Hello,
I am using LAMMPS/01FEB14 to do some compression test. I want to use restart command to generate restart file. However, i find if i put the restart command into the input syntax, it will affect the final result. (the results are different from the results generated from syntax without the restart command) I do not know which caused this problem.
reset_timestep 0
timestep 0.001 # 1 femtosecond
velocity all create 300 12345 mom yes rot no
fix 1 all nvt temp 300 300 1
dump pre all xyz 100000 dump_pre_*
thermo 1000
thermo_style custom step lx ly lz press pxx pyy pzz pxy pxz pyz pe temp
thermo_modify lost ignore
dump pre all cfg 10000 dump.pre.com.*.cfg mass type xs ys zs x y z c_csym c_peratom fx fy fz #restart 10000 restart.cu
run 10000
On any timestep you write a restarf file, the code forces a re-neighboring
to occur. This is to make a restarted run more nearly identical to the
original run, since the restarted run will perform a reneighboing when
it starts up. But if you compare to an original run that did not do
reneighboing on that timestep (which it may not if you don’t use the restart
command), then the two trajectories will diverge more rapidly due
to round-off issues. For a well-formulated model, the differences should
only be statistical, but that is the nature of MD.
On any timestep you write a restarf file, the code forces a re-neighboring
to occur. This is to make a restarted run more nearly identical to the
original run, since the restarted run will perform a reneighboing when
it starts up. But if you compare to an original run that did not do
reneighboing on that timestep (which it may not if you don't use the restart
command), then the two trajectories will diverge more rapidly due
to round-off issues. For a well-formulated model, the differences should
only be statistical, but that is the nature of MD.
all of this assumes that the simulation is properly set up, the
results are statistically well converged, and it has been run with
reasonable settings, e.g. for time step, neighbor list rebuild, kspace
convergence and more.
the big question now is here: *how* different are those different
results. if they are too far apart, that may indicate that either the
simulation is not well set up or the results you observe have a large
statistical error or that you have too aggressive simulation settings
or a combination of those.
How about write_restart command, will it perform reneighbouring and disturb the results?
it doesn't matter how you trigger the restarting.
rather than worrying about restarting affecting your results, you
should worry about why and how much they are affected. like steve and
i pointed out: either the difference is within reason and thus should
not be worried about, or there is something wrong with your simulation
and *neither* result (i.e. with or without restarting) is trustworthy.