Question about the setting of DFT (Density Function Theory)-NEB (Nudged Elastic Band) calculation

Question

I'm attempting to calculate Mg ion vacancy migration inside Pyrope unit cell. System has total 159 atoms, beginning and end geometry is optimized. For intermediate image, I'm using one coordinate that I acquired from crashed NEB case. This is not perfect intermediate geometry, but I thought this should be enough one to guide NEB.

1) Is there any rule of thumb to decide "num_of_images" based on moving distance of target atom? In my script, I wish to move one Mg atom from one position to next vacancy which is about 3.5 angstrom away. In this case, how many images for NEB would proper? For example, if I choose 1 image per 0.5 angstrom, then 8 images should be chosen including the beginning and the end. But I'm not sure if 1 image per 0.5 angstrom would be the proper value. How do you decide "number of images" per given distance?

2) How many core number should be chosen for NEB? I have some sense of core number decision based on my experience, but this is for relaxation. For NEB, I'm not sure how much core number is proper for given atom number or given number of images setting. I wish to learn if there are rule of thumb to decide core number for given NEB conditions.

3) How to decide intermediate image? First, I tried my best guess of Mg ion location between two vacancies as an intermediate image. Then, I'm using one from one of the crashed NEB simulation result. Is this good approach?

4) Can freezing atom helps convergence? (freezing = fixing x, y, z coordinate of atom) Is this better to freeze (put 1 1 1 beside the atom coordinate) the target atom for the beginning, intermediate, and end geometry? Or surrounding atoms around target atom? For now I'm not inducing any freeze to any atom in my NEB setting, but I'm curious if freezing helps convergence of NEB calculation.

Thank you

Answer 1

There is not a general rule for choosing the right displacement between images. For very big systems, I sometimes launch two NEB calculations: one with shallow parameters (eg: steps of 0.5 angstrom on the reaction coordinate), and later I use the almost converged geometries near the TS for another NEB calculation, with finer (less than 0.1 angstrom) steps. Of course the nature of the system might help you choose: sometimes things behave funny, and you have to choose a strategy for each system. Anticipating your last question, freezing or constraining sometime helps: while it doesn't improve convergence, it can speed up the calculation (by telling the program that some movements are forbidden, so that the NEB algorithm doesn't need to try them before discarding them). The intermediate image should not be mandatory for a NEB calculation (if not given, it is generated by interpolation and optimized at every cycle), but it can speed up a calculation, if provided. It can be taken, for instance, from a geometrical scan, or from an approximate TS geometry. As for the core number, if you are talking about processors, then generally speaking as much as possibile is a good choice (edit: provided that the software you are using is capable of parallelizing efficiently the slowest step of the calculation: your choice could be the same you use for geometrical optimizations at the same level of theory)