Dear users,
Is there any restriction on the ‘SEED’ value(integer) of ‘velocity’ command in LAMMPS?
Because just only changing the seed value (keeping 6 digit integer) when I calculate Thermal Conductivity of Si at room temperature it varies from 200-550 W/mK.
Please suggest me If anybody have any idea.
Amal
Hard to tell what your problem is from just this info. Certainly not
the SEED issue but something else.
Which method are you using to calculate the conductivity? Equilibrium
methods (EMD) using the GK formalism (as the example provided with the
lammps distribution) take longer to reach convergence. If you are not
waiting long enough that could be the reason why different SEEDs show
different final outcomes. Non-equilibrium methods (NEMD) will be, in
general, faster. It is easy to implement NEMD in lammps by extending
the compute/heat flux to get the flux in a certain spatial region.
Combined with a temperature gradient created using two thermostated
regions this could help you troubleshoot your current approach.
However, keep in mind that you will always have statistical deviations
(your 150% error bars feel way too large). I imagine they cannot be
very large for such a simple system as crystalline Si. Look in the
lit.
Carlos
Thank you for your reply Dr.Carlos.
I am using EMD and from auto-correlation function it is seen that it takes around 50 ps for convergence whereas the correlation I have chosen is 250ps. so there should not be any question as per as relaxation time is concerned.
Dear users,
Is there any restriction on the 'SEED' value(integer) of 'velocity' command
in LAMMPS?
no. why should that matter?
Because just only changing the seed value (keeping 6 digit integer) when I
calculate Thermal Conductivity of Si at room temperature it varies from
200-550 W/mK.
Please suggest me If anybody have any idea.
have you checked how well converged each of
these calculations are? what is the statistical
error estimate on them?
even though you may get a number with very
many digits as a result, it doesn't mean that
all of these digits are valid. people often underestimate
how badly MD simulations converge to results
and sometimes it may depend very much on
what you are looking at or the size of your system.
i've looked into this many, many years back.
http://klein-group.icms.temple.edu/akohlmey/files/talk-trieste2004-water.pdf
in some cases (MSD) results may "look" converged,
but they may be subject to low frequency fluctuations
which in turn may depend on the starting conditions.
axel.
If you say so. Guess you need to dig deeper into your problem then. If
you were to consider my suggestion about implementing and using NEMD
for troubleshooting, please notice that this is not the
reverse-non-equilibrium (RNEMD) approach coded in lammps. RNEMD works
on the principle of swapping atomic velocities between a cold and a
hot region. I would not use this method when dealing with systems with
bonded atoms such as solids. In a bonded system, the integration of
the equation of motions with constant timestep can become unstable
when suddenly a slow atom gets a large velocity. The natural way of
dealing with this problem would be to reduce the timestep but on the
other hand you are trying to calculate a property that implies running
the system for a while before reaching convergence.
I think in principle the RNEMD is usable in any system, but its
practical use in bonded-ones would be quickly compromised. I am far
from being an expert in non-equilibrium simulations thus an expert in
the topic might have further/better advice.
Carlos