You are on page 1of 14

Available online at www.sciencedirect.

com

Acta Materialia 57 (2009) 5862–5875


www.elsevier.com/locate/actamat

Phase-field modelling of as-cast microstructure evolution


in nickel-based superalloys
N. Warnken a,*, D. Ma b, A. Drevermann c, R.C. Reed a, S.G. Fries d,e, I. Steinbach e
a
University of Birmingham, Department of Metallurgy and Materials, Edgbaston, Birmingham B15 2TT, UK
b
Foundry Institute of the RWTH-Aachen, Intzestr. 5, 52072 Aachen, Germany
c
ACCESS e.V., Intzestr. 5, 52072 Aachen, Germany
d
SGF Consultancy, 52064 Aachen, Germany
e
ICAMS, Ruhr University Bochum, Stiepeler Strasse 129, D-44780 Bochum, Germany

Received 4 April 2009; received in revised form 4 August 2009; accepted 4 August 2009
Available online 10 September 2009

Abstract

A modelling approach is presented for the prediction of microstructure evolution during directional solidification of nickel-based
superalloys. A phase-field model is coupled to CALPHAD thermodynamic and kinetic (diffusion) databases, so that a multicomponent
alloy representative of those used in industrial practice can be handled. Dendritic growth and the formation of interdendritic phases in an
isothermal (2-D) cross-section are simulated for a range of solidification parameters. The sensitivity of the model to changes in the solid-
ification input parameters is investigated. It is demonstrated that the predicted patterns of microsegregation obtained from the simula-
tions compare well to the experimental ones; moreover, an experimentally observed change in the solidification sequence is correctly
predicted. The extension of the model to 3-D simulations is demonstrated. Simulations of the homogenization of the as-cast structure
during heat treatment are presented.
Ó 2009 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.

Keywords: Multicomponent-solidification; Directional solidification; Homogenisation/solutionisation; Phase field models; Nickel alloys

1. Introduction the differential equations which need to be solved are


strongly coupled, so that their non-linearity places signifi-
Prediction of phase transformations in metals and alloys cant demands on computing resources. Theory for trans-
requires models for the underlying thermodynamics and formations in multiphase, multicomponent systems has
kinetics, and good estimates of critical parameters such also been lacking.
as driving forces, diffusion coefficients and surface tensions But this situation is altering. The theory of transforma-
[1,2]. Traditionally, a lack of accurate materials data has tions in metals and alloys has advanced at pace in recent
hampered this effort. The calculation of microstructure years, and methods are available which are capable of
evolution in technical alloys—which usually possess more treating multicomponent, multiphase reactions in a realis-
than three or four alloying additions—have been possible tic way, with the underlying theory rigorously accounted
only with limited accuracy due to the practical necessity for [5–12]. Secondly, concerted effort over many years—
to reduce the problem to a binary or ternary system, con- for example by the CALPHAD (Calculation of Phase Dia-
sidering only those elements that are assumed to be of gram) community—means that for most systems reason-
importance [3,4]. An additional difficulty has been that ably accurate thermodynamic data are now available and
coded into thermodynamic databases [13,14]. Related
*
Corresponding author. Tel.: +44 121 414 3527. efforts have built databases of diffusion coefficients and
E-mail address: n.warnken@bham.ac.uk (N. Warnken). mobilities for the relevant phases [15–17]. Furthermore,

1359-6454/$36.00 Ó 2009 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.actamat.2009.08.013
N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875 5863
Z
there is now widespread availability of computer hardware,
F ¼ ½f intf þ f chem  dx3 ð1Þ
storage and memory at low cost, so that the possibility of
!
building models is no longer prevented by any lack of X 4rab g2ab
resources. These factors suggest that the time is ripe for sys- f intf ¼  r/a  r/b þ /a /b ð2Þ
gab p2
tematic attempts to model the kinetics of phase transfor- a;b¼1;...;N ;ab
!
mations in complex, metallic systems with all the relevant X X
compositional complexity accounted for. f chem ¼ /a f a f~
ca g þ l i
~ ~ c i
/a ca ð3Þ
a¼1;...;N a¼1;...;N
The research presented in this paper was carried out
with this goal in mind. Nickel-based superalloys—which where rab is the interface energy between phase a and b.
find widespread application in high-temperature applica- Different grains of the same phase, but differing in orienta-
tions such as the turbines used for jet propulsion and elec- tion so that a grain boundary is formed between them, are
tricity generation—are considered [18–20]. Attention is associated with separate phase fields. The interface energy
focused on the liquid to solid transformation; this is of will be anisotropic with respect to the relative orientation
practical importance since many critical components, e.g. between the phases. The term gab is the interface width that
turbine blades and guide vanes, are manufactured to near will be treated equal for all interfaces. The chemical free en-
net shape using casting methods [21–25]. Solidification ergy is built from the bulk free energies of the individual
can then be the cause of melt-related defects in these com- phases fa fc i
@g
 a g, dependent on the phase concentrations
ponents, which predictive methods would help to avoid. At i i
~ ¼ @ci is termed the diffusion potential of component
ca ; l
present, no model exists which addresses the full complex- i [9] to distinguish it from the chemical potential (g is the
ity of alloying in these alloys, the associated morphological chemical free energy density), introduced as a Lagrange
evolution of the dendritic structure and the subsequent multiplier to conserve the mass balance between the phases.
solid-state decomposition upon cooling. An attempt to The local number of phases is N ¼ N fxg, which is subject
do this is now described. to the constraints:
X
2. Multicomponent multi-phase-field model /a ¼ 1 ð4Þ
a¼1;...;N
X
The approach employed is based upon the multicompo- /a cia ¼ ci ð5Þ
nent multi-phase-field model reported in Refs. [9,26], with a¼1;...;N

some adaptations as follows. The governing equations for The multi-phase-field equations are given by (for details see
the phase-field /a ; 0 < /a < 1 and concentration ci of com- Ref. [27]):
ponent i of a multicomponent material are derived from
X  
the principle of Gibbs energy minimization, where the p2 dF dF
global free energy F is defined from the interfacial energy /_ a ¼  l  ð6Þ
b¼1;...;N
8gN ab d/a d/b
density f intf and the chemical energy density f chem . Both
are expanded in the phase-field variables, their gradients
which is a superposition of dual phase changes between
and the concentrations in the individual phases cia , accord-
pairs of phases. The interface mobility lab is defined
ing to:
separately for each pair of phases. The antisymmetric
approximation for the phase-field Eq. (6) [10] is used with

@fa @f
1374
the chemical driving force defined by: Dgab ¼ @/a
 @/bb
X  
p2
Dendrite Tip Temperature [°C]

/_ a ¼ lab rab /b r2 /a  /a r2 /b þ 2 ð/a  /b Þ


1372 b¼1;...;N
2g

p qffiffiffiffiffiffiffiffiffiffiffi
þ /a /b Dgab ð7Þ
g
1370
where r is the interface stiffness, which considers the inter-
face anisotropy via the Herring torque terms [28]. For a
1368 multicomponent system, a set of k diffusion equations for
every solute component is required; these are in general
not independent but linked by cross-terms. These equa-
0 2 4 6 8 tions are derived by considering the superposition of the
Velocity [mm/min]
fluxes in the individual phases:
Fig. 1. Calculated decrease in the dendrite tip temperature with the ! !
dendrite tip velocity, due to variation in the pile-up of alloying elements X N
dF X N

ahead of the tip. Estimates made using the modified marginal stability
c_i ¼ r ij
/a M a r j ¼ r ij j
/a Da rca ð8Þ
a¼1 dca a¼1
theory of the Appendix.
5864 N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875

where the chemical mobilities are denoted M ija and the dif- alloys [37,38]. It should be noted that is has been demon-
2
fusion coefficients by Dija ¼ /a M ika @c@k @c
F
j . The Gibbs energy
strated that the marginal stability theory does not capture
a a
the full physics of the dendritic growth problem, unlike the
data fa and chemical mobilities M ik are estimated from more up-to-date microscopic solvability theory [39,40].
the CALPHAD databases published by NIST [29,16]. An However, the final equations are very similar in appear-
extrapolation scheme for the thermodynamic calculations ance. Details on this are given in the Appendix. The micro-
needed at the interfaces is used to speed up the calculations structure evolution within the UC is then simulated using
[9]. The phase-field model is implemented in the MICRESS the multicomponent multi-phase-field approach.
software [30] which was coupled with the ThermoCalc The UC definition allows the smallest possible UC to
package [31]. analyzed, thus reducing computational cost. One advan-
tage of using the phase-field method to describe the micro-
3. Details of calculations for constrained dendritic growth structure evolution is that it is straightforward to enhance
the UC to simulate more complex situations. Possible
Nickel-based superalloys which are undergoing direc- enhancement in two dimensions can be achieved by using
tional solidification show mushy zone lengths on the milli- non-square-shaped UCs and larger UCs with more than
meter scale and dendrite tip radii of the order of one dendrite. These are compatible with the description
micrometers. A simulation which resolves both length given above if the following condition is fulfilled:
scales properly is currently still not feasible, in particular rffiffiffi
not for alloys which are multicomponent in nature. A
k1 ¼ ð10Þ
Approximations are therefore necessary. n
In order to overcome this limitation in this work, an which represents a correlation between the area A of the UC
approach based on a simple analysis of directionally grow- and the number of dendrites n within it. The same relation is
ing dendritic arrays is used, which is based on work pub- also commonly used to determine the PDAS of dendritic
lished by Ma [32]. In the following it is assumed that the arrays. Furthermore the UC can be enhanced by extending
directionally growing dendrites are perfectly aligned with it to three dimensions, with the size in the third dimension
the thermal gradient. The primary dendrite arm spacing chosen to reflect the secondary dendrite arm spacing (SDAS)
(PDAS), k1 , defines the characteristic length scale of direc- at the end of the solidification, see Section 5.5.
tionally growing dendritic arrays. The PDAS is related to
the thermal gradient G and solidification rate v according 4. Experiments
to:
k1 ¼ c G1=2 v1=4 ð9Þ A number of experiments were carried out with a super-
alloy designed specifically for the purposes of validating the
This correlation has been confirmed by several researchers model; the alloy composition is given in Table 1. The
[33–35]. The coefficient c is alloy dependent. In a recent results provide reference data for validating the model
study it was also shown that it depends on the anisotropy and testing its accuracy.
of the interfacial energy [36]. However, in this work it Cylindrical samples 8 mm in diameter were directionally
has been measured experimentally for the alloy analyzed, solidified. In order to achieve different thermal gradients,
see Section 4. It should be pointed out that the PDAS is Bridgman furnaces with three different experimental
the only characteristic length of the dendrites that does arrangements were used. This allowed processing at ther-
not change during the processing of the material, i.e. the mal gradients of 4, 10 and 20 Kmm1 with solidification
PDAS is the same at the dendrite tip during freezing and rates between 0.08 and 5:0 mmmin1 . A typical microstruc-
in the solidified microstructure. Thus, in order to achieve ture of samples solidified under industrially relevant casting
a fully homogeneous microstructure, microsegregation conditions consists of c dendrites and interdendritic c0 ; for
whose maximum wavelength is at the length scale of the very low solidification rates only the c phase is observed. In
PDAS has to be removed during the solution heat treat- the high gradient arrangements (10 and 20 Kmm1 ) sam-
ment, see Section 7. ples were quenched to freeze in the mushy zone morphol-
Given the above it is apparent that the PDAS is the most ogy. From these samples the planar–cellular transition
important characteristic length of the dendritic array. A was found to lie at G=v ¼ 224:4 K min1 mm1 and the cel-
representative volume element (or unit cell, UC) is there- lular–dendritic transition at G=v ¼ 126:3 K min1 mm1 .
fore defined on a transverse section with the size given by Mushy zone characteristics were determined from longitu-
2
ð0:5 k1 Þ . The factor 0.5 appears because—due to symme-
try—only one-quarter of the dendrite has to be taken into
account. The UC is isothermal, but the cooling rate T_ Table 1
depends on the solidification parameters consistent with Composition of the alloy used for the validation studies.
T_ ¼ Gv. The initial temperature is given by the correspond- Element Al Cr Ta W Ni
ing dendrite tip temperature, which is calculated using the at.% 13.06 10.49 2.67 2.92 Bal.
marginal stability approach, extended to multicomponent wt.% 5.80 8.98 7.94 8.84 Bal.
N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875 5865

dinal sections of quenched samples; the distance between To characterize the microsegregation patterns inherited
the dendrite tip and the bottom of the mushy zone (mushy from processing, wavelength dispersive X-ray spectroscopy
zone length) xsl , see Fig. 2, as well as the first occurrence of (WDX) was used. For this alloy, it was found that the ele-
interdendritic c0 were measured. The dimension of the ments Al, Cr and Ta segregate to the melt and W to the
mushy zone, but not the point of c0 formation, was found dendrites during the primary solidification, although the
to be solidification rate dependent. Furthermore, the frac- partitioning of Cr was weak, implying a partitioning coef-
tion of solid evolution along the mushy zone length was ficient k Cr close to unity. The solubility of Al and Ta in c0
measured from sections taken on quenched samples at was very high, while that of Cr and W very low. This leads
intervals of 0.9 mm. Finally the PDAS was determined to the interesting observation that the highest Cr concen-
from transverse sections and this allowed the coefficient c tration is observed in the interdendritic region, next to
in Eq. (9) to be determined as 1915 m3=4 K1=2 s1=4 (Fig. 3). the lowest Cr concentration, which is observed in the inter-
dendritic c0 . The same behaviour was recently also
observed for CMSX-10 alloy [41]. The volume fraction of
the interdendritic c0 was found to be around 1 vol.%, signif-
icantly lower than for other single-crystal superalloys (e.g.
CMSX-4  10 vol.%).
The liquidus temperature was estimated from differen-
tial thermal analysis measurements as 1386 °C, which com-
pares well with the results obtained from thermodynamic
calculations with the NIST database, which predict a liqui-
dus temperature of 1375 °C, a solidus temperature of
1340 °C and c0 solvus at 1263 °C.

5. Simulations

5.1. Simulation conditions

The UC was discretized using 100  100 square cells and


grid spacings of the order 1 lm. This corresponds to a
PDAS of 200 lm. In order to adjust the size of the UC size
to the required PDAS, the grid spacing was chosen appro-
priately, i.e. a grid spacing of Dx ¼ 0:9 lm, for example,
would be chosen for a PDAS of 180 lm. For the 3-D sim-
ulation it was assumed that the SDAS is one-third of the
PDAS and therefore a grid of the dimensions of
100  100  31 cells was used. The interfacial width gab in
Eq. (2) was chosen as 4 Dx for all interfaces.
Fig. 2. Micrograph of the experimental alloy indicating the distance xsl Consistent with the experimental work, simulations
between the dendrite tip and the base of the mushy zone. were performed for the three different thermal gradients
of 4, 10 and 20 K mm1 and a number of different solidifi-
500
cation rates, ranging from 0.08 to 8 mm min1 .
Treatment of the nucleation of interdendritic c0 at the c/
400 liquid interface was carried out as reported in Ref. [42]. The
critical undercooling for nucleation was assumed to be 7 K.
300
The thermodynamic and kinetic databases from NIST were
used [29,16].
λ1 [μm]

200
Fit 5.2. Simulation results
G=4 K/mm
100 G=10 K/mm The growth of the dendritic phase within the UC and the
G=20 K/mm
evolution of microsegregation for conditions correspond-
0
ing to a the thermal gradient of G ¼ 20 K mm1 and a
0 0.1 0.2 0.3 0.4 0.5 0.6 solidification rate v ¼ 5:0 mm min1 , is illustrated in
-0.5 -0.25 -0.5 -0.25
G v [K/mm] [mm/min] Fig. 4. The initial circular c dendrite tip quickly becomes
Fig. 3. Variation of the primary dendrite arm spacing as a function of unstable, developing the typical 4-fold symmetry expected.
thermal gradient G and solidification rate v, as obtained from the After the final portion of liquid has solidified (260 s), the
directional solidification experiments. dendritic morphology can still be identified via the segrega-
5866 N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875

Fig. 4. Simulated growth of the c dendrites, interdendritic c0 and formation of microsegregation in isothermal sections, for G ¼ 20:0 K mm1 and
v ¼ 5:0 mm min1 .

tion pattern which is inherited from processing. Al segre- concentration is observed in the dendrite core and the low-
gates to the melt during primary ðcÞ solidification est in the interdendritic c0 , which is the reverse for that seen
ðk < 1Þ, and enriches in the c0 phase. Thus the lowest Al for Al and Ta. Thus, a steady segregation behaviour is
concentration is observed in the dendrite core and the high- observed for the elements Al, Ta and W and an unsteady
est in the interdendritic c0 . Due to back-diffusion, the den- segregation for Cr. Direct comparison between measured
drite appears rather diffuse when visualized using the Al and calculated microsegregation patterns confirms these
segregation pattern. During the primary solidification, Cr observations (Fig. 5). Good agreement is also found for
segregates to the melt, but the partitioning coefficient is the position of the interdendritic c0 particles.
very close to unity. Thus the Cr concentration in the inter- In Fig. 6 the fraction of solid as function of temperature
dendritic region is higher than in the dendrite core. How- during directional solidification is shown for the case of lever
ever, the solubility of Cr in c0 is very low, and thus the rule, Scheil’s model, phase-field simulations and experiments.
lowest Cr concentration is found in the c0 particles. This Phase-field and experimental results are given for two differ-
contradicts most solidification and microsegregation theo- ent solidification rates (v ¼ 0:5 and 2:0 mm min1 ) at the
ries which assume constant and steady elemental partition- same thermal gradient ðG ¼ 10 K mm1 Þ. Naturally the
ing from the beginning to the end of solidification. Ta lever rule and Scheil’s model results are not sensitive to the
shows a partitioning behaviour very similar to that of Al, solidification conditions. Due to the initial undercooling
except that the dendritic structure appears sharp and clear assumed in the phase-field simulations, the fraction solid
due to its slower diffusion in the solid phase. W partitions close to the liquidus temperature (1375 °C) lies below the pre-
to the c phase during the primary solidification and shows diction of the Scheil’s model. After a steep increase, the
a very low solubility in the interdendritic c0 . The highest W phase-field result stays between the limits given by the lever
N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875 5867

Fig. 5. Comparison of calculated Al, Cr, Ta and W element distribution maps (top) with measured ones (bottom).

100

80
Fractionsolid[%]

60 Lever Rule
Scheil
0.5 mm/min
40 2.0 mm/min
MICRESS 0.5 mm/min
MICRESS 2.0 mm/min
20 G = 10 K/mm

0
1250 1260 1270 1280 1290 1300 1310 1320 1330 1340 1350 1360 1370
Temperature [°C]

Fig. 6. Estimates of fraction solid as a function of temperature, obtained from the Scheil model, lever rule and phase field simulations and comparison
with experimental results.

rule and the Scheil’s model, and good overall agreement 3 mm min1 and (ii) G ¼ 20 K mm1 ; v ¼ 5 mm min1 .
between phase-field simulation and experiments is found. For these conditions the PDAS k1 and the standard devia-
The results for the higher solidification rate ð2:0 mm min1 Þ tion in its value Dk1 were available from the measurements,
are slightly closer to the Scheil curve and exhibit a longer sol- and therefore in the simulation the PDAS was varied as
idus–liquidus interval. At higher temperatures (>1335 °C) k1  2Dk1 ; this resulted in simulations covering the upper
the measured fraction solid values are higher than the calcu- and lower end of the observed distribution. The cooling
lated ones. This could be due to a systematic measurement rate and initial temperature is to a first-order approxima-
error, as at these temperatures the dendrites grow during tion the same for each dendrite in an array. The results
the quench. show smaller c0 fractions at the lower, and higher c0 frac-
In practice, the PDAS in dendritic arrays exhibits a dis- tions at the upper end of the distribution, when compared
tribution rather than a unique value as indicated by the to the fraction at the average PDAS (Fig. 7). The variation
error bars in Fig. 3. For this reason the influence of the at the lower end is more pronounced than at upper end and
variations in the PDAS on the simulated microstructure the same trend is also observed for the freezing range.
was examined. Simulations for two sets of solidification These results are consistent for the two sets of solidification
parameters were performed: (i) G ¼ 4 K mm1 ; v ¼ parameters and can be explained by back-diffusion in the
5868 N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875

1.3 used for the assessment. The diagram for a thermal gradi-
ent of 10 K mm1 is shown in Fig. 8; further assessments
1.2 would be required for an extension to other thermal gradi-
20 K/mm ents. The three major regimes for directional solidification
1.1
γ’ fraction [%]

5 mm/min morphologies are indicated: planar front, cellular and den-


4 K/mm dritic growth. Each of these regimes are now discussed.
1 3 mm/min In the planar front regime, solidification proceeds with-
out a mushy zone. In this case, the growing solid phase
0.9
consists entirely of microsegregation-free c phase, the com-
0.8 position of which corresponds to the mean alloy composi-
tion; the solid phase forms at the solidus temperature of the
0.7 alloy [33]. Upon further cooling, the temperature reaches
-2 -1 0 1 2
the c0 solvus temperature and the c0 phase precipitates from
λ1 ±Δλ1
the c matrix via a solid-state transformation. For this case
Fig. 7. Variation of the calculated c0 fraction with the primary dendrite the transition temperatures, liquidus and c0 solvus, can be
spacing for fixed solidification conditions; Dk1 denotes the standard obtained from equilibrium thermodynamic calculations.
deviation of the measured primary spacing. Note that the solidification rates in this regime are approx-
imately two orders of magnitude lower that those in indus-
solid phase. Lower spacings reduce the diffusion distance trial-scale Bridgman processes used for turbine blade
and therefore increase the effect of solid-state diffusion. manufacturing.
This moves the solidification path away from the Scheil’s In the cellular front regime, the solidification front
model and closer towards that predicted by the lever rule. morphology commonly shows cells with a finite depth,
which can be measured from quenched samples. The
5.3. Kinetic phase diagram growing solid phase consists entirely of c phase, which
exhibits microsegregation due to solutal pile-up ahead
The modelling work has provided new insights into the and between the cells. This becomes more pronounced
solidification sequences occurring in these alloys, and in as the velocity increases. Upon further cooling, c0 forms
particular the construction of a kinetic phase diagram via solid-state precipitation, and due to the increasing
which correlates the phase transition temperatures with degree of microsegregation, the reaction starts at higher
the solidification conditions (see Fig. 8). Note that the dia- temperatures as the solidification rate increases. Unlike
gram is valid for the dynamic equilibrium of steady-state the case of planar front growth, the solidification front
directional solidification, as distinct from the static equilib- temperature and c0 formation temperature cannot be
rium of a thermodynamic phase diagram. Data obtained determined from simple analytical models and thermody-
from thermodynamic calculations, analytical models of namic calculations. Therefore the lines in the cellular
solidification, numerical simulations and experiments were regime in Fig. 8 are given as dashed lines, being meant
to indicate trends only.
The dendritic growth regime is where the phase-field
1400 model is particularly powerful; here transformation is asso-
Exp. Solidus liquid ciated with the formation of a substantial mushy region.
Exp. γ’ The dendrite tip temperature ðT dendrite Tip Þ was obtained
Sim. Solidus from the multicomponent marginal stability theory, which
Sim. γ’
liquid + γ gives the highest temperature at the cellular–dendritic tran-
1350
Temperature [°C]

sitions and a monotonic decrease with solidification rate, as


liq. + γ + γ’ the solidification rate increases. This constitutes the liqui-
dus temperature of the dendritic regime. The dendrites con-
γ sist of the c phase and for solidification rates higher than
1300 0:3 mm min1 the formation of interdendritic c0 is
observed. The mushy zone length xsl , as measured from
γ ’+ γ
longitudinal sections of quenched samples, multiplied by
planar cells dendrites G=10 K/mm the thermal gradient yields the non-equilibrium freezing
1250 range DT 0 . Using this in the relationship:
0.01 0.1 1 10
T 0 ¼ T dendrite Tip  xsl G ¼ T dendrite Tip  DT 0 ð11Þ
Fig. 8. The kinetic phase diagram which correlates the phase transition gives experimental estimates of the solidus line of the den-
temperatures—obtained from simulations and experiments—with the
dritic regime of the diagram in Fig. 8 (closed square sym-
solidification rate; the dashed horizontal lines mark (from the top) the
equilibrium liquidus, solidus and the Scheil c0 formation, as well as the bols). The temperature for the formation of the
equilibrium c0 solvus temperatures. interdendritic c0 is determined in a similar way.
N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875 5869

The simulated solidus temperature and the temperature thermal gradients (4, 10 and 20 K mm1 ) and different
of interdendritic c0 formation can directly be read from the solidification rates ð0:5–8:0 mm min1 Þ. The calculated
simulation results, shown as open symbols in Fig. 8. The fraction of interdendritic c0 is plotted as a function of the
temperature at which interdendritic c0 forms according to primary dendrite spacing in Fig. 9. The solid lines are
the Scheil model (1293 °C) is also given in Fig. 8 (thin hor- intended as visual guides to illustrate the trends observed.
izontal dashed line) for comparison. An increase in the The scatter in the results originates from the nucleation
solidification rate results in higher solute pile-up at the model used for the interdendritic c0 . The graphs show a
solid–liquid interface, which then in turn leads to finer den- decrease in the interdendritic c0 fraction with increasing
drites but also stronger microsegregation. This causes not PDAS. The slopes of the lines are proportional to the ther-
only the mushy zone length to increase but also triggers mal gradient, with the results for G ¼ 20 K mm1 showing
the formation of the interdendritic c0 . Moreover microseg- the steepest and for G ¼ 4 K mm1 the shallowest slopes.
regation also shifts the solid-state c0 precipitation to higher For the same range of solidification rates, a lower thermal
temperatures. Detailed information about this has not been gradient leads to a larger variation in the PDAS than a
extracted from the results available here, but it is reason- higher thermal gradient. At the same time, for the same
able to assume that the upper limit for this temperature range of solidification rates, a lower thermal gradient
shift is the formation temperature of the interdendritic c0 . results in a smaller variation in the cooling rate than a
This is supported by the observation that the interdendritic higher thermal gradient. A large PDAS and fast cooling
c0 nucleates via a solid-state reaction in the c dendrites as would weaken the effect of back-diffusion. This explains
described in Ref. [42]. why the high thermal gradient simulations show a larger
The results from simulations and experiments indicate variation of the interdendritic c0 fraction than the low-gra-
that the dendritic regime in Fig. 8 consists of two regions, dient simulations.
which are distinguished by a vertical dashed-dotted line at
0:3 mm min1 on the diagram. For lower velocities, only 5.5. Extension to 3-D calculations
the c phase grows from the melt, and the phase-transforma-
tion sequence can be summarized as liq ! liq þ c ! c The diagram in Fig. 10a illustrates the morphology of
! c þ c0s , where c0s refers to c0 formed by solid-state pre- the c dendrite in a partially solidified 3-D UC simulation,
cipitation. Above 0:3 mm min1 the formation of interden- together with the arrangement of the UC considered in this
dritic c0 is observed, and therefore the transformation case. The centre of the primary dendrite arm is located in
sequence changes to liq ! liq þ c ! liq þ c þ c0l ! c the upper left corner of the UC, while the bottom and
þc0l þ c0s , where c0l refers to c0 formed by solidification, i.e. the top plane run through the secondary dendrite arm
interdendritic c0 . Interestingly, the temperature of the inter- and the middle between two adjacent secondary arms,
dendritic c0 formation is independent of the solidification respectively. The crystallographic directions and the direc-
rate and approximately 10 K higher than the temperature tion of the thermal gradient are indicated; solidification
given by the Scheil model. The latter can be explained by proceeds in the direction of the latter. On top of the sec-
the considerable back-diffusion of the alloying elements. ondary arm the formation of a fin is observed, which would
The fast-diffusing elements Al and Cr [29,16] are most in a more freely growing dendrite eventually becoming a
likely to be responsible for this. tertiary arm.
Both simulation and experimental results show a signif- In the lower portion of Fig. 10b concentration profiles
icant increase in the liquid–solid temperature interval as the of W along the h1 0 0i direction, starting from the dendrite
solidification rate increases; this effect is particularly pro-
nounced below 0:3 mm min1 . It is also observed that in
the experiment the solidification finishes above or just
slightly below the equilibrium solidus temperature (thin 1.2
dashed line). This can be regarded as unrealistic and 1.1
indicates that the dendrite tip temperature given by the G = 4 K/mm
analytical model is too high. Above 0:3 mm min1 good 1
γ ’ fraction [%]

agreement between simulations and experiments is 0.9


observed.
0.8 G = 10 K/mm

5.4. Sensitivity studies 0.7 G = 20 K/mm

0.6
Given the assumptions in the UC model approach used
here, it is not obvious how the model will react to changes 0.5
100 200 300 400 500
in the solidification parameters G and v; one could expect
primary spacing λ1 [μm]
that changes are compensated by the cooling rate and the
primary dendrite spacing. To investigate this, a series of Fig. 9. Variation of the calculated c0 fraction with the primary dendrite
simulations was carried out for combinations of different spacing, for different combinations of G and v.
5870 N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875

a ic c0 is shown as black spotted regions, whilst the dendrite


morphology is revealed by plotting W-isosurfaces, corre-
sponding to 2.1 and 3.5 at.%. In order to generate the
image shown in Fig. 11, 16 UCs were combined and visu-
alized. It is clear that the interdendritic c0 forms in the plane
between the secondary dendrite arms and that the larger c0
particles are found along the h1 1 0i direction. The overall
secondary c0 fraction (0.37 vol.%) is slightly lower than in
the corresponding 2-D simulation (0.60 vol.%).
Naturally the computational time of these simulations
depends on the complexity of the problem and the size of
the calculation domain. Therefore the enhancement of
the UC to three dimensions has a significant influence on
b 4 the CPU time. A simulation for G ¼ 20 K mm1 and
v ¼ 5:0 mm min1 takes less than 1 day on a current PC
3D bottom
server. In the case of 3-D simulations the simulation time
3 also increases linearly with the number of cells in the third
direction. It is also found that the solidification time has a
W [at%]

3D top 2D
2

0 50 100 150 200


Dendrite core Distance [μm] Interdenditic

Fig. 10. 3-D representation of the unit cell (UC) (a) and comparison of
linescans from the bottom and top of the 3-D UC with the corresponding
2-D result (b).

core and running towards the interdendritic region in the


bottom and top plane of the 3-D UC, are given. The same
scan obtained from a 2-D UC simulation is also shown for
comparison. Both simulations were performed for the same
solidification parameters (G ¼ 10:0 K mm1 and v ¼ 0:5
mm min1 ). The curve taken from the bottom of the 3-D
simulation shows a steady decrease in the W concentration
towards the interdendritic region, although a significant
slope of the curve is observed only in the last 30 lm
towards the interdendritic region. The curve from the top
plane of the UC shows a very different concentration pro-
file. An initially steep drop of the W concentration at the
core of the dendrite is followed by a local maximum and
a steep decrease in the interdendritic region. The shapes
of the two curves reflect the morphology shown in
Fig. 10a. The local maximum of the bottom plane curve
marks the position at which the fin of the secondary arm
inpinges on the top plane of the UC (or more precisely with
the fin of the adjacent secondary arm). The curve from the
2-D simulation by way of comparison shows a steadier
decrease of the W concentration and even indicates the
location of interdendritic c0 in the interdendritic region.
The 2-D curve also stays well between the limits of the
curves from the 3-D UC.
The location of the interdendritic c0 can be clearly iden- Fig. 11. Predicted microstructure from 3D calculations, revealed by
tified from the simulation results (see Fig. 11). Interdendrit- appropriate choice of W isosurfaces (2.1 and 3.5 at.%).
N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875 5871

major influence on the CPU time: the two scale linearly. as one could also expect that this would be compensated for
Thus, simulations for lower thermal gradients and lower by a changes in the PDAS. The results confirm that different
withdrawal rates take longer. combinations of G and v result in the same PDAS but differ-
ent microstructures. This can be understood by looking at
6. Discussion the different relationship between PDAS and the cooling
rate, with the latter mainly determining the solidification
Despite the assumptions made, it has been demonstrated time. This influences significantly the dendrite morphology
that the model predicts much of the microstructural infor- and the amount of back-diffusion. However, the model is less
mation relevant to the directional solidification of the convincing in the cellular-dendritic regime; there, the
superalloys. In particular, the complex interaction between assumption of independently growing dendrites is no longer
elemental partitioning, evolution of dendritic morphology, valid. Furthermore a significant solutal pile-up will build up
back-diffusion, thermodynamic interactions and competing ahead of the solid–liquid interface. This contradicts the
length scales are accounted for. The model can be extended model assumptions and explains the difference between the
readily to other grades of superalloy, to include (i) those simulations and experimental results. However, these condi-
commercial superalloys (CMSX-4, Rene N5) which find tions are unlikely to occur in technical directional solidifica-
commercial application; and (ii) other prototype grades tion processes of turbine blades, since withdrawal rates of
which are being considered for use in advanced turbine sys- 3 mm min1 and thermal gradients of 4 K mm1 are
tems. Therefore, it is tremendously useful both for assess- employed. Therefore, this effect is unlikely to hamper the
ing the microstructure expected during processing and application of the model in practice.
hence the castability of alloys already in service, but also
for the purposes of anticipating manufacturing issues
which might arise with new grades of alloy, at the alloy 7. On the possibility of through-process modelling of
design stage. microstructure evolution of turbine blades
An important finding is that the fraction-solid evolution,
the solidification interval and the amount of interdendritic The simulation results raise the possibility of taking a
phase which forms are affected strongly by back-diffusion. significant step towards the goal of estimating the micro-
This supports findings reported earlier by other researchers structural condition at service entry, and thus the
[43]. It is found that the degree to which back-diffusion through-process modelling of microstructure evolution in
changes the solidification path depends not only on the dif- these alloys. Clearly, in order to simulate the solution heat
fusion kinetics, but also on the diffusion distance. Thus due treatment of single-crystal components, an accurate
to the formation of dendrite arms, the diffusion distance description of the solidification microstructure is necessary,
shortens and back-diffusion becomes more important. This as this defines the starting condition for the simulation; the
leads to the observed velocity-dependent solidification model presented provides this. As back-diffusion is taken
interval and explains why interdendritic c0 forms only into account, it is relatively straightforward in principle
above a given threshold of solidification rate. Classical to simulate the homogenization of the solidification micro-
1-D microsegregation models often use the secondary den- structure during subsequent solution heat treatment. In
drite arm spacing as the characteristic length, but this principle, calculations performed in this way would help
approach does not account for the composition difference to anticipate the heat treatment required for any new grade
between the dendrite core and the interdendritic region of superalloy. They might also aid in the optimization of
with the associated length, the PDAS. Subsequent simula- the design of the heat treatment process.
tions of solution heat treatment have to take this into To illustrate that this is indeed possible, calculations
account. The UC approach demonstrated here allows both have been performed for the microstructure evolution
lengths to be captured, as the dendritic morphology allows expected during a 10 hour isothermal solution heat treat-
for short back-diffusion distance while the PDAS remains ment at 1285 °C (see Fig. 12). It is apparent that after
the characteristic length scale. Finally back-diffusion also 30 min the fast-diffusing elements Al and Cr already show
raises the temperature of the interdendritic c0 formation very little remaining microsegregation, while the microseg-
by 10 K when compared to the Scheil model. The experi- regation of the slow-diffusing elements Ta and W is hardly
ments and previously reported results [44] support the affected. After 1 h new regions of inhomogeneity appear in
phase-field results. The 3-D extention of the UC leads to the dendrite side arms (Al), and move to the dendrite core
greater surface to volume rations and therefore to even after 2 h (Al and Cr). This phenomenon is attributed to the
more back-diffusion. cross-diffusion effects present in the system. The gradient in
Further points relate to the use of the UC for modelling a the W field drives the diffusion in the Al and Cr field. Thus
single representative dendrite in the array. Whilst there is the overall homogenization kinetics is controlled by the dif-
confidence assumptions will be relaxed as greater computa- fusion of W. After 10 h of solution heat treatment Al, Cr
tional resources become available, the results show already and Ta are microsegregation free, while W is still not fully
a good sensitivity to changes in the solidification parameters. homogenized. The amount of interdendritic c0 is noticeably
This is not immediately clear from the governing equations, reduced.
5872 N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875

Fig. 12. Simulated microstructure evolution during solution heat treatment at 1285 °C, illustrating the effect of homogenization of as-cast structure by
diffusion.

The efficiency of solution heat treatments is often evalu- tal values—which correspond to the standard deviation of
ated from the fraction of residual interdendritic c0 present the measured c0 fractions—appear rather large since the
at the end of the heat treatment, and this can also be pre- overall c0 fraction is not very high. However, the absolute
dicted. Fig. 13 shows the predicted evolution of the fraction variation of the measured c0 fraction is within reasonable
of the interdendritic c0 as a function of time for different limits. This provides confidence in the model as a useful
solution heat treatment temperatures, ranging from 1275 predictive tool.
to 1305 °C, as obtained from simulations. It is seen that Interestingly, the simulations predict an increase of the
interdendritic c0 dissolves significantly faster at higher tem- c0 fraction during the early stages of the heat treatment at
peratures. To test these calculations, experimental heat lower temperatures (1275 and 1285 °C)—see Fig. 13. The
treatments were carried out and these show the similar dissolution of the interdendritic c0 is predicted to occur
overall dissolution kinetics as the simulations—see only after an incubation period. This is due to the pres-
Fig. 13 symbols. Note that the error bars on the experimen- ence of microsegregation, which has a stabilizing effect on
N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875 5873

2
1285 °C
1.75 1295 °C

1.5

Fraction γ’ [%]
1.25 1275 °C

0.75
1285 °C
0.5
1295 °C
0.25
1305 °C
0
0 1 2 3 4 5 6 7 8 9 10 11
time [h]

Fig. 13. Simulated fraction of interdendritic c0 phase during solution heat treatment at various temperatures, and comparison with experimental results.

the interdendritic c0 ; the dissolution is found to start only Application of the model has provided new insight into
after the microsegregation has been sufficiently homoge- the phase transformations occurring during solidifica-
nized. This effect is not observed in the experiments; here tion; back-diffusion is important and a threshold for
the interdendritic c0 is surrounded instead by a c=c0 two- the c0 formation temperature has been predicted.
phase mixture, which formed during cooling of the solid- Extension of the methods to include homogenization of
ified material. More work is needed to account for these the as-cast microsegregation is demonstrated. This offers
effects. the hope that the service-entry microstructure can be
estimated as a function of all relevant processing steps
(casting, heat treatment); this represents a through-pro-
8. Summary and conclusions cess modelling capability.
A model, based upon on the phase-field method, was
presented for the prediction of microstructural evolution Appendix A. Dendrite tip growth parameters
during solidification of nickel-based superalloys. The fol-
lowing conclusions have been drawn: The dendrite tip radius for a given tip velocity vtip and
composition of the melt at the tip ci is given by:
It is demonstrated that the model can account for the sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
full compositional complexity of technically relevant 2C 1
Rtip ¼ 2p P mi ð1kÞc  ð12Þ
alloys which contain substantial quantities of Al, Ta, i vtip
Di
W and Cr among other elements.
The microsegregation, phase fractions in the as-cast where Di are the individual liquid diffusion coefficients of
microstructures, the solidification-rate-dependent dimen- the alloying elements and C ¼ r=DS f the Gibbs–Thomson
sions of the mushy zone and the sequence of phase forma- coefficient, r is the interfacial energy and DS f the entropy
tion are correctly predicted. The model also shows good of fusion per volume. The dendrite tip radius Rtip and the
response to changes in the casting conditions. The tip undercooling are correlated by the Ivantsov solution I v :
enhancement of the simulation approach to the 3-D UC  
c  c0i Rvtip
is demonstrated. X¼ i ¼ I v ð13Þ
ð1  kÞci 2Di
Since the model is connected to established thermody-
namic and kinetic databases of parameters, it has the with ci the interface liquid concentration of element i and
advantage of using well-assessed information; moreover c0i the corresponding far field (alloy) concentration. Eq.
its fidelity will improve as the accuracy of the underlying (13) is evaluated for each element separately. For a given
databases is improved. tip radius, the tip concentration is given by:
The model represents a valuable tool for the prediction c0i
of microstructure in commercially relevant superalloys ci ¼ ð14Þ
Xð1  kÞ  1
but also those which are under development. Therefore,
it is useful both for process modelling/simulation for The actual dendrite tip temperature is obtained when the
assessing manufacturability but also for the purposes temperature equivalent of the solutal pile-up is summed
of the design of new alloys. up over all alloying elements:
5874 N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875
X
T tip ¼ mli ðci  c0i Þ ð15Þ culate the dendrite tip temperature. Starting from an
initial estimate of the tip radius and the alloy composi-
The dimensionless undercooling X is calculated with the tion, the dimensionless undercooling X and a new tip
Ivantsov solution [45], which is different for 2-D and 3-D concentration ci is calculated with Eqs. (13) and (14).
problems: This new concentration is inserted into Eq. (12) and the
pffiffiffiffiffiffi pffiffiffi
2-D : X ¼ pP exp P erfc P ð16Þ result is compared to the previous estimate of the tip
radius. If the agreement is poor, the result is used as
3-D : X ¼ P exp P E1 P ð17Þ
the new guess for the next iteration. Usually convergence
where P ¼ Rv=2D is the solutal Pclet number, erfc denotes is reached after two or three iterations. It is found that
the error-function-complement and E1 the exponential the dendrite tip temperature (Fig. 1) and tip radius
integral. (Fig. 14) decrease with the solidification rate. The den-
The Ivantsov solution was derived for an isolated den- drite tip temperature defines the starting condition for
drite tip. Thus for a dendrite array with overlapping solutal the UC simulations.
fields, deviations in the tip radius and temperature are
expected. In the case of sufficiently fast propagation of References
the dendrite tips, the diffusion length ld ¼ 2D=vtip becomes
smaller than half the primary dendritic spacing ld < k1 =2. [1] Boettinger WJ, Coriell SR, Greer AL, Karma A, Kurz W, Rappaz M,
et al. Acta Mater 2000;48(1):43–70.
In this case dendrites growing in dendritic arrays can be
[2] Hecht U, Granasy L, Pusztai T, Bottger B, Apel M, Witusiewicz V,
treated as for the isolated case [46]. et al. Mater Sci Eng R: Reports 2004;46(1–2):1–49.
With Eq. (12) the influence of a single alloying element [3] Wang W, Lee PD, McLean M. Acta Mater 2003;51(10):2971–87.
of the dendrite tip radius can be estimated. The contribu- [4] Ludwig A, Gruber-Pretzler M, Mayer F, Ishmurzin A, Wu M. Mater
tions of the alloying elements are summed up and it is Sci Eng A 2005;413–414:485–9.
[5] Asta M, Beckermann C, Karma A, Kurz W, Napolitano R, Plapp M,
found that there is little or no influence elements with
et al. Acta Mater 2009;57(4):941–71.
k i 1 (no partitioning), mi 0 (no influence on the liqui- [6] Warren JA, Boettinger WJ. Acta Metall Mater 1995;43(2):689–703.
dus temperature) and ci 0 (vanishing concentration), or [7] Moelans Nele, Blanpain Bart, Wollants Patrick. Calphad 2008;
respectively combinations of these. 32(2):268–94.
An iterative scheme based on the linearized phase dia- [8] Kim Seong Gyoon, Kim Won Tae, Suzuki Toshio. Phys Rev E
1999;60(6):7186–97.
gram given in Table 2 and Eqs. (12)–(14) was used to cal-
[9] Eiken J, Bottger B, Steinbach I. Phys Rev E 2006;73(6). 066122–9.
[10] Steinbach I, Pezzolla F, Nestler B, Seeßelberg M, Prieler R, Schmitz
GJ, et al. Phys D Nonlinear Phenom 1996;94(3):135–47.
Table 2 [11] Grafe U, Böttger B, Tiaden J, Fries SG. Scripta Mater 2000;42:
Linearized phase diagram, reference temperature is T 0 ¼ 1375:2 C. 1179–86.
Concentrations are given in [at.%] and slopes in ½K at:%1  [12] Jacot A, Rappaz M. Acta Mater 2002;50(8):1909–26.
Element cliquid csolid mliquidus msolidus k [13] Kattner UR. JOM 1997;49(12):14–9.
[14] Lukas HL, Fries SG, Sundman B. Computational thermodynam-
Al 13.06 10.63 12.7 13.8 0.81 ics. Cambridge: Cambridge University Press; 2007.
Cr 10.49 9.60 36.5 29.2 0.92 [15] Campbell CE. Acta Mater 2008;56(16):4277–90.
Ta 2.66 2.00 44.4 52.3 0.75 [16] Campbell CE, Boettinger WJ, Kattner UR. Acta Mater 2002;
W 2.92 3.71 33.0 43.0 1.27 50(4):775–92.
[17] Engstrom A, Morral JE, Agren J. Acta Mater 1997;45(3):1189–99.
[18] Reed RC. The superalloys. Cambridge: Cambridge University Press;
2006.
[19] Y. Koizumi, T. Kobayashi, T. Yokokawa, Z. Jianxin, M. Osawa, H.
25 Harada, et al. In: Superalloys 2004. TMS; 2004.
[20] Durand-Charre Madeleine. The microstructure of superalloys. Lon-
don: Gordon and Breach; 1997.
Dendrite Tip Radius [μm]

20
[21] Husseini Naji S, Kumah Divine P, Yi Jian Z, Torbet Christopher J,
Arms Dohn A, Dufresne Eric M, et al. Acta Mater
15 2008;56(17):4715–23.
[22] Zhou YZ, Volek A, Green NR. Acta Mater 2008;56(11):2631–7.
[23] Hobbs R, Tin S, Rae C. Metall Mater Trans A 2005;36(10):2761–73.
10
[24] Ma Dexin, Grafe Uwe. Mater Sci Eng A 1999;270(2):339–42.
[25] Zhang J, Lou LH. J Mater Sci Technol 2007;23(3):289–300.
5 [26] Steinbach I, Böttger B, Eiken J, Warnken N, Fries SG. J Phase
Equilibria Diffus 2007;28(1):101–6.
[27] Steinbach I, Pezzolla F. Phys D Nonlinear Phenom 1999;134(4):
0 385–93.
0 2 4 6 8
Velocity [mm/min] [28] Herring C. In: Gomer R, Smith CS, editors. Structure and properties
of solid surfaces. Chicago: University of Chicago Press; 1952.
Fig. 14. As finer dendrite tips can grow faster than coarse ones, the tip [29] Kattner UR. In: Shull RD, Turchi PEA, Gonis A, editors. CALP-
radius decreases with increasing dendrite tip velocity. The curves was HAD and alloy thermodynamics. Warrendale, PA: TMS; 2002. p.
calculated using a modified marginal stability theory. 147–64.
N. Warnken et al. / Acta Materialia 57 (2009) 5862–5875 5875

[30] MICRESS. Web-page: www.micress.de. [40] Müller-Krumbhaar H, Kurz W, Brener E. In: Kostorz G, editor.
[31] Sundman B, Jansson B, Anderson JO. CALPHAD 1985;9:153–90. Phase transformations in materials. Weinheim: Wiley-VCH; 2001. p.
[32] Ma D. Gießereiforschung 1998;1:29–34. 81–170.
[33] Kurz W, Fisher DJ. Fundamentals of solidification. Zurich: Trans [41] Seong Moon Seo, Je-Hyun Lee, Young-Soo Yoo, Chang-Yong Jo,
Tech; 1998. Hirofumi Miyahara, Keisaku Ogi. In: Reed RC, Green KA, Caron P,
[34] Hunt JD. Solidification and casting of metals, vol. 192. London: The Gabb TP, Fahrmann MG, Huron ES, editors. Superalloys 2008.
Metals Society; 1979. TMS; 2008. p. 277–6.
[35] Ma D, Sahm PR. Metall Mater Trans A 1998;29(3A):1113–9. [42] Warnken N, Ma D, Mathes M, Steinbach I. Mater Sci Eng A
[36] Steinbach I. Acta Mater 2008;56(18):4965–71. 2005;413–414:267–71.
[37] Bobadilla M, Lacaze J, Lesoult G. J Cryst Growth 1988;89(4): [43] Thirumalai A, Akhtar A, Reed RC. Mater Sci Technol 2006;
531–44. 22(1):1–13.
[38] Diepers H, Eiken J, Steinbach I. In: Jolly MR, Stefanescu DM, [44] Walter C, Hallstedt B, Warnken N. Mater Sci Eng A 2005;397(1–2):
Warren JA, Krane MJ, editors. Modeling of casting, welding and 385–90.
advanced solidification processes X, vol. 10; 2003. p. 29–36. [45] Ivantsov GP. Doklady Akademii Nauk SSSR 1947;58:567.
[39] Langer JS. Phys Rev A 1986;33(1):435–41. [46] Warren James A, Langer JS. Phys Rev E 1993;47(4):2702–12.

You might also like