Encyclopedia of Nanotechnology

DOI 10.1007/978-94-007-6178-0_100946-1
# Springer Science+Business Media Dordrecht 2015

Car–Parrinello Molecular Dynamics

Mauro Boeroa* and Atsushi Oshiyamab
a
Institut de Physique et Chimie des Materiaux de Strasbourg (IPCMS), University of Strasbourg and CNRS, UMR 7504,
Strasbourg, France
b
Department of Applied Physics, The University of Tokyo, Tokyo, Japan

Synonyms
Ab initio molecular dynamics; CPMD; DFT-based molecular dynamics; First principles molecular
dynamics

Definition
The Car–Parrinello molecular dynamics (CPMD) is an extension of the Lagrangian formalism of classical
molecular dynamics in which the model potential describing the interaction among atoms is replaced by
the total energy functional of the system as provided by the Density Functional Theory (DFT). The
electronic wavefunctions are explicitly introduced as new dynamical variables. The simultaneous Euler-
Lagrange equations of motion for both sets of dynamical variables, atomic coordinates and electronic
wavefunctions, avoid the explicit minimization of the DFT total energy at each step of the dynamics.
Instead, they introduce a fictitious dynamics of the wavefunctions representing an adiabatic updating
on-the-fly of the electronic structure along the atomic dynamics.

Introduction
The main target in atomic-scale simulations is to reproduce in a realistic way physical and chemical events
occurring in materials. Specifically, the scope of First Principles Molecular Dynamics (FPMD) is to study
a system of interacting nuclei and electrons by recreating it on a computer in a way as close as possible to
nature and by simulating its dynamics over a physical length of time relevant to the properties of interest.
The inherent complexity of the simulated systems, from solids to biological macromolecules, calls for
methods able to go beyond the simple calculation of the electronic structure of a given set of coordinates
RI representing the positions of atoms. This is exactly the idea that started the entire field of Molecular
Dynamics (MD).
From an historical point of view, the MD approach was introduced by Alder and Wainwright [1] in the
late 1950s to study the interactions of hard spheres. Many important insights concerning the behavior of
simple liquids emerged from their studies, but due to the limitations of the computational facilities and the
pioneering stage of the MD, it was only in 1964 that the first dynamical simulation could be done. That
milestone case focused on liquid Ar with the interatomic interaction modeled by a truncated Lennard-
Jones potential [2]. In a nutshell, any MD method is an iterative numerical scheme for solving some
equations of motion (EOM), representing the physical evolution of the system under study. Modeling the
interaction of atoms with an analytic potential V(RI), especially when chemical bonds evolve in time and
they are broken or formed is a hard task not yet solved apart from a very limited class of chemical species.

*Email: mauro.boero@ipcms.unistra.fr

Page 1 of 10
Encyclopedia of Nanotechnology
DOI 10.1007/978-94-007-6178-0_100946-1
# Springer Science+Business Media Dordrecht 2015

On the other hand, the electronic structure for a general many-body system can be determined with a
computationally reasonable workload by means of the density functional theory (DFT), originally
proposed in the early 1960s by Kohn, Hohenberg, and Sham [3, 4]. Its importance in the advancement
of computational quantum chemistry and related fields was acknowledged by the Nobel Prize in
Chemistry in 1998 awarded jointly to Walter Kohn and John A. Pople. Joining the two fields, MD and
DFT, is exactly what the Car–Parrinello method is about, extending the range of both concepts [5, 6].

A Brief Overview of Density Functional Theory: The CPMD Potential

The DFT is a formulation of the many-body quantum mechanics in terms of an electron density
distribution, r(x), which describes the ground state of a system composed of interacting electrons and
point-like nuclei having positions {RI} [7]. All along the text, atomic units will be used for simplicity. In
practice, single-particle wavefunctions ci(x) are used to express the many-body mathematical function r
(x). The dramatic simplification, is the fact that not even the specific analytic form of the complex function
ci(x) matters, but only its square modulus, so that the electron density reads

X
N occ
rðxÞ ¼ fi jci ðxÞj2 (1)
i¼1

This expression is clearly a single Slater determinant constructed from wavefunctions representing all the
Nocc occupied orbitals. The coefficients fi are the (integer) occupation numbers, and they are equal to 1 in
the case in which the spin is explicitly considered (spin-unrestricted) or equal to 2 if the spin is neglected
and energy levels are considered as doubly-occupied (spin-restricted). Furthermore, the wavefunctions
ci(x) are subject to the orthonormality constraint
ð
ci ðxÞcj ðxÞd 3 x ¼ dij (2)

as in any quantum mechanics approach. The Kohn-Sham (KS) DFT total energy of the system in its
ground state is then written as

E KS ½fci g ¼ E k ½fci g þ E H ½r þ E xc ½r þ E eI ½r þ EII (3)

In Eq. 3, the first three terms on the right-hand side (Ek, EH, Exc) describe all the electron–electron
interactions, the fourth term (EeI) refers to the electron–nucleus interaction, and the fifth one (EII)
corresponds to the nucleus–nucleus interaction. More explicitly, Ek is the Schrödinger-like kinetic energy
expressed in terms of the single-particle wavefunctions ci(x) as

N occ ð
X  
 1 2
E k ½f c i g  ¼ fi ci ðxÞ  ∇ ci ðxÞ d 3 x (4)
i¼1
2

It must be remarked that this expression for the kinetic energy does not depend on the density r(x) but
directly on the wavefunctions. The second term, EH, is the Hartree energy, i.e., the Coulomb electrostatic
interaction between two charge distributions

Page 2 of 10
Encyclopedia of Nanotechnology
DOI 10.1007/978-94-007-6178-0_100946-1
# Springer Science+Business Media Dordrecht 2015

ðð
1 rðxÞrðyÞ 3 3
E H ½r  ¼ d xd y (5)
2 jx  y j

The exchange interaction and the electron correlations due to many-body effects are represented by the
term Exc[r], whose exact analytical expression is unknown. There are good approximations derived from
the homogeneous electron gas limit for the exchange interaction [7], the so-called local density approx-
imation (LDA), whose name comes from the fact that a homogeneous distribution of interacting electrons
is assumed, in which r(x) depends just on the local point x. Similarly, in the LDA version of the
correlation energy [7], the explicit analytic form of the functional comes from a parameterization of the
results of quantum Monte Carlo calculations. Due to the insufficiency of a simple LDA approximation for
many real systems, nonlocal approximations, including the gradient of the density, are often adopted and
the exchange-correlation functional becomes Exc[r, ∇r]. In practical applications, however, the gradient
enters only with its modulus, thus adding only a modest computational cost. The electrostatic interaction
between electrons and nuclei, is then
ðXM
Z I rðxÞ 3
E eI ½r ¼  d x (6)
I¼1
jx-RI j

where ZI is the charge of the Ith nucleus. However, in practice, this expression “as is” is computationally
expensive. In fact, two different length scales come into play: a small one for the core electrons,
characterized by rapidly varying wavefunctions, especially in the region very close to the nucleus, and
a longer one for the valence electrons that form chemical bonds and vary more smoothly. Clearly, the first
one would dominate and add a computational workload that would make impractical simulations of large
systems. To overcome this problem, one can observe that core electrons are generally inert and do not
participate to chemical bonds. This crucial observation led to the use of pseudopotentials [6]. Namely,
core electrons are eliminated and a potential describing the core–valence interaction is built by fitting to
the all-electron solutions of the Schrödinger or Dirac equation for the single atom of the chemical species
considered. In a pseudopotential (PP) approach, the electron–nucleus interaction is rewritten as
ð
EeI ½r ¼ d 3 x V ps ðx  RI Þ  rðxÞ (7)

Finally, the fifth and last term in right-hand side of Eq. 3 is simply the Coulomb interaction between two
classical nuclei I and J and is written as

X
M
ZI ZJ
E II ¼ (8)
I<J
jR I  R J j

where ZI and ZJ are the net valence charge in a PP approach. The total energy EKS of the ground state of
such a system of interacting electrons and nuclei is obtained by minimizing the KS functional with respect
to the single-particle orbitals ci(x), which, in practice, means solving the KS Schrödinger-like equations

d EKS
  H ci ðxÞ ¼ ei ci ðxÞ
KS
(9)
d ci

Page 3 of 10
Encyclopedia of Nanotechnology
DOI 10.1007/978-94-007-6178-0_100946-1
# Springer Science+Business Media Dordrecht 2015

The Basis Set Issue

A somehow arbitrary issue is the proper definition of ci(x). The answer is the selection of a proper basis
set on which orbitals can be expanded. One possible choice is a localized basis set expressing the
one-electron wavefunctions as

X
M
ci ðxÞ ¼ cki fk ðx; fRI gÞ (10)
k¼1

and the number of analytic functions used, M, is also an indicator of the computational cost of the quantum
calculation, in the obvious sense that the larger the basis set, the higher the computational workload. One
of the most popular basis sets is represented by Gaussian-type orbitals (GTO)
 
fk ðx  RI Þ ¼ fk ðrÞ ¼ N k  rkx x rky y rkz z  exp ak  r2 (11)

where r = xRI. When such a basis set is used, the constants Nk, and ak are kept fixed during the
electronic structure calculation, whereas the coefficients cki are allowed to vary until they are fully
optimized [8]. It must be remarked that orbitals expanded in a localized basis set depend on the atomic
positions RI. As a consequence, in any calculation in which the forces acting on the ions are required, the
explicit derivatives of these wavefunctions with respect to RI must be computed, leading to
non-Hellmann-Feynman force components known in the literature as Pulay forces [6, 8]. An alternative
basis set rather popular in physics is represented by plane waves (PW)

X
Gmax
ci ðxÞ ¼ ci ðGÞeiGx (12)
G¼0

where the sum is truncated at a suitable cut-off Gmax. In this case, no dependence on the atomic
coordinates and no arbitrariness in the increase in the number of basis functions exist.

First Principles Molecular Dynamics

Until the early 1980s, few applications of DFT went beyond the static calculations of the electronic
structure. Nonetheless, finite temperature and entropy effects are two of the dominant features in matter
and their role is often far from negligible. In this respect, the FPMD has represented a huge step forward.
In this particular combination of DFT and MD, the interactions among atoms, instead of being described
by an analytical function V({RI}) of the atomic coordinates RI, is directly computed from the DFT total
energy EKS, which is simultaneously a function of the electron wavefunctions and the atomic coordinates.
The Born–Oppenheimer (BO) approximation [9] allows to disentangle the motions of the electrons and
the nuclei, and each time the nuclei RI(t) are displaced from given positions at time t to new positions
RI(t + Dt) at a subsequent time t + Dt, an optimization of the electronic structure has to be performed. Then
the forces acting on the nuclei are estimated from the gradient of EKS with respect to the ionic position and
the variables RI(t) are updated to RI(t + Dt) by solving via finite differences the Newton-like EOM

€ I ¼ ∇RI min E KS ½fci g, fRI g

MIR (13)
fci g

Page 4 of 10
Encyclopedia of Nanotechnology
DOI 10.1007/978-94-007-6178-0_100946-1
# Springer Science+Business Media Dordrecht 2015

This iterative procedure assumes that the electronic structure is recomputed and the full diagonalization of
the Hamiltonian is performed at each time step t along the discrete trajectory {RI (t)}.

Car–Parrinello Molecular Dynamics

An alternative to this scheme, which has represented a real breakthrough in first principles dynamical
simulations, was proposed in 1985 by R. Car and M. Parrinello [5]. The scope (and driving force) was to
overcome the two major efforts arising in FPMD: On one hand one has to integrate the equations of
motion for the nuclear positions as in Eq. 13, which represent the long timescale part to the problem. On
the other hand, one has to propagate dynamically the smooth time-evolving (ground state) electronic
subsystem. The Car–Parrinello molecular dynamics (CPMD) is able to satisfy this second requirement in
a numerically stable way and makes an acceptable compromise for the time step length of the nuclear
motion. The formulation is an extension of a classical molecular dynamics Lagrangian in which the
electronic degrees of freedom (wavefunctions) are added to the system, along with any other dynamical
variable qa(t), i.e., thermostats, barostats, reaction coordinates, etc.

1X X ð 2 1X
_ m c_ i ðxÞ d 3 x þ
2 2
L CP
¼ M I RI þ a q_ a  E KS ½r, fRI g, qa 
2 I i
2 a
X  ð 
þ lij d 3 x ci ðxÞcj ðxÞ  dij (14)
ij

The first term on the right-hand side of Eq. 14 is the kinetic energy of the nuclei, the second one the
fictitious kinetic energy of the electrons representing the update of the wavefunctions during the
dynamics, the third one the kinetic term of any further dynamical variable used in the sense specified
above, the fourth one is the DFT total energy, and the last addendum is the orthonormality constraint for
the wavefunctions. The kinetic energy for the electronic degrees of freedom is the main novelty of the
CPMD approach: A strategy to update on-the-fly the wavefunctions when ions undergo a dynamical
displacement, avoiding expensive iterative diagonalization required by the BO approach at each time step.
The Euler–Lagrange EOM are as follows:

dE KS X
m c€i ðxÞ ¼  þ lij cj ðxÞ (15)
dci j

€ I ¼ ∇RI E KS
MIR (16)

@E KS
a q€a ¼  (17)
@qa

The fictitious mass m assigned to the orbitals ci(x) is the parameter that controls the speed of the updating
of the wavefunctions with respect to the nuclear positions. For this reason, it determines the degree of
adiabaticity of the two subsystems, electrons and nuclei.
It is straightforward to give a Hamiltonian, instead of a Lagrangian, formulation of the CPMD method,
via a simple Legendre transform after defining the momenta

Page 5 of 10
Encyclopedia of Nanotechnology
DOI 10.1007/978-94-007-6178-0_100946-1
# Springer Science+Business Media Dordrecht 2015

dLCP d LCP 
pi ðxÞ ¼ ðxÞ ¼ m c_ i ðxÞ pi ðxÞ ¼ ðxÞ ¼ m c_ i ðxÞ (18)
_
dc d c_
i i

pi ¼ ∇R_ I LCP ¼ M I R_ I (19)

@L CP
z_a ¼ ¼ a q_a (20)
@ q_ a

so that the Hamiltonian reads

X p_ 2 X ð p ðxÞ pi ðxÞ X z_2

H CP
¼ I
þ i
d xþ
3 a
 E KS ½r, fRI g, qa 
2M I m 2
X ð  a
I i a
(21)
 lij ci ðxÞcj ðxÞ d 3 x  dij
ij

and the CPMD equations of motion (Fig. 1) will be given by the corresponding Hamilton EOMs.
A rigorous mathematical proof of this scheme has been given by Bornemann and Sch€
utte [10], showing
CP BO
that the CPMD trajectory {R (t)} stays close to the BO one {R (t)} and the upper bound is
proportional to the square root of the fictitious mass m
 CP 
R ðt Þ  RBO ðt Þ < C  pffiffiffi
m (22)

where C is a positive constant (Fig. 2).

The fact that the CPMD was a milestone step forward in realistic simulations of materials at various
thermodynamics conditions can be easily seen by the number of publications in first principles molecular
dynamics (FPMD) before and after 1985. Indeed, the original Car–Parrinello publication has more than
7,000 citations in 2014 (source: ISI Web of Science), and, to acknowledge the importance of the method,

Fig. 1 Atomic structure (sticks; red = O, white = H) of a water molecule and electronic wavefunction of an O–Hs-bond in
terms of density map (blue: |ci(x)|2 = 0, red: |ci(x)|2 = maximum). The two sets of dynamical variable evolve in time according
to the equations of motions indicated, coupled via EKS

Page 6 of 10
Encyclopedia of Nanotechnology
DOI 10.1007/978-94-007-6178-0_100946-1
# Springer Science+Business Media Dordrecht 2015

Fig. 2 Schematic representation of a Car–Parrinello trajectory (red line) with respect to a Born–Oppenheimer dynamics (blue
line) one a given DFT-based potential energy surface

the international Physics and Astronomy Classification Scheme (PACS) introduced in 1996 a new
identification number, 71.15.Pd, to classify Car–Parrinello related publications. Since then, the method
has been applied to a wide variety of materials, ranging from solids, to liquids, and to biological systems
[11, 12].

Numerical Details
Although it is not restriction neither of DFT [13] nor of CPMD [14], PWs are often used as a convenient
basis set for the coding of CPMD, since they have several good properties: (i) Accuracy can be
systematically improved in a fully variational way, (ii) PWs are independent from atomic positions
(i.e., no Pulay forces [6]), (iii) PWs can be easily distributed in parallel processing. However, it must be
observed that the fact that PWs are not localized can lead to inefficiencies for small clusters or surfaces
placed in a large simulation cell. The equations of motion are discretized by finite differences, via a Verlet
or velocity-Verlet algorithm [15]. The ionic variables RI(t) are updated at a rate Dt, while the electronic
degrees of freedom are updated at a rate Dt/m1/2, i.e.,
m X   X
½ c i ð G, t þ Dt Þ þ ð G, t  Dt Þ  ð G, t Þ ¼  GH CP G0 ci ðG0 Þ þ Lij cj ðGÞ (23)
Dt2 G 0 j

For most of the applications, Dt and m fall in the range 3–5 au and 300–600 au, respectively. Of course, the
(quantum) time scale of electrons is dominating in this kind of approaches and simulations times are of the
order of few tens or, at very best, hundreds of ps. As far as the system size is concerned, with N electrons
and Gmax PWs, Gmax being integer, the scaling of the various parts composing the CPMD algorithm is
O(N Gmax) for the kinetic term, O(N Gmax log Gmax) for the local potential and O(N2 Gmax) for both the
nonlocal term and wavefunctions orthogonalization procedure.

Second-Generation Car–Parrinello Molecular Dynamics

The inherent high computational cost associated to the electronic structure calculations has limited the
affordable timescales for several years. Only the most advanced high performing computer platforms
recently available have allowed to increase the system size to about a thousand of atoms and simulation
times towards hundreds of picosecond. Yet, many phenomena still call for a substantial boost. These are,
for instance, diffusion in solids or, in the case of glasses generation from the melt, a less rapid cooling rate
suitable to avoid numerically induced structural problems. While linear scaling methods can be a viable

Page 7 of 10
Encyclopedia of Nanotechnology
DOI 10.1007/978-94-007-6178-0_100946-1
# Springer Science+Business Media Dordrecht 2015

way to access larger system sizes, they still have to face the problem of the simulation timescale.
Moreover, the crossover point at which linear scaling methods become advantageous has remained fairly
large, especially if high accuracy is needed. An interesting attempt at overcoming these limitations has
been proposed in 2007 [16]. The basic idea is to join the advantages of both the BO approach and the
CPMD; in a “nutshell”
CPMD BO
Conservation of constants of motion Good Convergence dependent
Electronic optimization Not needed Needed
Hamiltonian diagonalization Not needed Needed
Integration step Dt Small Large
Minimum of the BO surface Approximate Exact

These two approaches have nearly complementary features as sketched above. Following the CPMD
formulation, it can be remarked that the Lagrangian formulation for the propagation of the wavefunctions
is stable by construction, thus providing a reliable integration. This stability feature must then be
preserved. Concerning the efficiency, large integration steps Dt are desirable and possibly a small, or
better zero, deviation from the BO surface should be kept all along the dynamics to get a high accuracy.
The mathematical result of this list of requirement resumes into a modified ionic EOM
!
@E X @ D E @ h c j dE X  E
€I ¼  Lij cj
N SC N SC
MIR þ Lij ci j cj  2 i
 (24)
@RI i, j
@R I @R I d hc i j j

While the first term in the right hand of the equation is clear, the rest seems a bit puzzling at a first glance.
Indeed, in the original formulation [16], the selected basis set is not PW, but a localized basis set as in
Eq. 10. Hence, the electronic wavefunctions depend also on the atomic coordinates RI and the request of
orthogonality at each step is released to save time, meaning that the scalar product <ci|cj> is no longer
vanishing. Analogously, the total energy EDFT is not reoptimized as in full self-consistent BO procedures
and for this reason is indicated as non-self-consistent energy ENSC. Wavefunctions are propagated
according to an algorithm which resembles the original CPMD formulation in the sense that second
order EOMs are used, but a damping term (first-order derivative) is present which reminds a sort of
steepest-descent algorithm typical of the BO dynamics. The net result is the electron dynamics,

d2 d d E N SC X
m j c i þ g j c i ¼  þ Lij jci i (25)
dt2 i
dt i
d hci j j

which is then solved via a predictor–corrector scheme. With no pretention of completeness, the procedure
can be summarized as follows. On a first instance, in a localized basis set {|q>}, the electronic
wavefunctions are expanded as

X
M
jc i i ¼ Ciq jqi (26)
q¼1

on the M functions composing the localized basis. Then the NxM matrix of the expansion coefficient is
written as

Page 8 of 10
Encyclopedia of Nanotechnology
DOI 10.1007/978-94-007-6178-0_100946-1
# Springer Science+Business Media Dordrecht 2015

0 1
C 11 : : C 1M
B : : : C
C ¼ C iq ¼B
@ :
C (27)
: : A
CN 1 : : C NM

and the density matrix becomes P = CCT = PSP. The MxM matrix indicated as S is just given by the
expansion coefficient and its matrix elements have the usual form

X
N
S qq0 ¼ C iq C iq (28)
i¼1

Hence, the (DFT) total energy can be rewritten as Etot[C,RI] which can be used in a straightforward way to
write a BO dynamics

€ I ¼ ∇RI min E tot ½C, RI 

MR 8CT  S  C ¼ 1 (29)

under the given constraint on C which resumes in an implicit orthogonality condition. However it must be
kept into account that: (i) Diagonalization and minimization of Etot are required in BO; (ii) Hellman-
Feynman forces are just one component since Pulay forces due to the local basis set are present. Residual
force components appear due to non-self-consistency (NSC) of the approach. To take into account all the
points above, the basic strategy can be summarized in four major points:

1. Propagate the electronic variables in time according to the CP original idea of updating on-the-fly to
avoid expensive full diagonalization operations
2. Use a good propagation algorithm C(tn) = f(C(tn-1), . . . ,C(tn-m)) depending on previous time steps
m  [1,K] time steps
3. Select the appropriate number of steps K to keep C(tn) as close as possible to the (electronic) ground
state
4. Enforce convergence on the BO surface, correct this propagation CPMD-like afterwards

Point 3 corresponds to the first move in the numerical integration procedure and it can be identified as
the “predictor” part directly deriving from a standard numerical integration of the CPMD type equations
of motion. Point 5, instead, is the “corrector” needed afterwards to better converge the wavefunctions and
to restore the neglected self-consistent loop. The use of not necessarily fully converged wavefunctions at
the predictor propagation stage allows for large integration steps, thus resulting in a remarkable boost in
the dynamics.

Cross-References
▶ Ab Initio DFT Simulations of Nanostructures
▶ Computer Modeling and Simulation of Materials
▶ Electronic Structure Calculations
▶ First Principles Calculations

Page 9 of 10
Encyclopedia of Nanotechnology
DOI 10.1007/978-94-007-6178-0_100946-1
# Springer Science+Business Media Dordrecht 2015

▶ Molecular Dynamics Method

▶ Molecular Dynamics Simulations of Interactions Between Biological Molecules and Nanomaterials
▶ Molecular Dynamics Simulations of Nano-Bio Materials
▶ Surface Electronic Structure

References
1. Alder, B.J., Wainwright, T.E.J.: Phase transition for a hard sphere system. Chem. Phys. 27, 1208
(1957)
2. Rahman, A.: Correlation in the motion of atoms in liquid argon. Phys. Rev. 136, A405 (1964)
3. Hohenberg, P., Kohn, W.: Inhomogeneous electron gas. Phys. Rev. 136, B864 (1964)
4. Kohn, W., Sham, L.J.: Self-consistent equations including exchange and correlation effects. Phys.
Rev. 140, A1133 (1965)
5. Car, R., Parrinello, M.: Unified approach for molecular dynamics and Density-Functional theory.
Phys. Rev. Lett. 55, 2471 (1985)
6. Marx, D., Hutter, J.: Ab Initio Molecular Dynamics: Basic Theory and Advanced Methods. Cam-
bridge University Press, New York (2009). ISBN 978-0521898638
7. Parr, R.G., Yang, W.: Density-Functional Theory of Atoms and Molecules. Oxford University Press,
New York (1989). ISBN 0-19-504279
8. Hehre, W.J., Radom, L., Schleyer, P.V.R., Pople, J.A.: Ab Initio Molecular Orbital Theory. Wiley,
New York (1986). ISBN 978-0471812418
9. Born, M., Oppenheimer, J.R.: Zur Quantentheorie der Molekeln. Ann. Phys. 84, 457 (1927)
10. Bornemann, F.A., Sch€ utte, C.: A mathematical investigation of the Car-Parrinello method. Numer.
Mat. 78, 359 (1998)
11. Boero, M., Tateno, M.: Quantum theoretical approaches to proteins and nucleic acids. In: Oxford
Handbook of Nanoscience and Technology, Vol. 1: Basic Aspects, pp. 549–598. Oxford University
Press, New York (2010). ISBN 978-0199533046
12. Boero, M.: Reactive simulations for biochemical processes. In: Atomic-Scale Modeling of
Nanosystems and Nanostructured Materials. Lecture Notes in Physics, vol. 795, pp. 81–98. Springer,
Berlin/Heidelberg (2010). ISBN 978-3-642-04650-6
13. Oshiyama, A., Iwata, J.: Large-scale electronic-structure calculations for nanomaterials in density
functional theory. J. Phys. Conf. Ser. 302, 012030 (2011)
14. Car, R., Parrinello, M.: The unified approach to density functional and molecular dynamics in real
space. Solid State Commun 62, 403 (1987)
15. Pang, T.: An Introduction to Computational Physics. Cambridge University Press, New York (2000).
ISBN 978-1840858839
16. K€uhne, T.D., Krack, M., Mohamed, F.R., Parrinello, M.: Efficient and accurate Car-Parrinello-like
approach to Born-Oppenheimer molecular dynamics. Phys. Rev. Lett. 98, 066401 (2007)

Page 10 of 10