Professional Documents
Culture Documents
Vor-Splines Natural Element Method
Vor-Splines Natural Element Method
Int. J. Numer. Meth. Engng 2007; 00:1–28 Prepared using nmeauth.cls [Version: 2002/09/18 v2.02]
SUMMARY
The problem of generalizing the Natural Element Method in terms of higher-order consistency and
continuity is addressed here. It is done by means of the de Boor algorithm, the same employed to obtain
B-splines by linear combinations of linear interpolants in one dimension. By noting that any form of
natural neighbour interpolation can be considered as a suitable generalization of linear interpolation
to two and higher dimensions, the de Boor algorithm is extended to two or higher dimensions, thus
obtaining a new form of interpolation that can be used in a Galerkin framework to develop a new
class of meshless methods.
This new class of meshless methods closely resembles the isogeometric analysis developed by
[T.J.R. Hughes et al. Isogeometric analysis: CAD, finite elements, NURBS, exact geometry, and
mesh refinement. Computer Methods in Applied Mechanics and Engineering, 194:4135-4195, 2005].
However, unlike B-splines, the new class of interpolants does not rely on an underlying tensor-product
quadrilateral mesh. It is based upon the Delaunay triangulation of the cloud of knots and does not
require any regularity on the connectivity.
In addition, the new method conserves many of the attractive features of the Natural Element
Method, such as strict interpolation on the boundary, and thus direct imposition of essential boundary
conditions. After a theoretical description of the proposed method, some numerical examples are shown
to test its performance in the context of Linear Elastostatics. Copyright
c 2007 John Wiley & Sons,
Ltd.
key words: Meshless, Natural Element Method, Sibson interpolation, Laplace interpolation,
isogeometric analysis, de Boor’s algorithm, B-splines.
∗ Correspondence to: Elı́as Cueto. Mechanical Engineering Department. Edificio Betancourt. University of
Zaragoza. Marı́a de Luna, 5. E-50018 Zaragoza, Spain. e-mail: ecueto@unizar.es
Contract/grant sponsor: Spanish Ministry of Education and Science; contract/grant number: CICYT-DPI2005-
08727-C02-01
Contract/grant sponsor: European Union, sixth Framework Program; contract/grant number: DeSSOS project,
STREP project number 027252.
Contents
1 INTRODUCTION 2
7 NUMERICAL EXAMPLES 20
7.1 L2 -fit of a quadratic polynomial . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
7.2 L2 -fit of a cubic polynomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
7.3 Beam bending . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
7.4 Infinite plate with a hole . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
8 CONCLUSIONS 25
1. INTRODUCTION
Nowadays it is widely recognised that meshless methods have gained an utmost importance
for the numerical simulation of certain classes of problems. This is particularly noteworthy
in problems involving large deformations, evolving cracks, free surfaces and many others.
Although certain problems persist in most meshless methods (imposition of essential boundary
conditions or numerical integration, for instance) they constitute, after more than a decade
of research performed by the Computational Mechanics community, an appealing alternative
that is now being incorporated to commercial software.
The Natural Element Method (NEM) is nowadays a well-established technique in the
family of meshless methods that has attracted attention of many researchers. Both from the
theoretical point of view [5] [29] [25] [7] [8] [30] and its application to different problems [3]
[1] [21] [15], the Natural Element Method has demonstrated its success for a wide variety
of problems, mainly those involving large strains, free surfaces, etc. Some of its interesting
characteristics are the “exact” (i.e., within the order of the approximation) imposition of
essential boundary conditions and the exact numerical integration (achieved by the use of
Stabilized Conforming Nodal Integration, see [6] [14]).
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 3
In essence, the NEM is a Galerkin method based on the use of natural neighbour
interpolation (either Thiessen [27], Laplace [4] or Sibson [22] interpolation schemes) instead
of piece-wise polynomials, employed in the case of the FEM. The fact that, unlike most other
meshless methods, NEM shape functions are strictly interpolant, clearly simplifies the issue of
imposing essential boundary conditions, as mentioned before. In fact, many authors consider
natural neighbour interpolation as the most suitable generalization of linear interpolation to
two or higher dimensions, without an underlying quadrilateral or tensor-product mesh [11].
Precisely the fact that the NEM possesses linear consistency and C 0 continuity only is
perhaps on the basis of its limited popularity, if compared with other meshless methods, which
easily achieve higher-order consistency and even C ∞ continuity. Only one attempt has been
made to overcome this difficulty, up to our knowledge, by applying a quadratic consistency
and C 1 interpolant based on natural neighbours that, however, does not seem to posses any
further generalization [24]. This interpolant can be used, for instance, for solving fourth-order
partial differential equations such as those arising from the theory of Kirchhoff plate bending.
In this paper an attempt to overcome these NEM limitations is performed, by going back
to the foundations of B-splines and how by linear combinations of linear interpolants, higher-
order curves can be obtained. B-spline curves can be obtained by means of the so-called de
Boor’s algorithm [11]. For the surface case, tensor product B-spline surfaces were initially
proposed in [9]. An extensive review of this topic can be found in [11, Ch. 16]. Tensor product
B-spline surfaces are, however, very rigid. For instance, no tensor product surface can have
the connectivity of a double torus. This algorithm is here generalized, without the use of
tensor products, to higher dimensions. This is done by employing different natural neighbour
interpolation schemes.
Recently, the use of NURBS or B-splines as basis functions for a Finite Element-like
simulation has been studied [19]. The main objective is clear: B-splines (or, more properly,
NURBS) are the standard for CAD systems, and they reproduce the geometry of the domain
exactly, which is not the case in Finite Element models. The use of the same approximation
for both the construction of the geometry and the approximation of the essential field of
the problem obviously simplifies the burdensome mesh generation process. Other interesting
properties of NURBS like the so-called “variation diminishing” [11] property also apply. This
means that, unlike high-order polynomials, B-splines or NURBS do not show the well-known
Gibbs effect [13].
However, it remains unclear whether the use of B-splines-like finite elements leads to
remeshing problems when large distortions of the mesh occur. Tensor-product B-splines,
as mentioned before, are quite rigid. In addition, fulfilment of (inhomogeneous) essential
(Dirichlet) boundary conditions should be done typically in an approximate sense, or by
imposing them weakly. This same problem is common for many meshless methods.
In this paper we develop a new family of natural neighbour interpolation schemes that
show many interesting properties. First of all, they are based on Sibson [22], Laplace[26] or
Hiyoshi-Sugihara [17] interpolants, but higher-order approximation is achieved by means of the
de Boor’s algorithm. We show that these interpolants behave as a suitable generalization of
linear interpolation to higher dimensions. We thus obtain a family of B-spline-like surfaces that
do not rely on an underlying tensor-product mesh to be built. They share many interesting
properties with natural neighbour interpolants, as will be discussed hereafter. The control of
the degree of continuity and consistency is also possible, as will be seen, in a B-spline manner.
In addition, imposition of essential boundary conditions can be done by means of repeating
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
4 GONZÁLEZ, CUETO AND DOBLARÉ
the boundary knots employed to build the approximation spaces. These new function spaces
are used, in a Galerkin framework, to generalize the Natural Element Method to an arbitrary
order of consistency, thus obtaining a method that closely resembles the isoparametric finite
elements by Hughes [19], but without any tensor-product underlying mesh.
In section 2 we describe the most popular natural neighbour-based interpolants and show
how they can be interpreted as a suitable generalization of linear interpolation to higher
dimensions. In section 3 we describe the de Boor’s algorithm to construct B-splines and
develop a similar one based on natural neighbour interpolants. Some classical examples are
then analyzed to show the performance of the proposed method.
There are various interpolation schemes based on the use of natural neighbours. Historically,
the first one is due to Thiessen [27] and possesses constant consistency only. It has been
employed to approximate pressures in Hellinger-Reissner-like variational principles [16],
although it is obviously not suitable for second-order partial differential equations. The most
popular natural neighbour interpolant is probably the one due to Sibson [22], but other
interpolants have been recently developed, such as the so-called Laplace Interpolant or the
Hiyoshi-Sugihara interpolant [17]. In this section we review some of them. Others, such as a
recently-developed MLS scheme based on the use of natural neighbour can be consulted in [2]
and references therein.
Prior to this review we introduce some basic geometrical entities that are needed for further
developments.
† Even in three-dimensional spaces, it is common to refer to the Delaunay tetrahedralisation with the word
triangulation in the vast majority of the literature
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 5
sharing a facet of their Voronoi cell are called natural neighbours and hence the name of the
technique.
Equivalently, the second-order Voronoi diagram of the cloud is defined as
TIJ = {x ∈ Rd : d(x, xI ) < d(x, xJ ) < d(x, xK ) ∀ J 6= I 6= K}. (2)
Based on these definitions, different natural neighbour interpolation schemes have been
proposed. We review some of the most popular.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
6 GONZÁLEZ, CUETO AND DOBLARÉ
6 4
a
e d
1 x f 3
b c
7
2
Sibson’s interpolation scheme possesses the usual reproducing properties for this class of
problems, that is, it verifies the partition of unity property (constant consistency) and the
linear consistency property. Thus, it is suitable for the solution of second-order PDE. Other
interesting properties such as the Kronecker delta property [25] and linear interpolation on the
boundary [7] [30] are also verified by the NEM.
where sJ represents the length of the Voronoi segment associated to node J and n represents
the number of natural neighbours of the point under consideration, x. n = 4 in this example.
hJ represents the distance from the evaluation point to the edge of the voronoi cell, see Fig. 3.
Derivatives of the Laplace shape function are not defined along the edges of the Delaunay
triangles that lie inside its support (see [26]), but the approximation is C 1 at any other point.
This same interpolation scheme is used in the Meshless Finite Element Method (MFEM)
context [20], but in this case, the space is divided into polyhedral elements and Laplace
interpolation is constructed within each polyhedron. Since the same properties mentioned
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 7
Z Z
Y Y
X X
(a) (b)
for Sibson interpolation (i.e., linear completeness, exact interpolation on the nodes and linear
interpolation along the boundaries) hold, a conforming method is achieved.
A comparison of Sibson and Laplace shape functions for the central node within a lattice of
5 × 5 nodes is shown in Fig. 4.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
8 GONZÁLEZ, CUETO AND DOBLARÉ
The Laguerre-Voronoi diagram will thus be defined as the decomposition of the space into
regions of the form:
VI = {x ∈ Rd : πI (x) < πJ (x) ∀ J 6= I}. (9)
The Laguerre-Voronoi diagrams share many of they properties with standard Voronoi
diagrams, but not all. For instance:
• The Laguerre-Voronoi cell associated to a given node does not necessarily exist.
• If it exists, it is a convex polytope.
• The Laguerre-Voronoi facet is perpendicular to the line joining two neihgbouring nodes.
See [17] and references therein for detailed proofs of these assertions. The key aspect in defining
their generalized system of coordinates is considering a set of nodes equipped with increasing
ball radii. Note that the Laguerre-Voronoi diagram for ̺I = 0, ∀I, coincides with the standard
Voronoi diagram. If, starting from zero, all ball radii are increased, except the one associated
to the evaluation point, its Laguerre-Voronoi cell will consequently shrink until it disappears
for sufficiently large ball radius.
x2 x1
r1
h1
u1
xn xI
Figure 5. Definition of the shrinking cells in Laguerre-Voronoi diagrams. The dashed-line cell represents
the shrunken original cell (continuous line) when the weight ̺ 6= 0.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 9
where the super-index k represents how many times the obtained shape function can be
continuously differentiated [17]. In the above definition, the ψI0 (x)’s are well defined only
when x does not coincide with any of the nodes xI . In that case,
ψI0 (xJ ) = δIJ . (13)
The coordinates defined above generalize the sequence Laplace → Sibson and provide
an interesting means of interpolate data with smoother shape functions. Some remarkable
properties of these so-called “generalized” coordinates are (see [17]):
• Exactness (or, equivalently, the so-called Kronecker’s delta property):
ψIk (xJ ) = δIJ
• C k continuity if x lies on the Delaunay sphere of the data sites (nodes). Continuity other
than C 0 at the nodes themselves has not been proved analytically, but computational
experiments performed in [17] and also by the authors suggest its fulfillment. C ∞
continuity is obtained elsewhere.
In Fig. 6 the shape functions with C 1 and C 2 continuity are shown for a set of nodes placed
on a regular lattice.
This set of interpolating functions has linear precision, at most. This is easily proven if we
consider the limiting case of only three neighbouring nodes around a given evaluation point.
It is impossible to define a quadratic surface given only three data.
So, the problem of generalizing natural neighbour-based interpolants includes the elevation
of the degree of consistency. In this work we propose a novel technique based on the application
of the well-know de Boor’s algorihtm [11]. This algorithm is the same used in the context of B-
spline curves to generate higher-order curves from linear combinations of linear interpolations.
The de Casteljau algorithm for Bézier curves states that such curves can be obtained by
successive application of linear interpolation [11], i.e., given some points b0 , b1 , . . . , bn ∈ E3
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
10 GONZÁLEZ, CUETO AND DOBLARÉ
Z Z
X Y X Y
(a) (b)
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 11
b[u1,u2]
b[u,u1]
b[u,u2]
b[u,u]
b[u0,u1] b[u2,u3]
u0 u1 u u2 u3
s
s U 11
U 13
s
U 12
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
12 GONZÁLEZ, CUETO AND DOBLARÉ
where φI (u) represent the natural neighbour coordinates of point u with respect to knot I
and ϕrI (u) represent the natural neighbour coordinates of point u with respect to knot I, but
computed over an interval Uir , i.e., by eliminating r − 1 natural neighbours of the interval. The
notation used is shown in Fig. 8.
b[u1,u2]
b[u,u1]
b[u,u2]
b[u,u]
b[u0,u1] b[u2,u3]
f1(U11) f2(U11)
u0 u1 u u2 u3
j0(U12) j2(U12)
u0 u u2 u3
j1(U22) j3(U22)
u0 u1 u u3
Figure 8. Schematic representation of the de Boor’s algorithm employing natural neighbours. Between
parentheses, the domain of each function.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 13
4.1. Definition
The de Boor’s algorithm thus presented can be extended to higher-dimensional cases as
follows. In the following development we employ Sibson coordinates, although the proposed
algorithm can also be applied to Laplace and Hiyoshi-Sugihara interpolants, as will be shown
later. Consider again, for simplicity, a set of nodes N = {x1 , x2 , . . . , xM } ⊂ R2 and a
quadratic surface (the extension to three or higher dimensions and higher-order surfaces is
straightforward). From now on, we will work in non-parametric form, since it is extremely
hard to find the two-dimensional counterparts of the intervals Uir for irregularly scattered
sites. Then, we define a new class of surfaces constructed in the way:
I
n X
X n
s(x) = NIJ (x)dIJ , with dIJ = dJI (17)
I=1 J=1
and dIJ represent the control points in B-spline terminology (i.e., the degrees of freedom).
φI (x) represents the natural neighbour (Sibson) coordinate of the point x with respect to site
I. Functions ϕIJ (x) represent the natural neighbour coordinates of point x with respect to site
J, in the original cloud of points, but without the I-th site (see Fig. 9), in the sense described
by the previous section. Finally, nI is the number of natural neighbours of the point x when
we eliminate the site I, similarly to the de Boor’s algorithm. Note that the number of degrees
of freedom of the proposed approximation is much less than M 2 /2, since the sums in Eq. (17)
extend only over natural neighbors of each node.
The typical shape of the functions NIJ described before is shown in Fig. 10 for a general set
of irregularly distributed sites.
Proof The proof is straightforward, since the shape functions NIJ are the product of two
Sibson interpolants, always positive.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
14 GONZÁLEZ, CUETO AND DOBLARÉ
L L
J J
I
x x
K K
M M
N N
(a) (b)
Figure 9. Schematic representation of the proposed algorithm. (a) Set of sites {I, . . . , N }. We consider
an evaluation point x, whose neighbours are depicted as filled circles. The support of the function φI
is highlighted. (b) After eliminating site I, the support of function ϕIJ is highlighted. Note the new
set of neighbouring sites, {J, K, L, M }.
X Y
0.3
0.25
0.2
0.15
Z
0.1
0.05
0
1 1
2 2
3 3
Y X
4 4
5 5
Figure 10. Shape of a typical function NIJ for a set of irregularly distributed sites.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 15
Proof By applying the definition of the new basis functions, given by Eq. (18) we have
I I I
X n
n X X n
n X n
X n
X
NIJ (x) = φI (x)ϕIJ (x) = φI (x) ϕIJ (x).
I=1 J=1 I=1 J=1 I=1 J=1
and, by using the well-known partition of unity property of standard Sibson coordinates,
I I
n
X X n
n X n
X
ϕIJ (x) =1 ⇒ NIJ (x) = φI (x) = 1. (20)
J=1 I=1 J=1 I=1
and, finally,
n
1 k 1 kX
s(x) = x + x φI (x) = xk .
2 2
I=1
Following a similar approach, the quadratic completeness can be deduced and, by induction,
the p-th order precision.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
16 GONZÁLEZ, CUETO AND DOBLARÉ
Proposition 4.4. The basis functions NIJ (x) span the space of quadratic polynomials.
for each quadratic product xk xl of the coordinates k, l, where xkI represents the k-coordinate
of the knot I neighbouring the generic evaluation point x. Thus,
I I
X n
n X n
X n
X
NIJ (x)xkI xlJ = φI (x)xkI ϕIJ (x)xlJ .
I=1 J=1 I=1 J=1
We obtain finally
s(x) = xk xl ,
thus obtaining, together with Props. 4.2 and 4.3, the desired proof for the completeness of the
quadratic basis.
The interested reader can consult the book by Farin [11] for the details on a similar proof
for higher-order consistency. Note that the approximation employed in Eq. (21) should be
substituted for the appropriate one with the desired consistency, see Eq. (17).
Proposition 4.5. The proposed surfaces are C p−1 , where p stands for the order of the surface,
except at the nodes, in the direction joining two neighbouring nodes, where they are C 0 .
Proof The derivatives of the proposed surfaces (again we refer to the quadratic case) can be
obtained straightforwardly by applying the chain’s rule:
I
n n
∂s(x) ∂ XX
= NIJ (x)dIJ =
∂xk ∂xk
I=1 J=1
I
n
n X
∂ϕIJ (x)
X ∂φI (x) I
= dIJ ϕ J (x) + φ I (x) , k = 1, 2, . . . , d. (22)
∂xk ∂xk
I=1 J=1
Thus, the only points where Eq. (22) is not well defined are precisely the data sites, the only
points where the functions φI are not continuously differentiable. This is in perfect agreement
with the B-splines case, that match smoothly at the data sites, despite having discontinuous
derivatives there.
For the proof of the smoothness of the surfaces, we refer again to the quadratic case, and
thus we limit ourselves to the proof of the C 1 continuity, in the absence of repeated knots (for
the case of repeated knots, see Prop. 4.6).
To demonstrate the smoothness of the proposed surfaces, it is clear that the pair NIJ + NJI
should be continuously differentiable. Equivalently,
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 17
or, equivalently,
φ′I (I)ϕIJ (I) + ϕ′I ′
J (I) + φJ (I).
The term ϕ′IJ (I) is continuous, since it represents the derivative of the Sibson coordinates
with respect to node J once node I has been eliminated from the cloud, evaluated at the
original position of node I and along any direction. This also holds for higher-consistency
cases. However, the other two terms are not continuous, and their value depends on the choice
of a particular direction. In particular, for the joint surface to be continuously differentiable,
the sum
φ′I (I)ϕIJ (I) + φ′J (I) (23)
should take a value independent of the direction from which we approach the site I. From [10]
it is well-known that when an evaluation point approaches one of the data points on a straight
line through one of its neighbours, the new Sibson coordinate starts from null value growing
linearly, thus with C 0 continuity. The term in Eq. (23) is, in general, not continuous, when
differentiated in the direction of a neighbouring data site.
The x- and y-derivatives of the function pairs NIJ + NJI are shown in Fig. 11 and 12. A
single function NIJ is depicted in Fig. 10. Note that a single function NIJ is clearly not C 1 .
Note also that a high gradient exists in this particular case, but not a discontinuity. In fact,
the closer sites I and J are located, the higher is the gradient in the derivative.
If, on the contrary, the derivative is computed along the line joining two neighbouring sites,
the pair NIJ + NJI value takes the form depicted in Fig. 13(a), with the derivative shown in
Fig. 13(b).
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
18 GONZÁLEZ, CUETO AND DOBLARÉ
X Y
Z
3 3
4 4
Y X
5 5
6 6
X Y
0
Z
-0.5
3 3
4 4
Y X
5 5
6 6
Since φI (xI )ϕII (xI ) = 1 the interpolation at the data site follows.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 19
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
1 1.5 2 2.5 3 3.5 4 4.5 5
(a)
0.08
0.06
0.04
0.02
−0.02
−0.04
−0.06
−0.08
−0.1
1 1.5 2 2.5 3 3.5 4 4.5 5
(b)
Figure 13. (a) NIJ + NJI value along the line joining two neighbouring sites. (b) Derivative of the
same expression. Note the small jump in the derivatives near x ≃ 2.8.
By repeating the knots on the boundary of the convex hull of the data sites, for instance,
one can make the surface to be (piecewise) quadratically interpolant along the boundary. The
resulting function NII for a site on the boundary is depicted in Fig. 14.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
20 GONZÁLEZ, CUETO AND DOBLARÉ
X Y
0.5
Z
0
1 1
1.5 1.5
2 2
2.5 2.5
Y X
3 3
3.5 3.5
4 4
Figure 14. Interpolating function NII (x) obtained by repeating knots in the boundary of the convex
hull of the data sites. A discontinuity on the derivative appears at the site location.
This last property has again a tremendous importance when using this kind of approximation
in the context of Galerkin procedures, as in the Finite Element method. The use of repeated
knots ensures interpolation (without loss of consistency of the approximation) and thus an easy
imposition of essential boundary conditions and consistent interpolation along the boundary
by simply fixing the value of the approximation at the node (data site).
On the other hand, there is one important property of spline curves that is lost in this
new kind of surfaces, which is the knot insertion. If a new knot is inserted, thus obtaining a
refined control mesh, the obtained surface will not be the same. This phenomenon occurs since
Sibson’s interpolation has not the idempotence property [12].
Remark 5. The same guidelines can be followed in order to prove C n−r continuity in a general
case.
Remark 6. Extension to three or higher dimensions is also straightforward.
7. NUMERICAL EXAMPLES
In this section we perform some numerical examples concerning the ability of the proposed
surfaces to approximate data over irregularly scattered knots and also in its application within
linear Elastostatics. All examples employ quadratic approximation only, but, as mentioned
before, extension to higher degrees of consistency is straightforward.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 21
The obtained surface is shown in Fig. 15(a). As expected, the error obtained vanish
identically everywhere in the domain, see Fig. 15(b).
Z ERROR Z
1E-12
7.5E-13
5E-13
2.5E-13
X Y X Y
0
-2.5E-13
-5E-13
4 -7.5E-13
-1E-12
0.01
2 0
V7
-0.01
1 1
2 2
1
Y X
3 3
0
4 4
1 1
1.5 1.5
2 2
2.5 2.5
Y X
3 3
3.5 3.5
4 4
(a) (b)
Figure 15. (a) Surface obtained by a least squares fitting of the polynomial (x − 2.5)2 + (y − 2.5)2 = 0.
Knots are depicted as red spheres. Boundary knots are repeated. (b) Error obtained in the L2 -fit.
The same L2 -fit has been performed with the derivative of the mentioned surface. With
the coefficients obtained before, the approximation of the derivative of the surface through
the derivatives of the NIJ basis functions is achieved. The error in the approximation of the
derivative is again within machine precision (less than 10−14 ), see Fig. 16.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
22 GONZÁLEZ, CUETO AND DOBLARÉ
Z Z
X Y X Y
3
4E-14
0
error
Z
2E-14
-1 0
1 1
1.5
-2
2 2
Y 2.5
3 3 X
3.5
-3 4 4
1 1
1.5 1.5
2 2
2.5 2.5
Y 3 3 X
3.5 3.5
4 4
(a) (b)
Figure 16. (a) Approximation of the x-derivative of the surface (x − 2.5)2 + (y − 2.5)2 = 0. (b) Error
obtained in the L2 -fit of the derivatives.
To test the performance of the proposed formulations, we compared the results of the
proposed method with quadratic consistency and the theoretical displacements at the beam
tip [28]. The analytical solution is given by:
P y(L − x)
σx = − (24)
Iz
σy = 0 (25)
P D2
σxy = − y2 (26)
2I 4
and the displacements are given by
Py h D2 i
ux (x, y) = − ′ (6L − 3x)x + (2 + ν ′ ) y 2 − (27)
6E Iz 4
2
P h ′ 2 ′ D x 2
i
uy (x, y) = 3ν y (L − x) + (4 + 5ν ) + (3L − x)x (28)
6E ′ Iz 4
where Iz represents the moment of inertia of the beam, given by Iz = D3 /12. For plane strain
the material parameters are defined as
E
E′ = (29)
(1 − ν 2 )
ν
ν′ = (30)
(1 − ν)
In this case, L = 4.0 and D = 1.0. Young’s modulus was fixed to 1.0 and Poisson’s ratio to
0.25. The discretized beam is shown in Fig. 20.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 23
X Y
60
40
Z
20
1 1
2 2
Y3 3X
4 4
5 5
Figure 17. Surface obtained by L2 -fit of the cubic polynomial x3 − 2x2 + y 2 − 2y + 1 = 0, approximated
by a quadratic surface with 100 knots, irregularly distributed.
For numerical integration purposes, in all the examples shown, a regular background grid of
50 × 50 cells was employed. Each cell was splitted into two triangles and standard three-point
quadrature rules were employed. The objective of this technique is to avoid the numerical
integration errors found in all meshless methods due to the non-polynomial character of the
approximation. The use of the Delaunay triangles as integration cells was found to clearly
insufficient in order to obtain good convergence results.
The convergence results are shown in Fig. 21. Note that the convergence ratio is higher than
a priori expected. This is due to the quadratic form of the solution in vertical direction, which
is reproduced analytically by the quadratic approximation employed. Nevertheless, it indicates
the good approximation properties of the proposed method.
a hr a a3 i
u1 (r, θ) = (κ + 1) cos θ + 2 ((1 + κ) cos θ + cos 3θ) − 2 3 cos 3θ (31)
8µ a r r
a r
h a a3 i
u2 (r, θ) = (κ − 3) sin θ + 2 ((1 − κ) sin θ + sin 3θ) − 2 3 sin 3θ (32)
8µ a r r
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
24 GONZÁLEZ, CUETO AND DOBLARÉ
Log||error||L∞ -1.2
-1.4
-1.6
-1.8
5 6 7 8 9 10
Number of knots1/2
Figure 18. Convergence of the error in L∞ norm for the cubic polynomial fitted by quadratic surfaces.
P
D
x
Material parameters were Young’s modulus E=1.0 and Poisson coefficient ν = 0.25. µ
represents the shear modulus and κ is the Kolosov constant, defined as
κ = 3 − 4ν (33)
3−ν
κ= (34)
1+ν
respectively, for plane strain and plane stress.
Applying symmetry conditions, only one quarter of the plate was modelled, and exact
tractions were applied at the boundary of the model. The geometry of the model is shown
in Fig. 22. The plate was discretized with quadratic consistency approximants and, again,
repeated nodes were employed along the essential boundary. The position of the knots
(following the standard B-spline notation, note that they are not “nodes” in the Finite Element
sense) is shown in Fig. 23.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 25
0.5 0.5
0 0
−0.5 −0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
(a) (b)
0.5 0.5
0 0
−0.5 −0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
(c) (d)
Figure 20. Discretization of the beam problem: (a) 323 d.o.f. (b) 547 d.o.f. (c) 824 d.o.f. and (d) 942
d.o.f. Note that only the knots (not the nodes or “control points”) have been represented.
-2.6
-2.7 1
-2.8 3.40
log||Error||L2
-2.9
-3 1
3.72
-3.1
-3.2 1
3.75
-3.3
-3.4
15 20 25 30 35
1/2
(Number of degrees of freedom)
Figure 21. Convergence of the results in L2 norm for the beam problem.
The obtained convergence rate (R = 2.96, see Fig. 24) is very close to the theoretical one,
for quadratic consistency.
8. CONCLUSIONS
In this paper we have proposed a new scheme to generate high-order consistency natural
neighbour approximations. The so called de Boor’s algorithm, which is precisely the same
employed to construct B-spline curves in Computational Geometry, has been employed.
This method is able to generate arbitrary order consistency approximations, conserving
many of the interesting properties of the Natural Element Method, notably the interpolant
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
26 GONZÁLEZ, CUETO AND DOBLARÉ
x2
s0=1
5
P
r
q
x1
a=1
Figure 22. Geometry of the problem of an infinite plate with a hole under traction.
5 5
4.5 4.5
4 4
3.5 3.5
3 3
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
0 1 2 3 4 5 0 1 2 3 4 5
(a) (b)
5 5
4.5 4.5
4 4
3.5 3.5
3 3
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
0 1 2 3 4 5 0 1 2 3 4 5
(c) (d)
Figure 23. Discretization of the plate with a hole problem: (a) 36 d.o.f. (b) 100 d.o.f. (c) 200 d.o.f. and
(d) 322 d.o.f. Note that only the knots (not the nodes or “control points”) have been represented.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
HIGHER-ORDER NATURAL ELEMENT METHODS 27
-1.2
-1.4
log ||error||L2
-1.6
-1.8
1
-2 2.96
-2.2
4 6 8 10 12 14 16 18 20
1/2
(degrees of freedom)
Figure 24. Convergence of the results in L2 norm for the plate with a hole problem.
character along the boundary. This is achieved, as in B-spline curves, by repeating the
boundary knots defining the surface. Conversely to the B-spline case, continuity is not improved
if standard Sibson interpolants are used, but this can be solved by employing Hiyoshi-Sugihara
interpolants, which are C 1 -continuous at least, but only linearly complete.
The proposed method resembles the isogeometric Finite Element Method by T. Hughes,
in the sense that a new approximation method has been proposed that shares many of its
properties with the B-spline curves. However, the obtention of B-spline surfaces is quite rigid,
since they are constructed by tensor-product patches. In the proposed method, there is no
need of a tensor-product structure, and hence we would speak about an isogeometric meshless
method.
The other objective of the method, namely the generation of arbitrary order consistency
in the context of the NEM, has been achieved. This opens the possibility of p-like adaptive
simulations with NE methods. Obtaining smooth approximations with simpler structure than
that of Hiyoshi and Sugihara [17] is now part of the authors’ current effort of research.
ACKNOWLEDGEMENT
The authors gratefully acknowledge the comments and suggestions made by prof. G. Farin, whose
“splines over iterated Voronoi diagrams” inspired this work. The support to the authors provided
by the Spanish Ministry of Education and Science and the VI Framework Program of the European
Union is also acknowledged.
REFERENCES
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls
28 GONZÁLEZ, CUETO AND DOBLARÉ
2. I. Alfaro, J. Yvonnet, F. Chinesta, and E. Cueto. A study on the performance of Natural Neighbour-based
Galerkin Methods. International Journal for Numerical Methods in Engineering, accepted for publication,
2006.
3. I. Alfaro, J. Yvonnet, E. Cueto, F. Chinesta, and M. Doblaré. Meshless methods with application to metal
forming. Computer Methods in Applied Mechanics and Engineering, 195(48-49):6661–6675, 2006.
4. V. V. Belikov, V. D. Ivanov, V. K. Kontorovich, S. A. Korytnik, and A. Yu. Semenov. The non-Sibsonian
interpolation: A new method of interpolation of the values of a function on an arbitrary set of points.
Computational Mathematics and Mathematical Physics, 37(1):9–15, 1997.
5. J. Braun and M. Sambridge. A numerical method for solving partial differential equations on highly
irregular evolving grids. Nature, 376:655–660, 1995.
6. Jiun-Shyan Chen, Cheng-Tang Wu, Sangpil Yoon, and Yang You. A stabilized conforming nodal
integration for galerkin mesh-free methods. International Journal for Numerical Methods in Engineering,
50:435–466, 2001.
7. E. Cueto, M. Doblaré, and L. Gracia. Imposing essential boundary conditions in the Natural
Element Method by means of density-scaled α-shapes. International Journal for Numerical Methods
in Engineering, 49-4:519–546, 2000.
8. E. Cueto, N. Sukumar, B. Calvo, M. A. Martı́nez, J. Cegoñino, and M. Doblaré. Overview and recent
advances in Natural Neighbour Galerkin methods. Archives of Computational Methods in Engineering,
10(4):307–384, 2003.
9. C. de Boor. Bicubic Spline Interpolation. J. Math and Physics, 41:212–218, 1962.
10. G. Farin. Surfaces over Dirichlet Tesselations. Computer Aided Geometric Design, 7(1-4):281–292, 1990.
11. G. Farin. Curves and surfaces for CAGD. Morgan Kaufmann, 2002.
12. G. Farin. Splines over iterated Voronoi diagrams. Technical report, Arizona State University, 2003.
13. J. W. Gibbs. Fourier Series. Nature, 59, 1898.
14. D. Gonzalez, E. Cueto, M. A. Martinez, and M. Doblare. Numerical integration in Natural Neighbour
Galerkin methods. International Journal for Numerical Methods in Engineering, 60(12):2077–2104, 2004.
15. D. González, E. Cueto, F. Chinesta, and M. Doblaré. A natural element updated Lagrangian strategy for
free-surface fluid dynamics. Journal of Computational Physics, 223(1):127–150, 2007.
16. D. González, E. Cueto, and M. Doblaré. Volumetric locking in Natural Neighbour Galerkin methods.
International Journal for Numerical Methods in Engineering, 61(4):611–632, 2004.
17. H. Hiyoshi. Study on interpolation based on Voronoi diagrams. PhD thesis, Tokyo University, 2000.
18. H. Hiyoshi and K. Sugihara. Two generalizations of an interpolant based on Voronoi diagrams.
International Journal of Shape Modeling, 5(2):219–231, 1999.
19. T. J. R. Hughes, J. A. Cottrell, and Y. Bazilevs. Isogeometric analysis: CAD, finite elements, NURBS,
exact geometry and mesh refinement. Computer Methods in Applied Mechanics and Engineering,
194:4135–4195, 2005.
20. S. R. Idelsohn, E. Oñate, N. Calvo, and F. del Pin. The meshless finite element method. International
Journal for Numerical Methods in Engineering, 58:893–912, 2003.
21. L. Illoul, J. Yvonnet, F. Chinesta, and S. Clenet. Application of the natural-element method to model
moving electromagnetic devices. IEEE TRANSACTIONS ON MAGNETICS, 42(4):727–730, 2006.
22. R. Sibson. A Vector Identity for the Dirichlet Tesselation. Mathematical Proceedings of the Cambridge
Philosophical Society, 87:151–155, 1980.
23. R. Sibson. A brief description of natural neighbour interpolation. In Interpreting Multivariate Data. V.
Barnett (Editor), pages 21–36. John Wiley, 1981.
24. N. Sukumar and B. Moran. C 1 Natural Neighbour Interpolant for Partial Differential Equations.
Numerical Methods for Partial Differential Equations, 15(4):417–447, 1999.
25. N. Sukumar, B. Moran, and T. Belytschko. The Natural Element Method in Solid Mechanics.
International Journal for Numerical Methods in Engineering, 43(5):839–887, 1998.
26. N. Sukumar, B. Moran, A. Yu Semenov, and V. V. Belikov. Natural Neighbor Galerkin Methods.
International Journal for Numerical Methods in Engineering, 50(1):1–27, 2001.
27. A. H. Thiessen. Precipitation averages for large areas. Monthly Weather Report, 39:1082–1084, 1911.
28. S. Timoshenko and J. N. Goodier. Teorı́a de la Elasticidad. Editorial Urmo, Spain, 1972.
29. L. Traversoni. Natural Neighbour Finite Elements. In Intl. Conference on Hydraulic Engineering
Software. Hydrosoft Proceedings, pages 291–297. Computational Mechanics publications, 1994.
30. J. Yvonnet, D. Ryckelynck, P. Lorong, and F. Chinesta. A new extension of the Natural Element method
for non-convex and discontnuous problems: the Constrained Natural Element method. International
Journal for Numerical Methods in Enginering, 60(8):1452–1474, 2004.
Copyright
c 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2007; 00:1–28
Prepared using nmeauth.cls