HELP NEB doesn't converge

[Nikosnikov]

Dear Vasp users,

I start to desperate with my NEB calculations, I had perfrom more than 700 ionic steps, and the calculations is not finished yet.
The initial and final images are well converged (no imaginary frequencies), and I used 8 images to interpolate them.

Here is my INCAR file:
[align=center]****************************
SYSTEM = Mo-edge+H2
ALGO = FAST
LREAL =Auto
IMAGES = 8
EDIFF = 1E-6
EDIFFG = -0.05
ISTART= 2
SPRING = -5
#Ionic Relaxation
GGA = 91
IBRION = 1
NSW = 60
#POTIM = .3
#PREC = high
NELMIN = 4
ISYM = 0
ISIF = 1
#IDIPOL = 3
# Electronic minimisation
ENCUT = 337.00 eV
#ENAUG = 833.00 eV
# DOS related values
ISMEAR = 1
SIGMA = .1
LPLANE = .TRUE.
NPAR = 16
LSCALU = .FALSE.
NSIM = 8
****************************[/align]
(I am using VASP 4.6.3)

I think that I went in a wrong way. Browsing the forum and the Vasp manual it seems I have to change strategy to reach the convergence.

1- I can use a small dimensionality by adding NFREE=2 in the INCAR file to pre-converge the images. Then after reaching a convergence re-run without the NFREE tag.
2- I can use the stepest decend method without line optimization (IBRION=3 and SMASS=2)
3- I can "play" on the SPRING constant, the NEB method works with negative values and I used the suggestion of -5, if I put -1, do the convergence will be quicker?

?1-Do anybody can suggest me the good strategy to adopt?

?2-Can you explain me exactly what does mean the SPRING constant? What is the effect of this constant?

Thank you all for your help

N.

[Nikosnikov]

dear all,

it seems that many parameters can affect the convergence of NEB calculations.

Using IBRION=1, the alogorithm can fail to converge so that it's necessary to use the flag NELMIN = 4 (value up to 8, see vasp manual 6.20) and also a value for IBRION , =0.1 for instance to preconverge the calculations.

The parameters to use are not clear to me.

For sure, in order to perform NEB calculations one need to include in the INCAR file:
IMAGE = #
SPRING = -5
IBRION = 1 (or 3)

Then it's not clear to me for using those parameters:
NFREE = 2
SMASS = 3
NELMIN = 4 (up to
POTIM = 0.2 (to preconverge)


I am learning NEB methods, is there a clear overview of this method with VASP, explaining the methods to make complex Minimum Energy Pathways converging?

[admin]

1) please mind that the in the NEB formalism the atoms are not entirely free to move: only the tangential force components perpendicular to the reaction path (which is fixed according to the geomtries of the images) are minimized. Therefore EDIFFG please choose EDIFFG > 0 (total energy differences) as convergence criterium
2) convergence sometimes is better if the number of images is not too large (4 usually is a good choice) you can still refine the path in the neighborhood of the barrier by doing another NEB run, starting from the images adjacent to the barrier as initial and final states.
3) please carefully check if the interpolation between the initial and final state gave reasonable results (equidistant images):
--) the positions of all ions should be interpolated
--) please also mind that due to the periodicity of the cells, a position of e.g -0.1 is equal to 0.9.--> therefore,if an atom moves from -0.1 to +0.1, it may well be that the interpolation gives a drag from 0.9 to 0.1 (i.e the ion moves through the complete cell) --> please check the geometries of all ions in all images.

[graeme]

I don't fully agree with the admin's comments.

(1) In the NEB method, the forces on the images along the path are projected out. The NEB forces should rigorously go to zero as the band converges. I recommend always using EDIFFG<0 because a force-based criteria gives a real measure of how close the band is to convergence. Using EDIFFG>0 does not - the quality of a converged band will be dependent upon the aggressiveness of the optimizer.

(2) Convergence should be better with more images (within reasonable limits), although at a greater cost because each extra image requires an extra force evaluation. Using a climbing-image allows for finding the saddle point with as few images as are required to resolve the path (more precisely, the direction of the negative mode at the saddle).

(3) A correct interpolation script will allow for atoms to move across the box boundary - there should be no problems with periodic boundary conditions and the NEB.

It is not clear from your post how the convergence of the NEB is getting stuck. I do agree with the admins that looking at the band and checking the forces will help sort out the problem. If the forces are high, and stay high, it could be a result of the small radius of convergence of the LBFGS (IBRION=1) method, due to how it is coupled to the NEB in vasp. Using IBRION=3 and a small POTIM=0.01 is generally a safe and quick way to bring the NEB forces down to the 0.5 eV/Ang range. Then you should be able to use a more aggressive time step (POTIM=0.1), or switch to a second order optimizer, such as IBRION=1.

[Nikosnikov]

thank you for your replies,

I am confident with the image interpolation I am usisng (obtained with the reaction-preview tool of Material Studio), ensuring that the peroidic boundary conditions don't drive atoms from side to side of the cell.

I decided to do force-constrained calculations because I palnned to compute the Transition state frequencies and with the experience I have, the forces have to be well converged.

In the other hand the energy constrain is less restrictive than the force constrain, so that I switched to energy constrain. After the convergence is reached, I will try to go back to force constrain to see if I can reach the force converged point.

Concerning the number of images, I agree with graeme, I think it's easier to reach the convergence if the energy path is sampled with a lot a images. Well it's more computational costly but also the images are closer each other and the constraint are less "agresive".

N