Galli 072809

Ab-initio
MolecularDynamics
Giulia Galli
gagalli@ucdavis.edu
http://angstrom.ucdavis.edu
- What is Molecular Dynamics (MD)?
- How does one perform a MD simulation ?
- What can one compute by performing MD simulations?
- What is ab-initio MD?
Outline
Molecular dynamics (MD) : Newtonian equations of motion
and determination of trajectories as a function of time.
Integrate Newtons equations of motion for N atoms assuming ergodicity
) , , ( ) , , (
, , 1 ) , , ( ) (
1 1
1
N i N i
N i i i
E
N i t m
R R R R F
R R F R

=
= =
MD is a technique to obtain trajectories of a given system obeying classical
dynamics. Time averages of the systems properties can then be calculated.
MD is a deterministic method, i.e. the state of the system at any future time can
be predicted from its current state.
MD makes use of finite difference integration algorithms: at a given time, forces
on atoms are computed and combined with current positions and velocities to
generate new positions and velocities a `short` time step ahead. The force
acting on each atom is assumed to be nearly constant during this time step.
New atomic positions are then predicted, an updated set of forces is computed
and the dynamics cycle goes on.
Alder, B. J. and Wainwright, T. E. J. Chem. Phys. 27, 1208 (1957) and J. Chem. Phys. 31, 459 (1959)
Rahman, A. Phys. Rev. A136, 405 (1964); Stillinger, F. H. and Rahman, A. J. Chem. Phys. 60, 1545 (1974)
McCammon, J. A., Gelin, B. R., and Karplus, M. Nature (Lond.) 267, 585 (1977)
Molecular Dynamics (MD): the ergodic hypothesis
=
=
0
)) ( ), ( (
1
lim
t
N N
ave
dt t r t p A A
Under the ergodic hypothesis, one can identify temporal and
ensemble averages and thus compute thermodynamic
properties of a given system, once the interatomic potential is
specified.

= ) , ( ) , (
N N N N N N
r p r p A dr dp A
Molecular Dynamics: The Beginning
Study of microscopic nature of liquids and solids (32 atoms).
Hard-sphere potentials were used.
How does one perform a MD simulation ?
Define the model representing the real system
Calculations are then done on the model
Determine or compute on the fly the interaction potential
between the particles of the system
Choose number of particles, i.e. size of simulation cell
Choose algorithm to solve, numerically, Newton equations of
motion
Choose Statistical ensemble:
Micro-canonical (NVE)
Constant Temperature (NVT)
Constant Pressure and Temperature (NPT)
) , , ( ) , , (
, , 1 ) , , ( ) (
1 1
1
N i N i
N i i i
E
N i t m
R R R R F
R R F R

=
= =

= ) , ( ) , (
N N N N N N
r p r p A dr dp A
between the particle of the system
Choose algorithm to solve, numerically, Newton equations of
motion
) , , ( ) , , (
, , 1 ) , , ( ) (
1 1
1
N i N i
N i i i
E
N i t m
R R R R F
R R F R

=
= =

= ) , ( ) , (
N N N N N N
r p r p A dr dp A
Inter-atomic potential is always approximate
Validation is a key step of MD simulations
) , , ( ) , , (
, , 1 ) , , ( ) (
1 1
1
N i N i
N i i i
E
N i t m
R R R R F
R R F R

=
= =
Inter-atomic Potentials
Two routes are typically followed:
Empirical Potentials:
Choose a parametric form for the potential, incorporating some
physical effects deemed relevant, and adjust the parameters in a
phenomenological fashion, by fitting numerical predictions to known
experimental results (e.g. results of scattering experiments,
thermodynamic quantities, including specific heat, compressibility,
sound velocity measurements etc.)
Ab initio derived, quantum-mechanical potentials:
Choose a parametric form for the potential and fit the parameter to
numerical predictions obtained ab-initio for `simple` few atom
systems, by using, e.g. DFT or Quantum Chemistry methods.
Compute many-body inter-atomic potential on the fly at each set
of atomic positions, e.g. by using DFT.
See, e.g.: M. P.Allen and D. J. Tildesley, Computer Simulation of Liquids, Oxford University Press (1994)
) , , ( ) , , (
, , 1 ) , , ( ) (
1 1
1
N i N i
N i i i
E
N i t m
R R R R F
R R F R

=
= =

V(r
1
,r
2
,,r
N
) = v(r
i
,r
j
)
j= i+1
N
i=1
N1
Ab-initio Molec. Dyn.

Lennard-Jones (LJ) Potential

t =
m
2
4
One of the most important and popular potentials is the Lennard-Jones potential:

v(r
12
) = 4
r
12
|
\

|
.
|
12
r
12
|
\

|
.
|
6

`

)

Two adjustable parameters, (length)
and (energy)
Usually one takes as the length unit and 4 as the energy unit.
The mass unit is simply the mass of the particles; with only one component in
the system, the mass disappears (formally) from the calculation.
All other units are derivable from the above ones

V(r
1
,r
2
,,r
N
) = v(r
i
,r
j
)
j= i+1
N
i=1
N1
Example of popular empirical potentials for Si

Stillinger-Weber (1985)
basic empirical potential, initially for bulk structure, with pair-wise (sum of
two exponentials) and three-body terms (quadratic well in the angle with an
exponential cutoff in radius)
F. H. Stillinger and T. A. Weber, Phys. Rev. B, 31, 5262, 1985.
Tersoff (1988)
J.Tersoff, Phys.Rev.B 38, 9902 (1988)
EDIP (1997)
Environmentally Dependent Inter-atomic Potential
M. Z. Bazant and E. Kaxiras, Phys. Rev. Lett. 77, 4370 (1996).
M. Z. Bazant, E. Kaxiras, J. F. Justo, Phys. Rev. B 56, 8542 (1997).
J. F. Justo, M. Z. Bazant, E. Kaxiras, V. V. Bulatov, and S. Yip, Phys. Rev. B 58, 2539 (1998).
Choose a relatively simple, parametric form for the potential,
incorporating some physical effects deemed relevant, and adjust the
parameters in a phenomenological fashion, by fitting numerical
predictions to known experimental results (e.g. results of scattering
experiments, thermodynamic quantities, including specific heat,
compressibility, sound velocity measurements etc.)
) , , ( ) , , (
, , 1 ) , , ( ) (
1 1
1
N i N i
N i i i
E
N i t m
R R R R F
R R F R

=
= =

V(r
1
,r
2
,,r
N
) = v(r
i
,r
j
)
j= i+1
N
i=1
N1
Ab-initio Molec. Dyn.

Choose algorithm to solve Newton equation of motion
Avoid surface effects; representative cell must include all
relevant bonds and correlation.
When a molecule passes through one boundary another
molecule enters the opposite boundary
Use minimum image convention to account for neighbors
Scaling of algorithms to compute neighbor lists

Periodic Boundary Conditions and size of simulation cell
Choose initial Conditions
Assign initial ionic configuration
e.g., from x-ray crystallography or randomly chosen (usually not
recommended)
Assign initial velocities
For example: At thermal equilibrium, the expected value of the kinetic
energy of the system at temperature T is:
This can be obtained by assigning the velocity components v
i
from a
random Gaussian distribution with mean 0 and standard deviation
(k
B
T/m
i
):
T k N v m E
B
N
i
i i kin
) 3 (
2
1
2
1
3
1
2
= =
=
i
B
i
m
T k
v =
2
Finite difference integration: advances the system by a
small time step t during which forces are nearly constant
Taylor series to second order:
) (
) (
2
1 ) (
) ( ) (
) (
2
) (
) ( ) ( ) (
3 2
3 2
t O t
dt
t d
m
t
m
t
t t t
t O t
m
t
t t t t t
+ + + = +
+ + + = +
f f
v v
f
v r r
Computing df(t)/dt is something to avoid (too costly, from a computational
standpoint)
One thing we could do is to use the first equation to compute r(t+t), and then
compute v(t+t) from the new position as v(t+t)=[(r+t)-r(t)]/t
This is a complete disaster - catastrophic energy drifts and algorithm is not
time reversible.
Verlet algorithm [ a is acceleration: a = F/M]
R(t+t) = R(t) + v(t)t + (1/2)a(t)t
2
+ (1/6) a(t) t
3
+ O(t
4
)
R(t-t) = R(t) v(t)t + (1/2)a(t)t
2
- (1/6) a(t) t
3
+ O(t
4
)
Summing these two equations yields:
R(t+t) = 2R(t) R(t- t) + a(t)t
2
+ O(t
4
)
v(t+t) = v(t) + a(t)t + (1/2)a(t)t
2
+ O(t
3
)
a(t+t) = a(t) + a (t)t + O(t
2
)
Plugging a(t) into the v(t+t) expression yields:
v(t+t) = v(t) + (1/2)[a(t) + a(t+t)] t + O(t
3
)
Finite difference integration: an order (t)
4
algorithm
Finite difference integration: an order (t)
4
algorithm
Verlet algorithm [ a is acceleration: a = F/M]
R(t+t) = R(t) + v(t)t + (1/2)a(t)t
2
+ (1/6) a(t) t
3
+ O(t
4
)
R(t-t) = R(t) v(t)t + (1/2)a(t)t
2
- (1/6) a(t) t
3
+ O(t
4
)
Summing these two equations yields:
R(t+t) = 2R(t) R(t- t) + a(t)t
2
+ O(t
4
)
v(t) = [R(t+t) - R(t- t )]/t + O(t
2
)
Leap-frog algorithm (equivalent to Verlet algorithm)
R(t+t) = R(t) + v(t+(1/2)t) t
v(t+(1/2)t) = v(t-(1/2)t) + a(t) t
Beemans algorithm (higher order than Verlet algorithm)
R(t+t) = R(t) + v(t)t + (2/3)a(t)t
2
(1/6)a(t-t)t
2
v(t+t) = v(t) + v(t)t + (1/3)a(t)t + (5/6)a(t)t(1/6)a(t-t)t
Finite difference integration: other algorithms
Comparison of different integration schemes
It turns out that the Verlet algorithm is nearly always the best
choice.
It is very fast computationally
It requires little memory
Long-term energy conservation is excellent
It is time-reversible
Other, Verlet-type schemes, such as the Leap-Frog algorithm
are also widely used.
Higher order methods, such as the Predictor-Corrector scheme
are less common. One can use larger time steps, but these
algorithms suffer from being slower, requiring more storage, and
not being time reversible.
Choice of Time Step
No hard rules: if t is too small, not enough of phase space is explored, and if t is
too large, one will observe instabilities in the integration
Typical time-steps fall in the range of 0.1-10 femtoseconds (10 au = 0.24
femtoseconds)
For a liquid, the time step should be small compared to the mean time between
collisions.
Rule of thumb: for a flexible molecule, the time step should be no greater than 1/10
the time of the shortest period of motionHowever rule not always applicable; for
example: for water, ~ 1/40 must be used to ensure proper energy conservation.
Some examples
*
(t in seconds) :
Atoms: translation, t ~ 10
-14
Rigid Molecules: translation and rotation, t ~ 5x10
-15
Flexible Molecules: translation, rotation, torsion, t ~ 2x10
-15
Flexible Molecules & Bonds: translation, rotation, torsion, vibration, t < 10
-15
* From Molecular Modeling, 2nd edition, A. Leach, Prentice Hall
Simplest way to modify T:
`Correct` the velocities every step to give desired temperature.
Multiply by :
T
W
is the temperature that you want and T(t) is the kinetic
temperature at time t. Algorithm is simple, but crude and may
inhibit equilibration.
) ( / t T T
W
=
Simulating at constant T: simple rescaling
T k N v m E
B
N
i
i i kin
) 3 (
2
1
2
1
3
1
2
= =
=
Bath supplies or removes heat from the
system as appropriate
Exponentially scale the velocities at each
time step by a factor :
where determines how fast the bath
influences the system
Simulating at constant T: the Berendsen scheme
syst em
Heat bat h
Berendsen et al. Molecular dynamics with coupling to an external bath. J. Chem. Phys. 81:3684 (1984)
|
.
|
\
|
=
T
T t
bath
1 1
T: kinetic temperature
Simple velocity rescaling and Berendsens scheme
do not sample the canonical ensemble
Extended systems: Nos thermostat
An extra degree of freedom s is introduced, which scales the velocities and
time step
We then associate a kinetic and potential energy to s (f is the number of
degrees of freedom in the system)
It can be shown that the partition function of the coupled system is:
t s t s
= = , r v

2
2
1
. . , ln ) 1 ( s Q E K s kT f V
s W s

= + =
( )
c
Z kT E
kT
Q
f
Z / exp
2
) 1 (
1
2 / 1
|
.
|
\
|
+
=
Q measures coupling between reservoir and system. It should not

be too high (slow flow) or too low (oscillations).
Method allows one to sample the canonical ensemble
Nos Thermostat
for a given property A.
Note that the total momentum and total angular momentum
deviate from canonical values by O(1/N).
( )
c
r A
s
A , , p r
p
=
|
.
|
\
|
Example of application of Nos Thermostat
System of 108 argon atoms.
S. Nos Mol. Phys. 52, 255 (1984).
Effect of thermostat
Hoover Thermostat
Start with the Nos method and redefine the time variable
Eliminate s from equations of motion
Hoover thermostat samples a canonical ensemble
new old
sdt dt =
) ) ( ( , ,
W
s
T t T
Q
fk
Q
p
= = =

p f p
Stochastic Collisions
Influence the system temperature by reassigning the
velocity of a random particle (a collision ). An element of
Monte-Carlo.
The new velocity is from the Maxwell-Boltzmann distribution
corresponding to the desired T
W
.
Between collisions sample a micro-canonical ensemble. It
can be shown that overall the canonical ensemble is
sampled.
Collision frequency is important.
Can also reassign some or all particle velocities.
Simulating at constant P: Berendsen scheme
Couple the system to a pressure bath
Exponentially scale the volume of the simulation box at each time
step by a factor :
where : isothermal compressibility
P
: coupling constant
syst em
pressure bat h
Berendsen et al. Molecular dynamics with coupling to an external bath. J. Chem. Phys. 81:3684 (1984)
( )
bath
P
P P
t
1
|
.
|
\
|
+ =
=
N
i
i i kin
F x E
V
P
1
3
2
where
V : volume
x
i
: position of particle i
F
i
: force on particle i
Andersen Barostat
Variables are scaled.
The piston has its own kinetic and potential energies.
It can be shown that the time average of the trajectories
derived using the Andersen barostat equals the
isoenthalpic-isobaric ensemble average to O(N
-2
).
r r r r
=

3 / 1 3 / 1
, V V
V P PE V Q K
W V V
= = ,
2
1
2
Q measures coupling between reservoir and system. It should not

be too high (slow flow) or too low (oscillations).
Constant Temperature and Pressure (combining
Nose and Andersen formulations)
For example in the Hoover formulation:
W W
kT V P t P
V V
2
/ ) ) ( (
, 3 /
,

=
=
=
p p f p
Brief Summary of Constant T and/or P
The temperature can be fixed by:
scaling the velocities (partially or completely);
changing some or all of the velocities of the particles to a
randomly selected member of the Maxwell- Boltzmann
distribution of the desired T
coupling the system to a heat bath
Analogous methods exist to chose/maintain a constant pressure.
Combinations of methods can be used to simulate a system at
constant temperature and pressure.
Which Method to control T and P
Scaling is simple and easy and in the simplest case requires
no parameters. Convergence may be a problem; does not
sample canonical/isobaric/isobaric-isothermal ensemble.
Good for initial phase of equilibration.
Stochastic approach is more stable than scaling but leads to
a method which is no longer deterministic.
Extended system methods are more complex and require
choice of parameters. Nos-Hoover thermostats enable the
canonical ensemble to be sampled.
Dynamical properties may not be correct
Variable Cell Shape (Parrinello-Rahman method)
Let the shape (not only the volume) of unit cell change
Method is useful to study transition between crystal
structures
Dynamics is unrealistic and unphysical cell shapes may
be obtained in simulations of liquids
From S.de Gironcolis lecture
For each time step:
Compute the force on each atom:
Solve Newtons equation of motion for each atom, to get new
coordinates and velocities:
Store coordinates, velocities,
Analyze data, compute statistical averages
How do we perform a MD simulation ?
X
E
X E X F
= = ) ( ) (
) ( X F X M =

X: cartesian vector
of dimension 3N
M diagonal mass matrix
Use stepwise numerical integration
How long, that is: how many time steps?
As long as it takes for the quantity of interest not to change, as a
function of time (i.e. to be converged) within desired error bars
Statistical predictions
The trajectory of any system is sensitive to the initial conditions.
This means that two trajectories that are initially very close will diverge
exponentially as time progresses, if ergodicity is satisfied
If we consider the integration error as the source of an initial (small) difference,
then we always have a simulated trajectory divergent from the true one.
Is this a devastating blow to molecular dynamics?
Not really - were interested in making statistical predictions, not in knowing
precisely what happens at a given time.
Interpretation of results
Results are numbers, as opposed to elegant equations; thus,
one has to make sense out of a set of numbers, i. e. interpret
the results.
Results are obtained for finite systems only, whereas often
one is seeking estimates for infinite systems, to eliminate
surface effects; thus, one has to assess the sensitivity of the
results to the system size (finite-size scaling).
Even if simulation were free from approximations, the results
are usually affected by systematic and statistical errors,
which have to be estimated.
Sources of Error
Statistical errors
They affect numerical simulations, just like they affect experimental
measurements. This is because the experiment (or the simulation) necessarily
takes a finite time.
Assigning statistical errors is a rather delicate part of a simulation.
Just like in a real experiment, in computer simulations the measured
quantity is assigned a numerical value, which is in principle affected by
statistical and systematic errors.
Sources of Error
Just like in a real experiment, in computer simulations the measured
quantity is assigned a numerical value, which is in principle affected by
statistical and systematic errors.
Systematic errors
Experimentally, systematic errors are generally associated with lack of
calibration, or incomplete calibration
In a computer simulation, typical sources of systematic errors include:
The finite size of the simulation sample
The finite time-step in Molecular Dynamics simulations
Round-off errors due to the finite precision of the computer (any
numerical calculation suffers from this)
Error Bars
Error bars determine a range, centered around the estimated <A>, within
which the true <A> is expected to lie with a specified probability.
Simulation results are not meaningful without an associated error bar
Once a robust procedure is defined to calculate error bars reliably, then one can
empirically attempt to establish how many initial data points should be discarded, in
other words how long the transient is, before the system reaches equilibrium.
Typically, one wants to make sure that as little data as possible from the transient
region is included in the average, but at the same time one does not want to throw
away too much data.
What is usually done is to calculate the average < A > by discarding an increasing
number of initial data points until the average does not change, within error bars.
Review: Error Bars
Averages are computed as:

< A>=
1
n
A
i
i=1
n
Statistical error is then:

()
2
(A)
2
1
2
(n 1)
|
\

|
.
|

(A)
2
=
(A
i
< A>)
i
2
n(n 1)
where:
Actually, this is only true when two criteria are met:
the data is drawn from a Gaussian probability distribution
the data are uncorrelated
Re-blocking Transformation
Blocking or binning method is the simplest way to eliminate data correlation.
In this method, the data set is transformed, for example into one that is 1/2 the
size:
The statistical error is then recomputed, and the re-blocking is repeated.
2
(A
(n / 2)
)
2
(A
(n / 21)
)
2
(A
''
)
2
(A
'
)
2
(A)
2
One can derive the following result (*):

A
1
, A
2
, A
3
,, A
n
{ } A
1
'
, A
2
'
, A
3
'
,, A
n / 2
'
{ }
* see, for example, H. Flyvbjerg and H. G. Petersen, J. Chem. Phys. 91, 461
(1989).
For each time step:
Compute the force on each atom:
Solve Newtons 2
nd
law of motion for each atom, to get new
coordinates and velocities:
Store coordinates, velocities,
Analyze data, compute statistical averages
How do we perform a MD simulation ?
X
E
X E X F
= = ) ( ) (
) ( X F X M =

X: cartesian vector
of the system
M diagonal mass matrix
Use stepwise numerical integration
Compute system properties, analyze data
Static properties
Example: pair distribution function g(r)
Dynamical properties
Example: transport coeff., in particular auto-diffusion
coeff. D (liquids)
Minimize complex potential energy surfaces:
Simulated annealing
Pair Distribution Functions
The pair distribution function provides a measure of the probability of finding
any two particles at a given physical distance.
Typically, this is defined (in 3 dimensions) as:

g(r) = (N1)
p(r)
r
2
The pre-factor is defined in such a way so that:

r
2
g(r)
dr
r
2
dr
=1
1
N
Examples:
Al
2
O
3
Pair Distribution Functions: error bars
J.Grossman et al. JCP 2004
Comparison with experiment is indirect: diffraction expt.
yield structure factors not partial correlation functions
Probe system with neutron or x-ray: the
scattering cross section (or intensity) is
given in terms of a weighted average
structure factor:

Q=k1k0

d
d
(Q)= b
2
S(Q)+ b
2
b
2
where

S(Q)= Wij
ij
Sij (Q)
The weighting factors are (Faber-Ziman definition) :
for neutrons:

W
ij
N
=
cic jbi bj
b
2
; for x-rays:

W
ij
X
=
cic j fi (Q)fj (Q)
f(Q)
2
Concentrations
Relation between structure factors and real space
correlation functions

Sij (Q)=1+
4
Q
[gij (r)1
0
]sinQr.rdr
Direct physical interpretation:
4r
2
c
j
g
ij
(r )r = average number of j atoms in a
range (r, r + r ) from an i atom at the origin
For i, j = 1, . n there are n(n+1)/2 partial pair
distribution functions

The partial structure factors are related to partial pair distribution functions
through the Fourier transform
Examples of structure factors
0 2 4 6 8 10
-1.0
-0.5
0.0
0.5
1.0
[Expt.: Soper,Bruni&Ricci, J. Chem. Phys. 106, 247 (1997)]
D
2
O by neutrons
S
(
Q
)
-
1
Q (A
-1
)
0 2 4 6 8 10
-0.2
-0.1
0.0
0.1
0.2
[Expt.: Thiessen&Narten, J. Chem. Phys. 77, 2656 (1982)]
DHO by neutrons
0 2 4 6 8 10
-1.0
-0.5
0.0
0.5
H
2
O by X-rays
[Expt.: Narten&Levy, J. Chem. Phys. 55, 2263 (1971)]
Water
0 5 10 15 20
-1.0
-0.5
0.0
0.5
1.0
Amorphous Si
Solid line: vapour deposited
Dashed line: Si-implanted
S
(
Q
)
-
1
Q(A
-1
)
Transport Properties
Differential Equations
Constitutive Equations
0
) , (
= +
j
t
t r c
0
) , (
= +
q
t
t r T
c
p
0
) , (
= +
v
Dt
t r D
c D = j T k q =
) (
x y xy
v v =
Mass Energy Momentum
Ficks Law Fouriers Law Newtons Law
Small gradients
Transport Properties: self-diffusion
Consider self-diffusion in a pure substance
Consider mass current when initial configurations are given as
Dirac delta function
Combine mass balance eqn. with Ficks Law
0 ) , (
) , (
2
=
t c D
t
t c
r
r
) ( ) , ( r r = t c
)
2
exp( ) 2 ( ) , (
2
/ 2
Dt
r
Dt t c
d
=
r
Dimensionality
of given system
>= < dr r t c t r
2 2
) , ( ) ( r
1 ) , ( =
dr t c r
0 ) , (
) , (
2
=
t c D
t
t c
r
r
dD
t
t r
2
) (
2
=
> <
dDt t r
N
t r
i
2 )) ( (
1
) (
2 2
= >= <
dDt t r
N
t r
i
2 )) ( (
1
) (
2 2
= >= <
2
r
t
Slope here gives D
An alternative formulation using velocity instead of particle position
=
t
i
d t
0
) ( ) ( v r
1
1
2
1 1 2 2
0 0
1 2 2 1
0 0
1 2 2 1
0 0
1 2 1 2
0 0
1
0 0
0
( ) ( ) ( )
( ) ( )
2 ( ) ( )
2 (0) ( )
2 (0) ( )
2 2 (0) ( )
t t
t t
t
t
t t
t
r t d d
d d
d d
d d
d d
dDt t d

=
=
=
=
=
=

v v
v v
v v
v v
v v
v v
Diffusion coefficient: convergence as a function of time
J.Grossman et al. JCP 2004
Transport coefficients as time integrals of
autocorrelation functions
Zero-shear viscosity
Thermal Conductivity
> < =
0
) ( ) 0 (
1
v v d
d
D
> < =
0
) ( ) 0 (
1

xy xy
d
VkT
> < =
0
2
) ( ) 0 (
1
q q d
VkT
T
(
+ =
i j i
ij i i
r u v m
dT
d
q ) (
2
1
2
(
+ =
i j i
ij y ij
y
i
x
i i xy
r f x v v m ) (
2
1
=
t
i
d t
0
) ( ) ( v r
For a pair-wise potential
MD as a tool for minimization of complex potential
energy surfaces
Energy
Energy minimization
stops at local minima
Molecular dynamics
uses thermal energy
to explore the energy
surface
State A
State B
Choose a relatively simple, parametric form for the potential,
incorporating some physical effects deemed relevant, and adjust the
parameters in a phenomenological fashion, by fitting numerical
predictions to known experimental results (e.g. results of scattering
experiments, thermodynamic quantities, including specific heat,
compressibility, sound velocity measurements etc.)
) , , ( ) , , (
, , 1 ) , , ( ) (
1 1
1
N i N i
N i i i
E
N i t m
R R R R F
R R F R

=
= =

V(r
1
,r
2
,,r
N
) = v(r
i
,r
j
)
j= i+1
N
i=1
N1
Energy and forces from DFT

DFT with local and semi-local density approximations is a
`practical` theory to compute static and dynamical properties of
molecules and solids
We can use forces derived from DFT to perform Molecular
Dynamics (MD) simulations
Born-Oppenheimer approximation
Hellman-Feynman theorem
From Shobbana Narasimhans lecture
Ab-initio Molecular Dynamics
t
from Density Functional Theory
At each MD step: solve Kohn-Sham equations set of
N coupled, non linear partial differential equations self-
consistently, using iterative algorithms, subject to
orthonormality constraints. N = # of electrons
solid
liquid
Ab-initio Molecular Dynamics
t
from Density Functional Theory
Fe
C
At each MD step: solve Kohn-Sham equations set of
N coupled, non linear partial differential equations self-
consistently, using iterative algorithms, subject to
orthonormality constraints. N = # of electrons
=
=
+
+ =
= = +
=
ij j
N
i
i
i i i i
d
V d V V
N i V
i

r r r
r r
r r r
r r
r
r r
r
) ( ) (
) ( ) (
)) ( ), ( (
) (
) ( ) , (
1 ) , (
el
1
2
XC ion
el
Complexity of ab-initio MD
Solve set of N coupled, non linear partial differential
equations self-consistently, using iterative algorithms,
subject to orthonormality constraints. N = # of
electrons
The cost of solving the Kohn-Sham
equations is eventually dominated by
orthogonalization (O(N
3
))
Most straightforward way to carry out ab-initio MD:
Born-Oppenheimer Dynamics
Time
E
n
e
r
g
y
In Born-Oppenheimer (BO)
dynamics, at each point on
the trajectory:
= min[E
KS
]
M
I
R
I
= -
I
..
Tomorrow, Prof. Gygis lecture:
How to implement BO dynamics
Car-Parrinello method to perform ab-initio MD
Large scale calculations
Ab initio Molecular Dynamics
Simulating matter on the atomic scale.
Compute trajectories of individual atoms.
Use forces derived from Quantum Mechanics.
Use only fundamental physical constants as input.
Classical vs ab initio molecular dynamics
Classical MD
At each MD step:
Compute forces from a given
model energy function
Update atomic positions
Ab initio MD
At each MD step:
Solve the electronic structure
problem
Derive forces from electronic
wavefunctions
Update atomic positions
Millions of atoms
Nanoseconds
Hundreds of atoms
(*)
Picoseconds
(*) Remember: the key variable is number of electrons !
Ab-initio fitted potentials
Ab-initio derived force fields
(*)
: qualitative agreement with
experiment
(*) Robert Bukowski, Krzysztof Szalewicz, Gerrit C. Groenenboom, Ad van der Avoird, Science 315, 1249 (2007)
?
Some results from ab-initio MD
Phase diagrams
Liquids on surfaces
Nanostructures
Understanding phase diagrams of elements
and simple compounds
Diamond: difficult to tame, does it melt under P ?
Hydrogen: a liquid ground state at zero T and high P?
Oxygen at high P: A spin glass of oxygen molecules ?
Boron: is there a crystalline ground state?
Water: Interplay between dissociation and melting?
E.Schwegler, M.Sharma, F.Gygi and G.G., PNAS 2008; A.Correa, S.Bonev and G.G. PNAS 2006; S.Bonev, E.Schwegler,
T.Ogitsu and G.G. Nature 2004; B.Militzer, F.Gygi and G.G. PRL 2003; T.Ogitsu, E.Schwegler, F.Gygi and G.G,. PRL 2003;
T.Ogitsu, F.Gygi, J.Reed, Y.Motome, E.Schwegler and G.G. JACS 2009 and submitted.
Understanding phase diagrams of elements and
simple compounds : Predictions
S.Bonev, E.Schwegler, T.Ogitsu and G.G, Nature 2004
Experiment confirming
existence of maximum in
T(P): Silveiras group,
Harvard, 2008
A. Correa, S.Bonev, and G.G, PNAS 2006
Experiment
(CEA and LLNL)
confirming HP
melting, 2007
T.Ogitsu, F.Gygi, J.Reed, Y.Motome, E.Schwegler and G.G. JACS 2009 and submitted
Understanding phase diagrams of elements
and simple compounds
T.Ogitsu, F.Gygi, J.Reed, E.Schwegler and G.G. JACS 2009 and submitted.
Large scale two-phase ab-initio simulations
Type II Superionic
conductor
Liquid
Ice VII
Dyn. Translat.
Disordered
Ice VII
An example for which MD is not good enough
E.Schwegler, M.Sharma, F.Gygi and G.G., PNAS 2008 (in press); A.Correa, S.Bonev and G.G. PNAS 2006; S.Bonev,
E.Schwegler, T.Ogitsu and G.G. Nature 2004; B.Militzer, F.Gygi and G.G. PRL 2003; T.Ogitsu, E.Schwegler, F.Gygi and
G.G,. PRL 2003; T.Ogitsu, F.Gygi, J.Reed, E.Schwegler and G.G. JACS 2009 and submitted.
Cluster-expanded Ising Hamiltonian with ab-initio derived parameters, and
lattice MC simulations:
-boron has a frustrated, disordered
structure with residual entropy
the only element in nature
analogous to the ice and
spine ice compounds
Hydrophilic
surface SiC
Hydrophobic surfaces graphite, CNT
and hydrogenated diamond
Hydrophobic
solutes (benzene)
Solvation of ions
and Si clusters
Water at interfaces
Water in different environments: from small ion and Si quantum dot solvation to
water on surfaces in extended and confined media.
We address issues regarding structural, electronic, vibrational and some
dynamical properties of interfacial water, using ab-initio simulations.
D.Prendergast et al., JACS 2004; D.Prendergast and G.G. PRL 2006; G.Cicero, J.Grossman, A.Catellani and G.G., JACS 2005;
M.Allesch, E.Schwegler and G.G., JPC-B 2007 and JCP 2008; G.G. PNAS 2007; G.Cicero, J.Grossman, E.Schwegler, F.Gygi and G.G.
JACS 2008 ; M.Sharma, D.Donadio, E.Schwegler and G.G. Nanolett. 2008, D.Donadio, G.Cicero, E.Schwegler, M.Sharma and G.G.,
JPC-B 2009; P.Huang, E.Schwegler and GG, JPC-C 2009.
Hydrophobic surfaces graphite, CNT and
hydrogenated diamond
Water at interfaces
We are studying water in different environments: from small ion and Si quantum
dot solvation to water on surfaces in extended and confined media.
G.Cicero, J.Grossman, E.Schwegler, F.Gygi and G.G. JACS 2008 ; M.Sharma, D.Donadio, E.Schwegler and G.G.
Nanolett. 2008, E.Schwegler, M.Sharma and G.G., JPC-B 2009; P.Huang, E.Schwegler and GG, JPC-C 2009.
Structure: Interplay
between confinement and
interfacial effects?
Is a hydropobic surface a
hard wall?
Is there vapor at an
hydrophobic interface?
How to compare with
experiment?
Non polar and hydrophobic surfaces
Deuterated
Diamond Graphite CNT
J.-Y.Raty, F.Gygi and G.Galli PRL 2005
Simulations of single wall nanotubes on 1-nm Fe
nanoparticles
The tube grows
capped on an iron
nanoparticle
catalyst.
H.Dai 2001
Carbon does not mix
with iron: growth does
not require the presence
of a supersaturated
solution.
`Local` interaction
between C and catalyst
surface may be tuned to
influence growth.
Still out of reach of ab-initio MD
For example: Nucleation processes
T.Li, D.Donaadio, L.Ghiringhelli and G.Galli Nat.Mat. 2009
Ab-initio MD
Time
E
n
e
r
g
y
In Born-Oppenheimer (BO)
dynamics, at each point on
the trajectory:
= min[E
KS
]
M
I
R
I
= -
I
..
Tomorrow, Prof. Gygis lecture:
How to implement BO dynamics
Car-Parrinello method to perform ab-initio MD
Large scale calculations

Galli 072809

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Galli 072809

Hochgeladen von

Copyright:

Verfügbare Formate

Ab-initio

Ab-initio Molec. Dyn.

Example of popular empirical potentials for Si

Ab-initio Molec. Dyn.

Q measures coupling between reservoir and system. It should not

Q measures coupling between reservoir and system. It should not

Statistical error is then:

Energy and forces from DFT

Das könnte Ihnen auch gefallen