INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng 2004; 61:990–1008 (DOI: 10.1002/nme.1097)

The fractal finite element method for unbounded problems

A. Y. T. Leung1, 2, ∗, † , H. Dai2 , S. L. Fok2 and R. K. L. Su3

1 Building and Construction Department, City University, Hong Kong, China
2 School of Engineering, University of Manchester, U.K.
3 Civil Engineering Department, University of Hong Kong, Hong Kong, China

SUMMARY
The fractal finite element method, previously developed for stress intensity factor calculation for crack
problems in fracture mechanics, is extended to analyse some unbounded problems in half space. The
concepts of geometrical similarity and two-level finite element mesh are applied to generate an infinite
number of self-similar layers in the far field with a similarity ratio bigger than one; that is, one layer
is bigger than the next in size but of the same shape. Only conventional finite elements are used
and no new elements are generated. The global interpolating functions in the form of a truncated
infinite series are employed to transform the infinite number of nodal variables to a small number
of unknown coefficients associated with the global interpolating functions. Taking the advantage of
geometrical similarity, transformation for one layer is enough because the relevant entries of the
transformed matrix after assembling all layers are infinite geometric series of the similarity ratio and
can be summed analytically. Accurate nodal displacements are obtained as shown in the numerical
examples. Copyright 䉷 2004 John Wiley & Sons, Ltd.

KEY WORDS: fractal finite element; unbounded problem; stress intensity factor; elastic half space;
concentrated load; ring load

1. INTRODUCTION

In engineering practices, one is often faced with problems where the media bounding the
structure is much larger than the domain of interest, such as those that arise in soil–structure
and fluid–structure interactions. These include an electromagnetic field around an electrical
machine, an aeroplane flying in the air, an island in the stream of water waves and a building

∗ Correspondence to: A. Y. T. Leung, Department of Building and Construction, City University of Hong Kong,
Tatchee Avenue, HKSAR, China.
† E-mail: [email protected]

Contract/grant sponsor: Hong Kong Research Grant Council; contract/grant numbers: CityU 1047/00E and HKU
7014/00E
Received 14 March 2003
Published online 24 August 2004 Revised 1 December 2003
Copyright 䉷 2004 John Wiley & Sons, Ltd. Accepted 12 February 2004
 THE FRACTAL FINITE ELEMENT METHOD 991

∞
Infinite Elements

Rigid boundary of FE Mesh

(a) (b)

Figure 1. Treatment of exterior problems: (a) with rigid truncated boundary;

and (b) with infinite elements.

supported by the ground. In all these cases, the surrounding media can be treated as infinite
and these are collectively called unbounded domain problems. If conventional finite element
method is used to tackle these problems, the finite element mesh, i.e. the computational domain,
must be terminated somewhere short of infinity.
Various techniques have been used to treat unbounded problems with various degrees of
sophistication. The crudest model is the rigid boundary mesh truncation illustrated in Figure 1.
It is efficient and easy to program but may be inaccurate. For many static problems, simple
truncation at a rigid artificial boundary may work satisfactorily, but it may be unclear where such
boundary should be placed. A computational compromise is often at work in these situations.
If the truncation boundary is placed too near, computational effort is saved but the solution
accuracy can suffer; and if the truncation boundary is placed far away, solution accuracy is
improved but the computation expense may become excessive. For dynamic problems, a damper
can be applied to the boundary of the finite element mesh.
One of the most rigorous methods is the boundary integral method, which treats the boundary
with exact integro-differential equations. For example, Cruse [1] applied this method to 3D stress
analysis, Wood [2] to ground water flow analysis and Lysmer and Kuhlemeyer [3] to dynamics
of structure–soil interaction. However, these methods tend to destroy the banded nature of the
system matrix and are comparably difficult to program.
Other methods combine the efficiency and accuracy of analytical solutions and the agility
of standard finite element method. The concept of global–local interpolating functions was
first suggested by Mote [4]. It was recognized that, while the shape functions within a finite
element reduced the infinite number of degrees of freedom of a continuum to a finite number
of degrees of freedom associated with the nodal variables, the global interpolating functions for
the nodal variables could also reduce the number of unknowns drastically. He combined both
the global and local dependent variables and coupled the conventional finite element method
and the Ritz method. New elements were required to be generated, with global interpolating
functions being introduced during element formulation and the unknowns involving both global
and local variables simultaneously.
Bettess [5] used a series of shape functions that were analogous to Langrange polynomials,
but included an exponential decay term, to formulate infinite elements for unbounded problems.

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
992 A. Y. T. LEUNG ET AL.

Lynn and Hadid [6] presented a family of 1/r n type infinite elements for the analysis of elastic
half-space static problems, while infinite elements for acoustic and wave problems have been
proposed by Astley [7] and Burnett [8]. A detailed review of infinite elements for static problems
has been provided by Bettess and Bettess [9].
The idea of using an infinite self-similar mesh with conventional finite elements was first
introduced by the mathematicians Ying [10] and Thatcher [11]. The nodal displacements between
adjacent layers of self-similar elements are assumed to be related by a constant transfer matrix.
The method requires solving a second-degree n-order matrix difference eigen-problem and has
been applied to unbounded problems as well as the determination of stress intensity factors
of cracks. The second-degree n-order matrix difference eigen-problem is solvable only if the
stiffness matrix satisfies certain requirements. Therefore, the method is not applicable to general
2D and 3D problems.
The fractal finite element (FFE) method also makes use of self-similar finite elements.
It was proposed by Leung et al. [12, 13] to model structural systems and crack problems.
Similar to Mote’s method [4], FFE is also a semi-analytical method using both local and
global interpolating functions. The major difference between Mote’s method and FFE is that
the former method uses the global interpolating functions to form the element matrix by
integration while the latter uses the same functions to reduce the problem size by transforming
the infinite number of nodal displacements to a finite set of generalized co-ordinates.
The objective of this paper is to extend the fractal finite element method to static unbounded
problems. As a first attempt, axisymmetric problems will be considered. For such problems,
the global interpolating functions can be obtained by solving the relevant Legendre’s equation
[14]. General 1/r n series solutions with Legendre polynomials are available in the semi-infinite
domain. Therefore, the nodal displacements in the unbounded domain of fractal finite elements
(FFE) can be represented by a set of generalized co-ordinates of the global interpolating
functions. The stiffness matrix associated with the generalized co-ordinates is obtained by
matrix transformation. Because the general solution is of the type 1/r n , which diminishes very
rapidly when r approaches to infinity, a small number of such terms is enough for an accurate
simulation of the far field. This kind of truncation of the series reduces the unknowns in the
unbounded region to a small set of coefficients, thus making the method very efficient. A
numerical example is given in this paper.

2. FFE MESH

Consider a semi-infinite problem, the half-space is divided into the regular FE domain R1
and the unbounded fractal finite element (FFE) domain R2 by the boundary S as shown in
Figure 2. In the regular FE region, conventional finite elements are used and the nodal dis-
placements are the unknowns. In the FFE region extending to infinity, the fractal finite element
technique is adopted. Taking the origin as the centre of similarity and using
( > 1) as the
similarity ratio (or proportional constant), an infinite set of curves, similar to S but with ratios

1 ,
2 ,
3 , . . . is generated in the unbounded domain. A set of straight lines that emanate from
the origin and pass through the corner nodes on S is then created, dividing the unbounded
region into a fractal mesh of similar elements. All nodes on S are called the master nodes
while those beyond are called the slave nodes.

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
 THE FRACTAL FINITE ELEMENT METHOD 993

Master Nodes
Slave Nodes
S

First Layer

Second Layer
R1
Third Layer

Etc.

R2
Infinite Layers

Figure 2. Fractal finite mesh for unbounded domain.

3. GLOBAL INTERPOLATION FUNCTIONS FOR TORSIONLESS AXISYMMETRY

In cylindrical co-ordinates r, , z, with the corresponding displacement components u, v, w,

the component v vanishes and u, w are independent of  when the problem is axisymmetric.
Introducing a stress function , the general equilibrium equation that needs to be solved
is [14]

2 2

2 2

* 1 * * *  1 * * 
+ + 2 + + 2 = ∇ 2∇ 2 = 0 (1)
*r 2 r *r *z *r 2 r *r *z

where

2
* 
2Gu = − (2)
*r*z
2
* 
2Gw = 2(1 − )∇ 2  − (3)
*z2
E
G= (4)
2(1 + )

G being the shear modulus, E the Young’s modulus and  the Poisson ratio.
If the spherical polar co-ordinates R and  (Figure 3) are used instead, where

R 2 = r 2 + z2 (5)
z
 = arctan (6)
r

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
994 A. Y. T. LEUNG ET AL.

O r

Figure 3. Spherical and cylindrical co-ordinates for axisymmetric problems.

it can be shown that

2 2 2
* * * 1 * 1 *
+ 2= 2+ + 2 (7)
*r 2 *z *R R *R R *2
 
1 * 1 * sin  * 1 * tan  *
= cos  − = − (8)
r *r R cos  *R R * R *R R 2 *

and Equation (1) becomes

2 2

2 2

* 2 * 1 * 1 * *  2 * 1 * 1 * 
+ − tan  + + − tan  + =0
*R 2 R *R R 2 * R 2 *2 *R 2 R *R R 2 * R 2 *2

(9)

Let us consider the solutions of the Laplace equation

2 2

*  2 * 1 * 1 * 
+ − tan  + =0 (10)
*R 2 R *R R 2 * R 2 *2
which are automatically the solutions of Equation (9). A particular solution of Equation (10)
has the form
(a) n = R −n−1 n
where n is a function of the angle  only. Substituting (a) into Equation (10), we find for
n the following ordinary differential equation:
 
1 d dn
(b) cos  + n(n + 1)n = 0
cos  d d
Letting x = sin , we have
d 2 n dn
(1 − x 2 ) − 2x + n(n + 1)n = 0 (11)
dx 2 dx

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
 THE FRACTAL FINITE ELEMENT METHOD 995

This is Legendre’s equation, for which, the Legendre polynomials Pn (x) are one of the two
basic solutions:

1 dn
Pn (x) = [(x 2 − 1)n ] (12)
2n n! dx n

2k + 1 k
Pn+1 (x) = xPn (x) − Pn−1 (x) (13)
k+1 k+1

Using Pn (x) for n , n = 0, 1, 2, 3, 4, 5 . . . , we obtain the following set of solutions of Equa-

tion (10).

0 =A0 /R

1 =A1 x/R 2

(3x 2 − 1) 1
2 =A2
2 R3
(5x 3 − 3x) 1
3 =A3
(c) 2 R4
(35x 4 − 30x 2 + 3) 1
4 =A4
8 R5
(63x 5 − 70x 3 + 15x) 1
5 =A5
8 R6
..
.

which are also solutions of Equation (9). From them we can get further solutions of Equation
(9), which are not solutions of Equation (10). If R −n−1 n is a solution of Equation (10), it
can be shown that R −n+1 n is a solution of Equation (9). Therefore, multiplying expressions
(c) by R 2 , we obtain another set of solutions of Equation (9), namely,

0 =B0 R

1 =B1 x

(3x 2 − 1) 1
2 =B2
2 R

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
996 A. Y. T. LEUNG ET AL.

(5x 3 − 3x) 1
3 =B3
2 R2
(35x 4 − 30x 2 + 3) 1
4 =B4
(d) 8 R3
(63x 5 − 70x 3 + 15x) 1
5 =B5
8 R4
..
.
Each of expressions (c) and (d), and any linear combination of them, can be taken as a stress
function, and by a suitable adjustment of the constants A1 , A2 , A3 , A4 . . . and B1 , B2 , B3 , B4 . . .
we can obtain solutions of many different problems.
Substituting the general expression for  into Equations (2) and (3), we have

* (R −n−1 Pn (x)) * (R −n+1 Pn (x))

2 2

2Gu = − An + Bn (14)
n=0 *r*z *r*z

* (R −n−1 Pn (x)) * (R −n+1 Pn (x))

2 2

2Gw = − An + Bn (15)
n=0 *z2 *z2

From (5) and (6), we can easily show that

2 2

2
  
* * 1 * 1 * * 1 *
(e) = cos 
2
+ sin 
2
+ − 2 sin  cos 
*z2 *R 2 R *R R 2 *2 *R R *
2

2 2
  
* 1 * 1 * * * 1 *
(f) − = sin  cos  + − − (cos  − sin )
2 2
*r*z R *R R 2 *2 *R 2 *R R *

Thus
∞ R −n−3 ∞ R −n−1
u= An G11 (n, ) + Bn G12 (n, ) (16)
n=0 2G n=0 2G

∞ R −n−3 ∞ R −n−1
w= An G21 (n, ) + Bn G22 (n, ) (17)
n=0 2G n=0 2G

where Gij (n, ), with i, j = 1, 2, are the angular functions as shown below

(g) G11 (n, ) = sin  cos [−(n + 1)(n + 3)Pn (x) + Pn ] + (cos2  − sin2 )(n + 2)Pn
(h) G12 (n, ) = sin  cos [−(n − 1)(n + 1)Pn (x) + Pn ] + (cos2  − sin2 )nPn
(i) G21 (n, ) = (n + 1)(n + 2) sin2 Pn (x) + cos2 [−(n + 1)Pn (x) + Pn ]
−2(n + 2) cos  sin Pn

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
 THE FRACTAL FINITE ELEMENT METHOD 997

( j) G22 (n, ) = (n − 1)n sin2 Pn (x) + cos2 [−(n − 1)Pn (x) + Pn ] − 2n cos  sin Pn

with

*Pn (x) *x *x
Pn = · = Pn(1) (x) ·
*x * *
 2
*P 
2
* x *x
Pn = n = Pn(1) (x) · + Pn(2) (x) ·
* *2 *

Many problems of solids of revolution deformed by loads symmetrical about the axis of
symmetry have Equations (16) and (17) as their solutions. For FFE, truncations of these
expressions will be used as the interpolating functions.

4. THE REGULAR REGION

In the regular region R1 in Figure 2, for m-noded linear continuum axisymmetric elements, the
shape functions and the local interpolation for the element displacements may be expressed in
an isoparametric form as

m 
m
u= Ni (, )ui , w= Ni (, )wi (18)
i=1 i=1

for displacement interpolation, and


m 
m
r= Ni (, )ri , z= Ni (, )zi (19)
i=1 i=1

for geometric relations. {u, w} and {r, z} are the nodal displacements and nodal co-ordinates of
an element, respectively, and Ni (, ) are the shape functions. The resulting element stiffness
matrix equation is

Ku = f (20)

where K is the stiffness matrix, u is the nodal displacement vector and f is the nodal force
vector. The stiffness matrix K is determined from the following expression:
 1  1
K= BT DB(2 · r)det J d · d (21)
−1 −1

Here J denotes the Jacobian matrix of {r, z} with respect to {, }

r, z,
J= (22)
r, z,

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
998 A. Y. T. LEUNG ET AL.

B = [B1 , B2 , . . . , Bn ] is the strain–displacement operator where Bi is given by Equation (23),

such that
= Bu, while D is the material properties matrix, such that  = D
:
 
*Ni  
 *r 0  D D D 0
 
   
 *Ni  D D D 0 
 0   
 *z   
Bi =   , D = D D D 0  (23)
 Ni   
 0   
 r   1− 
 
 *Ni *Ni  0 0 0 D
2
*z *r

where D = E/(1 − 2 ).
The stiffness matrix equation for the regular FEM domain can be assembled and then
partitioned as
   
Kim Kim ui fi
= (24)
1 um 1
Kmi Kmm fm

where i denotes the inner nodes of the FEM domain, m the master nodes on S and the
1
superscript ‘1’ in Kmm 1 indicates the regular domain R .
and fm 1

5. THE UNBOUNDED REGION

In the unbounded region R2 , each layer is created from the adjacent inner layer using the same
similarity ratio throughout. The infinite number of nodal displacements will be transformed
into a finite set of generalized co-ordinates using the global interpolation functions derived in
Section 3. Since there is a simple relation between the stiffness matrices of similar elements,
this transformation can be performed at the local element level and the results summed up as a
geometrical progression series to be assembled into the global stiffness matrix. The contributions
to the global stiffness matrix from the infinite number of elements in the unbounded domain
are therefore fully accounted for while the number of degrees of freedom remains finite.
To begin with, the stiffness matrix K of the first layer is partitioned with respect to the
slave nodes (s) and master nodes (m):
2    2
Kmm Kms um fm
= (25)
Ksm Kss us fs

2
The superscript ‘2’ in Kmm and fm2 indicates the unbounded domain R .
2
Using Equations (16) and (17), one can transform the nodal displacements u to the gener-
alized co-ordinates a as

u = Ta (26)

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
 THE FRACTAL FINITE ELEMENT METHOD 999

where T is the transformation matrix evaluated from Equations (16) and (17). For the first
layer, only the displacements of the slave nodes are transformed so as to provide continuity
with the regular region. Thus
   
um I 0 um
= (27)
us 0 Ts a
where Ts is the transformation matrix for the slave nodes. Substituting Equation (27) into (25)
gives
2    2
Kmm Kms Ts um fm
= (28)
Ksm Kss Ts a fs
 
Pre-multiplying the equation by the transformation matrix 0I T0T on both sides then gives
s

2    2 
Kmm Kms Ts um fm
= (29)
TTs Ksm TTs Kss Ts a TTs fs
Similarly, for the second and subsequent layers in the unbounded domain, i.e. k
2, we have
TTk Kk Tk a = TTk fk (30)
Therefore, the stiffness matrix of the unbounded domain R2 can be assembled as
 2   
Kmm Kms Ts    2
fm 
u  
  m
 T ∞  = 
∞ (31)
Ts Ksm TTs Kss Ts + TTk Kk Tk a 
TTs fs + TTk fk 

k=2 k=2

6. THE GLOBAL STIFFNESS MATRIX AND FRACTAL TRANSFORMATION

By combining Equations (24) and (31), one has the equation of the complete half-space
containing both domains R1 and R2
   
Kii Kim 0    fi 
 ui 
 

  
 
 
 

Kmi Kmm 1
+ K 2
K T  f m
 mm ms s  um = (32)
    
 ∞  
 
 ∞ 

0 TTs Ksm TTs Kss Ts + TTk Kk Tk a 
TTs fs + TTk fk 

k=2 k=2

where i represents the inner nodes of the FEM domain, m the master nodes, s the slave nodes,
k indicates the kth layer of mesh beyond the boundary S, and T is the transformation
∞ matrix.
∞ T T
To solve Equation (32), one has yet to calculate the sum k=2 Tk Kk Tk and k=2 Tk fk .
Consider two elements, denoted by 1 and 2, which are similar in the r and z directions with

being the proportional constant. By Equations (22) and (23), one has
det J2 =
2 det J1 (33)

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
1000 A. Y. T. LEUNG ET AL.

and
1
B2 = B1 (34)


Therefore, for a continuum axisymmetric element
 1 1
K2 = BT2 DB2 (2 · r2 )det J2 d · d =
· K1 (35)
−1 −1

It follows that
Kk =
k−2 · K2 (36)
and
Tk = T2 Diag[
i ] = T2 Diag[
−(k−2)ni ] (37)
where

i =
−(k−2)ni (38)
is a scaling function and
i + 5

 for i = 1, 3, . . .
ni = 2 (39)

 i
for i = 2, 4, . . .
2
So,

∞ 
∞
TTk Kk Tk = Diag[
i ]T TT2
k−2 K2 T2 Diag[
j ] = [¯
ij k̄ij ] (40)
k=2 k=2

where

∞

¯ ij =
−(k−2)ni ·
(k−2) ·
−(k−2)nj = [1 −
−(ni +nj −1) ]−1 (41)
k=2

and
k̄ij = TT2 K2 T2 (42)
∞
k=2 Tk Fk . Hence, one
A similar procedure can be applied to the global generalized force T

can fully account for the contributions from all the layers in the infinite domain.

7. NUMERICAL EXAMPLES

Example 1 (The Boussinesq Problem)

To verify the method, FFE was applied to the problem of a point load in a half-space. The
results were compared with those from ABAQUS [15] using both infinite elements and a

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
 THE FRACTAL FINITE ELEMENT METHOD 1001

p
p
r
r

4.0

5.0
Zero displacement
imposed at radius 5.0
4.0

z
z

Figure 4. Finite element mesh with rigid Figure 5. Finite element mesh with infinite
truncated boundary. elements.

truncated rigid boundary. The analytical solution due to Boussinesq [14] was also used for
comparison.

Similar mesh configurations were used for all three FE models. The finite element mesh
for the rigid boundary model in Figure 4 is made up of 16 four-noded axisymmetric elements
(CAX4) [15] truncated at a radius of 5.0, where fully fixed boundary conditions are applied.
The infinite element mesh in Figure 5 is similar to that of the rigid boundary model. The
regular inner region is composed of the same twelve axisymmetric elements extending to a
radius of 4.0, but the outermost layer of elements are now replaced by four infinite elements
(CINA4) [15] to model the far-field domain. Similarly, the FFE mesh in Figure 6 has the
far-field domain being modelled by an infinite number of self-similar fractal finite elements.
For convergence study of the FFE method, we considered two other coarser meshes. Together
with the one already described, these are arbitrarily labelled as coarse, medium and fine,
respectively, as shown in Figure 7 for the regular region. Both fractal and infinite elements
were used in conjunction with these meshes to analyse the Boussinesq problem, and their speed
of convergence compared. The number of elements and number of nodes for these models are
detailed in Table I.
The following data were used throughout the analysis unless otherwise stated:
(i) The material is linearly elastic, with Young’s modulus E = 1.0 and Poisson’s ratio  = 0.1
(ii) A unit load is applied, i.e. p = 1.0
Boussinesq’s analytical solution for the problem of a point load in half-space gives the vertical
displacement as [14]
p
w= [(1 + )z2 (r 2 + z2 )−3/2 + 2(1 − 2 )(r 2 + z2 )−1/2 ] (43)
2 E

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
1002 A. Y. T. LEUNG ET AL.

4.0

1st layer
2nd layer
3rd layer
4th layer

Etc.

Figure 6. Finite element mesh with fractal finite elements.

Figure 7. Meshes used in convergence study (regular region).

Table I. Details of finite element meshes.

Total Number of Number of slave
Total number number of elements in 1st nodes in 1st
Mesh of elements nodes layer layer

Coarse 3 7 2 3
Medium 8 14 4 5
Fine 12 19 4 5

where r and z are the radial and vertical distances from the point load, respectively.
This equation clearly shows the 1/r singularity at the point of application of the load (r = 0).
Here one has the displacement variation along the vertical line beneath the point

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
 THE FRACTAL FINITE ELEMENT METHOD 1003

Figure 8. The Boussinesq problem: displacement results at r = 0 (FFE: 6

transformation terms and
= 1.09).

load as

0.49
w= (44)
z

Figure 8 shows the displacement results obtained from the three models in Figures 4–6. The
analytical solution from Equation (44) is also shown for comparison. Here, the similarity
constant
= 1.09 was used for the construction of the FFE mesh and the series solutions of
Equations (16) and (17) were truncated at the 6th term. As can be seen, the FFE results are
very similar to those from the infinite elements, and both sets of results are in good agreement
with the analytical solution for the depth considered. The rigid boundary model, as expected,
performed the worst.
Figure 9 shows the convergence of the FFE results with the number of transformation terms.
The mesh used was that of Figure 6 and the similarity ratio employed was again 1.09. The
results show that the displacements in the area of interest approach the analytical solution when
the number of transformation terms increases from 2 to 5.
The effect of the similarity ratio on the accuracy of the FFE solution is shown in
Figure 10. Here, the medium mesh in Figure 7 was used together with eight transforma-
tion terms. The results show that, the closer to unity is the similarity ratio, i.e. the larger

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
1004 A. Y. T. LEUNG ET AL.

Figure 9. Convergence of FFE results with number of transformation terms (
= 1.09).

the number of fractal elements used per radial distance in the unbounded domain, the more
accurate is the solution. Similar observations have been made in the analysis of crack problems
using FFE [16]. Since the value of the similarity ratio has no effect on the computational costs,
a value close to unity should be the automatic choice.
Figures 11 and 12 show the respective convergence of the infinite element and FFE results
with mesh density. Eight transformation terms and
= 1.25 were used with the FFE method.
As can be seen, despite the coarseness of the meshes used (Figure 7), both methods can capture
the far-field displacements accurately. The results become less accurate as one approaches the
loading point because of the singularity. However, this is not a problem for FFE which is well
suited for modelling this type of problems as well, as demonstrated in Reference [17]. Thus,
it is possible to obtain accurate results for the whole semi-infinite domain by creating just one
layer of finite elements explicitly, with the regular region within the curve S being modelled
with self-similar fractal finite elements with decreasing size, i.e. using a similarity ratio less
than unity.
Figure 13 shows the reduction of the fractional errors with increasing number of degrees
of freedom for both the FFE and infinite element methods, using displacement results at the
sampling point (0, −0.75) in the regular region. As can be seen, FFE is as efficient as, if not
better than, ABAQUS’ infinite elements. The FFE results were obtained by using a similarity
ratio of 1.25. As indicated in Figure 10, even better results could be achieved, without incurring
additional computational costs, if a smaller similarity ratio had been used.

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
 THE FRACTAL FINITE ELEMENT METHOD 1005

Figure 10. Effect of similarity ratio on FFE results (medium mesh with 8 transformation terms).

Example 2 (An elastic half-space subjected to a concentrated ring load)

As a second example, the surface deflection of an elastic half-space under a total ring load P
with radius r0 is considered. The non-dimensional surface deflection in terms of Er0 /(P (1−2 ))
is shown in Figure 14. In the regular region, a FE mesh including 44 elements is used for the
analysis. The calculated results are very close to the exact solution [14].

8. CONCLUSION

The fractal finite element method has been extended to static, axisymmetric unbounded prob-
lems. The far-field behaviour is modelled by using an infinite number of self-similar con-
ventional finite elements. No new elements are required. A general solution with Lengen-
dre polynomials is taken as the global interpolating functions to condense the infinite num-
ber of nodal displacements in the far field to a finite set of generalized co-ordinates by
means of the fractal transformation technique. Accurate numerical results are obtained for the
Boussinesq problem. For optimal performance, a similarity ratio close to unity should be used
with FFE.

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
1006 A. Y. T. LEUNG ET AL.

Figure 11. Convergence with mesh density using infinite elements.

Figure 12. Convergence with mesh density using FFE (8 transformation terms and
= 1.25).

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
 THE FRACTAL FINITE ELEMENT METHOD 1007

Figure 13. Convergence with DOF (FFE:
= 1.25).

Figure 14. Surface deflection due to a ring load (FFE: 8 transformation terms and
= 1.20).

ACKNOWLEDGEMENTS
The research is supported by the Hong Kong Research Grant Council #CityU 1047/00E and HKU
7014/00E.

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008
1008 A. Y. T. LEUNG ET AL.

REFERENCES
1. Cruse TA. Application of the boundary-integral equation method to three dimensional stress analysis.
Computers and Structures 1973; 3:509–527.
2. Wood WL. On the finite element solution of an exterior boundary value problem. International Journal for
Numerical Methods in Engineering 1976; 10:885–891.
3. Lysmer J, Kuhlemeyer RL. Finite dynamic model for infinite media. Journal of Engineering Mechanics
ASCE 1969; 859–877.
4. Mote CD. Global–local finite element. International Journal for Numerical Methods in Engineering 1971;
3:565–574.
5. Bettess P. Infinite elements. International Journal for Numerical Methods in Engineering 1977; 11:53–64.
6. Lynn PP, Hadid HA. Infinite elements with 1/r n type decay. International Journal for Numerical Methods
in Engineering 1981; 17:347–355.
7. Astley RJ. Infinite elements for wave problems: a review of current formulations and an assessment of
accuracy. International Journal for Numerical Methods in Engineering 2000; 49:951–976.
8. Burnett DS, Holford RL. An ellipsoidal acoustic infinite element. Computer Methods in Applied Mechanics
and Engineering 1998; 164:49–76.
9. Bettess P, Bettess JA. Infinite Elements for Static Problems. Engineering Computations 1984; 1:4–16.
10. Long-an Ying. Infinite Element Method. Peking University Press: Peking; Vieweg Publishing: Braunschweig,
1995.
11. Thatcher RW. Singularities in the solution of Laplace’s equation in two dimensions. Journal of the Institute
of Mathematics and its Applications 1975; 16:303–319.
12. Leung YT, Cheung YK. Dynamic analysis of frames by a two-level finite element method. Journal of Sound
and Vibration 1981; 74(1):1–9.
13. Leung AYT, Su RKL. Mode I crack problems by fractal two level finite element methods. Engineering
Fracture Mechanics 1994; 48(6):847–856.
14. Timoshenko SP, Goodier JN. Theory of Elasticity. McGraw-Hill: New York, 1970.
15. ABAQUS User’s Manual, Version 6.1, Hibbit, Karsson and Sorenson, Inc.: Pawtucket, USA, 2000.
16. Xie JF, Fok SL, Leung AYT. A parametric study on the fractal finite element method for two-dimensional
crack problems. International Journal for Numerical Methods in Engineering, to appear.
17. Leung AYT, Wong SC. Two-level finite element method for thin plates subject to concentrated loads.
Microcomputers in Civil Engineering 1988; 3(2):127–136.

Copyright 䉷 2004 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2004; 61:990–1008