Big force components after relaxation

Dear LAMMPS users,
I’m a rather new LAMMPS user, and I have a question about geometry optimization, and I hope for your help.
When I did relaxation of my structure using cg optimizer a noticed that the final maximal force component is so big
Force max component initial, final = 152.668 1.82271
But, as I think, convergence has been achieved.
When I try to relax simple structure like a diamond structure I obtained good force components (~1e-08).
It would be very kind of you if somebody can help me to fix this problem.

thank you.

Here is my input file

units metal
newton on
dimension 3
boundary p p p
atom_style atomic
read_data coo
pair_style airebo 3.0 0 0
pair_coeff * * CH.airebo C
neigh_modify every 1 delay 0 check yes
compute _rg all gyration
min_style cg
min_modify line quadratic
timestep 0.0005
dump 1 all xyz 10000 out.xyz
dump_modify 1 append yes
minimize 1.0e-16 1.0e-8 15000 30000
variable e equal etotal
print “@ $e”
undump 1

and my log.file after optimization

LAMMPS (5 Mar 2012)

units metal
newton on
dimension 3
boundary p p p
atom_style atomic

read_data coo
orthogonal box = (-43.219 -43.219 -43.219) to (0 0 0)
2 by 2 by 3 MPI processor grid
13918 atoms

pair_style airebo 3.0 0 0
pair_coeff * * CH.airebo C

neigh_modify every 1 delay 0 check yes

compute _rg all gyration

Dear LAMMPS users,
I'm a rather new LAMMPS user, and I have a question about geometry
optimization, and I hope for your help.
When I did relaxation of my structure using cg optimizer a noticed that the
final maximal force component is so big
Force max component initial, final = 152.668 1.82271
But, as I think, convergence has been achieved.

well, you never know. high-dimensional optimization
is a non-trivial thing. you may have a strong gradient
in just one direction and a flat potential hyper surface
in all other dimensions, e.g. one atom is "jammed".

there are a number of ways to test this:

- try a different optimizer (e.g. hftn)

- try different options to the minimizer (e.g. quadratic line search)

- do a small randomization of positions
  and optimize again and see, if you reach
  the same minimum

cheers,
    axel.

Dear Axel,

Thank you for your advices!

I tried to use different optimizers like hftn, sd, quickmin and fire, but cg did better relaxation than others because it showed smaller forces.

Now I will try to do what you said with style of optimization and randomization of atomic positions

Thank you

Alexander,

You didn’t provide full information about your system. May be you have a bad initial geometry, e.g. overlapped atoms.
Otherwise, why do you think that force component is rather big? If your structure is unstable during the simulation, try to use smaller timestep on initial trajectory.

Grigory Smirnov .

2012/4/16 Alexander G. Kvashnin <agkvashnin@…24…>

Grigory,

Thank you for your answer.
My structure is good, I checked it very careful and there are no overlopped atoms and so on.

I tried to decrease timestep but didn’t happen.

Why do you think that this force component is good?
I think that relaxation was succesfull when the forces is less than 10e-5 eV/Ang.

Dear Axel,

I tried to use different options to the minimizer, but it was nothing good, forces became much large than with quadratic line search.

I have one more question related to this problem. Is there any criteria that the structure is optimized well by LAMMPS? Because as I think the structure is optimized well if atomic forces is less than 10e-5 eV/Ang.

Thank you!

I tried to decrease timestep but didn’t happen.

If I understand you correctly, you ran only minimization procedure with different timesteps. Try to run dynamic calculation with a small or limited timestep (fix nve/limit) .

Why do you think that this force component is good?
I think that relaxation was succesfull when the forces is less than 10e-5 eV/Ang.

There is no such deterministic criterion for forces in MD. Result depends on geometry, interaction potential and simulation settings.
If your system is stable during the long simulation and force component is big, then this is correct.

Grigory Smirnov