A FUNDAMENTAL SOLUTION METHOD FOR INVERSE HEAT CONDUCTION PROBLEM

Y. C. HON
1

& T. WEI

Abstract. In this paper, we develop a new meshless and integration-free numerical scheme for solving an inverse heat conduction problem. The numerical scheme is developed based on the use of the fundamental solution as a radial basis function. To regularize the resultant ill-conditioned linear system of equations, we apply successfully both the Tikhonov regularization technique and the L-curve method to obtain a stable numerical approximation to the solution. The approach is readily extendable to solve high-dimensional problems under irregular domain.

1. Introduction Inverse heat conduction problem(IHCP) arises in the modeling and control of processes with heat propagation in thermophysics and mechanics of continuous media. Some kinds of non-destructive evaluation techniques to determine the interior heat ows in an inaccessible domain from exterior scattered data are indispensable for industrial and engineering applications. Studies of the complicated phenomena are particularly challenging for quantitative analysis and modeling. Inverse problems are in nature unstable because the unknown solutions and parameters have to be determined from indirect observable data which contain measurement error. The major diculty in establishing any numerical algorithm for approximating the solution is due to the severe ill-posedness of the problem and the ill-conditioning of the resultant discretized matrix. The studies of inverse heat conduction problems are even more dicult. Although heat conduction process is very smooth, the process is irreducible. This means that the characteristic of the solution (for instance, the shape of the interior heat ow) may not be aected by the observed data. On the other hand, the heat conduction process has no nite propagation speed and thus an ecient nondestructive testing technique can be achieved at a comparably much lower cost. The lack of mathematical analysis and ecient algorithm, however, hinders the development of such low cost and ecient non-destructive testing technique. Inverse heat conduction problem is highly ill-posed in the sense that any small change on the input data can result in a dramatic change to the solution. During the last decades, some numerical techniques have been proposed to solve a
Department of Mathematics, City University of Hong Kong, Hong Kong. Department of Mathematics, City University of Hong Kong, Hong Kong & Department of Mathematics, Lanzhou University, China. The work described in this paper was partially supported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project No. CityU 1178/02P), a China NSF grant (No. 10271050) and a NSF grant of the Gansu Province, China (ZS021-A25001-Z).
2 1

Y. C. HON

& T. WEI

one-dimensional IHCP [13, 16, 18, 19]. Among these methods the nite dierence method (FDM) and the nite element method (FEM) are so far the principal numerical tool of choice for the modeling and simulation of the IHCP. A major disadvantage of these methods, however, is their mesh-dependent characteristics which normally requires enormous computational eort and induces numerical instability when large number of grids or elements are required. The boundary element method (BEM) does not require domain discretization and hence the temperature data measurement can be chosen comparatively in an arbitrary way [17]. Frankel et al. [3] recently gave an unied space-and-time treatment of the inverse heat conduction problem by using the Chebyshev polynomials as basis functions. In this paper a meshless approach based on the use of fundamental solution as radial basis functions for solving the IHCP is devised . Since the basis functions generated from fundamental solution are the general solution of considered heat equation, the approximation to the temperature distribution only needs to satisfy the boundary condition and the given measurement data. Comparing with FEM and FDM, our proposed method does not require any domain discretization. Unlike BEM which requires taking functions interpolation on the boundary and integrations on the boundary element, our proposed method givens an approximated solution on the whole temporal and spatial domain without the need for extra quadratures. Furthermore, the approximations of the surface temperature and heat ux can be deduced by simple and direct function evaluations. In the last decade, the radial basis functions(RBF) have been under intensive research in the areas of multivariate function interpolation and partial dierential equations solver. Most of existing works are devoted for solving direct problem [9, 10, 11, 14, 21]. Hon and Wu [12] gave a rst approach in applying the RBF to solve the Cauchy problem for Laplace equation, which is well known inverse and ill-posed problem. Based on the combination of the RBFs and the method of fundamental solution (MFS) [1, 2], we propose in this paper an alternative method for solving some inverse heat conduction problems. The proposed method in this paper is to apply the fundamental solution of heat equation as basis functions. This approach is dierent from most existing numerical algorithms in solving dynamical problems where the nite dierence quotient will be used to discretize the time variable. Our method is also simple and feasible to handle various informal boundary conditions. 2. Formulation of the problem Consider a linear inverse heat conduction problem where the boundary heat ux has to be determined from the temperature measurements in the interior or on part of the boundary. Consider the following initial-boundary value problem: (1) (2) (3) (4) ut a2 uxx = 0, 0 < x < 1, 0 < t < tmax , u(x, 0) = (x), 0 x 1, u x (1, t) + u(1, t) = g (t),

u(x , t) = h(t),

where u(x, t) is the temperature distribution and tmax represents the maximum time of interest for the time evolution of the problem. The constants and are positive. A general IHCP is then to determine the temperature and heat ux on the surface x = 0. This problem is well known for being highly ill-posed.

x [0, 1],

0 t tmax ,

0 t tmax ,

INVERSE HEAT CONDUCTION PROBLEM

In practical application, if the input function g (t) are given only at discrete (1) values of times ti , i = 1, 2, , m and the internal measurement temperature h(t) are located at times t i , i = 1, 2, , l , then the boundary condition (3) and internal measurement condition (4) will be replaced as follows: (5) and (6)
(1) l {ti }m i=1 , {ti }i=1

u (1) (1) (1, ti ) + u(1, ti ) = gi , x u(x , t i ) = hi ,

i = 1, 2, , m,

i = 1, 2, , l,

where denote the discrete values of times in the internal (0, tmax ]. In this paper, we only consider the discrete situation. In fact, the proposed approach will illustrate that the standard pre-smoothing treatment of the initial data is unnecessary. In the following, we present a new approach to solve this inverse heat conduction problem. Notice that if the internal point x = 0, then the problem (1)(4) is a well-posed forward heat conduction problem. Furthermore, if x = 1 and = 0, the problem is called a Cauchy problem for heat equation. The following section will rst explain the basic idea how to apply the RBF-MFS to solve this kind of highly ill-posed inverse heat conduction problem. For stable computation, the standard Tikhonov regularization technique and L-curve method [4, 5, 6, 8] are adopted to solve the resultant ill-conditioned linear systems. 3. Fundamental solution method The fundamental solution of equation (1) is given as (7) F (x, t) =
x2 1 e 4a2 t H (t), 2a t

where H (t) is the Heaviside function. Assume that T > tmax is a constant. The time shift function (8) (x, t) = F (x, t + T ) is also a general solution of equation (1) on the domain [0, 1] [0, tmax ]. j 1 Now, assume that the known data are given at the points (xj , tj ) = ( n 1 , 0), j = 1, 2, , n satisfying the initial condition (2) on the line t = 0 and the points (1) (xj , tj ) = (1, tj n ), j = n + 1, n + 2, , n + m on the boundary x = 1 satisfy(1)

ing the boundary condition (5), where {ti }m i=1 are measurement times given in (5). We choose the scattered points on x = x at the discrete times (xj , tj ) = l (x , t j = n + m + 1, n + m + 2, , n + m + l, where x and {t j nm ), i }i=1 are n+m+l are pairwisely distinct in specied in equation (6). All the points {(xj , tj )}j =1 the two-dimensional space [0, 1] [0, tmax ]. Following the idea of the method of fundamental solution and radial basis functions for solving elliptic equation in [1, 2], we assume that an approximation to the solution of the IHCP problem (1),(2),(5),(6) can be expressed by the following basis functions:
n+m+l

(9)

u (x, t) =
j =1

j (x xj , t tj ),

Y. C. HON

& T. WEI

where the basis function (x, t) is given by equation (8) and j are unknown coefcients to be determined. For this choice of basis functions , the approximated solution u already satises the heat equation (1). The resultant linear system of equations for the unknown coecients j can then be obtained from the following simple collocation. Collocating (9) into (2),(6) and (5), we obtain
n+m+l

u (xi , ti ) =
j =1

j (xi xj , ti tj ), i = 1, 2, , n, i = n + m + 1, n + m + 2, , n + m + l,
n+m+l

= and

(xi ), hinm ,

u (xi , ti ) + u (xi , ti ) = x

j [
j =1

(xi xj , ti tj ) + (xi xj , ti tj )], x

= gin ,

i = n + 1, n + 2, , n + m.

In matrix form, the values of the unknown coecients j can be obtained from solving the following matrix equation: (10) A = b,

where A is a (n + m + l) (n + m + l) square matrix: (xi xj , ti tj ) , (11) A = x (xk xj , tk tj ) + (xk xj , tk tj ) (xs xj , ts tj ) and b is a (n + m + l) vector: (12) (xi ) , b = gkn hsnm

where i = 1, , n, k = n + 1, , n + m, s = n + m + 1, , n + m + l and j = 1, , n + m + l. In the next section, we successfully apply both the Tikhonov regularization technique and the L-curve method to obtain a stable numerical solution to the linear system (10). The solution to the IHCP problem (1),(2),(5) and (6) can then be obtained by substituting the coecient vector into equation (9). 4. Regularization methods Since the original IHCP is ill-posed, the ill-conditioning of the matrix A in equation (10) still persists. In other words, most standard numerical methods cannot achieve good accuracy in solving the matrix equation (10) due to the large condition number of the matrix A. In fact, the condition number of matrix A increases dramatically with respect to the total number of collocation points. Several regularization methods have been developed for solving these kinds of ill-conditional problems [4, 5, 6, 8]. In our computation we adapt the Tikhonov regularization

INVERSE HEAT CONDUCTION PROBLEM

technique [20] to solve the matrix equation (10). The Tikhonov regularized solution for equation (10) is dened to be the solution to the following least square problem: (13) min { A b
2

+ 2 2 },

where denotes the usual Euclidean norm and is called the regularization parameter. The determination of a suitable value for the regularization parameter is crucial and is still under intensive research [20]. In our computation we use the L-curve method to determine a suitable value of . The L-curve method was rstly applied by Lawson and Hanson [15] and more recently by Hansen and OLeary [8] to investigate the properties of regularized systems under dierent values of the regularization parameter . The L-curve method is sketched in the following: Dene the following curve (14) L = { (log (
2

),

log ( A b 2 )),

> 0}.

The curve is known as L-curve and a suitable regularization parameter corresponds to a regularized solution near the corner of the L-curve [4, 5, 6]. In our computation, we used the Matlab code developed by Hansen [7] for solving the discrete ill-conditioned system (10). Denote the regularized solution of (10) by . The approximated solution u for problem (1),(2),(5) and (6) is then given as
n+m+l

(15)

u (x, t) =
j =1

j (x xj , t tj ).

The temperature and heat ux at surface x = 0 are then given respectively by

n+m+l

(16) and

f (t) =
j =1

j (0 xj , t tj ),

n+m+l

(17)

q (t) =

j
j =1

(0 xj , t tj ). x

The numerical results in the following section indicate that the proposed scheme is stable, feasible, and ecient. 5. Numerical results For simplicity, we assume that the heat conduction coecient a = 1 and tmax = 1 or 2. For numerical error estimation , we dene the root mean square error by the following formula (18) RM S (u) = 1 N
N 2 (ui u i) ,

i=1

where N is the total number of uniformly distributed points on the domain [0, 1] [0, tmax ], ui and u i are respectively the exact and approximated temperature value at these points. The value of RM S ( u x ) is the RMS error for the heat ux on the

Y. C. HON

& T. WEI

whole domain. Similarly, we denote RM S (u(0, t)) and RM S ( u x (0, t)) as the RMS errors for the surface temperature and heat ux on x = 0 respectively. In the cases when the measurement data include some random noises, we use noisy data gi = gi + rand(i), where gi is the exact data and rand(i) is a random number between [1, 1] and the magnitude indicates the error level. Example 1: Let the exact solution for the problem (1),(2),(5) and (6) be 1 4 x ). 12 We consider two cases under dierent boundary conditions and measurement locations. In Case 1, we take tmax = 1, x = 1 and = 1, = 0. The number of collocation points is n = m = l = 11. All the given boundary data gi , hi and initial condition (x) can be computed from the exact solution (19). The numerical results obtained by our method are presented in Figures 1, 2, and 3. For Case 2, we take tmax = 1, x = 0.9 and = 1, = 1. The number of collocation point is the same. The comparisons between the exact solution and the approximation are given in Figure 4 and Figure 5 on their values of temperature and heat ux on the boundary x = 0. The numerical results show that the proposed scheme is accurate and eective method in comparing with most of the existing numerical methods. (19) u(x, t) = 2e4t (sin2x + cos2x) + 3(t2 + tx2 +

parameter T=2.9 x 10 2 error of u(x,t) 0 2 4 6 1 RMS(u)= 3.3044e4 1 0.5 t 0 0 0.5 x 0.5 t error of dudx(x,t)
3

parameter T=2.9

0.02 RMS(dudx)= 0.0015 0.01 0 0.01 1 1 0.5 0 0

u x (x, t)

x
in Case

Figure 1. The error distribution for u(x, t) and 1 with noiseless data.

Example 2: To further explore the applicability of the proposed method for solving IHCP, we consider the classical Becks problem [3] in which a triangular surface heat ux is imposed at x = 0. Assume tmax = 1, x = 1 and = 1, = 0 and an exact heat ux on x = 0 is given as t [0, 0.25] [0.75, 1], 0, u t0.25 , t [0.25, 0.5], (20) q (t) = (0, t) = 0.25 t x 0.75 , t [0.5, 0.75]. 0.25

INVERSE HEAT CONDUCTION PROBLEM

parameter T=2.9 du*/dx(0,t) and du/dx(0,t) 4 u*(0,t) and u(0,t) 3 2 1 0 RMS(u(0,t)) =9.1538e4 5 4 3 2 1

parameter T=2.9 RMS(dudx(0,t)) =0.0039

0 0.5 1 0 0.5 1 t t Figure 2. The boundary temperature u(0, t) and heat ux u x (0, t) for Case 1 with noiseless data. The solid line represents the exact solution and the dotted line represents the approximation. 0 parameter T=2.9 du*/dx(0,t) and du/dx(0,t) 4 u*(0,t) and u(0,t) 3 2 1 0 RMS(u(0,t)) =0.0047 5 4 3 2 1 parameter T=2.9 RMS(dudx(0,t)) =0.0119

0 0.5 1 0 0.5 1 t t Figure 3. The boundary temperature u(0, t) and heat ux u x (0, t) for Case 1 with discrete noisy data gi , hi , gi = gi + rand(i), hi = hi + rand(i), = 0.01. The solid line represents the exact solution and the dotted line represents the approximation. 0
The given data g (t) in (3) and the initial condition in (2) are assumed to be zeroes. By using the standard technique in the theory of partial dierential equations, the measurement data h(t) in (4) of problem is given as
t t

(21)

h(t) = u(1, t) =
0

q ( )d +
n=1

2(1)n

q ( )e(n)
0

(t )

d .

From the data h(t), we can then reconstruct the triangular surface heat ux q (t). The predicted surface heat ux q (t) and the exact q (t) are displayed in Figures 6 and 7 in which the total number of points is taken as n = 11, m = 20, l = 20. From the numerical results demonstrated in Figures 6 and 7, we can observe that

Y. C. HON

& T. WEI

parameter T=1.2 du*/dx(0,t) and du/dx(0,t) 4 u*(0,t) and u(0,t) 3 2 1 0 RMS(u(0,t)) =0.0016 5 4 3 2 1

parameter T=1.2 RMS(dudx(0,t)) =0.0072

0 0.5 1 0 0.5 1 t t Figure 4. The boundary temperature u(0, t) and heat ux u x (0, t) for zero noise data in Case 2. The solid line represents the exact solution and the dotted line represents the approximation. 0 parameter T=1.2 du*/dx(0,t) and du/dx(0,t) 4 u*(0,t) and u(0,t) 3 2 1 0 RMS(u(0,t)) =0.0412 5 4 3 2 1 parameter T=1.2 RMS(dudx(0,t)) =0.1361

0 0.5 1 0 0.5 1 t t Figure 5. The boundary temperature u(0, t) and heat ux u x (0, t) for noisy data gi , hi in Case 2, gi = gi + rand(i), hi = hi + rand(i), = 0.01. The solid line represents the exact solution and the dotted line represents the approximation. 0
the proposed scheme still gives a reasonable approximation to the exact solution q (t) which possesses a sharp triangular point Example 3: Consider an even more severe test function [17], namely, (22) v (x, t), t [0, 0.5) v (x, t) 2v (x, t 0.5), t [0.5, 1) u(x, t) = v ( x, t ) 2 v ( x, t 0 . 5) + 2 v ( x, t 1), t [1, 1.5) v (x, t) 2v (x, t 0.5) + 2v (x, t 1) 2v (x, t 1.5), t [1.5, 2], where (23) v (x, t) =
2 2 3(1 x)2 1 (1)n +t2 cos(n (1 x))en t 2 2 6 n n=1

INVERSE HEAT CONDUCTION PROBLEM

parameter T=1.1 0.6 RMS(u(0,t))=0.026 du*/dx(0,t) and du/dx(0,t) 0.5 u*(0,t) and u(0,t) 0.4 0.3 0.2 0.1 0 0.1 0 0.5 t 1 0.8 0.6 0.4 0.2 0 0.2 0 1.2

parameter T=1.1 RMS(dudx(0,t))=0.1249

0.5 t

Figure 6. The boundary temperature u(0, t) and heat ux u x (0, t) in the case of noiseless data h(t) for Example 2. The solid line represents the exact solution and the dotted line represents the approximation. which includes a discontinuous surface heat ux on x = 0. Figure 8 illustrates the numerical results for the temperature and heat ux on x = 0 when tmax = 2, x = 0.25 and = 1, = 0. All the measurement data gi , hi are computed by the exact solution (22). Here, we add a noise to the data hi = hi + rand(i), = 0.1. The numerical results, where the total number of collocation points is n = 31, m = l = 65, are displayed in Figure 9. Despite the discontinuities of the temperature in form of sharp corner and the heat ux in the form of square-wave step functions as shown in Figure 8 and Figure 9, the numerical approximations are in reasonable agreement with the exact solutions of the temperature and the heat ux. 6. Conclusion The universal approach for solving transient phenomena, directly or inversely, involves a time-marching procedure, i.e. taking one step in time and solving the problem in the spatial domain. In this paper, by inverting the resultant matrix, an approximated solution on the entire domain is obtained. This ensures that the time consuming iterative solving process is eectively formulated. We believe that the proposed RBF-MFS approach provides an ecient global time-space scheme for solving these kinds of inverse heat conduction problems. References
[1] Alves, C. J. S., Chen, C. S & S aler, B., The method of fundamental solutions for solving Poisson problems, International Series on Advances in Boundary Elements, vol. 13, Bounday Elements XXIV, pp. 6776, 2002.

10

Y. C. HON

& T. WEI

parameter T=1.4 0.6 RMS(u(0,t))=0.0778 du*/dx(0,t) and du/dx(0,t) 0.5 u*(0,t) and u(0,t) 0.4 0.3 0.2 0.1 0 0.1 0 0.5 t 0.5 1

parameter T=1.4

0.5 RMS(dudx(0,t))=0.2854

0.5 t

Figure 7. The boundary temperature u(0, t) and heat ux u x (0, t) with noisy data gi , hi for Example 2, gi = gi + rand(i), hi = hi + rand(i), = 0.001 The solid line represents the exact solution and the dotted line represents the approximation.

[2] Chen, C. S., Golberg, M. A. & Hon, Y. C., The method of fundamental Solutions and QuasiMonte-Carlo Method for Diusion Equations, International Journal for Numerical Methods in Engineering, 43, no. 8, pp. 14211435, 1998. [3] Frankel, J.I. & Keyhani. M, A Global Time Treatment for Inverse Heat Conduction Problems, International Journal of Heat Transfer, 119, pp. 673683, 1997. [4] Hansen, P.C., Analysis of discrete ill-posed problems by means of the L-curve, SIAM Review, 34 pp. 561580, 1992. [5] Hansen, P.C., Numerical tools for analysis and solution of Fredholm integral equations of the rst kind, Inverse Problems, 8, pp. 849872, 1992. [6] Hansen, P.C., The L-curve and its use in the numerical treatment of inverse problems, Computational Inverse Problems in Electrocardiology, 4, ed. Johnston, P., Advances in Computational Bioengineering Series, WIT Press, Southampton, 2000. [7] Hansen, P.C., Regularization tools: a Matlab package for analysis and solution of discrete ill-posed problems, Numerical Algorithms, 6, pp. 135, 1994. [8] Hansen, P.C. & OLeary Dianne Prost, The use of the L-curve in the regularization of discrete ill-posed problems. SIAM Journal of Scientic Computing, 14 , pp. 14871503, 1993. [9] Hon, Y.C., Cheung, K.F., Mao, X.Z. & Kansa, E.D., Multiquadric solution for shallow water equations, ASCE Journal of Hydraulic Engineering, 125 ,pp. 524533, 1999. [10] Hon, Y.C., Lu, M.W., Xue, W.M. & Zhou, X , A new formulation and computation of the triphasic model for mechano-electrochemical mixtures, Computational Mechanics, 24, pp. 155 165, 1999. [11] Hon, Y.C. & Mao, X.Z., A radial basis function method for solving options pricing models, The Journal of Financial Engineering , 8, pp. 3149, 1999. [12] Hon, Y.C. & Wu, Z.M., A numerical computation for inverse boundary determination problem, Engineering Analysis with Boundary Elements , 24, pp. 599606, 2000 . [13] Jonas, P & Louis, A.k., Approximate inverse for a one-dimensional inverse heat conduction problem, Inverse Problems, 16 , pp. 175185, 2000.

INVERSE HEAT CONDUCTION PROBLEM

11

parameter T=2.7 1 0.8 u*(0,t) and u(0,t) 0.6 0.4 0.2 0 0.2 0.4 0 1 t 2 du*/dx(0,t) and du/dx(0,t) RMS(u(0,t))=0.0886 1.5 1 0.5 0 0.5 1 1.5

parameter T=2.7

RMS(dudx(0,t))=0.4425 0 1 2 t

Figure 8. The boundary temperature u(0, t) and heat ux u x (0, t) with noiseless data gi , hi for Example 3, The solid line represents the exact solution and the dotted line represents the approximation.
[14] Kansa, E.J. & Hon, Y.C., Circumventing the ill-conditioning problem with multiquadric radial basis functions: applications to elliptic partial dierential equations, Computers and Mathematics with Applications, 39 ,pp. 123137, 2000. [15] Lawson, C.L. & Hanson, R.J., Solving Least Squares Problems, SIAM, Philadelphia, PA, 1995. First published by Prentice-Hall, 1974. [16] Lesnic, D. & Elliott. L, The decomposition approach to inverse heat conduction, Journal of Mathematical Analysis and Application, 232, pp. 8298, 1999. [17] Lesnic, D., Elliott, L. & Ingham, D. B., Application of the boundary element method to inverse heat conduction problems, International Communications in Heat and Mass Transfer 39 , no.7, pp. 1503-1517, 1996. [18] Liu, J., A stability analysis on Becks procedure for inverse heat conduction problem, Journal of Computational Physics, 123, pp. 6573, 1996. [19] Shen, S.Y., A numerical study of inverse heat conduction problems, Computers and Mathematics with Applications, 38, pp. 173188, 1999. [20] Tikhonov, A. N. & Arsenin, V.Y., On the solution of ill-posed problems , John Wiley and Sons, New York, 1977. [21] Zerroukat, M, Power, H. & Chen, C.S., A numerical method for heat transfer problems using collocation and radial basis functions, International Journal for numerical methods in Engineering, 42, pp. 12631278, 1998.

12

Y. C. HON

& T. WEI

parameter T=2.7 1 0.8 0.6 u*(0,t) and u(0,t) 0.4 0.2 0 0.2 0.4 0.6 0 1 t 2 du*/dx(0,t) and du/dx(0,t) RMS(u(0,t))=0.2046 3

parameter T=2.7 RMS(dudx(0,t))=0.7092 2 1 0 1 2 3

1 t

Figure 9. The boundary temperature u(0, t) and heat ux u x (0, t) with noise data hi for Example 3, hi = hi + rand(i), = 0.1. The solid line represents the exact solution and the dotted line represents the approximation.