructed using a shape function W which is smooth,
compactly supported around the origin, and strictly positive on its support,

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 3

for example, a radially symmetric cubic spline, or the d-fold tensor product
of a one-dimensional compactly supported function (in the first case, supp W
is a d-dimensional ball, and in the second case, the support is an axis parallel
cube). Then given a suitable set of points xi ∈ Rd , i ∈ I and some h > 0, we
define scaled and shifted versions of W whose supports cover Rd
 
x − xi [
Wi (x) = W , supp Wi = Rd . (2)
h i

Finally, using Shephard’s method [22], the partition of unity is built

Wi (x) X
ψi (x) = , σ(x) = Wk (x), x ∈ Rd (3)
σ(x)
k

and the partition functions ψi are used as test functions for equation (1).
Multiplying (1) with ψi and integrating over Rd , we obtain after integration
by parts
d
Z Z
ψi u dx − F (u) · ∇ψi dx = 0. (4)
dt Rd Rd
In order to split the flux integral into pairwise flux contributions
P between
particle
P i and its neighboring particles j, we use the fact that j ψ j = 1 and
∇( j ψj ) = 0 which leads to

d
Z XZ
ψi u dx − F (u) · (ψj ∇ψi − ψi ∇ψj ) dx = 0. (5)
dt Rd j Rd

The first integral is obviously related to a local average of u

1
Z Z
uni = ψi u dx|t=tn , Vi = ψi dx. (6)
Vi R d Rd

To approximate the second integral in (5), we note that, if u varies only slightly
around ū on the intersection of the supports of ψi and ψj , we have
XZ
− F (u) · (ψj ∇ψi − ψi ∇ψj ) dx ≈ F (ū) · βij ,
j Rd

where Z
βij = ψi ∇ψj − ψj ∇ψi dx. (7)
Rd
Following the usual procedure in the Finite Volume discretization we approx-
imate the flux
β ij
F (ū) · βij = |β ij |F (ū) · nij , nij = if |β ij | 6= 0
|β ij |

in terms of the discrete values with the help of a numerical flux function

preprint -- preprint -- preprint -- preprint -- preprint

 4 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari

F (ū) · nij ≈ g(ui , uj , nij ).

Finally, using an explicit Euler discretization of the time derivative, we obtain

a fully discrete approximation of (5)

1
X Z
0
un+1
i V i = u n
V
i i − ∆t |β ij |g(u n
i , u n
j , n ij ), u i = ψi u0 dx. (8)
j
V i R d

Observe that (8) has exactly the structure of the classical Finite Volume
Method [6, 14]. The only difference appears in the definition of the geomet-
ric parameters Vi and β ij . While they are integral expressions involving the
partition of unity functions ψi in the case of FVPM, they are mesh related
quantities in the classical Finite Volume case: Vi is the volume of cell Ci and
β ij = |Γij |nij is the product of interface area |Γij | and interface normal vector
nij (see figure 1). In other words, FVPM is very similar to FVM – only grid

Ci
PSfrag replacements
Γij

nij
Cj

Fig. 1. Control volume Ci with interfaces Γij and outer normals nij

generation is replaced by integration. From the update rule in (8) it is clear

that the value ui at particle i is only influenced by those values uj for which
β ij 6= 0. In view of (7), this can only happen if the particles i and j overlap.
Hence, it is natural to call ψi and ψj (interacting) neighbors if β ij 6= 0. In
particular, the sum in (8) only involves the particles from the neighbor list
Ni = {j : βij 6= 0}. Note also that particle i does not interact with itself since
β ii = 0 according to (7).
Before investigating the scheme (8) more closely, let us remark that the
derivation works similarly in the case when the partition of unity is time
dependent (moving particles). For simplicity, we assume here that the time
dependence is such that ψi satisfy the advection equation
∂ψi
+ a · ∇ψi = 0 (9)
∂t
where a is a smooth and bounded velocity field. Note that {ψi (t, ·)}i∈I is auto-
matically a partition of unity if this is true initially. However, the parameters

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 5

Vi and β ij now depend on time. If we use these functions ψi as test functions,

the time derivative gives rise to an additional term
 
∂u d
Z Z Z
ψi + divF (u) dx = ψi u dx − (F (u) − ua) · ∇ψi dx
Rd ∂t dt Rd Rd

which is of the same form as (4) if F (u) is replaced by the Lagrangian flux
G(t, x, u) = F (u)−ua(t, x). Another possibility to generate a time dependent
partition of unity is to move the points xi according to the vector field a and
to continuously apply Shephard’s method to Wi (x) = W ((x − xi )/h). Also
in this case, the modifications can be incorporated into the flux function by
going over from F to G (for details, we refer to [11]).
In the following, we study mathematical properties of the Finite Volume
Particle Method. For numerical results, we refer to [7, 12, 24, 31].

2.2 Geometric parameters

Inspecting the convergence proof for classical Finite Volume schemes (e.g.
[1, 3, 14, 29]), one observes that there are only a few requirements on the
geometric parameters β ij , Vi : the number of cell faces should be bounded, the
cell surface area j |βij | should be of the order hd−1 , and the cell volumes
P

Vi should be bounded from below by αhd . Moreover, there are two important
algebraic requirements. The first one reflects the fact that, if two cells Ci , Cj
share a common interface, then the surface area |β ij | = |βji | is equal but
the orientation β ij /|β ij | = −βji /|βji | is opposite. The second requirement
is related to the divergence theorem. PIf β ij is the product of interface area
and interface normal, then the sum j βij is nothing but the integral of the
normal vector over the surface of cell Ci . According to the divergence theorem,
this surface integral can Pbe written as a volume integral of the divergence of
constant fields so that j βij = 0.
In the following, we show that the same conditions are satisfied if the
geometric parameters β ij , Vi are not based on a mesh but are obtained by
integration from a partition of unity. As a consequence, many aspects of con-
vergence proofs for classical Finite Volume schemes can be taken over without
modification (examples are given in the following section). We start by a pre-
cise statement of our assumptions concerning the partition of unity.
(P1) The initial partition of unity is constructed according to (2), (3), based
on a locally finite point distribution {xi }. We assume that the maximal number
of overlapping particles is bounded, i.e.

max #{j : supp Wi ∩ supp Wj 6= ∅} = K < ∞. (10)

i∈I

Moreover, there should be a minimal overlap such that

X
Wi (x) ≥ µ > 0 ∀x ∈ Rd . (11)
i∈I

preprint -- preprint -- preprint -- preprint -- preprint

 6 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari

If we consider a sequence of partitions based on Wih (x) = W ((x−xhi )/h) with

h → 0, we assume that K and µ are h-independent.
(P2) The functions ψi (t, ·), t > 0 are obtained by solving (9) with initial
values given by the functions of the initial partition. The field a is assumed
to be bounded and continuously differentiable with bounded derivatives.
Using standard results from the theory of ordinary differential equations,
we conclude that the initial value problem ẋ(t) = a(t, x(t)), x(τ ) = ξ admits
a unique global solution which we denote by t → X(t, ξ, τ ). The function X
is smooth and invertible with respect to the ξ-variable, where

X(t, X(τ, y, t), τ ) = y ∀y ∈ Rd .

The Jacobian determinant of X(t, ξ, τ ) is given by

Z t 
J(t, ξ, τ ) = exp (diva)(s, X(s, ξ, τ )) ds .
τ

According to the method of characteristics, the solution of (9) satisfies

ψi (t, x) = ψ(0, X(0, x, t)) (12)

which immediately implies that {ψi (t, ·)}i∈I is a partition of unity. We want
to stress again that the restriction to the construction (9) is not compulsory. It
only helps us to avoid additional technical arguments and assumptions which
are necessary, for example, if we use moving points xi and construct the
partition from the functions Wi which have moving supports of fixed shape.
Also in that more technical situation, the following result can be shown.
Proposition 1. Assume (P1) and (P2) and let Vi (t), β ij (t) be defined by (6)
and (7) based on the partition functions {ψi (t, ·)}i∈I . Then there exist con-
stants α, K, C > 0 such that for t ∈ [0, T ] the number of neighbors is bounded
#Ni (t) ≤ K and
X
Vi (t) ≥ αhd , |β ij | ≤ Chd−1 . (13)
j

Moreover, the relations

X
β ij = −βji , βij = 0, i, j ∈ I (14)
j

are satisfied.
Proof. The algebraic conditions (14) follow directly from the skew-symmetric
definition
P of β ij .RThe vanishing sum over β ij is a consequence of the fact that
j ψ j = 1, and ∇ψi dx = 0
X Z X Z
β ij = ψi ∇ ψj dx − ∇ψi dx = 0.
j Rd j Rd

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 7

While these arguments do not depend on the structure of the partition, the
estimates (13) require more details. We begin with the investigation of the
volume.
Z Z Z
Vi (t) = ψi (t, x) dx = ψi (t, X(0, x, t)) dx = ψi (t, y)J(t, y, 0) dy.
Rd Rd Rd

If m is a lower bound for diva, then J(t, y, 0) ≥ exp(tm) and it suffices to

estimate ψi (0, y) = ψi0 (y) from below. In view of (10) and (12), the maximal
number of overlapping particles is K (which also proves the estimate of #Ni )
and in connection with (11), we conclude
X
µ ≤ σ(y) = Wk (y) ≤ KkW k∞ , ∀y ∈ Rd . (15)
k

Continuing the estimate of Vi , we have

Wi (y) exp(T m) d
Z Z
Vi (t) ≥ exp(tm) dy ≥ h W (z) dz = αhd
R d σ(y) KkW k ∞ R d

Next, we turn to the estimate R of the sum over |β ij |. Allying integration by

parts to (7), we have β ij = 2 ψj ∇ψi dx, so that
X Z Z
|β ij | ≤ 2 |∇ψi | dx ≤ 2 |∇ψi | dx.
j Rd Rd

Using chain rule and the same change of coordinates as above, we obtain
X Z
|βij | ≤ 2 |(∇X)T | |∇ψi0 |J dy.
j Rd

Gronwall’s lemma implies that the derivatives of X can be bounded in terms

of the derivatives of a. If |∇a| ≤ M then |(∇X)T | ≤ exp(tM ) and J ≤
exp(tM d). Hence, it suffices to estimate the integral over |∇ψi0 |. Definition
(3) implies P
∇Wi ∇Wk
∇ψi0 = − ψi k ,
σ σ
so that with (15)

2 2 d−1
Z Z Z
0
|∇ψi | dy ≤ |∇W |(y/h) dy = h |∇W (z)| dz
Rd hµ Rd µ Rd

and hence
X 2(K + 1) d−1
|β ij | ≤ exp((d + 1)T M ) h .
j
µ

This completes the proof. t

u

preprint -- preprint -- preprint -- preprint -- preprint

 8 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari

2.3 Stability
The aim of this section is to demonstrate that local averages uni generated
with FVPM can be estimated in the same way as in the mesh based Finite
Volume approach. Our main assumption is that the flux function g is mono-
tone. Note, however, that we allow for moving particles so that g should be
an approximation of the Lagrangian flux F (u) − ua(t, x). Consequently, the
flux between two particles will also depend on their location and time. Since
the dependence on location can be introduced in several ways (see below),
we do not specify it at this stage. More precisely, we incorporate the possible
movement of the particles into (8) by assuming the form
X
un+1
i Vin+1 = uni Vin − ∆t Gnij (uni , unj ) (16)
j

with initial values

1
Z
u0i
= 0 u0 (x)ψ(0, x) dx.
Vi R d
where the flux function Gnij should satisfy the following requirements.
(Z) Local interaction is ensured by
Gnij (u, v) = 0 for j 6∈ Nin .
(S) The flux function should be antisymmetric in the following sense
Gnij (u, v) = −Gnji (v, u).
(C) The relation to the Lagrangian flux is ensured by
Gnij (u, u) = F (u) · βnij − uAnij · β nij
where Anij approximates a on the support of ψi
|Anij − a(tn , xni )| ≤ ch, j ∈ Nin (17)
where xni is the barycenter of particle i
1
Z
xni = n xψ(tn , x) dx.
Vi R d
(L) The flux function should be locally Lipschitz continuous, i.e. for a ≤
u, v, w ≤ b we assume
|Gnij (u, w) − Gnij (v, w)| ≤ |βij |L(a, b)|u − v|.
(M) Finally, we assume monotonicity in the arguments of Gnij . For u 6= v
Gnij (u, w) − Gnij (v, w) Gnij (w, u) − Gnij (w, v)
≥ 0, ≤ 0.
u−v u−v
In order to prove a conservation property of (16) we assume that the flux
F (u) in (1) is defined for u = 0 (which can always be achieved by a simple
transformation).

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 9

Proposition 2. Assume (Z), (S), (C). If the initial value u0 of (1) has a
compact support then uni 6= 0 for only finitely many indices i and
X X
un+1
i Vin+1 = uni Vin .
i i

Proof. According to (P1), the point distribution {xi } is locally finite which
implies that the number of particles ψi (0, ·) whose support intersect with the
one of u0 is finite. In P
particular, only finitely many average values u0i are
non-zero and the sum i u0i Vi0 is well defined. Considering a particle i with
vanishing average u0i = 0 and also u0j = 0 for all neighbors j ∈ Ni0 , we have
X X
u1i Vi1 = −∆t G0ij (0, 0) = −∆t F (0) · βnij − 0 · Anij · β nij


j j

where (C) has been used. Since the sum over β ij vanishes according to (14),
we see that u1i = 0. This situation occurs forP all but finitely many particles
and hence u1i 6= 0 only finitely often so that i u1i Vi1 is well defined. Summing
over (16), we get
X X X
u1i Vi1 − u0i Vi0 = −∆t G0ij (u0i , u0j )
i i i,j
 
1X 1X
= −∆t  G0ij (u0i , u0j ) + G0ji (u0j , u0i ) .
2 2
i,j j,i

Using (S), we conclude that the right hand side vanishes which yields conser-
vation. Using an induction argument, the general statement follows. t
u
The second important property is monotonicity of the scheme.
Proposition 3. Assume (L), (M) and let a ≤ uni , win ≤ b. If for some ξ ∈
(0, 1) the CFL condition
ξVin
∆t ≤ min P
i∈I L(a, b) j |βij |

is satisfied then un+1

i ≤ win+1 .
Proof. Assuming first that uni 6= win , we obtain from (16)
X
(win+1 − un+1 )Vin+1 = −∆t Gnij (uni , wjn ) − Gnij (uni , unj )


i
j
 
X Gnij (win , wjn ) − Gnij (uni , wjn )
+ Vin − ∆t  (win − uni ) (18)
j
win − uni

Using the Lipschitz property

preprint -- preprint -- preprint -- preprint -- preprint

 10 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari
 
X Gnij (win , wjn ) − Gnij (uni , wjn ) ∆t X
Vin − ∆t ≥ Vin 1 − n |β ij |L(a, b)
j
win − uni Vi j

so that the second term on the right of (18) is positive if the CFL condition
is satisfied. The positivity of the first term follows immediately from (M). In
the case uni = win , the second term is not present and non-negativity follows
from the first term, again with the help of (M). t
u
An immediate consequence of monotonicity is L∞ -stability. Due to the
possible movement of the particles, however, the maximum may increase in
time.
Proposition 4. There exist h-independent constants γ1 , γ2 such that under
the assumptions of Proposition 2 together with (C) and the additional restric-
tion ∆t ≤ α/(2cC), where α, C are from (13) and c from (17), we have

max un+1
i ≤ exp(γ1 ∆t) max uni if max uni ≥ 0,
i∈I i∈I i

min un+1
i ≤ exp(γ1 ∆t) min uni if min uni ≤ 0,
i∈I i∈I i
n+1
max ui ≤ exp(γ2 ∆t) max uni if max uni ≤ 0,
i∈I i∈I i
n+1
min ui ≤ exp(γ2 ∆t) min uni if min uni ≥ 0.
i∈I i∈I i

Proof. Using Proposition 2 with win = maxj unj = M ≥ 0, we obtain un+1 i ≤

win+1 which yields the desired estimate if win+1 ≤ exp(γ1 ∆t)M . Since win =
M , we obtain with (C)
X
win+1 Vin+1 = M Vin − ∆t F (M ) · β nij − M Anij · βnij


j
 
∆t X
= Vin 1 − n Anij · β nij  M.
Vi j

The extra information about Anij in (C) implies

   
∆t X n n
 ∆t X n n
 ∆t X n
A ij · β ij ≤ a(t n , xi ) · β ij  + V n ch |β ij |
Vin  j  V n
i j i j

where the first term on the right vanishes because of (14). We conclude with
(13)  
∆t X n n
 cC
n Aij · β ij  ≤ ∆t.
Vi j α

Due to the movement of the particles, ψi (tn+1 , x) = ψi (tn , X(tn , x, tn+1 )

and since the Jacobian determinant follows the estimate J(tn+1 , bf y, tn ) ≥
exp(δ∆t) with δ being the minimum of diva, we have

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 11
Z
Vin+1 = ψi (tn , y)J(tn+1 , y, tn ) dy ≥ exp(δ∆t)Vin .
Rd

Combining these results and setting γ1 = cC/α − δ, we finally get win+1 ≤

exp(γ1 ∆t)M . Similarly, we can estimate the behavior of the minimum m =
mini uni ≥ 0. Setting win = m, we conclude from uni ≥ win that also un+1
i ≥
win+1 , and as above,
 
cC
win+1 ≥ 1 − ∆t exp(δ∆t)m.
α

Setting D = α/(2cC), we obtain 1 − ∆t/(2D) ≥ exp(− ln(2)∆t/D) for 0 ≤

∆t ≤ D so that the result follows with γ2 = δ − ln(2)/D. The remaining cases
where maximum or minimum are negative can be shown in the same way. t u.
We remark that no exponential factors appear in the estimates for max-
imum and minimum if the volumes are approximately calculated from the
formula  
∆t X
Vin+1 = V1n 1 + n Anij · β nij  .
Vi j

This is an approximation to the true volume evolution because the sum over
Anij · β nij can be regarded as approximation of diva if Anij is an evaluation of
a at a suitable point xnij (see [12] for details).
Apart from L∞ -estimates, the monotonicity can also be used to derive
a discrete Krushkov-entropy estimate. Moreover, a weak BV-estimate can be
shown along the lines of the proof in [1]. We will not go into further details here
but conclude with some comments on the construction of the flux function
Gnij . A particular example is based on the Lax-Friedrichs flux function

F (u) + F (v) n u − v u v
Gnij (u, v) = ·βij + − a(tn , xni )·β nij − a(tn , xnj )·β nij (19)
2 2λ 2 2
where xni , xnj are, for example, the barycenters of particles i and j at time tn .
If F is locally Lipschitz continuous and if the parameter λ is chosen such that
λ|F 0 (u) − a(t, x)| ≤ 1 for a ≤ u ≤ b, t ∈ [0, T ], and x ∈ Rd , then definition
(19) satisfies the requirements (Z), (S), (C), (L), (M).
More generally, if g0 (u, v, n) is any monotone and locally Lipschitz con-
tinuous numerical flux function which is consistent to F , then

M − a(tn , xni ) M + a(tn , xnj )

 
Gnij (u, v) = |β nij |g0 (u, v, nnij ) + u −v · β nij
2 2

satisfies all the requirements on the numerical flux function if M ≥ |a(t, x)|
for any t ∈ [0, T ], and x ∈ Rd . Instead of using two separate points, one can
also base the construction on a common point xnij for each pair of particles
such that xnij = xnji which replaces xni and xnj in the construction above.

preprint -- preprint -- preprint -- preprint -- preprint

 12 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari

We close with the remark that a full convergence proof of the scheme (16)
does not automatically follow from the estimates of the local averages u ni . This
is due to the fact that the reconstruction of an approximate solution uh from
the discrete values is not so straightforward as in the classical Finite Volume
case, where one sets
X
uh (t, x) = uni 1Ci (x)1[tn ,tn+1 ) (t). (20)
i∈I

The intuitive choice

X
uh (t, x) = uni ψi (x)1[tn ,tn+1 ) (t) (21)
i∈I

seems to be promising because uh is bounded in L∞ by the maximum of |uni |

and the reconstruction is conservative
Z X X Z
uh (tn , x) dx = uni Vin = u0i Vi0 = u0 (x) dx
Rd i i Rd

(although locally uh (tn , x)ψi (tn , x) dx 6= uni Vin , in general). The main diffi-
R

culty with the reconstruction (21) in the convergence proof is that it does not
commute with nonlinear functions. While in the FV approach (20) the flux of
the reconstruction is equal to the reconstruction of the fluxes,
X
F (uh (t, x)) = F (uni )1Ci (x)1[tn ,tn+1 ) (t),
i∈I

a similar relation is not available for the reconstruction (21). Therefore, the
estimates in terms of uni cannot immediately be used for the reconstruction
and further analysis is required for a complete convergence proof.

3 The Finite Pointset Method

As we have seen in section 2.1, the Finite Volume Particle Method can be in-
terpreted as a discretization of the weak form of conservation laws. The only
difference to the classical Finite Volume Method is that the test functions are
not constructed from mesh cells but are chosen as smooth partition of unity
functions. The flux-divergence divF is approximated with the help of geomet-
ric parameters which depend on the relative location of the particles (similar
to the classical Finite Volume Method where divF is replaced by an approx-
imate surface integral over the flux – reflecting the weak formulation). While
the weak approximation has the advantage of ensuring discrete conservation
and stability, it also requires a large numerical effort for the evaluation of the
geometric parameters and is, in its basic form, of low approximation order.

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 13

In the Finite Pointset Method (FPM) these advantages and disadvantages

are essentially reversed. The method is based on the strong form of the equa-
tion and therefore has the basic flavor of a Finite Difference scheme. Deriva-
tives are approximated using a least squares approach which requires only
little information about the relative location of the particles and easily allows
high order approximations. However, as a consequence of the higher geomet-
rical flexibility, discrete conservation cannot be proved (there is no volume
associated to the particles and therefore integral values are not defined as
naturally as in the Finite Volume approach), and also stability estimates are
not available through standard techniques, eventhough the method has been
successfully applied to a wide variety of problems.
In the following, we describe the adaption of the method to the case of
incompressible, viscous, multiphase flows.

3.1 Mathematical model and numerical scheme

We consider two immiscible fluids, for example, liquid and gas in a situa-
tion where the motion can be described by the incompressible Navier-Stokes
equations. They are, in Lagrangian form,
Dv 1 1
= − ∇ p + ∇ · (2µD) + g, (22)
Dt ρ ρ
∇·v = 0 (23)
where v is the fluid velocity vector, ρ is the fluid density, µ is the fluid viscosity,
D is the viscous stress tensor D = 21 (∇v + ∇T v) and g is the body force
acceleration vector.
For a numerical simulation of the process, we assume that the flow domain
is filled by a set of discretization points (point cloud). Each point (also referred
to as particle) locally represents a lump of fluid. The points carry all necessary
pieces of information (state variables) in order to completely describe the
local state of flow, such as density, velocity, pressure, etc. The particles move
exactly with fluid velocity, and, of course, the task of the numerical scheme
is to evolve the state variables at the particle locations in an accurate way.
The point cloud itself may be locally adaptive, i.e. the mean distance between
points may change locally due to certain criteria. That also means that, during
computation, new particles have to be created in sparse regions and removed
in dense areas.
In this paper, we initially provide each particle a flag representing the fluid
phase it belongs to. These flags do not change throughout the computation.
Moreover, the density and viscosity are constant on each particle path, so we
have
∂ρ
+ v · ∇ρ = 0 (24)
∂t
∂µ
+ v · ∇µ = 0. (25)
∂t

preprint -- preprint -- preprint -- preprint -- preprint

 14 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari

Therefore, each fluid particle has constant ρ and µ. Since ρ and µ are discon-
tinuous across the interface, the numerical scheme might have instabilities in
that particular region. Therefore, we consider a smooth density and viscosity
in the vicinity of an interface. The interface region can be detected by check-
ing the flags of particles in the neighborhood. If there are flags of only one
type in the neighbor list of some particle, then it is considered to be far from
the interface region. Near the interface, particles will find both types of flags
in the neighbor list. We modify the density and viscosity in each time step at
each particle position x near the interface by using the Shepard interpolation
Pn
wi ρ i
ρ̃(x) = Pi=1n (26)
wi
Pni=1
w i µi
µ̃(x) = Pi=1n , (27)
i=1 wi

where n is the total number of neighbor particles xi at x inside the compact

support of the weight function wi which is defined by (38). Occasionally, the
smoothing procedure of density and viscosity has to be repeated several times
in order to gain stability. Equations (22–23) are solved togehter with some
initial and boundary conditions.

Numerical scheme

Since the viscosity is smoothed near the interface, we can rewrite the momen-
tum equations whose spatial components are given by
du 1 ∂p 1 µ̃ 1 ∂ µ̃ ∂u ∂ µ̃ ∂v ∂ µ̃ ∂w
= gx − + ∇µ̃ · ∇u + ∆u + ( + + )
dt ρ̃ ∂x ρ̃ ρ̃ ρ̃ ∂x ∂x ∂y ∂x ∂z ∂x

dv 1 ∂p 1 µ̃ 1 ∂ µ̃ ∂u ∂ µ̃ ∂v ∂ µ̃ ∂w
= gy − + ∇µ̃ · ∇v + ∆v + ( + + )
dt ρ̃ ∂y ρ̃ ρ̃ ρ̃ ∂x ∂y ∂y ∂y ∂z ∂y

dw 1 ∂p 1 µ̃ 1 ∂ µ̃ ∂u ∂ µ̃ ∂v ∂ µ̃ ∂w
= gx − + ∇µ̃ · ∇w + ∆w + ( + + ).
dt ρ̃ ∂z ρ̃ ρ̃ ρ̃ ∂x ∂z ∂y ∂z ∂z ∂z

To ensure incompressibility, we use Chorin’s projection method [2]. Since we

consider the fully Lagrangian method, we first move particles with their old
velocities. The new positions are given by

xn+1 = xn + ∆t v n .

At each new particle position we modify the density and viscosity according
to (26) and (27), respectively and then compute the intermediate velocities
u∗ , v ∗ and w∗ implicitly by

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 15
n
∆t ∆t ∂v
u∗ − (∇µ̃ · ∇u∗ − µ̃∆u∗ ) = un + ∆t gx + ∇µ̃ · (28)
ρ̃ ρ̃ ∂x

∆t ∆t ∂v n
v∗ − (∇µ̃ · ∇v ∗ − µ̃∆v ∗ ) = v n + ∆t gy + ∇µ̃ · (29)
ρ̃ ρ̃ ∂y

∆t ∆t ∂v n
w∗ − (∇µ̃ · ∇w∗ − µ̃∆w∗ ) = wn + ∆t gz + ∇µ̃ · . (30)
ρ̃ ρ̃ ∂z
Then, at the second step, we correct v ∗ = (u∗ , v ∗ , w∗ ) by solving the
equation
∇pn+1
v n+1 = v ∗ − ∆t (31)
ρ̃
for pn+1 , with the incompressibility constraint

∇ · v n+1 = 0. (32)
By taking the divergence of equation (31) and by making use of (32), we
obtain the Poisson equation for the pressure
∇ · v∗
 n+1 
∇p
∇· = . (33)
ρ̃ ∆t
The boundary condition for p is obtained by projecting the equation (31) on
the outward unit normal vector n to the boundary Γ . Thus, we obtain the
Neumann boundary condition
 n+1
∂p ρ̃
= − (v n+1 − v ∗Γ ) · n, (34)
∂n ∆t Γ
where v Γ is the value of v on Γ . Assuming v · n = 0 on Γ , we obtain
 n+1
∂p
=0 (35)
∂n
on Γ .
We note that particle positions change only in the first step. The inter-
mediate velocity v ∗ is obtained on the new particle positions. The pressure
Poisson equation and the divergence free velocity vector are also computed
on the new particle positions.
We solve the above implicit equations (28–30) for the velocity vector v ∗
and pressure (33) by the constraint weighted least squares method which is
described in the following section.

3.2 FPM for solving general elliptic partial differential equations

To generalize the pressure Poisson problem, we consider the following linear

partial differential equation of second order

preprint -- preprint -- preprint -- preprint -- preprint

 16 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari

Aψ + B · ∇ψ + C∆ψ = f, (36)

where A, B, C and f are given. Note that for the pressure Poisson equation
(33), we have A = 0. The equation is solved with Dirichlet or Neumann
boundary conditions
∂ψ
ψ=g or = φ. (37)
∂n
We use the method proposed in [26] to solve the elliptic equation in a meshfree
framework. According to [9], it is more stable than the approach presented in
[17] and it easily handles Neumann boundary conditions.
Consider the computational domain Ω ∈ Rd , d ∈ {1, 2, 3}. Distribute N
particles xj ∈ Ω, j = 1, · · · , N , which are the discretization points and might
be irregular. Let x be an arbitrary particle in Ω, and we determine its neigh-
boring cloud of points. We introduce the weight function w = w(xi − x, h)
with small compact support of radius h. The weight function can be arbi-
trary but in our computation, we consider a Gaussian weight function of the
following form
( 2
exp(−α kxih−xk
2 ), if kxih−xk ≤ 1
w(xi − x; h) = (38)
0, else,

where α is a positive constant. Usually we choose α = 6.25 but in case of

Shepard interpolation α = 2. The size of h defines a set of neighboring particles
around x. Let P (x, h) = {xi : i = 1, 2, . . . , m} be the set of m neighboring
points of x in a ball of radius h.
A basic idea in FPM is to construct approximate derivatives from point
values using a least squares approach [15, 16, 25, 27, 28]. This leads to some
obvious restrictions on the number and the location of points in P (x, h).
Since we consider general second order equations, all derivatives up to order
two (and the function value itself) should be constructable so that, in 3D, a
minimal requirement is that P (x) contains at least ten points.
To explain the least squares approach, consider the Taylor expansions of
ψ(xi ) around x
X ∂ψ |j| 1
ψ(xi ) = (xi − x)j1 (yi − y)j2 (zi − z)j3 + ei , (39)
∂xj1 ∂y j2 ∂z j3 j!
|j|≤2

for i = 1, . . . , m, where ei is the truncation error (we expand to second order

for simplicity – higher order expansions are, of course, possible). Denote the
coefficients
∂ψ ∂ψ ∂ψ ∂2ψ
a0 = ψ(x), a1 = , a2 = , a3 = , a4 = ,
∂x ∂y ∂z ∂x2
∂2ψ ∂2ψ ∂2ψ ∂2ψ ∂2ψ
a5 = , a6 = , a7 = , a8 = , a9 = .
∂x∂y ∂x∂z ∂y 2 ∂y∂z ∂z 2

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 17

To the m equations (39) we add the equations (36) and (37) which are re-
expressed as

Aa0 + B1 a1 + B2 a2 + B3 a3 + C(a4 + a7 + a9 ) = f
nx a1 + ny a2 + nz a3 = φ,

where nx , ny , nz are the spatial components of the unit normal vector n on

the boundary Γ .
Note that, for the Dirichlet boundary condition, we have only m + 1 equa-
tions, where we directly prescribe the boundary conditions on the boundary
particles.
Now, we have to solve m + 2 equations. For m + 2 > 10 this system is
over-determined with respect to the unknowns ai and can be written in matrix
form as
e = M a − b,
where
1 2
h1,1 h2,1 h1,1 h3,1 21 h22,1 h2,1 h3,1 21 h23,1
 
1 h1,1 h2,1 h3,1 2 h1,1
 .. .. .. .. .. .. .. .. .. .. 
. . . . . . . . . . 
M =1 ,
 1 2 1 2 1 2

 h1,m h2,m h3,m 2 h1,m h1,m h2,m h1,m h3,m 2 h2,m h2,m h3,m 2 h3,m  
A B 1 B2 B3 C 0 0 C 0 C 
0 n x ny nz 0 0 0 0 0 0
T T T
a = (a0 , a1 , . . . , a9 ) , b = (ψ1 , . . . , ψm , f, g) , e = (e1 , . . . , em , em+1 , em+2 )
and h1,i = xi − x, h2,i = yi − y, h3,i = zi − z.
The unknowns a are computed by minimizing a weighted error over the
neighboring points. Thus, we have to minimize the following quadratic form
m+2
X
J= wi e2i , (40)
i=1
 
∂ψ
where em+1 = (Aψ + B · ∇ψ + C∆ψ − f ), em+2 = ∂n −φ and the cor-
responding weights wm+1 = wm+2 = 1. The above equation (40) can be
re-expressed in the form

J = (M a − b)T W (M a − b)

with the diagonal matrix

W = diag (w1 , w2 , . . . , wm , 1, 1) .

The minimization of J with respect to a formally yields ( if M T W M is non-

singular)
a = (M T W M )−1 (M T W )b. (41)

preprint -- preprint -- preprint -- preprint -- preprint

 18 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari

In (41) the vector (M T W )b is explicitly given by

m
X m
X
(M T W )b = wi ψi + Af, wi h1,i ψi + B1 f + nx φ,
i=1 i=1
Xm m
X
wi h2,i ψi + B2 f + ny φ, wi h3,i ψi + B3 f + nz φ,
i=1 i=1
m m m
1 X X X
wi h21,i ψi + Cf, wi h1,i h2,i ψi , wi h1,i h3,i ψi ,
2 i=1 i=1 i=1
m m m
!T
1X X 1X
wi h22,i ψi + Cf, wi h2,i h3,i ψi , wi h23,i ψi + Cf .
2 i=1 i=1
2 i=1

Thus we obtain from equation (41) that

m
! m
!
X X
ψ = Q11 wi ψi + Af + Q12 wi h1,i ψi + B1 f + nx φ
i=1 i=1
m
! m
!
X X
+ Q13 wi h2,i ψi + B2 f + ny φ + Q14 wi h3,i ψi + B3 f + nz φ
i=1 i=1
m
! m
!
1X 2
X
+ Q15 wi h1,i ψi + Cf + Q16 wi h1,i h2,i ψi
2 i=1 i=1
m
! m
!
X 1X
+ Q17 wi h1,i h3,i ψi + Q18 wi h22,i ψi + Cf
i=1
2 i=1
m
! m
!
X 1X 2
+ Q19 wi h2,i h3,i ψi + + Q1,10 wi h3,i ψi + Cf ,
i=1
2 i=1

where Q11 , Q12 , . . . , Q1,10 is the first row of the matrix (M T W M )−1 . Rear-
ranging the terms, we have
m
X h21,i
ψ− wi Q11 + Q12 h1,i + Q13 h2,i + Q14 h3,i + Q15 +
i=1
2
!
h22,i h23,i
Q16 h1,i h2,i + Q17 h1,i h3,i + Q18 + Q19 h2,i h3,i + Q1,10 ψi =
2 2

(Q11 A + Q12 B1 + Q13 B2 + Q14 B3 + Q15 C + Q18 C + Q1,10 C) f +

(Q12 nx + Q13 ny + Q14 nz ) φ.

Hence, if x is one of the N particles, say xj and xji its neighbors of number
m(j), where xj is distinct from xji , then we have the following sparse system
of equations for the unknowns ψj , j = 1, . . . , N

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 19

m(j)
X h21,ji
ψj − w ji Q11 + Q12 h1,ji + Q13 h2,ji + Q14 h3,ji + Q15 +
i=1
2
!
h22,ji h23,i
Q16 h1,ji h2,ji + Q17 h1,ji h3,ji + Q18 + Q19 h2,ji h3,ji + Q1,10 ψji =
2 2

(Q11 A + Q12 B1 + Q13 B2 + Q14 B3 + Q15 C + Q18 C + Q1,10 C) f +

(Q12 nx + Q13 ny + Q14 nz ) φ.

and in matrix form

LΨ = R. (42)
We have solved the above sparse system (42) using iterative methods like
Gauss-Seidel and SOR. Since the values of the velocities and the pressure
from time step n can be taken as initial values in the iterative methods for
the time step n + 1, only very few iteration steps are required if the variation
in time is not too large.

3.3 Numerical Tests

In the following we consider three examples in the two dimensional case. The
test cases are given in dimensionless form but can be interpreted in SI-units.

Rayleigh-Taylor instability

In order to test our numerical scheme we first consider the Rayleigh-Taylor

instability computed by the meshfree method SPH [4]. The authors have com-
pared the SPH results with those from the VOF method. In this test case the
heavy fluid lies above the light fluid. The computational domain is the rect-
angle [0, 1] × [0, 2]. The densities of two fluids are 1.8 and 1. The dynamical
viscosity of both fluids is µ = 0.4286 × 10−3. The gravity acts downwards
√ with
g = 1. The Reynolds number is Re = ρU L/µ = 420 with U = gL, L = 1.
Initially, we have considered 5147 particles with smoothing length h = 0.6.
The initial interface is given by 1 − 0.15 sin(2πx). The heavy particles (stars)
lie on and above this interface and the rest are light ones (dots). No-slip
boundary conditions are applied at the solid walls.
The time evolutions of the simulation are plotted in Fig. 2. Qualitatively,
the results seem to be better than those obtained by the SPH method in [4].

Breaking dam problem

This is a classical and simple test case to validate numerical schemes for the
simulation of free surface flows. In [19] experimental data are given and several
authors have reported their numerical results [13, 18, 26, 27].

preprint -- preprint -- preprint -- preprint -- preprint

 20 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari
2 2 2 2

1.8 1.8 1.8 1.8

1.6 1.6 1.6 1.6

1.4 1.4 1.4 1.4

1.2 1.2 1.2 1.2

1 1 1 1

0.8 0.8 0.8 0.8

0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Fig. 2. Rayleigh-Taylor instability from left to right at time t = 0, 3, 4, 5

The computational domain is a rectangle with the size of [0, 0.6] × [0, 0.3].
Consider a rectangular column of water with a width of a = 0.1 and a height
of 0.2. The rest of the domain is filled with air. In Fig. 3 the star particles
represent the water and the dot particles represent the air. No-slip boundary
conditions are applied at all boundaries.
Initially, 5624 particles are distributed with smoothing length h = 0.015.
The gravity with g = 9.81 acts downwards. The density of air is 1 and the
viscosity is 1.81 × 10.−5 . For water the density is 1000 and the viscosity is
1.005 × 10−3 .
In Fig. 3 we have plotted the simulation results at different times. Since
the density ratio of the fluids is very high, we have to smooth the density and
viscosity three times near the interface in order to avoid numerical instabilities.
If the density ratio is 100:1, it is enough to smooth the density at the interface
only once. For the viscosity, also three smoothing steps are used, but the
number of smoothing steps is less significant in this case.
In Fig. 4 the positions of the leading fluid front versus time as well
as the heights of the water column versus time are compared with exper-
imental results provided by [19]. The front position is computed as the
maximum distance among the water particle from the origin. Similarly, the
height is obtained from the maximum height of the water particles. The data
are plotted before the particles hit the right wall. The numerical computa-
tions are performed with different values of h. The results are plotted for
h = 0.04, 0.02, 0.01, 0.005 which correspond to the number of particles 400,
3117, 12593 and 49262 respectively. These figures show a good agreement
between the numerical and experimental results for small values of h.
The CPU time in a PC Pentium 4, 2.66GHz for different values of h are
presented in table 1. In all cases the time step ∆t = 0.002 has chosen and the
program was terminated at time t = 0.4.

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 21
0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15

0.1 0.1

0.05 0.05

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6

0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15

0.1 0.1

0.05 0.05

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6

Fig. 3. star: water, dot: air particles at time t = 0, 0.12, 0.24, 0.72.

8 1
Experiment Experiment
h=0.04 h=0.04
h=0.02 h=0.02
7 h=0.01 0.9 h=0.01
h=0.005 h=0.005

6 0.8

5 0.7

4 0.6

3 0.5

2 0.4

1 0.3

0 0.2
0 1 2 3 4 5 0 0.5 1 1.5 2 2.5 3 3.5

p
Fig. 4. Left: dimensionless front position x/a versus dimensionless time
p t 2g/a,
right: dimensionless column height x/(2a) versus dimensionless time t g/a.

Table 1.
h No of Particles CPU time
0.04 800 0 Min 31 Sec
0.02 3117 2 Min 10 Sec
0.01 12593 9 Min 48 Sec
0.005 49262 46 Min 14 Sec

Droplet splash

Finally, we consider a water droplet falling through air onto a water surface.
This problem was originally proposed in [20] and later reconsidered in [21] to
test numerical schemes for variable density flows. The computational domain

preprint -- preprint -- preprint -- preprint -- preprint

 22 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari

and the resolution is the same as in the test case of the Rayleigh-Taylor
instability. The particles inside the circle of radius 0.2 with center (0.5, 1.6) and
below the line y = 1 are considered as water particles and the rest air particles.
The other data are chosen as in the case of the breaking dam problem. These
results are consistent with the results shown in [20, 21].

2 2 2

1.8 1.8 1.8

1.6 1.6 1.6

1.4 1.4 1.4

1.2 1.2 1.2

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

2 2 2

1.8 1.8 1.8

1.6 1.6 1.6

1.4 1.4 1.4

1.2 1.2 1.2

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Fig. 5. Falling water droplet through air onto water surface from left to right and
top to bottom at times t = 0.2, 0.32, 0.48, 0.64, 0.88, 1.0.

preprint -- preprint -- preprint -- preprint -- preprint

 Meshless methods for conservation laws 23

Acknowledgement

This work has been carried out in the project Particle Methods for Conser-
vation Systems NE 269/11-3 which is part of the DFG – Priority Research
Program Analysis and Numerics for Conservation Laws.

References
1. C. Chainais-Hillairet. Finite volume schemes for a nonlinear hyperbolic equa-
tion. Convergence towards the entropy solution and error estimate. M 2 AN ,
33:129–156, 1999.
2. A. Chorin. Numerical solution of the Navier-Stokes equations. J. Math. Com-
put., 22:745–762, 1968.
3. B. Cockburn, F. Coquel, and P. LeFloch. An error estimate for finite volume
methods for multidimensional conservation laws. Math. Comput., 63:77–103,
1994.
4. S. J. Cummins and M. Rudmann. An SPH projection method. J. Comput.
Phys., 152:284–607, 1999.
5. G. A. Dilts. Moving least squared particle hydrodynamics – i. consistency and
stability. Int. J. Numer. Mech. Eng., 44:1115–1155, 1999.
6. E. Godlewski and P.-A. Raviart. Numerical Approximation of Hyperbolic Sys-
tems of Conservation Laws. Springer, 1996.
7. D. Hietel, K. Steiner, and J. Struckmeier. A finite-volume particle method for
compressible flows. Math. Models Methods Appl. Sci., 10:1363–1382, 2000.
8. J.M. Melenk I. Babushka, I. The partition of unity method. International
Journal of Numerical Methods in Engineering, 40:727–758, 1997.
9. O. Iliev and Tiwari S. A generalized (meshfree) finite difference discretization for
elliptic interface problems. In I. Dimov, I. Lirkov, S. Margenov, and Z. Zlatev,
editors, Numerical Methods and Applications, Lecture notes in Computer Sci-
ences. Springer, to appear.
10. Monaghan J. J. Smoothed particle hydrodynamics 1990. Annu. Rev. Astron.
Astrop, 30:543–574, 1992.
11. M. Junk and J. Struckmeier. Consistency analysis for mesh-free methods for
conservation laws. Mitt. Ges. Angew. Math. Mech., 24:99–126, 2001.
12. R. Keck. The finite volume particle method. PhD thesis, Universität Kaiser-
slautern, 2003.
13. F. J. Kelecy and Pletcher R. H. The development of free surface capturing
approach for multidimensional free surface flows in closed containers. J. Comput.
Phys., 138:939, 1997.
14. D. Kröner. Numerical Schemes for Conservation Laws. Wiley Teubner, 1997.
15. J. Kuhnert. General smoothed particle hydrodynamics. PhD thesis, Universität
Kaiserslautern, 1999.
16. J. Kuhnert. An upwind finite pointset method for compressible Euler and
Navier-Stokes equations. In M. Griebel and M. A. Schweitzer, editors, Mesh-
free methods for Partial Differential Equations, volume 26 of Lecture Notes in
Computational Science and Engineering. Springer, 2002.

preprint -- preprint -- preprint -- preprint -- preprint

 24 D. Hietel, M. Junk, J. Kuhnert, and S. Tiwari

17. T. Liszka and J. Orkisz. The finite difference method on arbitrary irregular grid
and its application in applied mechanics. Computers & Structures, 11:83–95,
1980.
18. V. Maronnier, Picasso. M., and J. Rappaz. Numerical simulation of free surface
flows. J. Comput. Phys., 155:439, 1999.
19. J. C. Martin and Moyce. M. J. An experimental study of the collapse of liquid
columns on a liquid horizontal plate. Phil. Trans. Roy. Soc. London Ser. A,
244:312, 1952.
20. E. G. Puckett, A. S. Almgren, J. B. Bell, D. L. Marcus, and W. J. Rider. A
high-order projection method for tracking fluid interfaces in variable density
incompressible flows. J. Comput. Phys., 130:269–282, 1997.
21. T. Schneider, N. Botta, K. J. Geratz, and R. Klein. Extension of finite vol-
ume compressible flow solvers to multi-dimensional. variable density zero Mach
number flows. J. Comput. Phys., 155:248–286, 1999.
22. D. Shepard. A two-dimensional interpolation function for irregularly spaced
points. Proceedings of A.C.M National Conference, pages 517–524, 1968.
23. Y. Y. Lu T. Belytschko and L. Gu. Element-free Galerkin methods. Int. J.
Num. Methods in Engineering, 37:229–256, 1994.
24. D. Teleaga. Numerical studies of a finite-volume particle method for conserva-
tion laws. Master’s thesis, Universität Kaiserslautern, 2000.
25. S. Tiwari. A LSQ-SPH approach for compressible viscous flows. In
H. Freistuehler and G. Warnecke, editors, proceedings of HYP2000, volume 141.
Birkhäuser, 2001.
26. S. Tiwari and J. Kuhnert. Grid free method for solving Poisson equation. Tech-
nical report, Fraunhofer ITWM, Kaiserslautern, Germany, 2001.
27. S. Tiwari and J. Kuhnert. Finite pointset method based on the projection
method for simulations of the incompressible Navier-Stokes equations. In
M. Griebel and M. A. Schweitzer, editors, Meshfree methods for Partial Dif-
ferential Equations, volume 26 of Lecture Notes in Computational Science and
Engineering. Springer, 2002.
28. S. Tiwari and S. Manservisi. Modeling incompressible Navier-Stokes flows by
lsq-sph. Nepal Mathematical Sciences Report, 20, 2003.
29. J.-P. Vila. Convergence and error estimates in finite volume schemes for gen-
eral multidimensional scalar conservation laws. I. Explicit monotone schemes.
M 2 AN , 28:267–295, 1994.
30. Y.F. Zhang W.K. Liu, S. Jun. Reproducing kernel particle method. Int. J.
Num. Meth. in Fluids, 20:1081–1106, 1995.
31. Z. Yang. Efficient calculation of geometric parameters in the finite volume
particle method. Master’s thesis, Universität Kaiserslautern, 2001.

preprint -- preprint -- preprint -- preprint -- preprint