Paper 1

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 27

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng 2008; 74:1421–1447


Published online 18 October 2007 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/nme.2216

Implicit boundary method for finite element analysis using


non-conforming mesh or grid

Ashok V. Kumar∗, †, ‡ , Sanjeev Padmanabhan§ and Ravi Burla§


Department of Mechanical and Aerospace Engineering, University of Florida, Gainesville, FL 32611, U.S.A.

SUMMARY
In the finite element method (FEM), a mesh is used for representing the geometry of the analysis and
for representing the test and trial functions by piece-wise interpolation. Recently, analysis techniques that
use structured grids have been developed to avoid the need for a conforming mesh. The boundaries of
the analysis domain are represented using implicit equations while a structured grid is used to interpolate
functions. Such a method for analysis using structured grids is presented here in which the analysis
domain is constructed by Boolean combination of step functions. Implicit equations of the boundary are
used in the construction of trial and test functions such that essential boundary conditions are guaranteed
to be satisfied. Furthermore, these functions are constructed such that internal elements, through which
no boundary passes, have the same stiffness matrix. This approach has been applied to solve linear
elastostatic problems and the results are compared with analytical and finite element analysis solutions to
show that the method gives solutions that are similar to the FEM in quality but is less computationally
expensive for dense mesh/grid and avoids the need for a conforming mesh. Copyright q 2007 John Wiley
& Sons, Ltd.

Received 15 February 2007; Revised 20 June 2007; Accepted 14 September 2007

KEY WORDS: non-conforming mesh; structured grid; level set method; implicit boundary method;
extended FEM (X-FEM)

1. INTRODUCTION

The finite element method (FEM) has been used extensively for engineering analysis problems not
only due to its applicability to solve a variety of physical problems but also due to its ability to handle
complex geometry and boundary conditions. For finite element analysis, the domain of analysis

∗ Correspondence to: Ashok V. Kumar, Department of Mechanical and Aerospace Engineering, University of Florida,
Gainesville, FL 32611, U.S.A.

E-mail: akumar@ufl.edu

Associate Professor.
§ Graduate Student.

Copyright q 2007 John Wiley & Sons, Ltd.


1422 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

Figure 1. (a) FEA mesh; (b) meshless approach; and (c) non-conforming mesh/grid.

is subdivided into elements (Figure 1(a)) and the resulting finite element mesh approximates the
geometry and is used to represent the trial and test functions by piece-wise interpolation within
each element. Using Galerkin’s approach, a symmetric set of linear simultaneous equations is
obtained. Advancements in computers have greatly reduced the computation time involved in
solving equations. However, creating a satisfactory mesh that conforms to the boundaries of the
domain still requires significant human input, and is therefore often, the most time-consuming
aspect of the overall analysis process. Mesh generation algorithms have been developed that work
acceptably for most 2D problems but can be unreliable for some 3D geometries often resulting in
poor or distorted elements in some regions that can lead to large error in the solution. A significant
amount of user intervention is sometimes needed to correct such problems.
Meshless methods of analysis avoid some of the problems associated with mesh generation by
using interpolation schemes that do not require connectivity between nodes. The nodes are scattered
within the problem domain as well as on the boundary of the domain as shown in Figure 1(b).
A variety of meshless methods have been proposed in the literature and they often differ in the
method of constructing the shape functions for the interpolation of the nodal variables. One of the
earliest approaches was smoothed particle hydrodyamics [1]. The most widely used shape functions
for meshless methods are based on moving least squares (MLS) approximation which originated
for the purpose of data fitting and surface construction. Nayroles et al. [2] used MLS in a method
called diffuse element method (DEM). Element-free Galerkin (EFG) [3] method, which evolved
from the DEM, also uses the MLS approximation and computes nodal stiffness matrices using a
background grid for integration. Krongauz and Belytschko [4] developed a technique where a string
of finite elements are used on the boundary of the domain on which essential boundary conditions

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1423

need to be applied. The shape functions of the edge finite elements are combined with the meshless
shape functions to impose the essential boundary conditions. The meshless local Petrov–Galerkin
(MLPG) method by Atluri and Zhu [5] uses MLS approximation and a local symmetric weak
form derived from weighted residual method confined to a small local subdomain of a node. The
weak form of the boundary value problem is integrated over local quadrature domains which
eliminates the need for a background grid. The point interpolation method by Liu and Gu [6]
is a way of constructing shape functions that possess the Kronecker delta property. This enables
the imposition of the essential boundary conditions. Melenk and Babuska [7] developed a more
generalized framework for meshless shape functions based on partition of unity that includes the
MLS approximation. Belytschko et al. [8] have provided a comprehensive review of the meshless
methods in a review paper. The method of finite spheres [9, 10] was developed as a meshless
technique based on partition of unity. Integration of weak form was performed on spheres using
specialized cubature rules. Macri and De [11] have presented efficient integration schemes for the
method of finite spheres. The coupling of the method of finite spheres with traditional FEM to
improve computational efficiency is also discussed. Sukumar et al. [12] developed natural neighbor
Galerkin method (NEM) which uses a non-sibsonian interpolation scheme based on Voronoi
tessellation. Dolbow and Belytschko [13] have investigated the errors arising in the numerical
integration of Galerkin weak form in mesh-free methods that use non-polynomial shape functions.
An alternative to the meshless approach has been to use non-conforming mesh, often a structured
grid, for the analysis where the geometry of the analysis domain is represented using implicit
equations and is therefore independent of the grid or mesh used. Figure 1(c) shows a structured
non-conforming grid. A structured grid is much easier to generate than a finite element mesh
and all the elements in the grid can have regular geometry (squares/rectangles/cubes). However,
since nodes are not guaranteed to be on the boundary, the traditional methods used in FEM for
applying essential boundary conditions cannot be used. The use of implicit equations for applying
boundary conditions was first proposed by Kantorovich and Krylov [14]. This approach, which
will be referred to as the implicit boundary method (IBM) in this paper, involves constructing
solution structures for functions such that the essential boundary conditions are satisfied. Let u be
a trial function defined over  ⊂ R2 or R3 that must satisfy the boundary condition u = a along a
which is part of the boundary of . If a (x) = 0 (x ∈  and a :  → R) is the implicit equation
of the boundary a , then the trial function can be defined as

u(x) = a (x)U (x)+a (1)

The solution structure for the trial function defined in Equation (1) is then guaranteed to
satisfy the condition u = a along the boundary defined by the implicit equation a (x) = 0 for any
U :  → R. The variable part of the solution structure is the function U (x) which can be replaced
by a finite dimensional approximate function U h (x) defined by piece-wise interpolation within
elements of a structured grid.
This approach for analysis using implicit equations and a structured grid has been described
by Hollig [15] in his book where R-functions are used to construct implicit equations of the
boundaries of the domain and B-spline interpolations are used for the variable part of the solution
structure. R-functions, which were originally developed by Rvachev and Shieko [16], have the
property that its sign depends only on the sign of its arguments and not their magnitude. This
property can be used to construct implicit functions for representing solids that are created by
a Boolean combination of half spaces. Shapiro and Tsukanov [17] and Shapiro [18] have also

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1424 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

used an R-Function-based representation to construct solution structures that satisfy a variety of


prescribed boundary conditions including essential, natural and convection boundary conditions.
Belytschko et al. [19] have developed an extended finite element method (XFEM) using a
structured grid and implicit boundary representation. Approximate implicit function of the solid
was constructed using radial basis functions by fitting a set of sample points on the boundary. Moes
et al. [20] have extended the XFEM for imposition of essential boundary conditions using carefully
selected Lagrange multiplier space on the boundaries. Laguardia et al. [21] have used NEM with
a structured placement of nodes using quadtree subdivision. They have also used R-functions for
the imposition of essential boundary conditions. Clark and Anderson [22] have used the penalty
boundary method to perform the finite element analysis based on a non-conforming mesh. The
weak form is constrained with the penalty factor to satisfy the prescribed boundary conditions.
The method presented here can be considered an XFEM [19] which uses the approach of
Kantorovich and Krylov [14] to apply essential boundary conditions. Solution structures are created
to represent the trial and test spaces using the implicit equations of boundaries such that essential
boundary conditions are automatically satisfied. The method can impose the boundary conditions
even if none of the nodes in the mesh or grid falls on the boundary. In this paper, the solution
structure is constructed using specially constructed weighting function, referred to here as essential
boundary function, which has a unit value everywhere except in the vicinity of boundaries on
which essential boundary conditions are applied. This function has zero value at the boundaries
and varies between zero and one over a very narrow band near the boundary. The limiting case is
studied in this paper when the width of this band tends to zero. This approach has been referred
to here as the IBM for applying boundary conditions because it uses implicit equations of the
boundary to construct the weighting function and the solution structure. The volume integrals in
the weak form are evaluated using step functions of the solid that have a unit value inside the solid
and zero outside. In Section 2, a brief summary is presented of the method for constructing the
implicit definition of the domain boundary using approximate step functions. Section 3 describes
the development of a modified weak form that incorporates the solution structure to satisfy the
essential boundary conditions. Several examples that illustrate the effectiveness of the method are
discussed in Section 4. The conclusions and inferences are provided in the final section of this
paper.

2. SOLID MODEL REPRESENTATION USING IMPLICIT EQUATIONS

Most modern solid modeling software use Boundary Representation (B-Rep) [23] to represent
solids where the geometry of curves and surfaces are typically represented using parametric
equations of curves and surfaces. Solid modeling using implicit equations has also been studied in
past research. The constructive solid geometry (CSG) technique [23] defines solids by the Boolean
combination of simple convex solids called primitives. These primitives are themselves defined as
intersections of half-spaces which are implicit equations of the form f (x)−c0 where x ∈ R2 for
planar curves and x ∈ R3 for surfaces. The implicit surface,  = f (x)−c = 0, divides space into
two regions: >0 and <0, which enables easy classification of a point as being inside or outside
depending on the sign of  computed at that point. The function  is often referred to as the
characteristic function.
As mentioned earlier, many recently developed structured grid analysis methods use Boolean
operations defined using R-functions to create the implicit equations needed for constructing

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1425

solution structures. Methods for the construction of implicit equations using Boolean operation
have been discussed by several authors [24–26]. In this paper, Boolean operations defined using
approximate step functions [27] are used to define the boundaries of the solid. Step function of a
solid is defined to have a value of 1 inside the solid and 0 outside. Step functions have been used
to perform volume integrations by several authors [19, 28–30]. In the level set method [30], signed
distance functions are used to define the implicit function. The distance function is then used to
construct step function of the solid for computing volume integrals. The level set method has been
extensively used for solving problems involving moving boundaries. The level set approach has
also been recently used for structural topology optimization [31, 32].
The approximate step function of a boundary represented by the implicit equation (x) can be
defined as follows:


⎪ 1, 


1 
h() = + , − (2)

⎪ 2 2


0, −
The function defined above is a linear approximate step function. Higher order approximate
step functions like quadratic, cubic step functions [33] or the smeared step function [30] can also
be used for defining the domain. The approximate step function tends to the exact Heaviside step
function as the range parameter  → 0. The volume integral of any function f (x, y, z) defined over
a domain  can be evaluated as [28–30]
 
Vf = f (x, y, z) dX = f (x, y, z)h(0 ) dX (3)
X0 X1

where 0 is the implicit curve bounding the volume of integration X0 and h(0 ) is the approximate
step function of the domain. Note that the function can be integrated over any arbitrary domain
X1 ⊃ X0 . If a grid overlaps the domain X0 then the integration can be carried over each element
of the grid and summed over the total number of elements n e .
ne 

Vf = f (x, y, z)h(0 ) dX (4)
i=1 Xe

The step function transitions from a zero value to unity at the boundary over a small distance
when  has small values. Therefore, for all internal elements the step function is unity and volume
integration can be done using Gauss Quadrature as in the traditional FEM. However, since a
structured grid is used for the analysis, all the internal elements have identical shape and size and
therefore the value of the integral is identical for all internal elements and can be even precomputed
analytically since the elements are rectangular or prismatic. For elements on the boundary, it often
helps to subdivide the elements further using the quadtree or octree subdivision scheme [23],
especially if the grid is coarse, to ensure accurate evaluation of the integrals. However, as illustrated
later in examples, as the grid is refined to ensure the accuracy of the results, the element’s size is
often small enough so that the subdivision of boundary elements does not significantly increase
the accuracy of the integration.
In the structured grid-based approach presented here, implicit equations of the boundaries of
the analysis domain are used for volume integration as well as for constructing solution structures.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1426 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

Figure 2. Multiple boundaries in a single boundary element.

The characteristic function, i for each boundary is defined by piece-wise interpolation over the
elements of the grid that the boundary passes through. When more than one curve is used to define
the boundary, elements that contain intersections of these curves have multiple curves passing
through them. Characteristic functions for all the curves passing through the element have to be
defined within it. Within each boundary element e ∈ E B , where E B is the set of gridelements
through which the boundary passes, the nodal density values are interpolated as i = j N j iej ,
where iej are nodal values of the ith characteristic function defined in element e, and N j are the
basis functions used for the interpolation. If equations of the surfaces bounding the solid are not
available in implicit form, then a convenient way to construct these characteristic functions is to use
signed distance function [30] so that iej is the distance of the jth node from the ith boundary with
a negative sign if the point is outside the solid. The step function of the solid region (inside of the
object) in that element is defined by Boolean operations between the step functions of individual
curves. In Figure 2 there are two boundaries 1 and 2 in a single boundary element. The solid
region in the element is defined using Boolean operations between the half-spaces defined by
these two boundaries. Step functions of regularized Boolean combinations are defined as shown
below [27].
Intersection operation:

H = h(1 )h(2 ) (5)

Difference operation:

H = h(1 )h(−2 ) (6)

Union operation:

H = h(1 )+h(2 )−h(1 )h(2 ) (7)

The operators given above tend to round sharp corners of an object for large values of the range
parameter  but is exact in the limit  → 0. The Boolean operations can be applied repeatedly when
more than two boundaries are present in a single boundary element. The characteristic function for

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1427

each boundary curve is only defined within the elements that it passes through; therefore, the step
function for internal and external elements is predefined to have unit and zero values, respectively.

3. IMPLICIT BOUNDARY FEM FOR ELASTOSTATIC PROBLEMS

The IBM is illustrated here for linear elastostatic problems even though the method can be
generalized and extended to any other linear boundary value problem without difficulty. The weak
form of the elastostatics boundary value problem is the principle of virtual work which is often
expressed as
  
dcT r d = duT T d+ duT b d (8)
 t 

In the above equation, r is the Cauchy stress tensor, T is the applied traction vector on the
boundary t and b is the body force vector. The test function or virtual displacement is denoted by
du and the corresponding virtual strain vector is dc. The natural and essential boundary conditions
are r· n̂ = T on t and u i = u ia on au , respectively, where n̂ is the unit vector normal to this
boundary and u ia the prescribed value of the ith displacement component on the ath Dirichlet
boundary au . The constitutive equation is the linear stress–strain relationship, r = Cc, where c is
the strain expressed as a column matrix and C is the elasticity matrix.
The space of trial functions Su and the space of test functions Su can be defined over  ⊂ Rm
as

Su = {u | u = Dug +ua } (9)


Su = {u | u = Dug } (10)

where u, u :  → Rm and m is the dimension of the problem. The components of


u = {u i } ∈ Su satisfy the condition u i = u ia along the boundary defined by the implicit equation
Di (x) = 0, where Di are components of D :  → Rm and u ia are components of ua :  → Rm ,
which will be referred to as the boundary value function. The variable part of the solution structure
g
is the function ug = {u i } which can be defined using a finite-dimensional approximation over the
g
grid. In this paper, u is referred to as the grid variable and is constructed by piece-wise inter-
polation using finite element shape functions (Lagrange interpolation functions). However, these
grid variables can be defined using B-spline interpolations or any of the meshless interpolation
schemes such as MLS.
The solution structure for admissible trial and test functions for displacement is expressed below
in index notation (with no implied summation over i):
g
u i = u is +u ia = Di u i +u ia (11)
g
u i = Di u i (12)

The functions u ia are the boundary value functions that must be constructed such that u ia = u ia on
au .
The functions, Di , which will be referred to as the Dirichlet functions or essential boundary

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1428 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

functions must satisfy the following conditions:

(i) Di = 0 on au
(ii) |∇ Di |  = 0 on au (13)
(iii) Di (x)>0 ∀x ∈ 

A method for constructing such Dirichlet functions using implicit equation of the boundaries
au is provided in the next subsection. Condition (i) is needed to ensure that u i = u ia on au , while
condition (ii) is necessary to ensure that |∇u i | is not constrained to be zero along the boundaries
and finally condition (iii) is necessary to ensure that u i is not constrained to be equal to u ia
anywhere within the domain of analysis. The homogeneous part of the solution structure is the
variable us = {u is } that satisfies homogeneous boundary conditions along au . Using this solution
structure the stresses and strains can also be decomposed into a homogeneous part and a boundary
value part:

r = Cc = C(cs +ca ) = rs +ra (14)

where
   
*u is *u j *u ia *u j
s a
1 1
is j = + and iaj = +
2 *x j *xi 2 *x j *xi

Substituting these relations into the weak form, the following relation is obtained, where the
known quantities have been moved to the right-hand side of the equation:
   
dc r d = −
T s T a
dc r d+ T
du T d+ duT b d (15)
  t 

The left-hand side involves the term rs , which depends on the variable part of the solution
structure. The volume integral can be computed as
 

dcT rs dV = dcT Ccs H de (16)
 e e

where e represents the volume of the eth element in the grid, H is the step function of the
domain of analysis  and the summation is over all the elements in the grid. The homogeneous
strain components are
 g
1 *Di g *u
g
*D j g *u j
is j = u i + Di i + u j +Dj (17)
2 *x j *x j *xi *xi

These strain components can be expressed using a matrix notation such as cs = BX, where X
contains the nodal values of the components of ug . For example, in 2D elastostatics, denoting the

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1429

components of us as (u s , v s ), the components of D as (Du , Dv ) and the grid variable ug = (u g , v g ),


the strain components corresponding to the homogeneous displacement us are
⎧ ⎫

⎪ *u s ⎪


⎪ ⎪


⎪ *x ⎪


⎪ ⎪

⎨ *vs ⎬
{cs } =

⎪ *y ⎪


⎪ ⎪


⎪ *u *v ⎪


⎪ s s ⎪

⎩ + ⎭
*y *x
⎡ ⎤
*N1 *Du *Nn *Du
⎢ Du *x + N1 *x 0 ··· Du
*x
+ Nn
*x
0 ⎥
⎢ ⎥
⎢ ⎥
⎢ *N1 *Dv *Nn *Dv ⎥
=⎢
⎢ 0 Dv + N1 ··· 0 Dv + Nn ⎥
⎢ *y *y *y *y ⎥

⎢ ⎥
⎣ *N1 *Du *N1 *Dv *Nn *Du *Nn *Dv ⎦
Du + N1 Dv + N1 ··· Du + Nn Dv + Nn
*y *y *x *x *y *y *x *x
⎧ ⎫
⎪ u
⎪ g1 ⎪
⎪ ⎪


⎪ ⎪
⎪ vg1 ⎪
⎪ ⎪


⎨ ⎪ ⎬
× .. (18)
⎪ . ⎪
⎪ ⎪

⎪ ⎪
⎪ u gn ⎪
⎪ ⎪


⎪ ⎪
⎩ ⎪ ⎭
vgn

The above equation is obtained


n assuming that the
components of the grid variable are interpolated
g n g
within each element as u g = i=1 Ni u i and v g = i=1 Ni vi , where Ni are the shape functions,
g g
(u i , vi ) are the nodal values of the grid variable at node i and n is the number of nodes in the
element. When strain is expressed as es = BX and substituted into the left-hand side of the weak
form, Equation (15), it can be reduced to a discrete form as in traditional finite element formulation
allowing the stiffness matrix for each element to be expressed as

K= BT CBH (e ) de (19)
e

However, it is advantageous to split the B matrix into two parts to separate terms that only
contain the gradient of the shape functions and terms that contain the gradients of the essential
boundary functions Di . This allows the strain to be expressed as

cs = (B1i +B2i )Xi (20)
i

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1430 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

where
⎡ ⎤ ⎡ ⎤
*Ni *Du
⎢ Du *x 0⎥ ⎢ Ni *x 0 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ *Ni ⎥ ⎢ *Dv ⎥
B1i = ⎢
⎢ 0 Dv ⎥, B2i = ⎢
⎢ 0 Ni ⎥ (21)
⎢ *y ⎥⎥ ⎢ *y ⎥

⎢ ⎥ ⎢ ⎥
⎣ *Ni *Ni ⎦ ⎣ *Du *Dv ⎦
Du Dv Ni Ni
*y *x *y *x
and
 g
ui
Xi = g
(22)
vi
Using a mixed index and matrix notation, the stiffness matrix can be expressed as shown below:

Ki j = Ki1j +(Ki2j +K2ji )+Ki3j (23)

where

Ki1j = BT1i CB1 j H dV
e

Ki2j = BT1i CB2 j H dV (24)
e

Ki3j = BT2i CB2 j H dV
e

It is clear from the above equation that for elements where the gradients of the essential boundary
functions Di vanish, the stiffness matrix Ki j = Ki1j . This provides an incentive to construct the Di
functions such that they have uniform values over much of the domain of analysis as discussed in
the next section. To evaluate the stiffness matrix Ke , the integral over the volume of the element can
be evaluated by Gauss quadrature approach for all the internal elements for which, by definition,
the step function H (e ) = 1. For boundary elements, it may be necessary to further subdivide the
element for accurate quadrature if the grid is coarse and the elements are relatively large. This can
be done most efficiently by using quadtree subdivision [23], followed by Gauss quadrature within
each subdivision.
When stiffness matrices are computed using structured grids, it is very likely that some elements
at the boundary have very small volume fraction (ratio of solid volume to element volume). Some
nodes of such elements that are not shared with any other elements will then have very small
stiffness resulting in ill-conditioned or singular global stiffness matrix. To avoid this problem, such
nodes need to be identified and constrained. In addition, the stiffness of elements whose volume
fraction is below a threshold (5%) is neglected and not added to the global stiffness matrix.
The element stiffness matrices are assembled in the traditional fashion to construct a global
stiffness matrix. There are two volume integrals and one surface integral term on the right side
of Equation (15). The curve/surface integral can be computed by approximating the curve and
surface using lines or surface triangulations, respectively, and performing numerical integration

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1431

using Gauss quadrature over line segments or triangles. This approach for curve/surface integration
has been used by other authors for similar grid-based approaches (see e.g. [19]).

3.1. Essential boundary function


The essential boundary functions Di must have a zero value on any boundary that has an essential
boundary condition specified. In addition, the function value must be positive and non-zero inside
the solid. Functions that satisfy the above requirements can be constructed in a number of different
ways. The method of constructing Di functions that were used for this study is described below
followed by a discussion of the advantages of this approach over several other possible choices
that have been considered. If essential boundary conditions are applied on the ith component of
displacement on boundaries defined by implicit equations  j = 0, where j ∈ J Di and J Di is the set
of boundaries on which these boundary conditions are applied, then Di (x) can be defined as
⎧ 
⎨ d( j ) if H <1.0
Di (x) = j∈J Di (25)

1.0 if H = 1.0
In the above definition, H is the step function of the solid region in the boundary element and
the functions d(i ) are constructed as

⎪ 1, >





⎪  2


⎨ − 2 +2 , 0
 
d() = (26)

⎪  2


⎪ − 2 −2 , −0



⎪  

1, <−
Figure 3 shows the plot of the d() function which shows that this function is defined such
that it is equal to zero at the boundary  = 0 and its value becomes equal to 1 outside the band
−<<. It has a positive value everywhere i  = 0, a discontinuous non-zero slope at  = 0
and has a continuous zero slope at  = ±. These properties of d() ensure that Di satisfies
the requirements of the essential boundary functions listed in Equation (13). The function d(i )
defined in Equation (26) is piece-wise quadratic in the interval i ∈ [0, ]. Similar piece-wise linear
or cubic functions can be created if a lower or higher degree of continuity is needed.
Note that the parameter , which defines the range over which d() varies, is the same that
is used to define step function h() in Equation (2) so that both Di and H reach unit values
at the same distance from the boundary. In addition, when  → 0, the Di function reaches unit
values very close to the boundary. This implies that for all the internal elements Di has a value
equal to unity which makes the internal elements exactly identical to traditional finite elements
in formulation and the computed stiffness matrix. Furthermore, since a regular structured grid is
used here for the analysis, all internal elements are of the same shape and size and have identical
stiffness matrices. Therefore, it is sufficient to compute stiffness for any one internal element and
use the same for all other internal elements, thereby saving computational time.
A very small value for the range parameter  implies that the Di function has very large
gradients near the boundary. This requires special treatment for evaluating the stiffness matrix

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1432 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

Figure 3. D-function plot.

Figure 4. Boundary band (0<<).

of the boundary elements that contain a boundary on which an essential boundary condition is
applied. The following discussion illustrates the method for the computation of stiffness matrix
for these elements.
In Equation (23), the stiffness matrix was decomposed into components Ki1j , Ki2j and Ki3j . The
first component Ki1j is computed in the same manner as for the boundary elements that do not have
essential boundary conditions. The second and third components contain gradients of essential
boundary functions Di which have large values near the boundary when  is made very small
and they vanish at points that are far from the boundary. Hence, the terms in the integrals will
be non-zero only within a very thin band along the boundary. Figure 4 shows an element on the
boundary of the domain and the band (0) near the boundary within which the gradient of
Di is non-zero.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1433

It is convenient to perform the integration with respect to a local coordinates system (t–n)
with axes along the tangent and normal direction. The integral along the normal direction can be
converted to an integral over  from 0 to  and the integral in the tangential direction can be
computed by Gauss quadrature over piece-wise linear approximation of the boundary. In the limit
as  → 0, the volume integrals can be converted into surface integrals as follows:
    
−1
Ki2j = BT1i CB2 j d = BT1i C B2 j (∇· n̂) d d (27)
  0
    
Ki3j = BT2i CB2 j d = BT2i CB2 j (∇· n̂)−1 d d (28)
  0

The above approach leads to fairly accurate and efficient computation of stiffness matrix for
boundary elements with essential boundaries. It also illustrates the advantage of constructing the
essential boundary functions as in Equations (25)–(26). Note that even though the function d()
has a discontinuous slope at  = 0 this does not cause a problem when the stiffness matrix is
computed as in Equations (27)–(28). In these equations, the inner integral over  can be done
analytically treating Ni , *Ni /*x and *Ni /*y as constants because for very small values of , the
band {x|0(x)} is very narrow. Since the integration is performed for  ∈ [0, ] we can use

2 
d() = − +2
 2 
according to Equation (28).
Several authors have suggested using R-functions [15–18] to construct the essential boundary
functions. However, the resulting function is non-linear throughout the domain of analysis and will
require integration over all internal elements. Owing to the highly non-linear nature of essential
boundary functions constructed using R-functions the resulting solution structure has difficulty
representing constant strains or stresses exactly and therefore does not pass the patch test. Another
approach for constructing essential boundary functions has been suggested by Hollig [15] as
follows:

w(x) = 1−max(0, 1−dist(x, *D)/) (29)

This function has some similarity to the D-function presented here in that it also plateaus to a
value of 1 at a fixed distance  from the boundary. Its definition is based on constructing a distance
function and cannot use the implicit equation of the boundary directly even when it is known in an
analytical form. This function can be used as an alternative to Equations (25) and (26) if the entire
geometry can be represented using distance functions. The limit of  → 0 has not been studied,
which is necessary to consider all internal elements identical.
In the limit  → 0, the approach presented in this paper inherits many of the well-known
advantages of the traditional FEM. Since the Di function becomes unity for all the internal elements,
these elements are identical to the conventional finite elements. For the boundary elements also Ki1j
has the same structure as the stiffness matrix of traditional finite elements; however, the integration
is carried out only over the region within the element that is solid. The additional terms Ki2j and
Ki3j are added to the stiffness of boundary elements only if they contain a boundary with essential
boundary conditions. These terms can be thought of as added stiffness that enforces the boundary

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1434 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

condition and as discussed earlier, these terms can be computed efficiently and accurately using
Equations (27)–(28).

3.2. Boundary value function


The boundary value functions u a must be defined such that they have a value equal to the imposed
essential boundary conditions on the appropriate boundaries. If the essential boundary conditions
are specified at multiple boundaries and the assigned values are not the same for all boundaries
then it is necessary to construct a continuous function that takes on appropriate values at respective
boundaries while transitioning smoothly in between. Transfinite interpolation schemes have been
proposed to achieve this goal [34]. However, the resulting function is rational and is too non-linear
to be used in the solution structure in Equations (9)–(10). Using such rational functions causes
difficulty in passing the patch test and the solution structure is often unable to accurately represent
constant strain fields.
While there is no unique method for constructing boundary value functions, it is beneficial
to ensure that the trial function is a polynomial of the same order as the shape functions used
for interpolating the grid variable. This can be accomplished by constructing the boundary value
function by piece-wise interpolation using the element shape functions. The nodal values of u a
can be interpolated similar to the grid variable u g using the grid element shape functions:

u a = Ni u ia (30)
i

In Equation (30) Ni are the shape functions of the grid element and u ia are the nodal values of
ua .It is important that the function u a should belong to the space spanned by the shape functions
used to represent the grid variable. If the shape functions Ni in Equation (30) are identical to
the shape functions used for interpolating the grid variable then this requirement is automatically

Figure 5. Boundary value function.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1435

satisfied. Otherwise, the grid variable cannot compensate for u a in the solution structure to better
approximate the exact solution leading to poor convergence. The solution structure also will not
satisfy completeness requirement for inappropriate selection of u a .
The nodes of the boundary elements should be assigned nodal values of u a such that upon
interpolating these values, the resulting function will have the desired values at the boundary
passing through the element. This is very easy to achieve when the assigned value is constant or
even linearly varying along the boundary. The rest of the nodes in the grid that are not part of any
boundary element can have any arbitrary value and therefore can be set equal to zero. Figure 5
shows a structured grid for an arbitrary analysis domain. To apply the essential boundary condition
u =  on the top edge of the domain, the nodal values of u a are set equal to  at all the nodes
shown in black color while at all other nodes the nodal values are set equal to zero. The boundary
value function u a contributes to the load computation on the right-hand side of Equation (15) for
all those elements in which the gradient of u a has non-zero magnitude.

4. RESULTS AND DISCUSSION

The IBM described in this paper was implemented by modifying a finite element program. The
shape of the analysis domains was defined using distance functions constructed from traditional
geometric/solid models. These distance functions, which are implicit equations of the boundary,
were used for constructing the solution structure and for the computation of the stiffness matrix
and load vectors. In this section we present few of the examples that were used to validate this
approach. In the 2D examples, bilinear 4-node quadrilateral elements were used with IBM (IBM
Q4) and compared with similar elements in FEA (FE Q4). Similarly, for the 3D example, 8-node
(trilinear) hexahedral elements were used (IBM H8 and FE H8). The first example is a plate
under uniaxial stress. While the problem itself is trivial, it is an important example that illustrates
fundamental aspects of the IBM. One could think of this example as being equivalent to the patch
test used for non-standard finite elements to verify whether they satisfy completeness condition
[35]. However, unlike in the patch test, only structured grids are used since our objective is to
eliminate the need for distorted elements and conforming mesh.

4.1. Example 1: rectangular plate with axial loading


A rectangular plate of dimensions 2.0 in by 1.0 in subjected to uniaxial loading of 1000 psi was
modeled as in Figure 6. Making use of symmetry conditions, only a quarter of the plate was
modeled as shown. The material of the plate is assumed to be steel with a modulus of elasticity
equal to 30×106 psi and a Poisson ratio of 0.3. Since the load is uniaxial and the stress is constant
within the plate, this problem can be modeled using a single 4-node bilinear element using the
traditional FEM. However, using IBM and its non-linear solution structure it is not obvious that
a single element can in fact provide a reasonable solution. To study this aspect, first, a model
consisting of a single 4-node quad element was constructed for analysis using the implicit boundary
approach. Figure 6(a) shows a single conforming element that models the plate while Figure 6(b)
shows another model where a non-conforming element that overlaps the plate is used.
The essential boundary functions Du and Dv are constructed such that the x-component of
displacement is fixed along the left edge of the plate and the y-component is fixed along the lower
edge. Therefore, the value of Du is equal to zero on the left edge and the value of Dv is zero

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1436 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

Figure 6. Plate under uniaxial loading: (a) conforming element and (b) non-conforming element.

Figure 7. Effect of range parameter on the computed X -displacement: (a) conforming


grid and (b) non-conforming grid.

at the lower edge. Since only homogeneous boundary conditions are applied, the boundary value
functions u a and v a are set equal to zero over the entire domain.
The analytical solution for this simple uniaxial stress problem is a linear displacement field with
constant strain x x = 3.333×10−5 in the X direction and  yy = −1.0×10−5 in the Y direction. The
horizontal component of displacement varies linearly from zero to the maximum displacement of
u max = 3.333×10−5 in in the X direction and vmax = −0.5×10−5 in in the Y direction.
Figure 7(a) and (b) compares the analytical solution and the numerical solution obtained using
a single conforming and non-conforming element, respectively, for a range of values of the range
parameter  used in computing Du and Dv . The plots show the x-displacement along the center of
the plate and as expected the displacement boundary condition is exactly imposed even when the
element is not conforming to the plate geometry for all values of the range parameter. The value of
 was varied from 1.0 to 10−4 to study its effect on the accuracy of the computed displacements.
Since this parameter determines the range over which the essential boundary functions vary non-
linearly from 0 to 1, large values of this parameter cause large errors. As the value of the range

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1437

parameter is reduced, the non-linearity of the solution is restricted to a small region and the results
obtained are closer to the theoretical results. For values equal to or less than 0.001, the error within
the element was found to be negligible. For the rest of the examples in this paper,  = 5×10−4
was used.
A single 4-node quadrilateral (bilinear) element using IBM passes the patch test provided the
range parameter,, is very small. For larger values of  (>0.001) a single element is not enough to
pass the patch test, but the solution does converge towards the exact solution with increasing grid
density. Obviously, for better computational efficiency and accuracy it makes sense to use very
small values for this parameter.

4.2. Example 2: curved beam


The curved beam shown in Figure 8 has a known analytical solution for stress distribution and
is studied here to illustrate the application of the IBM when the grid is truly non-conforming
at all the boundaries. The internal radius of the beam is 0.3 m and it has a square cross-section
(0.2 m×0.2 m). A load of 6 kN is applied on the end in the vertical direction. This beam is modeled
as a plane stress problem and is assumed to be made of steel with Young’s modulus of 200 GPa
and a Poisson ratio of 0.29.
The circumferential stress at points denoted as A and B was calculated from analytical solutions
[36] as A = −4.45 MPa and  B = 2.91 MPa. Figure 9 shows a structured grid of 4-node quad

elements with a total of 554 nodes that was used to model this curved beam. The grid encloses the
problem domain and the edges where boundary conditions are applied do not have any nodes on
them. The traction applied on the boundary was integrated according to Equation (15) to compute
the equivalent nodal forces that are shown schematically in Figure 9.
The stiffness matrices were computed and assembled to form the global system of equa-
tions that were solved to obtain the nodal values of the grid variable. These grid variables were
then used to compute the actual displacement field based on the solution structure in Equation
(9). The computed normal stress component in the Y direction is plotted in Figure 10. The
computed values of stresses are at the points A and B, respectively, are  A = −4.34 MPa and

 = 2.89 MPa. It can be seen that these values are very close to the theoretical values stated
B

earlier.

Figure 8. Curved beam model.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1438 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

Figure 9. Non-conforming structured grid and equivalent nodal forces.

Figure 10. Stress component  yy .

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1439

10
FE Q4
IBM Q4

10
log(Error in Strain Energy)

10

10

10 1 2 3 4
10 10 10 10
log (Num Nodes)

Figure 11. Plot of error in strain energy with respect to the number of nodes.

Convergence analysis was performed for this problem by successive refinement of the model
and plotting the error in strain energy versus number of nodes as shown in Figure 11. The error was
computed by comparing with the strain energy obtained using a very dense finite element mesh
(95 344 nodes). For the convergence plot, the coarsest grid used had 79 nodes and 54 elements,
while the finest grid considered in this example had 4194 nodes and 3933 elements. The values
are plotted on a log–log scale, where the x-axis represents the log of the number of nodes in the
model and the y-axis represents the error in the strain energy. The plot shows results obtained
using the traditional finite element approach and the IBM both using 4-node bilinear quadrilateral
element. As the number of the nodes in the model is increased, the error in strain energy obtained
by both approaches converges at approximately the same rate.
One possible source of error in the implicit boundary approach described in this paper is during
volume integration using step functions to compute stiffness matrix for boundary elements as in
Equation (19). Of course, the integration could be performed more accurately and perhaps faster
by subdividing the boundary elements into triangles for quadrature such that the boundary is
approximated by the edges of the triangles. However, the step function-based approach was used
here due to the ease of implementation and because, for sufficiently dense grids, the errors are
not larger than the errors incurred in approximating the curved edges with straight lines as is
done for linear and bilinear finite elements. To study the accuracy of the integration using step
function, three non-conforming grids with different element sizes (Grid1: 0.075, Grid2: 0.03 and
Grid3: 0.01) were used to compute the area of the curved beam structure using Equation (3).
The number of quadtree subdivisions for the volume integration (area computation) was varied
from 0 to 3. Since the actual area of the structure is a known quantity, the percentage error in
the computed value of volume integral was determined. Figure 12 shows the percentage error

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1440 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

Figure 12. Effect of subdivisions on volume integral.

in the computed area plotted as a function of number of quadtree subdivisions for each grid. It
can be seen from the figure that for coarse grids the error decreases with increasing number of
quadtree subdivisions but for fine grids there is no need for subdivision since the error is negligible
even without any subdivisions. In fact for grids dense enough to accurately represent the solution,
quadtree subdivisions are often not needed and will increase computational time unnecessarily, if
used.
Using the same grids described above (Grid1–Grid3), the plane stress problem described in
Figure 8 was solved using IBM and, for each grid, the number of quadtree subdivisions used
in stiffness computation was varied from 0 to 3. A finite element mesh with approximately the
same element sizes as the structured grid used for IBM was created and used to analyze the same
problem using the traditional FEM. The y-component of the normal stress along the horizontal
line from point A to B obtained for Grid 2 and Grid 3 is plotted in Figure 13. The theoretical stress
distribution is also plotted for comparison purpose. While there is improvement in the solution
with increasing quadtree subdivisions for the coarse grids, as the grid density increases there is no
significant improvement in the quality of the solution because the solution is very accurate even
without subdivision. It is clear from Figure 13 that quadtree subdivision is not necessary when the
grid is sufficiently dense.
The computational time comparison between IBM and traditional FEM was conducted to under-
stand the trade-off between not having to compute stiffness matrix for internal elements and the
extra cost associated with evaluating the stiffness of the boundary elements using quadtree subdi-
visions. Both methods used the same sparse matrix solver [37] to eliminate differences arising
from the differences in time taken for equation solving. In addition, identical algorithms and
implementation were used for assembling the stiffness matrices. The central processing unit (CPU)
time for stiffness computation and assembling the global matrix was recorded for each grid with

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1441

Figure 13. Distributed of  yy through thickness: (a) Grid 2 (element size 0.03)
and (b) Grid 3 (element size 0.01).

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1442 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

Figure 14. Comparison of CPU time taken: (a) stiffness computation time and (b) total analysis time.

varying subdivisions. The total analysis time was also noted. The total analysis time is the sum
of stiffness computation time and solver time. The same procedure was also repeated with FEM
grids. Figure 14(a) shows the CPU time (in milliseconds) for stiffness computation as a function
of number of elements. Figure 14(b) shows the total analysis time as a function of number of
elements.
The solver time is small compared with the stiffness computation and assembly time due to the
small size of the problem and the high efficiency of the sparse matrix solver. Hence, Figure 14(b)
looks almost identical to Figure 14(a). It is seen that for grids with very large number of elements
the time taken by the IBM is less than the FEM stiffness computation time, when fewer than three
quadtree subdivisions are performed. This is due to the fact that stiffness for all internal elements
is identical and, hence, is computed only once when a structured grid is used with the IBM, which
results in significant saving of computational time.

4.3. Example 3: 3D bracket


The bracket shown in Figure 15 was modeled using a 3D model with 8-node hexahedral elements to
demonstrate some advantages of using a structured grid with undistorted elements. The structure is
clamped on one end while the other end is subjected to traction in the Y direction. The dimensions

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1443

Figure 15. Model of 3D bracket.

Plot of Error in Strain Energy with respect to the Number of Nodes


10
FE H8
IBM H8
log(Error in Strain Energy)

10

10

10
2 3 4 5
10 10 10 10
log (Num Nodes)

Figure 16. Convergence analysis of 3D model.

of the component are shown in the figure and the magnitude of the applied traction is 10 000 Pa.
Young’s modulus is assumed to be 2×1011 Pa while the Poisson ratio is taken as 0.3.
This problem was again analyzed using both IBM and using traditional finite element software
(ABAQUS). Convergence analysis for this problem was performed by computing strain energy
for a series of models with varying mesh densities as plotted in Figure 16. The plot is shown on
a log–log scale, where the x-axis represents the number of nodes and the y-axis represents the

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1444 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

Figure 17. Comparison of bending stress.

error in strain energy. The error is computed using a reference solution obtained using a finite
element model consisting of 43 140 nodes and 9410 quadratic hexahedral elements. As the number
of nodes increases both methods converge to the same solution. In this example, the IBM provides
a better estimate of the strain energy when fewer nodes and elements are used. This is perhaps

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1445

because it is harder to accurately represent the geometry without having distorted elements in the
finite element mesh when fewer elements are used, whereas the structured grid has regular-shaped
elements regardless of the grid density.
The distribution of stress component x x shown in Figure 17(a) was computed using IBM and
a grid that has 8900 nodes and 7281 elements. The same result computed using finite element
mesh is shown in Figure 17(b) where the mesh has 50 598 nodes and 45 424 trilinear hexahedral
elements. The results obtained using IBM are close to the answer obtained by a higher density
FEM mesh. An FEM mesh with same size elements as the one used in implicit method gave a
lower value of stress perhaps because for lower mesh density the elements were more distorted to
conform to the geometry. In IBM the elements are all perfect cubes while the conforming mesh
often has significantly distorted elements, especially in 3D, which leads to errors during numerical
quadrature over the elements.

5. CONCLUSIONS

In this paper, a method for applying essential boundary conditions using implicit equations of the
boundary is discussed that enables finite element type analysis using a non-conforming structured
grid. The weak form for linear elasticity was modified using a solution structure such that essential
boundary conditions are automatically imposed. Even though this method has only been applied
to linear problems in this paper, it could be used to apply boundary conditions for non-linear
problems as well. The primary motivation for the method is the desire to eliminate the mesh
generation process and use implicit equations of curves/surfaces to represent the geometry instead
of using a conforming mesh to approximate it. Generating a uniform structured grid that encloses
the geometry is easier than to generate a conforming mesh. In addition, the method would enable
design engineers to perform analysis directly using solid models instead of first generating a mesh
and then applying boundary conditions on the mesh. One of the advantages of this method is that
all internal elements have the same stiffness matrix resulting in significant reduction in computation
required to determine the global stiffness matrix. However, the computation of the stiffness matrix
of boundary elements requires quadtree subdivisions if the mesh is coarse. The computational
efficiency has been compared only with traditional FEM and not with other approaches such as
boundary integral equation methods with fast multipole acceleration [38] which may be faster for
linear problems. The IBM enables the imposition of essential boundary conditions on boundaries
even if there are no nodes on it. This enables the use of structured grids and any scheme for
representing the solution where boundary conditions cannot be specified by assigning nodal values.
For example, most meshless methods use MLS approximation which does not satisfy the Kronecker
delta condition and therefore fixing the nodal values along the boundary is not sufficient to enforce
the boundary condition. Similarly, B-spline based representations of trial functions also pose
difficulty in applying essential boundary conditions due to the same reason. In all these cases, the
IBM described here is a promising approach.

REFERENCES
1. Monaghan JJ. An introduction to SPH. Computer Physics Communications 1987; 48(1):89–96.
2. Nayroles B, Touzat G, Villon P. Generalizing the finite element method: diffuse approximation and diffuse
elements. Computational Mechanics 1992; 10:307–318.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
1446 A. V. KUMAR, S. PADMANABHAN AND R. BURLA

3. Belytschko T, Lu YY, Gu L. Element free Galerkin methods. International Journal for Numerical Methods in
Engineering 1994; 37:229–256.
4. Krongauz Y, Belytschko T. Enforcement of essential boundary conditions in meshless approximations using finite
elements. Computer Methods in Applied Mechanics and Engineering 1996; 131:133–145.
5. Atluri SN, Zhu TL. The meshless local Petrov–Galerkin (MLPG) approach for solving problems in elasto-statics.
Computational Mechanics 2000; 25:169–179.
6. Liu GR, Gu YT. A point interpolation method for two-dimensional solids. International Journal for Numerical
Methods in Engineering 2001; 50:937–951.
7. Melenk JM, Babuska I. The partition of unity finite element method: basic theory and applications. Computer
Methods in Applied Mechanics and Engineering 1996; 139:289–314.
8. Belytschko T, Krongauz Y, Organ D, Fleming M, Krysl P. Meshless methods: an overview and recent developments.
Computer Methods in Applied Mechanics and Engineering 1996; 139:3–47.
9. De S, Bathe KJ. Method of finite spheres. Computational Mechanics 2000; 25:329–345.
10. De S, Bathe KJ. The method of finite spheres with improved numerical integration. Computers and Structures
2001; 79:2183–2196.
11. Macri M, De S. Towards an automatic discretization scheme for the method of finite spheres and its coupling
with the finite element method. Computers and Structures 2005; 83:1429–1447.
12. Sukumar N, Moran B, Yu Semenov A, Belikov VV. Natural neighbor Galerkin methods. International Journal
for Numerical Methods in Engineering 2001; 50:1–27.
13. Dolbow J, Belytschko T. Numerical integration of the Galerkin weak form in meshfree methods. Computational
Mechanics 1999; 23:219–230.
14. Kantorovich LV, Krylov VI. Approximate Methods for Higher Analysis. Interscience Publishers: New York, 1958.
15. Hollig K. Finite Element Methods with B-splines. SIAM: Philadelphia, 2003.
16. Rvachev VL, Shieko TI. R-functions in boundary value problems in mechanics. Applied Mechanics Reviews
1995; 48:151–188.
17. Shapiro V, Tsukanov I. Meshfree simulation of deforming domains. Computer-Aided Design 1999; 31:459–471.
18. Shapiro V. Theory of R-functions and applications: a primer. CPA Technical Report CPA88-3, Cornell
Programmable Automation. Sibley School of Mechanical Engineering, Ithaca, NY, 1998.
19. Belytschko T, Parimi C, Moes N, Sukumar N, Usui S. Structured extended finite element methods for solids
defined by implicit surfaces. International Journal for Numerical Methods in Engineering 2003; 56:609–635.
20. Moes N, Béchet E, Tourbier M. Imposing Dirichlet boundary conditions in the extended finite element method.
International Journal for Numerical Methods in Engineering 2006; 67:1641–1669.
21. Laguardia JJ, Cueto E, Doblare M. A natural neighbour Galerkin method with quadtree structure. International
Journal for Numerical Methods in Engineering 2005; 63:789–812.
22. Clark BW, Anderson DC. Finite element analysis in 3D using penalty boundary method. Proceedings of ASME
International Design Engineering Technical Conferences, Montreal, Canada, vol. 2, 29 September –2 October
2002; 289–297.
23. Mortenson ME. Geometric Modeling. Wiley: New York, 1997.
24. Bloomenthal J. Introduction to Implicit Surfaces. Morgan Kaufmann: San Francisco, CA, 1997.
25. Pasko A, Adzhiev V, Sourin A, Savchenko V. Function representation in geometric modeling: concepts,
implementation and applications. The Visual Computer 1995; 11(8):429–446.
26. Shapiro V, Tsukanov I. Implicit functions with guaranteed differential properties. Proceedings of the 5th ACM
Symposium on Solid Modeling and Applications, Ann Arbor, MI, 1999; 258–269.
27. Kumar AV, Lee JH. Step function representation of solid models and application to mesh free engineering
analysis. Journal of Mechanical Design 2006; 128(1):46–56.
28. Chang Y, Hou T, Merriman B, Osher S. A level set formulation of Eulerian interface capturing methods for
incompressible fluid flows. Journal of Computational Physics 1996; 124:449.
29. Kohno H, Tanahashi T. Numerical analysis of moving interfaces using a level set method coupled with adaptive
mesh refinement. International Journal for Numerical Methods in Fluids 2004; 45(9):921–944.
30. Osher S, Fedkiw R. Level Set Methods and Dynamic Implicit Surfaces. Springer: Philadelphia, 2002.
31. Wang X, Wang MY, Guo D. Structural shape and topology optimization in a level-set-based framework of region
representation. Structural and Multidisciplinary Optimization 2004; 27:1–19.
32. Wang MY, Wang X, Guo D. A level set method for structural topology optimization. Computer Methods in
Applied Mechanics and Engineering 2003; 192:227–246.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme
IMPLICIT BOUNDARY METHOD FOR FINITE ELEMENT ANALYSIS 1447

33. Kumar AV, Lee JH, Burla RK. Implicit solid modeling for mesh free analysis. Proceedings of International
Design Engineering Technical Conferences & Computers in Engineering IDETC/CIE, Long Beach, CA, U.S.A.,
2005.
34. Rvachev VL, Sheiko TI, Shapiro V, Tsukanov I. Transfinite interpolation over implicitly defined sets. Computer-
Aided Geometric Design 2001; 18:195–220.
35. Hughes TJR. The Finite Element Method: Linear Static and Dynamic Finite Element Analysis. Dover: New York,
2000.
36. Boresi AP, Schmidt RJ, Sidebottom OM. Advanced Mechanics of Materials. Wiley: New York, 2003.
37. Amestoy PR, Davis TA, Duff IS. An approximate minimum degree ordering algorithm. SIAM Journal on Matrix
Analysis and Applications 1996; 17(4):886–905.
38. Nishimura N. Fast multipole accelerated boundary integral equation methods. Applied Mechanics Review 2002;
55(4):299–324.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2008; 74:1421–1447
DOI: 10.1002/nme

You might also like