Applied Scientific Research 4 4 : 6 1 - 9 2 (1987) 61

© Martinus Nijhoff Publishers, Dordrecht - Printed in the Netherlands

Finite element simulation of solidification problems

R.W. LEWIS & P.M. ROBERTS

Institute for Numerical Methods in Engineering University College of Swansea, University of Wales,
Singleton Park, Swansea SA2 8PP, United Kingdom

Abstract. The modelling of liquid-solid phase change p h e n o m e n a is extremely important in m a n y

areas of science and engineering. In particular, the solidification of molten metals during various
casting methods in the foundry, provides a source of important practical problems which m a y be
resolved economically with the aid of computational models of the heat transfer processes
involved. Experimental design analysis is often prohibitively expensive, and the geometries and
complex boundary conditions encountered preclude any analytical solutions to the problems
posed. Thus the motivation for numerical simulation and computer aided design (CAD) systems is
clear, and several mathematical/computational modelling techniques have been brought to bear
in this area during recent years.
This paper reports on the application of the finite element method to solidification problems,
principally concerning industrial casting processes. Although convective heat transfer has been
modelled, the work herein considers only heat conduction, for clarity. The heat transfer model has
also been coupled with thermal stress analysis packages to predict mechanical behaviour including
cracking and eventual failure, but this is reported elsewhere.
Following the introduction, the mathematical and computational modelling tools are described
in detail, for completeness. A discussion on the handling of the phase change interface and latent
heat effects is then presented. Some aspects of the solution procedures are examined next, together
with special techniques for dealing with the mold-metal interface. Finally, some numerical
examples are presented which substantiate the capabilities of the finite element model, in both two
and three dimensions.

Nomenclature
c = heat capacity TL = liquidus temperature
C = capacitance matrix To = initial temperature
f = time function T~ = solidus temperature
F = loading term x = space coordinates
h = heat convection coefficient a = interface heat transfer coefficient
H = specific enthalpy 3' = iteration parameter
[ J[ = Jacobian determinant F = boundary of domain
[ J[ = patch approximation to [ J I 8T = solidification range
k = thermal Conductivity At = timestep magnitude
K = conductance matrix V = vector gradient operator
L = latent heat e = convergence tolerance
= unit outward normal 0 = timestepping parameter
N~ = nodal shape function 0 = known vector in alternating-direction
q = known heat flux formulation
R/ = nodal heat capacity X = Laplace modifying parameter
S = phase change interface (~, ~ ) = local space coordinates
t = time P = density
T = temperature ~ = time limit
= known boundary temperature q~(~) = shape function factor
T = vector of nodal temperatures +(~) = shape function factor
Ta = ambient temperature a = domain of interest
Tc = solidification temperature
62 R.W. Lewis and P.M. Roberts

1. Introduction
The growing use of computational modelling techniques in recent years
reflects the potential economic benefits they offer to many industrial processes,
including casting methods in the foundry. Numerical simulations of the
thermal, and mechanical, behaviour of both castings and moulds, will permit
designers to create robust, effective products, whilst simultaneously minimis-
ing process costs and waste. To achieve a practical simulation package, a heat
transfer analysis program is required at the core of an overall flexible CAD
system, incorporating pre- and post-processing graphics and other geometric
modelling tools. With such a system the designer can reliably predict hot spots
which could result in shrinkage porosity; alterations in geometric configura-
tion or to process parameters will eliminate such risks and ensure directional
solidification in advance of any casting being poured. The desired mechanical
properties of the finished casting can be controlled by predicting, via numeri-
cal experimentation, the effects of using differing mould materials and the
placement of various chills and/or cooling lines. Waste associated with risers
can be significantly reduced by optimising their magnitude and location.
This paper focusses on the mathematical model which is the basis of the
numerical heat transfer analysis program. The development of a non-linear
heat conduction simulator which uses the finite element method is presented.
Finite elements are employed because of their facility in modelling complex
domain configurations, and handling of non-linear boundary conditions. For
completeness some details of the discretisation process are given together with
a brief description of the mathematical model, i.e. differential equation and
boundary conditions, involved.
Various timestepping methods are considered for use in conjunction with
the Galerkin semi-discrete weighted residual form of the differential equation,
and the treatment of the non-linear terms which arise is also discussed. The
representation of latent heat effects is an important feature in the modelling of
solidification problems. The tracking of the phase change boundary is an
important numerical problem, which is resolved in our work by use of the
enthalpy method which incorporates latent heat in a smooth manner.
Techniques for improving the efficiency of the simulation program are then
looked at, including an alternating-direction implicit, and a 'mixed'
implicit/explicit solution algorithm coupled with a direct profile solver. A
quadratic conductivity approximation is also presented which speeds up
matrix generation.
A crucial component in simulating casting solidification is the handling of
the heat transfer at the interface between the mould and the metal, a critical
process parameter acutely affecting the solidification ratel The heat transfer
coefficient can vary in response to type and thickness of die coat, to the
pressure, temperature and cooling rate of the metal, and due to the formation
of an air gap. A coincident node technique is illustrated for dealing with the
air gap problem. A short description of an infinite element approach for
representing heat losses to an external medium is also given.
 Finite element simulation of sofidification problems 63

The numerical results from a series of simulations of practical problems are

then shown. The problems solved for include the techniques presented earlier,
and involve both two and three-dimensional meshes.

2. Heat transfer models

2.1. Mathematical models

The mathematical model consists of the differential equation governing the

conduction of heat through a medium, coupled with appropriate boundary
conditions. The Fourier law of conservation of energy may be expressed as
v'(kvT)=ocT; (x, t) e f X [ 0 , ~'1. (1)
The form of boundary condition is usually of two types, i.e. Difichlet (known
temperature) or Neuman (known heat flux), or can be a mixture of both. If we
denote the distinct boundary regions F1, £2 of the domain of ~ (see Fig. 1)
where Difichlet and Neuman conditions are known, then we have
T(x, t) = 7~(x, t); (x, t) ~ rl x [0, ,] (2)
and
kVT.h+q+h(T-T~)=O; (x,t)~F2x[O, ~-] (3)
where F 1 • F2 = d?.
Since we are dealing with a transient problem we must also specify the
initial state, or temperature distribution, giving the initial condition
T(x, O)= ro(x); x~2. (4)
In situations involving a change of phase, the classical Stefan problem is used

I I A

Fig. 1. Domain of interest and phase changeinterface.

64 R.W. Lewis and P.M. Roberts

to describe the conduction in a domain comprising two separate phases. Thus

on the subdomains f~l, ~-]2we have (f]l U f~2 = f], see Fig. 1)
V" (k i vTi) =loiciTi; (x, t) ~ a iX [0, T] (5)
Ti(x, t)= Tii(x)'fi(t); (x, t ) ~ F l i X [ 0 , T] (6)
kivZi'n+qi}-hi(Ti-rai)-~O; (x, t ) ~ r 2 i x [0 , 0"] (7)
V/(x, 0)= (s)
Furthermore, since the interface S is moving, we must have the interface
condition
kl VTI" ~s - k2 VT2" ~s = pLS. (9)
In the above form, the Stefan problem requires a front tracking algorithm
to determine the interface S at any given time - this requires special
timestepping or deforming mesh techniques. However, a 'weak' form of the
above problem can be posed by defining the enthalpy as

H= forpc dT (10)

or
dH
p c - dT" (11)
Then Eq. (1) (or Eq. (5)) becomes

V" (k V T ) = _Td__d_~H
_• (12)

or
V - (k V T ) =/~r. (13)
Thus in Eq. (11) we are only solving for temperature; the front tracking
problem has been 'removed', and the latent heat L will be included in the
definition of enthalpy H (see below).

2.2. Numerical models

2.2.1. Spatial discretisation. The finite element method has been chosen to
discretise the domain f] in space, rather than by using finite differences, on the
basis of its inherent flexibility in handling complicated shapes and boundary
conditions. Thus the domain 62 (2- or 3-dimensional) is divided up into
distinct regions or elements ~2e such that
M
U ~e=~.
e=l
Nodal points are then distributed at element vertices, along the element
edges, and possibly in the element interiors, as in Fig. 2. The usual element
 Finite element simulation of solidification problems 65

<ZZ22
Fig. 2. Schematic discretisation of general domain into finite elements.

shapes consist of triangles and quadrilaterals, with tetrahedra or parallele-

pipeds in 3-dimensions. The application of the finite element method to the
heat conduction equation is well documented, and indeed appears as a
standard example in nearly all finite element textbooks [1]. Therefore, only a
brief description is warranted here. We approximate the temperature distribu-
tion by a weighted series of locally supported polynomial functions - the
weights/amplitudes correspond to the temperatures at the nodal points or
'nodes'. Thus we have the approximation
N
l~(x, t) = E Ni(x)T~(t) (14)

where N~ are the basis, or shape functions and T~ the nodal temperatures.
Substituting this approximation into the Galerkin weighted residual form of
the differential equation, Eq. (12), we obtain the semi-discrete system
CT + KT = F (15)
where T is the vector { T/} of unknown nodal temperatures. The capacitance
matrix C, the conductance matrix K, and the thermal loading term F are
defined as follows:
dH
c,, = £-a-f u,u, d a (16)

Kij= fovu,. (k vNi) df~ + fr2hN~Nj dF 2 (17)

= - fr2( q - hr,)n, dr2. (18)

The appearance of the boundary integral terms in Eqs. (17), (18) takes into
account the flux boundary condition Eq. (3). The symmetric form of K
reflects the implicit use of a Green's identity ('integration by parts') to obtain
the weak form of the residual equation. Given the integrals of Eqs. (16)-(18),
66 R.W. Lewis and P.M. Roberts

we still have to evaluate them numerically. Analytical integration can only be

performed on simple, regular elements, and thus recourse to numerical integra-
tion is necessary. To facilitate this, the irregular element configurations are
mapped onto regular domains, typically [ - 1 , 1]", via isoparametric transfor-
mations [1]. The isoparametric mappings use the same functions Ni which
form the basis of the approximation space. Thus boundary curves of the same
order as the polynomial functions N/ can be represented exactly. Having
mapped each element onto the same convenient domain we then have to
choose a quadrature rule which is sufficiently accurate, but which does not
require excessive, even redundant, computation. A popular choice meeting
these requirements is that of the Gauss-Legendre quadrature rule which places
the sampling points at the zeroes of Legendre polynomials. In one dimension a
single point integrates exactly a straight line, two points a cubic polynomial,
three points a quintic polynomial and so on. 'Standard' Ganss-Legendre
integration rules have been employed throughout in the heat transfer programs
developed, although for 3-dimensional elements cheaper schemes with fewer
points but similar accuracy could have been utilised.

2.2.2. Timestepping algorithms. So far we have only considered the spatial

discretisation of our governing partial differential equation, hence giving the
'semi-discrete', ordinary, differential system equation (15)
C1" + K T = F
in which the time derivative remains. We now have to address ourselves to the
question of temporal discretisation. Many possible timestepping algorithms
are possible for this first order equation. However, we have decided on the use
of a straightforward two-level scheme which can vary between explicit (no
matrix inversion, timestep limited by stability) and implicit (matrix inversion,
unconditional stability) solution strategies. The scheme is frequently referred
to as the 'O-method' after the usual notation used for the free parameter. The
algorithm can be expressed in the form

c(Tn+I-T¢) + O K T n + I + ( 1 - O ) K T ~=F; 0~[0,1]

At
or, more concisely
[C-}-O AtK]T n+l= [C--}- ( 0 - - 1 ) AtK]Tn + AtF (19)
where the superscript denotes the time level, i.e. T n= T(n At). Note that
0 = 0 corresponds to an explicit scheme, requiring no matrix inversion provid-
ing C is diagonalised or 'lumped', and 0 = 1 corresponds to a fully implicit
scheme. Both these schemes have first order, 0(At), accuracy. The choice 0 = ½
corresponds to the well known Crank-Nicolson method, which is second order,
0(At2), accurate but can be suspect to spurious oscillations, or 'noise', for
larger timesteps - it is marginally stable. It is a two-level scheme since the
unknown values at t" + 1 are determined completely from the previous, known,
 Finite element simulation of solidification problems 67

values at t". Other values for 0 can of course be used, 0 = 2 has been
recommended, having been derived from finite element timestepping - where
the time domain is notionally discretised by 'shape' functions in the time
'direction'. Values of 0 >/½ lead to an unconditionally stable scheme - i.e.
errors in the solution do not grow, but for 0 < 1 the time step magnitude, At,
must be kept within the Courant limit - proportional to the element size and
diffusion term k, and inversely proportional to the capacitance term Oc or
(dH/dr).
The '0-method' has been used extensively in the examples to be presented
later, but other algorithms have been investigated during the development of
the heat conduction program [2,3]. These includes the three level scheme of
Lees [4], which can be expressed as

c +x =r
2 At 3
or

(20)
The three-level recurrence relation (20) solves for T at t n+2, given two
previous values T "+1, T n. It is unconditionally stable, but requires some form
of start-up e.g. by use of the above two level scheme, or by using T 1 = T o and
a very small initial timestep (the use of different timesteps At between
t" --* t "+1 and t "+1 --* t "+2 requires some modification to Eq. (20)). Although
stable, our experience has shown that solutions exhibited oscillatory, 'noisy'
behaviour - this was, however, considerably attenuated by redefining T ~ in
equation (20) to be the average [3]
(r + r o+ r
3
Another timestepping algorithm which has been utilised is that of a Laplace
modified scheme [5], used in conjunction with the finite element alternating-
direction method [2,6,7]. More details regarding the solution strategy of the
alternating-direction technique will be given below. However, it is pertinent to
include the Laplace modified algorithm in this section on timestepping proce-
dures. Applied to Eq. (12), a first order Laplace-modified algorithm, within a
Galerkin weighted residual form, gives

fa(r"+a - r " ) . N i l [ d £ + X A t £ v ' ( r "+1 - T " ) . V 'N[f[ da

.2 /" 32 t,~n+l 32N

+X2A, y, a-Ug-fiyt I - Tn) ~--U~y I a~l d £

= - A t £ dHk-/dT'VT n. v N I J I d £ (21)
 68 R.W. Lewis and P.M. Roberts

where the domain is transformed onto a union of rectangular (cuboids in

3-dimensions) regions. On the left hand side it can be seen that we are solving
for differences in temperature over the interval [t n, tn+X]. The term ~ - the
Laplace modifier - is an approximation to the diffusive term k. For constant
coefficients, and ignoring the higher order, O(At2), term, setting X = 0, ½k, k
reduces Eq. (20) to explicit, Crank-Nicolson, and implicit schemes respec-
tively. The perturbation term, aX2, is essential to the formulation of the
,.alternating-direction finite element scheme, and its form closely resembles that
' w h i c h can be derived from the analogous finite difference method. I J[
denotes the Jacobian of the mesh transformation onto the regular form, and
I f ] denotes its 'patch' approximation [6], discussed below.

2.2.3. Nonlinear problems. Although tacitly assumed, our notation so far has
ignored the possible dependencies of the coefficients k, p, c, H on the
independent variables space, time and temperatures. Thus Eq. (12) should be
written as
dH
V" (k(x, t, T) V T ) = -d-~(x, t, T). T. (22)

Variations in space and time do not present any special difficulties - apart
from discontinuities. More serious, however, is the dependence of the coeffi-
cients k, (dH/dT) on the actual solution, i.e. the temperature distribution.
We are faced with the problem of how to determine the coefficients, i.e. what
temperature values should be used. The matrix problem is given by (ignoring
spatial/temporal coefficient dependencies)
C(T)I"+ K ( T ) T = F(T). (23)
Put into the timestepping algorithm given by Eq. (19), we have
[ C ( r ) + O A t x ( r ) ] - r "+' = [ C ( T ) + ( O - 1) AtK(T)]. T" + Ate(T)
(24)

or, simply,
L (T) = P ( T ) T - G(T) = 0. (25)
For such nonlinear problems some form of iterative solution procedure is
required (linear problems can be considered as only requiring one iteration!).
Taking a truncated Taylor series expansion of Eq. (25) we obtain the well
known Newton-Raphson iteration scheme

{ T~=T"
T~%+~= T• + ATk k = O, 1, 2 . . .
Ark = - ( ~ L ( T ; ) ) - I L ( T ; )
(26)

where k denotes the number of iterations. The 'Jacobian' matrix (or 'tangen-
 Finite element simulation of solidification problems 69

tial' stiffness matrix) requires the calculation of derivatives. Only in special

circumstances can these derivatives be determined analytically - for 'real'
problems the only recourse is for approximate numerical derivatives, which
represents a substantial computing cost. Furthermore, the Jacobian matrix has
to be reconstructed at every iteration. Thus, although the Newton-Raphson
scheme gives quadratic convergence rates and is guaranteed to converge
(provided the initial 'guess' To~ is 'close enough' to the solution Tn+l), it is
very expensive to operate. Modified Newton-Raphson schemes, which iterate
directly using the first Jacobian constructed, only require one matrix construc-
tion per time step, but the convergence characteristics are frequently little
better than direct iteration which costs less.
In our experience the successive substitution procedure has proved ade-
quate, but Newton-Raphson type schemes may be required for highly non-lin-
ear problems to achieve acceptable convergence characteristics.
The scheme we have employed is based on direct, successive substitution
iteration. For reasonable timesteps, T n is often close enough to T n+l tO
ensure convergence by the direct approach. The algorithm adopted is of a
'predictor-corrector' form, viz (from Eq. (24)) [8].
Predictor:
[ C ( T " ) + AtOX(T~)] r : +1 = [ c ( r ~) + Z~t(0 - 1 ) K ( T " ) ] r ~ + AtF(T ~)
or

[C" + AtOK"]T~ +] = [C" + At(O- 1 ) K " ] T " + AtF" (27)

Corrector:
[C(TT) + AtOK(TT) ] Tp++11= [C( Tp~) + At( O - 1)K( Tp")] T" + 2xtF(Tp)
or

[C2 + AtOKp]T;++I1= [Cf + A t ( O - 1)K~] T" + AtF7 (28)

where p = O, 1, 2 . . . . (i.e. p is the iteration counter)
T7 = T U +]+(1-T)T"; r ~(0,1] (29)
r g + l = r 2 +1"
The weighting parameter y, which allows for variation in the choice of
temperature values for coefficient calculation, is chosen to take the same value
as 0 and in most of our practical tests both are set at ½ giving improved
accuracy and convergence.
With this predictor-corrector scheme, we control the timestep by the
number of iterations (corrections) required to reach convergence. Convergence
is deemed to have been reached when II TT++a1 - r p n + l II < e, for some p and a
user specified tolerance e. If the number of iterations required becomes
excessive then the timestep At is reduced; conversely, if very few iterations are
needed then At may be increased. This straightforward, direct iteration,
method has proven to be successful in all our tests, and is very economical
compared with Newton type iteration techniques.
70 R.W. Lewis and P.M. Roberts

3. Latent heat representation and phase change

The above heat conduction model must take into account the effect of latent
heat if it is to be used to simulate solidification. The latent heat L may be
used to determine the progression of the phase change interface S in the
formulation of the Stefan problem Eq. (5)-(9). As we have noted this
formulation requires some form of sophisticated front tracking algorithm.
However there is a simpler approach we can follow [9]. The latent heat effect
is approximated by a sharp increase in heat capacity within a narrow tempera-
ture range about the solidification point, Tc, as illustrated schematically in Fig.
3. The temperature range ~T from the solidus temperature, Ts, to the liquids
value, TL, can be extremely small so that for many substances the heat
capacity curve oc(T) exhibits a discontinuous Dirac function at To. Since we
are using numerical integration, which is only adequate for relatively smooth
curves, the direct calculation of Oc as a coefficient will lead to very poor
results.
However, the integral of the heat capacity, the physical quantity known as
the enthalpy H, as given by Eq. (10)

H= forOC dT

is a smooth function of temperature in the phase change zone. Thus, it is

preferable to use the enthalpy curve H(T), rather than the specific heat

~:~H /H
/
/
/
/
/

/i

f
/
f
/
f /
I I I >T
%%T.
Fig. 3. Variation of heat capacity and enthalpy near a phase change.
 Finite element simulation of solidification problems 71

capacity pc(T) curve. Hence we interpolate for the enthalpy in the same way
as for the temperature, i.e.
N

H = E Ni(x)Hi(t)" (30)
i=l

The enthalpy values are then utihsed in deriving the temperature derivative
(dH/dT) for the integrand of entries in the capacitance matrix, see Eq. (16).
Several averaging techniques have been used to determine this derivative term,
such as
d H _ = 1 _~I{~H
_~_ _ff~_~0T'
(31)
dr 3, v..i / v . . , ,

or

1/2
E (OH/Ox,) 2
dH i=1
, 1/2 "

However, a simple backward difference approximation [10]

(dH)'H'-H "-1
-'~ ~ r n- Tn_l (33)

has proven to be successful, giving excellent results for reasonable timestep

sizes. This last scheme is computationally quicker than the averaging ap-
proach, but does require the storage of previous enthalpy values at nodes. The
interpolation of H onto its nodal values saves a significant amount of
computing since H does not have to be evaluated from the H(T) curve at
every integration sampling point - the practice of interpolating for coefficients
using the same basis functions as for the unknown variables is sometimes
referred to as a 'group formulation'.
The use of the enthalpy method maintains a correct heat balance provided
the timestep is not so large as to involve great temperature changes at the
nodes, over one time increment, when 8T is very small. For many metals and
alloys 8T is relatively large so that the above technique works well, as will be
illustrated by the examples given below.

4. Novel solution techniques

In this section we detail some of the methods employed to increase the

computing efficiency of the overall solution process associated with the finite
element method of discretisation, as applied to solidification problems and
casting techniques in particular.
72 R. IV. Lewis and P.M. Roberts

4.1. Alternating-direction finite element method

The use of alternating-direction implicit (ADI) solution algorithms is well

known to finite difference modellers due to their efficiency both in solving
transient problems, and as a rapid iteration scheme for steady-state, elliptic,
problems. An alternating-direction (AD) finite element scheme has been
utilised by the present authors to compare its efficiency with standard solution
methods [2]. The AD method involves the solution of matrix systems con-
gruent to its finite difference namesake, but the derivation of these systems is
not analogous. The finite difference approach is via operator splitting, i.e.
solving or 'sweeping' along orthogonal directions, resulting in reduced band-
width systems - typically tridiagonal matrices for standard central difference
approximations.
However, the AD finite element algorithm relies on the tensor product form
of the shape functions N associated with the Lagrangian 'family' of elements,
i.e. the shape functions can be factored into directional parts, N~(~, ~/)=
~i(~)ffi(~/). Using this fact, and introducing some reasonable numerical
artifices, we can factor the overall system matrix into directional components,
e.g. K = KxKyK ~, where each of Kx, Ky, K~ has very narrow bandwidth
(tridiagonal for linear elements) for suitable nodal orderings.
We have already met the associated timestepping scheme, a Laplace mod-
ified algorithm in Eq. (21). The scheme given there could use standard finite
element methodology to give a perfectly acceptable discrete system. However,
to understand how it represents an AD formulation it is necessary to detail the
underlying 'philosophy' of the discretization process. Such detail can be found
in references [2,5,6] - only a brief sketch of the technique will be given here.
Firstly, the AD technique is based on regular, rectangular in two di-
mensions, meshes. However, unions of such meshes can be handled, and
coupled with the isoparametric transformation a great variety of mesh config-
urations can be undertaken. To facilitate description, a simple rectangle is
taken, as in Fig. 4. The general region f~g, Fig. 4(a), in 'global' coordinates
(x, y) is mapped onto the rectangular mesh f~, Fig. 4(b), in coordinates
(X, Y). Thereafter each element is mapped onto [ - 1 , 1] 2 in (4, 77) coordi-
nates for numerical integration (Gauss-Legendre). The reason for the ap-
pearance of the 'patch' approximation I JI to the actualJacobian I J I, for the
transformation of ag onto ~2, is that IJI is not of the tensor product from
required to factorise the integrals of (on the left hand side of) Eq. (21) into
directional components. Following [6], we adopt the following patch ap-
proximation

where I J l i is the mean value of I J I at node i, with the average being

determined over the patch of elements intersecting at node i. The system Eq,
(21) can be written in matrix notation as
K ( r °+1 - r " ) = AtO" (35)
 Finite element simulation of solidification problems 73

t2.

3 7"

-f2.

YT,
Fig. 4. Finite element mesh for alternating-direction formulation (a) general region ag, (b)
isoparametric base mesh fL

where
kn
- fu (dH/dT)n V T ~" V N [ J I d~

+ fa(lJl-lfl)( T"- T"-X) Ndf~ " (36)

The term involving (I J I - I f [) is an error-correction term which can be

omitted for rectangular meshes, or for meshes in which the elements are not
badly distorted. Now, because of the form of I J [ (used throughout in the left
hand side of equations (21), (35)) the system matrix can be factored as
= D 1 / 2 K D 1/2 (37)
where D = diag( I J I i). Furthermore the tensor product form of the shape
functions, and the inclusion of the 0(At 2) perturbation term enables the
following factorisation, with N,.( X, Y) = ~i ( X ) q~i(Y), in two dimensions,

KiJ= fn,j NiNj+X At vN~'v-NJ+~k2 At2oxoY OXOY] d X d Y

= fo (d~il~jq- ~k Ateo~eO~) dx.f, (,,% + X At~;+5) OY (38)

, O~i , O~i
where Oi = ~ and ~ = -~-~. You will notice that the perturbation term is
necessary to 'complete the square' in the integrand.
Thus,
K=Kx.Ky (39)
74 R.W. Lewis and P.M. Roberts

where Kx, Ky have very narrow band structure - tridiagonal for four-noded
bilinear elements, or pentadiagonal for nine-noded biquadratic elements and
so on. The system given by Eq. (35) may be solved as follows. Firstly O" is
determined at each time step - this involves standard finite element integrals.
Secondly, we solve for v" from K x v ' = D -1/2 AtO n, and then for w" from
Kyw" = v'. Finally we have T "+1 = T" + O-1/Zwn. The solution stages can be
performed rapidly using well known algorithms, and furthermore, K~, Ky are
block structured making for efficient vectorisation. K~, Ky are also time
independent and so need only be constructed once.
The above AD algorithm has been successfully used to solve solidification
problems on relatively simple geometries; but further program development
will enable the potential savings to be realised for quite complex configura-
tions.

4.2. Mixed impficit / explicit solution

Explicit time stepping requires no matrix inversion, and is thus very quick to
operate, but is limited by timesteps constrained by stability criteria. Implicit
schemes, on the other hand, are unconditionally stable and can take quite
large timesteps (but is still first order, 0(At), accurate); however, they require
matrix inversion/factorisation which can be time consuming for problems, as
in three dimensions, which involve large bandwidths. The mixed implicit-ex-
plicit technique, pioneered by Hughes and Liu [11,12], seeks to combine the
advantages of each scheme whilst minimising their deficiencies. Hence, explicit
solution can be carried out cheaply in areas of the domain where the solution
is changing slowly, whereas, to maintain stability, implicit treatment of ther-
mally 'active' regions is required [13].
Taking the first order system given in Eq. (15)
Ci" + K T = F
we can classify the elements of the mesh as being explicit (0 = 0, C 'lumped'
in Eq. (19)) or implicit (0.5 ~ 0 ~< 1.0 in (19)), giving
K = K I + K E]
/

C = (7t + Ce ~. (40)
r=r,+r~ ]
Setting V= T, and /~.+1 as an intermediate nodal temperature vector, we can
express Eq. (16) in the form, for a single element,
C V "+1 + K , T ~+1 + KEir'+ 1 = F "+1 (41)

where
ir "+a = T" + AtV'(1 - 0) (42)
 Finite element simulation o f solidification p r o b l e m s 75

and
Tn+l = ~ n + l + Atvn+lo. (43)
Substitution of Eq. (43) into (41) results, after some manipulation, in the form
[ C + AtOKI IT n+ l = [AgO(lN n + l - KE ~n+ l ) + C~/~n+l] . (44)
For an explicit element, i.e. 0 = 0, Eq. (42) becomes
T n+l = T n + A t V n
but from Eq. (16)
v.=c-l[v.-xr.]
and hence
C T "+1 = C T " + A t [ F " - KT"].

This equation is exactly the same given by the predictor of Eq. (27), again with
0 = 0, for Tg + 1. Setting 0 = 0 in Eq. (44) gives
C T n+ l = C # n+ l = C T ~ + t [ F ~ - K T ~ ]

where now C, K, F are determined using some intermediate 'new' tempera-

ture T ~ = 7T "+1 + (1 - ,{)T n. Once more, this is identical to the form of the
corrector Eq. (28). Similar arguments show the algorithm of Eqs. (41)-(44) is
identical to the predictor-corrector form for the implicit case 0.5 ~< 0 ~< 1.0.
Thus, an implicit-explicit algorithm can be implemented by specifying 0 for
each element [13], rather than globally, within the existing predictor-corrector
solution algorithm. However, to take full advantage of the presence of explicit
elements, which only contribute to the diagonal entries of the system (C is
diagonalised for explicit elements), an active column profile solver is used.
Essentially this performs Gaussian elimination only for those matrix entries
which have off-diagonal terms in the equation. Thus, with many explicit
elements, significant reductions in processing time can be achieved. One such
case could be the use of explicit elements to represent the area defined by the
sand mould, since it is not the heat transfer in the mould, but that in the cast
which is of primary concern.

4.3. Quadratic conductivity approximation

The governing Eq. (1) is nonlinear, and, as we have noted, this necessitates
some form of iterative solution procedure. This requires the re-formulation of
the system matrices C ( T ) and K ( T ) for each iteration due to their tempera-
ture dependence. This is particularly expensive for finite elements since this
would involve, adopting a simplistic approach, costly re-integration of all the
terms of Eqs. (16)-(18).
However, we can partly alleviate this computational burden by making
assumptions about the form of the temperature dependent coefficients -
76 R.W. Lewis and P.M. Roberts

namely the conductivity k ( T ) and the heat capacity ( d H / d T ) . We have

already discussed the interpolation of the enthalpy H using the finite element
shape functions and nodal values Hi, as in Eq. (30). If we denote the heat
capacity ( d H / d T ) by R, and assume a finite element discretisation for R, i.e.
N
R ( x , t) -~ ~, N i ( x ) R i. (45)
i=1

Then the capacitance matrix, after lumping by summation of rows, becomes

N
Cii = E Rj[NiNjd a
j=l ~'fa

or

c=I(c0R) (46)

where

Co,./= f N / N / d ~ (47)

and R is the vector of heat capacities (derived from enthalpy values H ) . The
matrix Co can be formed once at the start of the problem by numerical
integration; subsequent calculations of the diagonalised capacitance matrix
can be obtained much more rapidly by matrix multiplication as in Eq. (46).
We now apply similar 'group' formulation ideas to the conductance matrix
K ( T ) . Samonds et al. [14] recently put forward a 'quadratic conductivity
approximation' whereby the finite element integrals involving k can be
evaluated initially, with subsequent evaluations of K by matrix multiplication
involving the nodal temperatures.
Hence, we approximate by a quadratic function, i.e.

k ( T ) = a + bT + eT 2 (48)
where a, b, c are material constants. If we ignore the boundary condition
contribution, for clarity, then the integral term which gives us the product
K ( T ) . T in Eq. (15) over a single element becomes

lke = .fern" k( r ) v r d a e (49)

which on substitution of the quadratic form Eq. (48) yields, after some use of
vector operator identities,
b
lke=afuvN" VT d~e+ -~fuevN" v(T2) d~e + ~c £evN" v(T3) d~e.
(50)
 Finite element simulation of solidification problems 77

The next simplifying assumption is to interpolate the square and cube of the
temperature onto the same finite element basis as T, i.e.
N N
T2 = E N~T~2, T3 = E N~T~3 (51)
i=1 i=1

which gives us

Ik= aKoT+ lbKoT2 + ½cKoT3 = K 0 ( aT+ ~b T 2 + -~T

c
3) (52)

where

/~0 = £vN~- vNj da. (53)

The conductance matrix K o need only be evaluated, by numerical integra-
tion, once at the initialisation stage of computation.
If we apply this procedure to explicit elements in the mixed implicit-explicit
formulation then the K. T contribution to the right hand side vector can be
rapidly established by the matrix multiplications of Eq. (52) for all iterations.
A similar procedure could be utilised for the implicit elements, using a K 0
on the left hand side for K, subtracting AtO(½bKoT~2+ ½cKoTn3) from the
right hand side of Eqs. (27), (28). However, since the consistent capacitance
matrix has to be evaluated, and it does not cost much more to calculate the
conductivity matrix at the same time, it is doubtful whether the reduction in
matrix assembly costs will justify the loss of accuracy for implicit elements.
The quadratic conductivity approximation given above has demonstrated
its effectiveness in reducing matrix generation times, particularly for explicit
timestepping. One important consideration is the need for piecewise quadratic
approximations should the conductivity of the m d t vary sharply on solidifica-
tion, except for dements which contain more than one phase. However, the
approximation given here has proved adequate in our test problems, resulting
in much faster runs, especially for three dimensional meshes.
The quadratic conductivity technique shown here is thus justified for
materials where k(t) can be reasonably approximated by a quadratic curve.
Notable exceptions include cases where k(T) exhibits near discontinuous
behaviour.

4.4. Mould-metal interface models - coincident node technique

A critical problem in the simulation of casting methods is the treatment of the

heat transfer between metal and mould at the relevant interface. Samonds et
al. [15] have investigated various schemes for handling the mould-metal
interface, including the use of thin elements at the interface, and the use of a
coincident node technique, to be described in this section. The coincident node
approach is, numerically, virtually equivalent to the use of thin elements, but
78 R.W. Lewis and P.M. Roberts

@ ® @ e

i 3 S ; ss }
@) (~)
Fig. 5. Linear dements with common interface (a) standard connections, (b) coincident interface
nodes.

offers greater flexibility in mesh design, and also significant savings in

computer time.
Consider two bilinear elements which are adjacent as shown in Fig. 5(a).
The nodal connecfivifies for these elements can be listed as
Element Nodes
1 1 3 4 2
2 3 5 6 4
Upon assembly of the system matrices, with this element topology, a condition
of perfect conduction will exist between elements 1 and 2 because of the
common interface defined by nodes 3 and 4.
If the elements are considered as disjoint, e.g. by the introduction of a
spatially coincident node at each interface node, as shown in Figure 5(b), the
nodal connections for the elements become
Element Nodes
1 1 3 4 2
2 5 7 8 6
In this case standard matrix assembly will assume a condition of perfect
insulation between these elements since they have no nodes in common.
Neither perfect conduction nor insulation is adequate to describe the heat
transfer between metal and mould. Common modelling practice is to assume a
convective type heat transfer, as given by the expression
q, = h (TM~TAL -- TMOULD) (54)
where q1 is the heat flux across the interface, and h is a controlling coefficient.
Experimental values for h and its variation with time, or temperature, as
solidification progresses are available. We now seek to incorporate such a heat
transfer procedure within the coincident node formulation. This is quite easily
effected by making the following addition to each element conductivity matrix
(i.e. for each element abutting the interface) [8],

Ke=Ke+faN,(Nj-½N,.)dFz
rz
i, j = l , 2 . . . . N, (55)
 Finite element simulation of solidification problems 79

where N~ is the number of nodes on the interface boundary Ix; m is the

number of the node coincident with node j, and a is the interface heat
transfer coefficient (c.f. h in Eq. (54)). The integrated contribution of Eq. (55)
must be applied to elements on both sides of the interface, and the N,Nm cross
term has a factor of ½since it will appear twice during the integration.
Thus, given accurate measurement of the transfer parameter a we can
expect good results from this coincident node approach, which is flexible
during mesh design and has demonstrated its efficiency through our numerical
tests.

4. 5. Infinite element heat loss calculations

When dealing with external boundaries which are distant from the main area
of interest it is impractical to enmesh the external region with finite elements.
The associated computational overhead would be great even for a truncated
external mesh. A viable alternative is to discretise the entire external region as
if it were of infinite dimension, using 'infinite elements' [16,17] which map the
infinite regions onto the usual prototype domain for numerical integration.
Although perhaps this may be of limited use in casting simulation, it has
important practical applications in many areas, including heat transfer, and
the infinite element technique has been used successfully by the present
authors in the context of heat losses from geothermal or petroleum reservoirs
to the surrounding rock strata [16].
Zienkiewicz et al. [17] put forward an infinite element method based on a
mapping of a semi-infinite region onto [ - 1 , 1 ] " - the usual finite element
domain for Gauss-Legendre integration. The one-dimensional mapping is
schematically illustrated in Fig. 6(a), with a two-dimensional infinite 'strip'
element shown in Fig. 6(b), and its use in conjunction with a finite element
mesh given in Fig. 6(c).
Referring to Fig. 6(a), the one dimensional mapping, which is at the heart
of the mapped infinite element method can be expressed as

x= (1- x0+ 1 + (1-~ x2 (56)

or

x=No()xo+N2()x2.
This map transforms the region [xl, x3) (x 3 at infinity) onto the interval
- 1 ~<~ ~< 1 with X 1 = (X 0 -t- X2)/2. Note that N0(~ ) + N2(~) -- 1, and so the
map is not affected by a shift of origin.
This transformation can be easily coupled with existing shape function
mappings (e.g. isoparametric) to derive 'strip' elements of the form illustrated
by Fig. 6(b). It should be noted that all that is required to utilise these infinite
elements into existing finite element programs is some minor modifications to
80 R.W. Lewis and P.M. Roberts

• • •

~3 (20
/
NAP / / /
/ /
/ ./
-I 0 t

(a.?

__ _ _ _ x-~. (~:1)

>~

~ ~l I

rL / i--.

Fig. 6. Infinite elements: (a) one dimensional mapping, (b) strip element, (c) coupling with finite
element mesh.

the routine which calculates the Jacobian matrix determinant for numerical
integration.

5. Numerical examples

In this section we present some of the test problems which have been
simulated using the finite element heat conduction program with its various
 Finite element simulation of solidification problems 81

enhancements discussed above. All the problems involve solidification, with

most representing real casting projects.
The first example, however, deals with a corner freezing problem and is
intended to demonstrate the effectiveness of the alternating-direction solution
algorithm. The remaining examples will use the predictor2corrector timestep-
ping procedure.
The mesh which discretises the square region is illustrated in Fig. 7, and
comprises 9-noded biquadratic Lagrangian elements - giving pentadiagonal
matrices in the A D formulation. The region is initially set at the solidus
temperature - 0 . 1 5 o C, with material properties pc = 1.0, k = 1.08, and latent
heat L = 70.26. The phase change temperature interval 8T (see Fig. 3) is taken
as 2 °, and a constant timestep At = 0.05 was employed. At t = 0 the faces
x = 0, y = 0 are instantaneously set at - 45 ° C to initiate freezing. The outer
faces x = 4, y = 4 are assumed to be perfectly insulated. Figure 8 compares
the temperature history profiles at x = y = 1 for the A D algorithm against
both the predictor-corrector (Crank-Nicotson), and Lees timestepping schemes.
The curves show very good agreement and temperature contours from all these
methods are almost indistinguishable, showing the A D formulation to be as
accurate as standard solution methods. However, the C P U time required per
timestep by the techniques was

Alternating-direction algorithm = 4.02 s

Lees' algorithm = 10.22 s
Crank-Nicolson algorithm = 10.10 s.

Thus the A D method is 2.5 times faster than the usual timestepping methods,

Fig. 7. Comer freezingproblem - finite element mesh

82 R.W. Lewis and P.M. Roberts

LEES ALGORITHM
ALTERNATING DIRECTION - - - - - -
CRANK-NICOLSON

-5 "1
t~
-io
l--l

-15

c~
uJ
-20

-25
.,~>x~...
-30

-35
' I ' I , I ' I , I , I ' I '

t, B 12 16 20 24 28
TIME [SECONDS) x 10-I
Fig. 8. Temperature h i s t o r i e s a t p o i n t x = y = 1.

this figure should rise to about 10 for vector processor machines. The above
times are for an ICL 2966 computer, a scalar machine with virtual memory
capability.
The AD algorithm has also been successfully utilised for problems involv-
ing non-rectangular domains; a wedge shaped domain with similar physical
properties and problem definition to the above example has been successfully
simulated [2].
The AD code, however, requires further development to model general
mesh configurations and to cater for three dimensional meshes. An existing
three-dimensional heat conduction program, incorporating the predictor-cor-
rector timestepping scheme, has been amended to investigate the various
procedures for improved casting simulation discussed in the previous section.
To demonstrate the effectiveness of the implicit-explicit method an experi-
mental casting of a tapered slab of aluminum bronze in a resin bonded silica
sand mould was simulated [13]. Table 1 gives the thermal properties of the
metal, where the latent heat figure is a weighted average of the constituent
pure metals. Table 2 gives the properties of the sand mould - two sets of
conductivity curves were investigated. The cross-sectional finite element mesh,
shown in Fig. 9, consists of 56, 8-noded quadratic, isoparametric elements,
with the sand elements shaded. The initial temperature of the molten metal
was 1156 o C, with the initial mould-metal interface set at the liquidus temper-
ature, 1080 ° C. No thermal resistance, i.e. perfect conduction, was assumed at
the interface, and the latent heat was taken to evolve linearly over the
solidification range. The external boundary, of the sand elements, was fixed at
room temperature, 20°C, with the initial temperature being the same
 Finite element simulation of solidification problems 83

Table 1. Material properties - aluminium bronze.

Composition (%): Copper 80.5

Aluminium 9.5
Iron 4.5
Nickel 5.5
Solidification Range : 1,050 * - 1,080 o C (ST = 30 o )
Latent Heat, L : 2.37 X 105 J / k g
Specific Heat, c : 452.2 J / k g ° C
Density, p : 7,600 k g / m 3
Conductivity, k ( T ) :
Temperature ( ° C) k ( J / s m ° C)
200 54.42
400 46.05
600 43.95
800 41.86
1000 41.86
1200 41.86

throughout the sand mould. For the riser, the top boundary was fixed at the
initial metal temperature, 1156 ° , with the sides assumed to be perfectly
insulated.
The initial timestep was taken as 1.0 s, but this was altered by the
predictor-corrector controlling algorithm, with an iteration tolerance of 0.5 o C.
The first run assumed 0 = 7 = ½ (Crank-Nicolson, implicit) for all elements,
and Fig. 10 depicts the movement of the solidus temperature contour (1050 o C)
at 90s intervals, for the high sand conductivity case. The shape and progres-
sion of the solidification front agrees well with physical expectations, even for
the coarse mesh used. A similar numerical experiment on a slab which is not
tapered, as in Fig. 11, shows that a large proportion of the slab solidifies
almost simultaneously, unlike the even progress of the solidification for the
tapered case. This suggests feeding problems in this region yielding shrinkage
porosity, an effect which occurs in practice.

Table 2. Material properties - sand.

Density, O : 1,444.7 k g / m 3
Conductivity, k ( T ) : (High) 0.6606 - 2.084 × 1 0 - 4 T + 7.741 × 10- 7T2
( J / s i n ° C) (Low) 0.2828-1.681 X 1 0 - 4 T + 1.352 × 10-7T 2
Specific heat, c(T)
Temperature ( ° C) c(J//kg o C)
200 975.7
400 1092.9
600 1151.5
800 1159.9
1000 1176.7
84 R.W. Lewis and P.M. Roberts

Fig. 9. Tapered slab casting - meshwith sand elementsshaded.

The problem was run again, this time making all the sand elements explicit
(0 = 0), i.e. 64% of the mesh is solved by explicit timestepping. When using the
active column profile solver this resulted in a reduction of 40% in the running
time, with little qualitative difference from the results already shown. Thus the
efficiency of the implicit-explicit scheme, coupled with the profile solver, can
provide substantial savings in casting simulation.
The tapered slab problem was solved for once again, but this time employ-
ing the quadratic conductivity approximation to achieve even further processing
time reduction. Figure 12 shows the temperature history profiles, at selected
points along the slab, with consistent capacitance matrices (lumped for explicit
elements) and a piecewise linear determination of k from the set of values
given in Tables 1, 2 at each integration point. Figure 13 shows the same
profiles, but using the quadratic conductivity approximation and the diagona-
lised capacitance technique [14]. There is very close agreement between the
two sets of curves - a maximum difference of only 2%. However the assembly
time of the quadratic approximation case was 46% that of the standard
method, and, since the assembly time makes up 85% of the total solution cost,
 Finite element simulation of solidification problems 85

"7

I
I
.q
I
I

7
7

\
\

/
Fig. 10. Tapered slab casting - solidification front history.

thus giving an overall saving of 39% in total processing time. This is for the
predictor-corrector algorithm running in mixed implicit-explicit mode, with
explicit elements in the sand mould.
To demonstrate the three dimensional simulation capability, investigations
were carried out [15] into a production gravity die casting of an aluminium-
silicon-copper alloy (BS.1490 : LM4) with material properties given in Table 3.

Fig. 11. Untapered slab - solidification front history.

86 R.W. Lewis and P.M. Roberts

BNF sand cast tapered stab

1200. numerical results
1150
1100
105(

u
100(
i!!
V
-- 158
~a 95o:
"~ 900 -116

E 850
-74
800

750 -3z

700
I ;0 ~O0 3'oo' z~O0500
' ;O0 700
. . BOO
. . 900
. . 10001100 12oo
time (St
Fig. 12. Temperature history curves for tapered slab - standard matrix assembly.

Half of the die assembly, with its finite element mesh, is shown in Fig. 14, with
the cast shaded - the internal geometry/mesh of the cast is shown in Fig. 15.
The die composed of Grade 17 grey iron, the material properties for which can
be found in Table 4.
Thermocouples were placed at various positions both in the die and in the
cavity. The resulting temperature histories at these points are given in Fig. 16,
for a melt which was poured at approximately 760 o C.
The finite element mesh consisted of 374 linear brick (8-noded) elements
and 645 nodes. The coincident node method was employed at the interface,

BNF sand cast tapered stab

1200 numerical results
1150
1100

1050

10001
950 : IS8 158
~a

ro 900 116

B50

BOO : ?4

750 2

700 , , , , , , , , , ,
100 200 3'00 400 500 600 700 800 900 1000 1100 1200
time (S)
Fig. 13. Temperature history curves for tapered slab - quadratic conductivity function.
 Finite element simulation of solidification problems 87

Table 3. Material properties - alloy BS.1490 : LM4.

Composition (%):Aluminium 89.41
Silicon 5.70
Copper 3.00
Iron 0.70
Manganese 0.37
Zinc 0.32
Titanium 0.20
Nickel 0.16
Magnesium 0.06
Lead 0.05
Tin 0.03
Solidification Range : 525 ° - 625 ° C (ST = 100 o )
Latent Heat, L : 92.3 cal/g
Density, 0 : 2.67 g / c m 3
Conductivity, k(T), Specific Heat, c(T):
Temperature ( o C) k(cal/sec-cm- ° C) c ( c a l / g - ° C)
225 0.393 0.234
325 0.400 0.247
425 0.390 0.268
525 0.390 0.301
625 0.442 0.301
725 0.442 0.301

w i t h a h e a t t r a n s f e r c o e f f i c i e n t a = 5.0 × 10 - 2 c a l / s e c - c m 2- ° C, r e d u c e d a f t e r
19 s e c o n d s b y 75% t o m o d e l t h e f o r m a t i o n o f a n air g a p . A r o u n d t h e r i s e r w e
t o o k a = 3.0 × 10 - 2 c a l / s e c - c m 2- ° C, r e d u c e d a f t e r 25 s e c o n d s b y 60%. T h e s e
v a l u e s w e r e b a s e d u p o n p r e v i o u s e x p e r i m e n t a l a n d c o m p u t a t i o n a l w o r k [15].

...._¢----
J ~
f J

i
J
J

Fig. 14. Gravity die casting assembly - 3D finite element mesh.

88 R . W . Lewis and P.M. Roberts

f
f

Fig. 15. Finite element mesh of cast plus riser.

T h e b o t t o m o f t h e die a n d t h e p l a n e o f s y m m e t r y w e r e a s s u m e d to b e
p e r f e c t l y i n s u l a t e d , w h e r e a s free c o n v e c t i o n was a s s u m e d 'at t h e r e m a i n i n g
e x t e r n a l faces. T h e t e m p e r a t u r e h i s t o r i e s c a l c u l a t e d at t h e t h e r m o c o u p l e
p o s i t i o n s a r e s h o w n in Fig. 17. T h e s e results w e r e o b t a i n e d u s i n g all i m p l i c i t

Table 4. Material properties - grade 17 grey iron.

Composition (%): Iron 93.55
Carbon 3.18
Silicon 2.03
Manganese 0.56
Phosphorus 0.59
Sulphur 0.09
Density, p: 7.23 g/cm 3
Quadratic conductivity approximation:
k(T) -- [1,622 × 10-1 _ 2.477 X 10-4T + 2.002 X 10- 7T2] cal/sec-cm- ° C
Conductivity, k(T), Specific Heat, c(T):
Temperature ( ° C) K(cal/sec-cm- ° C) c(c.al/g- o C)
225 0.118 0.13
325 0.101 0.14
425 0.092 0.15
525 0.089 0.16
625 0.084 0.18
675 0.082 0.19
 Finite element simulation of solidification problems 89

700 . B.R. brake vatve, 3-13

experimenfat resuLfs

650.

600 _

I - -
550

oJ
50O . . . . . . . . 18
ro
L
Ig

E 450 18 22 * 20

L~O0 22

350 , , , i i , i i i

0 20 ~0 60 80 100 120
finle IS}

Fig. 16. Experimental temperature histories for die casting.

elements and standard conductivity matrix assembly. Comparing Figs. 15, 17

we note a good general agreement between the two sets of curves, except a
numerical prediction of re-heating in the projecting wing of the casting after
the air gap has been formed. Using an implicit-explicit formulation with all
the die elements explicit lead to instabifity. Instead, a single layer 'coating' of
implicit die elements around the cast was employed - making 64% of the
elements explicit. The results of this mixed form, with the quadratic conductiv-
ity approximation, are given in Fig. 18 - note that the difference from the

700 B.R. brake valve, 3 D

I numericalresulfs

~
650 1

6oo

242 255

l+O0- 242

350 , . . . . . , ,
0 20 40 60 80 100 120
lime (S)

Fig. 1 7. Numerical temperature histories for die casting - standard matrix assembly.
90 R.W. Lewis and P.M. Roberts

700 B.R. brake valve, 3-D

numerical resulfs

650

600

QI
B 500 a61

- 31s
~- /+50 - 46~. ~.2 .'z

I+00 _ 2~2

35O
20 /+0 60 80 100 120
time (S)

Fig. 18. Numerical temperature histories for die casting - explicit time stepping and quadratic
conductivity function.

curves of Fig. 17 is negligible. However the assembly time was reduced by

40%, and the mixed timestepping algorithm yielded a reduction of 63%
compared with the fully implicit run.
The use of the mixed implicit-explicit formulation, coupled with a quadratic
conductivity approximation, thus effectively halved the CPU time required to
model the gravity die cast simulation. Figure 19 depicts the movement of the
575 °C contour, which represents a solidification fraction of 80% - taken as
the point at which the liquid metal feeding has ceased. It is apparent that the
simulation predicts an underfed region in the casting.
Finally, the close agreement between experimental and numerical results
demonstrates the effectiveness of the coincident node technique for modelling
the mould-metal heat transfer at the interface. Further work needs to be
carried out to model the air gap formation, in particular by the variation of
as above and in other test problems investigated [8].

Fig. 19. ISO-solidification contour history - 80% at 3-second intervals.

 Finite element simulation of solidification problems 91

6. Summary

The development of a finite element program for modelling heat conduction

with phase change has been presented. The program has been used success-
fully to model a wide range of solidification problems including the casting
processes illustrated by the examples given. Efforts to improve the efficiency
of the finite element algorithm have succeeded in achieving substantial reduc-
tions in computer costs. Specifically the mixed implicit-explicit timestepping
technique has been incorporated to speed up solution times, and the use of a
quadratic conductivity approximation, which has significantly reduced the
matrix assembly time associated with finite elements. The reduction of the
computing overhead thus achieved makes for much cheaper numerical simula-
tion of casting methods, particularly for three-dimensional discretisation,
which is nearly always necessary due to the complex configurations involved.
A flexible technique for modelling the critical heat transfer processes
associated with the mould-metal interface has been developed - namely the
coincident node method. It has already established its viability, and its
adaptability, via the variation of the heat transfer coefficient, make it a useful
artifice in casting simulation. Further work needs to be carried out, however,
to properly take into account the nature of air gap formation and how it
affects the heat transfer coefficient.
It is hoped that further research and development of our simulation
capability in this area, will make computer modelling of casting processes a
practical foundry tool with which to optimise process and design features
without recourse to much more expensive experimental analysis. Thus numeri-
cal simulation is envisaged as playing a prominent role in making industrial
forming processes more economical and hence more effectual in a highly
competitive market.
Our future objectives for our finite element model must include the simula-
tion of convective heat transfer, where appropriate, and the coupling of the
above heat transfer model with a comprehensive stress analysis package
including crack propagation. A shorter term aim, however, is a comparison of
cooling zones in the cast incorporating a model for analysing crystalline
growth during solidification.
Once these objectives have been achieved a powerful simulation program
will be available which can predict most, if not all, of the difficulties presently
met in casting processes.

Acknowledgements

The authors wish to acknowledge the substantial contribution to the work

presented in this paper made by Dr M. Samonds, Dr K. Morgan and Mr R.
Symberlist.
92 R.W. Lewis and P.M. Roberts

References

1. O.C. Zienkiewicz and K. Morgan: Finite Elements and Approximation, 1st edn., p. 328, John
Wiley & Sons Ltd., New York (1983).
2. R.W. Lewis, K. Morgan and P.M. Roberts: Application of an alternating-direction finite
element method to heat transfer problems involving a change of phase, Numerical Heat
Transfer, Vol. 7 (1984).
3. R.W. Lewis, K. Morgan and P.M. Roberts: Determination of thermal stresses in solidification
problems. In: J.F.T. Pittman, O.C. Zienkiewicz, R.D. Wood and J.M. Alexander (eds)
Numerical Analysis of Forming Processes, Chap. 15, John Wiley & Sons Ltd., (1984).
4. M. Lees: A linear three-level difference scheme for quasi-linear parabolic equations, Math.
Comp. 20 (1966) 516-622.
5. J. Douglas Jr. and T. Dupont: Alternating-direction Galerkin methods on rectangles. In: B.
Hubbard (ed.), Proc. Numerical Solution of Partial Differential Equations, Vol. 2, pp. 133-214,
Academic Press, New York (1977).
6. L.J. Hayes: Implementation of finite element alternating-direction methods on non-rectangu-
lar regions, lnt. J. Numerical Methods in Engineering 16 (1980) 35-49.
7. P.M. Roberts: The development of alternating-direction finite element methods for enhanced
oil recovery simulation, Ph.D. Thesis, University College Swansea, University of Wales,
Swansea (1984).
8. M. Samonds: Finite element simulation of solidification in sand mould and gravity die
castings, Ph.D. Thesis, University College Swansea, University of Wales, Swansea (1985).
9. G. Comini, S. Del Guidice, R.W. Lewis and O.C. Zienkiewicz: Finite element solution of
non-linear heat conduction problems with special reference to phase change, Int. J. Numerical
Methods in Engineering 8 (1974) 613-624.
10. K. Morgan, R.W. Lewis and O.C. Zienldewicz: An improved algorithm for heat conduction
problems with phase change, lnt. J. Numerical Methods in Engineering 12, (1978) 1191-1195.
11. T.J.R. Hughes and W.K. Liu: Implicit-explicit finite elements in transient analysis: implemen-
tation and numerical examples, ASME. J. Appl. Mech. 45 (1978) 375-378.
12. W.K. Lin: Development of mixed time partition procedures for thermal analysis of structures,
Int. J. Numerical Methods in Engineering, 19, (1983) 125-140.
13. M. Samonds, K. Morgan and R.W. Lewis: Finite element modelling of solidification in sand
castings employing an implicit-explicit algorithm. Applied Mathematical Modelling, 9 (1985)
170-174.
14. M. Samonds, R.W. Lewis, K. Morgan and R. Symberlist: Efficient three-dimensional finite
element metal casting simulation employing a quadratic conductivity approximation, Proc.
Fourth lnt. Conf. on Numerical Methods in Thermal Problems, Part 1, pp. 64-77, Pineridge
Press, Swansea (1985).
15. M. Samonds, R.W. Lewis, K. Morgan and R. Symberlist: Finite element modelling of the
mould-metal interface in casting simulation with coincident nodes on thin elements. In: R.W.
Lewis, K. Morgan, J.A. Johnson, W.R. Smith (eds) Computational Techniques in Heat
Transfer, Vol. 1, Pineridge Press, Swansea (1985).
16. R.W. Lewis, K. Morgan and P.M. Roberts: Infinite element modelling of heat losses during
thermal recovery processes, paper SPE 13515, 8th SPE Symp. Reservoir Simulation, Dallas,
Texas (1985).
17. O.C. Zienkiewicz, C. Emson and P. Bettess: A novel boundary infinite element, Int. J.
Numerical Methods in Engineering, Vol. 19, pp. 393-404 (1983).