Artem R. Oganov and Mario Valle: How To Quantify Energy Landscapes of Solids
Artem R. Oganov and Mario Valle: How To Quantify Energy Landscapes of Solids
Artem R. Oganov and Mario Valle: How To Quantify Energy Landscapes of Solids
1
Laboratory of Crystallography, Department of Materials ETH Zurich, HCI G 515, Wolfgang-Pauli-Str.
* Now at Department of Geosciences and New York Center for Computational Science, Stony Brook
obeys any rules, and what these rules are. To answer this and related questions we use
several tools: (i) reduced energy surface and its density of states, (ii) descriptor of structure
approach can be used for rationalizing large databases of crystal structures and for tuning
notions and rules. Our analysis confirms the expectations that low-energy minima are
clustered in compact regions of configuration space („funnels“) and chemical systems tend
to have very few funnels, sometimes only one. This analysis can be applied to the physical
rule.
I. Introduction.
Energy landscapes determine the behaviour and properties of chemical systems (molecules,
fluids, solids) – such as the melting temperature, strong/fragile behavious of liquids, the
glass transition, dynamics and kinetics of structural transformations (e.g. displacive, order-
disorder and reconstructive phase transitions, protein folding); see [1,2]. However, being a
storing all values of this function on the computer would be impractical. The only way
forward is to apply some reduction procedure, whereby only the most essential information
What kind of information is extracted and how the discriminators are defined depends on a
particular purpose, and several approaches have been proposed in the past [e.g. 3-8]. We
mention in particular landscape statistics (see, e.g., [3-5]), the disconnectivity graphs [7],
orientational bond order parameters [8], and various distance metrics [6,9,10]. Previous
works focussed mainly on kinetics and physical properties of clusters, liquids and glasses.
Our focus is on crystalline solids and simple ways of translating the immense information
II. Energy surface, reduced energy surface and its density of states.
The most important piece of information is the set of energy minima (including, but not
restricted to, the global minimum). While for some purposes it is necessary also to explore
the saddle points (i.e. the transition states) of the landscape, here we analyse only the
minima – i.e. metastable and stable structures, which bear relevant chemical information,
The resulting reduced energy surface (i.e. the imaginary surface formed by local energy
minima) is much simpler, and often has an overall funnel-like shape (Fig. 1): the further a
local minimum is from the global minimum, the higher its energy. In more complex cases,
elements in the structure. The emergence of funnels on the energy landscape can be
The reduced energy surface is still multidimensional and overwhelmingly complex. Below
we will show how to map it in one or few dimensions and to establish whether it has one or
more energy funnels. Before that, we would like to dwell briefly on the useful information
contained in the density of states (DOS) of the reduced energy surface – see Fig. 2.
Such DOS plots provide a wealth of important information. Typically, there is a tall
liquid-like structures, and one or several smaller maxima at lower energies separated by
gaps (denoted as “region 1” in Fig. 2) and corresponding to ordered crystal structures and
their defective variants1. The melting temperature of the crystal is related to the energy
As system size increases, the number of possible energy minima increases exponentially.
One might think that with more degrees of freedom, larger systems would have more
uniform DOS with filled energy gaps. However, just the opposite happens (Fig. 2b): the
liquid-like maximum (“region 2” of the DOS) becomes higher (at the expense of ordered
structures, which become more rare), narrower and more Gaussian-like (as argued in Ref. 5
1
At sufficiently large system size, structures corresponding to the 2nd, 3rd, ... lowest energy minima will
simply be defective versions of the lowest-energy structure.
using the central limit theorem), and shifted to higher energies (disorder costs energy) – see
Fig. 2b. What we observe by examining energies and fingerprint functions (defined below)
is that all disordered structures become nearly identical for large system sizes. In the limit
of infinite system size all randomly produced structures will be identical and representative
of the amorphous state, DOS will look like a delta-function, and the corresponding energy
can be expected to correlate with the melting temperature. Finding the ground state with
1. Fingerprint functions.
Here we define a function (actually, one can propose a number of similarly defined
functions), which we call a fingerprint function and which uniquely characterises the
structure. Such a function should be (i) derived from the structure itself, rather than its
properties (such as energy), (ii) invariant with respect to shifts of the coordinate system,
rotations and reflections, (iii) sensitive to different orderings of the atoms (e.g. one should
be able to distinguish between different orderings in f.c.c. alloys), (iv) formally related
[5,9,10]. The definition introduced in [15] involves a sum over all interatomic distances Rij:
N bi b j
Φ( R) = ∑∑ δ ( R − Rij ) − 1 (1)
∑ N i N j bi b j
cell
i , cell j
4πRij
2
Nj
Δ
V
The factor before the double sum is just a normalisation factor involving the total number N
of atoms in the unit cell, the numbers Ni and Nj of atoms of each type and their scalar
atomic properties b (which can be set, e.g., to the atomic numbers, electronegativities,
chemical scale or Mendeleev numbers [13,14], or neutron scattering lengths). The double
sum runs over all atoms i within the unit cell and j within the threshold distance Rmax of the
i-th atom. V the unit cell volume, and δ ( R − Rij ) is a Gaussian-smeared delta-function,
absorbing numerical errors and making Φ(R) a smooth function. To be robust against
numerical errors and computationally convenient, the Φ(R) function is discretised over bins
of width Δ. This Φ-fingerprint is short-ranged, i.e. tends to zero at infinite R, and Φ(0) = -1.
When b-parameters in (1) are set to atomic neutron scattering lengths, Φ(R) is related to the
structure factor:
sin(QR)
S (Q) = 4π ∫ R 2Φ ( R ) dR (2)
QR
When atomic numbers are used as b-parameters, relation (2) gives a zeroth-order
This fingerprint, however, is not very sensitive to ordering of the atoms on a given structure
components of the fingerprint function coming from different atomic type pairs A-B,
δ ( R − Rij )
FAB(R) = ∑∑ NANB
− 1 = g AB ( R) − 1 , (3)
Ai , cell B j
4πR 2
ij Δ
V cell
where the double sum runs over all i-th atoms of type A within the unit cell and all j-th
atoms of type B within the distance Rmax. In eq. (3), gAB(R) is the pair correlation function;
subtracting 1 from it makes it short-ranged. The F-fingerprint components have the same
components can be obtained from diffraction experiments. F-fingerprint does not contain
∫F
0
AB ( R) R 2 dR = 0 (4)
Discretisation of fingerprints (1) and (3) allows them to be represented as vectors, values of
the fingerprint FAB(k) in each k-th bin (of width Δ) being vector coordinates. The
dimensionality of the fingerprint space equals the number of bins. Each structure can thus
between structures can be calculated using one of distance definitions given below.
When there is only one atomic type, both Φ- and F-fingerprints reduce to g(R)-1 (where
g(R) is the pair correlation function), and the Minkowski distance is:
1/ p
⎡ p⎤
DMinkowski = ⎢∑ ( F1 (k ) − F2 (k )) ⎥ (5)
⎣k ⎦
When p = 2, Minkowski distances are the usual Cartesian distances. However, we found
1 F *F 1
∑ F (k )F (k )
1 2
Dcosine = (1 − 1 2 ) = (1 − k
) (6a)
2 F1 F2 2 ∑F
k
1
2
(k ) ∑F
k
2
2
(k )
are more robust to the “curse of dimensionality” effect of distance concentration2. They are
related to Cartesian distances in a non-linear way, but conveniently can only take values
between 0 and 1 and are much more robust to small numerical errors. Cosine distances
2
The high-dimensional space is almost empty and the whole concept of neighborhood is meaningless in high
dimensions. Randomly generated points in highly-dimensional spaces tend to be all at the same distance (the
relative contrast (dmax-dmin)/dmin → 0 as dimensionality tends to infinity) and under small perturbation the
nearest point could change into the farthest one. These phenomena are not intrinsic to the high-dimensional
space, but depend strongly on the distance function used. In this respect cosine distance are superior to
Cartesian distance (or in general the Minkowski distances with p > 1)
measure the similarity between structures using the angle between their F-vectors. For
finite systems two F-vectors can be parallel only if they are identical. The generalisation of
1
∑∑ F 1, AB (k ) F2, AB (k ) w AB
Dcosine = (1 − AB k
) , (6b)
2 ∑∑ F
AB k
2
1, AB (k ) w AB ∑∑ F
AB k
2
2 , AB (k ) w AB
where the sums are over all FAB components, taking each component (note that FAB=FBA)
only once. The importance weight wAB of each fingerprint component is defined as:
NANB
w AB = (7)
∑ NANB
cell
Computing the distance between two structures one quantifies the difference between them.
This can be used for analyzing crystal structure databases and for visualizing results of
3. Degree of order.
Both Φ- and F-fingerprints describe correlations between atomic positions– i.e. the non-
randomness of the structure, responsible for its diffraction signal (eq. (2)). For an ideal gas
both Φ- and F- fingerprints are zero, and therefore their deviation from zero level can be
where λ is a characteristic length (for instance, λ=R0, the distance at which F(R) first
becomes zero3). While the angles between fingerprint vectors measure structural
differences, the lengths of these vectors show the degree of order of each structure. For an
3
An alternative useful definition would be λ=V1/3.
ideal gas Π = 0, for a gas of hard-sphere atoms Π = 1, for solids Π > 1. We use an
alternative definition of the degree of order: modifying the sum rule (4) to obtain the next
P = ∑ w AB PAB , (9b)
A, B
Roughly, P is related to the total scattering power of the structure (cf. eq. (2)), and simpler
structures have higher P values. As we will show below on concrete cases, statistically, the
higher P the lower the energy of the structure and therefore the degree of order P is an
We can relate the total energy to F-fingerprint in the approximation that the total energy is
∞
1 4π
E= ∑
2 A, B
U AB ( R ) =E RAN +
2
∑∫ F
A, B 0
AB ( R )w ABU AB ( R ) R 2 dR , (10)
where ERAN is the energy of a totally random structure with the same volume. Since ERAN is
a function of volume only, energy differences at constant volume can be written as:
∞
ΔEV = 2π ∑ ∫ ΔFAB ( R )w ABU AB ( R) R 2 dR (11)
A, B 0
This expression shows what is intuitively expected - the more similar a structure is to the
Based on our distance metric, and noting that cosine distances, just like occupation
as a measure of collective diversity. Calculating Scoll for a set of structures gives a single
number measuring the diversity of that set. This is very useful for tuning global
For each given structure, one may also compute its own quasi-entropy as a measure of
disorder and complexity of that structure (alternative to definitions (8) and (9)):
NA
S str = −∑ ln(1 − D Ai A j ) , (13)
A N cell
where distances DAi A j are measured between fingerprints4 of all i-th and j-th sites occupied
by chemical species A, and the total quasi-entropy is a weighted sum over all chemical
inequivalent atomic positions, but does not provide any measure of physical difference
We have proposed two families of closely related fingerprint functions (1) and (3). Based
structures, and a new parameter (9ab) quantifying the simplicity and order of a crystal
4
For this, one needs to compute a fingerprint for each i-th atom (rather than atomic species), defined
δ ( R − Rij )
by analogy with eq. (3) as FAi B = ∑ NB
−1.
Bj
4πR 2
ij Δ
V cell
structure. Quasi-entropies (12) and (13) can be used for analysing the performance of
global optimization methods and provide another measure of structural complexity. Below
we show how these quantities can be used to map energy landscapes of solids and
Chemical intuition tells us that since low-energy structures share many similarities (similar
coordination numbers, bond lengths and angles, sometimes even whole structural blocks),
they have to be clustered in the same region of parameter space. This would lead to an
overall bowl-like shape of the reduced energy surface (Fig. 1) and would have many
consequences of paramount importance. First, this would imply that low-energy structures
tend to be connected by relatively low energy barriers and short transition pathways (Bell-
Evans-Polanyi principle [24]). If there are several funnels, transitions between them are
likely to involve large activation barriers. Second, funnel-like topology of the landscape is
important for modern structure prediction methods (e.g. the evolutionary algorithm [16-19]
If the funnel-like shape of the reduced energy landscape is valid, there would be a
correlation between the energy and the distance from the global minimum: the further from
the global minimum, the higher the energy. This correlation can now be easily checked and
5
In all illustrations given below, we used F-fingerprints (eq. (3)) with Rmax=15 Å, Δ = 0.05 Å, and Gaussian-
smeared δ-function with σ = 0.075 Å. We use cosine distances and degree of order P (eq. 9a,b) throughout.
The data analysed here include 2997 distinct local minima for GaAs (8 atoms/cell), 249 for Au8Pd4 (12
atoms/cell), 967 for MgO (32 atoms/cell), 949 for H2O (12 atoms/cell), 6977 for MgNH (12 atoms/cell), and
1949 distinct local minima for the AB2 Lennard-Jones crystal. All data were obtained in evolutionary and
random sampling runs, duplicate minima were pruned using a clustering procedure described in [15]. Local
optimizations for most systems were done within the generalized gradient approximation [20] (only for
Au8Pd4, calculations were done within the local density approximation) using the VASP code [21].
Exceptions are MgO, for which we used the interatomic potential [22] and Lennard-Jones (see caption to Fig.
2 for the description of the potential model) systems – for both, structure relaxations and energy calculations
were performed using the GULP code [23].
turns out to be excellent in many chemically simple systems – this is shown in Fig. 5 for
At first surprisingly, for MgO with 32 atoms/cell the landscape turns out to be more
complex and contains three broad funnels with 6-, 5- and 4-coordinate Mg atoms,
respectively (Fig. 7a). The potential [22] used here correctly finds the ground state to have
the NaCl-type structure, and among the lowest-energy local minima we find an exciting
metastable structure with both Mg and O atoms in the fivefold coordination (Fig. 7b). This
structure was previously predicted in [26], but has not yet been identified experimentally.
Among other low-energy metastable structures are zincblende, wurtzite and their
polytypes. The ability of Mg atoms to adopt very different coordination numbers (ranging
from 4 to 8) in oxides and silicates is well known to mineralogists, and is the key to the
multi-funnel structure of the landscape. Fig. 7a shows that the funnels almost “touch” each
other and there are low-energy structures close to the boundaries between funnels. This
means that there may be relatively low-energy transition pathways between the funnels,
and the landscape can be considered intermediate between single-funnel and multi-funnel
types. The non-monotonic energy-distance plot reflects the complicated topology of the
12 atoms/cell. This landscape contains (at least) three clearly separated and chemically
distinct funnels (Fig. 8). These funnels have very different energies – structures in the
deepest funnel consist of H2O molecules, whereas structures belonging to the two higher-
energy funnels contain various mixtures of molecular groups (H2O, H2, O2, OH-, many
structures contain H2O2 and some H2O3 molecules). In many structures we find OH- ions,
which is at variance with previous calculations [27] where these ions were not seen at all.
The energy-distance correlation is clearly seen across the landscape and, of course, within
Among the systems we explored, the most complex landscapes are found for MgNH and
AB2 Lennard-Jones crystal with the potential described in caption to Fig. 2. Both systems,
though small (12 atoms/cell), possess remarkable complexity with multiple degeneracies.
For instance, for MgNH we found 25 structures with energies less than 2.5 meV/atom
above the ground state. Most of these structures contain NH2- groups, though some have
(NH2)- and N3- ions - low energies of the latter structures suggest (in agreement with
experiment) that MgNH is only marginally stable with respect to decomposition into
Mg3N2+Mg(NH2)2. Generally, the existence of funnels with energies close to the ground
state indicates that the system is close either to a phase transition or to decomposition. The
2D-map of the landscape and the energy-distance correlation plot for MgNH (Fig. 9a) show
low-energy funnels, some of which show very obvious structural differences. One of the
funnels consists entirely of layered structures, whereas in another funnel all structures
contain either parallel or antiparallel NH- groups. Finally, we note the gap in the energy
contain atomic/ionic hydrogen, the formation of which requires breaking N-H bonds.
For the AB2 Lennard-Jones crystal the energy-distance plot (Fig. 9b) has several flat
will see in the next section that several families of structures exist here, and within each
family the energy is nearly constant. The degeneracies originate from the short-rangedness
of interatomic interactions, where the length scale of typical geometric features (in this
case, interplanar distances) is longer than the range of interatomic interactions, and in the
competition between different interactions (in this model, A-A, B-B and A-B interactions
have the same strength and thus compete). Such degeneracies and complex energy
landscapes (probably less extreme than the case presented in Fig. 9b) may be expected in
systems forming quasicrystals and incommensurate phases, as well as systems unstable (as
the AB2 crystal here) or nearly unstable (as MgNH) agaist decomposition.
Pauling’s fifth rule [28] states that “the number of essentially different kinds of constituents
in a crystal tends to be small” – in other words, structures tend to be simple. Some have
accepted this rule, some criticized it. Our approach enables its systematic analysis from the
viewpoint of structural energies. If Pauling’s rule is valid, one should see a clear correlation
between the energy and degree of order. Finding cases where the correlation breaks down
A vast majority of cases that we analysed shows an excellent correlation between the
energy and the degree of order. This correlation is equally good for single- and multi-
funneled landscapes; in other words, it is more fundamental. For GaAs, MgO and H2O
ground states are the most ordered structures among the multitude of structures that we
generated. For Au8Pd4 the ground state is among very few most ordered structures, and
even for heavily frustrated systems, such as MgNH and AB2 Lennard-Jones crystal, the
ground states still belong to a small number of most ordered structures. Fig. 10 shows such
correlation plots for MgO, Au8Pd4, H2O and Lennard-Jones AB2 crystal.
Furthermore, we see that all chemically interesting structures are on the high-order side of
the correlation plots. We illustrate this point for MgO (Fig. 10a), where a number of
interesting and even unexpected low-energy ordered structures were found with the help of
such a correlation plot. For AB2 Lennard-Jones crystal, we again see families of structures
with nearly identical energies. The two most ordered families include a family of defective
NaCl-type structures (cubic close-packings of the B atoms, with a half of octahedral voids
occupies by the A atoms – the well-known CdCl2 structure type also belongs to this family)
and a family of layered structures, where the B atoms form a defective hexagonal close
packing, interrupted by layers of the A atoms in the trigonal prismatic coordination (Fig.
11). It is the latter family that contains several degenerate ground-state structures. The A-
layers are topologically identical to the graphene sheet, and there is a striking structural
similarity between the ground states of the AB2 Lennard-Jones crystal and Al-C alloys
[29]. Another similarity is that both Al-C alloys and the AB2 Lennard-Jones crystal are
Recently, Hart [30] found that for the special case of ordering in alloys the simplest
ordering schemes have the highest probability of occurring6. Here we have generalised this
conclusion to all structures (i.e. beyond the special case of ordering in alloys) and
confirmed it by analysing structural energies. We can conclude that: The ground state
normally adopts one of the simplest structures compatible with the chemistry of the
compound. Such structures tend to have lower energies. This rule is nothing else than an
energetics-based reformulation of Pauling’s fifth rule. The main limitation of this rule is the
presence of competing interactions, as in the case of AB2 Lennard-Jones crystal (Fig. 10 e).
Structural quasi-entropy Sstr, which is zero when each atomic species occupies only one
Wyckoff orbit and increases as the physical difference between the occupied sites
increases, is another and perhaps even more direct measure of structural simplicity. It more
directly relates to the original [28] formulation of Pauling’s 5th rule. Unlike the degree of
order, Sstr does not depend critically on the smoothing function δ(R-Rij) in the fingerprint
definition (eqs. 1,3) and its absolute (not only relative) values are meaningful. For all cases
investigated here, we found an excellent correlation between the energy and quasi-entropy
6
To quantify the simplicity of an alloy structure, Hart [30] introduced an index measuring the non-
randomness of the distribution of atoms over sites. His index is valid only for the case of ordering in
alloys. Our degree of order (eq. 9a,b) provides a universal criterion, equally valid for ordering in alloys
of a given structure, as well as for the comparison of completely different structures.
Sstr – this is shown in Fig. 12 for MgO (32 atoms/cell) and MgNH (12 atoms/cell). Like
with the degree of order P, the correlation is worst for the AB2 Lennard-Jones crystal, the
reason being its instability to decomposition that leads to long-period layered structures
that contain quite diverse atomic sites. The same can be expected for other complex and
frustrated systems. A good example of which is the recently discovered [31] high-pressure
stable phase of boron, the structure of which contains atomic sites that are so different that
there is charge transfer between them [31] and the structure possesses large Sstr=0.18.
occupy. Our conclusion is that: In the ground state and low-energy structures, atoms of
The degree of order P and structural quasi-entropy Sstr are not the only possible measures
of structural simplicity. One alternative is the set of orientational bond order parameters Qn
and crystal-like configurations and could be expected to be good predictors of the energy.
However, for all systems examined here (even as simple as MgO) the correlation of these
parameters with the energy is either very weak or non-existent – Fig. 13 shows this for the
Q6 parameter (we also checked for Q4, as well as W4 and W6 parameters [3132], and arrived
at the same conclusion). Thus, the degree of order P and structural quasi-entropy Sstr are
strong predictors of the energy, whereas the orientational bond order parameters Qn and Wn
are not.
Conclusions.
landscapes of crystalline systems. These tools can be used for rationalising structural and
valuable for many purposes – as two examples, we mention monitoring of the progress of
structure prediction simulations [16] and the emerging field of crystallographic genomics
[33], where such analysis is central for the rationalisation of large databases of crystal
structures.
Typical dimensionalities of fingerprint vector spaces for cases studied here, 102-103, are
atoms in the unit cell). However, even d overestimates the true dimensionality of the
landscape, because it ignores short-range order that leads to certain constraints κ in relative
d* = 3N + 3 – κ (14)
is generally a non-integer number and can be computed from the distance distribution.
(d = 39) for Au8Pd4, d* = 32.5 (d = 39) for MgNH and d* = 11.6 (d = 99) for MgO. The
intrinsic dimensionality gives the minimum number of dimensions sufficient for mapping
the data exactly, but approximate mappings can be done in lower dimensions.
(energy-distance plots extensively used here) to 2 (landscape “maps” also presented here)
to as many dimensions as needed, such analysis leads to a new level of chemical insight.
Based on the fingerprint function, we introduced two indices of structural simplicity, the
degree of order P (eq. 9ab) and structural quasi-entropy Sstr (eq. 13), which enabled us to
directly verify the well-known Pauling’s fifth rule (the parsimony rule): statistically,
simpler structures tend to have lower energies. Parameters P and Sstr are powerful tools for
analyzing results of crystal structure prediction simulations (see [15]) and crystal structure
databases. This analysis can be brought further: for instance, energy-order correlations for
order components PAB (eq. 9b) may indicate the dominant structure-forming interactions
The tools developed here are helpful in monitoring structure prediction simulations, as
shown in Fig. 14. Furthermore, these tools, as well as improved understanding of energy
landscapes, can be (and already have been [35]) used to improve performance of the
The same analysis can be done for the physical properties of crystals, leading to the
observe funnel-shapes landscapes, which would indicate types of structures that are of
particular interest. Many properties (e.g. the electrical conductivity) are expected to be well
correlated with the degree of order P and structural quasi-entropy Sstr. The analysis
presented here can thus provide the general link between the structure, properties and
stability of solids.
Acknowledgments. The formalism discussed here has been implemented by M.V. in the
work was supported by Swiss National Science Foundation (grants 200021-111847/1 and
200021-116219).
References:
1. Stillinger F.H. (1995). A topographic view of supercooled liquids and glass formation. Science. 267,
1935-1939.
2. Wales D.J., Miller M.A., Walsh T.R. (1998). Archetypal energy landscapes. Nature 394, 758-760.
3. Angelani L., Ruocco G., Sampoli M. (2003). General features of the energy landscape in Lennard-Jones-
4. Shah P., Chakravarty C. (2002). Potential-energy landscapes of simple solids. Phys. Rev. Lett. 88,
255501.
5. Büchner S., Heuer A. (1999). Potential energy landscape of a model glass former: Thermodynamics,
6. Heuer A. (1997). Properties of a glass-forming system as derived from its potential energy landscape.
7. Becker O.M., Karplus M. (1997). The topology of multidimensional potential energy surfaces: theory
and application to peptide structure and kinetics. J Chem. Phys. 106, 1495-1517.
8. Steinhardt P.J., Nelson D.R., Ronchetti M. (1983) Bond-orientational order in liquids and glasses. Phys.
9. Willighagen E.L., Wehrens R., Verwer P., de Gelder R., and Buydens L.M.C. (2005). Method for the
10. de Gelder R. (2006). Quantifying the similarity of crystal structures. IUCr CompComm Newsletter, 7,
59–69.
11. Doye J.P.K. (2002). Network topology of a potential energy landscape: a static scale-free network. Phys.
12. Doye J.P.K., Miller M.A., Wales D.J. (1999). The double-funnel energy landscape of the 38-atom
13. Pettifor D.G. (1984). A chemical scale for crystal structure maps. Solid State Commun. 51, 31-34.
14. Pettifor D.G. (1986). The structures of binary compounds: I. Phenomenological structure maps. J. Phys.
C19, 285-313.
15. Valle M., Oganov A.R. (2008). Crystal structure classifier for an evolutionary algorithm structure
predictor. IEEE Symposium on Visual Analytics Science and Technology (October 21 - 23, Columbus,
of the USPEX method and some of its applications. Psi-k Newsletter, number 84, Highlight of the
Month, 142-171.
17. Oganov A.R., Glass C.W., Ono S. (2006). High-pressure phases of CaCO3: crystal structure prediction
18. Oganov A.R., Glass C.W. (2006). Crystal structure prediction using ab initio evolutionary techniques:
19. Glass C.W., Oganov A.R., Hansen N. (2006). USPEX – evolutionary crystal structure prediction. Comp.
20. Perdew J.P., Burke K., Ernzerhof M. (1996). Generalized gradient approximation made simple. Phys.
21. Kresse G. & Furthmüller J. (1996). Efficiency of ab initio total-energy caclulations for metals and
22. Lewis G.V. & Catlow C.R.A. (1985). Potential models for ionic oxides. J. Phys. C.: Solid State Phys. 18,
1149-1161.
23. Gale J.D. (2005). GULP: Capabilities and prospects. Z. Krist. 220, 552-554.
24. Roy S., Goedecker S., Hellmann V. (2008). Bell-Evans-Polanyi principle for molecular dynamics
trajectories and its implications for global optimization. Phys. Rev. E77, 056707.
25. Oganov A.R., Ma Y., Glass C.W., Valle M. (2007). Evolutionary crystal structure prediction: overview
of the USPEX method and some of its applications. Psi-k Newsletter, number 84, Highlight of the
Month, 142-171.
26. Schon J.C., Jansen M. (1996). First step towards planning of syntheses in solid-state chemistry:
determination of promising structure candidates by global optimisation. Angew. Chem. – Int. Ed. 35,
1287-1304.
27. Pickard C.J., Needs R.J. (2007). When is H2O not water? J. Chem. Phys. 127, 244503.
28. Pauling L. (1928). The principles determining the structure of complex ionic crystals. J. Am. Chem. Soc.
51, 1010-1026.
29. Oganov A.R., Glass C.W. (2008). Evolutionary crystal structure prediction as a tool in materials design.
30. Hart G.L.W. (2007). Where are Nature's missing structures? Nature Materials 6, 941-945.
31. Oganov A.R., Chen J., Gatti C., Ma Y.-M., Yu T., Liu Z., Glass C.W., Ma Y.-Z., Kurakevich O.O.,
Solozhenko V.L. (2009). Ionic high-pressure form of elemental boron. Nature, in press.
32. van Duijneveldt J.S., Frenkel D. (1992). Computer simulation study of free energy barriers in crystal
33. Parkin A., Barr G., Dong W., Gilmore C.J., Jayatilaka D., MCKinnon J.J., Spackman M.A., Wilson C.C.
(2007). Comparing entire crystal structures: structural genetic fingerprinting. CrystEngComm 9, 648-652.
34. Camastra F., Vinciarelli A. (2001). Intrinsic dimension estimation of data: an approach based on
35. Lyakhov A.O., Oganov A.R., Valle M. (2009). How to predict large and complex crystal structures. In
prep.
FIGURE CAPTIONS:
Fig. 1. Conceptual depiction of an energy landscape (solid line) and reduced energy
Fig. 2. DOS of energy minima for (a) GaAs with 8 atoms/cell and (b) binary Lennard-
Jones crystal AB2 for several system sizes. Energies are shown relative to the ground
state. Data were obtained by locally optimising random structures, without pruning
identical structures. In (a), 3000 random structures were sampled, while in (b) we typically
sampled 5000 random structures at each system size. Calculation (a) was done using the
generalized gradient approximation [20] of density functional theory. In (b), for each
⎡ Rmin,ij 12 Rmin,ij 6 ⎤
atomic pair the Lennard-Jones potential was written as U ij = ε ij ⎢( ) − 2( ) ⎥,
⎣ R R ⎦
where Rmin,ij is the distance at which the potential reaches minimum, and ε is the depth of
the minimum). For all pairs we used the same ε, but different ideal lengths:
interactions leads to very complex energy landscapes and non-trivial ground states.
[16].
prediction simulation for Au8Pd4 (with 12 atoms/cell) at 1 atm. Each (N,M) pixel shows
the distance between the N-th and M-th structures. Note the increase of similarity between
the structures towards the end of the simulation – this is a consequence of “learning” in
evolutionary simulations. This matrix is by construction symmetric with respect to its
diagonal, and all diagonal (N,N) elements are zero (structure N is identical to itself).
atoms/cell). The minima were found in several ab initio evolutionary and random
sampling simulations. Energy differences (per unit cell) and distances in this and all
obtained in an ab initio evolutionary run [24]. Most structures are different decorations
Fig. 7. Energy landscape for MgO with 32 atoms/cell. (a) 2D-mapping of the landscape,
where each point represents a structure and distances between points on the graph are
maximally close to the distances between the corresponding fingerprints7. Darker points
and O atoms in the trigonal bipyramidal coordination. It has the space group P63/mmc, cell
parameters a=b=3.46 Å and c=4.18 Å, and atomic positions Mg (1/3, 2/3, 3/4) and O (2/3,
Fig. 8. Energy landscape of H2O with 12 atoms/cell. (a) 2D-map of the landscape with
darker points indicating lower-energy structures. (b) Energy-distance correlation. The inset
7
Of course, such mapping with reduced dimensionality cannot fully preserve distances (for the same
reason, 2D-maps of the world show distorted distances). However, the topology of the landscape is
correctly reproduced.
Fig. 9. Complex energy landscapes : (a) MgNH (12 atoms/cell) and (b) AB2 Lennard-
Fig. 10. Energy-order correlation plots. (a) MgO (32 atoms/cell), (b) Au8Pd4 (12
atoms/cell), (c) H2O (12 atoms/cell), with inset showing only the 580 structures containing
only H2O molecules, (d) MgNH (12 atoms/cell) and (2) AB2 Lennard-Jones crystal (12
Fig. 11. Families of nearly degenerate structures for AB2 Lennard-Jones crystal (12
atoms/cell). (a-c) ground states with layers of the A atoms in the trigonal prismatic
structure we found.
Fig. 12. Energy – structural quasi-entropy Sstr correlation plots. (a) MgO (32
atoms/cell), (b) MgNH (12 atoms/cell). Note a clear energetic preference for simple
Fig. 13. Energy-Q6 correlation plots. (a) MgO (32 atoms/cell), (b) AB2 Lennard-Jones
crystal (12 atoms/cell). It is obvious that, unlike the degree of order P (eq. 9ab) or
structural quasi-entropy (eq. 13), the orientational bond order parameter Q6 is not a good
predictor of energy.
this system (the ground state was found only once after relaxing 5000 random structures),
the evolutionary algorithm USPEX [1619] finds the ground state after only 110 structure
relaxations. The initial population consisted of 50 random structures, and each subsequent
generation consisted of 10 best old non-identical structures and 10 new structures produced
from them using the variation operators described in [1619]. In each generation the
collective quasi-entropy Scoll was computed for a set of structures that were used for
producing the next generation. Scoll should not decrease too fast, and when it becomes
sufficiently small, the simulation can be terminated. Inset shows the lowest energy as a
Fig. 3.
Fig. 4.
Fig. 5.
Fig. 6.
Fig. 7.
Fig. 8.
Fig. 9.
Fig. 10.
Fig. 11.
Fig. 12.
Fig. 13.
Fig. 14.