Brazilian Journal of Physics, vol. 29, no. 1, March, 1999 199
Sto chastic Classical Molecular Dynamics Coupled
to Functional Density Theory: Applications
to Large Molecular Systems
K. C. Mundim
Institute of Physics, Federal University of Bahia,
Salvador, Ba, Brazil
and
D. E. Ellis
Dept. of Chemistry and Materials Research Center
Northwestern University, Evanston IL 60208, USA
Received 07 Decemb er, 1998
Ahybrid approach is describ ed, which combines sto chastic classical molecular dynamics and
rst principles DensityFunctional theory to mo del the atomic and electronic structure of
large molecular and solid-state systems. The sto chastic molecular dynamics using Gener-
alized Simulated Annealing GSA is based on the nonextensive statistical mechanics and
thermo dynamics. Examples are given of applications in linear-chain p olymers, structural
ceramics, impurities in metals, and pharmacological molecule-protein interactions.
scrib e each of the comp onent pro cedures. I Intro duction
In complex materials and biophysics problems the num-
b er of degrees of freedom in nuclear and electronic co or-
dinates is currently to o large for e ective treatmentby
purely rst principles computation. Alternative tech-
niques whichinvolvea hybrid mix of classical and quan-
tum metho dologies can provide a p owerful to ol for anal-
ysis of structure, b onding, mechanical, electrical, and
sp ectroscopic prop erties. In this rep ort we describ e an
implementation which has b een evolved to deal sp ecif-
ically with protein folding, pharmacological molecule
do cking, impurities, defects, interfaces, grain b ound-
aries in crystals and related problems of 'real' solids.
As in anyevolving scheme, there is still muchroom
for improvement; however the guiding principles are
simple: to obtain a lo cal, chemically intuitive descrip-
tion of complex systems, which can b e extended in
a systematic way to the nanometer size scale. The
computational scheme is illustrated in Fig. 1, show-
Figure 1. Flowchart describing the hybrid MD+DV+GSA
ing classical Molecular Dynamics MD, Monte Carlo
approach.
MC-GSA sto chastic sampling, and Discrete Varia-
tional DV DensityFunctional DF quantum mechan-
ics pro cedures coupled together. We next brie y de-
200 K.C. Mundim and D. E. Ellis
II Classical Dynamics, Sto chas- metho ds.
- In MM, it is very simple to intro duce time evolu-
tic Metho ds, and Quantum
tion;
Clusters
- In MM, it is p ossible to intro duce the temp erature
as an external p erturbation, and trace its e ects.
Two imp ortant questions arise which are unfavor-
I I.1 Classical Metho dology: Molecular Dynam-
able to the MM metho ds: First, there do not exist well-
ics
de ned rules to evaluate the force constants; and sec-
ond, in order to cho ose the b est force eld is necessary
The idea of molecular mo deling is an attempt to de-
to have a priori some considerable knowledge ab out the
scrib e the quantum chemical b onds in terms of a clas-
molecular system. In molecular mechanics the atom is
sical force eld in Newton's equations. In a Molecular
represented by a spherical b o dy with a particle mass
Mechanics MM mo del the atomic b onds are repre-
equal, in general, to the resp ective atomic mass. In
sented by springs joining the atoms; i.e., the molecule
to day's molecular mechanics, several force eld mo dels
is assumed to b e a collection of masses and springs.
have b een prop osed. In general the molecular energy
In this case is necessary nd a set of parameters, for
p otential function related with the classical force eld
this force eld, that t the quantum atomic interac-
can b e expressed by the following sum [3],[4];
tions with sucient accuracy. Ideally one hop es that
such re nements will eventually lead toward a uni ed
computational mo del that can successfully mimic ob-
V = V + V + V + V + V + V 1
H b tor i tor p C vdW
served molecular prop erties. Academic research e orts
and the pharmaceutical industry interest in developing
N
H
X
1
new comp ounds in biological molecular systems have
2
K r r V =
Hn n o H
2
stimulated the app earance of di erent computational
n
co des based on classical force elds.
The rst molecular dynamics applications was p er-
N
X
1
2
K V = formed by Alder and Wainwright, [1] who used a p er-
n n o b
2
n
fectly elastic hard sphere mo del to represent the atomic
interactions. One of the most widely used force eld
N
ones is called MM1 Molecular Mechanics, prop osed
X
1
2
V = K
tor i n n o
by Allinger [2]. At the moment there are an uncount-
2
n
able numb er of MM computational co des using a classi-
cal force eld. Each one of these uses a particular force
N
X
1
eld to describ e di erent molecular prop erties and to
2
K [1 + cosm ] V =
n n n tor p
2
t some exp erimental results.
n
Reasons which justify the increasing use of MM can
b e listed, as for example:
Nc
X
1 q q
i j
Vc =
-Possibly the chief reason is the relatively short
4 r
ij
i computational time, which for MM metho ds increases 2 as M , where M is the numb er of atoms in the molecule. " Nv In contrast, the use of the ab-initio quantum metho ds X C ij C ij 12 6 V = vdW 12 6 in such molecular systems is computationally impracti- r r ij ij i cable b ecause the computer time to evaluate the inter- 4 electronic repulsion integrals increases as N or more Where, in summary, the rst term of eq. 1 V H rapidly, when correlation is taken into account, where represent the energy necessary to stretch or compress N is the numb er of basis orbitals. Normally, there are the atomic b ond; Vb is the p otential contribution that at least several basis functions p er atomic orbital shell- represents the b ond-angle b ending interaction; V , tor i 1s, 2s, 2p, etc.. V are the torsional contributions that represent tor p the harmonic dihedral b ending interaction and the si- - The MM metho d and its results are conceptually nusoidal dihedral torsion interactions. The last terms easier to understand than quantum mechanical QM Brazilian Journal of Physics, vol. 29, no. 1, March, 1999 201 V and V represent the non-b onded interactions the time t + t. In the same pro cedure the forces and C vdW such as Coulombic repulsion, hydrogen b onding and the new velo cities are used to calculate the new atomic van der Waals interactions. Each of these p otential en- p osition at time t + t, and so on. ergy functions represents a molecular deformation from Once the force on each atom has b een calculated for a reference geometry. The interactions given in equa- time t, the p osition of each atom at some later time t tion 1 are available in our Molecular Mechanics co de, + t is then given by and are typical of other currently available co des. To ! ! ! ! 2 study molecular prop erties in solid systems, additional r r + v t+ a t 3 = t t t t+t p otentials have b een intro duced, including Born-typ e ! ! a where = /m . Here t = n t, and t 0.001 pico F exp onential repulsion and Morse p otentials. ! r may b e taken seconds. The initial co ordinates t=0 Molecular dynamics MD calculations consist of from exp erimental data or may b e randomized; the analyzing the evolution of the molecular system with ! v are derived from the Maxwell- initial velo cities t=0 time. In this case the atoms are continuously moving, Boltzmann distribution, in general, and temp erature T the b onds are vibrating, the angles are b ending and is controlled by a dynamic scaling pro cedure. A co oling the whole molecule is rotating. In MD, successive con- or annealing pro cedure of mo difying T provides one way gurations of the system are generated byintegrating to approach stationary p oints, and p ossibly the equi- Newton's equations of motion. The result is a tra jec- librium state of the system. Once the p otentials are tory that sp eci es how the p ositions and velo cities of calculated for each new geometry, the new forces and the atoms vary with time. The tra jectory is obtained p ositions of each atom can again found. This pro cedure by solving the following di erential equations; is rep eated until convergence in the energy is reached. ! ! 2 There are di erent algorithms or numerical metho ds d r F i i 2 = 2 available to solve di erential equation 2. All meth- dt m i ! ! o ds are based on nite di erences and solve the equa- r where is the force over atom i along the co ordinate F i tion step by step in time. Often the step size is taken and m is the atomic mass. i constant, but adaptive metho ds have sometimes b een Equation 2 is solved by a di erence nite tech- found useful. nique with continuous p otential mo dels 1, which are The simplest and most straightforward but unfor- generally assumed to b e pairwise additive. The essen- tunately not suciently accurate approachistouse tial idea is that the integration is broken down into the Taylor expansion based on equation 2. many small stages, each separated in time by a xed increment dt. For each step the total force on each ! ! ! ! 2 r r = + v t+ a t 4 t+t t t t atom is calculated as the vector sum of the interactions with some or all atoms b elonging to the molecular sys- ! ! ! tem. The force is calculated by taking the gradientof v v = + a t t+t t t the p otential function represented by eq. 1. Using Another metho d sometimes used but computation- the forces and accelerations, all atomic p ositions and ally very exp ensive is the Taylor predictor; velo cities are determined using Newton's equation at c ! ! ! 2 3 r r r ! ! d 1 d d 1 2 3 4 r r = + t + t + t + O r 5 t+t t 2 2 dt 2 dt 6 dt ! ! ! 2 3 r r r 1 d 1 d ! ! d 2 3 4 r r t + t t + O r = t t t 2 2 dt 2 dt 6 dt 202 K.C. Mundim and D. E. Ellis This algorithm contains no explicit velo cities, and This metho dology, though p edagogically correct, al- we see that also the third derivative terms can b e made lows the system to get \trapp ed" in lo cal energy min- to cancel. In this case the velo cities can b e approxi- ima. There is no guarantee that this geometry corre- mated by sp onds to the global minimum energy. So the usual MD pro cedure may not b e the b est way to search for ! ! ! r r v =2t 6 = the global minimum p oint on the complex energy hy- t 1 t+1 t p ersurface. Another very well known algorithm is the Verlet In this context, if the molecular hyp ersurface en- leap-frog scheme given as; ergy is non-convex, sto chastic molecular dynamics of- ! ! fers a more ecientway of nding b oth lo cal minimaas v v = +F t 7 t t+t=2 t t=2 well as the global one. Thus we next discuss the more ! ! ! imp ortant concepts needed to implement a sto chastic v r =r + t t+t t t+t=2 molecular dynamics pro cedure. One consideration for molecular dynamics that im- mediately invalidates a numb er of these metho ds is the I I.2 Sto chastic Dynamics: Generalized Simu- cost of calculation of the force, or gradient of the p oten- lated Annealing tial. Computation of the force is extremely lab orious compared to any manipulationinvolved in up dating the variables to take one step forward in time. This means Simulated Annealing [5][6] metho ds have b een ap- that any metho d that involves more than one force eval- plied successfully in the description of a variety of global uation p er step cannot b e a go o d choice. This rules out extremization problems. Simulated Annealing metho ds the well-known Runge-Kutta metho d and its variants, have attracted signi cant attention due to their suit- that require up to four force evaluations p er step, so ability for large scale optimization problems, esp ecially the simple Verlet scheme is generally preferred. for those in which a desired global minimum is hidden The MD pro cedure can b e brie y summarized in the among many lo cal minima. The basic asp ect of the Sim- following steps: ulated Annealing metho d is that it is analogous to ther- i- Guess the initial molecular geometry, the initial mo dynamics, esp ecially concerning the way that liquids atomic velo cities and environment temp erature. freeze and crystallize, or that metals co ol and anneal. ii- Calculate the force, as the gradient of the p oten- At high temp eratures, the molecules move freely with tial 1. resp ect to one another. If the system is co oled slowly, iii- Calculate the new atomic p osition or molecu- thermal mobility is lost. The atoms are often able to lar geometry using one of the algorithms to integrate line themselves up and assume a molecular geometry Newton's equations. that is in general a lo cal equilibrium state. The simu- iv- Rede ne the atomic velo cities using the virial lated annealing pro cedure is actually more complicated theorem than the combinatory one, since the familiar problem of + * X 3 1 2 long, narrow p otential valleys again asserts itself. Sim- nk T = m v B i i 2 2 ulated annealing, as we will see, tries random steps, but i cy cle in a long, narrowvalley, almost all random steps are up- r hill. The amazing fact is that, for a slowly co oled sys- T v o new = v T tem, nature is able to nd this minimum energy state. old So the essence of the pro cess is slow co oling, allowing Where k is the Boltzmann's constant, T is the ini- B o ample time for redistribution of the atoms as they lose tial or \bath" temp erature and T is the lo cal temp era- their mobility. This is the technical de nition of an- ture. v and v are the velo cities of two consecutive old new nealing, and it is essential for ensuring that the lowest steps in the MD pro cedure. energy state will b e achieved. v- Saveintermediate information for statistics and go back to step i until the stabilization of the molec- The rst nontrivial solution along this line was pro- ular energy is reached. At the end of the MD pro ce- vided by Kirkpatrick et al. [5],[6] for classical systems, dure the total energy kinetic and p otential contribu- and also extended by Cep erley et al. [7] to quan- tion must b e constant. tum systems. It strictly follows the quasi-equilibrium Brazilian Journal of Physics, vol. 29, no. 1, March, 1999 203 Boltzmann-Gibbs statistics using a Gaussian visiting sition. distribution, and is sometimes referred to as Classical To generate the random vector x the present GSA t Simulated Annealing CSA or the Boltzmann machine. routine uses an extension of the pro cedure given in Ref.