Académique Documents
Professionnel Documents
Culture Documents
, we can use a
one grid algorithm, even as a base for future extension to multigrid compu-
tation [12], which applies an appropriate number m of sweeps of a relaxation
procedure R suitable for nonlinear problems [11,13], that is
X
h
R
m
X
h
0
; L
h
; K
h
; U
h
: 7
Obviously, the goodness of the solution of the elliptic grid generator (6)
depends on the required grid neness and on the domain geometry.
3. Transnite orthogonal algebraic generator
Let X R
2
be such that oX [
4
i1
oX
i
, with oX
1
\ oX
3
; oX
2
\ oX
4
;
where oX
1
, oX
2
, oX
3
, oX
4
are the supports of four regular curves
c
i
: 0; 1 ! oX
i
, i 1; . . . ; 4, taken counterclockwise. Furthermore, we assume
that the curve intersections occur only at the end points of the boundary curves
c
i
, i 1; . . . ; 4 i.e.
c
1
0 c
4
1; c
1
1 c
2
0; c
2
1 c
3
0; c
4
0 c
3
1:
We dene /
1
n : c
1
n, /
2
n : c
3
1 n, n 2 0; 1 and w
1
g :
c
4
1 g, w
2
g : c
2
g, g 2 0; 1 and four additional curves by computing
the derivatives of the / and w-curves, i.e.
/
i2
n
K
n
i2
k/
0
i
k
2
/
y
i
n
0
; /
x
i
n
0
; i 1; 2;
w
j2
g
K
g
j2
kw
0
j
k
2
w
y
j
g
0
; w
x
j
g
0
; j 1; 2;
8
denoting by /
x
, /
y
and w
x
, w
y
the components of the /-curves and w-curves,
respectively. The symbol k k
2
stands for the Euclidean norm and K
n
l
, K
g
l
,
l 3; 4 are suitable constant values (see Section 4 for a specic denition). As
we are going to deal with orthogonal lines emanating from the boundary of the
domain, we assume the following conditions on the boundary curves
18 C. Conti et al. / Appl. Math. Comput. 162 (2005) 1527
/
i2
0 w
0
1
u
i
; /
i2
1 w
0
2
u
i
;
w
i2
0 /
0
1
u
i
; w
i2
1 /
0
2
u
i
; i 1; 2;
/
00
i
0 w
00
1
u
i
; /
00
i
1 w
00
2
u
i
;
9
where u
1
0, u
2
1.
Then, we introduce the linear operators
P
1
/n; g :
X
4
i1
a
i
g/
i
n; P
2
wn; g :
X
4
j1
a
j
nw
j
g;
P
1
P
2
/; wn; g :
X
2
i1
a
i
gP
2
wn; u
i
a
i2
g
oP
2
wn; u
i
og
:
10
The functions a
j
n, j 1; . . . ; 4 and a
i
g, i 1; . . . ; 4, in (10) are the so
called blending functions satisfying cardinal interpolation
a
j1
l d
jl
; a
0
j1
l 0; j 0; 1; l 0; 1;
a
j3
l 0; a
0
j1
l d
jl
; j 0; 1; l 0; 1:
Thus, the Hermite blending function surface is
P
1
P
2
/; wn; g P
1
/n; g P
2
wn; g P
1
P
2
/; wn; g; 11
where P
1
P
2
is the so called Boolean sum operator. As well known, because
of the blending function properties, it holds
P
1
P
2
u
j
; g w
j
g; j 1; 2; P
1
P
2
n; u
i
/
i
n; i 1; 2;
oP
1
P
2
u
j2
; g
on
w
j
g; j 3; 4;
oP
1
P
2
n; u
i2
og
/
i
n; i 3; 4:
12
3.1. Mixed orthogonal algebraic method
As already noticed, transnite method do not get much degrees of freedom
to obtain workable meshes. Thus, as suggested in [4] and further investigated in
[5,10], they can be protably combined with tensor product techniques which,
involving control points, allow us a control of the grid on the interior of the
domain X. Thus, the idea is to work with blending functions a
i
n, i 1; . . . ; 4,
a
j
g, j 1; . . . ; 4, with local support on 0; 1. Here, they are dened as the
dilated versions of the classical Hermite bases with support on I
n
0
: 0; u
n
and
on I
n
1
: 1 ~u
n
; 1 being 0 < u
n
< 1 and 0 < ~u
n
< 1, i.e.
C. Conti et al. / Appl. Math. Comput. 162 (2005) 1527 19
a
1
n : 1
2
n
u
n
!
1
n
u
n
!
2
; a
3
n : n 1
n
u
n
!
2
; n 2 I
n
0
;
a
2
n : 3
2
n ~u
n
1
~u
n
!
n ~u
n
1
~u
n
!
2
;
a
4
n : n 1
n ~u
n
1
~u
n
!
2
; n 2 I
n
1
:
13
The blending functions a
j
g, j 1; . . . ; 4 are analogously dened with
support on I
g
0
: 0; u
g
and on I
g
1
: 1 ~u
g
; 1. Next, we dene the linear
transformation X : 0; 1
2
!R
2
, Xn; g xn; g; yn; g as
Xn; g : T
P
n; g P
1
P
2
/; w T
P
n; g; 14
where
T
P
n; g :
X
m
p
i1
X
n
p
j1
Q
ij
B
i;3
nB
j;3
g 15
with B
i;3
n and B
j;3
g denoting the usual cubic B-splines with knots
fn
i2
; n
i1
; n
i
; n
i1
; n
i2
g, fg
j2
; g
j1
; g
j
; g
j1
; g
j2
g, respectively. The set
Q fQ
ij
g
m
p
;n
p
i;j1
is the set of control points. For the B-spline knot choice and the
control point choice the reader can refer to [1,2].
The Boolean sum operator P
1
P
2
in (14) is also acting on the surface
T
P
n; g taking into account the eight boundary curves T
P
0; g, T
P
1; g,
T
P
n; 0, T
P
n; 1,
oT
P
0;g
on
,
oT
P
1;g
on
,
oT
P
n;0
og
,
oT
P
n;1
og
.
As discussed in [2], X satises
Xu
j
; g w
j
g; j 1; 2; Xn; u
i
/
i
n; i 1; 2;
oXu
j
; g
on
w
j
g; j 3; 4;
oXn; u
i
og
/
i
n; i 3; 4:
16
Moreover, setting u
n
n
3
n
2
, u
g
g
3
g
2
, ~u
n
n
m
p
1
n
m
p
2
,
~u
g
g
n
p
1
g
n
p
2
, because of the blending function locality, it holds
Xn; g T
P
n; g for all n; g 2 n
3
; n
m
p
2
g
3
; g
n
p
2
.
The grid X
h
0
is then obtained by sampling X at a given set of parameter
values fn
i
; g
j
g
m
g
;n
g
i;j1
i.e.
X
h
0
: Xn
i
; g
j
xn
i
; g
j
; yn
i
; g
j
m
g
;n
g
i;j1
:
20 C. Conti et al. / Appl. Math. Comput. 162 (2005) 1527
4. Numerical algorithm and computational parameters
Before to present the algorithm for constructing the algebraicelliptic grid
with orthogonal feature to the boundary, we need to make clearer how to set
some of the parameters introduced in the previous sections.
4.1. Choice of B-splines knots
Concerning the knots of the B-splines involved in (15), that is fn
i
g
m
p
2
i1
and
fg
l
g
n
p
2
l1
, they are set with the property
0 n
1
n
0
n
1
n
2
< < n
m
p
1
n
m
p
n
m
p
1
n
m
p
2
1;
with n
i1
n
i
1
m
p
3
; i 2; . . . ; m
p
2: 17
Similarly,
0 g
1
g
0
g
1
< g
2
< < g
n
p
1
g
n
p
g
n
p
1
g
n
p
2
1;
with g
l1
g
l
1
n
p
3
; l 2; . . . ; n
p
2: 18
4.2. Choice of the modulus of the boundary normal derivatives
Now, we discuss about the constants K
n
i
, K
g
i
, i 1; 2 to be used in (8). By
using the so called uniformity property (see [2] for a detailed discussion about
this property) we arrive at the following choice
K
n
i
P
m
p
2
r3
/
x
2
n
r
/
x
1
n
r
k/
0
i
k
2
P
m
p
2
r3
/
y
i
n
r
0
; i 3; 4 19
and
K
g
j
P
n
p
2
l3
w
y
2
g
l
w
y
1
g
l
kw
0
j
k
2
P
n
p
2
l3
w
x
j
g
l
0
; j 3; 4: 20
4.3. Choice of the control points
Let us consider the bilinear blending surface interpolating the four boundary
curves
Ls; t 1 t/
1
s t/
2
s 1 sw
1
t sw
2
t 1 t1 sw
1
0
sw
2
0 t1 sw
1
1 sw
2
1 21
C. Conti et al. / Appl. Math. Comput. 162 (2005) 1527 21
and the parameter values, in 0; 1 fr
i
g
m
p
i1
and fs
j
g
n
p
j1
dened as
r
1
: 0; r
2
: r
1
2
3m
p
2
; r
3
: r
2
3
3m
p
2
;
r
i1
: r
i
3
3m
p
2
; i 3; . . . ; m
p
3;
r
m
p
1
: r
m
p
2
2
3m
p
2
; r
m
p
: r
m
p
1
1
3m
p
2
1;
s
1
: 0; s
2
: s
1
2
3n
p
2
; s
3
: s
2
3
3n
p
2
;
s
j1
: s
j
3
3n
p
2
; j 3; . . . ; n
p
3;
s
n
p
1
: s
n
p
2
2
3n
p
2
; s
n
p
: s
n
p
1
1
3n
p
2
1:
22
The control points are obtained by sampling Ls; t, at the previous
parameter values that is
Q
ij
Lr
i
; s
j
; i 1; . . . ; m
p
; j 1; . . . ; n
p
: 23
4.4. Choice of the ghost points
In order to impose boundary orthogonality to the elliptic generator of the
algorithm, we exploit the boundary orthogonal grid constructed by the dened
mixed Hermite method in step 2 to
(i) set the boundary grid point distribution as Dirichlet conditions for the
problem (6) in step 3.1 (even taking into account physical boundary condi-
tions from step 1 on),
(ii) compute the orthogonal control functions (5) in step 3.3.
The algebraic grid is assumed to have the rst interior coordinate lines
correctly positioned so that appropriate exterior curves of ghost points can be
derived by orthogonal symmetry (algorithmstep 3.2). Leaving the ghost
points unchanged allows us to use the information on boundary orthogonality
in (5) and save it during the iterations. Permitting free positions to all the
interior points allows us to achieve smoothness. Orthogonality at corner points
is not imposed.
Algorithm
1. Input the boundary curves /
1
, /
2
, w
1
, w
2
, the grid size m
g
; n
g
, the number of
control points m
p
; n
p
, the tolerances s
d
, s
r
and s
o
and the max iteration num-
ber n
max
22 C. Conti et al. / Appl. Math. Comput. 162 (2005) 1527
2. Construct the algebraic grid through the following steps:
2.1. compute the orthogonal directions /
3
, /
4
, w
3
, w
4
by means of (8) and
(19), (20)
2.2. compute the control points fQ
ij
g
m
p
;n
p
i;j1
with the strategy in
(23)
2.3. compute the knots of the B-splines as in (17) and (18)
2.4. use (14) to generate the mixed algebraic surface Xn; g
2.5. compute the discrete grid X
h
: fXn
i
; g
j
g
m
g
;n
g
i;j1
with neness
h h
n
; h
g
, where h
n
:
1
m
g
1
, h
g
:
1
n
g
1
and n
i
: ih
n
, g
j
: jh
g
for
i; j 1; . . . ; m
g
; n
g
2.6. set the initial grid X
h
0
: X
h
3. compute a smoother grid by the elliptic procedure:
3.1. set prescribed boundary grid points as Dirichlet boundary condi-
tions for the problem (6), that is U
h
: fX
h
0
1; j; X
h
0
m
g
; j; X
h
0
i; 1;
X
h
0
i; n
g
g
m
g
;n
g
i;j1
3.2. construct an exterior curve of points by orthogonal symmetry from
fX
h
0
2; j; X
h
0
m
g
1; j; X
h
0
i; 2; X
h
0
i; n
g
1g
m
g
1;n
g
1
i;j2
3.3. compute Pj
o0;1
2 and Qj
o0;1
2 using (5)
3.4. extend P and Q on 0; 1
2
by bilinear transnite interpolation
3.5. Solve the system (6) by GaussSeidel relaxation procedure:
3.5.1. for k 1; . . . ; n
max
solve X
h
k
R
k
X
h
k1
; L
h
; K
h
; U
h
where R
k
is the kth iteration of the GS-
relaxation procedure
compute d
k
: kX
h
k
X
h
k1
k
2
and the residual r
k
: kL
h
X
h
k
k
2
compute the vector Ort
k
containing the scalar product of the discrete
boundary curves and the grid lines at the boundary
compute o
k
: kOrt
k
k
2
if d
k
6s
d
& r
k
6s
r
& o
k
6s
o
go to 4
4. Set m : k and X
h
: X
h
k
5. Print the nal grid X
h