PHYSICAL REVIEW B VOLUME 46, NUMBER 18 1 NOVEMBER 1992-II

Vibrational densities of states from molecular-dynamics calculations

S. Gongalves and H. Bonadeo
Deyartamento de Fxsica, Comisi6n Nacional de Energca At6mica, Avenida del Libertador 8850,
A/99 Buenos Aires, Aryentina
(Received 11 October 1991; revised manuscript received 29 June 1992)

A method for calculating the vibrational densities of states of crystalline materials by use of
molecular dynamics (MD) is presented. It consists mainly of starting the calculation with all atoms
at rest in their equilibrium positions, except one which is given an initial velocity; the time evolution
of this velocity is Fourier transformed to give the local vibrational density of states, and the procedure
repeated for nonequivalent atoms in the lattice. It is shown with examples that the vibrational
densities of states are unbiased and show very low noise and that calculations with this method
require much fewer MD steps than those performed using standard random initializations.

Molecular dynamics (MD) is a classical, non-
perturbative method for studying structural, dynamical,
and thermodynamical properties of complex systems. Its
use has been increasingly widespread since the 1960s, and
it has been applied to liquids, gases, and ordered and dis-
ordered solids. ~ In the case of solids it has the distinct
advantage of incorporating anharmonic forces and struc-
tural relaxations naturally; for disordered systems it is
able to replace restrained and cumbersome Green's func-
tions approaches.
One of the important features of a solid system is its
vibrational density of states, on which optical and ther-
modynamical properties depend. It is well known that
in the harmonic approximation, the power spectrum of
the velocity autocorrelation is the vibrational density of
states, i.e. , that
where the sum is taken over all atoms in the system.
It is to be noted that, strictly speaking, the equiva-
lence between classic power-spectrum and quantum den-
sity of states is valid only in the harmonic approxima-
tion: the power spectrum of an anharmonic oscillator
will yield peaks at the symmetry-allowed multiples of the
fundamental frequency, where no quantum states exist.
The position of the fundamental peaks, however, will be
correctly given, and Eq. (1) has been universally used
for the calculation of anharmonic densities of states. In
practice the calculation is performed by giving random
velocities to the atoms of the system, and following its
evolution for what is judged a sufBciently long period;
usually several initializations are used, and the results
averaged. The numerical Fourier transform is evaluated
over a finite interval, and usually a Gaussian apodization,
producing a Gaussian broadening of the resulting spec-
trum, is used. The random initializations are supposed
to assure that, according to the definition, all vibrational
modes are equally excited. This, however, may well not
be the case, especially in weakly anharmonic systems,
where energy transfer between modes is slow, and initial
biases tend to perpetuate. As a result, noisy and biased
spectra are obtained even for long runs and apparently
good statistics. s Exaggerating the Gaussian broadenings
only results in embellishing the spectra at the expense of
blurring the existing information.
We have found an alternative way of initializing a sys-
tem that gives excellent results at a much reduced cost in
computing time: it consists of starting from the equilib-
rium positions for all the atoms in the MD cell, and giving
a nonzero initial velocity to just one of the atoms. The
Fourier transform of the time evolution of the velocity of
this atom gives the local density of states at the atom;
the procedure is repeated for all nonequivalent atoms in
the crystal. It is quite evident that in this way, we ob-
tain a numerical Green's function for the system, since
we are measuring the system's response to a b-function
input; although the idea is very simple indeed, we have
not seen it mentioned yet in the literature.
The proof is straightforward, when the same harmonic
approximation used for Eq. (1) is implemented. The dis-
placements and velocities corresponding to the nth coor-
dinate of the ith atom in the mth cell are written as
(2)
where e' (kj) is the eigenvector for the normal coordinate
Qkj, N is the number of atoms and the M, 's are their
masses. With the initial conditions
u '(0)=0, Vm, i, n,
'(0) = 0,
v Vm g 0, i g i', o. P n',
v ',
(0) $0,
one obtains for the normal coordinates and velocities the
initial conditions

46 12 019 1992 The American Physical Society

12 020 BRIEF REPORTS 46

Q»(0) =0, particular form of exc't'

exci ing t e system,
(4) ( g v, (0)v, (t) &
o'
v l (0)e*', (kj), ( Q v, (0)v, (0) &

This, in the linear'near regime, is formall y

where * indicates th e complex
p onjugate; and since Q kg
con
e uctuation-dissi p atio
oscillates with f o in crested in followin

Q» (t) = Q» (0) cos(4)»t) .

Replacing Eqs. (4) and (5) in
'
Eq.. (2),
~ ~
ob
of tll
11 it o
bl

1
t,bt'd
i
W
d p 1'h
is this, we must
'
U
t rti fo t}1e equi'1'i rium positions
v l = l — e, (kj)e', (kj) cos(ui, ~t).
~~ a om in the unit cell a
d o , recor d t e time evolut'
dFo t s form it; we follow the
Taking initial velocities for all direc b i i h o rn a ve 1ocity in one an
t1M Mt a ization of the eigen-
11 an d using the normaliz
1 di
'
io ns; h e sum of the ourier transforms
vectors one obtains of the time evoluti u ions of these veloc't'
1 ld ' of tt es; fll
o owin g thes po
(7) a a oms in the cr stal
s a es. Of course the c
Fourier transform
whose F is the vi
b fo do 1y for d'irections and atom s th at are not
Fro ()
Eqs.. jl) and (7 , we have that for this

I
i
I \
\

I
I
I
I
I
I
I
I
I

I
I

'L- V3
O
A
1 ~
i
\

i i i I }JIi
1
FIG. 2. Local densities of vibr ational states (LDOS)
s impurity (0.75 of thee hoost mass) in an
' '
FIG. 1. Den
ensities of vibrational states ffcc monoatomic a ice oof 500 atoms, Lenn
lattice
li of 20 atoms; 2000 M
h i hb ons, wit'h aussian a od'
()S d d al cu 1a ion with 50 random
w'
'
'
n
() t th d (180 MD steps). Dashed li
oigi ad5i dp d t
ia excitation. Solid lines:
D hdl i 'h" '
gust resolution" n Gauss'
aussian apodization. '
ota MD ste ps, s 20 time origins and
r 1t for an infinite li
c resu
otted lines: exact 25 independentn initializations (90 000 total
'
quencies normali ed too th e cutofF frequency of the system quencies normalizede too tthee cutofF
cu frequency of the host lattice 4
46 BRIEF REPORTS
equivalent by symmetry.
It is to be noted that any other initialization will ini-
tially give biased results: even in the linear case, if more
than one atom is initially excited, each will excite all nor-
mal modes, but the phases of these will be difFerent, and
therefore they will interfere. An extreme case is when a
system is excited in just one normal mode, whose evo-
lution may then be followed, as is usually done in non-
equilibrium MD studies.
In order to illustrate the efficiency of the method, we
show in Fig. 1 the results of a calculation on a harmonic
linear chain of 20 identical atoms in the MD cell, com-
pared to the standard method, where 50 random initial-
izations were used; to obtain high resolution, 2000 MD
steps were used for each run. The figure shows the results
obtained by Fourier transforming p(t) without apodiza-
tion, and using a Gaussian apodization with a width that
satisfies the criterion of just resolution, i.e. , two adjacent
lines of equal intensity give rise to two bands without a
dip in between; these results are compared to the exact
ones. This system is intrinsically linear, and no thermal-
ization will correct initial biases, which are striking in the
latter case.
For imperfect crystals, the sum in Eq. (7) should be ex-
'Molecular Dynamics Simulation of Statistical-Mechanical
Systems, Proceedings of the International School of Physics,
course XCVII, edited by G. Ciccotti and G. Hoover (North-
Holland, Amsterdam, 1986).
J. M. Dickey and A. Paskin, Phys. Rev. 188, 1407 (1969).
12 021
tended, in principle, to all atoms in the MD cell; in prac-
tice, however, when, for instance, an impurity is present,
only a few atoms around it show a local vibrational den-
sity of states significantly difFerent from the bulk. 4 In
Fig. 2 we present the results for a mass defect in an
fcc monoatomic lattice, with 500 atoms in the MD cell
and Lennard- Jones Brst-neighbor interactions. A total of
4500 [Fig. 2(a)] and 90000 [Fig. 2(b)] MD steps have been
used in the standard calculation, and 180 in the present
scheme. The difference both in quality of the results and
time consumed is very relevant.
In conclusion, we think that the simple initialization
method presented here is a significant improvement over
the standard one for MD calculations of vibrational den-
sities of states of crystals. It will probably be most useful
in calculations on relatively simple systems with compli-
cated force laws, but in any case the fundamental ac-
complishment is to have complete assurance of obtaining
unbiased spectra.
One of us (S.G.) was supported by the Consejo Na-
cional de Investigaciones Cientificas y Tecnicas of Ar-
gentina.
3 J. Vermesse, D. Levesque, J. J. Weis, and M. L. Klein, Can.
J. Phys. 61, 1545 (1983).
S. Gongalves and H. Bonadeo, Phys. Rev. B 46, 10738
(1992).