Article

pubs.acs.org/cm

Multivariate Method-Assisted Ab Initio Study of Olivine-Type LiMXO4

(Main Group M2+−X5+ and M3+−X4+) Compositions as Potential Solid
Electrolytes
Randy Jalem, Takahiro Aoyama, Masanobu Nakayama,* and Masayuki Nogami
Department of Materials Science and Engineering, Nagoya Institute of Technology, Room 604A, Bldg. 2, Gokiso-cho, Showa-ku,
Nagoya City, Aichi, Japan 466-8555
*
S Supporting Information

ABSTRACT: Lithium-ion conductive solid oxide electrolytes are receiving increasing attention for the development of high-
performance all-solid state rechargeable Li-ion batteries. In this study, we report an effective search method that uses a
combination of ab initio calculations and multivariate analysis to find potential solid oxide electrolyte materials with low Li-ion
hopping energies (EAs) among 66 olivine-type oxides with an ordered structure LiMXO4 (main group M2+−X5+, M3+−X4+). The
ionic size of M, with its associated octahedral distortion leading to the appreciable local lattice distortion during the actual Li ion
hop, made the most significant contribution to the variation in EA values. Promising M−X pairs (<0.30 eV) with experimental
data in the Inorganic Crystal Structure Database include Mg−As (0.17 eV), Sc−Ge (0.22 eV), In−Ge (0.28 eV), and Mg−P
(0.30 eV); Mg−As and Sc−Ge are reported here for the first time as battery materials. Promising theoretical compositions are
also determined for Group 13 Al−X, Ga−X, and In−X pairs, and for Ca−X pairs.
KEYWORDS: Li ionic conductors, solid electrolyte, olivine materials, lithium-ion battery, density functional theory, multivariate analysis

■ INTRODUCTION
There has been intense interest in the development of lithium-
ion conductive oxides (solid electrolytes) as base materials for
all-solid-state rechargeable Li-ion batteries. Solid electrolytes
offer a variety of advantages over conventional electrolytes:
nonflammability, operating stability because of an oxide-based
composition, corrosion resistance (which leads to high safety),
a wide range of operating temperatures, and longer service life.
The solid electrolyte is a critical component; however, there are
an extremely limited set of potential materials available, because
of problems with the Li-ion hopping. This phenomenon is
caused by the strong interaction between neighboring ions in a
densely packed crystal structure, as a consequence of the low
polarizability of Li+ and O2−.1 Known oxides that can be used as
solid electrolytes include garnet-type Li5La3Ta2O12, perovskite-
type (Li,La)TiO3, Li β-Al2O3, LISICON, and NASICON.2−6
Finding suitable new solid electrolyte materials will provide the
next major advance in this field.
Olivine oxides were investigated in this study because it was
experimentally confirmed that such materialsone example is
the LiFePO4 compositioncould display high lithium bulk
mobility and ultrafast charging/discharging capabilities, indicat-
ing the possibility for use as a fast ionic conductor.7 However,
electronic conductivity has prevented compositions with
transition metals from becoming a viable solid electrolyte.8−10
Designing an olivine-type electrolyte material with high lithium
bulk mobility and suppressed electronic conductivity would be
a major technological advance, especially if it is paired with
attractive cathode materials of the same structure such as
LiFePO4 for the reduction of interfacial resistance. In fact, we
recently reported that LiInSiO4 and LiInGeO4 with olivine-type
structure show enormously small Li-ion hopping energies
0.36 and 0.23 eV, respectivelyusing ab initio density
functional theory (DFT) framework.11 Hence, the olivine-
type structure offers a fast Li+ conduction pathway.
Thus, an attempt has been made here to search the best
stoichiometric compositions of ordered LiMXO4 olivine
structures for use as a Li-ion conductor, based on main group
M octahedral atoms (Group 2, Group 3, lanthanides, and
Group 13) and X tetrahedral atoms (Group 14 and Group 15).
A total of 72 M−X pairs are conceivable, although most of the
compositions are not yet reported. This is an indispensable tool
for guiding critical decisions, such as the selection of
compositions to be prioritized, and it can reduce the time
and cost of experiments.
In this study, technical difficulty arising from computational
expense for a large number of compositions also is solved by a
new approach in Li-ion battery materials. A systematic search of
materials by ab initio computations and multivariate analysis
was used to assess olivine-type LiMXO4 compositions. We
devised this approach because an efficient method is needed for
predicting the Li-ion hopping energy (EA) for a large number
of isostructural compounds. Stabilization of crystal structures
via first-principle computation can provide accurate structural
data with minimal computing time; however, the opposite is
true for EA computations. A stochastic approach using
multivariate partial least-squares (PLS) regression reduced the
computing time for hopping energies. The advantage originates
from the training of the predictive model, where the number of
actual observations or response variables (i.e., EA) can be less
Received: January 5, 2012
Revised: March 13, 2012
Published: March 14, 2012

© 2012 American Chemical Society 1357 dx.doi.org/10.1021/cm3000427 | Chem. Mater. 2012, 24, 1357−1364
Chemistry of Materials Article

than the number of predictors. Thus, fewer reference samples variation of θ, the largest internal O−M−O angle, relative to
could be used to determine the trends in the structure- the b-axis (see Figure 1c).
dependent EA values. This technique has been widely adopted The DFT framework was used to calculate the different
in other areas of research.12−14 crystal structure lattice parameters for all the compositions

■ COMPUTATIONAL DETAILS
We initialize the atomistic model for the projected augmented wave
(Figure 2). When different cation pairs constitute the skeleton

(PAW) method15 through the Vienna Ab Initio Simulation Package

(VASP)16 and used published structural information on experimentally
synthesized ordered olivine materials from the Inorganic Crystal
Structure Database (ICSD). The atomic position data of the known
materials was utilized in the structural relaxation of virtual
compositions.17 An energy cutoff of 500 eV is chosen to ensure that
the total energies are converged within 3 meV per formula unit. Test
calculations showed that the results for a 2 × 3 × 3 Monkhorst-Pack k-
point grid in a unit cell showed a reasonable convergence and an error
of <1 meV/LiMXO4 unit in the total energy. Moreover, the associated
error for the calculated lattice parameters was ∼1.5% on average,
which is well within the standard generalized gradient approximation
(GGA) error when using the parametrization of Perdew−Burke−
Ernzernhof for solids (PBEsol) for the exchange correlation energy.18
The nudged-elastic-band (NEB) method19 was used to determine
migration pathway and energy, in which good accordance with
experiments for Li ion conductors has been reported.20 For the NEB
method, a 1 × 2 × 2 supercell with a formula unit of Li15M16X16O64 (1
Li vacancy created along the diffusion channel), and numerical
integration was performed over the Brillouin zone by sampling the Γ-
point. Convergence was determined to be within 30 meV/formula
unit. Additional descriptions for the methodology are available in
Sections 1 and 2 of the Supporting Information, as well as in the
literature.11 All the multivariate analyses were carried out using the
JMP statistical software.

■
RESULTS AND DISCUSSION
Olivine Structure, Structural Variation, and the
Relation to Li-Ion Hopping Energy. The olivine structure
(Figure 1) has the typical orthorhombic symmetry, which
belongs to the Pnma space group and has a slight distortion in
the close-packed arrays of oxygen ions. The distortion in the M
octahedron can be explained by two main factors: (i) M ions
being off-center, which is expressed relative to the tilt of the
octahedron from the a-axis (see Figure 1b) and (ii) the
Figure 1. (a) Projected view of the olivine structure. M octahedral
distortion is described by (b) M becoming off-center due to repulsion
from the Li channel, and (c) the variation of the largest internal O−
M−O angle, θ, relative to the b-axis. Green spheres represent Li ions,
purple octahedra represent M atoms, and blue tetrahedra represent X
atoms.
1358
Figure 2. (a) Unit-cell volume, (b) lattice parameter a, (c) lattice
parameter b, and (d) lattice parameter c of DFT-relaxed structures
with different M−X pair compositions, sorted according to X
tetrahedral cation.
host structure, observing how the structural parameters change
may clarify the determining factors in the value of the Li-ion
hopping energy. Figure 2a shows the direct relationship
between the unit-cell volume and the ionic size (Shannon
radii) of the inserted cations in both the M and X sites.21 For
simplicity, the different LiMXO4 samples are denoted by their
M−X pairing. The largest volume was determined for Ba−Sb
(434.87 Å3), and the smallest volume was found for Al−Si
(255.13 Å3). The Ba−Sb pair consists of large ions, whereas the
Al−Si pair consists of relatively small-sized ions. The main
cluster of data points belongs to pairs containing rare-earth
cations, with a volume that gradually increases with the ionic
size of M. Further investigation into the variation along the
major axes (Figures 2b−d) revealed that lattice parameter a
changes with volume, implying that the volume could be
governed primarily by the expansion or contraction in the a-
axis, as suggested by the distortion shown in Figure 1b. The
alternating Li and M octahedron along the a-axis are expected
to have strong repulsion because they are both highly ionic,
especially for large M. This repulsion also will result in M
becoming more off-center and farther from the Li channel,
directly affecting the value of a. The lattice parameter b is
directly influenced by the largest internal angle, θ (Figure 1c).
Because the variation in θ causes a strain component along the
b-axis, increasing or decreasing it should also alter the value of
lattice parameter b. Furthermore, no significant variation in
lattice parameter b was observed for different cation sizes of X.
The contribution along the b-axis from the X tetrahedra may
have been minimal; these sites are generally regarded as
dx.doi.org/10.1021/cm3000427 | Chem. Mater. 2012, 24, 1357−1364
Chemistry of Materials Article

interstices for small, strong directionality cations, which cause

minimal strain along the major directions. Furthermore, there is
no abrupt change in lattice parameter c for different M ionic
radii, where the bulk of the distortion was already taken in
along the a- and b-axes.
Another important observation here is that large-sized M
ions often lead to more distortion of the related octahedron
from the coordinated oxygen ions. As evidence, we calculated
the shape unideality factor for polyhedrons, formally defined by
the quadratic elongation ⟨λ⟩ of the form:22
n 1 ⎛ ⎞2
1 l
⟨λ⟩ = ∑∑⎜ ⎟
n ⎝ l0 ⎠ (1)
1 n
where l and l0 are the center-to-vertex distance of the actual and
regular polyhedron of the same volume, respectively. ⟨λ⟩ is
dimensionless, giving a quantitative measure of distortion that
is independent of the effective polyhedron size. For the M
octahedron with Group 2 cations such as MgP, CaP, SrP, and
BaP pairs, the ⟨λ⟩ values are 1.0318, 1.0538, 1.1157, and
1.1574, respectively. Conversely, X cations of Group 13, such as
Ca−P, Ca−As, Ca−Sb, Ba−P, Ba−As, and Ba−Sb pairs, follow
a similar pattern but to a lesser extent: 1.0032, 1.0080, 1.0203,
1.0008, 1.0017 and 1.0127, respectively. Later, we will discuss
these numerical values, among others, and use them to extract
the underlying factor(s) to build a prediction model for the Li-
ion hopping energy, which is characteristic of a particular
composition.
Li-Ion Hopping Energy via the NEB Method. To build
the working dataset for the PLS model, the Li-ion hopping
energy was computed for several M−X pairs by the NEB
method. Different intermediate positions were considered, with
the peak located midway between the Li-occupied site and the
neighboring Li-vacancy site. As a reference point, the criteria for Figure 3. (a) Migration pathway of the Li ion into an Li vacancy site
the Li-ion hopping energy was set using the experimental value (from position 1 to position 2) via an interstitial route (In−Ge pair
obtained for the (Li,La)TiO3 perovskite material (∼0.30 eV), was used as a model). (b) Energy and interatomic distance profiles of
ICSD-derived LiMXO4 compositions (from NEB method results) with
which has the highest experimental Li-ion conductivity value of (i) EA profiles and the corresponding nearest neighbor interatomic
the oxides (∼1 × 10−3 S/cm at 300 K).3 This material offers a distances during Li ion migration ((ii) Li−O1, (iii) Li−M2, (iv) Li−X1,
good benchmark for our study in which the associated Li-ion and (v) M3−X1 distance profiles).
jump is through equivalent sites, whereas a vacancy-driven
mechanism along a one-dimensional (1D)-diffusion channel interstitial intermediate routes along the diffusion channel, as
was determined through the study on LiFePO4 by using first- pictured in the inset of Figure 3a. One is through the interstitial
principles DFT,23 a force-field approach,9 and neutron site (dashed yellow outline) face-shared with the first-nearest
diffraction.24 This 1D conduction in olivines is likely to limit tetrahedral cation X1. Another is through the interstitial site
conductivity in polycrystalline pellets. However, this may be (solid yellow outline) edge-shared with the second-nearest
addressed by controlling the alignment of the grains to octahedral M cation M2. The path curvature can then be
minimize intergrain resistance. Also, the existence of antisite described as minimizing the long-range interaction between the
defects can give rise to pathway blocking, which, in turn, may mobile Li ion and X1 by maximizing their interatomic distance
further decrease conductivity.25,26
through passage into the immediate interstitial site associated
We then explore the possibility of entering into a low-EA-
with M2. To elaborate, we took In−Ge as a case material. The
value regime (<0.30 eV) through modification of the M:X
cation size ratio. The transport properties of the olivine Li−X1 distance (panel iv in Figure 3b) when Li is still at
structure are said to be largely determined by topology, where position 1 is ∼1 Å shorter than Li−M2. Another point is that
Li interchain routes, which utilize interstitial sites for Li-ion the X1-related route has a face-shared interstitial site with the X1
migration, become much more difficult, compared to intrachain tetrahedron while the M2-related route has an edge-shared
motion, because of the very close proximity of Li ions along the interstitial site with the M 2 octahedron. With these
chains, as opposed to between chains and the complex energy considerations, the X1-related route should be energetically
landscape between chains caused by the two types of unfavorable, because of the expected increased Li−X
polyhedron neighbors (in our study, ionic M octahedra and Coulombic interaction from the decreasing Li−X1 distance
covalent X tetrahedra). The pathway of a migrating Li ion in and energy penalty of face sharing in that route. The energy
the Li intrachain along the b-axis, based on DFT results, is profile and the corresponding variation in interatomic distances
depicted in Figure 3a. Position 1 is the starting point and between different neighboring ions for a mobile Li ion are
position 2 represents the Li vacancy site. There are two possible plotted for compositions with reports according to ICSD,
1359 dx.doi.org/10.1021/cm3000427 | Chem. Mater. 2012, 24, 1357−1364
Chemistry of Materials Article

namely, In−Ge, In−Si, Sc−Si, Sc−Ge, Mg−P, Mg−As, and
Tm−Si (Figure 3b). Other M−X pairings that are not included
in the plot demonstrated a similar trajectory and trending in
interatomic distance variation. As observed, the interatomic
distances for Li−O1 and Li−M2 decrease up to path midway
(see panels ii and iii in Figure 3b). For Li−O1, the variation in
the distance profile is relatively small which indicates that the Li
octahedral cages are relatively unchanged at different M and X
ions. For Li−M2, the variation is much more pronounced,
which could be directly attributed to the varying size and
distortion of the M octahedron itself. The M3−X1 distance only
showed a slight increase along the pathway (see panel v in
Figure 3b).
The Born effective charge calculations confirmed that the
dynamic charge of the X cations is considerably smaller than
their nominal charge, which confirms the strong covalent
character of X−O bonds. For the M cation, the case is opposite
where the actual charge tends to be larger than the nominal
value. It would be related to the polarizability of M and the
partially covalent interaction between M and O which is
responsible for the anomalous effective charges and for the
anisotropy observed in the oxygen charge tensor.27,28 For the Li
cation, the dynamic charge is almost isotropic, which is
reflective of its low polarizability and similarity with other
oxides containing alkali metals.29 These relationships are
collected in Table 1, where the ionic size dependency of the
Table 1. DFT-Calculated Born Effective (Dynamic) Charges
of Selected M−X Pair Compositions
Effective Charge
M−X pair Li M X
Al−Si +1.09 +3.06 +3.04
Al−Ge +1.10 +3.09 +3.21
Al−Sn +1.12 +3.01 +3.66
Ga−Si +1.11 +3.34 +3.05
Ga−Ge +1.12 +3.38 +3.20
Ga−Sn +1.14 +3.33 +3.59
In−Si +1.12 +3.41 +2.94
In−Ge +1.14 +3.44 +3.10
In−Sn +1.15 +3.40 +3.33
charges for M and X is made very clear. It is also apparent that
larger M tends to have higher polarizability, explaining the
observed M distortion and M off-centering (see Figure 1b).
From the analysis of structural variations in the previous
section, it is found out that the M octahedron tends to have
more distortion than the X tetrahedron. Since the pathway
maximizes the Li−X1 distance by passing through an M2 edge-
shared interstitial site, the M octahedron distortion, as a
function of M cation size, may have a direct influence on the Li
migration path. However, identifying the exact relationship to
EA must be clarified at this point.
One crucial parameter that comes into mind is the steric
hindrance of the interstitial oxygen cage edge-shared with M2
octahedron. However, this factor cannot account for the
variation in EA, since the change in Li−O distance with
different M during Li migration is negligible (panel ii in Figure
3b).
Another parameter is the electrostatic repulsive interaction
between Li and M2 (panel i in Figure 2b) over the course of
migration. This repulsive interaction increases as the
interatomic distance of Li−M2 decreases at the transition
1360
state. In other words, EA increases with the reduction of the
Li−M2 distance. However, the order of EA in panel i in Figure
2b seems to be independent of the Li−M2 distance. Hence, the
electrostatic repulsive interaction is not a crucial factor that
affects EA.
Because of this, there must be another important parameter
that underpins the EA variation observed at different M size.
Here, we propose that the contribution of local lattice
distortion would be an important factor for EA. To verify
this claim, we investigated the local lattice distortion that occurs
in the vicinity of the migration pathway. We calculated the total
quasi-volume of path for the migrating Li ion at the initial state
when Li ion is still at position 1 (Figure 3a) and at the
transition state when the Li ion is at the interstitial site
associated with M2. The path volumes at the initial state (Vi)
and the transition state (Vt), respectively, are calculated as
follows:
Vi = Poly1i + Poly2 i + Poly3i (2)
Vt = Poly1t + Poly2 t + Poly3t (3)
where Vi and Vt are the quasi-volume of the path at the initial
state and the transition state, respectively. The path volume is
taken as the sum of Poly1, Poly2, and Poly3, which are the Li-
ion octahedron volume, interstitial tetrahedron volume, and
neighboring octahedron Li vacancy volume, respectively, during
the initial and transition state. The findings are given in Figure
4 (with the three polyhedrons shown in the inset) using Al−
Figure 4. Local lattice distortion along the Li-ion migration pathway
expressed in terms of volume difference of the quasi-volume of the
path between the transition state (mobile Li ion at interstitial site) and
the initial state (before migration). Inset shows the derivation of the
path volume as the sum of the polyhedrons Poly1, Poly2, and Poly3.
Large gray spheres in Poly1 and Poly2 depict Li ions while the small
gray sphere in Poly3 depicts a Li vacancy. Poly1 is occupied with Li
ion at the initial state and empty at the transition state. Poly2 is empty
at the initial state and is occupied at the transition state. Sample pairs
are Al−Ge, Ga−Ge, and In−Ge.
Ge, Ga−Ge, and In−Ge pairs. The volume difference ΔV of the
two states is shown to be positive and increasing with
increasing M size, suggesting that Li-ion migration is
accompanied by significant local lattice distortion. It also
indicates that at larger M size, the degree of lattice distortion is
increased. Such types of distortion usually involve an energy
penalty and complements well with the observed EA variation.
The following candidates were identified among the
compositions, using data based on the ICSD and with reference
to our benchmark material: Mg−As (0.17 eV), Sc−Ge (0.22
eV), In−Ge (0.28 eV), and Mg−P (0.30 eV). To the best of
our knowledge, Mg−As and Sc−Ge have not been investigated
dx.doi.org/10.1021/cm3000427 | Chem. Mater. 2012, 24, 1357−1364
Chemistry of Materials
as materials for solid electrolytes. In−Ge has been evaluated
theoretically by Rajkumar et al., with similar results.11 Mg−P
was electrochemically characterized by Fehr et al., and
LiMgPO4 showed enhanced DC conductivity with an EA of
∼1.45 eV, which is consistent with Li-ion hopping without a Li-
ion deficiency.30 The difference of the latter from the DFT
result arises from the fact that the experimental result came
from the sum of the Li vacancy formation energy and the Li-ion
hopping energy.
It is worth mentioning that other critical factors also
contribute toward a successful solid electrolyte candidate
material. These are electrical conductivity, stability for forming
the olivine framework, and vacancy content, which mediates Li-
ion hopping in the diffusion channel. Since the substituted M
cations here are non-transition metals, electrical conduction
should be minimal in all the investigated LiMXO4 compositions
and the bulk of the conduction should originate mainly from
the ionic component. In terms of stability, some of the
compositions being sampled have experimental information in
ICSD, some of which are calculated to have low EA values.
These are LiInGeO4, LiScGeO4, LiMgPO4, and LiMgAsO4.
Some of the nonreported olivine-type materials may be
synthesized via hydrothermal methods or high-pressure
conditions, and so on. For the vacancy creation in the Li
channels, the introduction of aliovalent dopant cations either in
the M or X site is suggested. Instances where large fractions of
blockage in 1D diffusion in the olivine structure arre observed,
i.e., antisite defects could be dominated by mobility of other
species which allow crossover between different 1D channels.
Even with this alternative route, this blocking is likely to reduce
the overall Li mobility as well.31
Li-Ion Hopping Energy by PLS Modeling. Density
functional theory (DFT) and its extensions have been shown to
model the ground state of a wide variety of compounds with
acceptable accuracy. Structural parameters and their related EA
values have been shown to be efficiently and accurately
calculated by DFT methods.32−34 The relaxed structures must
first be determined in accordance with the type of site and the
known charges of the ions. The sampling scheme (see Section 1
in the Supporting Information) was based on the charge
neutrality requirement of the ideal LiMXO4 olivine structure;
therefore, only the M2+−X5+ and M3+−X4+ configurations were
considered. Group 2 (Mg2+, Ca2+, Sr2+, and Ba2+), Group 3
(Sc3+, Y3+, and rare-earth (RE3+)), and Group 13 (Al3+, Ga3+,
and In3+) ions were substituted into the octahedral M site, and
Group 14 (Si4+, Ge4+, and Sn4+) and Group 15 (P5+, As5+, and
Sb5+) ions were substituted into the tetrahedral X site. Some of
the stabilized compounds (15 M−X pairs) were then evaluated
in terms of EA values along the energetically favored b-axis by
the nudged elastic band (NEB) method within the DFT
framework.20 The accurate DFT EAs were then used as the
input response variable while the structural information
extracted from DFT-relaxed unit cells (lattice parameters,
Born effective charges, intrapolyhedron parameters, and
interpolyhedron parameters) served as predictors for training
the PLS model.
Considered in this study is the diffusing species residing
within a host and in which diffusion is mediated by site
vacancies. The energy when an ion is in an activated state
during migration depends on the surrounding configuration
and can be described very well by the information of the
stabilized lattice model with no diffusion computed by first
principles. This way, rapid calculation of EA is achievable for
1361
Article
any arbitrary composition. The different descriptors for Li-ion
hopping energy, in this case, will be called original variables
(OVs). By linearly combining these variables, formalism can be
made possible to describe the energy of ion migration, EA, of
the solid as35
n
EA = ∑ βiXi + β0
i=1 (4)
where EA represents the Li-ion hopping energy, Xi represents
the OVs, βi represents the model coefficients describing the
degree of contribution of each variable to the value of EA, and
β0 is the intercept of the function, which can be determined
directly from the model fit. The different OVs used in building
the PLS model can be grouped into structural parameters
(lattice parameters a, b, c, and cell volume), Born effective
charges of cations (for Li, M, and X), and intrapolyhedron
parameters and interpolyhedron parameters of DFT-relaxed
unit cells. Details regarding the parameters are listed in Table 2.
Visual representation for the interpolyhedron parameters
within the olivine structure is displayed in Figure 5.
The OVs are then constructed into “projected” variables or
latent variables (LVs) through linear combinations. The final
model showed that 84.13% of the variation in the response
variable Y (in this case, EA) can be explained by LV1 and
96.26% can be explained by three LVs (root-mean-square error,
RMSE = 0.316); see Section 3 of the Supporting Information.
The fit of the model shown in Figure 6 shows good accordance
between PLS and DFT. In other words, the use of linear
function formalism is justified to reproduce the DFT EA.
In order to demonstrate the prediction capability of the PLS
model, a test dataset was evaluated using M−X pairs that are
not used during the actual model building. The results are listed
in Table 3 for the DFT-derived and PLS-derived EA values and
their energy differences. It can be stated here that, indeed, there
is a significant predictive power based from the five tested
samples, namely, Al−Si, Dy−Ge, Y−Si, Ho−Si, and Sm−Ge.
The energy difference between DFT and PLS results falls
within 35 meV. The Al−Si pair is of practical interest here for
actual synthesis in which the EA value is significantly low (∼0.2
eV).
The EA values of the 66 M−X pair compositions were
predicted using the PLS function derived from the model. The
results are displayed in Figure 7, with data points sorted
according to the M and X ionic radii. The plot shows the
dependence of the Li-ion hopping energy on the ionic size of M
and X. An increase in the size of M increases the EA, whereas
an increase in the size of X leads to a decrease in EA.
Collectively, results from both DFT calculations and PLS
predictions showed low EA values for small-sized Group 13 M
cations (Al and Ga). Large-sized Group 2 cations (Sr and Ba)
and most lanthanide ions gave large barrier energy values. With
respect to our benchmark material, LiAlSiO4, LiAlGeO4,
LiGaSiO4, LiGaGeO4 and LiGaSnO4 can be treated as
candidate materials with predicted EA values of 0.23, 0.12,
0.28, 0.16, and 0.01 eV, respectively. From the standpoint of
mineral resource abundance, compositions containing Al and Si
should be regarded as a priority for new material synthesis.
We would like to point out a curious result that we
determined for the Al−Sn pair. From the PLS model, an
unrealistic value of −0.03 eV was estimated; this type of result
must be interpreted with caution in any regression fitting. In
this case, the data point was not discarded for being at the
dx.doi.org/10.1021/cm3000427 | Chem. Mater. 2012, 24, 1357−1364
Chemistry of Materials Article

Table 2. Description and Code Used for the Different PLS

Predictor Variables, Taken from Relaxed Unit Cells
symbol description
a lattice parameter a
b lattice parameter b
c lattice parameter c
v unit-cell volume
b1 born effective charge of Li ion
b2 born effective charge of octahedral M ion
b3 born effective charge of tetrahedral X ion
l1 average bond length of Li octahedron
p1 polyhedral volume of Li octahedron
i1 distortion index of Li octahedron
e1 quadratic elongation of Li octahedron
v1 bond angle variance of Li octahedron
c1 effective coordination number of Li ion
l2 average bond length of M octahedron
p2 polyhedral volume of M octahedron
Figure 5. Interpolyhedron parameters within the unit cell that were
i2 distortion index of M octahedron
used in PLS model building.
e2 quadratic elongation of M octahedron
v2 bond angle variance of M octahedron
c2 effective coordination number of M ion
l3 average bond length of X tetrahedron
p3 polyhedral volume of X tetrahedron
i3 distortion index of X tetrahedron
e3 quadratic elongation of X tetrahedron
v3 bond angle variance of X tetrahedron
c3 effective coordination number of X ion
d1 Li octahedron−M octahedron distance at edge sharing
d2 Li octahedron−X octahedron distance at edge sharing
d3 M octahedron−X tetrahedron distance at edge sharing
a1 Li−O−M angle at edge sharing
a2 Li−O−X angle at edge sharing
a3 M−Li−X angle at edge sharing Figure 6. Plot of DFT-calculated EA against PLS-predicted EA for
d4 M octahedron−X tetrahedron distance at corner sharing different M−X samples. Blue data points represent M−X pairs with
d5 Li octahedron−M octahedron distance at corner sharing published experimental data in ICSD.
d6 Li octahedron−X tetrahedron distance at corner sharing
a4 M−O−X angle at corner sharing Table 3. Comparison of DFT-Derived and PLS-Derived EA
a5 Li−O−M angle at corner sharing Values for M−X Pairs Which Are Excluded from the Model
a6 Li−O−X angle at corner sharing Building Step
d7 M−M octahedron pair distance at corner sharing
d8 M−M octahedron pair distance near midplane EA (eV)
d9 M−M octahedron pair distance along diagonal axis M−X pair DFT PLS difference
a7 O−O−O angle at corner-shared M octahedron pair Al−Si 0.1958 0.2296 −0.03379
d0 X−X tetrahedron pair distance near midplane Dy−Ge 0.3366 0.3388 −0.00225
Y−Si 0.5150 0.4824 −0.03253
extreme end of the dataset, and it was interpreted as a Ho−Si 0.4773 0.4552 +0.02526
theoretically promising composition.14 However, when NEB Sm−Ge 0.3937 0.4190 −0.02526
calculation was performed for verification, the integrity of the
olivine structure with the Al−Sn pair could not be maintained
anymore during structural relaxation when the needed Li
vacancy site for the hop was introduced. Although the original
unit cell was stabilized by the ab initio method, the structure
tended to collapse to form a more stable configuration at defect
incorporation. This was the only exception that was found
among the M−X pairings. Compositions such as LiAlSnO4,
LiGaSnO4, and LiInSnO4 may be difficult to synthesize,
because Sn does not highly favor a tetrahedral coordination.
Nevertheless, such results may still be beneficial in terms of
critical insights that could be extended when designing
Figure 7. Predicted Li-ion hopping EA for different M−X pair
materials to control microscopic and even macroscopic compositions, according to the final PLS model.
properties.
1362 dx.doi.org/10.1021/cm3000427 | Chem. Mater. 2012, 24, 1357−1364
Chemistry of Materials
The prediction of Li-ion hopping energy was initially based
on model training with 42 OVs in order to generate the first
PLS model (refer to Section 3.1 in the Supporting
Information). For the model refinement, attempts were made
to identify OVs that are of low importance to model building.
This is carried out by evaluating two important criteria: one is
the OV absolute coefficients used to fit the DFT EA values and
the other one is the parameter for variable importance in
projection (VIP) for the LV construction.36 The model
coefficient plot and VIP plot for the reduced final model
(only 33 OVs left) are illustrated in Figures 8a and 8b,
Figure 8. (a) Model coefficient plot and (b) variable importance in
projection (VIP) plot of the remaining 33 original variables for the
final PLS model (9 deleted unimportant variables).
respectively. There is no appreciable change observed after
deletion of unimportant OVs to the ordering of VIP between
the first and the reduced final model.
Aside from the efficient prediction of EA, another merit of
PLS approach is the means to deduce factors that directly affect
the EA value from the list of OVs used. Major contributions to
the EA value were identified to be from the M octahedron
descriptors, based on Figure 8, the three highest OVs (of which
are quadratic elongation of the M octahedron (e2)), the bond
angle variance of the M octahedron (v2), and the Li−O−M
angle at the edge-sharing octahedra along the diffusion channel
(a1). It is also worth noting that, in the final PLS model, LV1,
which contributes the largest magnitude for EA, has been built
from OVs that are also closely related to M octahedron
descriptors (please refer to Section 3.2 in the Supporting
Information). These agree well with our proposed crucial
parameter for Li-ion migration, which is the local lattice
distortion around the M octahedron.
■
CONCLUSIONS
In summary, we have successfully generated a robust model for
the prediction of the Li-ion hopping energy for 66 olivine-type
LiMXO4 compositions with main group M and X cations, using
a combination of ab initio computations and the PLS method.
An increase in the ionic size of M results in a significant
distortion of the M octahedron, which, in turn, leads to an
increase in the energy-penalizing local lattice distortion around
the migration pathway during the actual Li ion hop. This is
1363
Article
proposed to be the probable cause of the increase in the value
of EA. Interesting M−X pairs with relatively low Li-ion hopping
energies (<0.30 eV), such as Mg−As, Sc−Ge, In−Ge, and Mg−
P, were identified among the compositions with experimental
data from ICSD. Group 13−X pairs and Ca−X pairs were
identified as promising compositions. M−X pairs with Al and Si
cations are especially attractive candidate materials, because of
the high abundance of these elements.
A particular challenge for olivine materials, since they are
driven by the vacancy mechanism along the b-Li chain, is the
creation of the Li vacancies itself. Theoretical calculations based
from the work of Rajkumar et al. revealed that the formation
energy for the intrinsic defects, including Li-ion vacancy sites, is
relatively large (>1.5 eV).11 Hence, finding a suitable aliovalent
dopant cation in the M site should be one of the practical
approaches to creating Li vacancies. By ionic size consideration,
we are currently investigating experimentally several dopants
for the M3+ site in ICSD-reported LiInSiO4 and LiInGeO4; the
substitutions are Ti4+ and Zr4+ to create Li monovacancy and
Ta5+ and Nb5+ for Li divacancies.
■
*
ASSOCIATED CONTENT
S Supporting Information
Additional texts, five figures, and two tables with detailed
descriptions related to LiMXO4 composition sampling, first-
principle calculations, and partial least-squares analysis are
provided as Supporting Information. This material is available
free of charge via the Internet at http://pubs.acs.org.
■ AUTHOR INFORMATION
Corresponding Author
*Tel.: +81-52-735-5189. Fax: +81-52-735-5189. E-mail address:
nakayama.masanobu@nitech.ac.jp.
Notes
The authors declare no competing financial interest.
■
REFERENCES
(1) Irvine, J. T. S.; West, A. R. In High Conductivity Solid Ionic
Conductors; Takahashi, T., Ed.; World Scientific Publishing Co.:
Singapore, 1989.
(2) Thangadurai, V.; Weppner, W. Adv. Funct. Mater. 2005, 15, 107.
(3) Inaguma, Y.; Liquan, C.; Itoh, M.; Nakamura, T.; Uchida, T.;
Ikuta, H.; Wakihara, M. Solid State Commun. 1993, 86, 689.
(4) Alpen, U. V.; Schulz, H.; Talat, G. H.; Böhm, H. Solid State
Commun. 1977, 23, 911.
(5) Hong, H. Y.-P. Mater. Res. Bull. 1978, 13, 117.
(6) Goodenough, J. B.; Hong, H. Y.-P.; Kafalas, J. A. Mater. Res. Bull.
1976, 11, 203.
(7) Kang, B.; Ceder, G. Nat. Lett. 2009, 458, 190.
(8) Zhou, F. E.; Cococcioni, M.; Kang, K.; Ceder, G. Electrochem.
Commun. 2004, 6, 1144.
(9) Islam, M. S.; Driscoll, D. J.; Fisher, C. A. J.; Slater, P. R. Chem.
Mater. 2005, 17, 5085.
(10) Dai, D.; Whangbo, M.-H. Inorg. Chem. 2005, 44, 2407.
(11) Rajkumar, T.; Nakayama, M.; Nogami, M. Solid State Commun.
2010, 150, 693−696.
(12) Spiegelman, C. H.; Bennett, J. F.; Vannucci, M.; McShane, M. J.;
Cote, G. L. Anal. Chem. 2000, 72, 135.
(13) Marini, F.; Roncaglioni, A.; Novič, M. J. Chem. Inf. Model. 2005,
45, 1507.
(14) Sunnerheim, K.; Nordqvist, A.; Norlander, G.; Karlson, A.-K. B.;
Unelius, C. R.; Bohman, B.; Nordenhem, H.; Hellqvist, C.; Karlén, A.
J. Agric. Food Chem. 2007, 55, 9365.
(15) Kresse, G.; Joubert, D. Phys. Rev. B 1998, 59, 1758.
(16) Kresse, G.; Furthmüller, J. Phys. Rev. B 1996, 54, 11169.
dx.doi.org/10.1021/cm3000427 | Chem. Mater. 2012, 24, 1357−1364
Chemistry of Materials Article

(17) ICSD, Inorganic Crystal Structure Database; http://icsd.fiz-

karlsruhe.de/icsd/, 2006.
(18) Csonka, G. I.; Perdew, J. P.; Ruzsinszky, A.; Philipsen, P. H. T.;
Lebègue, S.; Paier, J.; Vydrov, O. A.; Á ngyán, J. G. Phys. Rev. B 2009,
79, 155107-1.
(19) Jonsson, H.; Mills, G.; Jacobsen, K. M. Nudged Elastic Band
Method for Finding Minimum Energy Paths of Transitions. In
Classical and Quantum Dynamics in Condensed Phase Simulations;
Berne, B. J., Ciccotti, G., Coker, D. F.; Eds.; World Scientific:
Singapore, 1998.
(20) Kishida, I.; Koyama, Y.; Kuwabara, A.; Yamamoto, T.; Oba, F.;
Tanaka, I. J. Phys. Chem. B 2006, 110, 8262.
(21) Shannon, R. D. Acta Crystallogr., Sect. A: Cryst. Phys. Diffr.,
Theor. Gen. Crystallogr. 1976, 32, 751.
(22) Robinson, K.; Gibbs, G. V.; Ribbe, P. H. Science 1971, 172, 567.
(23) Morgan, D.; Van der Ven, A.; Ceder, G. Electrochem. Solid-State
Lett. 2004, 7, A30.
(24) Nishimura, S.; Kobayashi, G.; Ohoyama, K.; Kanno, R.;
Yashima, M.; Yamada, A. Nat. Mater. 2008, 7, 707.
(25) Malik, R.; Burch, D.; Bazant, M.; Ceder, G. Nano Lett. 2010, 10,
4123.
(26) Yang, J.; Tse, J. S. J. Phys. Chem. A 2011, 115, 13045.
(27) Ghosez, Ph.; Michenaud, J.-P.; Gonze, X. Phys. Rev. B 1998, 58,
6224.
(28) Veithen, M.; Ghosez, Ph. Phys. Rev. B 2002, 65, 214302.
(29) Zhong, W.; King-Smith, R. D.; Vanderbilt, D. Phys. Rev. Lett.
1994, 72, 3618.
(30) Fehr, K. T.; Hochleitner, R.; Schmidbauer, E.; Schneider, J. Phys.
Chem. Miner. 2007, 34, 485.
(31) Yang, S. F.; Song, Y. N.; Zavalij, P. Y.; Whittingham, M. S.
Electrochem. Commun. 2002, 4, 239.
(32) Hautier, G.; Fischer, C.; Ehrlacher, V.; Jain, A.; Ceder, G. Inorg.
Chem. 2011, 50, 656.
(33) Hautier, G.; Fischer, C. C.; Jain, A.; Meuller, T.; Ceder, G.
Chem. Mater. 2010, 22, 3762.
(34) Woodley, S. M.; Catlow, R. Nat. Mater. 2008, 7, 937.
(35) Geladi, P.; Kowalski, B. R. Anal. Chim. Acta 1986, 185, 1.
(36) Chong, I.-G.; Jun, C.-H. Chemom. Intell. Lab. 2005, 78, 103.

1364 dx.doi.org/10.1021/cm3000427 | Chem. Mater. 2012, 24, 1357−1364