Beruflich Dokumente
Kultur Dokumente
=
n
n
n x a x y ) ( ) ( -------------(2)
where ) ( ) ( n x x
n
are some sub-domain basis functions
that are either piecewise linear or quadratic or still higher order
functions.
Now, the Weighted Residual statement of the equation (1) with
Galerkin formulation helps generate a set of simultaneous
algebraic equations that can very easily be solved numerically
[Ziekenwicz83].
USING WAVELETS
If ) (x
n
are chosen to be the scaling functions associated with
the wavelets, they too can be used as the basis functions for an
approximate solution. The advantage is that one could now
solve the problem in stages where one first obtains the
approximation coefficients and later obtains the detail
coefficients. However, for problems over a bounded interval,
one needs to carefully choose the wavelets that span the
interval in question [Bertoluzza94]. It is essential to follow an
2nd World Engineering Congress 2002 WEC
approach in which the approximate solution satisfies the
boundary conditions exactly. In practice, one keeps those basis
functions whose support is internal to the interval and adds to
them some edge functions that allow us to have the same order
of accuracy that we have on line. So, we can now impose the
boundary condition without introducing instabilities.
In this work we have chosen to work with Daubechies family
of wavelets and developed edge functions for scaling and its
associated wavelet functions. The scaling functions are known
to be orthogonal to their translates and also to the wavelets.
They are also orthogonal across the scales as well. The small
length members are quite irregular in shape and so have not
been used for solving differential equations in past. But, we
find that even though the shorter length wavelets are very
irregular in shape, they work well even for very low number of
elements without too many problems.
For the given Helmohltz equation, with ) ( x y as the
approximate solution, the residual
R is stated as
0 ) ( ) (
) (
2
2
2
+ =
x f x y
dx
x y d
R -----(3)
In order to reduce this residual over the entire domain, , we
call upon Galerkin weighted residual statement leading to:
N l
L
dx x f x y
dx
x y d
l
=
|
|
.
|
\
|
+ 1 0
0
) ( ) (
2
2
) (
2
--(4)
where N is the number of unknowns determined by the
resolution and
l
is the scaling function. N is a choice that, for
equal sub-divisions, determines the resolution.
Implementing the boundary conditions (y = 0 at x = 0 & y =1 at
x = L) using penalty approach, the appropriate weighted
residual statement, here, becomes
| | | | 0 ) ( ) ( ) ( ) (
) (
0
1
0
2
2
2
= + +
|
|
.
|
\
|
=
= L x
l l x l l
x y x y dx x f x y
dx
x y d
-----(5)
where is the ve + penalty number to satisfy the boundary
constraints of the problem. So, the large value of penalty
number could provide the satisfaction of the boundary
constrains and hence the better approximation is possible to the
solution. These set of equations for known ) (x
n
and f(x) can
be easily solved for
n
a s.
If, now, one-scale-better approximation is sought (one with
twice the resolution), it can be achieved by adding to hitherto
obtained solution, say,
) 0 (
y , (which is
=
n
n n
x a y ) (
) 0 (
) an
orthogonal complement
n
n n
x d ) ( based on the wavelets,
) (x
n
. The one-scale-better approximation is given by
+ =
n
n n
x d y y ) (
) 0 ( ) 1 (
------------- (6)
Now the weighted residual statement of Helmohltz equation
with detailed approximation becomes
g L J
f I K
d a
d a
= +
= +
------------- (7)
where
{ } { }
+
+
|
|
.
|
\
|
+
=
=
=
L x
l n
x
l n
n
n
L
l
n a
dx
dx
d
a K
0
2
2
2
0
{ } { }
+
+
|
|
.
|
\
|
+
=
=
=
L x
l n
x
l n
n
n
L
l
n d
dx
dx
d
d I
0
2
2
2
0
{ } { }
+
+
|
|
.
|
\
|
+
=
=
=
L x
l n
x
l n
n
n
L
l
n a
dx
dx
d
a J
0
2
2
2
0
{ } { }
+
+
|
|
.
|
\
|
+
=
=
=
L x
l n
x
l n
n
n
L
l
n d
dx
dx
d
d L
0
2
2
2
0
{ }
=
+ =
L x
l l
dx x f f ) ( .
{ }
=
+ =
L x
l l
dx x f g ) ( .
---(8)
The unknown coefficients
n
a
and
n
d
can now be obtained in
much the same way as (5). This process can be repeated and
the accuracy can be improved.
RESULTS
We choose the orthogonal Daubechies family of wavelets and
their associated scaling functions over the interval
[Bertoluza94].
We present two sets of results. In one, we solve for potential
distribution across a PN junction and in the other that for a
typical coaxial cable.
The equation to be solved for the PN junction depletion layer
potential V is obtained by choosing
0 =
and
= ) (x f
in the scalar Helmohltz equation:
) (
2
2
x
dx
V d
= ------ (9)
2nd World Engineering Congress 2002 WEC
with mixed boundary conditions
0 ) 10 ( , 0 ) 0 (
7
= = = =
x
dx
dV
x V ,
where (junction charge density) =
0
for 0 < x < 10
-7
,
(junction charge density) = -
0
for -10
-7
< x < 0 and
0
=
5x10
3
C/m
-3
.
Figure 1 and Table (1) provide the potential distribution and the
corresponding junction voltage & capacitance value for the
junction width equal to 2 x10
-7
m. We consider the domain to be
divided into (N=2 for db2and N=4 for db3) minimum
subdivisions to begin with. We employ scaling function and
determine the approximation coefficients
n
a s. Having done
that, we employ the corresponding wavelet function and solve
for the detail coefficients
n
d s. This exercise gives us one
scale better solution corresponding to 2N. We go up the scale
until N corresponds to N=64. We do this for db2, and db3 as
well and compare the results as shown in Fig.1 and table 1.
These results show that the method can provide crude
approximation as one begins with very few number of element,
N (say N = 2 for db2 & N = 4 for db3) but it improves in
accuracy as more and more details are added.
In Fig.1, The db2 begins with very crude approximation with
N =2 and evolves to the actual solution. On the other hand the
db3 starts with better approximation even with fewer number
of elements N and converges to almost exact solution when it
reaches third level detailed approximation itself with equivalent
N =32. So, it should be noted here that it is possible to get
almost exact solution just in few number of elements in the first
level crude approximation itself, by selecting a good choice of
higher unit length basis function, say db4, db5 or still
higher.
In Table 2, we take the case of co-axial cable and tabulate the
mean squared error computed from 20 nodes. We have taken
wavelets of length 3, 5, 7, 9 & 11 designated by & db2, db3 so
on, and have compared their MSE against conventional FEM
employing piecewise linear shape function. We show that our
wavelet based FEM provides less mean squared error with
higher (db3 and above) unit length of scaling function of db
wavelet compared to the same number of nodal values of
conventional FEM. In addition to that, it is also observed that
the computational effort in using higher dbs is nearly same as
lower dbs. The global matrix is 20 x 20 for all wavelets. The
matrix is more sparse for lower order wavelets than for higher
order wavelets, yet, they are all sparse. So, by choosing longer
Daubechies wavelets we could get significantly better solution
than the conventional one for very little extra cost.
Table 2 Mean square error calculation for
Proposed wavelet based FEM
For 20 Nodal values, Conventional FEM Mean Square
Error,
(MSE) = 1.229 x10
-4
Chosen db
Number
Proposed Wavelet Based FEM MSE
(For the Same 20 Nodal Values)
2
3
4
5
6
7
8
3.9600 x10
-1
2.6640 x10
-4
1.6220 x10
-5
1.0666 x10
-6
8.6789 x10
-7
1.7367 x10
-7
4.6594 x10
-8
In order to get a feel of the computational effort, let us look at
the equation (5), (7) and (8). Equation (5) is the conventional
FEM if shape functions are piece-wise linear. If one required
better accuracy, say, that corresponding to 2 N subdivisions, the
matrix elements have to be recomputed. The matrix is indeed a
sparse one. Consider now the multiresolution formulation. The
effort for N subdivisions is same. However, the effort
corresponding to 2N sub-divisions is different. We only append
the matrix K by I, J and L almost by inspection to arrive at new
global matrix and retain K from the previous calculations. It is
a hierarchical formulation.
0BTable 1: p-n junction potential with proposed FEM
1BActual Junction voltage = 0.5647 volt
2BActual Junction capacitance = 8.8542 x10
-13
F
db2 with 5 detailed levels 3Bdb3 with 4 detailed levels
Equivalent
of elements
Junction Voltage
(volt)
Junction
Capacitance
(Farad)
Junction Voltage
(volt)
Junction Capacitance
(Farad)
N = 2
4BN = 4
N = 8
N=16
N=32
N=64
0.0863
0.3571
0.4978
0.5198
0.5417
0.5462
5.7949 x10
-12
1.4001 x10
-12
1.0045 x10
-12
9.6190 x10
-12
9.2303 x10
-13
9.1547 x10
-13
--------
0.2794
0.5669
0.5661
0.5647
0.5647
--------------
1.7894 x10
-12
8.8193 x10
-12
8.8331 x10
-12
8.8544 x10
-13
8.8544 x10
-13
2nd World Engineering Congress 2002 WEC
Further, it is not difficult to see that the formulation is not
critically dependent on equal size of the sub-domains. It is
actually possible to choose the nodes according to any other
suitable criteria and fit the shape functions accordingly
[Ching01]. This exercise will truly reflect the FEM capability
for problems with arbitrary geometry.
CONCLUSION
This work, thus, establishes the feasibility of a multiresolution
Finite Element Method, albeit for one-dimensional problems. It
is observed that the solution, that uses longer scaling and
wavelet functions, provides less mean squared error as
compared to the shorter length wavelets. Moreover it shows
that longer length wavelets do not require much computational
effort than the short wavelets. This formulation is hierarchical
in nature, i.e. in each level of approximation, the global matrix
is constructed from the previous level of global matrix. One
clear advantage of using the formulation is that it requires far
fewer calculations than the conventional FEM formulation.
ACKNOWLEDGEMENT
The financial assistance from the Universiti Teknologi
PETRONAS to present this paper is kindly acknowledged.
REFERENCES
Ziekenwicz, O.C., and Morgan, K., Finite Elements and
Approximations, Wiley, New York, 1983.
Strang, G., and Nguyen, T., Wavelets and Filter Banks,
Wellesley-Cambridge Press, 1996.
Balanis, C.A., Advanced Engineering Electromagnetics, John
Wiley, 1989.
N. Narayan Rao, Elements of Engineering Electromagnetics,
Prentice Hall, 2000.
Bertoluzza, S., Naldi, G. and Ravel, J. C. (1994). Wavelet
Methods for the Numerical Solution of Boundary Value
Problems on the Interval, in Wavelets: Theory, Algorithms,
and Applications, Charles K.Chui, Laura Montefusco, and
Luigia Puccio (Eds.), Academic Press, pp. 425-448
V. Selvapandi, Improved Wavelet based FEM, Masters Thesis,
E. & E. Engg., Universiti Sains Malaysia (2001)
Bong Ching Ching, Wavelet Based Finite Element Method for
Transmission Line Problems, B.E. Final Year Project thesis, E
& E Engg. , Universiti Sains Malaysia (2001)
-1 -0.5 0 0.5 1
-4
-3
-2
-1
0
1
2
3
x 10
13
-1 -0.5 0 0.5 1
-4
-3
-2
-1
0
1
2
3
x 10
13
Legends
: - N=2;
: - N=4;
+ + : - N=8;
0 0 0 0 0 : - N=16;
>>> : - N=32;
Figure 1: PN junction potential with proposed FEM A) Using db2 B) Using db3
Legends
: - N=4;
: - N=8;
+ + : - N=16;
0 0 0 0 0 : - N=32;
x10
-7
x10
-7
Junction width (meter)
P
o
t
e
n
t
i
a
l
(
V
o
l
t
)
Junction width (meter)
P
o
t
e
n
t
i
a
l
(
V
o
l
t
)
10
-1
10
-1