UT theoretical chemistry code forum

Hello,
I would like to ask, how do VTSTtools define max force on an atom in CI-NEB calculations? (which we can see at lines 'FORCES: max atom, RMS' in OUTCAR)
It is strange to me that the max force on an atom is converged to a value below abs(EDIFFG) but when I check forces on individual atoms by grepping TOTAL-FORCE, I see forces on some atoms are still large ( large than abs(EDIFFG) several times).

Do you have an explanation for this problem?

Thank you.
Tran

The NEB uses a force projection - only the component of the force on each image perpendicular to the path is used for optimization and the convergence criterion. Yes, the maximum (NEB) force on any atom is what is reported.

Thank you very much for this useful information.
Have a nice day.
Tran

From this answer, I came to another question/idea: Since in a common manner to do NEB calculations, all images are optimized simultaneously and this way is very expensive, both CPU and Memory highly demanding. So can we do like this:
1. Step 1: Optimizing individually each image
2. Step 2: Copy CONTCAR from the previous optimization as new POSCAR and use WAVECAR to start the NEB calculations as normally (optimizing all images simultaneously).
Because, after step 1, forces on each atom are already converged so it is likely that the project forces on each image, in the direction perpendicular to the migration path, are smaller than abs(EDIFFG) and thus step 2 becomes much easier.

I have seen someone asked a similar question before but proposed to do only individual optimizations and take energies to define the energy path and the saddle point. In that way, I think the obtained saddle point maybe not accurate as there are no interactions between images. So, do you think it will be more efficient to do NEB calculations with the two steps suggested above?

No, that's not going to work. If you have access to sufficient computing resources, you can put the calculation of each image on a separate computer with communication through MPI. Then, when you have a reasonable path, you can switch to the dimer method using the neb2dim.pl script and converge to the saddle with a single image.

Thank you very much for your comments and advice :)