# find local minimum of PEL using cg and sd

Dear lammps users,

I am minimizing several snapshots of my system (amorphous silica, BKS
potential) from a long dynamic run. I want to find the configuration
that is at (or very close to) a local minimum of the potential energy
landscape. I used the "cg" and "sd" min styles. I have two questions
about the results, it confused me.

For both styles, I used the default backtrack line search, dmax=0.1,
etol=0.0, and ftol=1.0e-6. Unit is metal. Time step is 0.001.

(1) "cg" gives slightly lower potential energy than "sd", and
converges faster. Below are the results of two min methods on same
initial configuration. I think I should trust cg for my purpose, is
this correct?

(2) Suppose cg is what I need, I would like to see the "Fmax" goes
down to order of 10^-4 or smaller, but I found it stops at Fmax =
0.038885496 in the case below.
The energy next-to-last, final = -8850.24384423, -8850.24384423, where
it seems I have reached the bottom, but why force is not very very
small ?
Also, it stops by "linesearch alpha is zero", i.e. no atom moves if I
am right, if so, will the 0.038885496 force cause one atom to move
more ?

The PEL is more rugged as zoomed in, but I still expect a little flat
bottom (energy not changing, force~zero) at local minimums. I cannot
make sure if I have really found it.
Is my setting wrong or my understanding wrong? should the minimum be
like that? Thank you for any hint.

Regards,
Long
QTP, UF

Aidan can probably comment on this.
The minimizer is stopping b/c it
can no longer make progress in the energy.
I'm not sure why the force (on one atom)
is still so big. How many atoms in
your system? Are you applying any
other constraints to the minimize via
fixes, e.g. extra forces that you are
turning off for minimization?

Steve

I have 648 atoms there, no other constrains for minimization.
Below is the input script, the data file is attached, I think this is
a very straight forward job.
Thanks a lot.

55118.data (25.2 KB)

Dear lammps users,

I am minimizing several snapshots of my system (amorphous silica, BKS
potential) from a long dynamic run. I want to find the configuration
that is at (or very close to) a local minimum of the potential energy
landscape. I used the "cg" and "sd" min styles. I have two questions
about the results, it confused me.

For both styles, I used the default backtrack line search, dmax=0.1,
etol=0.0, and ftol=1.0e-6. Unit is metal. Time step is 0.001.

(1) "cg" gives slightly lower potential energy than "sd", and
converges faster. Below are the results of two min methods on same
initial configuration. I think I should trust cg for my purpose, is
this correct?

steepest descend "sd" is a simple to implement and
fast (per iteration) minimization method that usually
works well when you are far away from the minimum,
but can be very slow to converge tightly.

a conjugate gradient "cg" method is almost always
the better choice.

(2) Suppose cg is what I need, I would like to see the "Fmax" goes
down to order of 10^-4 or smaller, but I found it stops at Fmax =
0.038885496 in the case below.
The energy next-to-last, final = -8850.24384423, -8850.24384423, where
it seems I have reached the bottom, but why force is not very very
small ?
Also, it stops by "linesearch alpha is zero", i.e. no atom moves if I
am right, if so, will the 0.038885496 force cause one atom to move
more ?

not necessarily. it can be prohibited by symmetry,
e.g. when an atom is jammed between two others.

The PEL is more rugged as zoomed in, but I still expect a little flat
bottom (energy not changing, force~zero) at local minimums. I cannot
make sure if I have really found it.
Is my setting wrong or my understanding wrong? should the minimum be
like that? Thank you for any hint.

a high-dimensional potential energy surface has almost
always some "traps" somewhere in some dimensions.

one thing to try would be the "hftn" minimizer.

and a second would be to do a small randomization
of the atomic positions (set random velocities and do
one MD time step with fix nve and reminimize.).
that could be used to break out "jammed" atoms,
but it can also get you stuck in a higher local
minimum, by chance.

axel.

I just tried "hftn", it does better, now the final Fmax is of order
10^-3, (0.0023240066 below).
My final purpose is to find two adjacent minimums, and run NEB to see
the barrier in between.
Do you think Fmax~10^-3 is good enough for my purpose ?

And, you said "it can be prohibited by symmetry, e.g. when an atom is
jammed between two others.".
Do you mean three atoms in a line, although the middle one has
non-zero net force (direction of force is also in the line) but
further moving will cause it to be too close to the other two so it
just stops there ? Can you please explain more ? I don't get it.

Thank you.