change ships_radial to acejl_radial with splines.

[wcwitt]

Hi @yury-lysogorskiy, I believe you heard from @casv2 that we are updating our Julia exports to PACE. We now have a version that works, so I'm starting a conversation here. Do you mind taking a quick look to see if anything looks objectionable? Thank you!

There is one lingering issue we should resolve before merging. Right now, we export separate, species-dependent pair potentials as LAMMPS splines, completely separate from PACE. In principle, we could export these in a separate yace, and then combine them in LAMMPS, which would be a bit cleaner. But do you have a better idea? Ideally it would be nice to have a single yace.

Also, if any of @casv2, @cortner, or @bernstei want to inspect these updates, you are more than welcome.

[yury-lysogorskiy]

HI @wcwitt , thank you for PR. The code looks fine for me.

Regarding separate pair potential, I have no strong opinion here, but for out recent carbon potential, we needed to add pair-wise D2 correction and also were thinking of having it in PACE, but end up to have it as separate pair_style table and combine with pair_style hybrid/overlay.
The obvious benefit of that is that you can use different cutoffs for pair-wise potential (larger cutoff ~ 10 A) and manybody ACE (smaller cutoff ~ 5-6 A). This is important for example, in KOKKOS implementation, since having one big cutoff results in large number of neighbours and would explode memory usage, reserved on GPU by KOKKOS-PACE.

On other hand side, that was one particular case of Carbon-ACE + D2 correction and not every day usage pattern...

[wcwitt]

Thanks!

to have it as separate pair_style table and combine with pair_style hybrid/overlay

This is indeed what we are doing now. It's fine, but becomes a bit annoying with, say, ternary potentials where there are many separate pairs one needs to overlay.

To take your example, would there be any disadvantages in exporting the D2 correction as a 2-body yace? And then overlaying the two yaces?

[wcwitt]

Hi @yury-lysogorskiy, friendly ping. Do you think a separate yace that consolidates all the two-body terms will be slower than LAMMPS pair_style table?

[yury-lysogorskiy]

I dont know excatly. But I agree that it could become annoying to have multiple pair table. So maybe yes, separate pair yace is better option

[wcwitt]

Hi @yury-lysogorskiy hope all is well. I think we can merge this if you're up for it.

[yury-lysogorskiy]

@wcwitt Would that break backward compatibility with older ACEjl models ?

[wcwitt]

Yes - we think that is okay, but does it bother you?

[yury-lysogorskiy]

I dont know exactly how many users of old yace format, that using SHIPs radial. Maybe one can have both options...
However, if it is not a problem on your side, then I'm also fine. But in that case I need to replace tests of old SHIPS radial with some new ACEjl radials.

So, could you provide me with some examples of new ACEjl potentials (single/multispecies) + some structure (any ASE compat format) + expected energy/forces, please ? I will make tests out of it in our internal core repo