PHYSICAL REVIEW LETTERS

VOLUME 55, NUMBER 22

Unified Approach for Molecular Dynamics and Density-Functional

25 NOVEMBER 1985

Theory

R. Car
International

School for Advanced Studies, Trieste, Italy

and

M. Parrinello
Dipartimento di Fisica Teorica, Universita di Trieste, Trieste, Italy, and
International School for Advanced Studies, Trieste, Italy
(Received 5 August 1985)

We present a unified scheme that, by combining molecular dynamics and density-functional

theory, profoundly extends the range of both concepts. Our approach extends molecular dynamics
beyond the usual pair-potential approximation, thereby making possible the simulation of both covalently bonded and metallic systems. In addition it permits the application of density-functional
theory to much larger systems than previously feasible. The new technique is demonstrated by the
calculation of some static and dynamic properties of crystalline silicon within a self-consistent pseudopotential framework.
PACS numbers:

71.10. +x, 65.50. +m, 71.45.Gm

very large and/or disordered systems and to the computation of interatomic forces for MD simulations.
We wish to present here a new method that is able
to overcome the above difficulties and to achieve the
following results: (i) compute ground-state electronic
properties of large andlor disordered systems at the
level of state-of-the-art electronic structure calculations; (ii) perform ah initio MD simulations where the
only assumptions are the validity of classical mechanthe Bornionic motion
and
ics to describe
Oppenheimer (BO) approximation to separate nuclear
and electronic coordinates.
Following Kohn and Sham3 (KS) we write the electron density in terms of occupied single-particle orthonormal orbitals: n(r) = X,. ~tlt;(r) ~2. A point of the
BO potential energy surface is given by the minimum
with respect to the Q;(r) of the energy functional,

Electronic structure calculations based on densitycomfunctional (DF) theory' and finite-temperature

puter simulations based on molecular dynamics (MD)
have greatly contributed
to our understanding
of
condensed-matter
systems. MD calculations are able
to predict equilibrium and nonequilibrium
properties
of condensed systems. However, in all practical applications MD calculations have used empirical interatomic potentials. This approach, while appropriate for
systems like the rare gases, may fail for covalent
andi or metallic systems. Furthermore, these calculations convey no information about electronic properties. On the other hand, DF calculations have provided an accurate, albeit approximate, description of the
chemical bond in a large variety of systems, ' but are
computationally very demanding. This has so far precluded the application of DF schemes to the study of
I

E[(p, }, (Rt), (ct}]= X,.

d3r

tlt

(r) [ (h2/2m)Vz]tlt;(r) + U[n(r), {Rt), (o, }].

Here (Ri) indicate the nuclear coordinates and (n} are

all the possible external constraints imposed on the
system, like the volume fl, the strain e, etc. The
functional U contains the internuclear Coulomb repulsion and the effective electronic potential energy, ineluding external nuclear, Hartree, and exchange and
correlation contributions.
In the conventional formulation, minimization of
the energy functional [Eq. (1)] with respect to the orbitals p;, subject to the orthonormality
constraint,
leads to the self-consistent KS equations, i.e. ,
2m

ri'+,
6n(r),

y, (r) = e;y;(r).

(2)

The solution of Eq. (2) involves repeated matrix diagonalizations with a computational effort rapidly grow-

of the problem. Since the whole procedure has to be repeat ed for any new atomic configuration, the theoretical prediction of the equilibrium
geometries, when these are not known from experiment, still remains an u nsolved problem in most cases.
We adopt a quite diff erent approach and regard the
minimization of the KS functional as a complex optimization problem which can be solved by applying the
concept of simulated an nealing, recently introduced by
Ktrkpatrtck, Gelatt, and Vecchi. 4 In this approach an
objective function O({p) ) is minimized relative to the
parameters
(p), by generation of a succession of
distribution
probability
(p) s with a Boltzman-type
~ exp( O((P) )/T) via a Monte Carlo procedure.
For T 0 the state of lowest O((p) ) is reached uning with the size

1985 The American Physical Society

PHYSICAL REVIEW LETTERS

VOLUME 55, NUMBER 22

less the system is trapped into some metastable state.

In our case the objective function is the total-energy
functional and the variational parameters are the coefficients of the expansion of the KS orbitals in some
convenient basis and possibly the ionic positions
and/or the (n}'s. We found that a simulated annealing strategy based on MD, rather than on the Metropolis Monte Carlo method of Kirkpatrick, Gelatt, and

= X,. 2 p,

d r I+;

t) =5,,

(4

pP;(r, t) = oE/o$, '(r, t) + XkA, P(r,t),

IRI =
~

PRIE,
,

(5a)
(Sb)

pn= (8 E/Bn),

(5c)

where A;k are Lagrange multipliers introduced in order

to satisfy the constraints in Eq. (4). The ion dynamics
in Eqs. (5) may have a real physical meaning, whereas
the dynamics associated with the (p;)'s and the {n}'s
is fictitious and has to be considered only as a tool to
perform the dynamical simulated annealing. Equation
(3) defines a potential energy E and a classical kinetic
energy L given by

K= X,. ,' p, JI

d r

'
IP;I'+ Xt ,'MtRt + X,
p, n.

The equilibrium value (K) of the classical kinetic energy can be calculated as the temporal average over the
trajectories generated by the equations of motion [Eqs.
(5)] and related to the temperature of the system by
suitable normalization.
By variation of the velocities,
i.e. , the {P;}'s, {Rt}'s, and (n)'s, the temperature of
0 the
the system can be slowly reduced and for T
equilibrium state of minimal E is reached. At equilibrium Q;=0, Eq. (Sa) is identical within a unitary
transformation to the KS equation [Eq. (2)], and the
eigenvalues of the A matrix coincide with the occupied
KS eigenvalues. Only when these conditions are satisfied does the Lagrangean in Eq. (3) describe a real
physical system whose representative point in configurational space lies on the BO surface. For large systems our scheme is more efficient than standard diagoin the present
nalization techniques. 5 Furthermore,
self-consistency, ionic reapproach, diagonalization,
2472

"

(3)

&

In Eq. (3) the dot indicates time derivative, Mt are the

physical ionic masses, and p, and p, are arbitrary
parameters of appropriate units.
The Lagrangean in Eq. (3) generates a dynamics for
the parameters (p;) 's, (Rt) 's, and {n)'s through the
equations of motion:

Vecchi, can be applied efficiently to minimize the KS

functional.
This approach, which may be called
"dynamical simulated annealing,
not only is useful as
a minimization
procedure but, as we demonstrate
here, it allows also the study of finite temperature
properties.
In our method we consider the parameters (p, )
[Eq. (1)] to be
(Rt), (n) in the energy-functional
dependent on time and introduce the Lagrangean

+ Xt MtRt + X2 p, n
E [(+;), (Rt), (n}],

where the Q, are subject to the holonomic constraints

J{ d'r y, (r, t)yj(r,

25 NOVEMBER 1985

are
and strain relaxation
laxation, and volume
Th.e amount of classical
achieved simultaneously
kinetic energy is a measure of the departure of a system from the self-consistent minimum of its total energy.
It should be stressed that the dynamical simulated
annealing technique introduced above is a method of
quite general applicability in the context of functional
As such it can be useful in many areas
minimization.
of physics. For instance, it can be applied to the study
of classical field theories or to obtain the ground-state
energy in Hartree-Fock or configuration interaction
schemes. We also observe that, as far as functional
minimization is concerned, Newtonian dynamics may
be conveniently replaced by Langevin or other types
of dynamics.
In order to illustrate how our method works in practice, we present results obtained for the ground-state
electronic structure of Si as follows. We have considered a simple cubic supercell containing eight atoms
We have
subject to periodic boundary conditions.
used a local pseudopotentials
and a local-density approximation to the exact exchange and correlation
functional.
The single-particle orbitals for the valence
electrons have been expanded in plane waves with an
energy cutoff of 8 Ry, which amounts to including 437
plane waves at the I point. For simplicity, only the I
point of the Brillouin zone (BZ) of the supercell has
been considered in the evaluation of the energy functional. 'o This leads to a total of 16&& 437 complex electronic variational parameters,
since sixteen is the
number of doubly occupied KS levels. A simulated
annealing run is illustrated in Fig. 1. The lattice
parameter was allowed to vary while the ions were kept
in their perfect diamond arrangement.
The total enerthe
lattice
and
the
of the
parameter,
eigenvalues
gy,
matrix of the Lagrangean multipliers are plotted as
functions of the simulation "time.
The initial conditions for the electronic orbitals were fixed by filling
the lowest available plane-wave states and giving a
Maxwellian
distribution
of velocities to the components of the fields. The value of p, was chosen to be
1 a. u. The mass p, & associated with variation in the
volume was taken to be 10 5 a. u. The deerlet algo-

"

PHYSICAL REVIEW LETTERS

VOLUME 55, NUMBER 22

25 NOVEMBER 1985
I~

5x IO
-

~
I- 3.5
~
UJ ~
CO

lo.o IWcn
I
9.5 Z Z

0D
: 9.0 oo
-

UJ

LLJ

~
I- X
~ 4.0

0 I-

8.5

~0
f

sec

a).

-.E

0.025 r25'
0.000 --------. --0.025-

b)

4.I925-

-
Ep

4.I9350.03

0.02
0.0I-

04.

0.5

Izs

Q Z
cf W

er,

25

0.5

and I multiplets for two different MD runs.

atom, and I
The lattice constant was taken to be equal to the experimental value of 10.26 a. u. ; At and p, were taken to be 10 and 300
a. u. , respectively. The dashed line in the second panel from
the top indicates the T=O ground-state energy. The triangles indicate a doubly degenerate level.
~

IA
C4

l00

tO

lA

O
O

200

00
0
300

FIG. 1. Evolution of total energy per atom, lattice constant, and eigenvalues of the A matrix, during a typical
dynamical annealing run. The partial averages of the classical kinetic energy E during each subsection of the run are
0 the
indicated in the lower part of the picture. For E
eigenvalues of the A matrix tend to the KS eigenvalues.
The various multiplets are labeled according to the symmetry of the diamond lattice.

rithm" with a time step of 0. 1 a. u. was used and the

values of the Lagrange multipliers were determined by
the method of Ryckaert, Ciccotti, and Berendsen. '
After some initial equilibration the temperature was
progressively reduced to very small values. A satisfactory degree of convergence is seen to be achieved after
200 time steps, when our results agree within nuselfmerical errors with those of a conventional
consistent calculation for the same model. '3
We can consider now a situation in which the ions,
to which we associate their actual physical masses, are
while the
allowed to move at a given temperature,
kinetic energy of the electronic variational parameters
remains equal to zero. In this case the electrons are at
any time in their ground state and the ions move
under the action of BO forces. This can be achieved
either by conveniently reoptimizing the electronic variational parameters or by realizing a metastable situation in which the kinetic energy associated with the
s remains always very small compared to the typical
variations of the potential energy of the system. This
is equivalent to giving the BO surface a finite thickness
proportional to the temperature associated with the
s. If this temperature remains very small, the ion

400
&/h, t

FIG. 2. From top to bottom, temporal evolution of average atomic displacement along
potential energy per

~o~
UJ

Lai

200

400

200
t /4&

zo

1"I

-0.5

CO &g)
LU

dynamics generated via Eqs. (5) provides a good

representation of the actual dynamics of a physical system.
In Fig. 2 we report the results of two different sets
of calculations in which we have performed dynamical
simulations for our model. In these calculations it was
not necessary to reoptimize the electronic variational
parameters at each point along the trajectory, since the
thickness of the BO surface never exceeded the value
of 7X 10 6 a. u. per atom, a rather small fraction of the
potential energy variation. This is perfectly adequate
to represent the interionic forces in this particular case.
If the thickness of the BO surface were too large, not
only would the forces be incorrectly estimated but also
they might depend upon the path along which a given
point of the potential energy surface is approached.
On the other hand, for very small thicknesses the low
velocities of the electronic variational
parameters
might lead to intolerably long relaxation times. In
such a case a compromise would be necessary. In Fig.
2(a) the atoms were initially displaced from their perfect lattice position according to the eigenmode er
25

corresponding to the optical phonon mode at the I

point of the diamond lattice. The system undergoes
slightly anharmonic oscillations whose frequency is 20
THz, in perfect agreement with the results of a static
frozen-phonon
calculation for the same model, '3
showing that the thickness of the BO surface was adequate. In Fig. 2 (a) we also report how the effect of
the ionic oscillatory motion is reflected in some electronic properties. Notice that the threefold degenerate

2473

VOLUME 55, NUMBER 22

PHYSICAL REVIEW LETTERS

I,

level splits, in perfect phase with the ioninto a singlet and a doublet, whereas the
low-lying I t state is hardly affected by the ionic
motion. These results are contrasted with those reported in Fig. 2(b). Here the ions were first randomly
displaced from their equilibrium position and a simulated annealing was performed in order to bring the
electrons in the corresponding ground state. The ions
were then allowed to move. After some equilibration
the average kinetic energy associated with ionic motion
250 K and the
had a value corresponding
to
behavior of the system was as illustrated in Fig. 2(b).
The projection of the ionic displacement along the er

topmost
ic motion,

25

eigenvector and the electronic properties do not show

correlation. The degeneracy of the 12s,
one-electron eigenstate is completely lifted by thermal
disorder, while the I t state still remains hardly affected by the ionic motion.
The calculations presented here can all be performed
on a VAX-like minicomputer.
Access to supercomputers can make possible the simulation of larger systems and more realistic models. Because of the simplicity of Newton's equations the computer code can be
fully vectorized with not much effort. However the
main advantages of the present approach lie in its ability to perform a global minimization of the energy DF
and, more importantly, in offering a convenient and,
in principle, exact tool for studying finite temperature
effects and dynamical properties.
We benefited from discussions with A. Baldereschi,
P. Carnevali, A. Nobile, S. T. Pantelides, A. Selloni,
E. Tosatti, and A. R. Williams. Special thanks are due
to S. Baroni for precious suggestions and valuable
help. This work has been supported by the Gruppo
any apparent

25 NOVEMBER 1985

Nazionale di Struttura della Materia del Consiglio Nazionale delle Richerche and by the Ministero della
Pubblica Istruzione.

tSee, for instance, Theory of the Inhomogeneous Electron

Gas, edited by S. Lundqvist and N. M. March (Plenum, New
York, 1983).
zA. Rahman, in Correlation Functions and Quasiparticle In
teraetions in Condensed Matter, edited by J. Woods Halley,
NATO Advanced Study Series Vol. 35 (Plenum, New York,

1977).
3W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 (1965).
4S. Kirkpatrick, C. D. Gelatt, Jr. , and M. P. Vecchi, Science 220, 671 (1983).
5For instance, if the %occupied single-particle orbitals are
expanded into M plane waves (M )& N), a standard diagonalization requires O(M3) operations, whereas Eq. (5a) requires both O(NMln M) and O(N M) operations.
P. Carnevali and A. Selloni, private communication.
P. J. Rossky, J. D. Doll, and H. L. Friedman, J. Chem.
Phys. 69, 4628 (1978); C. Pangali, M. Rao, and B. J. Berne,
Chem. Phys. Lett. 55, 413 (1978).
J. A. Appelbaum and D. R. Hamann, Phys. Rev. B 8,

1777 (1973).
9J. P. Perdew and A. Zunger,

(1981).

Phys. Rev.

8 23, 5048

~
The sampling of the BZ and the pseudopotential used in
this paper are not very accurate and should be replaced with
better ones in more realistic calculations.
t tL. Verlet, Phys. Rev. 159, 98 (1967).
J. P. Ryckaert, G. Ciccotti, and H. J. C. Berendsen, J.
Comput. Phys. 23, 327 (1977).
&3When accurate pseudopotentials
are used together with
an accurate sampling of the BZ, local-density calculations
agree very well with experiment [see M. T. Yin and M. L.
Cohen, Phys. Rev. B 26, 3259 (1982)].
&