Modeling Thermal Transport and Viscosity

with Molecular Dynamics

Steve Plimpton
Sandia National Labs
[email protected]

LAMMPS Users and Developers Workshop

International Centre for Theoretical Physics (ICTP)
March 2014 - Trieste, Italy

Presentation: SAND2014-2225C
Tackling a new problem with LAMMPS

Features: lammps.sandia.gov/features.html
Commands page: doc/Section Commands.html
Papers: lammps.sandia.gov/papers.html
Mail list: lammps.sandia.gov/mail.html
Adding new features: doc/Section modify.html
Howto explanations in manual:
doc/Section howto.html
6.20 Calculating thermal conductivity
6.21 Calculating viscosity
What is thermal conductivity?

Propensity of a material to transmit heat (thermal energy)

Solids or liquids or gases
Temperature and density dependent
High = good heat sink, low = good insulator
What is thermal conductivity?

Propensity of a material to transmit heat (thermal energy)

Solids or liquids or gases
Temperature and density dependent
High = good heat sink, low = good insulator
Fundamental equation:

J = T

KE
J = heat flux =
Area time

T = temperature gradient = dT /dz

= thermal conductivity = W / m K
What is viscosity?

Propensity of a fluid to transmit momentum perpendicular to

direction of momentum flow (shear direction)
Fluid friction or resistance to flow
Fluid = Liquids and gases
High = honey, low = water
What is viscosity?

Propensity of a fluid to transmit momentum perpendicular to

direction of momentum flow (shear direction)
Fluid friction or resistance to flow
Fluid = Liquids and gases
High = honey, low = water
Fundamental equation:

Vx
Jz (px ) =
z

Jz (px ) = momentum flux in perpendicular direction

Vx
= transverse velocity gradient
z

= shear viscosity
4 methods for computing thermal conductivity

Non-equilibrium methods:
basic idea: induce a temperature gradient or heat flux and
monitor the other quantity
direct thermostatting method of Ikeshoji and Hafskjold
reverse perturbation method of Muller-Plathe
aggregate variant of Muller-Plathe method
Equilibrium method:
Green-Kubo formalism
4 methods for computing thermal conductivity

Non-equilibrium methods:
basic idea: induce a temperature gradient or heat flux and
monitor the other quantity
direct thermostatting method of Ikeshoji and Hafskjold
reverse perturbation method of Muller-Plathe
aggregate variant of Muller-Plathe method
Equilibrium method:
Green-Kubo formalism

See examples/KAPPA for 4 sample scripts

3d LJ fluid, but adaptable to other systems (e.g. solids)
4 methods for computing shear viscosity

Non-equilibrium methods:
basic idea: induce a flow gradient or momentum flux and
monitor the other quantity
drag wall over fluid to induce shear
NEMD shear deformation with SLLOD thermostatting
Muller-Plathe reverse perturbation method
Equilibrium method:
Green-Kubo formalism
auto-correlation of pressure tensor component
4 methods for computing shear viscosity

Non-equilibrium methods:
basic idea: induce a flow gradient or momentum flux and
monitor the other quantity
drag wall over fluid to induce shear
NEMD shear deformation with SLLOD thermostatting
Muller-Plathe reverse perturbation method
Equilibrium method:
Green-Kubo formalism
auto-correlation of pressure tensor component

See examples/VISCOSITY for 4 sample scripts

2d LJ fluid, but adaptable to other systems
Caveats for atomistic MD

1 Missing electronic effects for

empirical atomistic simulations
heat is transported by phonons
electronic effects included only indirectly in potential
if electrons make large contribution to , wont see it
2 Homogeneous vs heterogeneous systems
formulas are for homogeneous bulk
in graphene sheets is 2d, possibly asymmetric
for fluid flowing thru CNTs is radial BC
3 Mis-match to experiment
MD has severe length- and time-scale constraints
temperature gradients & shear rates are typically
orders of magnitude larger than expt
(1) Direct thermostatting method

Ikeshoji and Hafskjold, Molecular Physics, 81, 251-261 (1994)

2 thermostats for 2 regions of simulation box
One hot, one cold
Monitor flux of energy needed to maintain T
Direct thermostatting method

LAMMPS implementation:
fix langevin using compute temp/region as bias
fix langevin can tally energy each thermostat adds/subtracts
fix ave/spatial monitors resulting temperature gradient

Q z
=
2Areat T
Script for direct thermostatting method

lattice fcc ${rho}

region box block 0 $x 0 $y 0 $z

# heat layers

region hot block INF INF INF INF 0 1

region cold block INF INF INF INF 10 11
compute Thot all temp/region hot
compute Tcold all temp/region cold

# 1st equilibration run

fix 1 all nvt temp $t $t 0.5

run 1000
unfix 1
More script for direct thermostatting method

# thermal conductivity calculation

compute ke all ke/atom

variable temp atom c ke/1.5

fix hot all langevin ${thi} ${thi} 1.0 59804 tally yes
fix cold all langevin ${tlo} ${tlo} 1.0 287859 ...
fix modify hot temp Thot
fix modify cold temp Tcold

fix 2 all ave/spatial 10 100 1000 z lower 0.05 v temp &

file tmp.profile units reduced

thermo style custom step temp c Thot c Tcold f hot f cold

run 20000
Output for direct thermostatting method

Step Temp Thot Tcold hot cold

...
30000 1.3011151 1.7275961 1.06067 -0.84589474 0.8965726
31000 1.3002026 1.5313418 1.0526131 -0.8964083 0.93984929
Loop time of 25.7381 on 8 procs for 20000 steps with 8000 atoms
(2) Muller-Plathe reverse perturbation method

Muller-Plathe, J Chem Phys, 106, 6082 (1997)

Define hot and cold regions of simulation box
Find hottest atom in cold region, coldest atom in hot region
Swap velocity vector of these 2 atoms (energy)
Tally heat flux due to KE exchanges
Monitor the induced temperature profile
Reverse of previous method
Muller-Plathe reverse perturbation method

LAMMPS implementation:
fix thermal/conductivity swaps KE and tallies heat flux
fix ave/spatial monitors induced temperature gradient

Q z
=
2Areat T
Script for Muller-Plathe reverse method

# thermal conductivity calculation

compute ke all ke/atom

variable temp atom c ke/1.5

fix 1 all nve

fix 2 all ave/spatial 10 100 1000 z lower 0.05 &
v temp file tmp.profile units reduced
fix 3 all thermal/conductivity 10 z 20

variable tdiff equal f 2[11][3]-f 2[1][3]

thermo style custom step temp epair etotal &
f 3 v tdiff

run 20000
Output for Muller-Plathe reverse method

Step Temp E pair TotEng 3 tdiff

...
40000 1.4071151 -3.8068479 -1.6964391 14307.339 1.1772366
41000 1.4126121 -3.8153948 -1.6967416 15087.11 1.1408062
Loop time of 23.9599 on 8 procs for 20000 steps with 8000 atoms
(3) Variant of Muller-Plathe reverse perturbation method

Define hot and cold regions of simulation box

Add/subtract energy continuously to all atoms in these regions
Equal and opposite heat flux
Monitor the induced temperature profile
Variant of Muller-Plathe method

LAMMPS implementation:
fix heat adds/subtracts KE in a region
fix ave/spatial monitors induced temperature gradient

Q z
=
2Areat T
Script for variant of Muller-Plathe method

# thermal conductivity calculation

fix hot all heat 1 100.0 region hot

fix cold all heat 1 -100.0 region cold

compute ke all ke/atom

variable temp atom c ke/1.5

fix 2 all ave/spatial 10 100 1000 z lower 0.05 &

v temp file tmp.heat.profile units reduced

variable tdiff equal f 2[11][3]-f 2[1][3]

run 20000
Output for variant of Muller-Plathe method

Step Temp Thot Tcold tdiff

...
30000 1.382101 1.9337034 1.0679145 -0.79821576
31000 1.3779178 1.8832819 1.0837774 -0.80611097
Loop time of 24.3193 on 8 procs for 20000 steps with 8000 atoms
(4) Green-Kubo equilibrium method

Relate ensemble average of auto-correlation of J to

Equilibrium J computable from per-atom KE, PE, virial

Z Z
V V
= hJx (0)Jx (t)i dt = hJ(0) J(t)i dt
kB T 2 0 3kB T 2 0

" #
1 X X
J = ei vi Si vi
V
i i

1 X X
= ei vi + (fij vj ) xij
V
i i<j

1 X 1 X
= ei vi + (fij (vi + vj )) xij
V 2
i i<j
Green-Kubo method

LAMMPS implementation:
compute heat/flux calculates J tensor
fix ave/correlate performs auto-correlation
variable trap() function performs time integration
Script for Green-Kubo method

compute myKE all ke/atom

compute myPE all pe/atom
compute myStress all stress/atom virial
compute flux all heat/flux myKE myPE myStress

fix JJ all ave/correlate $s $p $d &

c flux[1] c flux[2] c flux[3] type auto &
file tmp.heatflux ave running

variable k11 equal trap(f JJ[3])*${scale}

variable k22 equal trap(f JJ[4])*${scale}
variable k33 equal trap(f JJ[5])*${scale}

run 100000
Output for Green-Kubo method

Step Temp k11 k22 k33

...
98000 1.3477904 3.2534428 2.8638625 3.8437754
100000 1.3583776 3.3351133 2.859474 3.7715301
Loop time of 52.1737 on 8 procs for 100000 steps with 4000 atoms

variable kappa equal (v k11+v k22+v k33)/3.0

print "thermal conductivity: ${kappa}"
Comparing the 4 methods for thermal conductivity

Liquid Argon at state point: * = 0.6, T * = 1.35, Rc = 2.5

D Evans, Phys Rev A, 34, 1449 (1986)

Method
Direct thermostat 3.41
Muller-Plathe 3.45
M-P with fix heat 3.39
Green-Kubo 3.78
Evans paper 3.3
Experiment agrees with Evans
Comparing the 4 methods for thermal conductivity

Liquid Argon at state point: * = 0.6, T * = 1.35, Rc = 2.5

D Evans, Phys Rev A, 34, 1449 (1986)

Method
Direct thermostat 3.41
Muller-Plathe 3.45
M-P with fix heat 3.39
Green-Kubo 3.78
Evans paper 3.3
Experiment agrees with Evans

Small systems have boundary effects

Need to monitor equilibration and statistical noise
Factors of 2 are easy to miss!
(1) Shearing via moving wall

LAMMPS methodology:
Rigid, moving wall
Fix addforce can apply load
if desired
Important to thermostat flow
since adding energy
fix langevin on non-sheared
dimensions
compute temp/profile to
subtract flow profile
Monitor Pxz and velocity
profile of flow
(2) Shearing via deforming box

LAMMPS methodology:
Fix deform for box deformation
Important to thermostat flow
since adding energy
fix nvt/sllod for SLLOD
equations of motion
Evans and Morriss,
Phys Rev A, 30, 1528 (1984)
Monitor Pxz and velocity
profile of flow
insure flow profile agrees
with box deformation
(3) Muller-Plathe reverse perturbation method

Muller-Plathe,
Phys Rev E, 59, 4894 (1999)
Define two slabs within
simulation box
J (p ) p v
Find max Vx in one region, z x
max Vx in other region
Fix viscosity swaps momenta of
these 2 atoms (or molecules)
Tally momentum flux due to
exchanges
v
Monitor the induced velocity
profile
Reverse of previous methods
(4) Green-Kubo equilibrium method

Relate ensemble average of auto-correlation of Pxz to

Z
V
= hPxz (0)Pxz (t)i dt
kB T 0

Pxz computable from virial

Fix ave/correlate performs auto-correlation
Variable trap() function performs time integration
Comparing the 4 methods for viscosity

LJ at state point: * = 0.6, T * = 1.0, Rc = 2.5

Woodcock, AIChE Journal, 52, 438 (2006)

Method
Moving wall 0.946
Deforming box 1.18
Muller-Plathe 0.997
Green-Kubo 1.07
literature value 1.0
Comparing the 4 methods for viscosity

LJ at state point: * = 0.6, T * = 1.0, Rc = 2.5

Woodcock, AIChE Journal, 52, 438 (2006)

Method
Moving wall 0.946
Deforming box 1.18
Muller-Plathe 0.997
Green-Kubo 1.07
literature value 1.0

Small systems have boundary effects

Need to monitor equilibration and statistical noise
Factors of 2 are easy to miss!
Shear viscosity for rigid-bodies in SRD fluid
Shear viscosity for aspherical bodies in SRD fluid
Shear viscosity for aspherical bodies in SRD fluid

Any of these examples could use short-chain polymer solvents

Trade-offs between methods

NEMD methods pros:

intuitive to understand
quick to converge
NEMD methods cons:
unphysically large temperature gradients and heat fluxes
bigger systems to allow for gradient
Trade-offs between methods

NEMD methods pros:

intuitive to understand
quick to converge
NEMD methods cons:
unphysically large temperature gradients and heat fluxes
bigger systems to allow for gradient

Green-Kubo method pros:

equilibrium simulation
can use smaller system
Green-Kubo method cons:
slow to converge
hard to tell when correlation integral has converged
Hands-on exercise #1

Focus on viscosity (or thermal conductivity) (or both!)

viscosity simulations are more visual to animate
Study scripts in examples/VISCOSITY (or examples/KAPPA)
4 scripts, for each of 4 methods
understand what each command and parameter represents
Figure out how to analyze output to get (or )
Hands-on exercise #1

Focus on viscosity (or thermal conductivity) (or both!)

viscosity simulations are more visual to animate
Study scripts in examples/VISCOSITY (or examples/KAPPA)
4 scripts, for each of 4 methods
understand what each command and parameter represents
Figure out how to analyze output to get (or )

Reproduce 4 values in examples/VISCOSITY/README

Do scripts run faster in parallel?
Do they produce the same answers in parallel?
Hands-on exercise #2

Change parameters in input scripts:

size of system, density, temperature
shear rate, cutoff of potential
IMPORTANT - When you change script and do a new run:
visualize to insure system dynamics are normal
monitor velocity (or temperature) profile
check convergence of G-K integrations
are you running long enough?
Otherwise your or values may be bogus
Hands-on exercise #2

Change parameters in input scripts:

size of system, density, temperature
shear rate, cutoff of potential
IMPORTANT - When you change script and do a new run:
visualize to insure system dynamics are normal
monitor velocity (or temperature) profile
check convergence of G-K integrations
are you running long enough?
Otherwise your or values may be bogus
Choose one larger/smaller value of a parameter
how much larger or smaller?
Does (or ) change with that parameter?
Do all methods still agree?
Does variation make physical sense?
Hands-on exercise #3

Make a plot as vary a parameter over a wide range

size of system, density, temperature
shear rate, cutoff of potential
E.g. versus shear-rate for shear-thinning
What other parameters should remain constant?
e.g. temperature, pressure
How much can parameter vary before dynamics break down?
e.g. liquid crystallizes at too high a density
Hands-on exercise #3

Make a plot as vary a parameter over a wide range

size of system, density, temperature
shear rate, cutoff of potential
E.g. versus shear-rate for shear-thinning
What other parameters should remain constant?
e.g. temperature, pressure
How much can parameter vary before dynamics break down?
e.g. liquid crystallizes at too high a density

Bonus: modify script to run series of simulations

as parameter varies
see Section howto.html 6.4 and variable command
Bonus: run/viz M-P viscosity scripts in
examples/ASPHERICAL