Future of JuLIP

[cortner]

CC @jameskermode @tjjarvinen --

As mentioned many times, it seems that it might be time to retire JuLIP, but before we do, can we write down which parts of it were particularly useful so we can move them to new packages or community packages.

Here is a start:

• convenient construction of crystal structures a la ASE (e.g. bulk(:Si, cubic = true) * 3), but maybe this should just be killed in favour of the excellent python packages out there?
• The EAM potential implementation > this can just go into a new package EAM.jl
• maybe also the Si implementation but this one is less clear
• The abstract interface for energies, forces etc > this could become part of AtomsBase, or a similar package.
• Decent optimizers, and most importantly the preconditioning features > could this become part of Molly, or a sister package to Molly?
• The MLIPs abstract interface can just be moved into ACEbase.jl until CESMIX finally agree to make their packages community-owned.

[jameskermode]

convenient construction of crystal structures a la ASE (e.g. bulk(:Si, cubic = true) * 3), but maybe this should just be killed in favour of the excellent python packages out there?

I vote for simply using ASE and friends for this. Building structures should not normally be performance critical.

The EAM potential implementation > this can just go into a new package EAM.jl

Agreed, this would be useful to keep around.

maybe also the Si implementation but this one is less clear

Don't have a strong feeling about this, it's more of a toy model useful for testing.

The abstract interface for energies, forces etc > this could become part of AtomsBase, or a similar package.

Yes, agreed. An abstract calculator interface would be a good addition to AtomsBase, in my view, but it may be difficult to find a suitably generic level of abstraction encompassing both atomistic codes (where forces/energy/stress is enough) and electronic structure (where one would like to expose wavefunctions etc too). ASE also struggles a bit with this.

Decent optimizers, and most importantly the preconditioning features > could this become part of Molly, or a sister package to Molly?

Or a separate package on top of the AtomsBase abstract calculator interface above? Would that maybe too fine grained?

The MLIPs abstract interface can just be moved into ACEbase.jl until CESMIX finally agree to make their packages community-owned.

Also agreed.