You are on page 1of 7

Available online at www.sciencedirect.

com

Acta Materialia 56 (2008) 1890–1896


www.elsevier.com/locate/actamat

Analytical treatment of diffusion during precipitate growth


in multicomponent systems
Qing Chen a,*, Johan Jeppsson b, John Ågren b
a
Thermo-Calc Software AB, Björnnäsvägen 21, 113 47 Stockhlom, Sweden
b
Department of Materials Science and Engineering, KTH, 100 44 Stockholm, Sweden

Received 17 October 2007; received in revised form 19 December 2007; accepted 19 December 2007
Available online 4 March 2008

Abstract

We propose an approximate growth rate equation that takes into account both cross-diffusion and high supersaturations for modeling
precipitation in multicomponent systems. We then apply it to an Fe-alloy in which interstitial C atoms diffuse much faster than substi-
tutional solutes, and predict a spontaneous transition from slow growth under ortho-equilibrium to fast growth under the non-partition-
ing local equilibrium condition. The transition is caused by the decrease in the Gibbs–Thomson effect as the growing particle becomes
larger. The results agree with DICTRA simulations where full diffusion fields are calculated.
Ó 2008 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.

Keywords: Precipitation; Diffusion; Kinetics; Thermodynamics; NPLE

1. Introduction In general, for multicomponent alloys the operating tie-


line at the phase interface has to be found numerically by
Precipitation processes may be modeled in detail by dif- seeking a common growth rate for all independent compo-
fuse interface methods such as phase-field [1] or sharp nents. In previous studies the growth rate was formulated
interface methods like DICTRA [2]. These simulations by either neglecting the cross-terms in the diffusivity matrix
are based on numerical calculations of full diffusion fields in order to consider high supersaturations [4], or by includ-
and may be quite time consuming. Consequently, only a ing cross-diffusion terms but restricting the calculations to
small number of particles may reasonably be handled. In small supersaturations [3]. Sometimes, even for small super-
many applications one is less interested in fine details but saturations, the cross-diffusion was ignored [5]. In Ref. [4], it
more interested in handling complex alloys with more than was felt impossible to use the analytic binary growth rate
10 elements and in the size distributions of several precipi- equation for multicomponent systems, and the consequence
tate phases. It is then necessary to take a simpler approach. of neglecting the cross-diffusion terms was severe deviation
Usually one avoids solving the full diffusion problem and of the modeled results from experimental observations so
rather applies analytical expressions for the diffusive fluxes that an ad hoc growth rate had to be found. The approach
of the solute elements. Such expressions may be obtained in Ref. [3] was not aimed specifically at high supersatura-
from the steady-state field approximation or similar tions. However, for a system with significantly different
approaches [3–6]. These solutions only use the precipitate, atomic mobilities, large supersaturations can occur for
the interface and the matrix average compositions to repre- slow-moving elements even though the overall supersatura-
sent the diffusion field [3–5]. In some approaches not even tion of the system is very small. In the extreme case of a non-
the interface compositions are used [6]. partitioning local equilibrium (NPLE) growth, the super-
saturations of substitutional elements approach unity. The
*
Corresponding author. Tel.: +46 8 54595939; fax: +46 8 6733718. approach is then not applicable. Therefore, it seems neces-
E-mail address: qing@thermocalc.se (Q. Chen). sary to find a multicomponent growth rate equation that

1359-6454/$34.00 Ó 2008 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.actamat.2007.12.037
Q. Chen et al. / Acta Materialia 56 (2008) 1890–1896 1891

can take both cross-diffusion terms and high supersatura- tðcP  cI Þ ¼ DðcM  cI Þ=nR: ð6Þ
tions into account at the same time. In this work, we present
If thepsupersaturation
ffiffiffiffiffiffiffiffiffi is very small, i.e. X ! 0, one obtains
such an equation and apply it to a steel to predict a sponta-
k ¼ X=2 and thus n ¼ 1, which recovers the well-known
neous transition from slow growth to fast growth due to the
steady-state equation used, for example, in coarsening the-
Gibbs–Thomson effect.
ory [9]:
2. Model DðcM  cI Þ
t¼ : ð7Þ
RðcP  cI Þ
For the sake of simplicity, we treat a spherical particle of
If the supersaturation is very large, i.e. X ! 1, one obtains
stoichiometric composition or with negligible atomic diffu- pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sivity growing under the diffusion-controlled condition. We k ¼ 3=2ð1  XÞ and thus n ¼ Xð1  XÞ=3 ! 0, i.e. the
start from the exact solution to binary systems and refor- effective diffusion distance becomes very small and the
mulate it by introducing an effective diffusion distance fac- growth rate very large.
tor, and then derive a general multicomponent growth rate
equation on the basis of the local equilibrium assumption 2.2. Multicomponent systems
and the flux balance equations that depends on individual
effective diffusion distances, which in turn varies with In an n-component alloy of concentration cM i , generally
changing supersaturations of individual components. speaking, the interface concentrations cIi and cPi in the
Finally, we break up diffusivities into mobilities and ther- matrix and precipitate phases are determined simulta-
modynamic factors, and present an equation to treat neously [10] with the interface velocity from the flux bal-
cross-diffusion directly by working with mobilities and ance equation for each element:
chemical potential gradients.   Xn1
t cPi  cIi ¼ Dij rcj ð8Þ
2.1. Binary systems j¼1

where Dij is the chemical diffusivity in the matrix phase and


Given an alloy of concentration cM, the concentration in rcj is the concentration gradient of element j at the inter-
the matrix at the phase interface cI and that in the precip- face. Both i and j go from 1 to n  1 because there are n  1
itate cP can be uniquely determined by the phase equilib- independent concentration variables in an n-component
rium tie-line under the isothermal local equilibrium system. So far we have only n  1 equations for 2n  1
assumption. With these well-defined interface conditions, unknowns. Invoking the local equilibrium assumption,
the diffusion-controlled growth of an isolated precipitate we obtain n more necessary equations:
in an infinite matrix can be solved exactly [7,8] and the
growth rate t can be written as lPi ¼ lIi ; ð9Þ

2Dk2 where lPi and lIi are chemical potentials of element i in the
t¼ ; ð1Þ precipitate and matrix at the interface, respectively, and are
R
functions of cPi and cIi . Now we have to define rcj in order
where D is the diffusivity in the matrix, R is the radius and to solve 2n  1 (Eqs. (8) and (9)). Supposing that the effec-
k is given by tive diffusion distance dj or the factor nj for each element j
pffiffiffi depends only on its own supersaturation Xj , we can obtain
2k2  2k3 p expðk2 ÞerfcðkÞ ¼ X; ð2Þ
a general multicomponent growth rate equation directly
where X ¼ ðcM  cI Þ=ðcP  cI Þ is the so-called dimension- from Eq. (6):
less supersaturation. Eqs. (1) and (2) are actually obtained
from the flux balance equation:   Xn1  .
t cPi  cIi ¼ Dij cM I
j  cj nj R: ð10Þ
tðcP  cI Þ ¼ Drc; ð3Þ j¼1

where rc is the concentration gradient close to the phase During the coarsening stage of a precipitate, Xj ! 0 and
interface. Let us introduce an effective diffusion distance thus nj ! 1, and we obtain the same multicomponent
d and rewrite the above equation as coarsening rate equation as that of Morral and Purdy [11].
tðcP  cI Þ ¼ DðcM  cI Þ=d: ð4Þ
2.3. Direct treatment of cross-diffusion with mobilites and
Comparing this result with Eq. (1), we get: chemical potentials
X
d¼ R ¼ nR; ð5Þ The off-diagonal terms in the diffusivity matrix are
2k2
mostly related to thermodynamic effects but also depend
where n ¼ X=2k2 is a factor for adjusting the effective diffu- on the choice of frame of reference. The first effect comes
sion distance from the radius as the supersaturation varies. from expressing fluxes as functions of concentration gradi-
Inserting the above expression into Eq. (4), we get: ents. The driving force for diffusion is essentially a gradient
1892 Q. Chen et al. / Acta Materialia 56 (2008) 1890–1896

in chemical potential and it can therefore be more suitable


to work directly with mobilities rather than with chemical
diffusivities. The diffusivity matrix can be divided into a
mobility M and a thermodynamic factor ol/oc, and the
large off-diagonal elements in the matrix come from the
thermodynamic factor:
X
n1
olj
Dik ¼ cj M ij : ð11Þ
j¼1
ock

The diffusional flux of element i can then be expressed as


X
n1 X
n1 X
n1
olj
Ji ¼  Dik rck ¼  cj M ij rck
k¼1 j¼1 k¼1
ock
X
n1
¼ cj M ij rlj : ð12Þ
j¼1
Fig. 1. The calculated growth rates of metastable M23C6 particles of
The off-diagonal terms in the last part of Eq. (12) are only different sizes in supersaturated ferrite alloys A (Fe–2 wt.% Cr–0.2 wt.%
related to the frame of reference; they are zero in the lattice C) and B (Fe–2 wt.% Cr–0.05 wt.% C).
fixed frame of reference and non-zero in any other frame of
reference due to the Kirkendall effect. The effect is, how-
ever, small compared to the approximations introduced a normal t vs. R curve, such as curve B, where the growth
earlier with the modified steady-state equation. Neglecting rate of a M23C6 particle from another supersaturated fer-
these cross-terms, Eq. (10) can be expressed with mobilities rite alloy B (Fe–2 wt.% Cr–0.05 wt.% C) at the same tem-
and chemical potential gradients as perature is shown to increase sharply to its maximum at
  around twice the critical nucleus size and then decreases
t cPi  cIi ¼ cIi M i ðlM I
i  li Þ=ni R: ð13Þ
slowly.
In the above equation, the cross-diffusion is accounted for We now examine the other results obtained at the same
directly by the dependence of chemical potentials on time during the simulation, i.e. the dependence of interface
concentrations. concentrations on the particle size for alloys A and B. The
results are depicted in Figs. 2 and 3, where u-fractions or
3. Results and discussion atomic fractions per mole of substitutional elements are
used. The C and Cr contents in the body-centered cubic
We now apply our model to calculate the growth rate of (bcc) matrix phase far away from the interface are shown
a M23C6 particle metastably formed from a supersaturated as dotted lines, which coincide with the left end points of
ferrite alloy A (Fe–2 wt.% Cr–0.2 wt.% C) at 1053 K. All those curves for the bcc phase, i.e. concentrations at the
necessary thermodynamic and kinetic data, i.e. the temper- interface in the bcc phase for a particle of the critical
ature- and composition-dependent Gibbs free energy and nucleus size. As can clearly be seen, for alloy A at a particle
atomic mobility, were taken from the well-assessed dat- size of about 2.1 nm, abrupt changes occur for the Cr con-
abases TCFE4 [12] and MOB2 [13]. When mobility varies tent in both bcc and M23C6 phases and for the C content in
with composition, an average value was used. The bcc. For particles larger than 2.1 nm in radius, their Cr
Gibbs–Thomson effect was taken into account during the content at the interface is almost the same as that in the
calculation of chemical potentials by adding a pressure dif- matrix. In contrast, no sudden changes exist for the varia-
ference term, 2rV m =R, to the Gibbs energy of the M23C6 tions of interface concentrations in alloy B, and the redis-
phase. The interfacial energy r and the molar volume tributions of solutes in the matrix and precipitate phases
V m , based on substitutional elements, were assumed con- are continuous and very smooth.
stant and equal to 0.4 J m2 and 6  106 m3 mol1, Returning to Fig. 1, we see from curve A that the
respectively. For each given particle radius R, the growth growth rates for small particles around 1–2 nm are several
rate t was obtained together with the interface concentra- orders of magnitude slower than those for particles larger
tions cPi and cIi (i = C, Cr) by numerically solving five non- than 2.1 nm. So the curve can be easily divided into two
linear equations that consist of Eqs. (9) and (13). parts, one corresponding to a slow growth mode and the
The calculated results are shown as curve A in Fig. 1. It other to a fast growth mode. In fact, from Figs. 2 and 3,
is interesting to see that the growth rate increases dramat- we know immediately that the slow growth mode involves
ically from zero as the particle size increases from the crit- redistribution of Cr in the matrix and precipitation under
ical radius, 0.78 nm, and then reaches almost a plateau, local ortho-equilibrium, and the fast growth mode involves
and then shoots up again and reaches a maximum, and no such redistribution as the Cr content in the particles is
finally decreases gradually. This is obviously different from the same as that in the matrix, which corresponds to a
Q. Chen et al. / Acta Materialia 56 (2008) 1890–1896 1893

Fig. 2. The calculated interface compositions for the metastable growth of Fig. 3. The calculated interface compositions for the metastable growth of
M23C6 particles in supersaturated ferrite alloy A (Fe–2 wt.% Cr–0.2 wt.% M23C6 particles in supersaturated ferrite alloy B (Fe–2 wt.% Cr–0.05 wt.%
C). The dotted line represents the matrix composition. C). The dotted line represents the matrix composition.

phase transformation under the so-called NPLE condition. the M23C6 phase, i.e. the contribution due to the so-called
For alloy B, the growth of the particle is governed by the Gibbs–Thomson effect. Therefore, for a series of given par-
slow mode and is not influenced by particle size. ticle radius, we may draw a series of corresponding bcc sol-
To find out why the behavior for alloys A and B are so vus lines and the related tie-lines. For clarity, only the lines
different and why the growth mode can change spontane- for the particle size of 2.1 nm, where the growth mode
ously as the particle grows larger, we need to examine the changes in alloy A, are plotted and they have the same col-
metastable phase equilibrium between the bcc and M23C6 ors, but are dashed. The two circle symbols mark the con-
phases and the influence of the Gibbs–Thomson effect on centrations of alloys A and B. If we calculate the bcc solvus
it. The isothermal section of the Fe–Cr–C system at lines for the M23C6 particles of critical sizes in alloys A and
1053 K is calculated by using Thermo-Calc software [14] B, they should certainly pass through points A and B,
and the thermodynamic database TCFE4 [12]. The result- respectively.
ing Fe-rich corner is shown in Fig. 4. To calculate the meta- By ignoring the M23C6 phase, the iso-carbon-activity
stable bcc and M23C6 phase equilibrium, all other phases in lines for the two alloys can also be calculated, and these
the database are suspended during the calculation. The are shown by the red lines passing through points A and
black solid line is the solvus of the bcc phase, and the green B. The former intersects at C with the solvus line of only
solid lines are the tie-lines of the two phases being consid- the 2.1 nm particles, and the latter crosses both solvus lines,
ered here. For particles with a finite radius, the phase equi- especially at E for that corresponding to an infinitely large
librium can be calculated by adding a curvature-induced particle. It should be mentioned that two special tie-lines
pressure difference term, 2rV m =R, to the Gibbs energy of for M23C6 containing the same amount of Cr as in the
1894 Q. Chen et al. / Acta Materialia 56 (2008) 1890–1896

value, which was calculated to be 2.8 nm, because there


exists no cross-point for the bcc solvus and iso-carbon-
activity lines above this critical value. For particles that
are smaller than 2.8 nm but larger than 2.1 nm, the bcc sol-
vus and iso-carbon-activity line do intersect with each
other, but this does not mean that the slow growth mode
is possible. In fact, the intersection points for particles in
this range of size would yield tie-lines suggesting that the
Cr content in M23C6 at the interface is lower than that in
the matrix. This is simply not viable for the growth of
M23C6 particles from the bcc matrix where the distribution
coefficient of Cr is greater than 1.

3.1. Comparison with DICTRA simulation results

The derived model is intended for multiparticle, multi-


Fig. 4. The calculated isothermal Fe–Cr–C phase diagram at 1053 K. The
phase simulations in complex alloys and is therefore a sim-
solid black line is the solvus of the bcc phase, and the solid green lines are plified approach to the diffusion problem. The model has
the conjugate tie-lines of the bcc and M23C6 phases. The dashed black and been compared to solutions of the full diffusion problem
green lines are similar ones corresponding to a particle size of 2.1 nm. The to show that the introduced approximations are not severe.
red lines are iso-carbon-activity lines for ferrite alloys A and B, The software DICTRA [2] was used to simulate the growth
respectively. The heavy blue lines are the loci of the compositions of the
bcc phase at the interface with M23C6 particles of different sizes.
of one spherical M23C6 particle in alloys A and B. All ther-
modynamic and kinetic data were taken from the databases
TCFE4 [12] and MOB2 [13], and the Gibbs–Thomson
matrix have also been calculated for infinitely large and effect was accounted for in the usual way.
2.1 nm particles. They are those lines starting from the It should be mentioned that the simulations in DICTRA
bcc side at D and C, respectively. The loci of the composi- were numerically tricky to perform due to large differences in
tions of the bcc phase at the interface shown in Figs. 2 the magnitude of both particle size and growth rate. For
and 3 are also superimposed in the diagram as heavy blue example, in the case of alloy A, concentration gradients in
lines. As we can see, for alloy B, the locus starts from B the range of fractions of nanometers have to be resolved
and follow the iso-carbon-activity line and end at E; for in the matrix close to the particle, and the matrix has to be
alloy A, the locus starts from A and follow the iso-car- of the order of micrometers to avoid impingement. The growth
bon-activity line to C, and then change its course abruptly rate also changes abruptly by five orders of magnitude.
and move almost horizontally to D. As expected, the locus
corresponding to the slow mode coincides with the iso-car-
bon-activity lines and this means that the chemical potential
gradient of carbon near the interface in the bcc phase is
almost negligible. However, despite the negligible gradient,
a small finite interface velocity, which matches that of the
very slow-moving Cr atoms, can be obtained for carbon
due to its extremely large atomic mobility. Hence, the
growth of the particles is controlled by the diffusion of Cr
in this mode.
In the case of alloy A, when the particle size is larger
than 2.1 nm, the fast growth mode is adopted because it
is now possible for the precipitates to inherit the Cr content
in the matrix, which means undergoing a NPLE during
growth that is fully controlled by the diffusion of carbon.
In this mode, the supersaturation of Cr is almost 1, so its
effective diffusion distance factor is approaching 0, and
the gradient of chemical potential of Cr becomes tremen-
dously large, which makes it possible for the slow-moving Fig. 5. The calculated growth rates of metastable M23C6 particles of
Cr to acquire a very high interface velocity, the same as different sizes in supersaturated ferrite alloys A (Fe–2 wt.% Cr–0.2 wt.%
C) and B (Fe–2 wt.% Cr–0.05 wt.% C). The solid black lines indicate
that for C, whose chemical potential gradient is now finite
results from direct treatment of cross-diffusion. The dashed lines indicate
and whose mobility is always very large. results from indirect treatment of cross-diffusion. The dotted lines indicate
It is understandable that the NPLE growth mechanism results from indirect treatment without use of cross-diffusivities. The
is the only feasible one for particles larger than a critical square and diamond symbols indicate results from DICTRA simulations.
Q. Chen et al. / Acta Materialia 56 (2008) 1890–1896 1895

the direct method or DICTRA. The indirect method of


treating cross-terms gives a similar result as the direct
method and agrees more or less with the DICTRA simula-
tion, which indicates that the chemical potential gradients
are described well with cross-diffusivities and composition
gradients for alloy B in our individual effective diffusion
distance approach.
In the case of alloy A, the results are different. By
neglecting the cross-diffusivities, the transition from
ortho-equilibrium to NPLE is captured but the growth
rates are overestimated and the transition is not at the cor-
rect particle size. This is due to the same reason as
discussed in the preceding paragraph, i.e. the wrong inter-
face compositions that follow the iso-carbon-composition
line and deviate from the iso-carbon-activity line. An even
worse, and maybe surprising, result is that the indirect
Fig. 6. Same as Fig. 4 except that the dashed blue lines are the results
method of treating cross-terms cannot capture the transi-
from indirect treatment of cross-diffusion, and the dotted blue lines are the tion to NPLE, and the growth rate is orders of magnitude
results from indirect treatment without use of cross-diffusivities, and the too low when the particle size is over 2.1 nm. In Fig. 6, we
square and diamond symbols are the results from DICTRA simulations. can see that the interface compositions never reach NPLE
and the growth therefore never gets fully controlled by car-
The results from the DICTRA simulations are indicated bon diffusion. The indirect method of treating cross-diffu-
by square and diamond symbols in Figs. 5 and 6, together sion estimates incorrectly the cross-term effect in this
with the results from our simplified approach indicated by extreme case where the diffusion distance of one compo-
solid lines. For both the growth rate and composition as a nent is very short and therefore the cross-terms become
function of particle size, the agreement is very good. The very important. The direct method of treating cross-diffu-
simplified model can even in the extreme case of alloy A sion estimates correctly the cross-term effect because it is
fully describe the switch of the growth mode from ortho- accounted for directly by the dependence of chemical
equilibrium to NPLE. potential on concentrations.

3.2. Method with diffusion coefficients 4. Summary and conclusions

In Section 2.2, we derived a method of treating multi- We have proposed an approximate growth rate equation
component particle growth including cross-diffusion with that takes into account both cross-diffusion and high super-
diffusion coefficients. The method is indirect in the sense saturations for modeling precipitation in multicomponent
that the cross-diffusion can be accounted for directly by systems. Although the derivation is confined to spherical
using mobilities and chemical potentials as shown in Sec- particles, it can be readily extended to other simple geom-
tion 2.3. etries. With this equation, different possible growth modes
The results from the simulations for alloys A and B can be captured automatically without any ad hoc treat-
using this indirect method, i.e. solving Eqs. (9) and (10), ment. We have applied the equation to a Fe-alloy where
are also shown in Figs. 5 and 6. For comparison, the results interstitial C atoms diffuse much faster than substitutional
from the calculations using diffusivities without cross-terms solutes and found a spontaneous transition from slow
have also been added. growth under ortho-equilibrium to fast growth under the
Starting with alloy B, it is easy to see that the cross- NPLE condition. From analysis of phase equilibrium and
terms are important. Neglecting cross-diffusivities results thermodynamics subjected to pressures induced by the cur-
in large deviations from the direct method and the DIC- vature of particles of finite sizes, we understand that the
TRA simulation results. The growth rates are overesti- transition is caused by the decrease in the Gibbs–Thomson
mated because the obtained operating compositions at effect as the growing particle becomes larger. DICTRA
the interface are wrong. Ignoring the cross-terms in Eq. simulations have also been carried out and the obtained
(10), we see that in order to achieve a slow growth mode, results corroborate that of our simplified approach. It has
the matrix carbon composition at the interface has to be also been found that the method using mobilities instead
about the same as that far away from the interface. Note of diffusivities within our individual effective diffusion dis-
that it is composition and not chemical potential or activity tance approach is probably always superior due to the fact
that is used. This is why, without cross-terms, the operating that chemical potential gradients instead of composition
interface compositions follow almost the iso-carbon-com- gradients are then used and the cross-term effect can there-
position line in Fig. 6. The correct interface compositions fore be accounted for directly by the composition depen-
should follow the iso-carbon-activity line as indicated by dence of chemical potentials.
1896 Q. Chen et al. / Acta Materialia 56 (2008) 1890–1896

Acknowledgments [5] Robson JD. Acta Mater 2004;52:1409.


[6] Svoboda J, Fischer FD, Fratzl P, Kozeschnik E. Mater Sci Eng A
2004;385:166.
One of the authors (J.J.) gratefully acknowledges sup- [7] Aaron HB, Fainstein D, Kotler R. J Appl Phys 1970;41:4404.
port from the Swedish foundation for strategic research [8] Zener C. J Appl Phys 1949;20:950.
(SSF) through the MATOP program. [9] Lifshitz IM, Slyozov VV. J Phys Chem Solid 1961;19:35.
[10] Andersson J-O, Hoglund L, Jonsson B, Agren J. Computer simula-
tion of multicomponent diffusional transformation in steel. In: Purdy
References GR, editor. Fundamentals and applications of ternary diffusion. New
York: Pergamon Press; 1990. p. 153.
[1] Chen LQ. Annu Rev Mater Res 2002;32:113. [11] Morral JE, Purdy GR. Scripta Metall Mater 1994;30:905.
[2] Borgenstam A, Engstrom A, Hoglund L, Agren J. J Phase Equilib [12] Thermo-Calc AB. TCFE4, Thermo-Calc Software AB Steel Data-
2000;21:269. base, Version 4. Stockholm: Thermo-Calc AB; 2006.
[3] Jou H-J, Voorhees P, Olson GB. Computer simulations for the [13] Thermo-Calc AB. MOB2, Thermo-Calc Software AB Mobility
prediction of microstructure/property variation in aeroturbine disks. Database, Version 2. Stockholm: Thermo-Calc AB; 1998.
Superalloy 2004. Champion, PA: TMS; 2004. p. 877. [14] Sundman B, Jansson B, Andersson J-O. CALPHAD 1985;9:
[4] Sourmail T. PhD thesis, University of Cambridge; 2002. 153.

You might also like