NEB Convergence for large systems

[Salman_101]

I want to calculate the path of diffusion of surface vacancies in ZnO. I was able to get converged results for a 2x2 3 layered cell using 5 images, but the same calculation didn't converge with 3 images. Now I want to do the same calculations, but with a bigger cell ( 3 layered 3x3 cell). I have tried using 5 and 7 images, but the calculations just won't converge. I preconverged all the images using IBRION = 3 and I converged the end points using IBRION =1 and I am using IBRION =1 for the NEB run. I am doing the simple NEB run i.e. not using the climbing image technique.
I am attaching the input files.

Any pointers?

[graeme]

There are a few odd things about this calculation.

First, I suggest freezing the bottom atoms in the slab to the same positions in your initial and final states. Then, when you make the neb path, these frozen atoms will be frozen in the same position.

Second, I don't see any need to pre-converge the intermediate images. Just create an initial path between your endpoints and relax it.

Third, If you want to use as have as few images as possible, the climbing image method is suggested.

Fouth, if computational expense is a problem, you can use the neb2dim.pl script to generate a dimer calculation from an NEB calculation. This will allow you to find a saddle with just one image.

Finally, to debug problems with convergence, you would have to post an example calculation. It is important to understand if convergence is just slow, or if there really is a problem with the calculation.

But anyway, start by freezing the bottom atoms in the slab and run a climbing image NEB calculation. If there really are convergence problems, I would be interested to see the output from the calculation so we can get to the bottom of it.

[Salman_101]

Thanks a lot for the quick reply.
I will do the calculations freezing the bottom atoms. I am attaching the results of my previous calculations, if that may help. I have removed the OUTCAR files, because of their bulky size.

The reason why I pre-converged the images was because the moving oxygen atom was getting scaringly close to a Zn atom. First I linearly interpolated between the starting and the end points and then for each image I did a selective dynamics convergence by fixing the x coordinate of the hopping oxygen atom and allowing all the other atoms to move.
Another reason why i was prompted to do so was because it is mentioned on the vasp website that to improve convergence in a NEB run the images should be preconverged. Here is the link
http://cms.mpi.univie.ac.at/vasp/vasp/E ... ethod.html

Thanks again.

[graeme]

I understand the point about atoms moving close together with a linear interpolation. You might try our nebavoid.pl script which pushes atoms apart to a specified distance. Another new and useful algorithms which is available through ASE is the NEB-TR script, which uses internal coordinates to interpolate an initial path.

The simplest solution, however, is to use a conservative optimizer, (IBRION=3 and POTIM=0.1) until the forces on the images are below 1 eV/Ang. Then, you can switch to a more aggressive and efficient optimizer.

For debugging, the OUTCARs are essential; don't worry about the size.

[Salman_101]

Thanks for all the advice. I hope I am not being a bother.
I am attaching files for a 7 image run with all the OUTCARS included.

[graeme]

I don't see anything wrong with this NEB calculations. While it did not reach convergence within the 300 iterations allowed, it does appear to be converging. You can speed things up using our vtstcode, the climbing image, and freezing the lowest layer of atoms.

[Salman_101]

Thanks for all the help

I was finally able to get a converged result by freezing the bottom layer, though the convergence was achieved after 941 steps. Next I will try doing the same calculation using the climbing image method.
Thanks again, you have been very helpful.