Dynamic Relaxation for siesta/flos NEB calculation

[rezabma]

Hi Nick,
I am working with NEB calculations with siesta/flos and siesta/ASE (https://wiki.fysik.dtu.dk/ase/ase/neb.html).
As you may know, ASE is a set of python codes that links to different calculators (such as siesta, openmx, fhi-aims, vasp, orca, gaussian, dmol3, castep, and many other codes) to do optimization, NEB calculation, phonon analyses, and other calculations. from two years ago, I used different calculators with ASE to do NEB calculations. The best one is siesta+ASE which is very fast and robust.

Comparing to siesta/ASE, siesta/flos needs more steps ( more than twice), but each similar steps (with the same n_images and the same SCF parameters) in siesta/flos is very faster than siesta/ASE. For example, one NEB calculation in FLOS+siesta is twice or more faster than ASE+siesta (maybe this is because that FLOS uses LUA and ASE is based on Python). Almost half of these very large NEB steps contain some images with very low forces (below the criteria I set; i.e. 0.05 eV/A). For example, with n_images equal to 12, in 151 th step of NEB, more than 7 images have forces of 0.07 or less. But in the next steps, these images change, and therefore, their forces change. One good idea is that, the structures of these images are fixed unless their forces becomes greater than the threshold value (0.05 eV/A). You can follow this idea if the following paper:

‘Scaled and Dynamic Optimizations of Nudged Elastic Bands’, P. Lindgren, G. Kastlunger and A. A. Peterson, J. Chem. Theory Comput. 15, 11, 5787-5793 (2019).

It has been proved that with this trick, speed of convergence increases more than twice (see the paper).

In ASE, it has been implemented as "DyNEB". You can see its source code:

https://wiki.fysik.dtu.dk/ase/_modules/ase/neb.html#DyNEB

Thanks
Reza

[zerothi]

Any takers would be very welcome, I don't have any time at the moment to commit... So any help is much appreciated! :)

[arsalan-akhtar]

@rezabma Tanx ill look at it

[zerothi]

@rezabma could you perhaps please give full details of the run where you did this? I.e. are you sure the convergence parameters are the same in flos/ase? In ase it is governed internally, whereas in flos it is governed by siesta.

Anything that could help us figure out where the performance hit comes from would be appreciated.