Dear Users,
dear xiang gu,
I have here a question about system relaxation by means of energy
minimization, i.e. regarding the Lammps minimize command.
My system contains several millions of Si atoms (I simulate with
Stillinger-Weber potential), in which a specified concentration of defects
in form of Frenkel-pairs are created. I intend to relax such a system to
its lowest energy before studying further physical processes we are
interested in. However, when I used minimize command together with style
cg (or sd) to relax it, I found such a relaxation proceeds very very
if you'd make a quick survey of the convergence characteristics
of a conjugate gradient and even more so of a steepest descend
algorithm, you'll see that none of them is well suited for your
project. from what i've seen, you best shot would be a linear
scaling BFGS variant. of course, one important question that you
should consider first is, how exhaustive isyour minimization
supposed should be, i.e. how tight you want to converge. any
minimization scheme will only lead you to the nearest local minimum
and depending how rugger you potential hypersurface is and how
energetically close various minimal are, you may be setting your
convergence parameter needlessly too tightly.
slowly if I set energy tolerance as 1e-11 or 1e-12; even when TOL = 1e-12,
the actual dE/|E| is still around 5e-12 after 400,000 timesteps are
...and you also should consider the limitations of double precision
floating point numbers. you have only 14 digits accuracy in your
mantissa. so going below a relative convergence of less than 1e-11 or
less for millions of particles is asking for a _lot_.
performed. As we want to relax dozens of samples, such a slow speed is not
that we can accept.
And something I found interesting is that, during most time of the
relaxation, energy falls very very slowly; but sometimes, within dozens or
hundreds of timesteps, energy can fall rather significantly--but as this
quick energy descending is not dominating the process, it doesn't help
much to speed up the relaxation.
that is just a sign of the characteristics of your potential
hypersurface in combination with the algorithm.
So I want to ask scientists who are experienced in this lammps
minimize: is the code relating to energy minimization in Lammps
efficiently optimized or not? What is the physical reason of the quick
energy descending sometimes happens within short time (perhaps because the
code is written not in the most optimized way)? And any good suggestions
optimization _cannot_ have anything to do with that, since
with or without optimization you should get the same result
(within the limitations of floating point accuracy).
for speeding up the relaxation process by means simulation setup
optimization (I mean avoiding modifying the lammps code)?
i have three suggestions:
a) don't go for an accuracy that is by far exeeding
the accuracy of your model and the quality of information
that the method you choose can provide.
b) you may want to try an MD with annealing and
a friction term as an alternative.
c) any "optimization" in that sense that you are looking for
would mean to implement a better algorithm and that requires
that you (or somebody else) get your hands "dirty",
i.e. modify lammps.
you have to keep in mind, that the purpose of minimizer algorithms
in most MD codes is simply to remove excessive potential
energy after assembling a system from fragments and there
CG and even SD are efficient enough, so that MD runs don't
crash or have atoms/molecules being "shot around" because of
excessively high forces.
cheers,
axel.