Basics of

Molecular Dynamics Simulation

CHEM 430
A Nobel Prize for Molecular
Dynamics and QM/MM
What is Classical Molecular Dynamics?

Simulation of explicit particles (atoms, ions, . . . )

Particles interact via relatively simple analytical potential
functions
Newton’s equations of motion are integrated for all
particles simultaneously
100-1,000,000’s of particles depending on model
Time 10 ps to 1 µs depending on model (typically ns)
Why Classical Molecular Dynamics?
Includes temperature and pressure e ects
Able to treat comparatively large systems for a relatively
long time
Help interpret experiments, and to provide alternative
interpretations
Test theories of intra- and intermolecular interaction
Obtain detailed molecular level information
Structure and dynamics (compared to QC, MM, MC . . . )
Improve our understanding of nature through
model-building (!!)
Phase Space and Time Averages
A system containing N atoms has 6 N values defining the
state of the system (3 N coordinates ( r ), 3 N momenta ( p))
Each combination of coordinates and momenta define a
point in the 6N-dimensional phase space : Γ N
Consider a property that can be described as a function of
the coordinates and momenta, A(Γ) , (for instance the total
energy or the (instantaneous) pressure)
The average value of the property can then be written as:
τ
1
Aobs = < A > time = < A(Γ( t)) > time = lim A(Γ( t)) dt
τ ->∞ τ t= 0

The ergodic hypothesis: “The long time average is equal to

the ensemble average in the limit of an infinite number of
members in the ensemble”
Molecular Dynamics
A MD simulation generates a sequence of points in phase
space connected in time
The result is a trajectory of all particles in the system as a
function of time
Time averages and other properties can be calculated
from this trajectory
Motion of the system through phase space is governed by
Hamiltonian equations of motion :
∂H pi
ṙ i = =
∂ pi mi
∂H
ṗ i = − = fi
∂ ri
The Hamiltonian :
H( r, p) = K ( p) + V (r)
N
p2i
K ( p) =
2m i
i
V is a potential energy function
Newtons equations of motion :
∂H px
ẋ = = = vx
∂ px m
∂H ∂ V ( x)
ṗ x = − = − = F = ma x
∂x ∂x
The Born-Oppenheimer approximation :
Electrons move much faster than the nuclei
¨
QM: Solve Schrodinger equation for static nuclei
MD: Nuclei experience electrons as an average field
Steps in Performing an MD Simulation
Selection of interaction model
Selection of boundary conditions Initialize: Initial particle
positions, velocities ...
Selection of initial conditions (positions,
velocities . . . )
Selection of ensemble (NVE, NVT, NPT . . . ) Calculate forces from
Selection of target temperature, particle positions
density/pressure . . .
Loop
Selection of integrator, thermostat, barostat . . . Solve Newton’s eqns. over
of motion (integrate) time-
Perform simulation until equilibration is steps
reached (property dependent)
Perform analysis /
Perform production simulation to collect Write data to disk
thermodynamic averages, positions, velocities
Analyze the results via post-processing
The Interaction Model
The potential energy can be divided into interactions
between pairs, triplets, . . . of particles :

V (r ) = Vij (r i , r j )+ Vijk ( r i , r j , r k )+ Vijkl (r i , r j , r k , r l ) . . .

ij ijk ijkl

i i
V(ri,rj)

V(ri,rk) j
V(ri,rj,rk) j

V(rj,rk)

k k
The interaction energy is (usually) largely dominated by the
pair term. Truncate the series and make the assumption of
pairwise additivity :

V (r ) ≈ Vij ( r i , r j )
ij

In simulations the “true” pair potential is usually replaced

by an effective pair potential, which incorporate the
average many-body e ects into the first term :

V (r) ≈ Vijeff ( r i , r j )
ij

These potentials are usually parameterized to reproduce

some experimental data
 The Lennard-Jones Potential
σij 12 σij 6
Vij (rij ) = 4 ij −
rij rij
rij = |r i − r j |
2

1.5 Lennard-Jones interaction (Argon)

1
V (kJ/mol)

0.5
σ (3.41 Å)
0

-0.5 ε (0.996 kJ/mol)

-1

-1.5
1 2 3 4 5 6 7 8 9
r (Å) Two dimensional simulation of argon
The Initial Configuration
For a solid it may be possible to start from the
experimentally determined structure
For a liquid it is also common to start from a “solid” lattice,
and letting the structure melt
For a liquid it is also possible to start from a “random”
placement of molecules (this may be tricky)
If previous simulation data is available, one might modify
a configuration to suit the new simulation, for instance by
replacing one or a few particles
Macromolecule in liquid: Use experimental coordinates of
macromolecule from solid phase structure determination,
surround it by liquid
The Initial Velocities
The velocities of the particles is related to the kinetic
energy :
N
m ivi2
K =
2
i

The kinetic energy is related to the temperature :

2 < K > = gkB T = 3NkB T
Assign velocities according to a Boltzmann distribution :
mi v 2
mi 1/ 2 − i
P(vi ) = e kB T
2πkB T
Alternatively assign particle velocities from a uniform
distribution. They will obtain a Boltzmann distribution
after a short time.
It is often necessary to adjust the initial velocities so the
total linear momentum of the system is zero :
N N
p= pi = m iv i = 0
i i
This momentum corresponds to translation of the whole
system
Having p = 0 in simulations of “liquid” water, has been
referred to as a “flying ice cube”
Newton’s equations of motion will preserve the linear
momentum resulting in a NVE P ensemble, resulting in
a system with three fewer degrees of freedom
For a system with periodic boundary conditions, the
angular momentum will not be conserved
Statistical Mechanical Ensembles
The “natural” ensemble for MD is the microcanonical
ensemble: NVE (constant number of particles, constant
volume, constant energy)
Typical experiments: Constant pressure and/or
temperature
Canonical ensemble: NVT (constant number of particles,
volume, temperature); Isobaric, isothermal (Gibbs): NPT
(constant number of particles, pressure, temperature)
Some ways to include such e ects in MD
Stochastic events (change kinetic energy of single particle)
Periodic rescaling of velocities (2 K = gk B T )
Ad hoc partial rescaling of velocities (Berendsen’s thermal
bath)
Extended system methods: Include extra variables in the
equations of motion such as the volume of the system
Periodic Boundary Conditions
A “real” system might contain about N A = 6 · 10 23 molecules
Simulation may contain about 100-1,000,000 molecules
A small cube of 10x10x10 molecules has about half of the molecules on the
surface
Not appropriate to study “bulk” properties
Periodic boundary conditions:
Small box replicated in all directions
A particle that leaves the box on one side
is replaced by an image particle that
enters from the other side
There are no walls and no surface
particles
Some systems inherently contains a boundary, for instance a liquid droplet, or a
surface
The Minimum Image Convention
Particles interact only with the closest periodic image of the other particles:

How to compute minimum image :

Simple and straightforward
rx = xi − x j ;
i f ( r x > 0 .5 * l e n g t h )
rx = rx − l e n g t h ;
i f ( r x < − 0.5 * l e n g t h )
rx = rx + l e n g t h ;

Fortran has a function , anint, which returns the nearest integer

rx = rx - length* anint (rx/length)

In C, the floor function can be used

rx = rx − length*floor(rx/length + 0.5);
Truncating the Potential
The most time consuming part: Computing the
nonbonded (vdw & electrostatic) energies and forces
For a pair additive potential there are N 2 such interactions
Consider the Lennard-Jones argon potential again:
0
-0.1
-0.2 Spherical nonbonded c
-0.3
V (kJ/mol)

-0.4
-0.5 V ij ( rij ) = V ij ( rij ) rij ≤ rc
-0.6
-0.7 V ij ( rij ) = 0 rij > rc
-0.8
-0.9
-1 Disadvantage: Discontinuity
0 1 2 3 4 5 6 7 8 9
r (Å) in the energy and forces
Switching function ( S(rij ) smoothly from 1 to 0)
V ij ( rij ) = V ij ( rij ) rij < rl
V ij ( rij ) = S ( rij ) V ij ( rij ) rl ≤ rij ≤ rc
V ij ( rij ) = 0 rij > rc
Minimum Image and Spherical Cutoff
Very commonly minimum images and spherical cuto s
are used together

rc
Searching for Neighbors
The search loop is a double loop over all particles ( n).
Execution time for the loop is proportional to n2
The Verlet neighbor list : Maintain list of neighbor pairs
closer than the radiu rc rskin

rc ∆t rc

rc+rskin rc+rskin

Update list from time to time ( about once every 25 timesteps,

but this can be automated)
Use of Linked Lists (Cell Method)
For large systems the execution time will be dominated by
the search loop (because building the Verlet neighbor lists
are still O( N 2 )
One solution is to divide the space into cells

a Finding which cell a

particle belongs to is
}
y
an O ( N ) problem. A
particle with coordinate
( x , y ) belongs to cell
}b ( floor (x/a ) , floor (y/b ) )
All particles are sorted
into so-called linked
x lists . There is one list
Cell (0,0) Cell (7,2)
for each cell .
When all particles are sorted into lists , the double loop can be
done over cells instead of particles . We don’t have to consider
all cells . Only cells with particles closer to each other than the
c distance must be considered :

For the particles inside the

central cell only particles
inside the central cell itself
and the shadowed cells are
considered if the cuto
distance is equal to , for
instance , 1.2 cell lengths
Domain Decomposition
For very large systems the memory requirement on each node
becomes large , when using a replicated data algorithm
Also the program will become slow due to that forces from all
particles must be sent to all processors each timestep

One solution for very large

systems is to use domain
decomposition . Particles are
assigned to processors
depending on their position
in space.

When domain decomposition

is used larger systems may
A 3x3 decomposition of a two be simulated simply by using
more processors
dimensional system
For large systems the interaction cuto radius, r c , is
substantially smaller than the size of the system

rc
 Particles closer to the
“edge” of the domain
than the cuto radius,
interact with particles
on the node.

UP/ UP/
LEFT UP RIGHT
8 neighbours in 2D ,
LEFT RIGHT
and 26 in 3D

DOWN/ DOWN/
LEFT DOWN RIGHT
Integrating the Equations of Motion
Motion of all particles is coupled (many-body problem),
This is impossible to solve analytically . . .
Solve using finite di erences
Positions, velocities . . . approximated as a Taylor series:
1 2 1 1
r ( t + ∆ t) = r ( t) + ∆ tv ( t) + ∆ t a ( t) + ∆ t 3 b( t) + ∆ t 4 c ( t) + . . .
2 6 24
1 2 1 3
v ( t + ∆ t) = v ( t) + ∆ ta ( t) + ∆ t b( t) + ∆ t c ( t) + . . .
2 6
1 2
a ( t + ∆ t) = a( t) + ∆ tb( t) + ∆ t c ( t) + . . .
2

b( t + ∆ t) = b( t) + ∆ tc ( t) + . . .
The simplest method truncate the series after the
first-order term. This gives Euler’s method :
r ( t + ∆ t) = r ( t) + ∆ tv ( t)
v ( t + ∆ t) = v ( t) + ∆ ta ( t)

This is a first order method (and should not be used)

The Verlet algorithm (1968, Loop Verlet) :
1 2
r ( t + ∆ t) = r ( t) + ∆ tv ( t) + ∆ t a ( t)
2
1
r ( t − ∆ t) = r ( t) − ∆ tv ( t) + ∆ t 2 a ( t)
2

Add the equations, and rearrange :

r ( t + ∆ t) = 2r ( t) − r ( t − ∆ t) + ∆ t 2 a( t)
The Verlet algorithm is a second order algorithm
Drawbacks of the Verlet algorithm :
Velocities do not appear as they have been eliminated
by the addition. They must be estimated, for instance
by taking another integration step :
r ( t + ∆ t) − r ( t − ∆ t)
v ( t) =
2 ∆t

Velocities are needed, for instance, when computing

the kinetic energy of the system
It is not a self-starting algorithm
A small term ( ∆t 2 a( t) ) is added to the di erence of two
large terms (2 r ( t) − r ( t − ∆ t) )
There are several variations of the Verlet method :
The Leap-Frog Verlet method :
1
r (t + ∆ t ) = r ( t) + ∆ t v ( t + ∆ t)
2
1 1
v ( t + ∆ t) = v ( t − ∆ t) + ∆ t a ( t)
2 2

Disadvantage: Velocities and positions not available at

the same time

The Velocity Verlet method :

1 2
r(t + ∆ t ) = r ( t) + ∆ t v( t) + ∆ t a( t)
2
∆t
v (t + ∆ t ) = v ( t) +
2 [a ( t) + ]
a ( t + ∆ t)

Solved in two steps with a force evaluation in between

Yields the positions, velocities and accelerations at the
same time
Apart from numerical di erences, the three Verlet methods
result in the same trajectories
The Gear predictor-corrector integrator :
First new positions, velocities and accelerations are
predicted according to the Taylor expansions :

1 2 1 3 1
r p( t + ∆ t ) = r ( t) + ∆ t v ( t) + ∆t a( t) + ∆ t b( t) + ∆ t 4 c ( t)
2 6 24

p 1 2 1 3
v (t + ∆ t ) = v ( t) + ∆ t a ( t) + ∆t b( t) + ∆ t c ( t)
2 6

p 1 2
a (t + ∆ t ) = a ( t) + ∆ t b( t) + ∆ t c ( t)
2

b p( t + ∆ t ) = b( t) + ∆ t c ( t)

Then the forces are evaluated to give the true accelerations

a( t + ∆ t)
The Gear predictor-corrector integrator (cont) :
The di erence between the predicted and calculated
accelerations is computed :
∆ a( t + ∆ t) = a ( t + ∆ t) − a p ( t + ∆ t)

The correction step :

r c ( t + ∆ t) = r p( t + ∆ t) + c 0 ∆a( t + ∆ t)
v c ( t + ∆ t) = v p( t + ∆ t) + c 1 ∆ a( t + ∆ t)
a c ( t + ∆ t) = a p( t + ∆ t) + c 2 ∆a( t + ∆ t)
b c ( t + ∆ t) = b p( t + ∆ t) + c 3 ∆a( t + ∆ t)
c c ( t + ∆ t) = c p( t + ∆ t) + c 4 ∆a( t + ∆ t)

The values of the coefficients c 0 , c 1 , . . . depend on the

19 , c = 3 , c = 1, c = 1 , c = 1 )
order of the method ( c 0 = 120 1 4 2 3 2 4 12
Choosing the Timestep
A balance is needed in choosing the timestep. A too short timestep lead to
a very expensive solution of the equations of motion, and leads to a limited
coverage of phase space. A much too small timestep might lead to
round-o errors.
Too large of a timestep leads to instabilities

Example: Simulations of the 2D Lennard-Jones Argon System

-356.029
Energy (kJ/mol)

-356.03

-356.031

-356.032
0 1 2 3 4 5 6 7 8 9 10

Time (ps)
Total energy as a function of time “Exact” trajectory of one argon atom
 Velocity Verlet Integrator ∆ t = 1 fs (blue)
-356.029

Energy (kJ/mol)
-356.03

-356.031

-356.032
0 1 2 3 4 5 6 7 8 9 10

Time (ps)

Velocity Verlet Integrator ∆ t = 10 fs (blue)

-355.99
-356
Energy (kJ/mol)

-356.01
-356.02
-356.03
-356.04
-356.05
-356.06
-356.07
-356.08
-356.09
0 1 2 3 4 5 6 7 8 9 10
Time (ps)
Gear Predictor-Corrector Method ∆ t = 1 fs (blue)
-356.029

Energy (kJ/mol)
-356.03

-356.031

-356.032
0 1 2 3 4 5 6 7 8 9 10

Time (ps)

Gear Predictor-Corrector Method ∆ t = 10 fs (blue)

-347
-348
-349
Energy (kJ/mol)

-350
-351
-352
-353
-354
-355
-356
-357
0 1 2 3 4 5 6 7 8 9 10
Time (ps)
Comparison of Velocity Verlet (red) and the Gear
Predictor-Corrector (blue) at ∆ t = 1 fs :
-356.022
-356.023
The Gear method has
short time fluctuations,
Energy (kJ/mol)

-356.024
-356.025
-356.026 but it has a (substantial)
-356.027
-356.028
long-time drift
-356.029
-356.03 The Verlet method uses
-356.031
-356.032 a central di erence and
0 100 200 300 400 500 600 700 800 9001000
is thus time-reversible
Time (ps)

The trajectories will eventually always diverge from the “true”

trajectory
The energy (Hamiltonian) of the system should be conserved
Use a shorter timestep for :
Lighter particles
Higher temperatures
Sti er potentials