A Method of Fundamental Solutions For Two-Dimensional Heat Conduction

A method of fundamental solutions for two-dimensional
heat conduction
B. Tomas Johansson, D. Lesnic, Thomas Henry Reeve
To cite this version:

B. Tomas Johansson, D. Lesnic, Thomas Henry Reeve. A method of fundamental solutions for two-
dimensional heat conduction. International Journal of Computer Mathematics, Taylor & Francis,
2011, pp.1. <10.1080/00207160.2010.522233>. <hal-00678795>
HAL Id: hal-00678795

https://hal.archives-ouvertes.fr/hal-00678795
Submitted on 14 Mar 2012
HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
International Journal of Computer Mathematics
Fo
rP
A method of fundamental solutions for two-dimensional

ee
heat conduction
rR
Journal: International Journal of Computer Mathematics
Manuscript ID: GCOM-2009-0885-B.R1
Manuscript Type: Original Article

ev
Date Submitted by the

28-Jul-2010
Author:
Complete List of Authors: Johansson, B.; University of Birmingham, School of Mathematics

ie
Lesnic, D.; University of Leeds, Applied Mathematics

Reeve, Thomas; University of Birmingham, School of Mathematics
w
Heat conduction, Method of fundamental solutions, Regularization,

Keywords:
Two-dimensional Domains, Collocation
On
Note: The following files were submitted by the author for peer review, but cannot be converted
to PDF. You must view these files (e.g. movies) online.
Research paper (2D).zip

ly
URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

July 27, 2010 13:31 International Journal of Computer Mathematics JohLesReeMFS
Page 1 of 17 International Journal of Computer Mathematics
International Journal of Computer Mathematics

1 Vol. 00, No. 00, January 2008, 1–17
2
3
4
5
6
7 RESEARCH ARTICLE
8
9
10 A method of fundamental solutions for
11 two-dimensional heat conduction
12
13
B. Tomas Johanssona∗ , Daniel Lesnicb and Thomas Reevec
14
15 a,c
16 School of Mathematics, University of Birmingham, Birmingham, B15 2TT, UK;
Fo
b
17 Department of Applied Mathematics, University of Leeds, Leeds, LS2 9JT, UK
18 (Received )
19
20
rP
We investigate an application of the method of fundamental solutions (MFS) to heat con-

21 duction in two-dimensional bodies, where the thermal diffusivity is piecewise constant. We
22 extend the MFS proposed in [15] for one-dimensional heat conduction with the sources placed
23 outside the space domain of interest, to the two-dimensional setting. Theoretical properties of
the method, as well as numerical investigations, are included, showing that accurate results
24 can be obtained efficiently with small computational cost.
ee
25
26
Keywords: heat conduction; method of fundamental solutions
27
28 AMS Subject Classification: 35K05; 35A35; 65N35
rR
29
30
31 1. Introduction
32
33
ev
In many engineering applications finding the solution to various heat conduction

34
problems is of fundamental importance. Examples include, heat exchangers, math-
35
36 ematical finance, in particular after transforming the Black–Scholes equation into
the heat equation, and various chemical and biological systems, including diffusion
iew
37
38 and transportation problems. Thus, due to its importance, many different numer-
39 ical techniques have been developed for calculating heat flow. The method of fun-
40 damental solutions (MFS) is a powerful numerical technique that has been used to
41 obtain highly accurate numerical approximations of solutions to linear partial dif-
42 ferential equations (PDEs) with small computational effort, see the reviews [8, 10].
43
On
However, this technique has mainly been applied to stationary heat flow governed
44
by elliptic partial differential equations [1, 2]. Recently, in [15], an MFS for the
45
46 time-dependent linear heat equation in one spatial dimension was proposed and
47 investigated. This method was extended to free surface Stefan problems in [5] and
to heat conduction in one-dimensional layered materials in [16]. Encouraged by
ly
48
49 these results, in this paper we extend the approach considered in [15] to heat con-
50 duction in two-dimensional bodies. We note that other formulations of the MFS
51 for the parabolic heat equation were given in [4, 11, 19, 25, 26].
52 We begin the work in Section 2 by introducing some notation and function spaces,
53 and formulate and review results for the linear heat conduction equation. In Sec-
54
tion 3, we prove some theoretical results that we will use in our MFS formulation, in
55
56 particular, that linear combinations of fundamental solutions are dense in the space
57
58 ∗ Corresponding author. Email: B.T.Johansson@bham.ac.uk
59
60
ISSN: 0020-7160 print/ISSN 1029-0265 online
c 2008 Taylor & Francis
DOI: 0020716YYxxxxxxxx
http://www.informaworld.com

International Journal of Computer Mathematics Page 2 of 17
2 B. Tomas Johansson et al.
1 of square integrable functions defined on the boundary of the solution domain. In

2 Section 4, we describe the MFS implementation for various (bounded) solution
3
domains. In Section 5, we present numerical results for both some direct problems
4
5 and also for one inverse problem. These results show that accurate approximations
6 can be obtained efficiently with small computational effort.
7
8
9 2. Preliminaries and notation
10
11 We let x = (x1 , x2 ) and y = (y1 , y2 ) be points in R2 and T > 0 be a fixed real
12
number. The conducting body D is a two-dimensional bounded domain in R2 with
13
14
smooth bounding surface Γ = ∂D, for example, C 2 -smooth is sufficient. The closure
15 of the body D is D̄ = D ∪ Γ. Composed with time we have the following cylinders
16 DT = D × (0, T ] and ΓT = Γ × (0, T ], respectively. The closures of DT and ΓT are
Fo
17 given by D̄T = D̄ × [0, T ] and Γ̄T = Γ × [0, T ], respectively.

18 We are interested in constructing the solution u to the heat equation in the
19 domain DT , supplied with initial and Dirichlet boundary conditions, that is u
20 solves
rP
21
22 ∂u(x, t)
23 − ∆u(x, t) = 0, (x, t) ∈ DT , (1)
∂t
24
ee
25
26
27
u(x, t) = h(x, t), (x, t) ∈ ΓT , (2)
28
rR
29
30 u(x, 0) = u0 (x), x ∈ D, (3)
31
32 where u0 (x) and h(x, t) are sufficiently smooth functions. We point out that, in
33
ev
principle, the MFS that we propose and investigate can be applied to other bound-
34
ary conditions, such as Neumann and mixed boundary conditions.
35
36 To guarantee the existence and uniqueness of a solution to (1)–(3) we impose
the following compatibility conditions:
iew
37
38
39 ∂h
u0 (x) = h(x, 0) and (x, 0) = ∆u0 (x), x ∈ Γ. (4)
40 ∂t
41
42 With these conditions the following uniqueness theorem holds, see, for example,
43 [9].
On
44
45 Theorem 2.1 Let u0 (x) ∈ C 2 (D̄) and h(x, t) ∈ C 1 (Γ̄T ) satisfy the compatibility
46 conditions (4). Then there exists a unique solution u ∈ C 2,1 (D̄T ), to the equations
47 (1)–(3), which depends continuously on the data.
ly
48
49 Theorem 2.1 tells us, in particular, that the problem given by the equations
50 (1)–(3) is well-posed.
51
52
53 3. Denseness properties of linear combinations of fundamental solutions
54
55 The fundamental solution of (1) in two-dimensions is given by
56
57
58 H(t − τ ) − |x−y| 2
F (x, t; y, τ ) = e 4(t−τ ) , (5)

59 4π(t − τ )
60

International Journal of Computer Mathematics 3
1 where H is the Heaviside function which is introduced in order to emphasize that

2 the fundamental solution is zero for t ≤ τ . We shall investigate some properties of
3
linear combinations of such functions for various source points y.
4
5 We begin by constructing a set of source points placed outside the region D̄. Let
6 DE (E for enclosed) be an open domain, containing D̄, with bounding surface ΓE ,
7 where the distance between the points on the surfaces Γ and ΓE is greater than
8 zero. Let {yj , τm }j,m=1,2,... be a denumerable, everywhere dense set of points in
9 ΓE × [−T, T ], (τm 6= 0) and set
10
11 (j)
12
vm (x, t) = F (x, t; yj , τm ). (6)
13
14 Figure 1 shows how the source points may be placed around a domain D, either
15 using a symmetric shape or shapes which take the general shape of Γ obtained by
16 dilatation.
Fo
17
18
19 Collocation points t
t Source points
20 ΓE × (−T, T )
rP
T D
21 T D
x2
22 x2
Γ
23 0
24 0 x1
x1
ee
25
26 −T
27 −T
DE
28
rR
29
30 Figure 1. MFS for two-dimensional heat conduction, source points located outside of the spatial domain
31 D̄ and in time [−T, T ].
32
33
ev
34 We now construct the following infinite series

35
36 X ∞
∞ X
u∞ (x, t) = c(j) (j)
m vm (x, t), (7)
iew
37
38 j=1 m=1
39
40 (j)
41 where cm are set equal to zero except for a finite number of values. Note that, due
42 to the Heaviside function in (5), we have u∞ (x, t) = 0 for t ≤ τ = minm,j:|c(j)
m |6=0
τm .
43 Also note that, since F solves the heat equation, u∞ also satisfies the heat equation
On
44 in DT .
45
46
47 3.1 Denseness on the lateral surface
ly
48
49 We prove the following denseness result on the lateral surface Γ × (−T, T ):
50
(j)
51 Theorem 3.1 The set of functions {vm (x, t)}∞ j,m=1 restricted on Γ×(−T, T ) form
52 2
a linearly independent and dense set in L (Γ × (−T, T )).
53
54 Proof A similar version of the proof of this theorem was given in one-dimension
55 in [15] and in three-dimensions in [21], and we follow those ideas here in the two-
56 dimensional case.
57
Linear independence: Assume that we do not have linear independence, then
58 (j )
59 there exist positive integers N, m0 , j0 ∈ {1, . . . , N }, and a coefficient cm00 6= 0 such
60

1 that
2
3 N X
X N
4 c(j) (j)
m vm (x, t) = 0, (x, t) ∈ Γ × (−T, T ). (8)
5 j=1 m=1
6
7
8
Define the function
9 N
N X
10 X
11
U (x, t) = c(j) (j)
m vm (x, t), (x, t) ∈ D × (−T, T ). (9)
12 j=1 m=1
13
14 Then U satisfies the following equations:
15
16 ∂U
Fo
17
− ∆U = 0, in D × (−T, T ), (10)
∂t
18
19
20 U (x, t) = 0, (x, t) ∈ Γ × (−T, T ), (11)
rP
21
22
23
24 U (x, −T ) = 0. (12)
ee
25
26 We have obtained the above equations by observing that the fundamental solu-
27 tion satisfies (10); the Heaviside function makes the fundamental solution equal to
28 zero in equation (12), and (8) gives us (11). By the uniqueness Theorem 2.1, the
rR
29
only solution to the equations (10)–(12) is U (x, t) ≡ 0 for (x, t) ∈ D × (−T, T ).
30
31
Because U is analytic in DE × (−T, T ), we also have U (x, t) = 0 for (x, t) ∈
32 DE × (−T, T ), see [24].
33 We now let the point (x, t) approach the point (yj0 , τm0 ) ∈ ΓE × (−T, T ) such
ev
34 that the ratio

35
36 |x − yj0 |2
(13)
iew
37 4(t − τm0 )
38
39 (j ) (j )
40 remains bounded. Then the summand cm00 vm00 (x, t) in (8) may be made as large
41 as we wish, while the other terms in the series (8) remain bounded; this gives
42 us a contradiction and thus, we have linear independence for the set of functions
(j)
43 {vm (x, t)}∞ 2
j,m=1 in L (Γ × (−T, T )).
On
44 (j)
45 Denseness: We next prove that the sequence {vm (x, t)}∞ j,m=1 is a dense set in
2
L (Γ × (−T, T )). Assume on the contrary that it is not a dense set. Then there
46
47 exists an element f (x, t) in L2 (Γ × (−T, T )), which we can assume is continuous,
ly
48 such that
49
50 Z T Z
(j)
51 vm (x, t)f (x, t) dx dt = 0, j, m = 1, 2, . . . (14)
52 −T Γ
53
(j)
54 To show that {vm (x, t)} is dense we have to show that f (x, t) ≡ 0 in (14). From
55 definition (6), equation (14) can be rewritten as
56
57 Z T Z
58 F (x, t; yj , τm )f (x, t) dx dt = 0, j, m = 1, 2, . . . , (15)
59 τm Γ
60

1 where the Heaviside function in (5) has been used to reduce the range of integration
2 with respect to the time variable. We introduce an equivalent form of the classical
3
single-layer heat potential given by
4
5 Z T Z
6
V (y, τ ) = F (x, t; y, τ )f (x, t) dx dt, y 6∈ Γ. (16)
7 τ Γ
8
9 It is well-known that V (y, τ ) is a smooth solution to the heat equation in the
10
exterior of D̄ × (−T, T ) and it cannot vanish on any surface in this exterior region
11
12 without being identically zero. Thus, by the continuity of F and (15), we find that
13 V (y, τ ) = 0 for (y, τ ) ∈ ΓE × (−T, T ), which is in the exterior of D̄ × (−T, T );
14 we then conclude that V = 0 in the exterior of D̄ × (−T, T ). Moreover, since V is
15 continuous across Γ × [−T, T ] we also have V (y, τ ) = 0 on Γ × [−T, T ]. This implies
16 that V = 0 also in D̄ × (−T, T ) since V satisfies the heat equation in D × (−T, T ).
Fo
17 Finally, using the jump relations for the normal derivative of V on Γ × [−T, T ], see
18 [9, p. 133], we get
19
20
rP
21
1 ∂V (x, t)
f (x, t) ± = 0, (x, t) ∈ Γ × (−T, T ), (17)
22 2 ∂ν
23
24 where ν represents the unit normal on the surface Γ × (−T, T ). Thus, f ≡ 0 and
(j)
ee
25 therefore, {vm (x, t)}∞ 2

j,m=1 is a dense set in L (Γ × (−T, T )).
26
27
28
3.2 Denseness on the base surface
rR
29
30 We now show that we also have denseness on the “base” surface D × {0}, where
31 the initial condition is imposed in (1)–(3).
32
33 (j) (j)
Theorem 3.2 The set of functions {vm (x, 0)}∞
ev
j,m=1 , where vm (x, t) is given by

34
35
(6) with τm < 0, form a linearly independent and dense set in L2 (D).
36 Proof The method of proof is similar to that used in [15] in one-dimension, and
iew
37 we give it here, for completeness, in higher dimensions.

38
Linear independence: Assume that we do not have linear independence, then
39 (j )
40 there exist positive integers N, m0 , j0 ∈ {1, . . . , N }, and a coefficient cm00 6= 0 such
41 that
42
43 N X
N
On
X
44 c(j) (j)
m vm (x, 0) = 0, x ∈ D. (18)
45 j=1 m=1
46
47 We shall use the corollary of Theorem 3 in [17], which guarantees that if u(x, t)
ly
48 2
is a smooth solution of the heat equation in R2 that is bounded by Beβ|x| and if
49
50 u(x, 0) = ǫ then, |u(x, t)| ≤ ǫ, for 0 ≤ t ≤ T. From (9) and (18) we have that
51
52 U (x, 0) = 0, x ∈ D. (19)
53
54 We also have that U satisfies (10) and (12). Now, since U (x1 , x2 , 0), where x =
55
(x1 , x2 ), is a real analytic function in each of the variables x1 and x2 , we find
56
57
that U (x, 0) = 0 for every x ∈ R2 , see [20, p. 14]. Moreover, since each τm < 0,
58 U is continuous on R2 × [0, T ] and is at least twice continuously differentiable in
59 R2 × [0, T ]. Furthermore, U also satisfies the heat equation (1), and the following
60

1 inequality clearly holds

2
3 2
4 |U (x, t)| ≤ Beβ|x| , t ∈ [0, T ], (20)
5
6
7
for some positive constants B and β. Thus, from Theorem 3 of [17], and its corollary,
8 we conclude that U (x, t) = 0 in R2 × [0, T ]. In particular, from [24], we may extend
9 U such that U (x, t) = 0 also in DE × [−T, T ].
10 We now let the point (x, t) approach the point (yj0 , τm0 ) ∈ ΓE × [−T, 0] such
11 (j ) (j )
that the ratio (13) remains bounded. Now, the summand cm00 vm00 (x, t) may be
12 made as large as we wish, while the other terms in the series (18) remain bounded;
13
this gives a contradiction and we have linear independence for the set of functions
14 (j)
15 {vm (x, 0)} in L2 (D).
(j)
16 Denseness: We shall show that the set of functions {vm (x, 0)}, where τm < 0,
Fo
17 2
is a dense set in L (D). Assume that this is not a dense set, then there exists a
18 function f ∈ C 2 (D) such that
19
20
rP
Z
21 (j)
vm (x, 0)f (x) dx = 0, j, m = 1, 2, . . . (21)
22 D
23
24
We let w be a weak solution of the heat equation (1), see [7], with initial condition
ee
25
26 w(x, 0) = f (x) and boundary condition w(x, t) = 0 for (x, t) ∈ ΓT . We may
27 transform (21) using Green’s identities, see [22], into the following form
28
rR
29 T
∂w(x, t)
Z Z
30 (j)
vm (x, t) dxdt = 0, j, m = 1, 2, . . .
31 0 Γ ∂ν
32
33 where ν is the outward pointing unit normal to Γ. From Theorem 3.1 we know
ev
34 (j)
that {vm (x, t)} restricted on ΓT is a dense set in L2 (ΓT ), and we may conclude
35
36 that the normal derivative of w is zero on ΓT . Therefore, both w and ∂w ∂ν are zero
on ΓT . From [24], we conclude that w(x, t) = 0 for (x, t) ∈ D̄T ; hence f ≡ 0, and
iew
37
(j)
38 {vm (x, 0)}, where τm < 0, is a dense set in L2 (D).
39
40
41
42 4.The MFS for the heat equation in two-dimensions
43
On
44 The denseness results proved in the previous section, Theorems 3.1 and 3.2, which
45
involved linear combinations of the fundamental solution (5) of the heat equation
46
47
(1), enable us to describe a method for approximating the solution to the problem
(1)–(3). We note that the MFS we propose may be applied to domains of general
ly
48
49 shape and size and the source points may also be placed arbitrarily, for example,
50 placed symmetrically on a circular pseudo-boundary, or to match similarly the
51 general shape of the domain D, see Figure 2 and [12]. The only restriction on the
52 placement of the source points is that they are located on the boundary ΓE outside
53 the domain D, and placed relatively close to D such that u has no singularity in
54 DE × [0, T ]. Also, it might be more practical to take the sources on the interval
55
(−ǫ, T ), where ǫ > 0 is small, instead of the full interval (−T, T ). However, some
56
57
preliminary numerical investigations in [14] for the backward heat conduction prob-
58 lem showed that ǫ cannot be chosen too small if no loss in accuracy and stability
59 is to be secured.
60

1 x2 x2
2 Source P oints
3
4
5 0 x1 0
D x1
6 D
7
ΓE ΓE
8
9
10 Figure 2. Arbitrary domain with varying source point locations restricted to any contour ΓE embracing
the given solution domain D.
11
12
4.1 A direct MFS for the two-dimensional heat equation
13
14 We search for an approximation to the solution of equations (1)–(3) in the following
15 form:
16
Fo
17 2M X
X N
18 uM,N (x, t) = c(j)
m F (x, t; yj , τm ), (x, t) ∈ D̄T . (22)
19 m=1 j=1
20
rP
21 For simplicity, let us describe the MFS in the case of circular domains. We shall
22
also consider rectangular domains in the next section, see Example 3 and 4.
23
24 We consider a two-dimensional circular domain D, with boundary Γ and radius
r0 > 0, centred at the origin, and let us place the source points (yj )j=1,N on a
ee
25
26 circle r0 + h, h > 0, also centred at the origin. The parameter h > 0 will be chosen
27 such that the error at the lateral and base surfaces is minimized (viz maximum
28 principle for the heat equation).
rR
29 Take the time points (τm )m=1,2M (each in the interval (−T, T )) as given by
30
31
2(m − M ) − 1
32 τm = T, m = 1, . . . , 2M,
33 2M
ev
34
35 and using polar coordinates, place the source points in space at
36
2πj
iew
37 yj = (r0 + h, θj ) = r0 + h, , j = 1, . . . , N.
38 N
39
40 In polar coordinates equation (22) is now represented by
41
42 2M X
X N
43 uM,N (r, θ, t) = c(j)
m F (r, θ, t; r0 + h, θj , τm ). (23)
On
44 m=1 j=1
45
46 We have located N × 2M source points in total outside the domain D and in
47
time; we place the same number of collocation points in total on Γ̄T ∪ (D × {0}),
ly
48
49 the boundary in time and the domain at time t = 0. Of course, the location of
50 source and collocation points may be chosen arbitrarily, here we choose points for
51 ease of calculation. Let
52
53 i
ti = T, i = 0, . . . , M,
54 M
55
56 and on Γ set
57
58 2πk
(r0 , θk ) = r0 , , k = 1, . . . , N.
59 N
60

1 x2
2
3
t Γ
4 Γ
r0+h
D
5 T
r0
6
7 x2 D x1
8
9 0
x1
10 t=0
11
12
13 DE
−T
Source points
14
15 ΓE Collocation points
16
Fo
17
Figure 3. Location of source and collocation points when D is a circular domain.
18
19
20
We have located N × (M + 1) collocation points on the boundary, the remaining
rP
21 N × (M − 1) points will be located on D when t = 0. We consider M − 1 circles of

22 radius
23 1
24 l 2
rl = r0 , l = 1, . . . , M − 1,
ee
25 M
26
27 where the square root has been introduced to spread the points out within the
28
domain, and not to cluster them at the centre. We place N equally spaced points
rR
29
30 on each circle such that
31
2πk
32 (rl , θk ) = rl , , k = 1, . . . , N,
33 N
ev
34
35 see Figure 3 for a detailed graphical representation of the position of the various
36 source and collocation points given above.
iew
37 We now impose the boundary and initial conditions (2) and (3) so that we can
38 (j)
determine the unknown coefficients cm in (23). In polar coordinates we obtain the
39
40
equations
41
42 uM,N (r0 , θk , ti ) = h(r0 , θk , ti ), (24)
43
On
44
45 uM,N (rl , θk , 0) = u0 (rl , θk , 0), (25)
46
47 where k = 1, . . . , N, i = 0, . . . , M and l = 1, . . . , M − 1.
ly
48
The system of equations (24) and (25) contains N ×(M −1)+N ×(M +1) = 2M N
49
50 equations and 2M N unknowns, therefore, we may obtain a unique solution. We
51 can represent this system of equations as
52
53 Ac = g, (26)
54
55 (j)
where c is the vector of unknowns cm , g is the vector representing the values of
56
57
the functions u0 and h at the respective collocation points, and A is the matrix
58 corresponding to the value of the fundamental solution at the points outlined above.
59 For certain boundary collocation and source points it might be possible to use the
60

1 properties of circulant matrices to develop a matrix decomposition algorithm [18]

2 for the solution of the system of equations (26) and thus, substantially reduce the
3
computational cost. This could be considered in a future investigation.
4
5 To solve this system it is possible to solve it directly by Gaussian elimination.
6 However, it is well-known that employing the MFS can yield matrices with large
7 condition numbers [6, 23], as h increases. Then, in such a situation instead of (26)
8 it may be necessary to consider the Tikhonov regularization
9
10 (Atr A + λI)c = Atr g, (27)
11
12 where the superscript tr denotes the transpose of a matrix and I is the identity
13
matrix. The linear system of equations (27) is solved using a Gaussian elimination
14
15 method (employed backslash “\” command in MATLAB). In (27), λ > 0 is a small
16 regularization parameter (usually in the interval [10−1 , 10−16 ]) can be chosen by
Fo
17 trial and error, namely start with a large value of λ, say λ = 10−1 , and then
18 decrease it gradually as λ = 10−2 , 10−3 , . . . until an oscillatory solution starts to
19 develop. To choose λ one can also use the L-curve criterion of [13]. In a future
20 study, it would be interesting to investigate the dependence of the solution on λ
rP
21 and h.
22
23
24
5. Numerical results
ee
25
26
27 In [15] it was shown that the direct MFS approximation applied to the one-
28 dimensional heat equation with source points located outside the domain and in
rR
29 time is accurate. Below, we present numerical results for approximations in two-

30 dimensional domains, such as circular and square domains. In order to assess the
31 accuracy of the numerical MFS solutions we compare them with the available exact
32 solutions for various benchmark test examples. Numerical results are presented for
33
ev
N = 20 and M = 30 points, which were found sufficiently large to ensure that any
34
further increase in these numbers did not significantly improve the accuracy of the
35
36
numerical solution without affecting its stability.
iew
37
38
5.1 Example 1
39
40 Let D = {x : |x|2 < 1}, DT = {(x, t) : |x|2 < 1, t ∈ (0, 1]}, and ΓT = {(x, t) :
41 |x|2 = 1, t ∈ (0, 1]}. We solve the following problem, using the direct MFS laid out
42
in the previous section,
43
On
44
45
∂u(x, t)
− ∆u(x, t) = 0, (x, t) ∈ DT , (28)
46 ∂t
47
ly
48
49 u(x, t) = 4t + 1, (x, t) ∈ Γ, (29)
50
51
52 u(x, 0) = |x|2 , x ∈ D. (30)
53
54 The exact solution of problem (28)–(30) is u(x, t) = 4t + |x|2 . The source points
55
are placed on a circle with radius 1 + h. The value of h > 0 will be chosen ap-
56
57
propriately. However, the accuracy of the approximation appears to decrease when
58 h < 0.25 or h > 4. In Figure 4 the exact solution and the MFS approximations are
59 plotted in one-dimension, x = (x1 , 0), for times t ∈ {0.2, 0.8} with λ = 10−8 in the
60

1 Tikhonov regularization. As can be seen in Figure 4 the approximation obtained

2 is accurate, and was stable even when considering other small positive values of λ,
3
however, for stable results we could not take λ = 0 due to the round-off precision
4
5 errors.
6
4.5
7
8 4 u(x1 , 0, 0.2)
9 3.5
10 3
11
2.5
12
13 2
14 1.5
u(x1 , 0, 0.8)
15 1
16
Fo
0.5
17 −1 −0.5 0 0.5 1
18 x1
19
20 Figure 4. The exact solution (—) and the approximate values (∗) for u(x1 , 0, 0.2) and u(x1 , 0, 0.8), for
rP
Example 1.
21
22
23 Figure 5 contains plots of the exact solution and the direct MFS approximations
24 for h ∈ {0.5, 4} and λ = 10−8 . From this figure it can be seen that the numerical
results obtained with h = 0.5 are slightly more accurate than those obtained with
ee
25
26 h = 4.
27
28 2
rR
29
30 1.8
31
u(x1 , 0, 0.2)
1.6
32
33
ev
1.4
34
35 1.2
36 1
iew
37
38 0.8
−1 −0.5 0 0.5 1
39 x1
40
41 Figure 5. The exact solution (—) and the approximate values for u(x1 , 0, 0.2) obtained with h = 0.5 (∗)
42 and h = 4 (◦), for Example 1.
43
On
44 Finally, we consider a three-dimensional plot of the exact solution u(x1 , x2 , 0.8)

45 in Figure 6(a), and the MFS approximation uM,N in Figure 6(b) obtained with
46 h = 1. Figure 6(c) shows the graph of the absolute error, and we note that the
47
approximation is very accurate with a maximum absolute error of O(10−5 ).
ly
48
49
50
5.2 Example 2
51
52 In this example we choose the same D, DT and Γ as in Example 1, but instead of
53 u(x, t) = 4t + |x|2 , we consider the exact solution of the equations (1)–(3) given by
54
55
u(x, t) = ex1 +x2 cos(x1 + x2 + 4t), (31)
56
57
58 where the boundary and initial equations (2) and (3) have been obtained from
59 (31). Note that this function is not constant on circles centred at the origin, and
60

1 (a) (b) (c)

2 −5
uerror (x1 , x2 , 0.8)

x 10
4.2 4.2 6
u(x1 , x2 , 0.8)
u(x1 , x2 , 0.8)
3 replacemen 4 4
4
4 3.8
3.6
3.8
3.6
2
5 3.4 3.4
0
6 3.2
1
1
3.2
1
1
1
1
0.5
7 x2
0
0
x1 x2
0
−0.5
0
x1 x2
0
0
−1 −1 −1 −1 −1 −1 x1
8
9 Figure 6. Three-dimensional plot of: (a) The exact solution u(x1 , x2 , 0.8), (b) the approximate solution
10 uM,N obtained with λ = 10−8 and h = 1, and (c) the absolute error, for Example 1.
11
12 not symmetric, thus being different in character compared with the solution in
13 Example 1.
14 Again, the source points will be placed on a circle with radius 1 + h, with final
15
time point T = 1. As an alternative to the strategy for choosing λ described at the
16
Fo
17 end of Section 4.1, we employ the L-curve criterion [13]. In Figure 7 we present a
18 plot of the L-curve for Example 2, where the residual is plotted against the 2-norm
19 of the solution c. We choose the regularization parameter λ, which corresponds to
20 the corner of the “L” in Figure 7; namely, in this example we take λ = 10−8 .
rP
21
22 1.4
10
λ = 10−12
23
24
ee
25 1.3
10
26
k c k2
27 1.2
10
28
rR
29 λ = 10−10
30 1.1
10
λ = 10−8 λ = 10−5
31 λ = 10−1
32
−5 −4 −3 −2 −1 0 1
33 10 10 10 10 10 10 10
ev
k Ac − g k2
34
35
36 Figure 7. Plot of the L-curve, for Example 2.
iew
37
38 In Figure 8 the exact solution and the direct MFS approximations for λ = 10−8
39 are plotted in one-dimension, x = (x1 , 0), for times t = {0.2, 0.8}, with h = 1.
40 It is clear from Figure 8 that the approximation is very accurate. Also, varying λ
41 between [10−12 , 10−4 ] only slightly changes the accuracy of the approximation.
42
43
On
1
44 u(x1 , 0, 0.2)
45
0.5
46
47
ly
0
48
49
−0.5
50
51
−1
52
53 u(x1 , 0, 0.8)
−1.5
54 −1 −0.5 0 0.5 1
55 x1
56
57 Figure 8. The exact (—) and the approximate values (∗) for u(x1 , 0, 0.2) and u(x1 , 0, 0.8), for Example 2.
58
59 Finally, we consider a three-dimensional plot of the exact solution u(x1 , x2 , 0.8)
60

1 in Figure 9(a), the MFS approximation uM,N in Figure 9(b), and the absolute error
2 in Figure 9(c), obtained with h = 1.
3
4 (a) (b) (c) x 10
5 1
−4
u(x1 , x2 , 0.2)
u(x1 , x2 , 0.2)
1
uerror (x1 , x2 , 0.2)

1.5
6 0
0
7 −1
−1 1
8 −2
−2
−3 0.5
9 −1
−0.5
−3
−1
−0.5
10 x1
0
0.5 0.5
1
x1
0
0.5 0
0.5
1
0 0
1
0 −1
11 1 −1
−0.5
x2 1 −1
−0.5
x2 0 1 −1 x2
x1
12
13 Figure 9. Three-dimensional plot of: (a) The exact solution u(x1 , x2 , 0.2), (b) the approximate solution
uM,N obtained with λ = 10−8 and h = 1, and (c) the absolute error, for Example 2.
14
15
16 The plots obtained in Examples 1 and 2 show that the MFS approximation is
Fo
17 accurate in circular domains, with errors usually in the interval [10−4 , 10−2 ], for a
18 wide range of parameters h and λ.
19
20
rP
21 5.3 Example 3
22
23 In the next two examples we consider square domains with edge length L. In
24 Examples 3 and 4 source points will be placed on both squares, see Figure 10, as
ee
25 well as circles; we vary the shapes where we place the source points to highlight
26 that the placement of the source points do not need to follow the shape of Γ.
27 The numerical implementation is the same as in the previous examples, including
28 the placement of the sources and boundary collocation points and their numbers
rR
29
N = 20, M = 30.
30
31
32 t x2
33 ΓE
ev
34 D
T
35 L Γ
36
iew
37 x1
38
39
40
41 −T
42
43 L+h
On
44
45 Figure 10. MFS in two-dimensions for square domains.
46
47
The following problem was considered in [11]. We take D = (−0.2, 0.2) ×
ly
48
49 (−0.2, 0.2) to be a square of edge length L = 0.4, take T = 0.9, and solve
50
51 ∂u(x, t)
52
− α∆u(x, t) = 0, (x, t) ∈ DT = D × (0, 0.9], (32)
∂t
53
54
55 u(x, t) = 0, (x, t) ∈ ΓT = ∂D × (0, 0.9], (33)
56
57
58
59 u(x, 0) = 1, x ∈ D. (34)
60

1 Here, we have thermal diffusivity α = 1000/5.8, and the fundamental solution of

2 equation (32) is now given by
3
4
H(t − τ ) − 4α(t−τ
|x−y|2
5 F (x, t; y, τ ) = e ). (35)
6 4πα(t − τ )
7
8 We note that in this example the compatibility conditions (4) are violated. The
9 exact solution to the problem (32)–(34) is given by, see Carslaw and Jaeger (1959),
10
"∞
11 16 X (−1)n −α(2n+1)2 tπ2 /(4(L/2)2 )

(2n + 1)πx1
#
12 u(x1 , x2 , t) = 2 e cos ×
13 π 2n + 1 2(L/2)
n=0
14 " ∞ # (36)
15 X (−1)m
(2m + 1)πx2
−α(2m+1)2 tπ 2 /(4(L/2)2 )
e cos .
16 2m + 1 2(L/2)
Fo
m=0
17
18
19
However, when we plot the exact (using 100 terms in the series expansion (36))
20 and the approximate solutions at the point (x1 , x2 ) = (0, 0) for time t ∈ [0, 0.9],
rP
21 with source points placed on a square with edge length 0.4 + h, h = 5, λ = 10−8 ,
22 we get a large discrepancy, see Figure 11.
23
24 1.2
ee
25 uM,N
1 u
26
27 0.8
28
u(0, 0, t)
rR
0.6
29
30 0.4
31 0.2
32
33 0
ev
34 −0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
35 t
36
iew
37 Figure 11. The exact solution u(0, 0, t) and the approximation uM,N , as functions of time t ∈ [0, 0.9], for
38 Example 3.
39
40 Changing the parameter λ in the Tikhonov regularization (27) did not seem to
41 improve the approximation, however, we observe that the exact solution (36) decays
42 very rapidly due to the exponential terms. This means that we should consider
43
On
a much smaller time interval. Figure 12 shows the exact and the approximate
44 solutions, with final time point T = 0.0006, plotted over t ∈ [0, 0.0004], where
45
46
λ = 10−6 and source points have now been placed on a circle of radius h = 0.84.
47 These figures show that the choice of the final time T , in particular when con-
sidering fast decaying functions, is also important when implementing the MFS.
ly
48
49 Time-marching methods, [25], could then perhaps be used to extend our approxi-
50 mation to larger time intervals.
51
52
53 5.4 Example 4
54
55 We consider D = (0, 1) × (0, 1) and DT = D × (0, 3], and we wish to solve the
56 following problem:
57
58 ∂u(x, t)
− ∆u(x, t) = 0, (x, t) ∈ DT , (37)
59 ∂t
60

1
1.4
2 uM,N
3 1.2 u
4
1
5
u(0, 0, t)
6 0.8
7 0.6
8
9 0.4
10 0.2
11
0
12 0 1 2 3
−4
4
x 10
13 t
14
15 Figure 12. The exact solution u(0, 0, t) and the approximation uM,N with T = 0.0006, for Example 3.
16
Fo
17 √
πx π 1

−π 2 t/4 1
18 u(x1 , 0, t) = u(x1 , 1, t) = 2e cos − + √ , x1 ∈ (0, 1), t ∈ (0, 3],
19 2 4 2
20 (38)
rP
21
√
πx
22 −π 2 t/4 2 π 1
u(0, x2 , t) = 2e cos − + √ , x2 ∈ (0, 1), t ∈ (0, 3], (39)
23 2 4 2
24
ee
25
26
∂u π 2
27 (0, x2 , t) = e−π t/4 , x2 ∈ (0, 1), t ∈ (0, 3], (40)
28 ∂x1 2
rR
29
30
31 √ h πx
1 π πx
2 π i
u(x1 , x2 , 0) = 2 cos − + cos − , (x1 , x2 ) ∈ D. (41)
32 2 4 2 4
33
ev
34 This is an inverse problem with missing boundary data at x1 = 1, which we wish

35 to determine using the Cauchy boundary data over-specification at x1 = 0. The
36 exact solution of problem (37)–(41) is
iew
37
38 √

−π 2 t/4 πx1 π πx2 π
39 u(x1 , x2 , t) = 2e cos − + cos − , (x, t) ∈ DT . (42)
40 2 4 2 4
41
42 In this example, we shall show results when the source points are placed on
43 squares, as well as when we place them on circles, to show that placement of the
On
44 sources does not need to follow the shape of the solution domain. When the source
45 points are placed on a square they will be located at (−h/2, 1 + h/2) × (−h/2, 1 +
46 h/2), whilst the source points placed on a circle will have radius h with centre at
47
(0.5, 0.5), see Figure 13.
ly
48
49
50 (a) (b)
1 + h2
51
1 1
52
53 h
54
0 0
55
56 − h2
57
58
Figure 13. Examples of source point location for the unit square solution domain of Example 4.
59
60

1 Figures 14(a) and 14(b) show the exact solution u(1, x2 , 1.5) and its normal
2 ∂u
derivative ∂x (1, x2 , 1.5), respectively, in comparison with the approximate solu-
3 1
4 tions, obtained with h = 3 and λ = 10−8 .

5
(a) (b)
6 0.06 −0.038
7 0.058
8
∂x1 (1, x2 , 1.5)

−0.0385
u(1, x2 , 1.5)
0.056
9
0.054 −0.039
10
11 0.052
∂u
−0.0395
12 0.05
13 0.048
0 0.2 0.4 0.6 0.8 1
−0.04
0 0.2 0.4 0.6 0.8 1
14 x2 x2
15 Figure 14. (a) The exact solution u(1, x2 , 1.5) (—) and the MFS approximation uM,N with h = 3 (∗).
16 ∂u
Fo
(b) The exact normal derivative ∂x (1, x2 , 1.5) (—) and the MFS approximation (◦), for Example 4.
1
17
18
19 Figure 15 shows the exact solution u(x1 , x2 , 0.5) and the MFS approximation
20 uM,N . Note that for the approximation uM,N in Figure 15(c) we have instead
rP
21 placed the source points on a circle and there is still good agreement with the
22 exact solution. In Figure 16 we present plots of the absolute error at time t = 2.5
23 for two different values of h. From this figure it can be seen that the error increases
24
when source points have been placed too close to the boundary.
ee
25
26
27 (a) (b) (c)
uM,N (x1 , x2 , 0.5)
uM,N (x1 , x2 , 0.5)

u(x1 , x2 , 0.5)
28 0.9
0.9 0.9
rR
0.8
29 0.7
0.8 0.8
0.7 0.7
30 0.6
0.6 0.6
31 0.5
1 0.5 0.5
1 1
32 0.5
0.5
1
0.5
0.5
1
0.5
0.5
1
33 x2 x1 x2 x1 x2 x1
ev
0 0
0 0 0 0
34
35 Figure 15. (a) The exact solution u(x1 , x2 , 0.5), (b) the MFS approximation uM,N using source points
36 placed on a square, h = 3, and (c) the MFS approximation uM,N using source points placed on a circle
with radius h = 3 and centre (0.5, 0.5), for Example 4.
iew
37
38
39
40 (a) (b)
41 x 10
−4
uerror (x1 , x2 , 2.5)
uerror (x1 , x2 , 2.5)
1.5 0.015
42
43
On
1 0.01
44 0.5 0.005
45
0 0
46 1
1
1
1
47 0.5
0.5
0.5
0.5
x2 x1 x2 x1
ly
0 0 0 0
48
49
50 Figure 16. The absolute error |u(x1 , x2 , 2.5) − uM,N (x1 , x2 , 2.5)| when the MFS approximation uM,N has
been generated using source points placed on a square with: (a) h = 3 and (b) h = 1, for Example 4.
51
52
53 Finally, in Figure 17, random noise simulating measurement errors, have been
54 added to the Dirichlet boundary data (39) as follows:
55
56 uδ (0, x2 , t) = u(0, x2 , t) + N (0, σ 2 ),
57
58
59 where N (0, σ 2 ) represents the normal distribution with mean zero and standard
60

1 deviation
2
3
4 σ=δ× max |u(0, x2 , t)|,
(x2 ,t)∈(0,1)×(0,3)
5
6 and δ is the relative noise level. A set of ten noisy random data functions
7
{uδk (0, x2 , t)}k=1,10 was generated, and the source points in the MFS have been
8
9 placed on a circle of radius h, centred at (1/2, 1/2). Figure 17(a) presents a plot
10 of the exact solution u(1, x2 , 0.5), and the best (∗) and the least accurate (◦) MFS
11 approximations from these ten data sets, obtained with δ = 3% noise, h = 3, T = 3
12 and λ = 10−4 . In Figure 17(b) we present a three-dimensional plot of the absolute
13 error when δ = 3%, h = 3, λ = 10−4 and t = 0.5.
14
15 (a) 0.74 (b)
16
Fo
0.72
17 0.03
uerror (x1 , x2 , 0.5)

0.7
18
u(1, x2 , 0.5)
0.68
19 0.02
0.66
20
rP
0.64 0.01
21
0.62
22 0
0.6 1
23 1
0.58
24 0.5
0.5
ee
25 0 0.2 0.4 0.6 0.8 1 x2 0 0 x1

x1
26
27 Figure 17. (a) The exact solution u(1, x2 , 0.5) (—) and the best (∗) and the least accurate (◦) MFS
28 approximations from ten different sets of noisy data with noise level δ = 3% and (b) the absolute error
rR
|u(x1 , x2 , 0.5) − uM,N (x1 , x2 , 0.5)| when δ = 3%, for one of the noisy data sets, for Example 4.
29
30
31 Adding more noise such as δ = 5% did not significantly change the stability
32 of the numerical results provided that regularization is applied appropriately. As
33 expected, the accuracy decreases when the noise level increases and the regularizing
ev
34 parameter usually has to take a larger value. Noise can also be added to the other
35 data functions, such as the Neumann data (40), and the same stable and accurate
36
numerical results are expected. Thus, for this inverse problem, the regularized MFS
iew
37
38 is a stable approximation with respect to noisy data.
39
40
41 6. Conclusions
42
43 We have investigated the application of the MFS for linear heat conduction in two-
On
44 dimensional conducting bodies with Dirichlet boundary conditions. The solution

45
is sought in the form of a linear combination of fundamental solutions of the heat
46
47
equation with the source points placed outside the body and in time as well. The-
oretical properties, such as denseness of the approximation on the boundary of the
ly
48
49 body, have been obtained in the space of square integrable functions. Moreover,
50 numerical results are presented for both circular and rectangular configurations
51 showing that accurate approximations can be obtained when the source points are
52 placed on circles or rectangles at an appropriate distance from the solution do-
53 main. The numerical results show that accurate approximations can be obtained
54 at small computational cost, and that the method is not too sensitive with respect
55
to the placement of the sources. Numerical results have also been included for an
56
57
inverse problem, where overspecified data was given on a part of the boundary of a
58 body and the solution was to be reconstructed on the remaining part. Good agree-
59 ment with the exact available analytical solution was also obtained in this case,
60

REFERENCES 17
1 showing that boundary conditions other than Dirichlet can be handled. Exten-
2 sions to the three-dimensional time-dependent heat equation are straightforward
3
by accordingly modifying the fundamental solution (5).
4
5
6
7
8 6.1 Acknowledgements
9 The authors would like to thank the referees for their suggestions. The third author
10
acknowledges financial support from EPSRC in the form of a Doctoral Training
11
12 Grant (DTG).
13
14
15 References
16
Fo
17 [1] K. Balakrishnan and P. Ramachandran, The method of fundamental solutions for linear diffusion-
reaction equations, Math. Computer Modelling 31 (2000), pp. 221–237.
18 [2] A. Bogomolny, Fundamental solutions method for elliptic boundary value problems, SIAM J. Numer.
19 Anal. 22 (1985), pp. 644–669.
20 [3] H. Carslaw and J. Jaeger, Conduction of Heat in Solids, Oxford University Press, London (1959).
rP
[4] S. Chantasiriwan, Methods of fundamental solutions for time-dependent heat conduction problems,
21 Int. J. Numer. Meth. Engng. 66 (2006), pp. 147–165.
22 [5] S. Chantasiriwan, B. Johansson, and D. Lesnic, The method of fundamental solutions for free surface
23 Stefan problems, Engng. Anal. Boundary Elements 33 (2009), pp. 529–538.
24 [6] C. Chen, H. Cho, and M. Golberg, Some comments on the ill-conditioning of the method of funda-
mental solutions, Engng. Anal. Boundary Elements 30 (2006), pp. 405–410.
ee
25 [7] L. Evans, Partial Differential Equations, American Mathematical Society, Providence, Rhode Island
26 (1998).
27 [8] G. Fairweather and A. Karageorghis, The method of fundamental solutions for elliptic boundary value
problems, Adv. Comput. Math. 9 (1998), pp. 69–95.
28 [9] A. Friedman, Partial Differential Equations of Parabolic Type, Prentice-Hall Inc., Englewood Cliffs,
rR
29 NJ (1964).
30 [10] M. Golberg and C. Chen, The method of fundamental solutions for potential, Helmholtz and diffu-
sion problems, in Boundary integral methods: numerical and mathematical aspects, M. Golberg, ed.,
31 Computational Mechanics Publications, Boston (1999), pp. 103–176.
32 [11] M. Golberg, C. Chen, and A. Muleshkov, The method of fundamental solutions for time-dependent
33 problems, in Proc. 13th Int. Conf. Boundary Element Technology, C.B. C.S. Chen and D. Pepper,
ev
eds., WIT Press, Southampton (1999), pp. 377–386.

34 [12] P. Gorzelanczyk and J.A. Kolodziej, Some remarks concerning the shape of the source contour with
35 application of the method of fundamental solutions to elastic torsion of prismatic rods, Engng. Anal.
36 Boundary Elements 32 (2008), pp. 64–75.
[13] P. Hansen, Analysis of discrete ill-posed problems by means of the L-curve, SIAM review 34 (1992),
iew
37
pp. 561–580.
38 [14] Y. Hon and M. Li, A discrepancy principle for the source points location in using the MFS for solving
39 the BHCP, Int. J. Comput. Meth. 6 (2009), pp. 181–197.
40 [15] B. Johansson and D. Lesnic, A method of fundamental solutions for transient heat conduction, Engng.
Anal. Boundary Elements 32 (2008), pp. 697–703.
41 [16] ———, A method of fundamental solutions for transient heat conduction in layered materials, Engng.
42 Anal. Boundary Elements 33 (2009), pp. 1362–1367.
43 [17] R. Johnson, A Priori Estimates and Unique Continuation Theorems for Second Order Parabolic
On
Equations, Trans AMS 158 (1971), pp. 167–177.

44 [18] A. Karageorghis, Y.S. Smyrlis, and G. Fairweather, Matrix decomposition MFS algorithms, Interna-
45 tional Workshop on MeshFree Methods, (6 pages).
46 [19] J. Kolodziej, J. Stefaniak, and M. Kleiber, Transient heat conduction by boundary collocation methods
47 and FEM - A comparison study, In: Notes on Numerical Fluid Mechanics, pp. 104-115, Braunschweig,
Vieweg Verlag (1995).
ly
48 [20] S. Krantz and H. Parks, A primer of real analytic functions, 2nd ed., Birkhauser, Boston (2002).
49 [21] V. Kupradze, A method for the approximate solution of limiting problems in mathematical physics,
50 USSR Comput. Maths. Math. Phys. 4 (1964), pp. 199–205.
[22] J.L. Lions and E. Magenes, Non-homogeneous Boundary Value Problems and Applications, vol. 1,
51 Springer, Berlin (1972).
52 [23] P. Ramachandran, Method of fundamental solutions: singular value decomposition analysis, Commun.
53 Numer. Methods Engng. 18 (2002), pp. 789–801.
[24] J.C. Saut and B. Scheurer, Unique continuation for some evolution equations, J. differential equations
54 66 (1987), pp. 118–139.
55 [25] S. Valtchev and N. Roberty, A time-marching MFS scheme for heat conduction problems, Engng.
56 Anal. Boundary Elements 32 (2008), pp. 480–493.
57 [26] D. Young, C. Tsai, K. Murugesan, and C. Fan, Time-dependent fundamental solutions for homoge-
neous diffusion problems, Engng. Anal. Boundary Elements 28 (2004), pp. 1463–1473.
58
59
60

A Method of Fundamental Solutions For Two-Dimensional Heat Conduction

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Method of Fundamental Solutions For Two-Dimensional Heat Conduction

Uploaded by

Copyright:

Available Formats

A method of fundamental solutions for two-dimensional

To cite this version:

HAL Id: hal-00678795

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

A method of fundamental solutions for two-dimensional

Journal: International Journal of Computer Mathematics

Manuscript ID: GCOM-2009-0885-B.R1

Manuscript Type: Original Article

Date Submitted by the

Complete List of Authors: Johansson, B.; University of Birmingham, School of Mathematics

Lesnic, D.; University of Leeds, Applied Mathematics

Heat conduction, Method of fundamental solutions, Regularization,

Research paper (2D).zip

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

International Journal of Computer Mathematics

We investigate an application of the method of fundamental solutions (MFS) to heat con-

In many engineering applications finding the solution to various heat conduction

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

1 of square integrable functions defined on the boundary of the solution domain. In

17 given by D̄T = D̄ × [0, T ] and Γ̄T = Γ × [0, T ], respectively.

F (x, t; y, τ ) = e 4(t−τ ) , (5)

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

1 where H is the Heaviside function which is introduced in order to emphasize that

34 We now construct the following infinite series

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

34 that the ratio

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

25 therefore, {vm (x, t)}∞ 2

j,m=1 , where vm (x, t) is given by

37 we give it here, for completeness, in higher dimensions.

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

1 inequality clearly holds

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

21 N × (M − 1) points will be located on D when t = 0. We consider M − 1 circles of

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

1 properties of circulant matrices to develop a matrix decomposition algorithm [18]

29 time is accurate. Below, we present numerical results for approximations in two-

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

1 Tikhonov regularization. As can be seen in Figure 4 the approximation obtained

44 Finally, we consider a three-dimensional plot of the exact solution u(x1 , x2 , 0.8)

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

1 (a) (b) (c)

uerror (x1 , x2 , 0.8)

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

uerror (x1 , x2 , 0.2)

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

1 Here, we have thermal diffusivity α = 1000/5.8, and the fundamental solution of

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

34 This is an inverse problem with missing boundary data at x1 = 1, which we wish

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

4 tions, obtained with h = 3 and λ = 10−8 .

∂x1 (1, x2 , 1.5)

uM,N (x1 , x2 , 0.5)

uerror (x1 , x2 , 2.5)

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

uerror (x1 , x2 , 0.5)

25 0 0.2 0.4 0.6 0.8 1 x2 0 0 x1

44 dimensional conducting bodies with Dirichlet boundary conditions. The solution

URL: http://mc.manuscriptcentral.com/gcom E-mail: ijcm@informa.com

eds., WIT Press, Southampton (1999), pp. 377–386.

Equations, Trans AMS 158 (1971), pp. 167–177.