INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING, VOL.

28, 1769-1785 (1989)

MODELLING LIQUID-SOLID PHASE CHANGES WITH

MELT CONVECTION

JONATHAN A. DANTZIG
Department of Mechanical and Industrial Engineering, University of Illinois, Urbana, Illinois 61801 U.S.A.

SUMMARY
Methods are described for modelling of phase change processes using the finite element method to simulate
freezing and melting including convection in the melt. Evaluation of several enthalpy/specific heat methods
and time marching schemes is also included. Suppression of velocities in the solid region is described, and
example problems are given. Comparison is made to simulationsperformed by other researchers using finite
differencemethods. Substantially different results were found for one of these problems, and this result is
shown to be caused by numerical problems in the earlier work.

INTRODUCTION
Melting and freezing problems are very important in the metals processing industry and in energy
storage systems. Understanding energy transport in these problems is important because heat
transfer during the process can determine the overall efficiency and the evolution of the process. At
the same time, phase change processes necessarily proceed with temperature gradients in the
liquid phase which, under the action of gravitational forces, lead to convection in the melt due to
buoyancy forces.
The interaction of convection in the melt along with the solidification process can have a very
strong effect on the resulting microstructure. A number of researchers have shown reorientation of
columnar grains,' alteration of the size and location of equiaxed zones' and macro~egregation,~.
all due to melt convection. Mathematical models have been used in attempts to better understand
the processes and thus control them. Although the most convenient mathematical models would
use analytical solutions to the coupled heat and momentum transport equations, very few such
solutions exist for these problems, and none would extend to the realistic problems where
complicated geometries and temperature dependent material properties are included. For this
reason, nearly all of the efforts in this area have been numerical.
There are different types of numerical methods which are appropriate to phase change
problems, depending on the kind of material involved. In pure materials, eutectics or congruent
melting phases, the liquid-solid interface is sharp and corresponds to an isotherm. For these kinds
of problems it may be appropriate to have part of the mesh coincide with the solidification front at
all times, and distort the mesh in both phases as the boundary moves. A number of these front-
tracking methods have been de~eloped,~,but none exists for three-dimensional problems.
For alloys which freeze over a range of temperatures, front-tracking methods are no longer
applicable. Instead, what is normally done is to specify the evolution of latent heat over a freezing
range as part of the material properties. The phase boundaries are then recovered from the
solution as the isotherms corresponding to the beginning and end of the transformation. The two

0029-5981/89/081769-17$08.50 Received 23 June 1988

0 1989 by John Wiley & Sons, Ltd. Revised 5 December 1988
1770 J. A. DANTZIG

most common means for inclusion of latent heat in the material propcrties are enthalpy,/speclfic
heat methods7- or heat source methods.'- l o
A further issue which applies in handling freezing range alloys, when coupled with fluid flow in
the melt, is to describe the development of resistance to flow in the solidifying material. One way to
treat this property is to introduce an increased viscosity as freezing proceeds-beginning with the
liquid phase viscosity at the start of freezing and increasing to a very large viscosity in the solid.''
A second method treats the mushy alloy as a porous medium, whose permeability decreases and
eventually vanishes with increasing amounts of solidification.' 2, '
Voller and co-workers", ' have published a series of recent papers describing new techniques
for modelling these problems. The methods were applied both to model problems and to
investigations of experiments performed by Gau and V i ~ k a n t a . 'Voller's
~ work used a finite
difference formulation, as has much of the other work in this area. This paper will describe finite
element implementations for the enthalpy/specific heat methods and flow problems, applied to the
same suite of problems. Finite element methods have certain advantages compared to finite
difference methods when dealing with complicated geometries and complex material response
functions, particularly if it is desired to later couple the heat and mass transport calculations to
stress analysis.

METHODS

Basic equations
The problems to be considered will examine liquid-solid phase changes in materials where the
liquid is assumed to be Newtonian and the heat conduction follows Fourier's law. It will also be
assumed that the fluid density depends only on temperature, following the Boussinesq approxima-
'
tion.' The balance equations for such materials are well established and are reproduced here only
to the extent necessary for the discussion at hand (cf. Reference 15).The mass balance equation in
such a system is given by

The momentum equations are given by

p;:(-+u.vu ) = -Vp+pV2U+flg(T-T,,,)

and the energy balance equation is

where U is the velocity vector, T is the temperature, p is the pressure, p is the density, p is the
viscosity, /3 is the volumetric compressibility referred to temperature T,,,, g is the gravity vector, cp
is the specific heat, k is the thermal conductivity and Q is an internal heat generation term. In the
finite element method, these equations are written as follows:1'

[C]T ( U ] =o (4)
 LIQUID SOLID PHASE CHANGES 1771

The definitions of the various matrices and vectors in terms of the element shape functions and
their derivatives are given in Reference 15, but the analogy with the terms in equations 1-3 is
apparent.

Enthalpylspecijic heat formulation

Since it is possible to include the latent heat in either the source term or in the specific heat term
in equation (3), it i s more a matter of computational efficiency than of physical meaning to choose
between them. When using the specific heat term, some have used an effective specific heat which
includes, over the freezing range, an extra specific heat corresponding to the heat evolved during
f r e e ~ i n gOthers
.~ favour formulations based on the enthalpy of the material, H , because they are
less prone to difficulties of lost latent heat during solidification. There are advantages and
disadvantages to nearly every method, but the formulations based on enthalpy seem to be more
robust in terms of less susceptibility to loss of latent heat, and relative insensitivity to time step size
and mesh density.
Even when using the methods based on enthalpy, there are still a number of possible
formulations. Some authors define enthalpy as a new variable and the energy equation is
reformulated in terms of the enthalpy.’O This method, called ‘full enthalpy’, necessitates that
temperatures be recovered later from the solution. Sometimes latent heat evolution is included in
the right hand side of equation (3) in a source term. Rolph and Bathe,16 Brent et a l l 3 and others
have used this method. My experience has been that, in finite element methods, inclusion of latent
heat on the right hand side of equation (6) tends to make the solution much less stable when
compared to inclusion of latent heat in the capacitance matrix on the left hand side of equation (6).
As an example of this, Bathe’s method can sometimes produce an unphysical ‘ringing’ near the
phase boundary.I7
In a recent review, Lewis and Roberts’ have given several formulations which provide for
inclusion of latent heat in the capacitance matrix in finite element formulations. They defined an
approximate effective specific heat, computed at each element integration point according to the
following equation:

In terms of the finite element formulation, equation (7) is written in terms of the nodal
temperatures and shape functions as follows:

where { T ) is the vector of nodal temperatures, { H ( T ) )is the corresponding vector of nodal
enthalpies and [ B ] is the matrix of derivatives of the element shape functions. This method was
formulated for the present work by first storing an enthalpy vs. temperature curve for the material,
from which the nodal values for enthalpy could be obtained by simply reading the curve. Linear
interpolation was used for values between the stored data points. The value for cp was then
computed at each integration point according to equation (8).
For transient problems, the change in enthalpy from the value at the previous time step and the
change in temperature from the value at the previous time step can also be used to construct an
effective specific heat.* The effective specific heat is given by the following expression:
dH/dt
c,=- (9)
dTfdt
1772 J. A. DANTZIG

In this case, the specific heats at the integration points were evaluated from the temperatures at
these points using the following equation:

where the subscripts refer to the time step number and the temperatures are given by
T,=[S] ( T )
where [S] is the element shape function matrix evaluated at an integration point. The enthalpies
were again evaluated by interpolation in the enthalpy temperature curve.
It should be noted that both this method and the spatial gradient method described above fail
when the temperature gradient within the element, or the change in temperature over a time step,
is very small. This is a common occurrence at the beginning of a simulation. In these cases, the
slope of the enthalpy vs. temperature curve at the temperature corresponding to the integration
poinl can be used to obtain the specific heat. This is equivalent to using an effective specific heat vs.
temperature curve. The method thus acquires one disadvantage of the specific heat curve method,
i.e. that even isothermal transformations must be spread over at least a small temperature range.
However, in this case the temperature range can be made very small.

Velocity extinction in the mushy zone

In Voller's formulation,13 a source term was added on the right hand side of equation (2) to
simulate the permeability change through the mushy zone. For the reasons described above
relating to solution stability, it was felt that use of an enhanced viscosity would be preferable in the
finite element solution procedures. It should be noted that these two models correspond to
different physical models.
The porous medium formulation corresponds physically to situations where the fluid and solid
do not move together, so that the enthalpy and the resistance to flow do not have the same
dependence on the field variables. For example in the growth of a columnar array of crystals, the
solid consists of attached dendrites and the interdendritic fluid trickies between them. The
enhanced viscosity formulation corresponds better to the physical situation of equiaxed solidi-
fication, where free-floating crystals are transported with the melt, with the proportion of these
crystals in the fluid increasing with increasing solidification. Both models correspond to real
physical events in different kinds of castings and, hence, are both valid. Enthalpy is convected with
the solid-liquid mixture in the case of free-floating crystals, while it is not convected with the fluid
if the solid is fixed, as in the porous medium model.
Following the work of Joly and Mehrabian'* in examining the rheology of partially solidified
alloys, the viscosity vs. temperature curve for mushy alloys was used with the viscosity increasing
very slowly from 0 to 40 per cent solid, then increasing rapidly to a very high value corresponding
to the solid. An example is included below.

Solution of the finite element equations

The following calculations were performed using the commercial package FIDAP, modified to
include the entAalpy/specific heat formulations described above. A number of different methods
were evaluated for both time integration and iterative solution of the equations generated at each
time step. It was found that the most robust and computationally efficient method was to use
trapezoidal (Crank-Nicholson) time integration along with Newton-Raphson iteration for
 LIQUID SOLID PHASE CHANGES 1773

solution at each time step. The time step size was selected using a variable time integration
procedure incorporated in FIDAP.” After convergence at the present step, the time step size for
the next time step was computed based on control of the local truncation error in the time
discretization. Specifically,

where E is a time step truncation error parameter and d , is the r.m.s. change in velocities and
temperatures at time step n, compared to step n- 1. The reader is referred to the reference manuals
for the code for further details. If the factor in the radical was smaller than 0.7, the time step size
was halved and the time step was repeated. If the value of factor exceeded one, the next time step
was increased, and if the factor was between 0.7 and 1.0, the time step size was not changed. The
optimum value for E was established by experimentation in the first set of problems described
below to be 1 . 0 ~
Iteration for solution was performed at each time step using a Newton-Raphson procedure,
also described in the FIDAP theoretical manual. In the most strongly non-linear problems, it was
sometimes necessary to include a relaxation factor of 0.5 in the Newton-Raphson iterations. At
each time step, convergence was declared when the relative errors in the nodal values of the
temperatures and velocity, and the relative errors in the residuals of the momentum and energy
equations, were both less than 0.001.

RESULTS

Enthalpylspecijic heat formulations

The enthalpy formulation was tested on a series of Stefan problems, formulated in two
dimensions. The material was contained in a domain of half-thickness L, and was initially all
liquid at the freezing point, T f . All faces were insulated, except for the left hand face (at x = 0), which
at time zero was set to a new temperature below the freezing point. The differential equation and
boundary conditions for this problem are given by

aT
pc --=kV2T (13)
p at
T(x, y, t=O)= (14)
T(x=O,y, t>O)=T, (15)
VT.n=O (all other faces) (16)

In addition, on the phase boundary, S(t),

T= Tf
dS

S(t =O) =o (19)

where L, is the latent heat of fusion.
To test the numerical method, it was convenient to non-dimensionalize the equations according
1774 J. A. DANTZIG

to the following scheme:

c1
7=-t
L2
where r is the thermal diffusivity. The governing equations and boundary conditions are then

Thus, the only parameter in the problem is the Stefan number, Ste. Simulation becomes
increasingly difficult with decreasing value of the Stefan number. This problem is particularly
useful for testing the formulations because it has an exact solution given by the following
equations: ’

and ^J is the solution of

A series of numerical experiments was carried out to test the various formulations using
different time step sizes, mesh densities and means for implementing the boundary conditions. The
best results were found when there werc at least four elements through the half-thickness and at
least 45 time steps were used to span the freezing time. Table I shows a comparison of computed
and analytical results for the solidification time for various methods. Previous studies have
indicated that the solidification time is an important parameter for testing the validity of the
solution, because the most likely problem is loss of latent heat, which results in error in
solidification time.” Crank-Nicholson time integration was preferred because its time truncation
error is smaller, resulting in fewer time steps and correspondingly shorter execution times.
Table I includes data from solutions using all three of the enthalpy/specific heat formulations
described above. It can be seen that the time differentiation method was most accurate, the spatial
gradient method worked consistently overestimated the freezing time significantly and the
 LlOUID SOLID PHASE CHANGES 1775

Table I. Summary of results from one-dimensional Stefan problems

Stefan Specific Number of Freezing Error CPU + 1/0 Number of

number heat elements time (per cent) (s) time steps

1 Temporal 3 0.529 - 18.6 9-7 46

t = 0650 4 0.603 - 7.2 13.6 54
(exact) 5 0.607 - 6.6 15.7 58
6 0.675 3.8 19.4 60
7 0.670 3.0 24.0 65
8 0.702 8.0 29.8 71

Spatial 3 0.645 - 0.8 9.8 37

4 0.741 14.9 123 42
5 0.780 19.9 15.8 45
6 0.799 23.0 18.1 48
I 0.802 23.5 20.2 51
8 0.803 23.5 31.3 55

Standard 3 0.435 -33.1 10.7 48

4 0.490 - 24.6 18.6 56
5 0561 - 13.6 21.0 65
6 0.586 - 9.9 22.7 67
7 0589 - 9.4 30.7- 74
8 0590 - 9.2 27.7 70

0.5 Temporal 3 0.923 - 20.2 13.2 51

t=1.156 4 1.057 - 8.5 16.5 61
(exact) 5 1.160 0.4 17.8 65
6 1.189 2.9 24.9 76

Spatial 3 1.254 8.5 11.7 42

4 1.404 21.4 12.9 47
5 1.475 27.6 15.6 51
6 1.495 29.3 19.1 54

Standard 3 0.815 -29.5 13.9 58

4 0.904 -21.8 19.4 64
5 0.967 - 16.4 23.5 73
6 1.019 -11 8 32.4 82
_~~ _ _ _ _ ~ _ _
0.33 Temporal 3 1.399 - 15.8 11.1 52
t= 1.661 4 1575 - 5.2 16.3 63
(exact) 5 1.623 - 2.3 23.4 74
6 1.676 0.9 26.0 80
_ ~ _ __ ~ _~_ -~ ~ _ _ ~

standard specific heat formulation (i.e. the effective specific heat formulation) consistently
underestimated the freezing time significantly. In all cases, a stable value of the solidification time
appeared once four or five elements were used.
The reason for the lack of accuracy in the standard specific heat formulation can be understood
by examining Figure 1, which shows a possible temperature distribution within a given element at
some time in the simulation. At this time, the freezing range happens to fall between integration
points. Therefore, the effective specific heat at each integration point would have a low value,
corresponding either to liquid or solid, even though freezing is taking place in the element. Because
1776 I. A. DANTZIG

of this, temperatures at the nodes would change significantly over the time step, and it is possible
that at the next time step the freezing range would jump past the integration points which were in
the liquid phase, unless the time steps are kept very small. If this occurs, some of the latent heat
disappears from the problem, resulting in a freezing time which is too short. The advantage of the
enthalpy/specific heat techniques is that they are much less sensitive to time step size, as both
methods would not have missed the latent heat for this case.
The simulations were coupled to the variable time stepping algorithm as described in the
previous section. After some experimentation with the parameters, the formulas given in the last
section were found to produce the best results. An interesting interaction was found between the
variable time step selected using this algorithm and the solution when using the enthalpy
formulations. As the freezing front moved through an element, the time step size increased rapidly

Freezing Range

Figure 1. Isoparametric element with freezing range inside

0.20

0.16

-= 0.12
Ic
-Y
L
U

0.08

0.04

I I 1
0.00

Tine
Figure 2. Time step size vs. time in simulation of Stefan problem using four elements
 LIQUID SOLID PHASE CHANGES 1777

until the phase boundary crossed an element boundary. At this point the d u t i o n changed rapidly
so that the variable time stepping algorithm cut the time step size very sharply. The evolution of
the time step size as a function of time is depicted in Figure 2. The sharp edges iii the curve
correspond to the phase boundary crossing the element boundaries. Nevertheless, because the
algorithm controls both the time truncation error and the error in the solution at. each time step,
the solution did not deteriorate.
The source of the rapid change in time step size is understood by examining Figure 3. In the
freezing element, the nodal .ialrre of the temperature is held at the freezing point until all of the
latent heat from the element is released. While this is occurring, the temperature distribution in the
solid portion becomes a straight line between the temperature at x = 0 and the freezing point at the
first node still liquid. When that node passes through the freezing point, the temperature
distribution changes rapidly to reach a similar quasi-steady value to the next node. In this period
of rapid change, the time stepping algorithm squeezes the time steps down to very small values.
Because the temperature distribution jumps between these positions, the time histories of the
nodes takes on a stepped appearance, as shown in Figure 4.

Succensivc rest states

T=T, dIU Diwce

Figure 3. Quasi-steady temperature distributions in the one-dimensional Stefan problem. Top, bottom and right hand
sides are all insulated

I 1 I

0.50 f.00 1.50 2.00

Ti=
Figure 4. Temperature vs. time at successive nodes in the one-dimensional Stefan problem using four elements and Stefan
number of 0.33
1778 J. A. DANTZIG

The problem just described is especially acute in the first element. If the boundary condition at
x = O (0 = 0) is strictly apphed, the temperature distribution in the element is constraiiied to be a
straight line (by choicc ofthe interpolation function), and regardless of the choice of time step size,
latent heat is lost. 'To overcome this problem, the boundary condition at x=O was weakened to
one with a large hcat transfer coeffjcient convection to the true boundary condition temperature.
In dimensionless form.

Experimentation showed that a Riot number (Bi)of 15 based o n the element dimension gave
satisfactory results.

Freezing of u mush!, nlloy

Voller and co-workers' described a test problem for their enthalpy/porosity formulation
which consisted of the freezing of a hypothetical mushy alloy. The geometry for this problem is
shown in Figure 5. The alloy, which freezes in the range -0.1 to +0.1, was initially held at
8= +0.5 in a square cavity of unit edge length. At t = O , the left hand edge was instantaneously
cooled to temperature 0 = -0.5 and solidification proceeded into the cavity.
This problem was modelled using the mesh of four-noded linear isoparametric elements shown
in Figure 6. The fluid flow equations werc solved on this mesh using the Crank-Nicholson time

Figure 5. Domain for freezing of mushy alloy, after Voller

Flgure 6. Finite element mesh used to simulate freezing of the mushy alloy
 LIQUID- SOLID PHASE CHANGES 1779

integration and Newton-Raphson iteration described earlier, and using the same convergence
criteria. The material properties used are given in Reference 13.
The Rayleigh number for this problem is moderately large, at lo4. However, because the
Prandtl number is also large (lo3),the Grashof number, which is the important parameter, is just
10 and the problem is not especially difficult from a computational standpoint. Figure 7 shows the
position of the computed liquidus and solidus lines at various times, including velocity vectors to
indicate the flow. These results are qualitatively very similar to those reported by Voller, in terms
of the shape and location of the liquidus and solidus isotherms. There was not sufficient detail in
Voller's paper to compare the two simulations quantitatively. The calculations used 276 time steps
to cover 2000 s of simulated time, which took a total of 18 cpu hours on a Sun 3/280 FPA.

Simulution of the melting of gallium

Gau and V i ~ k a n t a performed
'~ a series of experiments where they melted and/or solidified
gallium in rectangular enclosures and observed the motion of the liquid-solid interface with time.
The geometry and pertinent experimental conditions are depicted in Figure 8. The experiment was
simulated in much the same way as the previous problem.
This problem represents a much more difficult numerical challenge than the previous one. As
the material melts from the left hand side, a cavity is generated whose aspect ratio is very high at
first when the cavity is thin, and decreases as time proceeds and more material is mclted. The
appropriate Rayleigh number i s one based on the thickness of the melted layer, and thus increases
throughout the experiment. The Prandtl number is small (0,0216)so that the importance of the

.........

,, , ,
,.
.. --_.- .- - _
-
- - 2

-_ . - - -

IS I ' ' .

I . .:.. . .
. , . .I:. : .:',:. :.. :-::::::::I
.....
. .. .. . . . . .. . . . .. - .-, ...... --
.. .. ... .. ... .. ... ... ... ... . ... ... ... ... .... . .. ... .. . . .
-__,,..I

Figure 7. Temperature and velocity vector fields computed at dimensionless times of 50,250,500 and 1OOO after beginning
of simulation. Note that velocities decrease with time, as effcctive Rayleigh number decreases. Maximum velocity in all
plots is 0.015 in dimensionless units, at time equal to 50
1780 J. A. DANTZIG

inertial terms in the momentum equations is accentuated. The Stefan number in the experiment
was also very small (0.05).
Four different meshes, shown in Figure 9, were used. spanning the intervals 0-49, 49-215,
21 5-542 and 542-630 s, respectively. The latest solution obtained on the preceding mesh was
interpolated onto the new mesh at each change. The progress of the melt interface was similar to
the results of Gau and Viskanta. The results obtained by Voller and the present work match this
apsect of the experimental data rather well.
The flow patterns obtained in this work, however, differed substantially from those of Voller.
Selected results are shown in Figures 1&12. Multiple convection cells were found to develop in

,,,Insulatcd

--I/
1
~ Liquid

T=7KC -- -
1

Solid

Figure 8. Schematic view of experimental apparatus used by Gau and Viskanta for melting of gallium.

0 - 44 s e c 49 - 215 sec

215 - 5 4 2 sec 542 - 630 sec

Figure 9. Meshes used in simulation of gallium melting. Applicable times are (beginning from the upper left): W9,49-215.
215-542 and 542-630 s
the meit about 30 s after the onset of melting. 'The number of cells changed with time, beginning
with six cells at 303, and decreasing to four, three, two and finally to one as the experiment
progresscd. The primary determining factor for the number ofcells, once they appeared, seemed to
be the aspect ratio of the cavity. The mechanism for d ~ r e a s i n gthe number of cells was for two

._

! i iet. 22 sc:

Figure 10. Evolution olniultiple cells in melted region from 16 to 32 s. Each figureshows streamlines of -0.1, --0.2, --0.3,
etc. to -2:60. along with the freezing point isotherm. Note the appearance of multiple co-rotating cells

", -' I e ' ' 1 se: i G ic: :9 iec ra ie.

Figure 11, Continued evolution of multiple cell pattern in melted region from 42 to SO s. Each figure shows slreamlines as
iiiFigure 10, and freezing point isotherm. Note the cell pattern change effected by consolidaiion of cells in :he ripper
portion oi the domain
1782 J . A DANWIG

Figure 12. Continued evolution of multiple cells in melted region from 350 t o 500 s. Each figure shows streamlines and
ireezing point isotherm as before. Note that the strong circulation causes some weak recirculation between cells, as well as
the sevcrc distortion of the melt interface

existing cells to join, usually at the top of the domain. See Figures 10 12.
Because the shape of the phase boundary was similar to both those seen in the experiment and
Voller’s simulation, one can conclude that the shape of the freezing front was not especially
sensitive to the exact nature of the convective flow. The melt is opaque, and no description of the
flow structure was reported by Gau and Viskanta. Thus, it is not possible to conclude from the
comparison with experimental data alone which of the simulations is correct. This point is
important, howewr, because one can anticipate a strong effeci on the segregation patterns if the
melt had had a second component.

DISCUSSION
The results shown in the previous sections indicate that the formulations developed for modelling
the phase change with natural convection using enthalpy/specific heat and enhanced viscosity
work rather well. The most interesting point to pursue is the difference in the nature of the flow
found in the gallium melting problem. Is the multiple cell solution or the single cell solution
correct?
The problem of natural convection in high aspect ratio chambers with a heated vertical wall has
been examined by a number of authors. These studies were stimulated by interest in heat losses
from thermopane windows. Chikhaoui et a1.” examined this problem both numerically and
experimentally for air. Lee and Korpela22 also examined a similar problem numerically. Both
groups of authors demonstrated that when the Grashof number exceeds 1.2 x lo4 multiple cells
appear in the flow for high aspect ratio cavities. In experiments in air (Prandtl number=0.7)
Chikhaoui ei ul. found three cells, and also concluded that there were important three-dimensional
effects. Lee and Korpela performed two-dimensional simulations, and found multiple cells at
Grashofnumber1.2 x lo4 with a Prandtl number of Lero. Figure 2 in their paper indicates six cells
and is remarkably similar to those in Figure 10.
It is possible to compute the expected time of onset of multiple cell convection in the gallium
melting problem using the experimental data provided by Gau and Viskanta. The thickness of the
solidified layer for very early time should grow approximately in proportion to the square root of
time (see equation (31)). This assumption probably remains valid so long as the melt front shows
little disturbance from the vertical. A t t = 120 s, Gau and Viskanta show the melt interface to be
 LIQUID SOLID PHASE CHANGES 1783

vertical at a distance of 0.96 mm from the left hand wall. Calling the melt thickness 6, and using the
properties of gallium given in Reference 13. the Crashof number (Cr) is given by
G r = 1.1 x t05d3 (33)
with 6 measured in mm. The time at which the (;rash& number should be equal to 1.2 x lo4 is then
30s. This time agrees very closely with the value found for the onset of multiple cells in the
simulations. It seems likely, therefore, that the multiple cells found in the present simulation are
correct.
Voller and co-workers used the SIMPLE algorithm,23 which includes upwinding for highly
convective problems, This scheme has been criticized for being overly diffusive and suppressing
some eddies because of artificially increased viscosity.24 To determine whether this was a possible
source of the discrepancy, upwinding was included in the present solution strategy. FIDAP uses
‘streamline upwinding’ which is purported to create less numerical diffusion than the upwinding
procedure in the SIMPLE algorithm. The amount of upwinding is controlled by a parameter
which the user can set. See Reference 15 for details. The result is shown in Figure 13, where the
suppression of some of the cells is quite apparent. l h i s suggests that the upwind differencing is a
likely source of the difference between the present work and that of Voller. Figure 14 shows the
computed maximum values of the stream function in Voller’s work compared to those foutid in the
present work, which resulted in values nearly three times as large. This further supports the
possibility that the extra diffusion due to upwind differencing caused the multiple cells to be
suppressed.
It is also possible that the mesh used in Voller’s work was not fine enough to resolve the multiple
cells. This could not be tested with the present code, however, because the solution failed to
converge beyond about 30 s (corresponding to the onset of the multi-cell convection) when using a
coarser mesh. Since the mesh density and time stepping required to resolve the complex flow
structure were much finer than those used by Voller, the execution of the sirnulation took much
longer. The simulation was run in several parts, some on a Sun 3/280 and others on a Cray X-MP.
The total number of time steps from 0 to 630 s was approximately 4500, and it is estimated that the
run would havc totalled about 50 cpu hours on the SUN3/280.

16 sec 20 sec 24 set 28 sec 32 sec

Figure 13. Evolution of cell patterns when upwinding was included in the solution. Times correspond to those in Figure
10. Note suppression of several cells
1784 J. A. DANTZIG

2.00

-8e
Lo
1.50

4
__2
W
S
x
1.00

0.50

0.00
0 200 400 60 0 800

Ti*
Figure 14. Comparison of the maximum absolute value of the stream function in the present work to that reported by
Voller. The ripples in the present work arise from the unstable nature of the flow (see Figure 12)

Gau and Viskanta' were unable to obtain reproducible results in their experiments when
doing solidification problems. They speculated that the large degree of anisotropy in the thermal
conductivity of gallium in the solid state was responsible for this. The present work suggests a
second factor which may have been important. The multiple cells which develop can have a strong
effect on the shape of the liquid-solid interface. Note that the interface shape computed in the
present work is somewhat more jagged than that reported by either Voller or Viskanta. The
evohtion of multiple cells has been shown to be very dependent on the initial state22and it is
therefore possible that a series of experiments might have produced slightly different multiple
cellular states, corresponding to different interface shapes. Experiments with better resolution of
the flow fields, or a material with conductivity closer to isotropic, would be necessary to resolve
this question.

CONCLUSIONS
Techniques have been implemented, using the a finite element method, to analyse liquid solid
phase change problems including natural convection. Several entha1py;specific heat formulations
were developed and tested for accuracy by comparing with analytical solutions. Enthalpy
methods using either spatial or temporal gradients were found to give the best results. Modelling
of the increased rigidity with solidification was accomplished using an enhanced viscosity model.
The results agreed very well with those of Voller, who used a source term to model the mushy zone
as a porous medium.
In a test problem relating to melting of gallium, corrcsponding to experimental results obtained
by Gau and Viskanta, the present solution resulted in convection patterns different from those
found by Voller. Comparison with other work in natural convection suggests that the present
 LIQUID SOLID PHASE CHANGES 1785

result showing multiple cells is correct, and that the source of the discrepancy is numerical
diffusion caused by the use of upwind differencing in Voller’s scheme. The progress of the phase
boundary, however, was not strongly affected by the nature of the cellular pattern.

ACKNOWLEDGEMENT

The author would like to thank Dr. Michael Engelman of Fluid Dynamics International for help
in implementing the enthalpy methods and stimulating discussions through the progress of the
work. He would also like to thank Professor Vaughan Voller of the Univeristy of Minnesota for
suggesting the problem and several discussions of the results. Finally, the support of the National
Center for Supercomputing Applications at the University of Illinois, and of Cray Research is
gratefully acknowledged.

REFERENCES
1. G. S. Cole, Transport Processes and Fluid Flow. In Solidijcation. Am. Sac. Metals, Metals Park, 1971.
2. S. C. Flood and J. D. Hunt, in S. Kou and R. Mehrabian (eds.), Modeling and Control of‘Casting and Welding Processes,
pp. 607-618, Warrendale, PA, 1986. TMS-AIME.
3. M. C. F-lcmings, Sofidijcation Processing. McGraw-Hill, New York, 1974.
4. W. D. Bennon and F. P. Incropera, ‘The evolution ofmicrostructure in statically cast binary ingots’, Met. Trans., 18B,
611-616 (1987).
5. J. Crank. Free and Motling Boundary Problems. Oxford Univ. Press, Oxford, 1984.
6. 11. M. Ettouney and R. A. Brown, ‘Finite-element methods for steady soliditication problems’, J . Comp. Phys.. 49,
118-150 (1983).
7. 0.C. Zienkiewicz, The Finite Element Method. McGraw-Hill, New York, 1977.
8. R. W. Lewis and P. M. Roberts, ‘Finite element simulation of solidification problems’, Applied Scientific Research, 44,
61-92 (1987).
9. C . W. Kim and G . K. Ruhlandt, ‘A 3-D finite element method program for casting soliditication simulation’. in S. Kou
and R. Mehrabian, (eds.), Modeling ond Control qfCasting und Welding Processes, pp. 557~~564, Warrendale, PA, 1986.
TMS-AIME.
10. V. R. Voller and C . Prakash, ‘A tixed grid numerical modelling methodology for convection--diffusionmushy region
phase change problems’, Int. J . Heat Mass Trans$&-, 30, 1709-1719 (1987).
11. D. K. Gartling. In Computer Methods in Fluids. Pentech, 1980.
12. R. Mehrabian, M. Keane and M. C. Flemings, ‘Interdendritic fluid flow and macrosegregation; influence of gravity’,
Met. Trans., IB, 1209-1220 (1970).
13. A. D. Brent, V. R. Voller and K. J. Reid, ‘Enthalpy-porosity technique for modeling convection-diffusion phase
change: application to the melting of a pure metal’, Num. Heat Transfer, 13, 297-318 (1988).
14. C. Gau and R. Viskanta, ‘Melting and solidification of a pure metal on a vertical wall’, J . Heat Transfer, 108, 174-181
(1986).
15. M. S. Engelman, F I D A P Theoretical Manual. Fluid Dynamics International, Evanston, IL, 1987.
16. W. D. Rolph 111 and K. J. Bathe, ‘An efficient algorithm for analysis of nonlinear heat transfer with phase changes’, Int.
j . numer. m ~ t h o d seng., 18, 119-1 34 (1982).
17. A. B. Shapiro, TOPAZ reference Manual. Lawrence Livermore National Laboratory, Livermore, CA, 1984.
18. P. A. Joly and R. Mehrabian, ‘The rheology of a partially solid alloy’, J . Mat. Sci., 11, 1393--1418 (1976).
19. J. Stefan, ‘Uber die theorie der eisbildung, inbesondere iiber die eisbildung im po!armeere’, Ann. Phys. and Chem., 42,
269-286 (1891).
20. J. A. Dantzig, Unpublished Research. University of Illinois, 1987.
21. A. Chikhaoui, P. Bontoux, M. K. Maslanik and R. L. Sani, ‘Steady threedimensional thermal convection in a vertical
rectangular enclosure: transition to multicellular flow’, in M. S. Engelman (ed.), FIDAP User’s ConJKrence, pp.
4.14.19. Fluid Dynamics International, 1987.
22. Y. Lee and S. A. Korpela, ‘Multicellular natural convection in a vertical slot’, .I. Fluid A4ech., 126, 91-121 (1983).
23. S. V. Patankar, Numerical Heat Tranger and Fhid Flow. Hemisphere, Washington, D.C., 1980.
24. A. N, Brooks and T. J. R. Hughes, ‘Streamline upwindiPetrov-Galcrkin formulations for convection dominated flows
with particular emphasis on the incompressible Navier-Stokes equations’, Comp. Meth. Appl. Mech., 32, 199- 259
(1984).