Hi users and admins,
I'm new to NEB and have a question about the most efficient way to do NEB calculations. The manual suggests using 1 image initially, and then increasing the number of images as necessary.
My question is, what is the most efficient way to run the calculation as the number of images grows? My choices are:
1) Run all images at once
2) Relax each image individually in its own separate calculation, using the pre-relaxed images on either side of it as fixed endpoints.
Option 2 is intuitively more efficient, as it would reduce the number of degrees of freedom in the overall calculation. Do you think this is a reasonable approach?
Efficient NEB setup
Moderators: Global Moderator, Moderator
-
duncand
Efficient NEB setup
Last edited by duncand on Wed Dec 04, 2013 1:52 am, edited 1 time in total.
-
kelum
- Newbie
- Posts: 42
- Joined: Wed Jul 27, 2011 12:22 pm
- Location: Vienna
Efficient NEB setup
Hi,
option 2 looks more efficient and reasonable, although it obviously needs more job keeping.
One of the issues with option 1 is that it can be difficult to converge. As all the images are moving, the forces between neighbours change all the time. It one of the images jumps a bit away (because of e.g. SCF did not converge and forces are not accurate) then its neighbours will get some kick and a lot of optimisation needs to be redone again.
But also the type of the problem is important. If you have a simple symmetric barrier, then the fastest way to get the barrier is just to run a single image (and continue with option 2 if you need the pathway). If it's not symmetric, you might spend a while locating the maximum. For this case it is probably more efficient to run all images at once first. Not up to the convergence but just some very cheap calculation to give you some rough info about the pathway. Then you can take the image with the highest energy and its two neighbours to optimise it properly (and use option 2 for the rest of the pathway as well).
To sum up: Option 1 is useful at the beginning if the pathway is not simple and option 2 is useful to refine the path and barrier.
I think you can find also some relevant information in our paper where we looked at this a bit more, also looking at DFT specific issues:
http://iopscience.iop.org/0953-8984/22/7/074203
jik
option 2 looks more efficient and reasonable, although it obviously needs more job keeping.
One of the issues with option 1 is that it can be difficult to converge. As all the images are moving, the forces between neighbours change all the time. It one of the images jumps a bit away (because of e.g. SCF did not converge and forces are not accurate) then its neighbours will get some kick and a lot of optimisation needs to be redone again.
But also the type of the problem is important. If you have a simple symmetric barrier, then the fastest way to get the barrier is just to run a single image (and continue with option 2 if you need the pathway). If it's not symmetric, you might spend a while locating the maximum. For this case it is probably more efficient to run all images at once first. Not up to the convergence but just some very cheap calculation to give you some rough info about the pathway. Then you can take the image with the highest energy and its two neighbours to optimise it properly (and use option 2 for the rest of the pathway as well).
To sum up: Option 1 is useful at the beginning if the pathway is not simple and option 2 is useful to refine the path and barrier.
I think you can find also some relevant information in our paper where we looked at this a bit more, also looking at DFT specific issues:
http://iopscience.iop.org/0953-8984/22/7/074203
jik
Last edited by kelum on Wed Dec 04, 2013 12:04 pm, edited 1 time in total.
-
duncand
Efficient NEB setup
Thanks for the information! I will look at your paper as well.
Last edited by duncand on Thu Dec 05, 2013 12:41 am, edited 1 time in total.