I’m doing an ABF simulation, I ran the system for a few nanoseconds and then I tried to continue that simulation with the respective restart files from the initial simulation, (I’m using colvars as a library in LAMMPS)
I launched two exactly the same simulations based on the config file and restart files and also number of processors, but the trend between two continuations starts to deviate after some time! Shouldn’t I expect exactly the same trends between these two runs? or is there any randomness in effect somewhere?
I’d really appreciate it if you clarify this to me.
Hello Kasra, if you disable ABF and run two unbiased MD simulations from the same restart file, and run for a few million MD steps, do you ever see the exact same trajectory between the two files?
Giacomo, I’ll try it and report it back, it may take a while as it was happening after 5ns, I’ve turned the ABF off and I’ll let you know as soon as I get the results.
Giacomo, I'll try it and report it back, it may take a while as it was
happening after 5ns, I've turned the ABF off and I'll let you know as soon
as I get the results.
you are missing the point. giacomo was asking a rhetorical question.
in fact, the issue is a FAQ on any MD mailing list. too many people
don't realize that MD is a chaotic system and thus *always* exhibits
exponential divergence (in popular science also known as "the
butterfly effect") unless you use fixed point math and no thermostat.
Thanks Axel, I thought using the same restart file and the same number of processors on the same cluster are enough conditions to not perturb the solutions and get the exact same results. So apparently what is causing the deviations in my case is using thermostat in my simulation?! but it’s a berendsen thermostat, i.e. it doesn’t involve any randomness as something like langevine, can that still cause the divergence?
Can still using the same number of processors on the same cluster (not the same nodes though but all nodes having the same specification) cause divergence?
Thanks Axel, I thought using the same restart file and the same number of
processors on the same cluster are enough conditions to not perturb the
solutions and get the exact same results. So apparently what is causing the
deviations in my case is using thermostat in my simulation?! but it's a
berendsen thermostat, i.e. it doesn't involve any randomness as something
like langevine, can that still cause the divergence?
simply using floating point math is sufficient.
Can still using the same number of processors on the same cluster (not the
same nodes though but all nodes having the same specification) cause
divergence?
yes. you just need a single flipped bit and unless you enforce ieee
754 style rounding after *every* single floating point operation
(which slows down things massively) and use exactly the same number of
processors you will get a divergence. even with fix nve only.
Crystal clear now…thanks. I was concerned as this divergence affects the biasing samples and pmf trends in ABF between two simulations big time (comparing them on the fly) so I have to expect them to show similar results after enough time on each has lapsed, basically comparing them on the fly has no meaning, right?
Crystal clear now...thanks. I was concerned as this divergence affects the
biasing samples and pmf trends in ABF between two simulations big time
(comparing them on the fly) so I have to expect them to show similar results
after enough time on each has lapsed, basically comparing them on the fly
has no meaning, right?
multiple independent simulations should converge to the same overall
result. some properties converge faster than others. sometimes you
have a very fast process overlaid with a very slow one, so you may
think you have converged results, but then you don't.
the bad thing is that you never really know for sure. and what was
acceptable some time ago may be considered too little now.
check out the term "ergodicity" and discussions of it in some stat
mech text book...
since you are using enhanced sampling methods, it is worth spending
more than just the minimal effort on understanding these issues. there
are some really good review papers and books on this around, too.