You are on page 1of 21

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/223573638

First-principles calculations of the structural and thermodynamic properties


of bcc, fcc and hcp solid solutions in the Al–TM (TM=Ti, Zr and Hf) systems: A
comparison of cluster ex...

Article  in  Acta Materialia · August 2008


DOI: 10.1016/j.actamat.2008.03.006

CITATIONS READS

94 434

3 authors:

Gautam Ghosh Axel van de Walle


Northwestern University Brown University
142 PUBLICATIONS   3,581 CITATIONS    134 PUBLICATIONS   4,261 CITATIONS   

SEE PROFILE SEE PROFILE

Mark Asta
University of California, Berkeley
334 PUBLICATIONS   12,372 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Mechanical Properties of Materials View project

First principles modeling of Ni-based superalloys View project

All content following this page was uploaded by Gautam Ghosh on 01 January 2018.

The user has requested enhancement of the downloaded file.


Available online at www.sciencedirect.com

Acta Materialia 56 (2008) 3202–3221


www.elsevier.com/locate/actamat

First-principles calculations of the structural and


thermodynamic properties of bcc, fcc and hcp solid solutions
in the Al–TM (TM = Ti, Zr and Hf) systems: A comparison
of cluster expansion and supercell methods
G. Ghosh a,*, A. van de Walle b, M. Asta c
a
Department of Materials Science and Engineering, Robert R. McCormick School of Engineering and Applied Science,
Northwestern University, 2220 Campus Drive, Evanston, IL 60208-3108, USA
b
Engineering and Applied Science Division, California Institute of Technology, Pasadena, CA 91125, USA
c
Department of Chemical Engineering and Materials Science, University of California, Davis, CA 95616, USA

Received 1 January 2008; received in revised form 1 March 2008; accepted 2 March 2008
Available online 7 April 2008

Abstract

The thermodynamic properties of solid solutions with body-centered cubic (bcc), face-centered cubic (fcc) and hexagonal close-packed
(hcp) structures in the Al–TM (TM = Ti, Zr and Hf) systems are calculated from first-principles using cluster expansion (CE), Monte-
Carlo simulation and supercell methods. The 32-atom special quasirandom structure (SQS) supercells are employed to compute prop-
erties at 25, 50 and 75 at.% TM compositions, and 64-atom supercells have been employed to compute properties of alloys in the dilute
concentration limit (one solute and 63 solvent atoms). In general, the energy of mixing (DmE) calculated by CE and dilute supercells agree
very well. In the concentrated region, the DmE values calculated by CE and SQS methods also agree well in many cases; however, note-
worthy discrepancies are found in some cases, which we argue originate from inherent elastic and dynamic instabilities of the relevant
parent lattice structures. The importance of short-range order on the calculated values of DmE for hcp Al–Ti alloys is demonstrated. We
also present calculated results for the composition dependence of the atomic volumes in random solid solutions with bcc, fcc and hcp
structures. The properties of solid solutions reported here may be integrated within the CALPHAD formalism to develop reliable ther-
modynamic databases in order to facilitate: (i) calculations of stable and metastable phase diagrams of binary and multicomponent sys-
tems, (ii) alloy design, and (iii) processing of Al–TM-based alloys.
Ó 2008 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.

Keywords: Ab initio electron theory; Cluster expansion; Phase stability; Special quasirandom structure (SQS); Thermodynamics

1. Introduction erences therein). At present, the three most widely employed


approaches for this purpose include: (i) the cluster expan-
Over the past 20 years significant efforts have been sion (CE) method ([2–6], and references therein), (ii) super-
devoted to developing computationally efficient approaches cell methods, including both conventional supercells and
for calculating the properties of disordered alloy solid-solu- special-quasirandom structures (SQS) (e.g. [7–11]), and
tion phases within the predictive framework of electronic (iii) approaches based on the coherent potential approxima-
density-functional theory (DFT) (see, e.g., Ref. [1] and ref- tion (CPA) (e.g. [12–14], and references therein). The latter
two methods have found widespread application in calcula-
tions of the electronic structure of disordered alloys, and
*
Corresponding author. Tel.: +1 847 467 2595; fax: +1 847 491 7820. supercell methods are also commonly employed to analyze
E-mail address: g-ghosh@northwestern.edu (G. Ghosh). local atomic structures and variations in bond lengths with

1359-6454/$34.00 Ó 2008 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.actamat.2008.03.006
G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221 3203

chemical composition. A somewhat more demanding appli- tions with such relatively small periodic structures remain
cation of these first-principles-based methods involves cal- under discussion (see below).
culation of the thermodynamic properties of solid At present the only technique that allows equilibrium
solutions. For this purpose computational approaches must SRO and elastic relaxations to be modeled in a self-consis-
generally be capable of modeling effects on mixing energies tent manner, without a priori knowledge of the values of
and entropies resulting from varying degrees of configura- the SRO parameters, is the CE method. However, due to
tional short-range-order (SRO) (i.e. correlations in the the significant computational demands associated with
occupancies of neighboring atoms) and atomic displace- the construction of a well-converged CE, particularly for
ments associated with elastic relaxations. systems with large elastic relaxations, the SQS approach
For each of the three approaches listed above, calcula- has found growing use as a computationally more efficient
tions of the energetic and thermodynamic properties of framework for estimating the energetics of solid solutions.
solid solutions are most easily performed assuming random It is noted that in most applications of the SQS approach
atomic configurations, although each can be adapted to the effects of SRO are neglected. The present work presents
consider effects of SRO with varying degrees of approxima- a comparison between the predictions of CE and SQS
tion. In the CE approach, accounting for SRO requires methods in the application to face-centered cubic (fcc),
coupling the CE model for energetics with Monte-Carlo body-centered cubic (bcc) and hexagonal close-packed
(MC) simulations to compute the equilibrium state of (hcp) solid solutions in three Al–TM (TM = Ti, Zr and
order and the associated energy or free energy at a given Hf) systems. From a practical standpoint these systems
composition and temperature (see below). SRO effects are of interest for potential high-temperature structural-
can be modeled in a similar way within CPA-based materials applications. From an alloy-theory perspective
approaches, provided they are combined with the formal- these systems are interesting due to the known importance
ism of concentration waves (e.g. [12,15]) or the generalized of SRO and atomic relaxation effects [27–29,6] on their
perturbation method (e.g. [16,17]) and appropriate statisti- structural and thermodynamic properties. Thus, these sys-
cal mechanical methods (MC simulations or mean-field tems provide a critical test of the relative accuracy of the
models) to describe atomic correlations. Recently it has different computational approaches for modeling the prop-
also been demonstrated how SRO effects can be incorpo- erties of solid solutions. The Al–TM systems are also of
rated in total energy calculations using a non-local modifi- interest because their phase diagrams exhibit all three of
cation of the CPA [14]. If the equilibrium state of SRO is the most common metallic solid solution structures,
known (e.g. either from experiment or from independent namely bcc, fcc and hcp; the results presented in this work
CE or CPA calculations), periodic SQS supercells can be may thus be useful for future refinements of their assessed
constructed to reproduce non-zero values of the SRO phase diagrams. The comparison between SQS and CE
parameters (e.g. [8]) out to several neighbor shells, and methods presented here complements earlier similar com-
used to compute associated effects on alloy energetics. parisons between CPA, CE and supercell methods, where
Modeling of elastic relaxation energies, resulting from good agreement between results obtained with these differ-
atomic displacements away from ideal lattice positions, ent methods were found in calculations where the effects of
represents a significant challenge for all of the methods SRO and elastic displacements were neglected [30–32].
described above. In the CE method, difficulties arise due The remainder of this paper is organized as follows. In
to the long-range nature of the elastically mediated interac- Section 2 we describe an integrated approach to modeling
tions resulting from atomic displacements (e.g. [18–20]), phase stability in multicomponent and multiphase systems.
giving rise to slowly converging expansions, and requiring In Section 3 the previously reported experimental and cal-
relatively large numbers of structural energy calculations culated thermodynamic data for relevant solid solutions is
to achieve acceptable accuracy. To circumvent this prob- briefly reviewed. In Section 4 the computational methodol-
lem several alternative formulations have been proposed ogies used in this study are described, and in Section 5 cal-
[4,21–23] for modeling elastically mediated interactions culated properties for fcc, bcc and hcp solid solutions are
explicitly, as a framework for improving the convergence presented. In Section 6 we discuss the results, including
properties of the CE approach for systems with large elastic comparisons of: (i) the data obtained with the different
relaxations (e.g. systems with large size mismatch). Within methods, (ii) first-principles calculated and experimental
the CPA-based methods no self-consistent calculation of data, and (iii) present and previous calculated results.
elastic relaxation energies has been demonstrated to date, Based on our computational results, in Section 7 we briefly
and such effects are generally treated using separate theo- discuss the scope for further experimental measurements.
retical frameworks [24–26]. For SQS methods the effects Finally, conclusions are summarized in Section 8.
of elastic relaxations on the energetics of a solid solution
can be modeled in a straightforward manner by relaxing 2. Modeling phase stability in multicomponent and
the positions of the atoms and the geometry of the supercell multiphase systems: an integrated approach
in a first-principles calculation; while this approach is rela-
tively straightforward, details related to the optimal One of our goals is to model phase stability in multicom-
method for modeling the relaxation energy of solid solu- ponent, multiphase systems in order to facilitate computa-
3204 G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221

tional design, processing and rapid prototyping of materi- ability of sufficient experimental data to generate accurate
als. In this context, our approach to modeling phase stabil- free energy functions. In the absence of adequate experi-
ity and the properties of interest is shown schematically in mental data, and in cases when experimental data cannot
Fig. 1. This approach integrates conventional experimental be accessed for various reasons (e.g. sluggish kinetics or
data, along with results from a large number of ‘‘virtual metastability), first-principles calculations offer an effective
experiments” performed within the predictive computa- strategy for supplementing thermodynamic databases,
tional framework of electronic DFT. The motivation for thereby increasing the range of applicability of computa-
integrating computational data into the modeling frame- tional thermodynamic methods, while limiting the reliance
work is primarily driven by the recent tremendous upon extensive experimental measurements. In addition to
advances in solid-state theory, computer hardware, algo- thermodynamic properties, several physical and elastic
rithm, software and user interfaces. The increasing avail- properties may also be modeled within the CALPHAD for-
ability of well-developed DFT software (for reviews see malism [38].
Ref. [1,33]) and high-performance computer hardware In this paper, it is not our goal to demonstrate the cal-
has led to a rapidly decreasing cost of computational mod- culation of all the properties listed in Fig. 1. Rather, we
eling at a time when the cost of performing experimental present selected properties (energy of mixing, DmE, and
measurements has been steadily increasing. atomic volumes) of bcc, fcc and hcp solid solutions in
It is important to point out that due to the multicompo- Al–TM (TM = Ti, Zr and Hf) systems calculated from
nent and multiphase nature of engineering alloys, the com- first principles. Experimental thermodynamic data for
putational cost of direct application of accurate first- solid solutions is scarce for these systems, yet it is needed
principles methods (i.e. methods using quantum-mechani- to calculate stable and metastable phases diagrams for
cal calculations requiring only the atomic numbers of the the binary and multicomponent systems containing Al
constituent elements and their arrangement as input) to and TM (TM = Ti, Zr and Hf) elements. We also com-
the direct modeling of phase stability in these systems is pare the calculated properties with experimental data,
prohibitive. A convenient way to overcome this limitation where available. The present work supplements our ear-
is to employ the CALPHAD-based [34,35] approaches of lier publications where we have reported the energy of
computational thermodynamics [36] and kinetics [37] formation (DfE) of intermetallic compounds (stable,
where multicomponent, multiphase equilibria and diffusion metastable and virtual) [39], the entropy of formation
problems can be solved rather rapidly, requiring minimal of selected fcc-based phases [40], and the entropy of mix-
computational resources. The predictive power of CALP- ing of fcc solid solutions [40] in Al–TM (TM = Ti, Zr
HAD-based approaches, however, is limited by the avail- and Hf) systems.

Fig. 1. A schematic representation of modeling phase stability and cohesive properties in multicomponent and multiphase systems by integrating virtual
experiments using modern computational tools and conventional experiments: electronic density functional theory (DFT); heat of formation (DfH); heat
 1 ); heat of mixing (DmH); free energy of mixing (DmG); heat of transformation (DtrH); special quasirandom structure
of solution at infinite dilution (DH
(SQS); effective cluster interaction (ECI); cluster variation method (CVM); Monte-Carlo simulation (MCS); ordering/clustering (SRO/clus); Redlich–
Kister (R–K); compound energy formalism (CEF); stacking fault (SF); antiphase boundary (APB).
G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221 3205

3. Thermodynamic data of solid solution phases in Al–TM report of first-principles calculations for solid solution
(TM = Ti, Zr and Hf) systems: a brief review phases in the Al–Hf system.

3.1. Experimental: calorimetry, galvanic cell and related 3.3. CALPHAD modeling of phase diagrams
measurements
As summarized very recently by Schuster and Palm [41],
Experimental data for the Al–Ti system available prior CALPHAD modeling of phase equilibria in the Al–Ti sys-
to 2005 has been compiled by Schuster and Palm [41]. tem has been reported at least seven times [60–66]. Murray
Kubaschewski and Dench [42] performed direct reaction [67] pointed out that thermodynamic modeling of the Al–
calorimetry of Al–Ti alloys for the entire composition Ti system presents several difficulties due to, among other
range, and reported the heats of formation (DfH) for factors, the lack of adequate thermochemical data. Due
solid-phase alloys. Hoch and Usell [43] carried out Knud- to the varied amount of experimental data available at
sen effusion mass spectrometry, and reported the activity the time of the different CALPHAD assessments, and also
of components in bcc solid solutions at 1780 K. Samokhval inherent to the CALPHAD-based optimization process,
et al. [44] measured electromotive forces using galvanic large variations in reported thermodynamic model param-
cells, and reported the heat (DmH) and free energy of mix- eters are found.
ing (DmG) of hcp solid solutions at 960 K. All of these CALPHAD modeling of both Al–Zr [65,68] and Al–Hf
results were critically assessed by Hultgren et al. [45], and [69,70] systems have each been reported twice.
also by Desai [46]. More recently, Eckert et al. [47] carried
out Knudsen cell-mass spectrometry, and reported the free 4. Computational methodology
energy of mixing in bcc and hcp solid solutions at 1473 K.
Murray et al. assessed the experimental information for 4.1. Ab initio total energy calculations
the Al–Zr [48] and Al–Hf [49] systems. For both of these
systems no experimentally measured thermochemical data The ab initio calculations presented here are based on
is presently available for the solid solution phases. electronic DFT, and have been carried out using the VASP
(Vienna Ab initio Simulation Package) code [71–73],
3.2. First-principles calculations employing Vanderbilt-type ultrasoft pseudopotentials [74]
as implemented in VASP [75], and an expansion of the elec-
Asta et al. employed first-principles methods to investi- tronic wavefunctions in plane waves with a kinetic-energy
gate phase stability and thermodynamic properties of fcc- cutoff of 220 eV. All calculated results were derived
[27,28] and hcp-based [28] phases in the Al–Ti system. Spe- employing the generalized gradient approximation
cifically, they employed the full-potential linear muffin tin (GGA) due to Perdew and Wang [76]. Brillouin-zone inte-
orbital (FP-LMTO) method [50,51] within the local density grations were performed using Monkhorst–Pack [77] k-
approximation (LDA) [52] to calculate total energies of 11 point meshes, and the Methfessel–Paxton [78] technique
phases based on the fcc lattice and nine phases based on the with a smearing parameter of 0.1 eV. The total energy
hcp structure. From these results, effective cluster interac- was converged numerically to less than 1  106 eV/atom
tions (ECIs) in fcc and hcp CEs were computed using the with respect to electronic, ionic and unit cell degrees of
structure inversion method (SIM) [53]. Using ECIs as input freedom; the latter two were relaxed using calculated forces
to cluster variation method (CVM) [54] calculations, solid- with a preconditioned conjugate gradient algorithm. After
state phase diagrams were calculated. structural optimization, calculated forces (which were not
Rubin and Finel [55] employed the linear muffin tin identically zero by symmetry) were converged to less than
orbital atomic-sphere-approximation (LMTO-ASA) 5 meV/Å in magnitude. All calculations were performed
method [50,56] within the LDA [57,58] to calculate total using the ‘‘high” precision setting within the VASP code.
energies of a set of bcc-based phases in the Al–Ti system, Further computational details may be found elsewhere [39].
including pure elements, DO3-Ti3Al, B2-TiAl, B32-TiAl
and DO3-TiAl3. Once again, using the formation energies 4.2. Cluster expansion
of these phases, ECIs were derived by the SIM, although
calculated phase diagrams were not reported. In the present study the energetics of solid-solution
In the case of the Al–Zr system, only solid solutions based phases in the Al–TM (TM = Ti, Zr and Hf) alloy systems
on the fcc lattice have been investigated using first-principles are derived within the first-principles cluster expansion
methods by Clouet et al. [59]. They used the FP-LMTO framework [2–4] coupled with MC simulations [79,80].
method [50,51] within the LDA [52] to calculate total ener- This approach allows incorporation of contributions to
gies of 26 fcc-based phases, from which 17 ECIs were derived solid-solution energies arising from short-range order as
using the SIM. Clouet et al. also reported calculations of sta- well as elastic relaxations. The codes used to perform these
ble and metastable solid solubilities for Zr in Al. calculations are available as part of the alloy theoretic
With the exception of the previous work of the authors automated toolkit (ATAT) developed by one of the co-
[40], presented in further detail below, there is no previous authors and his collaborators [81,82].
3206 G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221

The CE [2] is a compact representation of the configura- same volume. Then, calculate the strain tensor eij mapping
tional dependence of an alloy’s energy. In the case of the the unit cell of the unrelaxed structure onto the unit cell of
binary alloy system studied here, the configuration of the the relaxed structure.PThe magnitude of the relaxation is
1=2
alloy system can be described by spin-like occupation vari- then defined as R ¼ ð ij e2ij Þ .
ables ri, which take values 1 or +1 depending on the type We have excluded from the cluster-expansion fits the
of atom occupying a given atomic site i. A particular energies of all structures whose relaxation magnitude R
arrangement of spins of the parent lattice is called a config- exceeds 9%. This threshold is motivated by the observation
uration (or structure) and can be represented by a vector r that, among the bcc and fcc structures considered, there is a
containing the value of the occupation variable rj for each noticeable gap in the distribution of values of R at around
site j in the parent lattice. The CE parametrizes the energy 9%. While a similar gap could not be located in the case of
E (per atom) as a polynomial in the occupation variables: the hcp structural energies, the same threshold was adopted
X   in this case as well. No detailed analysis of the effect of this
EðrÞ ¼ ma J a Pi2b ri a ð1Þ choice upon the CE predictions was conducted in the pres-
a
ent study since this approximation is expected to affect
where a is a cluster (a set of sites i, corresponding to pairs, most strongly the values of mixing energies calculated far
triplets, tetrahedra, etc.). The sum is taken over all clusters from the ranges of composition and temperature where
a that are inequivalent by a symmetry operation of the the solid-solution structures are thermodynamically stable.
space group of the parent lattice, while the average h  ia It is also extremely important to note that since our ‘‘struc-
is taken over all clusters b that are equivalent to a by sym- ture rejection rule” was based on the magnitude of the
metry. The coefficients Ja in this expansion embody the structural relaxations rather than, say, on how poorly the
information regarding the energetics of the alloy and are CE fitted these structural energies, we preserve the validity
called ECIs. The multiplicities, ma, indicate the number of any statistical assessment of the accuracy of the model
of clusters that are equivalent by symmetry to a (divided (as will be described below).
by the number of lattice sites). The ECIs are obtained by Once the CE is constructed, it can be used to perform
a fit to a set of configurations whose energies have been cal- highly efficient lattice model MC simulations that provide
culated from first principles. The accuracy of the resulting heats of formation effects in solid solutions. These MC sim-
CE can be systematically improved by increasing the num- ulations were performed using the Easy Monte-Carlo Code
ber of clusters a in the expansion and/or the number of (EMC2) described in Ref. [86]. The free energies were
structural energies (i.e. the energies of various configura- obtained using thermodynamic integration [79,80]. In some
tions r) used in the fit. of the procedures described below, we also make use of the
A difficulty associated with the generation of the CE of high-temperature limit (random alloy) of the enthalpy,
Al–TM (TM = Ti, Zr and Hf) alloys was found to arise which is defined as
due to the presence of structural instabilities. As was first X jaj
pointed out by Craievich et al. [83,84] and Grimvall and EðrÞ ¼ ma J a ðhri iÞ ð2Þ
co-workers [85], for elemental transition metals with stable a

low-temperature close-packed structures the bcc structure where jaj denotes the number of sites in cluster a.
is typically mechanically unstable, and vice versa. Consis-
tent with this trend, in the structural optimization (i.e. 4.2.1. Estimated accuracy
the minimization of the total energy with respect to symme- The present calculations involve several approxima-
try-allowed distortions of the lattice vectors and atomic tions. The first are associated with the use of the general-
positions) of low-symmetry bcc- and fcc-based superstruc- ized gradient (GGA) [87,88] and pseudopotential
tures we have found extremely large relaxations away from approximations in the electronic structure calculations.
ideal geometries. Based upon a comparison between our previously pub-
Such large structural relaxations present a problem in lished formation energies for Al–TM systems with avail-
the application of the CE formalism, which partitions able calorimetry data for ordered intermetallic
microscopic alloy degrees of freedom according to configu- compounds [39], these errors are expected to be on the
rational states defined relative to undistorted parent lat- order of 1 kJ mol1. Further sources of approximation
tices. For example, a tetragonal fcc-based crystal involve the use of a cluster expansion to calculate energet-
structure may relax to form a crystal structure with a c/a ics of disordered solid solutions. One source of uncertainty
ratio corresponding to the bcc lattice; in such a case the associated with this approach is the level of convergence
structural energy should formally be included within the realized in the expansion. This can be estimated using the
set of allowed bcc configurations rather than as an fcc cross-validation [82] score listed in Table 3. The cross-val-
superstructure. To avoid such ambiguities in the present idation score is defined as
approach we quantify the magnitude of the relaxations
!1=2
experienced by a given structure according to a metric 1X n
defined as follows. First, scale the relaxed and unrelaxed CV ¼ b ðiÞ Þ2
ðEi  E ð3Þ
n i¼1
structures isotropically so that their unit cells have the
G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221 3207

where Ei denotes the calculated energy of the ith structure MC sampling is also sometimes used to generate large
while Eb ðiÞ denotes the energy of structure i predicted from a SQSs [9,11].
CE fit to the remaining n  1 energies E1, . . . , Ei1, Ei, Tables 1–3 list the lattice vectors and the ideal positions
Ei+1 . . . En. The CV score, which is a quantity analogous of atoms in 32-atom SQS supercells employed in this work.
to the root mean square error, is specifically designed to Starting with these ideal SQS structures, relaxations are
provide a measure of the predictive power of the least- carried out in the following three steps: (i) step I involves
squares fit for data not included in the fit (such as the only volume relaxation, i.e. without altering the atomic
enthalpies of fully disordered solid solutions). This gives positions (keeping them fixed at the ideal lattice positions
estimates of uncertainties listed in Table 4.
While the use of MC simulations involves statistical
sampling errors, these errors can be made negligible com- Table 1
pared to the errors attributable to the fit of the CE. In par- Structural data (ideal positions) of 32-atom bcc-SQS at 25 and 50 at.% B,
ticular, as reported in Table 5, the MC simulations were with a0 being the lattice parameter of the bcc unit cell, used in this study
carried out using a number of independent samples suffi- A3B AB
cient to converge the average concentrations and energies Lattice vectors
to within 0.1%, an error that is imperceptible on the scale aX = [2, 0, 1]a0 aX = [2, 0, 1]a0
of our calculated energy curves. A likely more significant aY = [0, 1, 2]a0 aY = [0, 1, 2]a0
source of error is the fact that we have neglected non-con- aZ = [0, 3, 2]a0 aZ = [0, 3, 2]a0
figurational sources of entropy, such as lattice vibrations Atomic coordinates
4 12 8 24
and electronic excitations, in our calculations of finite-tem- A – ð0:0; 0:0; 0:0Þ; A – ð0:0; 32 ; 32Þ; A – ð0:0; 32 ; 32Þ; A – ð0:0; 20 28
32 ; 32Þ;
8 24
A – ð0:0; 32 ; 32Þ; A – ð0:0; 10 ; 28
32 32Þ; A – ð0:0; 28 ; 20
Þ; A – ð 1 7 29
; ; Þ;
perature properties. This introduces systematic errors in 32 32 4 32 32
A – ð0:0; 16 16
32 ; 32Þ; A – ð0:0; 32 ; 32Þ;
24 8
A – ð14 ; 15 21 1 6
32 ; 32Þ; A – ð2 ; 32 ; 32Þ;
2
the temperature dependence of the calculated free energies, A – ð0:0; 28 20
Þ; ð 1 2 22
A – ð12 ; 14 26 1 22 18
;
32 32 A – 2 32 ; 32Þ;
; ;
32 32 Þ; A – ð2 32 32Þ;
; ;
enthalpies and entropies, which has been observed to result A – ð12 ; 32
6 2
; 32 Þ; A – ð34 ; 32 5 7
; 32 Þ; A – ð14 ; 23 ;
32 32
13
Þ; A – ð34 ; 32
5 7
; 32 Þ;
in an overestimation of the temperature scale in numerous A – ð14 ; 32
7 29
; 32Þ; A – ð14 ; 11 32 32Þ;
; 9
A – ð34 ; 13 ;
32 32
31
Þ; A – ð 3 9 19
4 32 32Þ;
; ;
systems (see, e.g., Refs. [89,90], and references therein). As A – ð34 ; 13 31 1 14 26
32 ; 32Þ; A – ð2 ; 32 ; 32Þ; A – ð12 ; 26 30 1 27 25
32 ; 32Þ; A – ð4 ; 32 ; 32Þ;
A – ð14 ; 15 ; 21
Þ; A – ð1 18 6
2 ; 32 ; 32Þ; A – ð34 ; 29 ; 15
Þ; A – ð 1 31 5
4 ; 32 ; 32Þ;
a consequence, the effects of SRO upon calculated solid- 32 32 32 32
A – ð14 ; 19 ;
32 32
1
Þ; A – ð3 21 23
4 ; 32 ; 32Þ; B – ð0:0; 0:0; 0:0Þ; B – ð0:0; 32 4 12
; 32Þ;
solution energies in this work are expected to represent A – ð12 ; 22 18
Þ; ð3 25 3
B – ð0:0; 12 4 16 16
;
32 32 A – 4 ; 32 ; 32Þ; ;
32 32 Þ; B – ð0:0; 32 32Þ;
;
upper bounds (since the magnitude of SRO effects gener- A – ð12 ; 26 ;
32 32
30
Þ; A – ð1 27 25
4 ; 32 ; 32Þ; B – ð0:0; 24 8 1 3 17
32 ; 32Þ; B – ð4 ; 32 ; 32Þ;
ally increases with decreasing temperature). A – ð34 ; 29 ;
32 32
15
Þ; A – ð1 31 5
4 ; 32 ; 32Þ; B – ð14 ; 11 ;
32 32
9
Þ; B – ð1 19 1
4 32 ; 32Þ;
;
B – ð0:0; 12 ;
32 32
4
Þ; B – ð14 ; 32
3 17
; 32Þ; B – ð12 ; 32
2 22
; 32Þ; B – ð12 ; 10 14
32 ; 32Þ;
B – ð14 ; 23 ; 13
Þ; B – ð1 10 14
2 32 ; 32Þ;
; B – ð12 ; 18 ; 6
Þ; B – ð1 30 10
2 32 32Þ;
; ;
4.3. Supercell methods 32 32 32 32
B – ð12 ; 15 10 3 1 27
16 ; 32Þ; B – ð4 ; 32 ; 32Þ; B – ð34 ; 32
1 27
; 32Þ; B – ð34 ; 17 11
32 ; 32Þ;
B – ð34 ; 32
9 19
; 32Þ; B – ð34 ; 17 11
32 ; 32Þ B – ð34 ; 21 ;
32 32
23
Þ; B – ð3 25 3
4 32 32Þ
; ;
The CE provides a complete description of an alloy’s
energetics (for a given lattice) and, in particular, of the fully
disordered state, via Eq. (2). However, when one only
Table 2
wishes to obtain the energetics of the fully disordered (i.e. Structural data (ideal positions) of 32-atom fcc-SQS at 25 and 50 at.% B,
random) configurational state, so-called SQSs can be used with a0 being the lattice parameter of the fcc unit cell, used in this study
to directly obtain the properties at a given composition A3B AB
from a single supercell calculation.
Lattice vectors
The SQS approach also relies on the existence of an aX = [0.5, 0.5, 1]a0 aX = [0.5, 0.5, 1]a0
expression of the form of Eq. (1) and assumes that truncat- aY = [1.5, 1, 1.5]a0 aY = [1.5, 1, 1.5]a0
ing the sum to a finite number of terms still provides an aZ = [2, 2, 0]a0 aZ = [2, 2, 0]a0
accurate parametrization the energy of the alloy. Under Atomic coordinates
this assumption, the energy of any structure having the A – ð0:0; 0:0; 0:0Þ; A – ð0:0; 0:0; 32 8
Þ; A – ð0:0; 0:0; 0:0Þ; A – ð0:0; 0:0; 32 8
Þ;
same correlations hPi2b ri ia as the disordered state (for all A – ð0:0; 0:0; 16 32 Þ; A – ð0:0; 0:0; 24
32Þ; A – ð0:0; 0:0; 16 32 Þ; A – ð0:0; 0:0; 24
32 Þ;
clusters a actually included in the sum) should provide an A – ð48 ; 0:0; 12 4
32Þ; A – ð8 ; 0:0; 32Þ;
20
A – ð48 ; 0:0; 32 4
Þ; A – ð48 ; 0:0; 20 32Þ;

accurate estimate the energy of the disordered state. A A – ð48 ; 0:0; 28 32 Þ; A – ð 3 2 1


;
8 8 32; Þ; A – ð48 ; 0:0; 28 1 3 3
32Þ; A – ð8 ; 4 ; 32Þ;
A – ð68 ; 48 ; 32
2
Þ; A – ð28 ; 48 ; 32 6
Þ; A – ð58 ; 34 ; 32
7
Þ; A – ð68 ; 24 ; 10 32Þ;
SQS attempts to match as many correlations as possible 3 2 9 7 2 13
A – ð8 ; 8 ; 32Þ; A – ð8 ; 8 ; 32Þ; A – ð28 ; 24 ; 14 5 3 15
32Þ; A – ð8 ; 4 ; 32Þ;
with a structure having as few atoms in the unit cell as pos- A – ð28 ; 48 ; 14 5 6 15
32Þ; A – ð8 ; 8 ; 32Þ; A – ð58 ; 34 ; 23 3 1 25
32Þ; A – ð8 ; 4 ; 32Þ;
sible. SQSs are traditionally generated, for a given unit cell A – ð38 ; 28 ; 17
32 Þ; A – ð6 4 18
8 ; 8 ; 32Þ; A – ð78 ; 14 ; 29
32 Þ; A – ð2 2 30
8 ; 4 ; 32Þ;
size, by enumerating every possible supercell of that size A – ð8 ; 8 ; 32Þ; A – ð28 ; 48 ; 22
1 6 19
32Þ; B – ð8 ; 0:0; 32Þ; B – ð18 ; 34 ; 19
4 12
32Þ;

and every possible configuration within that unit cell in A – ð58 ; 68 ; 23 3 2 25


32Þ; A – ð8 ; 8 ; 32Þ; B – ð18 ; 34 ; 27 2 2 6
32Þ; B – ð8 ; 4 ; 32Þ;
A – ð18 ; 68 ; 27 7 2 29
32Þ; A – ð8 ; 8 ; 32Þ; B – ð28 ; 24 ; 22 3 1 1
32Þ; B – ð8 ; 4 ; 32Þ;
search of the structure having as many pair correlations A – ð28 ; 48 ; 30 Þ; ð5 6 31
B – ð38 ; 14 ; 32
9
Þ; B – ð38 ; 14 ; 17
32 A – 8 ; 8 ; 32Þ; 32Þ;
agreeing with the ones of the disordered state as possible. B – ð8 ; 0:0; 32Þ; B – ð18 ; 68 ; 11
4 4
32Þ; B – ð18 ; 34 ; 11 5 3 31
32Þ; B – ð8 ; 4 ; 32Þ;
When multiple configurations match the same number of B – ð18 ; 68 ; 32
3
Þ; B – ð58 ; 68 ; 327
Þ; B – ð68 ; 24 ; 32
2
Þ; B – ð68 ; 24 ; 1832Þ;
pair correlations, matching multibody correlations of the B – ð68 ; 48 ; 10 6 4 26
32Þ; B – ð8 ; 8 ; 32Þ; B – ð8 ; 4 ; 32Þ; B – ð78 ; 14 ; 32
6 2 26 5
Þ;
disordered state can be used to identify the best SQS. B – ð78 ; 28 ; 32
5
Þ; B – ð78 ; 28 ; 21
32Þ B – ð78 ; 14 ; 13
32 Þ; B – ð7 1 21
8 4 32Þ
; ;
3208 G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221

Table 3 AlTM compositions, the relevant PCFs are identical up


Structural data (ideal positions) of 32-atom hcp-SQS at 25 and 50 at.% B, to the 4-NN distance, and (iii) for the hcp-SQS at Al3TM
with a0 being the lattice parameter and a c/a ratio of 1.64605 of the hcp
unit cell, used in this study
(also AlTM3) the relevant PCFs are identical up to the 4-
NN distance, while at AlTM composition the relevant
A3B AB
PCFs are identical up to the 5-NN distance. Equivalently,
Lattice vectors
pffiffiffi pffiffiffi the Warren–Cowley SRO parameter is zero within the
aX ¼ ½3:5; 3=2; pffiffi0a
ffi 0 aX ¼ ½3:5; 3=2; pffiffi0a
ffi 0
aY ¼ ½2:5; 3 3=2; 0a0 aY ¼ ½2:5; 3 3=2; 0a0
aforementioned first n-NN shells in each of the respective
aZ ¼ ½0; 0; 3:2921a0 aZ ¼ ½0; 0; 3:2921a0 cases.
The energetics of a solid solution in the dilute limit was
Atomic coordinates
A – ð0:0; 0:0; 0:0Þ; A – ð0:0; 0:0; 24Þ; A – ð0:0; 0:0; 0:0Þ; A – ð24 9 3
; 24 ; 0:0Þ; also investigated by placing a single isolated impurity in a
9
A – ð24 3
; 24 ; 0:0Þ; A – ð18 6
24 ; 24 ; 0:0Þ; A – ð18 ;
24 24
6
; 0:0Þ; A – ð21 15
24 24 0:0Þ;
; ; sufficiently large supercell (to minimize interactions
3 9 12 12
A – ð24 ; 24 ; 0:0Þ; A – ð24 ; 24 ; 0:0Þ; 9
A – ð24 6 2
; 24 ; 4Þ; A – ð24 2 22 1
; 24 ; 4Þ; between the periodic images of the dilute impurity). For
A – ð21 15 2 22 1
24 ; 24 ; 0:0Þ; A – ð24 ; 24 ; 4Þ; A – ð11 ; 1 3
24 24 4 ; Þ; A – ð12 12 2
24 ; 24 ; 4Þ; this purpose, we have used 64-atom supercells, based on
2 22 3 3 9 2
A – ð24 ; 24 ; 4Þ; A – ð24 ; 24 ; 4Þ; A – ð17 19 3 2 22 3
24 ; 24 ; 4Þ; A – ð24 ; 24 ; 4Þ;
5 7 3 6 18 2
bcc, fcc and hcp lattices, containing one impurity atom.
A – ð24 ; 24 ; 4Þ; A – ð24 ; 24 ; 4Þ; A – ð20 ; 4 1
; Þ; A – ð20 4 3
24 ; 24 ; 4Þ;
8 16 1
A – ð24 ; 24 ; 4Þ; A – ð248 16 3
; 24 ; 4Þ;
24 24 4
A – ð21 15 2 5 7 3 The dilute supercells consisted of 4  4  4 supercells of
24 ; 24 ; 4Þ; A – ð24 ; 24 ; 4Þ;
9
A – ð24 3 2
; 24 ; 4Þ; A – ð11 1 3
24 ; 24 ; 4Þ; A – ð23 ; 13 3
24 24 4 ; Þ; A – ð 6 18 2
24 24 ; 4Þ;
; the corresponding primitive unit cell of bcc or fcc, and a
14 10 1 14 10 3
A – ð24 ; 24 ; 4Þ; A – ð24 ; 24 ; 4Þ; B – ð0:0; 0:0; 24Þ; B – ð24 3
; 249
; 0:0Þ; 4  4  2 supercell of hcp. In the case of dilute supercells
A – ð17 19 3 18 6 2
24 ; 24 ; 4Þ; A – ð24 ; 24 ; 4Þ; B – ð12 ;
24 24
12
; 0:0Þ; B – ð 6 18
24 24 ; 0:0Þ;
; of bcc and fcc, a k-point mesh used was 7  7  7, while
A – ð20 ; 4 1
; Þ; A – ð20 4 3
24 ; 24 ; 4Þ; B – ð15 21 3 9 2
24 ; 24 ; 0:0Þ; B – ð24 ; 24 ; 4Þ;
24 24 4 a k-point mesh 6  6  6 was used for the hcp supercell.
A – ð21 ; 15 2
24 24 4 ; Þ; A – ð23 13 3
24 ; 24 ; 4Þ;
5
B – ð24 7 1
; 24 ; 4Þ; B – ð248 16 1
; 24 ; 4Þ;
6 18
B – ð24 ; 24 ; 0:0Þ; B – ð15 21 8 16 3
B – ð24 ; 24 ; 4Þ; B – ð11 1 1 In the energy calculations for the dilute supercells, full
24 ; 24 ; 0:0Þ; 24 ; 24 ; 4Þ;
5
B – ð24 7 1
; 24 ; 4Þ; B – ð11 ; 1 1
24 24 ; 4Þ; B – ð14 ; 10 1
24 24 4 ; Þ; B – ð14 10 3
24 24 ; 4Þ;
; geometry relaxation was performed.
B – ð12 12 2 15 21 2
24 ; 24 ; 4Þ; B – ð24 ; 24 ; 4Þ; B – ð17 19 1 18 6 2
24 ; 24 ; 4Þ; B – ð24 ; 24 ; 4Þ;
B – ð17 ; 19 1
24 24 4 ; Þ; B – ð23 13 1
24 ; 24 ; 4Þ B – ð15 ; 21 2
24 24 4 ; Þ; B – ð23 13 1
24 ; 24 ; 4Þ 5. Computational results

In the process of developing a well-converged CE it is


for the given parent structure); (ii) step II involves both important to verify that the fitted expansion is sufficiently
volume and ionic relaxations, keeping the shape of the unit accurate so that the predicted ground states (according to
cell fixed with the geometry corresponding to the ideal lat- the energies predicted from the ECIs) agree with the actual
tice structure; and (iii) step III involves full relaxation, i.e. ground states (according to the first-principles energies). In
all degrees of freedom (volume, internal atom positions Fig. 2, this verification procedure is illustrated for the case
and unit-cell shapes) are relaxed. The k-point meshes for of the fcc Al–Zr system. The open symbols indicate the
bcc-, fcc- and hcp-SQS supercells were 6  6  4, 9  5  structures whose energy is known from first principles
4 and 5  5  6, respectively. and that were used for CE construction, while the crosses
Since we are dealing with binary alloys, it is instructive correspond to a database of additional structures generated
to investigate the deviations from ideal SQS structures,
after each relaxation step I–III, in terms of partial pair cor-
relation functions (PCFs), gij(r). The PCFs represent homo 0
and hetero coordinations, and also indicate the nature of
Formation Energy, eV/atom

chemical order in the alloy. The use of PCFs is well estab- -0.1
lished in the investigation of SRO in disordered systems.
The PCF is defined as -0.2

1 dN ðrÞ -0.3
gij ðrÞ ¼ ð4Þ
4pqr2 dr
where N(r) is the number of atoms of type j within a sphere -0.4
of radius r around a selected atom of type i, and q is the Predicted
bulk density of atoms of type j. The Al–Al, Al–TM and -0.5 Known str
TM–TM PCFs are calculated after the above-mentioned Known gs
-0.6
relaxation steps by selecting each atom and then the num- 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ber of atoms of type j in thin spherical shells with radii r Al Zr
Atomic Fraction Zr
and (r + dr) around these atoms. We have evaluated dNdrðrÞ
up to 7.8 Å with a step size of 0.05 Å. Fig. 2. A typical output of our cluster expansion (CE), showing the
The SQSs listed in Tables 1–3 have the following charac- formation energy of various structures used in CE of fcc Al–Zr alloys.
Diamonds denote the energies for which the ab initio energy is known.
teristics: (i) for the bcc-SQS at Al3TM (also AlTM3) and
The solid line defines the convex hull of these energies from which the
AlTM compositions, the Al–Al, Al–TM and TM–TM ground states (squares) can be identified. The crosses denote predicted
PCFs are identical up to the 6-NN (nearest neighbor) dis- energies for a large database of generated structures to verify that all
tance, (ii) for the fcc-SQS at Al3TM (also AlTM3) and ground states are found. The reference states are fcc Al and fcc Zr.
G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221 3209

Table 4
Characteristics of the cluster expansions (CEs)
System Number of structures used Number of atoms in the Number of clusters Cross-validation score (meV/atom or
lattice (excluded) unit cell (pair + triplet + quad.) nm3/atom)
Al–Ti
bcc 27 (45) 2–20 Energy: 8 + 9 + 1 52.7 (5.0 kJ/mol-atom)
Volume: 5 + 1 + 0 0.00083
fcc 78 (20) 4–32 Energy: 11 + 12 + 16 28.7 (2.7 kJ/mol-atom)
Volume: 6 + 5 + 6 0.00011
hcp 39 (9) 2–18 Energy: 8 + 0 + 0 19.9 (1.9 kJ/mol-atom)
Volume: 6 + 0 + 0 0.00011
Al–Zr
bcc 18 (42) 2–16 Energy: 9 + 0 + 0 57.2 (5.5 kJ/mol-atom)
Volume: 3 + 0 + 0 0.00081
fcc 41 (15) 4–32 Energy: 8 + 7 + 1 22.9 (2.2 kJ/mol-atom)
Volume: 5 + 2 + 0 0.00015
hcp 55 (27) 2–20 Energy: 15 + 15 + 6 37.4 (3.6 kJ/mol-atom)
Volume: 8 + 4 + 1 0.00013
Al–Hf
bcc 24 (35) 2–21 Energy: 13 + 4 + 0 43.1 (4.1 kJ/mol-atom)
Volume: 1 + 4 + 0 0.00017
fcc 34 (6) 4–12 Energy: 8 + 7 + 1 29.1 (2.8 kJ/mol-atom)
Volume: 5 + 2 + 1 0.00011
hcp 30 (2) 2–12 Energy: 7 + 15 + 0 29.5 (2.8 kJ/mol-atom)
Volume: 5 + 2 + 0 0.00007
The number of structures is the number of first-principles energies used as an input to the fit. The number of atoms in the unit cell refers to the structures
used in the CE. The number of clusters represents the number of terms in the CE. The cross-validation score (Eq. (3)) is an estimator of the root mean
square error of the out-of-sample predictive accuracy of the CE.

for the purpose of checking that no new ground states are Figs. 6, 8 and 10 plot the mixing energies obtained by
predicted. The ability of the CE to reproduce the ground- CE, SQS method and dilute supercell calculations. The
state structures is confirmed by noting that none of the mixing energy (DmE) is defined as
crosses lie below the convex hull of the predicted energy  
Dm Eh ðxTM Þ ¼ Eh ðxTM Þ  ð1  xTM ÞEhAl þ xTM EhTM ð5Þ
of the known ground states (indicated by open boxes).
Table 4 describes the characteristics of the CEs gener- where Eh, EhAl , EhTM denote energies of solid solution, ele-
ated in this work. In addition to the number of structures, mental Al, elemental TM (Ti, Hf and Zr) in the same crys-
and the number of atoms in the unit cell of each of the tal structure h (bcc or fcc or hcp). In order to quantify the
structures used in the fit of the CE, we report the number effect of relaxation on SQS energetics, we consider the mix-
of structures whose energy was calculated but excluded ing energies obtained after each of the relaxation steps I–III
based on the relaxation magnitude criterion described in described in Section 4.3. Figs. 7, 9 and 11 compare the par-
Section 4.2. This is a measure of the system’s tendency to tial PCFs after each type of relaxation of SQS supercells at
relax away from the intended lattice. The cross-validation equiatomic composition. These PCF results provide graph-
score is calculated from Eq. (3) using only the nonexcluded ical representations of the effect of relaxations on the rear-
structures. Figs. 3–5 shows the effective cluster interactions rangement of atoms. For the purpose demonstration, Figs.
for all the CEs constructed in the present study. As 7, 9 and 11 represent only one case for each of the bcc-, fcc-
expected for a well-converged CE, the magnitudes of the and hcp-SQS calculations.
ECIs are found to decay with distance and with cluster size. Like formation energy, volume is also a scalar quantity.
It is noteworthy that the convergence of the ECIs is rela- Therefore, we have used the CE method to calculate the
tively slow: pair interactions beyond the 10th neighbor composition dependence of the atomic volume in Al–TM
shell are required to obtain the desired CV scores for some systems of interest. For this purpose, the structures used
of the systems, and a relatively large number of many-body to generate the CE of the energy are the same as those used
(triplet and four-body) clusters are required. for the CE of volume. In each case, starting with the set of
Table 5 indicates the input parameters of the MC simu- clusters used for the CE of energy, a set of expansion coef-
lations that were used to calculate heats of formation (i.e. ficients for volume is obtained by minimizing the CV score
formation energies). These MC simulations, which auto- for volume. The results are summarized in Table 4. It is
matically account for the presence of SRO, give the results seen that the CE coefficients for the volume correspond
shown in Figs. 12–14. For comparison, these figures also to clusters that are a subset of those required to obtain a
plot, as a point of reference, the energy for a completely well-converged CE for energy. These results demonstrate
random solid solution, i.e. without chemical SRO. that in general the number of clusters required to obtain
3210 G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221

70 70
Al-Ti BCC 60 Al-Zr BCC
60
50
50
40
40

ECI, meV
30
ECI, meV

30 20
20 10
10 0
-10
0
-20
-10 -30

40 FCC 60 FCC
30 50
20 40

ECI, meV
10 30
ECI, meV

20
0
10
-10
0
-20
-10
-30
-20
-40 -30

HCP HCP
50 60
40
40
30
ECI, meV
ECI, meV

20
20
0
10
0 -20

-10 -40

-20 -60
1 2 3 4 1 2 3 4 1 2 1 2 3 4 1 2 3 4 1 2
Pairs Triplets Quads Pairs Triplets Quads
d/d nn d/dnn
Fig. 3. Calculated ECIs vs. cluster diameter (normalized by nearest Fig. 4. Calculated ECIs vs. cluster diameter (normalized by nearest
neighbor distance) in bcc, fcc and hcp alloys of the Al–Ti system. neighbor distance) in bcc, fcc and hcp alloys of the Al–Zr system.

a well-converged CE depends on the nature of the physical 6. Discussion


property being expanded.
Fig. 15a–c shows the composition dependence of atomic 6.1. Comparison between CE and SQS methods
volume in random solid solutions of Al–Ti, Al–Zr and Al–
Hf alloys. Irrespective of the lattice, a negative deviation One of the main objectives of this work is to assess the
from an ideal behavior is seen in all three binary systems. relative abilities of the SQS and the CE methods to predict
This is also consistent with the trend of equilibrium volume the energetics of disordered solid solutions. It may be sur-
of Al–TM (TM = Ti, Zr and Hf) intermetallics presented prising at first that the two methods do not agree perfectly
in earlier publications [27,28,39,59], and also for the fcc since they rely on the same basic approximation of a trun-
[27,28] and hcp [28] solid solutions in the Al–Ti system. cated CE. The discrepancy can be attributed to two factors:
In general, a knowledge of the composition dependence (i) the CE and the SQS are typically not based upon a clus-
of atomic volume is important in the study of supersatu- ter expansion truncated to the same number of clusters;
rated metastable alloys, phase transformations and the and (ii) both methods neglect the effect of ‘‘larger” clusters
design of controlled misfit alloys. not included in the truncated CE. The first issue is in prac-
G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221 3211

20 0
CE
10 Al-Hf BCC -5 Dilute supercell BCC
SQS: volume-relaxed
0 -10 SQS: (volume+ion)-relaxed
-15 SQS: fully-relaxed
ECI, meV

-10
-20
-20
-25
-30
-30
-40

Energy of Mixing, kJ/mol-atom


CE
-50 Dilute supercell FCC
-5
SQS: volume-relaxed
-60 SQS: (volume+ion)-relaxed
-10
SQS: fully-relaxed
-15
30 FCC
-20
20 -25
ECI, meV

10
CE
Dilute supercell HCP
0 -5
SQS: volume-relaxed
-10 SQS: (volume+ion)-relaxed
-10 SQS: fully-relaxed

-20 -15

-30 -20
-25
-30
HCP 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
50
Al Ti
40 Atomic Fraction Ti

30 Fig. 6. Comparison of calculated energy of mixing by cluster expansion


and supercell methods of Al–Ti alloys: bcc (top), fcc (middle) and hcp
ECI, meV

20 (bottom) phase.
10
0
tice unavoidable because SQS are typically constructed by
-10
imposing a unit cell size and attempting to match as many
-20 correlations as possible under that constrain, rather than
1 2 3 4 1 2 3 4 1 2
Pairs Triplets Quads
the other way around (i.e. setting a target number of corre-
lations and scanning all unit cell sizes). Similarly, CEs are
d/d nn
typically built iteratively by adding clusters using the
Fig. 5. Calculated ECIs vs. cluster diameter (normalized by nearest cross-validation criterion to dictate the best clusters to
neighbor distance) in bcc, fcc and hcp alloys of the Al–Hf system. add, rather than imposing a certain number of clusters a
priori. Hence, in general, the clusters accounted for in the
CE and SQS methods will differ. The issue that clusters
Table 5 beyond a certain size are always neglected is unavoidable
Input parameters for the Monte-Carlo simulations and merely reflects that the goal is to obtain a reasonable
Lattice Cell size Target precision Chemical potential step approximation to the true disordered state with finite com-
(at.%) (meV) putational resources.
bcc 20  20  20 0.1 2 We observe that the CE results generally agree better
fcc 20  20  20 0.1 2 with the SQS energies obtained without cell shape relaxa-
hcp 21  21  12 0.1 2 tion than with the fully relaxed SQS energies. The SQS
The cell size reports the size of the simulation cell (expressed in multiples energies obtained after relaxing ion positions and cell vol-
of the primitive cell). The target precision indicates that the system was ume (but not cell shape) are arguably the most representa-
thermally equilibrated and that the thermodynamic functions were aver- tive of the true disordered state, since the latter does not
aged over a number of steps sufficient to provide the atomic composition
exhibit any macroscopic anisotropy in the cell shape relax-
of the system with an accuracy 0.1 at.% for any given externally imposed
chemical potentials. The chemical potential step is the increment used to ations. Other authors have previously argued that this
perform the numerical integration of the thermodynamic functions in the restricted amount of relaxation provides the best model
thermodynamic integration process. of the disordered state [91]. In this sense, the CE energies
3212 G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221

16 0
FCC-SQS at AlTi Al-Al CE
14 -5 Dilute supercell
BCC
Volume-relaxed Al-Ti
12 Ti-Ti -10 SQS: volume-relaxed
SQS: (volume+ion)-relaxed
10 -15 SQS: fully relaxed
8 -20
6 -25
4 -30
Partial pair correlation function, g (r)

2 -35
ij

Energy of Mixing, kJ/mol-atom


CE
FCC-SQS at AlTi Al-Al FCC
10 (Vol+ion)-relaxed
Al-Ti
-5 Dilute supercell
SQS: volume-relaxed
Ti-Ti -10 SQS: (volume+ion)-relaxed
8 SQS: fully-relaxed
-15
6
-20
4 -25
2 -30

FCC-SQS at AlTi CE
Fully-relaxed
Al-Al
-5 Dilute supercell HCP
10 Al-Ti SQS: volume-relaxed
Ti-Ti -10 SQS: (volume+ion)-relaxed
8 SQS: fully-relaxed
-15
6
-20
4 -25
2 -30
0 -35
0 1 2 3 4 5 6 7 8 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Interatomic Distance (r), Å Al Atomic Fraction Zr Zr

Fig. 7. Partial atomic pair correlation function, gij(r), in fcc SQS at Al–Ti: Fig. 8. Comparison of calculated energy of mixing by cluster expansion
after volume relaxation (top), after volume and ionic relaxations (middle), and supercell methods of Al–Zr alloys: bcc (top), fcc (middle) and hcp
and after full relaxations (bottom). (bottom) phase.

agree best with the most physically relevant of the SQS they relax away from the intended lattice (see first column
energies. of Table 4). This is probably the simplest advance warning
Although the disordered state energies predicted by the of such problems: after calculations of only a few small-cell
two methods sometimes differ substantially (notably, on structures, it becomes rapidly apparent that a large fraction
the Al-rich side of bcc Al–Ti and Al–Zr), they do so by a exhibit considerable cell shape change during relaxations.
predictable amount. The margin of error is entirely consis- This test only involves looking at cell shapes and is there-
tent with the CV score of the corresponding CE (see last fore easy to implement in practice. A second manifestation
column of Table 4 and Figs. 6, 8 and 10). of the problem is the large difference in the volume-relaxed
Perhaps not surprisingly, the largest errors occur in SQS energy (i.e. method I in the terminology of Section
composition ranges where the alloy does not adopt, in 4.3) and the two ions-relaxed SQS energies (methods II
equilibrium, the type of lattice considered (notably, the and III) for most Al-rich bcc SQS. This suggests that local
bcc lattice in the Al-rich region). The CE (and the SQS) Al-rich clusters of atoms exhibit substantial local relax-
for the bcc lattice in the systems considered are the most ations away from the bcc lattice. Interestingly, the global
difficult to employ because the bcc phase is only stable over cell shape relaxations do not appear to change the energy
a relatively narrow range of compositions and tempera- significantly in these cases. We conjecture that this is due
tures. In fact, even when the bcc phase is present in the to the fact that the different clusters of Al atoms have ran-
phase diagram, it represents a dynamic mixture of many domly oriented directions of preference for Bain path dis-
lower-symmetry variants that results in bcc symmetry only tortion that do not lead to a macroscopic cell shape change.
in an average sense [92–94]. Pure bcc Ti, Zr and Hf are all Further evidence of considerable local relaxations is
mechanically unstable at absolute zero, and the average bcc provided in the PCFs (see Fig. 9 for the bcc AlZr SQS)
symmetry only manifests itself thanks to entropy-driven and noting that both (volume + ion)-relaxed and fully
stabilization at high temperature. relaxed geometry have PCFs that bear no qualitative
The mechanical stability problem has a number of man- resemblance to the corresponding plot for the volume-
ifestations. First, in the construction of the bcc CE, a num- relaxed-only geometry. As a point of comparison, consider
ber of structures have to be eliminated from the fit because the PCFs of the fcc AlTi SQS (Fig. 7), where the PCFs of
G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221 3213

12 0
CE
BCC-SQS at AlZr Al-Al -5 BCC
10 Volume-relaxed Al-Zr
Dilute supercell
Zr-Zr -10 SQS: volume-relaxed
8 -15
SQS: (volume+ion)-relaxed
SQS: fully-relaxed
6 -20
4 -25
-30
Partial pair correlation function, g (r)

2 -35
ij

Energy of Mixing, kJ/mol-atom


CE
BCC-SQS at AlZr Al-Al Dilute supercell
FCC
6 (Vol+ion)-relaxed -5
Al-Zr SQS: volume-relaxed
5 Zr-Zr
-10 SQS: (volume+ion)-relaxed
SQS: fully-relaxed
4
-15
3
-20
2
1 -25

CE
BCC-SQS at AlZr Al-Al HCP
7 Fully-relaxed Dilute supercell
Al-Zr -5 SQS: volume-relaxed
6 Zr-Zr SQS: (volume+ion)-relaxed
5 -10 SQS: fully-relaxed
4
3 -15
2 -20
1
0 -25
0 1 2 3 4 5 6 7 8 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Interatomic Distance (r), Å Al Hf
Atomic Fraction Hf
Fig. 9. Partial atomic pair correlation function, gij(r), in bcc SQS at Al– Fig. 10. Comparison of calculated energy of mixing by cluster expansion
Zr: after volume relaxation (top), after volume and ionic relaxations and supercell methods of Al–Hf alloys: bcc (top), fcc (middle) and hcp
(middle), and after full relaxations (bottom). (bottom) phase.

volume-relaxed, (volume + ion)-relaxed and fully relaxed relations encode the energetic effects of local relaxations).
geometries are very similar. This is what we have attempted in the SQSs generated in
The lattice relaxation problem is not entirely limited to this work and, in that sense, our SQSs are perhaps more
bcc structures. The Al-rich portion of the hcp Al–Hf sys- suitable for highly relaxing alloys systems than SQSs
tem (Fig. 10) is also affected by this problem, as indicated matching only pair correlations of the disordered state.
by the poor agreement between the SQS and CE results One important limitation of SQSs is that they typically
as well as the considerable changes in the PCFs upon relax- are generated to model perfectly disordered solid solutions,
ations (Fig. 11). ignoring the presence of potential SRO. This approxima-
It is sometimes argued that SQS can provide a better tion is justified when the temperature range of interest is
representation of the disordered state than a CE because much higher than that of any order–disorder transitions,
the SQS’s low symmetry better samples the variety of local but rigorously testing that the temperature is ‘‘high
relaxations taking place in the disordered state, while the enough” is best done by constructing a CE and using it
CE ‘‘ extrapolates” the effect of relaxations from high-sym- as the Hamiltonian of a lattice–gas MC simulation.
metry structures. However, this argument assumes that the Accordingly, we have calculated the ‘‘ exact” heat of for-
CE only contains high-symmetry structures, which is not mation of these alloys at various temperatures by MC sim-
typically the case for well-converged CEs, as assessed by ulations and compared the results with those
the CV score. Here, we find that, except for cases where corresponding to the fully disordered alloy formation
the mechanical stability of the lattice is questionable, the energy obtained from the same CE in Figs. 12–14. For
two methods agree rather well and that the discrepancies the MC results, indicated by the symbols, the presence of
are consistent with the accuracy of the CE, as estimated concentration ‘‘gaps” in the data, and the appearance of
by the CV score. ‘‘cusps” in the energy vs. concentration indicate the pres-
Another consideration is that many SQS reported in the ence of long-range-ordered intermetallic phases which have
literature only attempt to match pair correlations. In the formed spontaneously in the simulations at concentrated
presence of substantial local relaxations, matching higher- compositions. Thus, for the temperatures which show these
order correlation is crucial (because these higher-order cor- features, only the uninterrupted sequences of symbols cor-
3214 G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221

16 0
14 HCP-SQS at AlHf Al-Al -5 CE BCC
Volume-relaxed Al-Hf -10 MCS at 1500 K
12 Hf-Hf -15
10 -20
8 -25
6 -30

Energy of Mixing, kJ/mol-atom


-35
Partial pair correlation function, g (r)

4
2 -40
ij

-45
HCP-SQS at AlHf -50
9 (Vol+ion)-relaxed
Al-Al
-55
8 Al-Hf
7 Hf-Hf CE
-5 HCP
6 MCS at 1000 K
5 -10 MCS at 1500 K
4 MCS at 2000 K
3 -15
2 -20
1 -25
11 HCP-SQS at AlHf Al-Al -30
10 Fully-relaxed
Al-Hf
9 -35
Hf-Hf
8 -40
7
6 -45
5 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
4 Al Atomic Fraction Ti Ti
3
2
1 Fig. 12. Comparison of calculated energy of mixing by cluster expansion
0 and Monte-Carlo simulations (MCS) of Al–Ti alloys: bcc (top) and hcp
0 1 2 3 4 5 6 7 8 (bottom) phase. The gaps in the top curve are indicative of first-order
Interatomic Distance (r), Å phase transformations on the bcc lattice. Therefore, fully disordered CE
results should only be compared with the leftmost and rightmost
Fig. 11. Partial atomic pair correlation function, gij(r), in hcp SQS at Al– uninterrupted portions of the MCS energy curves, corresponding to
Hf: after volume relaxation (top), after volume and ionic relaxations disordered solid solution potentially exhibiting SRO (but no LRO). While
(middle), and after full relaxations (bottom). the bottom curve (hcp) is uninterrupted (implying the absence of first-
order transitions), an analysis of fluctuations in energy and composition
(not shown) reveals peaks indicating that second-order transitions take
place at around x = 0.15, 0.38, 0.62, 0.84 for T = 1000 K and T = 1500 K.
responding to the most TM- or Al-rich compositions corre- Hence, the comparison with the fully disordered CE results is only
spond to disordered solid solutions. Specifically, for Al–Ti, meaningful for the composition intervals 0 < xTi < 0.15 and 0.84 < xTi < 1
the data points corresponding to xTi < 0.2 and xTi > 0.8 for T = 1000 K, 1500 K. For T = 2000 K a solid solution is stable (and the
have been confirmed to correspond to disordered solid- comparison meaningful) throughout the whole composition range.
solution phases; for Al–Zr all of the MC data for the bcc
structure, and results for xZr < 0.15 and xZr > 0.85 corre-
spond to hcp solid solutions; for Al–Hf solid solutions effects in the temperature range of interest. If this test suc-
are confirmed to be stable at xHf < 0.18 and xHf > 0.82 ceeds, the SQS energy can be considered reliable, but if the
for bcc and xHf < 0.05 and xHf > 0.52 for hcp structures. test fails, a properly converged CE must be constructed and
A comparison of the energetics for solid solutions with MC simulations are needed to obtain quantitative results.
SRO and those without (given by the solid lines in Figs.
12–14) shows clear quantitative differences which are as 6.2. Comparison between computational and experimental
large as 30–50% of the total magnitude of the mixing results
energy (cf. the results for bcc Zr and hcp Al–Hf) at the
most concentrated compositions. To check the accuracy of our calculated alloy energetics
Our findings thus demonstrate the potential significance it is useful to compare these with available calorimetric
of SRO effects in modeling the energetics of solid-solution data. Of the nine cases (i.e. three crystal structures and
phases. Clearly, SQS structures which model random three alloy systems) considered in our computations, mea-
atomic configurations should be used to estimate the mix- sured values of DfH are available only for hcp Al–Ti solid
ing energies of a solid solution phase only after validating solutions [42]. The formation energy (DfE) of a solid solu-
the accuracy of neglecting SRO effects. Without necessarily tion is calculated as
constructing a fully fledged CE, this could be straightfor-  
Df Eh ðxTM Þ ¼ Eh ðxTM Þ  ð1  xTM ÞE/Al þ xTM EuTM ð6Þ
wardly done, for instance, by fitting a simple nearest-neigh-
bor cluster expansion to an SQS formation energy and where Eh, E/Al , EuTM denote energies of solid solutions with
using MC simulations to quantify the importance of SRO the structure h, elemental Al with the structure /, and ele-
G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221 3215

0 0
-5 CE -5 CE BCC
BCC
MCS at 2000 K -10 MCS at 2000 K
-10
-15
-15 -20
-20 -25
-25 -30
Energy of Mixing, kJ/mol-atom

Energy of Mixing, kJ/mol-atom


-30 -35
-40
-35
-45
-40 -50
-45 -55
CE HCP HCP
-5 CE
MCS at 1500 K -5
-10 MCS at 1500 K
-15 -10
-20
-15
-25
-30 -20
-35
-25
-40
-45 -30
-50
-35
-55
-60 -40
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Al Atomic Fraction Zr Zr Al Atomic Fraction Hf Hf

Fig. 13. Comparison of calculated energy of mixing by cluster expansion Fig. 14. Comparison of calculated energy of mixing by cluster expansion
and Monte-Carlo simulations (MCS) of Al–Zr alloys: bcc (top) and hcp and Monte-Carlo simulations (MCS) of Al–Hf alloys: bcc (top) and hcp
(bottom) phase. The bcc solid solution (top) is stable over the whole (bottom) phase. Note that first order phase transitions take place on both
composition range as evidenced by the absence of gaps in the MCS results, the bcc and hcp lattice. Therefore, in both panels, only the leftmost and
thus making the comparison with fully random CE results possible at all rightmost uninterrupted portions of the MCS energy curves should be
compositions. The gaps in the bottom curve are indicative of first-order compared with the fully disordered CE results.
phase transformations on the hcp lattice. Therefore, only the leftmost and
rightmost uninterrupted portions of the MCS energy curves should be
compared with the fully disordered CE results. the available experimental data [42,44], which was later
critically assessed by others [45,46]. Kubaschewski and
Dench [42] performed direct reaction calorimetry, while
mental TM (Ti, Hf and Zr) with the structure u. Our com- Samokhval et al. [44] derived DfH indirectly based on gal-
putational results are summarized in Table 6, along with vanic cell measurements of electromotive forces (EMF)

Table 6
A comparison of calculated energy of formation (DfE) (both cluster expansion (CE) or random solid solution and Monte-Carlo simulation (MCS) at
1000 K to include the effect of SRO) and experimental (calorimetry) heat of formation (DfH, kJ/mol-atom) of hcp Al–Ti solid solution
At. Fr. Ti DfE from first-principles DfH from calorimetry
CE (random) MCS at 1000 K w/o vibration MCS at 1000 K with vibration
0.9145 8.03 ± 1.9 8.86 ± 1.9 8.75 ± 1.9 9.46 ± 1.0a
0.8355 14.06 ± 1.9 17.97 ± 1.9 17.36 ± 1.9 18.38 ± 1.0
0.8355 16.87 ± 1.0
0.7615 18.52 ± 1.9 26.48 ± 1.9 b 23.95 ± 1.0
0.6925 21.50 ± 1.9 29.68 ± 1.9 b 29.05 ± 1.0
0.6925 29.85 ± 1.5
0.9450 5.29 ± 1.9 5.69 ± 1.9 5.64 ± 1.9 8.08 ± 0.8c
0.9220 7.40 ± 1.9 8.07 ± 1.9 7.99 ± 1.9 10.68 ± 1.0
0.8990 9.32 ± 1.9 10.46 ± 1.9 10.33 ± 1.9 13.36 ± 1.3
0.8800 10.84 ± 1.9 12.49 ± 1.9 12.26 ± 1.9 14.23 ± 2.0d
0.8920 9.89 ± 1.9 11.22 ± 1.9 11.04 ± 1.9 12.00 ± 2.0e
The reference states are fcc Al and hcp Ti. Two MCS results are reported, one based on this work’s CE and one based on the CE from Ref. [6] that
accounts for lattice vibrations via temperature-dependent ECI.
a
Ref. [42].
b
Composition lies in a two-phase field [6].
c
Derived from galvanic cell data [44].
d
Critical assessment of experimental data [45].
e
Critical assessment of experimental data [46].
3216 G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221

using a solid electrolyte. Although Kubaschewski and is not entirely surprising. In fact, a number of physical
Dench [42] reported DfH at 298 K, it is important to note property, mechanical property and microstructure investi-
that direct reaction calorimetry was performed in the tem- gations by transmission electron microscopy (TEM) con-
perature range 1073–1123 K. Therefore, in Table 6 we have firm the presence of SRO in hcp Al–Ti alloys. These
included DfE calculated by MC simulations at 1000 K. In include electrical resistivity and magnetic susceptibility
considering the DfH data of Kubaschewski and Dench measurements [95–100], mechanical properties showing
[42], it is important to note two additional points, which transition from homogeneous slip to planar slip [101,102],
were also pointed out by Zhang et al. [64] in their thermo- and detailed analysis of stacking fault fringe contrast in
dynamic assessment of the Al–Ti system. First, Kubas- TEM micrographs [103,104].
chewski and Dench identified the structure of Al–Ti In comparing computational results with experimental
alloys, after direct reaction calorimetry, as being hcp for data, clear advantages of the CE (combined with MC sim-
compositions containing up to about 30 at.% Al; according ulations) over the SQS method are apparent. For example,
to recent equilibrium phase diagram assessments, some of in the CE method DmE or DfE can be calculated as a func-
these alloys should be either in a hcp + DO19 two-phase tion of composition in a continuous manner. By contrast,
or in the DO19 single-phase fields. Second, they assigned in the SQS method DmE or DfE are evaluated at discrete
an uncertainty of ±1 kJ/mol-atom for their DfH data; how- compositions, and as a result one has to rely on interpola-
ever, according to critical assessments of their data a more tion/extrapolation to compare the calculated results with
realistic uncertainty would be ±2 kJ/mol-atom [45,46]. We
note that the uncertainty in the measurements is compara-
ble in magnitude to the CV score (i.e. estimated precision)
0.0172 BCC Al-Ti
of ±1.9 kJ/mol-atom in our CE calculations. Atomic Volume, nm3 FCC
From the results listed in Table 6 it is clear that calcu- 0.0170
HCP
lated values of DfE, which include SRO effects (i.e. MC 0.0168
simulations at 1000 K), show a remarkably good agree-
0.0166
ment with calorimetry data, even if we restrict the compo-
sition range to Al concentrations less than 15 at.% Al. The 0.0164
importance of SRO effects is clear when a comparison is 0.0162
made between the MC results and calculations based on
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
the random-alloy approximation, which are considerably Al Atomic Fraction Ti Ti
smaller in magnitude and in poorer agreement with exper-
imental data. The data of Samokhval et al. [44] represent
derived (rather than directly measured) quantities and are 0.023 BCC Al-Zr
Atomic Volume, nm3

expected to contain larger uncertainties relative to the FCC


0.022
direct calorimetry results; nevertheless, these results are HCP
found to show a similar trend. 0.021
The differences between the results in the second and 0.020
third columns of Table 6 reflect the magnitude of SRO 0.019
effects on the calculated formation energies as predicted 0.018
by the MC simulations at 1000 K. It is important to note
0.017
that these SRO effects are generally temperature dependent
(becoming smaller in magnitude at higher temperatures) 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Al Atomic Fraction Zr Zr
and that a source of error in our MC calculations of forma-
tion energies thus arises due to our neglect of vibrational
contributions to the alloy energetics, which is known to
0.022 BCC Al-Hf
lead to a large overestimation of order–disorder transition
Atomic Volume, nm3

FCC
temperatures in hcp Ti–Al alloys [6]. To estimate the mag- 0.021 HCP
nitude of this error, we included vibrational contributions 0.020
to the ECIs based on the vibrational entropy calculations
presented in Ref. [6], and used them to perform additional 0.019
MC simulations, the results of which are listed in the fourth 0.018
column of Table 6. A comparison of these results with
0.017
those obtained neglecting vibrational contributions show
relatively small differences, and establish that the SRO 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Al Atomic Fraction Hf Hf
effects on the calculated formation energies are indeed
significant. Fig. 15. Atomic volume (at 0 K) in random solid solutions with bcc, fcc
The importance of including SRO effects to obtain good and hcp structures obtained from cluster expansion: (a) Al–Ti, (b) Al–Zr
agreement between calculated and measured values of DfE and (c) Al–Hf system.
G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221 3217

0.00 ordered phases and only the data at the lower and higher
-0.04 at 2000 K Ti compositions correspond to stable solid solutions.
at 1500 K
-0.08
sro parameter

at 1000 K
-0.12 6.3. Comparison between present and previous CE results,
-0.16 and CALPHAD modeling
-0.20
-0.24 Table 7 compares the DmE of random solid solutions
-0.28 calculated in this study and also those based on ECIs
-0.32 reported in previous work [55,59]. For the ease of compar-
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 ison, calculated values are listed at 25, 50 and 75 at.% TM
Al Atomic Fraction Ti Ti (TM = Ti, Hf and Zr).
In the case of bcc Al–Ti alloys, our values differ by as
Fig. 16. Calculated short-range order (SRO) parameter in hcp Al–Ti solid
solution as a function of temperature. Note that for T = 1000 and 1500 K much 10 kJ/mol-atom from those obtained in previous
the alloy is actually long-range ordered (LRO) in the composition interval work by Rubin and Finel [55]. In the previous work, a rel-
[0.15, 0.84] (see Fig. 12 caption). The cusps in the SRO in fact point to atively limited CE was considered: six structures were con-
order phases in this case. Note that no LRO takes place at T = 2000 K. sidered, and used to derive two pairs, a triplet and a four-
body ECI. By comparison, in the present work the ECIs
experimental data, which can limit the accuracy of the cal- are derived from 27 structures, considering fairly large unit
culated results. Additionally, through thermodynamic inte- cells (see Table 4) and relativley long-ranged ECIs. The dis-
gration methods, the CE approach can give access to free crepancies between the two sets of calculated results prob-
energy (and entropy) as well as enthalpy results, which ably stems from a better convergence of the CEs in the
facilitates direct comparisons to data derived from EMF present work. It is noteworthy that for the results in Table
measurements (e.g. [6]). 7, the DmE is skewed towards the Al-side in the previous
Fig. 16 shows the calculated SRO parameter as a func- calculations, while it is almost symmetric in the present
tion of temperature in hcp Al–Ti solid solution. As dis- study. In the case of fcc Al–Ti alloys, our values agree well
cussed in the previous section, the data for the most with those reported by Asta et al. [27,28] although they dif-
concentrated compositions correspond to long-range- fer somewhat for Al-rich compositions. In the case of hcp

Table 7
A comparison between calculated (both cluster expansion (CE) and SQS methods) and CALPHAD assessments of heat of mixing (DmH, kJ/mol-atom) of
bcc, fcc and hcp solid solutions in Al–TM (TM = Ti, Zr and Hf) systems
Lattice At. Fr. TM Al–Ti Al–Zr Al–Hf
CEa SQSb
Calphad c
CE a
SQSb
Calphad d
CEa SQSb Calphade
bcc 0.25 20.5 25.9 21.8 to 34.8 20.5 29.6 23.2 to 33.6 25.7 26.9 27.8 to 32.0
30.7f
0.50 26.9 33.0 29.1 to 34.9 27.4 36.2 30.5 to 41.3 33.6 33.9 36.2 to 37.0
32.8f
0.75 21.3 22.3 21.8 to 24.3 20.5 22.8 21.2 to 26.8 24.6 23.8 22.6 to 27.8
17.7f
fcc 0.25 19.3 20.4 18.9 to 30.7 23.8 21.4 22.7 to 28.7 17.3 17.9 21.0
25.7g 20.5h
0.50 25.7 26.7 25.3 to 38.6 27.6 27.7 30.0 to 38.2 18.9 23.6 28.0
27.4g 29.2h
0.75 20.7 19.5 15.1 to 23.3 17.8 19.4 20.9 to 28.7 11.8 17.1 21.0
18.3g 24.2h
hcp 0.25 19.9 20.8 21.7 to 32.3 25.5 22.5 22.9 to 33.8 15.2 18.2 28.6 to 29.2
21.3g
0.50 26.6 27.6 30.0 to 37.6 28.5 33.7 30.6 to 41.6 17.5 22.2 32.4 to 38.1
26.8g
0.75 19.9 18.1 22.5 to 27.8 17.9 18.9 21.4 to 26.9 10.9 12.9 19.8 to 28.6
20.2g
a
Random solid solution.
b
After volume and ionic relaxations.
c
Based on CALPHAD modeling [60–66].
d
Based on CALPHAD modeling [65,68].
e
Based on CALPHAD modeling [69,70].
f
Calculated using the ECIs reported in Ref. [55].
g
Ref. [28].
h
Calculated using the ECIs reported in Ref. [59].
3218 G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221

Al–Ti alloys, our values agree well with those reported by Table 8
Asta et al. [28] over the entire composition range. In the Ordering energies (DEOrd, kJ/mol-atom), from random solid solutions
(SS) to ordered phases, calculated from first-principles
case of fcc Al–Zr alloys, our values show only a modest
agreement with those reported Clouet et al. [59], and the Ordering At. Fr. TM DEOrd in Al–TM systems
composition dependence of DmE differ markedly. While Al–Ti Al–Zr Al–Hf
our results show that DmE is skewed towards the Al-rich bcc(SS) ? DO3 0.25 14.34 12.94 18.88
side, the results of Clouet et al. [59] suggest the opposite 23.58a
trend. bcc(SS) ? B2 0.50 8.26 10.05 0.95
7.37a
The results derived from CALPHAD modeling in Table bcc(SS) ? DO3 0.75 2.77 0.91 6.22
7 feature a large range of mixing energies, which is primar- 6.06a
ily due to lack experimental DmH data for the solid solution fcc(SS) ? L12 0.25 18.53 23.44 21.12
phases. As a result, the energy and entropy contributions to 15.54b 30.74c
Gibbs energy could not be evaluated uniquely from the fcc(SS) ? DO22 0.25 20.85 23.57 22.99
17.66b 30.79c
experimental phase boundary data, leading to a wide range fcc(SS) ? DO23 0.25 21.61 26.13 23.98
of estimates for the different thermodynamic assessments. 33.00c
fcc(SS) ? L10 0.50 16.56 18.90 19.60
6.4. Ordering energies 17.46b 19.49c
fcc(SS) ? L12 0.75 9.10 15.71 17.36
13.86b 11.42c
Since we have calculated the formation energies for both fcc(SS) ? DO22 0.75 7.85 12.09 14.85
disordered solid solutions and ordered intermetallic com- 11.26b 6.66c
pounds, it is possible to calculate values for the ordering fcc(SS) ? DO23 0.75 8.46 13.82 15.99
energies (DEOrd), defined as the difference between the for- 8.88c
mation energy of an ordered phase (AlmTMn with structure hcp(SS) ? DO19 0.25 13.19 16.58 18.18
15.29b
/) and the random solid solution (h) with the same parent hcp(SS) ? B19 0.50 14.34 18.03 19.33
lattice structure at the same composition: hcp(SS) ? DO19 0.75 7.31 12.58 12.76
10.22b
DEOrd ¼ Df E/ ðAlm TMn Þ  Df Eh ðCE at Alm TMn Þ ð7Þ
a
Derived from the parameters in Ref. [55].
The formation energies on the right-hand side of Eq. (7) b
Derived from the parameters in Ref. [28].
c
are calculated at the equilibrium volume of the ordered Derived from the parameters in Ref. [59].
phase and of the random alloy, respectively. Here, we have
considered the following ordering processes: (i) from bcc sponding DEOrd values are considerably smaller in magni-
solid solution to B2 and DO3 structures, (ii) from fcc solid tude. Differences between the current and previous results
solution to L12, DO22, DO23 and L10 structures, and (iii) are apparent in the calculated magnitudes of DEOrd. In
from hcp solid solution to B19 and DO19 structures. the case of fcc alloys, the DEOrd values in this study differ
The calculated DEOrds, which are listed in Table 8, have from those derived by Asta et al [27,28] by as much as
also been compared with previous studies of alloy energet- 4.7 kJ/mol-atom in the Al–Ti system, and from Clouet
ics using first-principles methods [27,28,55,59]. In these pre- et al. [59] by as much as 7.3 kJ/mol-atom in Al–Zr system.
vious publications values of DEOrd were not reported In the case of hcp alloys, the DEOrd values in this study dif-
explicitly; instead we have derived them using the reported fer from those of Asta et al. [27,28] by less than 3 kJ/mol-
parameters. For example, Rubin and Finel [55] and Clouet atom in Al–Ti.
et al. [59] reported the formation energies of ordered The differences in magnitudes of DEOrd obtained in the
phases, as well as the values of the relevant ECIs, from present vs. previous studies may be attributed to the fol-
which the formation energies of random alloys could be lowing factors: (i) the calculation of total energies in the
derived. Asta et al. [27,28] reported the formation energies different studies made use of different methods, e.g.
of both ordered phases and the random alloy, allowing LMTO-ASA, FP-LMTO and USPP (VASP), involving dif-
ordering energies to be derived by subtraction. ferent numerical approximations; (ii) the role of exchange-
The results in Table 8 obtained in the present and previ- correlation, i.e. the use of LDA vs. GGA; (iii) the level of
ous calculations show very similar trends in the composi- convergence of the CEs, which in turn directly affects the
tion dependence of DEOrd for a given system and solid predicted formation energy of the random alloy; and (iv)
solution. For example, in the case of bcc Al–Ti alloys the the level of accuracy involved in the modeling of elastic
maximum values of DEOrd is found at 50 at.% Ti (corre- relaxation energies. While factors (iii) and (iv) are difficult
sponding to ordering into the B2 structure), while in Al- to assess due to lack of detail available in previous work
rich alloys bcc-based chemical ordering to the DO3 phase (e.g. the estimated precision of the cluster expansions as
is strongly disfavored. Similarly, when fcc Al–Ti and Al– estimated by the CV score), we note that (i) and (ii) can
Zr alloys are considered, we find that the maximum in lead to significant quantifiable differences. For example,
DEOrd occurs at 25 at.% Ti or Zr, leading to the formation the values of DfE for L12, DO22, and DO23 structures of
of DO23 structure, while in Ti- or Zr-rich alloys the corre- Al3Zr calculated by the FP-LMTO–LDA [59] method dif-
G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221 3219

fer from those using USPP–GGA (VASP) [39] by about (ii) DmE values have been calculated for both completely
4 kJ/mol-atom, with the former method giving more nega- random solid solutions and solid solutions with their
tive values of DfE. equilibrium state of SRO. It is shown that SRO has a
significant effect on the alloy energetics in the Al–TM
7. Suggestions for further experimental measurements systems considered, and calculated results which
include these effects are found to be in excellent agree-
In view of the calculated results presented in Figs. 6, 8, ment with experimental calorimetry data available
10, 12–14, and the discussions in the previous section, it for hcp Al–Ti alloys.
would be very useful to have access to expanded calorimet- (iii) In the dilute limit (i.e. in the absence of solute–solute
ric data for bcc and hcp solution phases in Al–TM interactions), the DmE values calculated using con-
(TM = Ti, Zr and Hf) systems. Specifically, calorimetric ventional supercells agree very well in most cases with
data would be very useful to establish the accuracy of alter- those obtained by the CE method which does not
native first-principles methods in the calculation of DmH explicitly consider structures with such dilute compo-
for the different solid solution phases. Comparisons of such sitions in the calculation of the ECIs. Notable excep-
experimental data with the results of theoretical calcula- tions are the results for bcc Ti(Al) and bcc Al(Zr),
tions performed with and without consideration of SRO where the solvent crystal structure is not mechani-
would help to establish the importance of these effects on cally stable.
solid-solution energetics. The DfH data for hcp Al–Ti (iv) In the concentrated region, the DmE calculated by CE
alloys [42], which validates the accuracy of the present and SQS methods are compared. While in many cases
first-principles-based MC results, helps to establish the we find very good agreement between these two
importance of including SRO in calculations of the ener- methods, a discrepancy between these methods may
getics of this phase; additional experimental data is highly arise in several cases where very large relaxation
desirable to ascertain the importance of SRO more gener- effects are obtained, presumably due to inherent elas-
ally in other phases with different crystal structures and tic and dynamic instabilities of the parent lattice
TM elements. In this respect, experimental data for bcc structures.
Ti–Al solid solutions would be particularly interesting for (v) Further calorimetric data for bcc and hcp solid solu-
testing the accuracy of the first-principles calculations, tions in the Al–TM (TM = Ti, Zr and Hf) binary sys-
since the solubility of Al in the bcc phase is much higher tems is desirable to test and validate the alloy
than in the hcp solid solutions, and as shown in Figs. 12– energetics calculated from first-principles, to establish
14 the SRO contributions to DmE, as reflected by the differ- the relative accuracy of the different methods
ences between random-alloy (CE) and MC results, are suf- described in Section 1, and to establish the impor-
ficiently large at the higher concentrations to lie outside tance of SRO effects on the energetics of solid
experimental and calculated uncertainties. One limitation solutions.
of our treatment the Ti-rich bcc phase is the inability to (vi) The thermodynamic properties of bcc, fcc and hcp
properly model its dynamic instabilities [92–94]. We have solid solution phases presented here, and also for
implicitly relied on the assumption that the free energy con- intermetallics and fcc solid solution phase in our pre-
tributions arising from these instabilities cancel out in the vious publications [39,40], may be integrated within
calculation of the formation free energies, but this approx- the CALPHAD formalism [34,35] to create reliable
imation clearly deserves further investigation. thermodynamic database in order to facilitate calcu-
lations of stable and metastable phase diagrams of
8. Conclusions binary and multicomponent systems containing Al
and TM (TM = Ti, Zr and Hf). This in turn may
A systematic and comprehensive study of phase stability facilitate design and processing of relevant materials
of bcc, fcc and hcp solid solutions in Al–TM (TM = Ti, Zr based on these alloy systems.
and Hf) systems has been carried out employing electronic
DFT methods and modern alloy theory tools. Specifically, Acknowledgements
the energetics of solid solutions in Al–TM systems have
been calculated from first principles using two different This research was supported by the US Department of
approaches based on the CE and supercell (both dilute Energy, Office of Basic Energy Sciences, under Contract
and concentrated) methods. The following conclusions Nos. DE-FG02-02ER45997 (G.G.) and DE-FG02-
are drawn: 01ER45910 (A.v.d.W. and M.A.) at Northwestern, and
DE-FG02-06ER46282 (M.A.) at UC Davis. In addition,
(i) The ECIs in bcc, fcc and hcp alloys for binary Al–TM this material is based upon work supported by the National
(TM = Ti, Zr and Hf) systems have been calculated Science Foundation under the following NSF programs:
from first principles. To achieve well-converged Partnerships for Advanced Computational Infrastructure,
CEs, relatively long-ranged interactions are required Distributed Terascale Facility (DTF) and Terascale Exten-
for these systems. sions: Enhancements to the Extensible Terascale Facility.
3220 G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221

Specific to NSF programs, two of us (G.G. and A.v.d.W.) [31] Abrikosov IA, Ruban AV, Johansson B, Skriver HL. Comput
have utilized Itanium clusters as a part of TeraGrid sites at Mater Sci 1998;10:302.
[32] Abrikosov IA, Johansson B. Phys Rev B 1998;57:14164.
the University of Illinois at Urbana-Champaign and at San [33] Turchi PEA, Abrikosov IA, Burton B, Fries SG, Grimvall G,
Diego Supercomputing Center. One of us (G.G.) would Kaufman L, et al. Calphad 2007;31:4.
like to thank Prof. I.A. Abrikosov, Linköping University, [34] Kaufman L, Bernstein H. Computer calculation of phase dia-
Sweden, for fruitful discussions. The authors thank Prof. grams. New York: Academic Press; 1970.
M. Widom, Carnegie Mellon University, for sharing his [35] Lukas HL, Fries SG, Sundman B. Computational thermodynamics.
The Calphad method. Cambridge: Cambridge University Press;
partial correlation function calculation code. 2007.
[36] Andersson JO, Helander T, Höglund L, Shi PF, Sundman B.
References Calphad 2002;26:273.
[37] Borgenstam A, Engström A, Höglund L, Ågren J. J Phase Equilib
[1] Hafner J, Wolverton C, Ceder G. MRS Bull 2006;31:659. 2000;21:269.
[2] Sanchez JM, Ducastelle F, Gratias D. Physica A 1984;128:334. [38] Ghosh G, Olson GB. Acta Mater 2007;55:3281.
[3] de Fontaine D. Solid State Phys 1994;47:33. [39] Ghosh G, Asta M. Acta Mater 2005;53:3225.
[4] Zunger A. In: Turchi PE, Finel A, editors. Statics and dynamics of [40] Liu JZ, Ghosh G, van de Walle A, Asta M. Phys Rev B
alloy phase transformations. NATO ASI Series. New York: Plenum 2007;75:104117.
Press; 1994. [41] Schuster JC, Palm M. J Phase Equilib Diff 2006;27:255.
[5] Hart GLW, Blum V, Walorski MJ, Zunger A. Nat Mater [42] Kubaschewski O, Dench WA. Acta Metall 1955;3:339.
2005;4:391. [43] Hoch M, Usell RJ. Metall Trans 1971;2:2627.
[6] van de Walle A, Ghosh G, Asta M. Ab initio modeling of alloy [44] Samokhval VV, Poleshchuk PA, Vecher AA. Russ J Phys Chem
phase stability. In: Bozzolo G, Noebe RD, Abel P, editors. Applied 1971;45:1174.
computational materials modeling: theory, simulation and experi- [45] Hultgren R, Desai P, Hawkins P, Gleiser M, Kelley K. Selected
ment. New York: Springer; 2007. p. 1–34. values of the thermodynamic properties of binary alloys. Metals
[7] Zunger A, Wei S-H, Ferreira LG, Bernard JE. Phys Rev Lett Park (OH): ASM International; 1973. p. 497.
1990;65:353. [46] Desai PD. J Phys Chem Ref Data 1987;16:109.
[8] Mäder KA, Zunger A. Phys Rev B 1995;51:10462. [47] Eckert M, Bencze L, Kath D, Nickel H, Hilpert K. Ber Bunsenges
[9] Abrikosov IA, Simak SI, Johansson B, Ruban AV, Skriver HL. Phys Chem 1996;100:418.
Phys Rev B 1997;56:9319. [48] Murray JL, Peruzzi A, Abriata JP. J Phase Equilib 1992;13:277.
[10] Jiang C, Wolverton C, Sofo J, Chen LQ, Liu Z-K. Phys Rev B [49] Murray JL, McAlister AJ, Kahan DJ. J Phase Equilib 1998;19:376.
2004;69:214202. [50] Andersen OK. Phys Rev B 1975;12:3060.
[11] Shin D, van de Walle A, Wang Y, Liu Z-K. Phys Rev B [51] Methfessel M. Phys Rev B 1988;38:1537.
2007;76:144204. [52] U von Barth, Hedin L. J Phys C 1972;5:1629.
[12] Stocks GM, Nicholson DM, Shelton WA, Gyorffy BL, Pinski FJ, [53] Connolly JWD, Williams AR. Phys Rev B 1983;27:5169.
Johnson DD, et al. In: Turchi PEA, Gonis A, editors. Statics and [54] Kikuchi R. Phys Rev 1951;81:988.
dynamics of alloy phase transformations. NATO ASI Series. New [55] Rubin G, Finel A. J Phys: Condens Mater 1995;7:3139.
York: Plenum Press; 1994. [56] Skriver HL. The LMTO method. Berlin: Springer; 1991.
[13] Vitos L, Abrikosov IA, Johansson B. Phys Rev Lett 2001;87:156401. [57] Hohenberg P, Kohn W. Phys Rev B 1964;136:864.
[14] Rowlands DA, Ernst A, Gyorffy BL, Staunton JB. Phys Rev B [58] Sham LJ, Kohn W. Phys Rev A 1965;140:1133.
2006;73:165122. [59] Clouet E, Sanchez JM, Sigli C. Phys Rev B 2002;65:094105.
[15] Chepulskii RV, Staunton JB, Bruno E, Ginatempo B, Johnson DD. [60] Kaufman L, Nesor H. Calphad 1978;2:325.
Phys Rev B 2002;65:064201. [61] Gros JP, Sundman B, Ansara I. Scripta Metall 1988;22:1587.
[16] Ducastelle F. Order and phase stability in alloys. Amster- [62] Murray JL. Metall Trans 1988;19A:243.
dam: North-Holland; 1991. [63] Kattner UR, Lin JC, Chang YA. Metall Trans 1992;23A:2081.
[17] Ruban AV, Shallcross S, Simak SI, Skriver HL. Phys Rev B [64] Zhang F, Chen LS, Chang YA, Kattner UR. Intermetallics
2004;70:125115. 1997;5:471.
[18] Khachaturyan AG. Theory and structural transformations in [65] Saunders N. In: Ansara I, Dinsdale AT, Rand MH, editors.
solids. New York: John Wiley; 1983. COST507: Thermochemical database for light metal alloys. vol.
[19] de Fontaine D. Solid State Phys 1979;34:73. 2. Brussels and Luxembourg: European Commission; 1998.
[20] Bugaev VN, Reichert H, Shchyglo O, Udyansky A, Sikula Y, Dosch [66] Ohnuma I, Fujita Y, Mitsui H, Ishikawa K, Kainuma R, Ishida K.
H. Phys Rev B 2002;65:180203. Acta Mater 2000;48:3113.
[21] Wolverton C, Zunger A. Phys Rev Lett 1995;75:3162. [67] Murray JL. In: Murray JL, editor. Phase diagram of binary titanium
[22] Shchyglo O, Bugaev VN, Drautz R, Udyansky A, Reichert H, alloys. Metals Park (OH): ASM International; 1987. p. 12.
Dosch H. Phys Rev B 2005;72:140201. [68] Wang T, Jin Z, Zhao JC. J Phase Equilib 2001;22:544.
[23] Geng HY, Sluiter MHF, Chen NX. Phys Rev B 2006;73:012202. [69] Kaufman L, Nesor H. Can Metall Quart 1975;14:221.
[24] Ruban AV, Simak SI, Shallcross S, Skriver HL. Phys Rev B [70] Wang T, Jin Z, Zhao JC. J Phase Equilib 2002;23:416.
2003;67:214302. [71] Kresse G, Hafner JJ. Phys Rev B 1994;49:14251.
[25] Kissavos AE, Shallcross S, Kaufman L, Granas O, Ruban AV, [72] Kresse G, Furthmüller J. Phys Rev B 1996;54:11169.
Abrikosov IA. Phys Rev B 2007;75:184203. [73] Kresse G, Furthmüller J. Comput Mater Sci 1996;6:15.
[26] Alling B, Ruban AV, Karimi A, Peil OE, Simak SI, Hultman L, [74] Vanderbilt D. Phys Rev B 1990;41:7892.
et al. Phys Rev B 2007;75:045123. [75] Kresse G, Hafner J. J Phys: Condens Mater 1994;6:8245.
[27] Asta M, de Fontaine D, van Schilfgaarde M, Sluiter M, Methfessel [76] Perdew JP. In: Ziesche P, Eschrig H, editors. Electronic structure of
M. Phys Rev B 1992;46:5055. solids’91. Berlin: Akademie Verlag; 1991. p. 11.
[28] Asta M, de Fontaine D, van Schilfgaarde M. J Mater Res [77] Monkhorst HJ, Pack JD. Phys Rev B 1976;13:5188.
1993;8:2554. [78] Methfessel M, Paxton AT. Phys Rev B 1989;40:3616.
[29] van de Walle A, Asta M. Met Mater Trans A 2002;33:735. [79] Binder K, Heermann DW. Monte-Carlo simulation in statistical
[30] Johnson DD, Asta M. Comput Mater Sci 1997;8:54. physics. New York: Springer; 1988.
G. Ghosh et al. / Acta Materialia 56 (2008) 3202–3221 3221

[80] Newman MEJ, Markema GT. Monte-Carlo methods in statistical [93] Heiming A, Petry W, Trampenau J, Alba M, Herzig C, Schober HR,
physics. Oxford: Clarendon Press; 1999. et al. Phys Rev B 1991;43:10948.
[81] van de Walle A, Asta M, Ceder G. Calphad 2003;26:539. [94] Trampenau J, Heiming A, Petry W, Alba M, Herzig C, Miekeley W,
[82] van de Walle A, Ceder G. J Phase Equilib 2002;23:348. et al. Phys Rev B 1991;43:10963.
[83] Craievich PJ, Weinert M, Sanchez JM, Watson RE. Phys Rev Lett [95] Münster A, Sagel K, Zwicker U. Acta Metall 1956;4:558.
1994;72:307. [96] Yao YL. Trans ASM 1961;54:241.
[84] Craievich PJ, Sanchez JM, Watson RE, Weinert M. Phys Rev B [97] Clark D, Jepson KS, Lewis GI. J Inst Met 1962/1963;91:197.
1997;55:787. [98] Crossley FA. Trans AIME 1966;236:1174.
[85] Grimvall G. Ber Bunsen Phys Chem 1998;102:1083. [99] Namboodhiri TKG, McMahan Jr CJ, Herman H. Metall Trans
[86] van de Walle A, Asta M. Modell Simul Mater Sci Eng 2002;10:521. 1973;4:1323.
[87] Perdew JP, Wang Y. Phys Rev B 1991;45:13244. [100] Swartzendruber LJ, Bennett LH, Ives LK, Shull RD. Mater Sci Eng
[88] Watson RE, Weinert M. Solid State Phys 2001;56:1. 1981;51:P1.
[89] Ozolins V, Asta M. Phys Rev Lett 2001;86:448. [101] Blackburn MJ, Williams JC. Trans ASM 1969;62:398.
[90] van de Walle A, Ceder G. Rev Mod Phys 2002;74:11. [102] Truax DJ, McMahan Jr CJ. Mater Sci Eng 1974;13:125.
[91] Liot F, Simak SI, Abrikosov IA. J Appl Phys 2006;99:08P906. [103] Neeraj T, Mills MJ. Mater Sci Eng A 2001;319–321:415.
[92] Petry W, Heiming A, Trampenau J, Alba M, Herzig C, Schober HR, [104] Neeraj T, Mills MJ. Philos Mag A 2002;82:779.
et al. Phys Rev B 1991;43:10933.

View publication stats

You might also like