Professional Documents
Culture Documents
Geometry Optimization of Periodic Systems Using Internal Coordinates
Geometry Optimization of Periodic Systems Using Internal Coordinates
An algorithm is proposed for the structural optimization of periodic systems in internal 共chemical兲
coordinates. Internal coordinates may include in addition to the usual bond lengths, bond angles,
out-of-plane and dihedral angles, various “lattice internal coordinates” such as cell edge lengths, cell
angles, cell volume, etc. The coordinate transformations between Cartesian 共or fractional兲 and
internal coordinates are performed by a generalized Wilson B-matrix, which in contrast to the
previous formulation by Kudin et al. 关J. Chem. Phys. 114, 2919 共2001兲兴 includes the explicit
dependence of the lattice parameters on the positions of all unit cell atoms. The performance of the
method, including constrained optimizations, is demonstrated on several examples, such as layered
and microporous materials 共gibbsite and chabazite兲 as well as the urea molecular crystal. The
calculations used energies and forces from the ab initio density functional theory plane wave
method in the projector-augmented wave formalism. © 2005 American Institute of Physics.
关DOI: 10.1063/1.1864932兴
Downloaded 19 Jul 2006 to 194.214.217.17. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
124508-2 Bučko, Hafner, and Ángyán J. Chem. Phys. 122, 124508 共2005兲
II. METHOD 1 dE
␣ = − 共6兲
A. Cartesian and fractional coordinates ⍀ d␣
The structure of a periodic system is characterized by by the relationship 共cf. the Appendix兲
three lattice vectors, arranged in the matrix h = 关a1 , a2 , a3兴
dE
and by the 3N fractional coordinates, s = 兵s␣a 其. The Cartesian
dh␣
=−⍀ 兺␥ ␣␥共ht兲␥
−1
. 共7兲
positions of atom a in the L = 共l1 , l2 , l3兲 unit cell of the solid
is given by the linear transformation Electronic structure codes usually provide Cartesian force
3 components, which can be transformed to fractional coordi-
r␣a,La = 兺
=1
h␣共sa + la 兲. 共1兲 nate forces as
dE E ra E
The three lattice vectors form, in the general case, an unnor- ds␣a
= 兺 ra sa = 兺 兺␥ ra sa h␥共s␥a + l␥a 兲
 ␣  ␣
malized and nonorthogonal basis. The fractional coordinates
can be obtained from the Cartesian ones by the inverse trans- E
formation in terms of the reciprocal lattice vectors, 共ht兲−1 = 兺 ra h␣ . 共8兲

= 关a*1 , a*2 , a*3兴, defined by a*i · a j = ␦ij as
vectors is conveniently described by the strain tensor ␣ as As we shall see later, in Sec. II D, the same relationships can
兺␥ 共␦␣␥ + ␣␥兲h␥ .
be obtained directly from the B-matrix formalism, to be de-
⬘ =
h␣ 共3兲
veloped below.
The extended Hessian matrix F in Eq. 共5兲 is built up
Any change in h␣ implies a variation of atomic Cartesian from several blocks: that of the second derivatives of the
coordinates by Eq. 共1兲 which obey an analogous relationship energy with respect to atomic positions 共related to the dy-
namical matrix兲, that of the stress-strain derivatives, as well
r␣⬘ = 兺 共␦␣ + ␣兲r . 共4兲 as cross terms.
Downloaded 19 Jul 2006 to 194.214.217.17. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
124508-3 Geometry optimization of periodic systems J. Chem. Phys. 122, 124508 共2005兲
冉 冊冉 冊冉 冊
Even for small systems, the number of generated internal
coordinates is usually much larger than the number of ionic ␦q Bqs Bqh ␦s
= q̂s q̂h , 共21兲
degrees of freedom 共i.e., 3N − 6 for molecules and 3N + 3 for ␦q̂ B B ␦h
systems with three-dimensional periodicity兲. The handling of
this set of redundant coordinates may lead to a significant where the individual blocks Bqs and Bq̂s describe the linear
increase of the computational time and may cause conver- transformation of the atomic positions s␣a , while the blocks
gence problems in the coordinate back-transformation step. Bqh and Bq̂h describe transformations involving the lattice
To avoid these problems, Baker et al. proposed the use of vector components h␣ to atomic qi and lattice internal q̂ j
nonredundant linear combinations of primitive internal coordinates. The blocks of the augmented B-matrix can be
coordinates.24 Taking the singular value decomposition of calculated from the relationships 共1兲 and 共2兲 using the chain
the B-matrix, as B = UtBDV, where U and V are unitary ma- rule.
trices and BD is a diagonal matrix formed from the eigenvec- An internal coordinate may depend on the Cartesian co-
tors associated with the nonzero eigenvalues of 共BBt兲. After ordinates of atoms in different unit cells. Therefore the
multiplication of the B-matrix equation from left by U, one B-matrix elements between fractional atomic distortions and
obtains unique internal coordinates should be calculated as
Downloaded 19 Jul 2006 to 194.214.217.17. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
124508-4 Bučko, Hafner, and Ángyán J. Chem. Phys. 122, 124508 共2005兲
q
= 兺 a,Li 共sa + la 兲,
a,L r
a
共23兲 E. Constraints
a ␣
Constrained geometry optimizations are helpful and
where the superscript qh refers to the type of B-matrix ele- even necessary in most of the applications of ab initio cal-
ment. culations to chemical reactions, phase transitions, etc. The
The “lattice internal” coordinates q̂, such as cell edge general strategy is to make vanish, exactly or at least ap-
lengths, a / b ratio, cell angles, volume, etc., do not depend proximately, forces along the constrained coordinates to
explicitly on the fractional atomic positions, therefore the avoid deformations involving the variation of the constrained
block of the extended B-matrix is zero, Bq̂s = 0. The transfor- coordinates. The principal advantage of using internal coor-
mation between the lattice vector changes and various “in- dinates is that one can impose exact internal coordinate con-
ternal” lattice parameters is described by the matrix elements straints during the optimization. Various algorithms, such as
the use of projection operator,30 orthogonalization,24 and
q̂i Lagrange multiplier31 techniques have been proposed in the
Bi,q̂h␣ = . 共24兲
h␣ past. Our implementation follows essentially the orthogonal-
ization algorithm of Ref. 24.
For instance, the following relationship holds for q̂i = ⍀, the
In the first step, the B-matrix is modified in such a way
unit cell volume:
that first derivatives of the active coordinates 共those coordi-
nates which are allowed to be relaxed兲 are orthogonalized
␣ = ⍀共h 兲␣ . 共25兲
q̂h t −1
B⍀,
with respect to each constrained coordinate qc. The rows B j
Other specific lattice B-matrix elements can also be derived are modified according to the formula
from the definition of lattice vector lengths and lattice 共B j · Bc兲 Bc
angles. B̄ j = B j − 兺c 兩Bc兩 兩Bc兩
, 共30兲
In order to appreciate the role of the extended B-matrix
in lattice optimizations, let us consider the special case of where the summation is over the constrained coordinates. If
Cartesian coordinates as internals, i.e., ␦ = 兵␦r , ␦h其. Accord- one of the rows of the original B-matrix is identical to a
ing to the coordinate transformation relationship, Eq. 共13兲, constrained coordinate, it is exactly annihilated in this step.
Delocalized internal coordinates and corresponding gra-
␦r = Brs␦s + Brh␦h 共26兲
dients are generated from the modified matrix B̄, as de-
and using the matrix elements given by Eqs. 共22兲 and 共23兲, scribed in previous sections. The delocalized coordinates are
Downloaded 19 Jul 2006 to 194.214.217.17. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
124508-5 Geometry optimization of periodic systems J. Chem. Phys. 122, 124508 共2005兲
now linear combinations of either constrained, or active co- Farkas and Schleger33 suggested an improved DIIS algo-
ordinates, but not of both. Finally, the gradients for delocal- rithm, which allows one to adjust automatically the dimen-
ized coordinates corresponding to constrained coordinates sion of the iterative subspace. The idea is that the ionic step
are set to zero. The optimization then proceeds as in the case produced by DIIS ⌬q̃ = q̃k+1 − q̃k is compared to a simple
of an unconstrained relaxation. Not only single primitive in- quasi-Newton step ⌬q̃QN = H̃−1˜ and should meet the follow-
ternal coordinates but also the ratios and sums of primitive ing criteria.
internal coordinates, the norms of vectors whose components
are primitive internal coordinates 共e.g., vibrational modes兲, 共i兲 The direction of the DIIS step ⌬q̃ deviates from ⌬q̃QN
lattice parameters, the volume of the unit cell and many other by an angle . The step is accepted if cos共兲 is larger than
coordinate types can be constrained. Note that Cartesian co- 0.97, 0.84, 0.71, 0.67, 0.62, 0.56, 0.49, and 0.41 for two to
ordinates can be taken as special 共single-body兲 primitive in- nine recent relaxation steps used in DIIS. For a dimension of
ternal coordinates. ten or higher this criterion is not taken into account.
Special attention should be paid to the constrained opti- 共ii兲 The norm of DIIS step is limited to be not more than
mizations where the lattice parameters are allowed to ten times larger than that of the reference step.
change, but some of the internal degrees of freedom are fro- 共iii兲 The sum of all positive coefficients ci should not
zen. The components of the Cartesian forces that correspond exceed the value of 15.
to the fixed internal coordinates should be subtracted from 共iv兲 The magnitude of c / 兩r兩2 should not exceed 108.
this expression of the stress tensor, in order to avoid their
If one of these criteria is not fulfilled, the step is not
“contamination.” Our B-matrix technique, if used consis-
accepted and the most remote vector is removed from the
tently, takes care of these effects.
iterative subspace. This procedure is repeated until all crite-
ria are fulfilled.
F. Optimization strategy
One of the major advantages of the use of internal in-
The minimization algorithm used in this study is based stead of Cartesian coordinates is that a reasonable guess for
on the geometrical direct inversion in the iterative subspace Hessian matrix can be constructed. Even a very simple
共DIIS兲 method by Császár and Pulay32 improved recently by model Hessian which is just a diagonal matrix with elements
Farkas and Schlegel.33 In this method the information on the 0.5, 0.2, and 0.1 a.u. for bonds, angles, and torsions, respec-
potential energy surface collected in the preceding M relax- tively, usually works very well. Lindh et al.34 proposed a
ation steps is used to minimize the norm of the error vector model Hessian, constructed from force constants that are
defined as a linear combination of gradients ˜ corresponding simple functions of nuclear positions
to delocalized internal coordinates q̃,
kij = krij , 共34兲
k
rk = 兺
i=k−M
ci˜i . 共31兲 kijk = kij jk , 共35兲
The coefficients ci are obtained by minimizing 兩r兩2 under kijkl = kij jkkl , 共36兲
the constraint 兺i=k−M
k
ci = 1 leading to a set of equations
冢 冣冢 冣 冢 冣
with
˜k−M
t
˜k−M . . . ˜k−M
t
˜k 1 ck−M 0
. ... . . . . ij = exp关␣ij共r20,ij − r2ij兲兴. 共37兲
. ... . . . . For the first three rows of the periodic table one needs alto-
= . gether 15 independent parameters for the quantities kr, k, k,
. ... . . . .
␣ij, and r0,ij that are collected in Table I. The model Hessian
˜tk˜k−M ... ˜tk˜k 1 ck 0
can be easily transformed to delocalized internal coordinates
1 ... 1 0 1 using the formula
共32兲
H̃ = UtHU. 共38兲
A new set of delocalized internal coordinates is calculated
using In the course of the relaxation, the Hessian matrix is
updated using the Broyden–Fletcher–Goldfarb–Shanno algo-
M M
rithm:
兺 兺
冉 冊
q̃k+1 = ciq̃i + H̃ −1
c i
˜ i. 共33兲
i=k−M i=k−M
⌬˜k⌬˜tk H̃k−1⌬q̃k−1⌬q̃k−1
t
H̃k−1
H̃k = H̃k−1 − + , 共39兲
The dimension M of the iterative subspace used in DIIS ⌬˜k−1⌬q̃k
t
⌬q̃k−1
t
H̃k−1⌬q̃k−1
has substantial impact on efficiency of the relaxation. When
the structure is close to the minimum 共i.e., in the harmonic where ⌬˜k = ˜k − ˜k−1 is the change of gradients associated
region兲 M should be relatively high 共ten or more兲. When with the the relaxation step ⌬q̃k = q̃k − q̃k−1.
starting from a poor guess and the landscape of the potential The performance of our optimization engine GADGET is
energy is not well described by a harmonic approximation, a checked against the “native” Vienna ab initio simulation
better performance can be achieved by using only a limited package 共VASP兲 optimizer using the conjugate gradient al-
number of previous relaxation steps. gorithm and performed in cartesian coordinates. The conju-
Downloaded 19 Jul 2006 to 194.214.217.17. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
124508-6 Bučko, Hafner, and Ángyán J. Chem. Phys. 122, 124508 共2005兲
TABLE I. Parameters of the model Hessian proposed by Lindh et al. 共Ref. 34兲. Indices i and j designate the
rows of the periodic table of elements to which the atoms correspond. The universal stretch, bend, and torsion
coefficients are kr = 0.45, k = 0.15, and k = 0.005. All quantities are given in atomic units.
gate gradient algorithm tries to improve a simple steepest the plane wave basis-set incompleteness error.37 The applied
descent step 共i.e., in the direction of gradient兲 by conjugating cutoff energy of 800 eV is pretty high and it is usually as-
the search direction from the most recent step. It starts with sumed that in this case the Pulay stress is negligible. Some-
the steepest descent direction, but in all the following steps what unexpectedly, we have found relatively small but nev-
the search direction 共sk+1兲 is given by ertheless significant residual stresses at the equilibrium
volume obtained by fitting a Murnaghan equation of state to
fk + fk−1 · fk
␥= , 共40兲 a series of fixed-volume relaxed structures. This residual
fk−1 · fk−1 stress 共Pulay stress兲 can be added as a constant in the auto-
matic relaxation procedure, in order to correct the finite-basis
sk+1 = fk + ␥sk . 共41兲 error in the direct lattice optimizations.
The conjugate gradient algorithm requires the line mini- Different kinds of optimizations were performed, all
mization along search direction. starting from the same initial geometry. 共i兲 Only the atomic
positions are optimized at constant lattice parameters, using
delocalized internal coordinates 共GADGET兲. 共ii兲 Atomic posi-
III. APPLICATIONS tion relaxation with the native VASP optimizer. 共iii兲 Simul-
taneous atomic parameter and lattice parameter optimization
The electronic structure calculations were done with the
in delocalized internal coordinates with GADGET. 共iv兲 Full
VASP,1 using the density functional theory in the generalized
atomic position and lattice parameter relaxation with VASP.
gradient approximation. The PW91 exchange-correlation
共v兲 Multistep geometry relaxation with GADGET using a se-
functional was used, using the projector-augmented wave
ries of fixed-volume relaxations to determine a Murnaghan-
formalism.35,36 The calculations were done with high preci-
type equation of state 共EOS兲 and reoptimizing the structure
sion, i.e., the wave function was developed on a plane wave
at the volume of the interpolated minimum. 共vi兲 Full delo-
basis with a cutoff of 800 eV, and the support grid for the
calized internal coordinate relaxation using the Pulay stress
representation of the charge density was sufficiently precise
deduced from the EOS optimization.
to avoid any wraparound errors. The electronic wave func-
tions were converged in each step to 2.7⫻ 10−8 hartree. It is
expected that these computational parameters allow us to A. Microporous material: SiO2 chabazite
obtain quite reliable forces. Chabazite 共Fig. 1兲 is microporous aluminosilicate min-
The geometry optimizations were done with the external eral 共zeolite兲 in which every Si 共Al兲 atom is coordinated by
optimizer GADGET, written in Python. GADGET reads the ge- four O atoms. The space groups of the purely siliceous cha-
ometry, energy, and gradients from the VASP output, sets up
internal coordinates, estimates an optimal move, calculates
the new set of lattice parameters and Cartesian coordinates,
and starts a new VASP calculation, until convergence. As we
shall see, the additional overhead related to the repeated re-
starts of VASP is largely compensated by the gain in the
number of iterations. Our optimizer can be easily interfaced
with other packages that calculate total energies and forces/
stresses, and such an interface is already operational with the
5
GAUSSIAN 03 package. Optimizations were carried out using
the geometrical DIIS method with an iterative subspace of
dimension 5. Convergence criteria involve the simultaneous
fulfillment of an energy change less than 1 ⫻ 10−6 a.u., a
maximal gradient less than 2 ⫻ 10−4 a.u., and a volume
change 共in lattice relaxations兲 that is smaller than 0.05%.
In lattice parameter optimizations, involving volume
changes, we are faced with the problem of “Pulay stress,” i.e. FIG. 1. Rhombohedral unit cell of purely siliceous chabazite, Si12O24.
Downloaded 19 Jul 2006 to 194.214.217.17. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
124508-7 Geometry optimization of periodic systems J. Chem. Phys. 122, 124508 共2005兲
TABLE II. Optimization of the chabazite structure: the initial configuration 共Initial兲; relaxation of the ionic degrees of freedom using delocalized internal
coordinates 共兵q其兲 and using the VASP native optimizer 共兵s其兲; relaxation of the lattice parameters and ionic positions using delocalized internal coordinates
共兵q , q̂其兲 and with the VASP native optimizer 共兵s , h其兲; minimum of the fitted Murnaghan equation of state 共EOS兲; and the full relaxation using delocalized
internal coordinates with stress corrected for a Pulay stress 共兵q , q̂ ; 其兲. Listed are cell parameters 关a, b, c 共Å兲; ␣, , ␥ 共deg.兲兴, energy 共hartree兲, gradient 共a.u.兲,
stress 共kB兲, cell volume 共Å3兲, and the number of relaxation steps.
bazite is R3̄m 共No. 166兲, however, we have chosen a slightly retical vibrational spectrum of gibbsite and bayerite in light
perturbed initial configuration, corresponding to a highly sili- of some new experimental results.42
ceous protonated framework.38 The primitive unit cell con- In this system the conventional primitive internal coor-
tains 12 Si 共Al兲 atoms and 24 O atoms giving a total of 288 dinates 共bonds, angles, and torsions兲 are insufficient to de-
primitive internal coordinates 共bonds, angles, and torsions兲. scribe all atomic degrees of freedom. In order to describe
The results obtained after various types of relaxations are correctly the mutual position of subsequent layers we have
summarized in Table II. In every case our algorithm outper- used inverse power distance coordinates 共5 / r coordinates兲
formed the native VASP optimizer by a factor of 2–3. The proposed by Baker and Pulay.43 The 5 / r coordinates were
minimum of the full relaxation is checked against a fit to the generated only for pairs of atoms lying in different structural
Murnaghan equation. In spite of the high plane wave cutoff, fragments 共layers兲 and satisfying the condition that the cor-
the effect of the Pulay stress is not negligible: the residual responding interatomic distance is smaller than 1.6 times the
stress was found to be around −0.6 kB. The cell volume for sum of the covalent radii. In this way we have generated 664
the resulting configuration is by ⬃1 Å3 too low compared to conventional primitive and 179 inverse-power distance coor-
the true minimum. dinates. Data for the initial configuration and for the relaxed
The underestimation of the equilibrium cell volume due structures are collected in Table III. The performance of our
to the Pulay stress is a quite general phenomenon observed method is by a factor of 10 better than the conjugate gradient
in all examples given here. The full relaxation with compen-
sation for a Pulay stress leads to the structure which is very
close to a minimum of the Murnaghan equation. The sym-
metry of the unit cell is not completely recovered 共a ⫽ b
⫽ c兲 in full relaxations, due to the fact that the lattice param-
eters are usually “softer” than the atomic degrees of freedom
and thus more difficult to relax. In order to relax to a true
symmetry, a more stringent relaxation criterion for the stress
tensor would be required.
Downloaded 19 Jul 2006 to 194.214.217.17. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
124508-8 Bučko, Hafner, and Ángyán J. Chem. Phys. 122, 124508 共2005兲
TABLE III. Optimization of the gibbsite structure: the initial configuration 共Initial兲; relaxation of the ionic degrees of freedom using delocalized internal
coordinates 共兵q其兲 and using the VASP native optimizer 共兵s其兲; relaxation of the lattice parameters and ionic positions using delocalized internal coordinates
共兵q , q̂其兲 and with the VASP native optimizer 共兵s , h其兲; minimum of the fitted Murnaghan equation of state 共EOS兲; and the full relaxation using delocalized
internal coordinates with stress corrected for a Pulay stress 共兵q , q̂ ; 其兲. Listed are cell parameters 关a, b, c 共Å兲; ␣, , ␥ 共deg.兲兴, energy 共hartree兲, gradient 共a.u.兲,
stress 共kB兲, cell volume 共Å3兲, and the number of relaxation steps.
algorithm implemented in VASP. The full relaxation with the tive and 160 inverse-power distance coordinates have been
compensation for a Pulay stress of 兵−0.64, −1.81, −0.50其 used to generate the delocalized internal coordinates. The
leads to the configuration close to the minimum obtained data concerning this example are collected in Table IV. The
from the Murnaghan equation of state. initial configuration has already the correct space group sym-
metry 共P4̄21m兲. As pointed out by Baker et al.,25 delocalized
C. Molecular crystal: Urea internal coordinates are automatically symmetry adapted.
Therefore the symmetry is preserved if the relaxation is
Urea 关共NH2兲2CO兴 共Fig. 3兲 is an example of a molecular started with correct symmetry. As in the previous examples,
crystal 共space group P4̄21m, No. 113兲, with strong electro- Pulay stress causes the underestimation of the equilibrium
static and hydrogen bond interactions. In the crystal, the mol- cell volume. The structure relaxed with stress compensated
ecules occupy a special position compatible with the full for this effect is very close to the minimum of the Mur-
molecular symmetry 共mm2兲. High-resolution x-ray44 and naghan fit. The performance of our algorithm is by a factor
neutron45,46 diffraction studies show small differences in the of 5 better than the native optimizer in VASP.
cell parameters, but some significant changes in the atomic
positions. D. Constrained relaxation: Urea
The conventional primitive internal coordinates are gen-
erated only for intramolecular degrees of freedom, the mu- To illustrate the ability of our algorithm to impose virtu-
tual positions of individual urea molecules are described us- ally any geometrical constraint on the structure, we have
ing the 5 / r coordinates generated as described in the chosen example of urea with fixed intramolecular angles and
preceding section. Total number of 152 conventional primi- torsions. The relaxations have been started from the same
configuration as in the previous example. The relaxation cri-
teria are the same as in the unconstrained relaxations. Note,
however, that the contributions to the cartesian gradients cor-
responding to the constrained coordinates had to be projected
out. Results of relaxations of atomic positions, atomic posi-
tions, and lattice parameters, data for the minimum of the
Murnaghan equation of state and for the relaxation with the
correction for the Pulay stress 共estimated in the uncon-
strained relaxations兲 are shown in Table V.
IV. CONCLUSIONS
We have shown that the geometry optimization of peri-
odic systems in internal coordinates offers a highly advanta-
FIG. 3. Tetragonal unit cell of urea, 共NH2兲2CO. geous alternative to perform structural relaxations of solids
Downloaded 19 Jul 2006 to 194.214.217.17. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
124508-9 Geometry optimization of periodic systems J. Chem. Phys. 122, 124508 共2005兲
TABLE IV. Optimization of the urea structure: the initial configuration 共Initial兲; relaxation of the ionic degrees of freedom using delocalized internal
coordinates 共兵q其兲 and using the VASP native optimizer 共兵s其兲; relaxation of the lattice parameters and ionic positions using delocalized internal coordinates
共兵q , q̂其兲 and with the VASP native optimizer 共兵s , h其兲; minimum of the fitted Murnaghan equation of state 共EOS兲; and the full relaxation using delocalized
internal coordinates with stress corrected for a Pulay stress 共兵q , q̂ ; 其兲. Listed are the cell parameters 关a, b, c 共Å兲; ␣, , ␥ 共deg.兲兴, energy 共hartree兲, gradient
共a.u.兲, stress 共kB兲, cell volume 共Å3兲, and the number of relaxation steps.
and surfaces. It has been demonstrated that the present opens a broad field of application of periodic electronic
implementation is able to handle not only atomic position but structure codes to complicated problems, such as phase tran-
also lattice parameter relaxations. The final geometries ob- sitions, solid state chemical reactions, etc.
tained by the native optimizer of VASP and the internal co- The internal coordinate optimizer GADGET is an indepen-
ordinate optimizer GADGET are the same within the numeri- dent software, easy to interface with various molecular and
cal precision of the electronic code. For the selected model solid state total energy codes. It allows the user to optimize
systems GADGET outperforms the native optimizer by a fac- not only periodic but also finite 共molecular兲 systems in delo-
tor of 2–10. One has to be aware of the fact that a consider- calized internal coordinates by ignoring the lattice parameter
able portion of this better performance can be ascribed to the block of the extended B-matrix, i.e., by the standard method
quality of the initial Hessian guess and to the use of more widely documented in the literature 共see, e.g., Ref. 12兲. In
efficient optimization algorithms. addition to the innovative treatment of periodic systems, the
The most significant advantage of the internal coordinate GADGET program includes a large selection of algorithms
optimization lies undeniably in the handling of constraints, permitting the efficient handling of Hessian update, transi-
as demonstrated by the simple examples of rigid molecule tion state optimization, etc. An exhaustive overview of these
optimization of the urea molecular crystal. In any case, features will be the subject of a forthcoming paper that de-
the possibility of constraining virtually any physically/ scribes the technical aspects of our implementation of these
chemically motivated combination of internal coordinates algorithms in GADGET.
TABLE V. Relaxation of urea with constrained intramolecular angles and torsions: the initial configuration 共Initial兲; relaxation of the ionic degrees of freedom
using delocalized internal coordinates 共兵q其兲; relaxation of the lattice parameters and ionic positions using delocalized internal coordinates 共兵q , q̂其兲; minimum
of the fitted Murnaghan equation of state 共EOS兲; and the relaxation of the lattice parameters and ionic positions using delocalized internal coordinates with
stress corrected for a Pulay stress 共兵q , q̂ ; 其兲. Listed are cell parameters 关a, b, c 共Å兲; ␣, , ␥ 共deg.兲兴, energy 共hartree兲, gradient 共a.u.兲, stress 共kB兲, cell volume
共Å3兲, and the number of relaxation steps.
Downloaded 19 Jul 2006 to 194.214.217.17. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
124508-10 Bučko, Hafner, and Ángyán J. Chem. Phys. 122, 124508 共2005兲
3
ACKNOWLEDGMENT M. D. Segall, P. L. D. Lindan, M. J. Probert, C. J. Pickard, P. J. Hasnip, S.
J. Clarke, and M. C. Payne, J. Phys.: Condens. Matter 14, 2717 共2002兲.
One of the authors 共J.G.A.兲 is indebted to the Computa- 4
J. M. Soler, E. Artacho, J. D. Gale, A. García, J. Junquera, P. Ordejón, and
tional Materials Science College for having supported his D. Sánchez-Portal, J. Phys.: Condens. Matter 14, 2745 共2002兲.
5
M. J. Frisch, G. W. Trucks, H. B. Schlegel et al., GAUSSIAN 03, 2004.
stay in Vienna. 6
V. R. Saunders, R. Dovesi, C. Roetti, M. Causà, N. M. Harrison, R.
Orlando, and C. N. Zicovitch-Wilson, CRYSTAL 98 User’s Manual 共Univer-
APPENDIX: THE STRESS TENSOR sity of Torino, Torino, Italy, 1998兲.
7
B. Civalleri, P. D’Arco, R. Orlando, V. R. Saunders, and R. Dovesi, Chem.
During a deformation of the lattice, all atomic positions Phys. Lett. 348, 131 共2001兲.
and virtual grid points, including the lattice vector endpoints, 8
J. D. Gale and A. L. Rohl, Mol. Simul. 29, 291 共2003兲.
change according to the relationship
9
G. G. Ferenczy and J. G. Ángyán, J. Comput. Chem. 22, 1679 共2001兲.
10
G. Fogarasi, X. Zhou, P. W. Taylor, and P. Pulay, J. Am. Chem. Soc. 114,
v␣⬘ = 兺 共␦␣ + ␣兲v , 共A1兲 11
8191 共1992兲.
J. Baker, J. Comput. Chem. 14, 1085 共1993兲.
12
V. Bakken and T. Helgaker, J. Chem. Phys. 117, 9160 共2002兲.
where ␣ are the elements of the strain tensor. The stress 13
M. V. Fernández-Serra, E. Artacho, and J. M. Soler, Phys. Rev. B 67,
tensor elements ␣ are usually defined as the volume- 100101 共2003兲.
14
K. Kudin, G. Scuseria, and H. B. Schlegel, J. Chem. Phys. 114, 2919
normalized negative strain derivatives of the energy: 共2001兲.
15
1 E J. Andzelm, R. D. King-Smith, and G. Fitzgerald, Chem. Phys. Lett. 335,
␣ = − . 共A2兲 321 共2001兲.
⍀ ␣ 16
E. B. J. Wilson, J. C. Decius, and P. C. Cross, Molecular Vibrations. The
Theory of Infrared and Raman Vibrational Spectra 共Dover, New York,
The Cartesian to fractional transformation matrix h␣ is 1955兲.
formed from the three column vectors a of the lattice as 17
S. Nosé and M. L. Klein, Mol. Phys. 50, 1055 共1983兲.
h␣ = 共a兲␣. Using Eq. 共A1兲 it follows that the variation of the
18
I. Souza and J. L. Martins, Phys. Rev. B 55, 8733 共1997兲.
19
B. Paizs, G. Fogarasi, and P. Pulay, J. Chem. Phys. 109, 6571 共1998兲.
lattice vectors can be expressed as 20
Ö. Farkas and H. B. Schlegel, J. Chem. Phys. 109, 7100 共1998兲.
兺␥ 共␦␣␥ + ␣␥兲h␥
21
S. R. Billeter, A. J. Turner, and W. Thiel, Phys. Chem. Chem. Phys. 2,
⬘ =
h␣ 共A3兲 2177 共2000兲.
22
B. Paizs, J. Baker, S. Suhai, and P. Pulay, J. Chem. Phys. 113, 6566
and the strain derivative of the lattice vectors is 共2000兲.
23
K. Németh, O. Coulaud, G. Monard, and J. G. Ángyán, J. Chem. Phys.
h␣ 113, 5598 共2000兲.
= ␦␣h . 共A4兲 24
J. Baker, A. Kessi, and B. Delley, J. Chem. Phys. 105, 192 共1996兲.
25
J. Baker, D. Kinghorn, and P. Pulay, J. Chem. Phys. 110, 4986 共1999兲.
26
K. N. Kudin and G. E. Scuseria, Phys. Rev. B 61, 5141 共2000兲.
The strain derivative of the energy can be written in terms of 27
G. J. Martyna, D. J. Tobias, and M. L. Klein, J. Chem. Phys. 101, 4177
the lattice vector component derivatives, using Eq. 共A4兲, as 共1994兲.
28
P. H. Hünenberger, J. Chem. Phys. 116, 6880 共2002兲.
E E h E
=兺 =兺 R. G. Winkler, J. Chem. Phys. 117, 2449 共2002兲.
29
h  共A5兲
␣ h ␣ h␣ D.-H. Lu, M. Zhao, and D. G. Truhlar, J. Comput. Chem. 12, 376 共1991兲.
30
31
J. Baker, J. Comput. Chem. 18, 1079 共1997兲.
leading to the alternative form of the stress tensor
32
P. Császár and P. Pulay, J. Mol. Struct.: THEOCHEM 114, 31 共1984兲.
33
Ö. Farkas and H. B. Schlegel, Phys. Chem. Chem. Phys. 4, 11 共2002兲.
E
34
1 R. Lindh, A. Bernhardsson, G. Karlström and P.-Å. Malmqvist, Chem.
␣ = −
⍀
兺␥ h␣␥ h␥
t
. 共A6兲 Phys. Lett. 241, 423 共1995兲.
35
P. E. Blöchl, Phys. Rev. B 50, 17953 共1994兲.
36
G. Kresse and D. Joubert, Phys. Rev. B 59, 1758 共1999兲.
The lattice-vector derivatives of the energy can be obtained 37
G. P. Francis and M. C. Payne, J. Phys.: Condens. Matter 2, 4395 共1990兲.
38
from the stress tensor, since L. J. Smith, A. Davidson, and A. K. Cheetham, Catal. Lett. 49, 143
共1997兲.
E 39
H. Saalfeld and M. Wedde, Z. Kristallogr. 139, 129 共1974兲.
−⍀ 兺 ␣共ht兲⑀−1 = 兺
␥ h␣␥
t
h␥ 共ht兲⑀
−1
, 共A7兲 40
J. D. Gale, A. L. Rohl, V. Milman, and M. C. Warren, J. Phys. Chem. B
105, 10236 共2001兲.
41
M. Digne, P. Sautet, P. Raybaud, H. Toulhoat, and E. Artacho, J. Phys.
so Chem. B 106, 5155 共2002兲.
E
42
J. G. Ángyán, H. Drider, B. Humbert, and T. Bučko 共in preparation兲.
= − ⍀ 兺 ␣␥共ht兲␥
−1
. 共A8兲 43
J. Baker and P. Pulay, J. Chem. Phys. 105, 11100 共1996兲.
h␣ ␥
44
S. Swaminathan, B. M. Craven, M. A. Spackman, and R. F. Stewart, Acta
Crystallogr., Sect. B: Struct. Sci. 40, 398 共1984兲.
45
S. Swaminathan, B. M. Craven, and R. K. McMullan, Acta Crystallogr.,
1
G. Kresse and J. Furthmüller, Comput. Mater. Sci. 6, 15 共1996兲. Sect. B: Struct. Sci. 40, 300 共1984兲.
2 46
K.-H. Schwarz, P. Blaha, and G. K. H. Madsen, Comput. Phys. Commun. V. Zavodnik, A. Stash, V. Tsirelson, R. D. Vries, and D. Feil, Acta Crys-
147, 71 共2002兲. tallogr., Sect. B: Struct. Sci. 55, 45 共1999兲.
Downloaded 19 Jul 2006 to 194.214.217.17. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp