CI-NEB Convergence Issue?

[mtgarvey]

UPDATE: The calculation has finished (it was set to run max 500 steps). I have put the final vaspout plot for one of the steps and the mep plot below. I am not sure if this just needs to be run more or if something should be changed? I noticed that the first and last images are lower in energy than the IS and FS. Could that cause problems?

P.S. This is should be the final calculation that we need to run for a publication, so I would really like to figure out how to get it to converge.
-----------------------------------------------------------

Hello All,

I am running a CI-NEB calculation of a system similar to one that I posted about in May viewtopic.php?f=2&t=1516

I have converged the calculation using gamma point only and am now running it with the desired k-point divisions. Things looked ok at first, but now some of the images seem to be oscillating in energy and force (see vaspout images below). My INCAR file is below. Is this behavior normal? Should I change optimizers or parameters?

Any hints or suggestions would be appreciated.

INCAR
-----------------------------------------------------------
SYSTEM = tilted to linked

PREC = Normal
ENCUT = 400
ALGO = Fast
LREAL = Auto
#LWAVE = .FALSE.

# NELMDL = -3
ISMEAR = -1
SIGMA = 0.03

ISYM = 0

IMAGES = 5 # number of images, forces NEB method
LCLIMB = .TRUE. # use climbing image method
IOPT = 1 # LBFGS = Limited-memory Broyden-Fletcher-Goldfarb-Shanno
INVCURV = 0.002
IBRION = 3 # tell VASP to do MD
POTIM = 0 # tell VASP zero time step

NELMIN = 4 # do a minimum of four electronic steps
EDIFF = 1E-6 # electronic convergence
EDIFFG = -0.02 # accuracy of ions -0.3 Low; -0.01 High
NSW = 500 # number of ionic steps

LPLANE = .TRUE.
LSCALU = .FALSE.
NSIM = 4
NCORE = 8
-----------------------------------------------------------

[mtgarvey]

I have restarted this calculation using the Quick-Min optimizer and it seems to be getting worse. The forces and energy for all images are increasing (see example below). Does anyone have a suggestion on what may help?

[graeme]

If you post your input files or a .tar.gz file of the run, I'll take a look.

[mtgarvey]

Professor Henkelman,

Thank you for looking at this. I have posted .tar.gz files of two runs. The file run4.tgz is the last run with the gamma point only version which converged. The file run5.tgz is the run started from the results of run4 and seemed to give strange behavior after ~320 steps. Please let me know if you need any other information.

[graeme]

I think I've looked at this system before -- it is a remarkable pathway.

So I see a few issues with this path. First, the initial and final states do not seem to be properly minimized since the neighboring images are lower in energy. It is understandable that this minimization will be difficult since there are soft modes associated with the molecular bending and stiffer modes associated with the Au-Au bonds.

More importantly for convergence, you can see a significant shift between images 2 and 3. This indicates that you have poor resolution of images near the saddle. Specifically, the overall path has large motion of the molecules and yet you see a significant shift of the Au adatom near the climbing image. Without sufficient resolution, the climbing image can oscillate around the saddle without ever being able to converge to it.

So my recommendation is to fully minimize the endpoints and then reconverge using ~8 moving images. That, or start a dimer (min-mode) calculation from your current best-guess at the saddle. The neb2dim.pl script can help with setting this up.