1. Introduction
For a fixed solution of a partial differential equation, it is usually restricted to a certain region. If this region varies with time, we call it the moving boundary problem. If part of the boundary of the fixed region is to be determined simultaneously with the solution of the fixed problem, we call it the free boundary problem, and the unknown boundary is called the free boundary. For free boundary problems, in addition to the usual fixed solution conditions, boundary conditions (Stefan conditions) must be added to the free boundary.
The Stefan problem is a class of heat conduction problems. It was first formulated by the Austrian physicist Joseph Stefan in the late 19th Century. The background of this problem is closely related to the industrial production during the industrial revolution of that time. In industrial production, many substances needed to be heat-treated, so the laws and methods of heat conduction needed to be studied to better control and utilize heat energy.
The difficulty of the Stefan problem is tracing the location of the interface, for example to model the process of change at the intersection of ice and water. Since the interface changes with time, these problems are also called free boundary problems. The study of free boundary problems has a wide practical background, such as plasma physics, percolation mechanics, and plasticity mechanics [
1,
2,
3,
4,
5,
6], which have presented various forms of constant and indeterminate free boundary problems. Furthermore, chemical vapor deposition [
7], the vapor permeation of thermally cracked carbon in chemistry [
8], tumor growth in medicine [
9,
10,
11,
12,
13], the expansion propagation problem of biological populations [
14,
15,
16,
17], and the U.S. option pricing problem [
18,
19] also have free boundary problems. In fact, all free boundary problems are nonlinear problems, and it is important to solve them with the solution of the free boundary, which will be determined together with the solution of the fixed problem. Since these problems are closely related to practical applications, the efficient algorithmic implementation of free boundary problems is of great importance for scientific research and production practice.
Currently, a variety of numerical methods to the Stefan problems have emerged. The boundary integral method numerically solves the integral equation at a moving interface [
20]. The interface tracing method explicitly represents the free boundary by using a fixed grid of points [
1]. The immersion interface method [
21] uses a fixed spatial grid for some physical quantities and a moving grid for the free boundary, and the information between the free boundary and the fixed grid is obtained by the immersion interface method. Segal [
22] proposed an adaptive grid method for solving the free boundary problem, where the movement of the grid is determined by the control equations. Murray and Landis [
23] compared the fixed-grid method with the adaptive-grid method. The adaptive-grid method can obtain the free boundary location more accurately, while the fixed-grid method can obtain certain physical quantities (temperature distribution over the whole solution area). More accurately, the enthalpy difference method [
24] is an implicit method that represents the heat of the whole system by introducing an enthalpy function with a jump discontinuity at the free boundary, and this discontinuity helps to determine the location of the interface. The moving grid method [
25] is a common method for solving free boundary problems, where the free boundary position is always fixed at the nth grid point, and the grid needs to be updated at each time step due to the movement of the free boundary. In recent years, the phase field method [
26] has became a popular method based on the phase field function, which corresponds to a fixed constant in each phase and the interface region between the two values. This method considers a fuzzy boundary between the two phases, which is different from the classical Stefan problem, where the phase transition occurs in this interface region, where the thickness of the region is an artificially given parameter. The level-set approach [
27] has also received increasing attention in recent years, by introducing a level-set function, which defines the interface position as a zero level set, obtained from the advection equation related to the velocity field, which varies considerably in different applications of the level-set approach. Sussman used the velocity of the fluid to model compressible two-phase flow [
28], and Chen extended the interface moving velocity to the whole region by the advection equation in the solidification problem [
27]. In contrast to the moving grid method, the level set method uses a fixed grid and avoids updating the grid at each time step. In contrast to the fixed-gridmethod, which finds the interface position at a fixed grid point, the phase-field method does not track the interface position precisely, and therefore, the discretization at the interface position is not as accurate as the fixed-grid method.
All the above methods have proven their high accuracy for some specific Stefan problems, and each has its own advantages and disadvantages. However, a general framework for solving the Stefan forward and inverse problems is still missing in all the methods at this stage.
In recent years, with the continuous development of neural networks, the use of neural networks to solve partial differential equations has gradually become popular [
29,
30,
31,
32]. Raissi et al. [
33] proposed physics-informed neural networks for solving forward and inverse problems of partial differential equations, and deep-learning-based physics-informed neural networks have also been proposed for solving free boundary problems [
34], as well as neural networks to solve the Stefan problems using a lattice-free grid-free automatic differentiation technique, which breaks the limitations of the above methods.
Deep learning models have achieved advanced results in solving various types of partial differential equations. However, the success of deep learning models relies heavily on a large amount of training data. In some specialized fields, the cost of data acquisition is very large. In addition, labeling samples requires much effort. Therefore, in recent years, a new learning approach, small sample learning, has gained popularity [
35]. Small sample learning has been successfully applied to many new fields, such as: neural networks translation, target detection, etc. By using small sample learning, the accuracy of the model can be improved with few labeled data.
In this paper, we extended the recently emerging physics-informed neural networks framework [
33] to solve the general Stefan problems. As we know, the original framework of physics-informed neural networks does not deal well with free boundary problems with time variation. To achieve this goal, we propose and modified the neural network framework proposed by Wang [
34]. The specific contributions of this paper can be summarized as follows: Firstly, we incorporated the idea of small sample learning into neural network training, i.e., a small sample loss is added to the loss function optimization in order to improve the training accuracy and correct the model. Secondly, we changed the loss function to cope with the outliers that may arise from free boundary shifts. Finally, we applied the proposed framework to irregular regions and irregular free boundaries to test its performance. In summary, the proposed method provides a new general framework for solving the Stefan problems.
This paper is structured as follows:
Section 2 introduces the Stefan problem and its mathematical model.
Section 3 introduces the knowledge of neural networks and the PINNs’ improvement strategy.
Section 4 verifies the accuracy and applicability of the proposed framework with numerical arithmetic examples. The conclusions and outlook are given in
Section 5.
2. Model Issues
In this section, we introduce the mathematical model of the Stefan problem and the corresponding boundary conditions using a one-dimensional single-phase Stefan problem in the solidification or melting process as an example.
As shown in
Figure 1, assume that a semi-infinite solid occupying
is in the process of solidification or melting. For any moment
, the
region consists of a solid and a liquid. The liquid is located in the
region, and the solid is located in the
region.
If the volume change due to solidification or melting is not considered and the region
is considered, the temperature
satisfies the classical diffusion equation over the region:
and for the initial and boundary conditions:
at the interface
, the following Stefan conditions need to be met:
where
denotes the free boundary, (4) denotes the initial position of the free boundary, and (5) denotes the temperature at the time of freezing. For the forward problem, in thermal physics, it is the simultaneous solution of the temperature distribution and the free boundary for which various parameters are known. Each Stefan problem corresponds to a class of inverse problems, and similar to other mathematical physics inverse problems, the inverse Stefan problem is not definite, while the uniqueness and stability of the solution are not always guaranteed.