INTRODUCTION

Stefan problems are a special class of moving boundary problems (also known as phase-change
problems). Some examples of these problems include freezing of food and water, solidification of
metals, etc. In such problems, the apriori marks the partition between solid and liquid, which is found
out from the solution. Some of these phase change phenomena are depicted in the figures below.
The direct Stefan problem requires the temperature and the moving boundary interface to be found,
provided the thermal properties of heat conducting object, along with the initial and boundary con-
ditions are given. In the Inverse Stefan Problem (ISP), as the name suggests, thermal properties and
initial or boundary conditions are determined using additional information like measurement of the
moving boundary or the temperature at selected points on domain. ISPs are one of the most important
classes of improperly posed problems.
Generally, it is difficult to determine the approximate solutions of both types of Stefan problems
using methods like finite element method, finite difference method, or boundary element method.
But the method of fundamental solutions (MFS) is the best way to solve these problems, due to the
advantages it holds over the above stated methods. In the governing equation, if the fundamental
solution of the operator is explicitly known, then the MFS can be applied.
Our aim is to reconstruct the boundary data in the one-phase ISP using MFS through MATLAB as
efficiently as possible.

(A) (B)

F IGURE 0.1. (A)Melting glacier (natural process),(B)Continuous casting process (in-

dustrial process)

1
PROBLEM FORMULATION
The first step is to familiarize ourselves with the governing equations of one-dimensional, one-
phase ISP. Assume x = s(t), where t belong to (0, T ] and s(t) > 0 in (0, T ], be the given smooth
moving boundary. Let ΩT = (0, s(t))x(0,t] represent the heat conduction domain where the closure
of the domain is ΩT = [0, s(t)]x(0, T ], where T represents the final timet.We have to find the solution
u(x,t) ∈ C2,1 (ΩT ) C(ΩT ), of the direct one-dimensional, one-phase Stefan problem, along with
T

the boundary which is x = s(t), which satisfies the one-dimensional heat equation,

∂u ∂ 2u
(0.1) ∂t (x,t) = ∂ x2 (x,t), (x,t)εΩT

with respect to the initial conditions

(0.2) s(0) ≥ 0

(0.3) u(x, 0) = uinit (x), xε[0, s(0)]

and for the moving boundary conditions at x = s(t), the Dirichlet and Neumann data are given by

(0.4) u(s(t),t) = h1 (t),tε(0, T ],

∂u
(0.5) ∂ x (s(t),t) = g1 (t),tε(0, T ],

where h1 (t) = 0, g1 (t) = −s’(t) satisfy : h1 (0) = u(s(0)) and g1 (0) = u0 (s(0)), which are essentially
compatibility conditions. If s(0) = 0, then the initial condition (0.3) is omitted. For the moving
boundary at x = 0, we have the following Dirichlet and Neumann data:

(0.6) u(0,t) = h0 (t),tε(0, T ],

(0.7) − ∂∂ ux (0,t) = g0 (t),tε(0, T ],

Note that the set of equations (0.1)–(0.6) with s(t) to be determined, or the set of equations (0.1)–
(0.5) and (0.7) with s(t) to be determined, are both direct one-phase Stefan problems, and that results
on their well-posedness can be found in [1, 2] and the references therein.
In this report, we will take the standard ISP into consideration. For this, we need to reconstruct the
Dirichlet and Neumann data at the fixed boundary (x = 0) and determine the solution u(x,t), satisfying
(0.1)–(0.5), with s(t) prescribed. Although this ISP generally has a unique solution [3, 4], it becomes
ill-posed if there are any perturbations in the input data. Hence, regularization is necessary in order
to maintain stability.

1
2

LITERATURE
Although, there are many articles available in the direction of numerical solution of direct Stefan
problems, the literature available in the direction of numerical solution of the inverse Stefan prob-
lem is very scarce. Below we present the literature, in particular for numerical solution of the one-
dimensional one-phase inverse Stefan problem and inverse Cauchy-Stefan problem using the MFS.

• Chantasiriwan et. al. [8] first used MFS to solve the direct inverse stephen problem. Addi-
tional information related to direct stephen problem can be found in [4].
• The boundary data for inverse stephen problem was reconstructed by Johansson et. al. [9]
using this method(MFS). By fixing location of source points from the boundary of domain and
also increasing source and collocation points there was no significant increase in accuracy.At a
later point of time they concluded that accurate and stable results could be actually obtained by
few collocation and source points and that this resulted in saving a lot of time in computation
and storage. Johansson et. al. [10] this was extended to 1-D inverse cauchy stephen problem
where initial data was missing.
• In [6] the authors considered the conclusion of [9] in a mathematical way as for a tolerance
that is given how the source points can be placed in an appropriate way so as to minimize the
effort in a computational sense and therefore this is dependent on number of source and collo-
cation points. In addition to this the error due to approximation doesn’t cross the tolerance.An
efficient method is obtained that reconstructs dirichlet boundary data and also the location of
source points is optimized. But this method doesn’t consider the accuracy of Neumann data.
• Later an adaptive MFS was given, Reddy et. al. [7] in which Neumann and Dirchlet were
reconstructed in an efficient way. Also the best suited regularization parameter is also obtained
for a given value of measurement error.
• Later motivated by [10] this was extended to one dimensional one phase inverse cauchy
stephen problem.

Till date there is no application for an MFS that could reconstruct the data at boundary for the two--
phase inverse Stefan problems which is our aim to study in this project.
 3

MFS
The key advantages and few disadvantages of the MFS are listed as such:

• It is relatively easy to code in MATLAB and cost effective.

• No complicated meshes need to be created as MFS is a collocation method.
• Stable and accurate results,in combination with appropriate regularization methods, were pro-
duced, to solve various problems like elliptic, time-dependent parabolic, free boundary, etc.

• The fundamental solution of the governing linear PDE has to be known in order to apply MFS.
• Accuracy can be greatly affected by the position and number of source and collocation points.

We will focus ontesting and applying the MFS for one-phase heat conduction problems.

In the MFS, an approximationt of the form

2 M+1
(0.8) uMFS (x,t) = ∑ ∑ cij F(x,t; yi(τ j ), τ j ), (x,t)ε Ω¯T ,
i=1 j=1

where F is the fundamental solution of equation (1) and is given by

H(t − τ) (x − y)2
(0.9) F(x,t; y, τ) = p exp(− )
4π(t − τ) 4(t − τ)

where H denotes the Heaviside function and cij are real coefficients that have to be determined.
Moreover, {yi }2i=1 are space singularities placed outside [0, s(t)],tε(0, T ] and {τi }M
j=1 are time singu-
larities that are located in the interval (−T, T ). Justification for the MFS approximation (0.8) comes
via denseness results, and these can be found in [6].
To compute theMFS matrix, we must first distribute the source and collocation points. For each of
the three examples that we shall consider, we place these in the same way as was done in [10].

We place the source points at (−h, τ) for τε(−T, T ), (s(τ) + h, τ) for τε(0, T ),and (s(0) + h, τ) for
τε(−T, 0), i.e. at y1 (τ j ) = −h and y2 (τ j ) = h + s(τ j H(τ j )),j =1, . . . ,M + 1, where the time points
{τ 
j } j=1,.....,M+1 ε(−T, T ) are given by
τ M
+1−m = − T − 2T m, m = 0, ..., M
2M M+1 2
2
3T 2T
τ M = 2M + M m, m = 0, ..., M2
−1
2 +2−m
On the other hand, for the collocation points, we let
ti = Mi T, x1i = s(ti ), i = 0, . . . ,M
4

F IGURE 0.1. The above√ figure demonstrates the domains of two-phase ISP, with
boundary s(t) = 3 − 3 − 2t. Red and yellow lines indicate source points. Black,
blue and magenta lines indicate collocation points. Green and blue lines indicate ini-
tial condition along base.

System of equations. The coefficients cij are calculated by imposing equations (0.4) and (0.5) at the
collocation points. This gives the following system of equations:

(0.10) uM (x1i ,ti ) = 0, ∂∂Mx (x1i ,ti ) = −s0 (ti ), i = 0, ..., M

The system of equations (0.10) contains the same number of equations as unknowns, i.e. 2(M + 1).
For the computation of the solution by MFS, the collocated data from the boundary is required.
Then, we add random noise to equation (5) in the form

(0.11) uδx (s(t),t) = ux (s(t,t)) + N(0, (σ δ )2 ),

where

σ δ = δ xmax(s(t),tε(0,T ] |ux (s(t),t)|,

and the random noise is denoted by N(0, (σ δ )2 , i.e. the normal distribution with zero mean and
standard deviation σv, where δ% represents the percentage of relative random noise. Although it is
also possible to add noise to some other quantity that is related to s(t), as suggested in [4], such cases
are not pursued here.
Finally, the system of equations can be written as

(0.12) Ac = f ,
 5

where A is the matrix denoting the value of the MFS solution at the corresponding collocation and
source points, c is the vector of unknown constants cij and f is the vector representing the boundary
and initial values at the collocation points. The matrix A can be expected to have a high condition
number [5, 6], and thus regularization is normally required. For this purpose, Tikhonov regularization
is used, leading to the modified system of equations

(0.13) (AT A + λ I)c = AT f ,

instead of the system of equations (0.12); in (0.13), AT denotes the transpose of the matrix A, I is
the identity matrix and λ ≥ 0 is the regularization parameter. We solve this well-conditioned, linear
system of equations for the coefficients of c using Gaussian elimination.
 CONCLUSION
Using MATLAB, the domain of two phase ISP was plotted. The effectiveness and adaptiveness of
the algorithm in [7] was proved by testing in MATLAB by help of various test problems. In the next
part of this course we aim to extend this to two-phase two inverse stephen problem. 1

1
 REFERENCES
(1) Cannon, J.R., Hill, C.D.: Existence, uniqueness, stability, and monotone dependence in a
Stefan problem for the heat equation. J. Math. Mech. 17(17), 1–19 (1967) ]
(2) Cannon, J.R., Primicerio, M.: Remarks on the one-phase Stefan problem for the heat equa-
tion with the flux prescribed on the fixed boundary. Journal of Mathematical Analysis and
Applications 35(2), 361–373 (1971)
(3) Cannon, J.R., Douglas Jr., J.: The Cauchy problem for the heat equation. SIAM J. Numer.
Anal. 4, 317–336 (1967)
(4) Ang, D.D., Dinh, A.P.N., Thanh, D.N.: An inverse Stefan problem: identification of boundary
value. Journal of Computational and Applied Mathematics 66(1), 75 – 84 (1996)
(5) Chen, C.S., Cho, H.A., Golberg, M.A.: Some comments on the ill-conditioning of the method
of fundamental solutions. Engineering Analysis with Boundary Elements 30(5), 405 – 410
(2006)
(6) Ramachandran, P.A.:Method of fundamental solutions: singular value decomposition analy-
sis. Communications in Numerical Methods in Engineering 18(11), 789–801 (2002)
(7) Thomas Henry Reeve :The Method of Fundamental Solutions for some direct and inverse
problems , University of Birmingham Research Archive(2013).
(8) S. Chantasiriwan, B. T. Johansson, and D. Lesnic, “The method of fundamental solutions for
free surface Stefan problems,” Eng. Anal. Bound. Elem., vol. 33, no. 4, pp. 529–538, 2009.
(9) B. T. Johansson, D. Lesnic, and T. Reeve, “A method of fundamental solutions for the onedi-
mensional inverse Stefan problem,” App. Math. Modelling, vol. 35, pp. 4367–4378, 2011.
(10) B. T. Johansson, D. Lesnic, and T. Reeve, “Numerical approximation of the one-dimensional
inverse Cauchy-Stefan problem using a method of fundamental solutions,” Inverse Probl. Sci.
Eng., vol. 19, no. 5, pp. 659–677, 2011.
(11) G. M. M. Reddy, M. Vynnycky, and J. A. Cuminato, “On efficient reconstruction of boundary
data with optimal placement of the source points in the MFS: application to inverse Stefan
problems,” Inverse Problems in Science and Engineering, vol. 26, no. 9, pp. 1249–1279,
2018.
(12) G. Reddy, M. Vynnycky, and J. Cuminato, “An efficient adaptive boundary algorithm to re-
construct Neumann boundary data in the MFS for the inverse Stefan problem,” Journal of
Computational and Applied Mathematics, vol. 349, pp. 21 – 40, 2019.