Molecular dynamics meets the physical world:

Thermostats and barostats

Justin Finnerty
j.finnerty@grs-sim.de, 2011
Scientific theories of physical world

•Energy is invariant

•Real world has quantum selection rules

•E = PV/nRT

•With constant energy we have PV = nRT

Molecular dynamics review
Phase-space and trajectories

• State of classical system (canonical): described by position and momenta of

all particles (notation q is position and p is momentum).

• Phase space point X = (q,p) gives 6N degrees of freedom

� �
• Extract property from ensemble �A� = A(q, p)P (q, p)dqdp

P (q, p) = Q−1 e−E(q,p)/kB T

� �
Q = e−E(q,p)/kB T dqdp

• Intractable to evaluate over phase space

• N = 100; q, p assume +1,-1 in each coordinate; 26N ≅ 4x10180 calc of A

and E
Molecular dynamics review
Properties of ensemble average

• MD ansatz. Probability function P(q,p) means phase space is “empty” except

around local energy minima. Means integrand is near-zero (if A does no go to
∞ with increasing energy) except around local energy minima

• Key is to pick “reasonable” starting geometry

• Role of MD “equilibration” phase is to generate low-energy (high-probability)

system - “equilibration” here does not refer to chemical “equilibrium” but to
relaxing the system to get energy minima, hopefully close to the chemical
“equilibrium”.

• If we start in region with high P(q,p) energy conserving evolution over time will
continue to sample high P(q,p) phase space (note these are unique and non-
intersecting trajectories)
Molecular dynamics review
Time average of trajectories

• Sampling periodically along trajectory evolving in time gives

M
1 �
�A� = A(ti )
M i
�
1 t0 +t
�A� = lim A(t)dt
t→∞ t t
0

• MD ansatz is starting from an initial low energy phase point (ie high P(q,p))
and moving along an iso-energetic time evolving trajectory will sample
properties of interest with high probability weights

�
• End of sampling uses property value convergence (�A� − A(ti ))2
Molecular dynamics review
pressure, temperature, volume

• Pressure (F is pairwise potential function):

  
N N
1  1 � � |pi | 
• P = Fij |qi − qj | +
V 3 i j>i
mi

• Temperature:
N
1 � |pi |2
• T =
3N kB i=1 mi

• Volume:

• parameters of the unit cell

Physically important conditions

• Basic MD is constant energy: micro-canonical nVE

• Constant temperature (volume and/or pressure vary) : nVT

• Biological reactions

• Constant pressure (volume and/or temperature vary) : nPT

• Chemical reaction occurring open to the atmosphere

Control of system variables to maintain
temperature or pressure

• Stochastic methods

• constrain a system variable to preset distribution function

• Strong-coupling methods

• scale system variable to give exact preset derived value

• Weak-coupling methods

• scale system variable in direction of desired derived value

• Extended system dynamics

• extend degrees of freedom to include temperature or pressure terms

Thermostat

• Stochastic Langevin thermostat apply friction and random force to momenta

�N
dpi
= (Fij |qi − qj |) − γpi + Ri (t)
dt j
• Andersen thermostat assign velocity of random particle to new velocity
from Maxwellian distribution

• Strong coupling isokinetic/Gaussian thermostat scale velocity

�N
dpi
= (Fij |qi − qj |) − αpi
dt j
• Weak coupling Berendsen bath unit-cell immersed in surrounding bath at T0
N
� � �
dpi p T0
= (Fij |qi − qj |) − i −1
dt j
τT T

• τ is relaxation constant
Nosé-Hoover thermostat

• Hoover, Phys. Rev. A 31 (1985) 1695; S. Nosé, J. Chem. Phys. 81 (1984) 511;
S. Nosé, Mol. Phys. 52 (1984) 255

• add extra term η to the equations of motion

N �
� N
dp pη
= Fij |qi − qj | − p
dt i j>i
Q
N
�
dpη |pi |2 3
= − nkB T
dt i
2mi 2

• The Q can be considered to be some fictional “heat bath mass”. Large Q

gives weak coupling, Nosé suggested Q ~ 6nkBT

• temperature is second-order differential

Example temperature control
6.5 Controlling the system 203
1.4

1.3
T (ε/kB)

1.2

1.1 LD
Berendsen
NoséHoover
1
0 1 2 3 4 5
*
t

Figure 6.12 The temperature response of a Lennard–Jones fluid under control of

three thermostats (solid line: Langevin; dotted line: weak-coupling; dashed line:
Nosé–Hoover) after a step change in the reference temperature (Hess, 2002a, and
by permission from van der Spoel et al., 2005.)

p2ηM −1
ṗηM = − kB T. (6.135)
Nosé-Hoover thermostat

• Energy (Hamiltonian H) of the simulated (physical) system fluctuates.

However, the energy of the system + heat bath (Hamiltonian H’) is conserved
Q 2
H� = H + pη + 3nkB T f (pη )
2

• If the system is ergodic the stationary state can then be shown to be a

canonical distribution
� � 2
��
p Q
exp −β V + + p2η
2m 2
Barostats

• Weak coupling Berendsen bath

• scale each dimension by μ: where β is isothermal compressibility - don’t

need to know exactly as it appears as ratio with τ (a relaxation constant)
  
� �1/3 N N
β∂t 1  1 � � |pi | 
µ= 1− (P0 − P ) P = F |q
ij i − q j | +
τ V 3 i j>i
mi

• This is similar to the Berendsen thermostat where we scale velocities by λ.

� � ��1/2
∂t T0
λ= 1+ −1
τ T
• Together give realistic fluctuations in T and P (J. Chem. Phys. 81 (1984) 3684)

• Note: pressure scale dimension, temperature scale velocity

Berendson barostat

Instantaneous pressure vs time and instantaneous temperature vs time in an

MD simulation of 256 particles at a density of 0.8442, using the Berendsen
barostat to impose an instantaneous pressure jump from 1.0 to 6.0. Each
curve corresponds to a different value of the rise time constant τ
Parrinello-Rahman Barostat

• The extended dimension of Nosé-Hoover thermostat can be applied to

pressure to give a Nosé-Hoover barostat

• The Parrinello-Rahman barostat (J. Appl. Phys. 52 (1981) 7182) extended this
further by making each unit vector of the unit-cell independent so that

• as with Nosé-Hoover the volume is a variable in the simulation.

• additionally allows dynamic shape change (gives control of stress as well

as pressure)

• The additional terms in the equations of motion are of similar type to that
shown for the Nosé-Hoover thermostat (though somewhat more complex!)
 does indic
might have been guessed by the title given to this subsection.
after quen
Figure 3 shows the details of the changes with the passage of
only the s
time. The MD cell, i.e., the h matrix, undergoes large and
but it also
swift changes which cannot possibly be described as elastic
shells sho
Parrinello-Rahman barostat deformations. In fact, as Fig. 3 shows, when equilibrium was
reached the average values of the components of h* were
cle positio
were pres
(h Tl) = 5.54 ± 0.03, (h!2) = 9.76 ± 0.06,
here that
the c axis
• Graph of MD simulation of a metal Havi
under a c
crystal after application of external .x TI = 20
stress. temperatu
was no st
veal the s
• Note second order decay of the ing the lo
three variables hii related to the quench te
unit-vectors ................. ,............................. "' .......

• (Temperature change is due to release

of elastic energy as crystal structure
changes) 0 •

ff
8

6
(a)
...................... ........... .......... ''' ............. " ... ...... .............

o 500 1000
Consequences of control method

method pro con

stochastic canonical disturb dynamics

non-Hamiltonian
strong coupling canonical in phase space disturb dynamic
accuracy

no well-defined ensemble ensemble average ok

weak coupling
good for non-equilibrium perturbation not ok

canonical in configuration more computation

extended dynamics
space disturb dynamics
Consequences of control method

method pro con

now considered less

Berendsen thermostat fast, smooth first-order
reliable for simulation at
and barostat approach to equilibrium
equilibrium

maintain canonical
ensemble.
Nosé-Hoover thermostat considered most reliable
slow, second-order
and Parinello-Rahman for simulation at
approach to equilibrium
barostat equilibrium and for
predicting thermo-
dynamic properties
Group discussion: MD calculation process

• What impacts will the different control methods have on parallel MD code?
(steps in right square not necessarily shown in actual order)

• Pressure and temperature are global properties calculated as ensemble averages

equilibrate solve equations of motion

pass start condition sample

experiment apply control

pass end condition propagate ensemble