Sei sulla pagina 1di 4

IEEE TRANSACTIONS ON MAGNETICS, VOL. 48, NO.

2, FEBRUARY 2012 607

Handling Material Discontinuities in the Generalized Finite Element


Method to Solve Wave Propagation Problems
Werley Gomes Facco1;2 , Elson Jos Silva3 , Alex Sander Moura2 , Na-sses Zoia Lima3 , and
Rodney Rezende Saldanha2

Department of Exact Sciences, Federal University of the Vales Jequitinhonha and Mucuri, Minas Gerais 39803-371, Brazil
Graduate Program in Electrical Engineering - Federal University of Minas Gerais, Minas Gerais 31270-901, Brazil
Department of Electrical Engineering, Federal University of Minas Gerais, Minas Gerais 31270-901, Brazil

To solve wave propagation problems involving change of medium, many authors employ the generalized finite element method with
plane wave enrichment and Lagrange multipliers to ensure interface constraints. However this approach produces ill conditioned and
nonpositive definite systems, making it hard to solve. This paper presents an approach based on the mortar element method that substi-
tutes the Lagrange multipliers with the advantage of generating sparse and positive definite systems. Various numerical aspects affecting
the generalized finite element method efficiency are analyzed by solving a 2-D scattering problem.
Index TermsGeneralized finite element method, Lagrange multipliers, mortar element methods, scattering problems.

I. INTRODUCTION piecewise homogeneous domain. Next we summarize the weak


formulation of an electromagnetic scattering problem and
formulate the GFEM enriched with plane waves. In Section III
HE generalized finite element method (GFEM) is a we address the LM and MEM issues in the GFEM formulation.
T combination of the finite element method (FEM) and the
method of partition of unity (PUM) [1], [2]. The main feature of
The computational results and the conclusions are presented in
Sections IV and V, respectively.
the PUM is to allow the basis functions to incorporate analytical
knowledge about the solution on a general mesh structure which
II. FORMULATION
supports a partition of unity [3]. The GFEM has been proved
to be suitable to deal with wave propagation problems, where In this paper, we solve the time harmonic electromagnetic
the classical FEM may require a prohibitive mesh [4][6]. It scattering problem in two dimensions to validate the proposed
is usual to solve Helmholtz equation using FEM mesh with a technique. Assuming that the obstacle is enclosed by a smooth
minimum resolution of ten nodal points per wavelength. Using artificial boundary , the annular domain, , be-
plane wave analytical solutions of the homogeneous Helmholtz tween and is defined by the presence of two media whose
equation, we can obtain accurate results with mesh size greater interface is , Fig. 1. Specifically, we consider transverse elec-
than one wavelength, what can reduce significantly the number tric field excitation and seek the z-component of the mag-
of unknowns. In this paper, the shape functions of the GFEM netic field that satisfies the Helmholtz equation in . We impose
are obtained by multiplying plane waves in different directions homogeneous Neumann boundary condition on , and specify
with the nodal piecewise linear basis functions of a finite an absorbing boundary condition (ABC) to impose the Sommer-
element triangulation. In problems where the domain is formed feld radiation condition on .
by regions of different materials, the GFEM spaces do not The GFEM is applied to the variational formulation of the
match at the interfaces with step change in the wave number. governing equation [1]. We derive a weak formulation by ap-
The usual solution is to introduce Lagrange multipliers (LM) to plying the method of weighted residuals in each subdomain. By
approximately enforce the interface conditions. However, this usual procedure, the weak forms are given by
method leads to ill conditioned and nonpositive definite matrix
which imposes severe restrictions on the methods to solve (1)
the system of equations [7], [8]. In [9], [10] the authors have
shown that there are advantages in using the mortar element
method (MEM) to enforce interface conditions in the FEM,
for instance, the resultant matrix is positive definite. This work
proposes an approach which can be interpreted as an adaptation
of the MEM for GFEM to handle material discontinuities in (2)

Manuscript received July 07, 2011; revised October 12, 2011; accepted Oc- where and are the restriction of the total mag-
tober 15, 2011. Date of current version January 25, 2012. Corresponding author: netic field and the test function on and , respectively.
W. G. Facco (e-mail: werleyfacco@yahoo.com.br). The wave numbers and are constant in each subdomain.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. The function on corresponds to trun-
Digital Object Identifier 10.1109/TMAG.2011.2173309 cate the domain with the second order Bayliss-Turkel ABC and
0018-9464/$31.00 2012 IEEE
608 IEEE TRANSACTIONS ON MAGNETICS, VOL. 48, NO. 2, FEBRUARY 2012

To approximate the LM, we use the same approach used for


the magnetic field approximation (5), i.e., given an element
with one of its edges belonging to , we write

(7)

where and are the Lagrange multipliers


for the node in direction . The matrix form of (7) is written
as , where is the vector with the Lagrange
multipliers and . Therefore, we can add the
continuity condition as
Fig. 1. Computational domain
between the PEC surface 0 and the artificial
boundary 0 .
(8)

is the magnetic field of the incident


plane wave. Finally, the following block symmetric system is obtained:

A. GFEM Function Space


To construct the approximation space for GFEM, we com-
bine the nodal FEM shape functions with the plane wave (9)
functions in different directions of the plane [4] at each node.
From element point of view, for each element , the local
space of linear combinations of plane wave with directions
, is given by
where , , and are the conventional stiffness matrices
and load vectors of each domain of the GFEM. and are
(3) matrices related to integral terms over material interface [8].
and are the vectors of the unknowns amplitudes on , while
and are the ones out of in and , respectively.
where
Note that the last diagonal block is formed by zero elements.
(4) The system (9) is ill conditioned and nonpositive definite [8]
and care must be taken to solve it. Following [10] in our im-
and is the number of wave directions. In this case the new plementation we avoid singular matrices by using a node num-
enriched shape functions generate the space of bering schema which guarantees that the degrees of freedom of
approximation functions of GFEM. Thus for each element the the nodes located on the common interface take the first places.
magnetic field is approximated by
B. Extension of the Mortar Element Method for GFEM
(5) In [10] the authors propose an aproach based on the MEM
to enforce continuity constraints in the classical FEM. Here we
extend their methodology to GFEM enriched with plane waves.
This approach is related to LM through the mortar condition,
III. CONTINUITY BETWEEN SUBDOMAINS for each plane wave direction we can write
Now we focus on enforcing constraints along the interface
. Since the wave number is an attribute of each region our
formulation must handle the step jumps on the interface when (10)
the subdomains are glued.

A. Lagrange Multipliers
where
In general, enforcing interface constraints is typically done
by means of LM. In this case the contour integral over , in (1) (11)
and (2) is evaluated with
and may be calculated by line integrals over
(6)
(12)
and a mixed formulation is obtained.
FACCO et al.: HANDLING MATERIAL DISCONTINUITIES IN THE GENERALIZED FINITE ELEMENT METHOD 609

(13)

The functions and are the nodal base functions of the


node when the element is inside and , respectively, and
is the number of nodes on the interface . The inversion
of the matrix is required for each wave direction. It does
not introduce significant overhead on the overall time to solve
the problem as its dimension is small and it is almost
diagonal.
The matrices are combined according to the wave direc-
Fig. 2. Triangular mesh used in the computations.
tion of each node and the following interface matrix relation is
obtained:

(14)

Using the condition (14) the unknowns over the whole do-
main can be linked as follows:

(15)

which can be expressed as

(16) Fig. 3. Real and imaginary parts of the magnetic field for GFEM-LM, along
the line C .
where is the global coupling matrix and is the identity
matrix.
Applying the coupling matrix on the generalized finite ele- A. GFEM With LM
ment system , we have Fig. 3 shows the real and imaginary parts of the magnetic
field solution evaluated numerically and analytically on the line
(17) . The magnetic
field calculated through the GFEM with directions of
To obtain the Hermitian condition the final system is premul- plane waves provides an excellent approximation to the analytic
tiplied by the conjugate of the coefficient matrix. Then, the re- solution [8]. It is observed that the absolute errors in the real
sultant system is sparse and positive definite and, therefore, it and imaginary parts are consistently below 0.0027 and 0.0015,
can be efficiently solved by a wide range of methods. respectively.

IV. RESULTS AND DISCUSSION B. GFEM With MEM

In this section, we present and analyze the computational re- The solution to the problem considering the GFEM with
sults obtained using the LM and MEM aproaches to enforce the MEM and directions of plane waves is shown in Fig. 4,
interface conditions in the GFEM. The model problem shown through which one observes that the solution for the magnetic
in Fig. 1 is solved by enriching the GFEM with plane waves field obtained by this method is as accurate as that of the LM
in different directions. The following data are employed in the with GFEM. The absolute error of the real and imaginary part
computations: does not exceed 0.0024 and 0.0016 along the line .
radius of the PEC surface, ;
C. -Convergence of the GFEM With MEM and LM
radius of the dielectric interface, ;
radius of truncation boundary, ; To analyze the convergence of the GFEM the accuracy of the
wave numbers, and . model is measured by the relative -norm error defined in
The radius lengths are normalized by the wavelength in the the whole computational domain as
subdomain . The computational domain was discretized
into 588 triangular elements, resulting in 991 edges and 353 (18)
nodes. Among these nodes, 90 are on the interface , as shown
in Fig. 2. To evaluate the coefficients of the system (9) we em- Fig. 6 shows the -convergence of real and imaginary parts
ployed Gauss-Legendre integration rule with 64 points inside of the magnetic field solution along the line by GFEM with
each triangle and 6 along each boundary edge. MEM and ML, , 4, 6, 8, 10. To obtain accurate results we
610 IEEE TRANSACTIONS ON MAGNETICS, VOL. 48, NO. 2, FEBRUARY 2012

TABLE I
COMPUTATION TIME

field is smoothly transmited to the inner medium. The matrix


patterns shown in Fig. 7 are obtained for mesh Fig. 2 with six
Fig. 4. Real and imaginary parts of the magnetic field for GFEM-MEM, along directions of plane waves. These patterns reveal that even with
the line C . a small system dimension the MEM has more than four times
nonzero elements than LM. The values in Table I report the com-
putation time, , to assemble and, , to solve the final system
with increasing wave directions. Since the MEM requires a ma-
trix inversion for each direction on the interface nodes it de-
mands more computational time to get the final system. How-
ever, in the overall, the MEM is much faster than LM, which
is explained by the ill conditioning of the GFEM-LM matrix
whose solution is very time consuming.
Fig. 5. Total magnetic field around the cylinder, q = 10,  = 1:9%, (left) real
part, (right) imaginary part. V. CONCLUSION
In this paper, a scattering problem was solved using the
GFEM with emphasis on methods to enforce interface condi-
tions. It was shown that LM can be computationally demanding
since it leads to ill conditioned matrices. A new approach based
on the MEM was proposed and its main advantage is to provide
a sparse definite positive matrix that significantly facilitates the
solution of the final equation system. The two approaches are
conceptually equivalents and presented the same convergence
rate.

ACKNOWLEDGMENT
This work was supported in part by the State of Minas Gerais
Research FoundationFAPEMIG, Brazil, the Brazilian gov-
Fig. 6. Real and imaginary q -convergence of the GFEM with MEM solution ernment agency CAPES and National Council for Scientific and
versus the corresponding convergence of the GFEM with LM, for Mesh Fig. 2,
for q =2 , 4, 6, 8, 10.
Technological DevelopmentCNPq, Brazil.

REFERENCES
[1] I. Babuska and J. M. Melenk, Int. J. Num. Meth. Eng, no. 40, pp.
727758, 1996.
[2] A. Plaks, I. Tsukerman, G. Friedman, and B. Yellen, IEEE Trans.
Magn., vol. 39, no. 3, pp. 14361439, May 2003.
[3] L. Proekt, S. Yuferev, I. Tsukerman, and N. Ida, IEEE Trans. Magn.,
vol. 38, no. 2, pp. 649652, Mar. 2002.
[4] T. Strouboulis, I. Babuska, and R. Hidajat, Comput. Meth. Appl. Mech.
Eng., no. 195, pp. 47114731, 2006.
[5] L. Proekt and I. Tsukerman, IEEE Trans. Magn., vol. 38, no. 2, pp.
Fig. 7. View of nonzero elements for (left) MEM and (right) LM methods. 741744, Mar. 2002.
[6] C. Lu and B. Shanker, IEEE Trans. Magn., vol. 43, no. 3, pp.
10021012, Mar. 2007.
[7] D. Rodger, H. C. Lai, and P. J. Leonard, IEEE Trans. Magn., vol. 26,
need a minimum of 6 wave directions which leads to (6 353) no. 2, pp. 548550, Mar. 1990.
degree of freedom in the MEM and for the LM. [8] O. Laghrouche, P. Bettess, E. Perrey-Debain, J. Trevelyan, and R. Hi-
dajat, Comput. Meth. Appl. Mech. Eng., no. 194, pp. 367381, 2005.
We see that the approach used to enforce boundary continuity [9] O. J. Antunes, J. P. A. Bastos, N. Sadowski, A. Razek, L. Santandrea,
does not affect the convergence. Fig. 5 shows the total magnetic F. Bouillault, and F. Rapetti, IEEE Trans. Magn., vol. 41, no. 5, pp.
field distribution around the diffracting cylinder. The numerical 14721475, May 2005.
[10] O. J. Antunes, J. P. A. Bastos, N. Sadowski, A. Razek, L. Santandrea,
results are reported on lower half surface of the mesh Fig. 2 and F. Bouillault, and F. Rapetti, IEEE Trans. Magn., vol. 42, no. 4, pp.
the analytical results are shown on the other half. Note that the 599602, Apr. 2006.

Potrebbero piacerti anche