Numerical methods for Stefan problems

Numerical methods for Stefan problems

Adaptive finite element methods for degenerate parabolic problems

A. Schmidt
Z. Chen, R.H. Nochetto, C. Verdi
Phase transitions (liquid-solid, e.g.) can be modelled by degenerate parabolic equations. The classical Stefan problem is one example. In our work group, we develope error estimators and adaptive finite element methods for such equations.

Classical Stefan problem

Eriksson and Johnson [ErikssonJohnson:91] developed the first a posteriori error estimators for finite element diskretizations of linear parabolic problems (like heat equation), where only computable terms like the discrete numerical solution enter. These methods were extended by both authors [ErikssonJohnson:95b] and by Verfürth [Verfuerth:95a] to classes of nonlinear parabolic equations. But degenerate parabolic problems like the classical Stefan problem

u_t - \Delta \beta(u) = f mit \beta(s) = min(s,0) + max(s-1, 0)

are not covered by them. As \beta vanishes in the whole interval [0,1], the equation is not uniformly parabolic, so the above techniques can not be applied. In our work groupt, adaptive methods for degenerate parabolic problems were developed [NSV:97b]. In [NSV:97a], a-posteriori error estimators for the Stefan problem are developed, which give upper bounds for the error in temperature in $L^2(L^2)$ and for the error in enthalpy in $L^\infty(H^{-1})$. The methods are based on the use of a (linear, degenerate) dual problem. The adaptive finite element method was implemented in two and three space dimensions, with local mesh refinements based on the bisecion of mesh elements (triangles or tetrahedra), based on the finite element toolbox ALBERT. In [NSV:97a] we demonstrate the application to a two dimensional model problem, more simulations in two and three space dimensions are presented in [NSV:97g,NSV:97e].

NSV4_1.gif, 17 kB
Moving interface for a 3D Stefan problem

Continuous Casting

The continuous casting problem is a Stefan problem with given, dominant convection. It can be used to model the casting of steel. $L^2$-norm based methods lead here to an exponential dependence of the constants in the error estimator on the maximal convection. In [CNS:98a] we present a posteriori estimators for the $L^1(L^1)$-norm of the temperature error, which results in a much weeker dependence of constants from the convection. The next figure shows the temperature graph and the corresponding adaptive mesh (with a vertical zoom by factor 16) from a two dimensional simulation.

CNS1_1.gif, 17 kB
Temperature graph from a 2D continuous casting simulation

CNS1_2.gif, 5 kB
Corresponding finite element mesh (vertical zoom by factor 16)

Phase relaxation problems

Relaxations of the classical Stefan problem were introduced and analyzed by Visintin [Visintin:85,Visintin:86]. They lead to a system with one equation for the temperature and one for the phase variable. In contrast to the classical Stefan problem, the temperature at the phase boundary does not necessarily coincide with the melting temperature of the substance. In [CNS:97] we present a posteriori error estimators and an adaptive finite element methods for such problems. The application to phase field equations is work in progress.

Back to the projects overview