Generation of configurations question

[MES-physics]

Dear Users and Developers,
What methods, software, etc. do most of you use to develop initial configurations to be trained,
before the minimization with DFT/VASP? (I'm not referring to any subsequent M.D. iterations.)
Or do you just start with the available crystal databases?
Many thanks.

[albapa]

Hi,

We have developed at data-efficient method for crystalline structures: https://iopscience.iop.org/article/10.1088/2632-2153/ac9ae7/pdf

Or you can run ML-assisted AIMD with CASTEP to collect relevant configurations: https://pubs.aip.org/aip/jcp/article/159/4/044803/2904492

And there are many active learning schemes out there! I hope this helps.

Cheers,
Albert

[MES-physics]

Thank you!

[gabor1]

Nowadays we would use a foundation ml model such as mace-omat (there are earlier versions, mace-mpa0 or mace-mpa0, or the mace-off family for organics). These are typically stable enough for random search, geometry optimisation, neb, and MD.

[MES-physics]

I looked into this a bit, but tell me if I understand. So if using a mace_omat model, one can choose the desired element , and this model can be directly plugged into LAMMPS ?
And it is not just a way of producing xyz files that I can train in GAP? Have we moved beyond GAP at this point?
I'm a student and would like to see the general picture of the issue.
Thanks very much.

[gabor1]

yes, the model can be plugged into lammps.

a large majority of research groups have indeed "moved beyond GAP". There are still some legacy users for a variety of reasons, e.g. a project started a long time ago and GAP works well so no reason to change, or the fact that GAP can be trained in a few minutes if you have a thousand CPU cores. but MACE CPU is also coming (only for evaluation not for training), so it's a moving landscape.