effect of seed number

Dear all
I have a problem in effect of “seed” number (that use in “velocity create” command) on simulation result.
as I found in documentations and mailing list, I should define the seed number by myself just to take same result for same seed number. I know that a few diffrence in results is not problem when I used diffrent seed #.
but I see a strange behavior for diffrent seed #. in a simple script as the following
when I used seed number like “234” the simulation continued for 1000000 step (end of simulation time) without any problem , but when I used seed number like " 235" the temprature increased after a few steps and the simulation diverge (following log). I repeated the simulations more and even on diffrent computers but the problem exist. I think this not real the simulation have big diffrent result just because of diffrent seed number. please advise me.

timestep 0.0002
fix 1 all box/relax aniso 0
min_style cg
min_modify dmax 1 line quadratic
minimize 0 1.0e-10 100000 1000000
minimize 0 1.0e-10 100000 1000000
unfix 1

write_restart step1.restart

#velocity all create 300.0 235 dist gaussian #Bad result
velocity all create 300.0 234 dist gaussian #Good result
fix 1 all nve
run 1000000
unfix 1

for seed 234:
Step PotEng Volume Temp Press Fmax
10 -23159.63 30267.564 300 5253.4625 3.6300581e-12
20 -23152.268 30267.564 285.17074 5837.8904 3.2628584
30 -23124.743 30267.564 229.7453 14482.083 7.9906086
.
40000 -23083.778 30267.564 147.18974 16118.865 5.5118631
40010 -23083.594 30267.564 146.81859 13905.321 5.9187778
.
999990 -23082.99 30267.564 145.54858 14911.027 6.1356478
1000000 -23082.866 30267.564 145.30059 15383.47 6.433465

Dear Hadi,

how many atoms do you have in your system?

Oleg Sergeev,
junior researcher, VNIIA.

24.04.13, 09:55, "Hadi Sedaghat" <[email protected]>":

Thanks for your replay, Oleg.

I have 3840 atom in simulation box.

Dear all
I have a problem in effect of "seed" number (that use in "velocity create"
command) on simulation result.
as I found in documentations and mailing list, I should define the seed
number by myself just to take same result for same seed number. I know
that a few diffrence in results is not problem when I used diffrent seed #.
but I see a strange behavior for diffrent seed #. in a simple script as
the following
when I used seed number like "234" the simulation continued for 1000000
step (end of simulation time) without any problem , but when I used seed
number like " 235" the temprature increased after a few steps and the
simulation diverge (following log). I repeated the simulations more and
even on diffrent computers but the problem exist. I think this not real the
simulation have big diffrent result just because of diffrent seed number.
please advise me.

your interpretation of the situation is flawed. it is not the one
simulation that "goes bad" that is the problematic one, but rather *all*
are problematic because one of them go wrong. one of the big problems in MD
is proper sampling of phase space, so if you see something bad happening on
one trajectory starting from a different but equivalent set of initial
conditions, then it is likely that the same thing would happen for all of
them, if you run long enough. certain events, like locally correlated
fluctuations, are very rare but have to be taken seriously nevertheless.
the fact that your temperature rises and the simulation diverges indicates
that either your potentials are bad, your timestep is too large or your
neighborlist settings inadequate or something else is wrong. since you can
reproduce the "bad" behavior, you should investigate through visualization
what happens exactly at the point where the problem begins to manifest and
then draw the proper conclusions.

axel.

Dear Axel
Thanks for your valuable comment. I do the simulation run for 10,000,000 steps (10 times bigger than the previus succesfull run) to better understanding what happend in simulation.
1- about “neighborlist settings inadequate” I used the “neighbor 2.5 bin” & “neigh_modify every 1 delay 0 check yes” and the radius cutoff is 4 angstrom and the time step is 0.0001 picosecond that I think is too small. I verify the setting but I think it is adequate. please any advise.

Thanks in advance

Dear Axel
Thanks for your valuable comment. I do the simulation run for 10,000,000
steps (10 times bigger than the previus succesfull run) to better
understanding what happend in simulation.

it doesn't matter for how many more steps you run. rare events are just
that: rare.

1- about "neighborlist settings inadequate" I used the "neighbor 2.5 bin"
& "neigh_modify every 1 delay 0 check yes" and the radius cutoff is 4
angstrom and the time step is 0.0001 picosecond that I think is too small.
I verify the setting but I think it is adequate. please any advise.

i already told you what to do: find out what exactly happens with the "bad"
simulation. your logic is the wrong way around. having reasonable settings
does not *guarantee* that all simulations will work. but if you can make a
"bad" simulation work, then you will have reasonable settings.

axel.