5th National Congress on Civil Engineering,

May 4-6, 2010, Ferdowsi University of Mashhad, Mashhad, Iran

Application of higher-order elements in scaled boundary finite

element method (SBFEM) to improve its accuracy and
efficiency

M.I. Khodakarami1, N. Khaji2

1- Ph.D. Student, Department of Civil Engineering, Tarbiat Modares University, Tehran, Iran
2- Assistant Professor, Department of Civil Engineering, Tarbiat Modares University, Tehran, Iran

khodakarami@modares.ac.ir
nkhaji@modares.ac.ir

Abstract
The SBFEM is a semi-analytical fundamental-solution-less boundary-element method based only on
finite elements, which combines advantages of the finite element and boundary element methods. This
method has numerical solution on the boundary and analytical solution on the domain of the problem. In
this paper, the Chebyshev shape functions with Clenshaw-Curtis quadrature are applied to the elements
on boundary, which leads to lumped-form and blocky-lumped-form of coefficient matrices and higher-
order elements. A set of wave propagation problems, subjected to various load forms are modeled using
the SBFEM with very small number of degrees of freedom. The numerical results agree very well with
the analytical solutions as well the results from other numerical methods.
Keywords: elastodynamics, scaled boundary finite element method, higher order element,
Clenshaw-Curtis quadrature, chebyshev shape function.

1. INTRODUCTION

Many numerical methods used in elastodynamics, the most widely used computational procedure in solid
mechanics are finite difference method, finite element method, boundary element method, spectral element
method with their own advantages and disadvantages. In computational wave propagation, considerable
efforts have been devoted for developing highly accurate numerical techniques for the solution of elastic
wave equations.
Finite difference methods, which have been widely used in computational elastodynamics, require a
large number of grid points to achieve the expected accuracy, even with high- order explicit or implicit
spatial operators. Free surface boundaries and complex configurations coarse modeling produce lack of
preesion in the simulation of some problems. In practice, a difficult trade-off between numerical dispersion
and computational cost is required [1].
In the finite element methods, the domain is spatially discretized. In each finite element, shape
functions in the form of polynomials interpolate the displacement. Standard numerical integration of regular
functions leads to the static-stiffness and mass matrices that are then assembled enforcing compatibility and
equilibrium. A great flexibility exists in representing the geometry and materials, but low-order finite element
methods exhibit poor dispersion properties. For the unbounded domain, the finite element method cannot
represent the radiation condition at infinity exactly [2].
In the boundary-element method, only the boundary is discretized, leading to a reduction of the spatial
dimension by one. The main advantages are that the solution is south over a domain one dimension lower
than the physical domain, and that the radiation is a priori satisfied. On the other hand, a fundamental
solution satisfying the governing equations in the domain must be available. This analytical solution is often
very complicated exhibiting singularities. Special numerical integration of the polynomials with the
fundamental solution, involving singularities yields non-symmetric coefficient matrices [3].
As already mentioned, the FEM and BEM converge to the exact for decreasing element site, and
special techniques and large number of elements are necessary to achieve high accuracy.
5th National Congress on Civil Engineering,
May 4-6, 2010, Ferdowsi University of Mashhad, Mashhad, Iran

Higher-order methods like spectral methods, that enable to achieve the expected accuracy using few
grid points per wavelength, have also been proposed for elastodynamics Problems. This method is suitable
for modeling complicated geometry and has exponential convergence rate. In SEM, the mass matrices are
diagonal, due to choice of specific shape function and Guass-Lobbato-Legender integration, and few number
of element and unknowns require obtaining high accuracy [4].
Few years ago, a novel semi-analytical procedure called the scaled boundary finite element method
has been developed. This method has both advantages of BEM and FEM, similar to BEM, only the boundary
is discretized, and solution on the domain is analytical. Similar to FEM the solution on the boundary is
numerical and this method can implement to complex geometry an any elastic properties (isotropic,
anisotropic, etc) and no fundamental solution required, the three main disadvantages of the SBFEM is that
there are no diagonal coefficient matrices coupled ODE and numerical disperses due to use of lower order
shape function.
In this paper, a novel procedure is developed to improve accuracy and efficiency of the scaled boundary
finite element method for elastodynamics problems. First, the scaled boundary finite element method is
summarized, then using the Chebyshev polynomials as shape function and evaluating the numerical integral
over the elements using Clenshaw-Curtis quadrature, lumpediform coefficient matrices and higher-order
elements are obtained. Finally, the accuracy and efficiency of the method is fully demonstrated by comparing
results obtained from the SBFEM with those reported in other studies. For this purpose, a set of wave
propagation problems, subjected to various load forms such as triangular load, and ramped load are modeled
using the SBFEM. Furthermore, support motion boundary conditions are examined using the SBFEM. Each
problem is successfully modeled using a very small number of degrees of freedom in comparison with other
numerical methods. The numerical results agree very well with the analytical solutions as well the results
from other numerical methods.

2. SUMMARY OF SCALED BOUNDARY FINITE ELEMENT METHOD

The derivation of the scaled boundary finite element method for elastodynamics problem is detailed in [5]. In
this section, only the concept and the equations that are necessary for explaining the new developments are
summarized. In the scaled boundary finite element method, point O is chosen as scaling center from which
the total boundary must be visible (Figure 1-a). For bounded medium, this point is selected inside the
domain, and for unbounded domain, outside the domain (Figure 1-b).

Figure 1. Special choice of scaling centre: (a) on boundary for bounded medium; (b) on extension of
boundary for unbounded medium

In scaled boundary finite element method, only the boundaries that visible from the scaling center is
discretize (Figure 1). Usually, the origin of the two- dimensional Cartesian coordinate system xˆ , yˆ is
selected at the scaling center. The coordinates of the nodes of an element in the Cartesian coordinate system
are x and y. In addition, the local co-ordinate system is defined using radial coordinate ξ from scaling center
to the boundaries (from ξ=1 to ξ=1 for bounded domain and from ξ=1 to ξ=∞ for unbounded domain) and
tangential coordinate η on the boundaries (-1≤η≤1). The transformation of the geometry from Cartesian
coordinate ( xˆ , yˆ ) to dimensionless coordinate (ξ , η) is performed using the shape function as shape function
[N(η)]. In scaled boundary finite element method, similar to FEM, the shape functions and shape functions
are equaled. The geometry of the elements in new coordinate is introduced by:
x (η ) = [N (η )]{x} (1-a)
y(η ) = [N (η )]{y} (1-b)

2
5th National Congress on Civil Engineering,
May 4-6, 2010, Ferdowsi University of Mashhad, Mashhad, Iran

The scaled boundary transformation relates any point in the domain with the coordinates denoted as
xˆ, yˆ to the corresponding point on the line element as:

x (ξ ,η ) = ξ .x(η ) (2-a)

y(ξ ,η ) = ξ . y(η ) (2-b)
To transform all governing equation to scaled boundary coordinate and produce the solution, standard
procedures as employed in constructing isoparametric finite element are followed. On the boundary, the
Jacobean matrix for the scaled boundary transformation equals:
⎡ x (η ) y (η ) ⎤
J (η ) = ⎢ (3)
⎣ x1η (η ) y1η (η ) ⎥⎦
The spatial derivatives in tow coordinate systems are related as:
⎧∂⎫
⎪⎪ ∂xˆ ⎪⎪ 1 ⎧ y1η (η ) ⎫ ∂ 1 ⎧− y (η )⎫ 1 ∂
⎨ ∂ ⎬= ⎨ ⎬ + ⎨ ⎬
⎪ ⎪ J (η ) − x
⎩ 1η ⎭(η ) ∂ξ J (η ) ⎩ x(η ) ⎭ ξ ∂η (4)
⎪⎩ ∂yˆ ⎪⎭
Long radial lines passing through the scaling center O and a node on the boundary, the nodal displacement
function {u(ξ)} are introduced. The displacement {u(ξ,η)} at a point (ξ,η) are obtained by interpolation of the
displacement function:
{u (ξ ,η )} = [u x (ξ ,η ) u y (ξ ,η ) ] = [N (η )]{u (ξ )}
T
(5)
Using equation 4, the strain {ε(ξ, η)} are obtained:
{ε (ξ ,η )} = [B1 (η )][u,ξ (ξ )]+ 1 [B 2 (η )]{u (ξ )} (6)
ξ
where the matrices [B1(η)] and [B2(η)]are defined as:
[ ]
B1 (η ) = {b1 (η )}[N (η )] (7-a)
[B (η )] = {b
2 2
(η ) [N ,η (η ) ]
} (7-b)
with
⎧ y1η ⎫
{b (η )} =
1 1
⎨ ⎬ (8-a)
J (η ) ⎩− x1η ⎭
1 ⎧− y ⎫
{b 2
}
(η ) = ⎨ ⎬
J (η ) ⎩ x ⎭
(8-b)

According to equations (7), (8); [B2], [B1], {b1} and {b2} only depend on the geometry of the element on
the boundary. Using the Hook's law, the stresses at a point (ξ, η) is defined as:
{σ (ξ ,η )} = [D ]{ε (ξ ,η )} (9)
where [D] is the elasticity matrix of the domain. To derive the governing equation in scaled boundary finite
element method, start from Navier equation that is:
{σ },i + P = ρ {u} (10)
where ρ is the mass density and P is vector of external forces, substituting, equation 9 in 10 and using
weighted residual method, the scaled boundary finite element governing equation in frequency-domain
expressed as:
[E ]ξ
o 2
{u (ξ )},ξξ + ([E 1 ]T − [E 1 ]+ [E 0 ])ξ {u (ξ )},ξ − [E 2 ]{u (ξ )}− [M o ]ω 2 {u (ξ )} = {F (ξ )} (11)
The coefficient matrices [Eo], [E1], [E2] and [Mo] are obtained by assembling the element coefficient
matrices calculated on the boundary. These coefficient matrices are written as:
[E ] = ∫ [B ] [D][B ] J dη
1 T
1 1
o
(12-a)
−1

[E ] = ∫ [B ] [D][B ] J dη
1 T
1 2 1
(12-b)
−1

[E ] = ∫ [B ] [D][B ] J dη
1 T
2 2 2
(12-c)
−1

3
5th National Congress on Civil Engineering,
May 4-6, 2010, Ferdowsi University of Mashhad, Mashhad, Iran

[M ] = ∫ [N ] ρ [N ] J dη
1 T
o
(12-d)
−1

{F (ξ )} = ξ 2 {F (ξ )}+ ξ {F (ξ )}= ξ ∫ [N ] p J dη + ξ ∫ [N ] {b } {σ } J dη
b t 2
1

−1
T 1

−1
T 1 T
(12-e)
where, p is the external body load and {σ } on side of boundary; also for two dimensional problems that have
one dimensional with line elements the matrix [D ] is simplified as shear modules G.
Note that no integrals over the domain are present in equation (11). On the other hand, this method on
the boundaries is numerical and in the domain is analytic. Similar to stiffness and mass matrices in finite
element method [Eo], [E1], [E2] and [Mo] are sparse, hence, equation (11) is a system of coupled differential
equation that only depend on radial coordinate ξ and not dependent on η. Results are obtain by solving
system of equations (11) and employ the initial and boundary conditions. The internal nodal forces on a
surface with constant ξ are obtained by integrating the surface traction over elements. They are expressed as
{Q(ξ )} = [E 0 ]ξ {u,ξ (ξ )}+ [E 1 ]T {u (ξ )} (13)

3. SHAPE FUNCTIONS AND NUMERICAL INTEGRATION

After discretizing the boundary, the next step is simulation of the geometry using shape functions [N (η )] . As
regards, numerical integration method only useful for special functions and weight functions under
integration sign and integration points, for calculate the numerical integration that present in this method,
Chebyshev polynomials used that using Clenshaw-Curtis integration method, the coefficient metrics [E0] and
[M0] been diagonal. This shape function as for Nη+1-nodes element is defined [8]:
Nη
2 1
N (η ) = ∑c c Tn (ηi )Tn (η ) (14)
Nη n =0 i n

where, 0 ≤ i ≤ Nη . There are Nη+1 nodes in the η direction. Ti are the Chebyshev polynomials and ci=1 for
0<i<Nη and 2 for i=0,Nη. Because of this definition, the Chebyshev polynomials are equal to either zero or
one at any given control point:
N αn (η )β = δ αβ (15)
where δ denotes the Kronecker Delta, in this method, the control points ηα, α=0,…,n, needed in the definition
(14) are chosen to be the n+1 Chebyshev-Lobatto-Legender (C-L-L) points, which are [8]:
⎛ kπ ⎞
η k = − cos⎜ ⎟ ; k = 0,1,..., N (16)
⎝N ⎠
where N is total number of the points. As will explained later, this choice is motivated by the fact that the
combination of Clenshaw-Curtis integration method and Chebyshev polynomials leads to exactly diagonal
coefficient matrices [E0] and [M0] and blocky-diagonal coefficient matrices [E1] and [E2].
To calculate the coefficient matrices, (equation (12-a) to (12-b)), integrals need to be evaluated
numerically over each element. As control points are extreme of Chebyshev-Lobatto-Legender polynomials,
the Clenshaw-Curtis integration method will be used, because it leads to diagonal coefficient matrices. When
used in conjunction with CLL interpolation points. In two-dimensional problems, integrations over elements
with boundary Γe hence given:
x
x

∫ f ( )d = ∫ f ( (η )) J (η )dη
1

−1 (17)
Γe

Using Clenshaw-Curtis integration method equation (17) calculated as [8]:

⎡1 1 1 ⎤
∫ f (η )dη = (b − a )⎢⎣ 2 c − ...⎥ = (b − a )W
b c2 k +1
− c3 − c5 − ... − (18)
a
1
3 15 (2k + 1)(2k − 1) ⎦
where cj defined as:
2 n+1 '' ⎛ π ( j − 1)k ⎞
cj = ∑ f (η k )cos⎜⎝ n + 1 ⎟⎠ ;
n + 1 k =0
j = 1,..., N (19)

4
5th National Congress on Civil Engineering,
May 4-6, 2010, Ferdowsi University of Mashhad, Mashhad, Iran

where, double prime in summation of equation (19) denotes that the first and last terms of the series is haled,
also, in equation (18), ηk are C-L-L points. Applying the Clenshaw-Curtis quadrature (equation 18) to
equations (12-a) and (12-d) and using equation (15), the coefficient matrices calculated in lumped-form as:

[M ] = ∑ c N (η
p +1
0
ij k i k )N j (η k )ρ J (η k ) = δ ij ck ρ J (η k ) (20-a)
k =1

[E ] = ∑ c N (η )N (η )[D] J (η
p +1
0
ij k i k j k k ) = δ ij ck D J (η k ) (20-b)
k =1

4. SOLUTION PROCEDURE

To solve the system of second-order ordinary differential equations (11) of order n analytically, a
transformation to first-order ordinary differential equations of order 2n is performed. However, in this paper,
the Frobenius solution procedure used to solve the non-homogeneous differential equations system (11).
Equation (11) is a second-order non-homogeneous differential equation system with respect to radial
coordinate ξ. This procedure explained widely in [6]. Here, the final equations that are useful to find the
solution of the equation (11) in frequency domain will be introduced.

5. NUMERICAL EXAMPLES

The aforementioned methodology has been implemented in a two-dimensional frequency-domain SBFEM

code is provided. In order to validate the nature and general behavior of the method, three numerical
examples have been considered. A plane stress condition is assumed for all the examples. Numerical results
are compared
# with those obtained by exact analytical solutions [9] and/or by other numerical methods to
assess the stability and the accuracy of the present method. No physical damping (i.e., pure elastic material
behavior) is considered in all the SBEFM analyses. All quantities are measured in SI units. In all examples,
the 5-node elements are used to discretizing the boundary.

5.1. Rectangular bar subjected to two different loadings

In the first example, verification of the algorithm especially in comparison with analytical solutions is the
main target. A rectangular bar (Figure 2a), whose length L is twice its width W, is fixed at its left end with
traction free on its top and bottom sides. The Poisson’s ratio is considered null to impose one-dimensional
condition. The right end side of bar is uniformly subjected to two various tensile loadings: a Heaviside step
function representing a suddenly applied load (P0 = 1) as shown in Fig. 1b, and a triangular function which
increases from zero at time t = 0 to P0 at t = 1, and then decreases to zero at t =2 as depicted in Fig. 1c. The
material constants are as follows: the Young's modulus E = 1 and the mass density ρ = 1. These material
properties yield the maximum propagation velocity of stress waves c = 1. In the triangular loading, the
primary wave front just reaches the fixed end at time t = 1, and bounces back to the right end at t = 2.

Figure 2. Rectangular bar subjected to prescribed loading: (a) geometry, boundary conditions and meshing, (b)
Heaviside loading, (c) triangular loading.

The time histories of horizontal displacement at point A and horizontal stress at point B (Figure 2a)
are investigated. The numerical results by the SBFEM are compared with that of analytical solution. Figure
3 shows the horizontal displacement of point A, for the Heaviside step function loading. From this figure,
one can see that the SBFEM give very good results and that the higher-order elements that used in this paper
give the better results. The horizontal stress histories at point B, for the Heaviside step function loading is
shown in Figure 4 Good agreement between the SBFEM results and the analytical solution can be

5
5th National Congress on Civil Engineering,
May 4-6, 2010, Ferdowsi University of Mashhad, Mashhad, Iran

observed, especially with higher-order elements. The FEM results 10 experience small oscillations around
the analytical solutions at moments when the stress jumps. This type of oscillations are caused by the sudden
application of the step load, and observed in other numerical studies such as domain boundary element
method, boundary element method, and hybrid BEMFEM approach.
Figure 5 shows the horizontal displacement of point A, for the triangular loading. As it is obvious
from this figure, the results from the higher-order SBFEM, SEM and analytical approaches are almost
identical. Excellent agreement can also be observed in the horizontal stress histories at point B as shown in
Figure 6.

Figure 3. The results of horizontal displacement at point A under Heaviside loading: (a) analytical, FEM and
SBFEM[6] (b) present

Figure 4. The results of horizontal stress at point B under Heaviside loading: (a) analytical, FEM and
SBFEM [6] (b) present

Figure 5. The results of horizontal displacement at point A under triangular loading: (a) analytical, SEM [7]
(b) present

Figure 6. The results of horizontal stress at point B under triangular loading: (a) SEM and analytical [7],
(b) present

5.2. Bi-material rectangular bar subjected to impulsive loading

6
5th National Congress on Civil Engineering,
May 4-6, 2010, Ferdowsi University of Mashhad, Mashhad, Iran

In this example, we study a one-dimensional rectangular bar consisting of two different materials loaded by a
sinusoidal impulsive loading, in order to check the accuracy of the higher-order SBFEM in analysis of
impact problems. Solutions based upon a FEM and SBFEM are available for comparison. The segmented bar
consists of two materials: one half is the steel and the other the aluminum (Figure 7a). This bar is fixed at its
left end with traction free on its top and bottom sides. The right end side of bar is uniformly subjected to a
tensile loading of sinusoidal pulse function (Figure 8b). The geometry and loading function that are
employed in this example as L = 2W = 50 mm, and P0 = 100 MPa. The material parameters are as follows:
steel Young's modulus E1 = 200 GPa, aluminum Young's modulus E2 = 70 GPa, ρ1 = 7860 kg/m3, ρ2 = 2710
kg/m3, and ν1 = ν2 = 0. These material properties yield the maximum propagation velocity of stress waves c1
= 5044 m/s and c2 = 5082 m/s.

Figure 9. Bi-material rectangular bar subjected to impulsive loading: (a) geometry, boundary conditions, and
meshing, (b) sinusoidal pulse loading.

The time histories of horizontal displacement at point A (Figure 9a) are investigated. Figure 10 shows
the horizontal displacement of point A, in this example. Again, higher order meshes provide almost
converged time variations of numerical results.

Figure 10. The results of horizontal displacement at point A

5.3. Plane portal frame subjected to lateral loading

In this example, we consider a two-dimensional portal frame in order to demonstrate the accuracy of the
SBFEM in modeling structures that are more realistic. As no analytical solution of this problem is available,
comparison is made with the results based upon a FEM and SBFEM method. We selected the same geometry
(Figure 12a) and material properties as: Eρ = 10000 and ν = 0.2. The structure is fixed at the two bottom
edges and is loaded by a lateral uniformly distributed loading with ramp time dependence (Figure 12b). The
mesh consists of five node elements are shown by dashed/solid lines in Figure 12a.
The time variation of horizontal displacement at point A (Figure 12a) is shown in Figure 13. The time
variation of shear displacement at point B (Figure 12a) is given in Figure 11. The results from SBFEM give
almost converged time variations of numerical results. The results corresponding to higher order elements are
identical to the results obtained FEM and SBFEM.

Figure 12. Portal frame subjected to lateral loading: (a) geometry, boundary conditions and meshing, (b) ramp
time dependence function.

7
5th National Congress on Civil Engineering,
May 4-6, 2010, Ferdowsi University of Mashhad, Mashhad, Iran

Figure 13. The results of horizontal displacement at point A under ramp time loading: (a) FEM and SBFEM
[6], (b) present

Figure 14. The results of horizontal displacement at point B under ramp time loading: (a) FEM and SBFEM
[6], (b) present

6. CONCLUSIONS

In this paper, detailed introduction of the scaled boundary finite element method (SBFEM) for modeling
wave propagation problems in frequency-domain has been presented. The goal of this paper is improving
accuracy and efficiency of this method using higher-order elements. As would be realized from the detailed
formulation to show a general picture of this method, the only difference between the higher-order SEM and
the classical SBFEM could be summarized in two features: adoption of specific shape functions and
numerical quadrature. Transient analyses of five examples have been successfully carried out using the
SBFEM. In these examples, various dynamic behaviors, geometries, materials properties, boundary
conditions, and transient load functions have been selected to illustrate the applicability and generality of this
method. One may note that all these examples have been successfully modeled with very small number of
DOFs, preserving very high accuracy comparing with other analytical and numerical solutions.

7. REFERENCES

1. Robertsson J. O. A. (1996), “A numerical free-surface condition for elastic/viscoelastic finite-difference

modeling in the presence of topography,” Geophysics, 50, pp 1444–1456.
2. Bao H., Bielak J., Ghattas O., Kallivokas L. F., O'Hallaron D. R., Shewchuk J. R., Xu J. (1998), “Large-
scale simulation of elastic wave propagation in heterogeneous media on parallel computers,” Comp. Meth.
Appl. Mech. Eng. 152, pp 85–102.
3. Bouchon M., Schultz C. A., Toksoz M. N. (1996), “Effect of three-dimensional topography on seismic
motion,” J. gheophys. Res. 101, pp 5835–5846.
4. Komatitsch D., Tromp J. (2008), “Review article- Spectral-Element and adjoint methods in seismology,”
Communications in computational physics 3, pp 1–32.
5. Wolf J.P. (2004), “The scaled boundary finite element method,” John Wiley & Sons Ltd.
6. Yang Z. J., Deek. A. (2007), “A frequency-domain approach for modelling transient elastodynamics using
scaled boundary finite element method”, Comput Mech., 40, pp 725–738.
7. Khaji N., Habibi M., Mirhashemian P. (2009), “Modeling transient elastodynamic problems using spectral
element method,” Asian Journal of Civil Engineering, 10, pp 361–380.
8. Press W. H., Teukolsky S.A., Vetterling W.T., (1997), “Numerical Recipes in Fortran,” Cambridge
university press.
9. Deeks Z.J., Hao H. (2006), “A Frobenius solution to the scaled boundary finite element equations in
frequency domain,” Int J Numer Methods Eng. DOI 10.1002/nme.1926s