Embeddings for (A)CE models

[cortner]

The initial prototype is a bit rigid in the class of particles systems. We need to import the very general ideas to construct embeddings from ACE.jl.

@zhanglw0521 you are a bit familiar with that, maybe you can give @CheukHinHoJerry an introduction? But don't get started on it yet.

[zhanglw0521]

Does it mean that we need a more general state structure instead of merely the positions of particles?

[cortner]

Related to this - should we bring the States idea to this package?

[zhanglw0521]

I can imagine it is useful to have the States structure in certain cases. Basically, it should not be hard to use a state as an input of a lux chain (e.g., split the input x as ( (|x.rr|, x.be), \hat{x.rr}) in the first layer and send them to radial basis and Ylm basis, respectively, in the second layer, the rest just follows), but not sure where the States structure should go in (this package or the new P4ML.jl, or elsewhere?) and whether or not is it elegant to have it.

Also, it seems that States is not the only thing to consider under this issue - should we discuss this in more detail at some point, say, next Tue?

[cortner]

Where did we leave the discussion about the State

[zhanglw0521]

We mentioned we can either have one more layer in between EquivariantModels.jl and real implementations to deal with the State or simply leave it in each implementation specified package.

[cortner]

So specifically - did we have a firm decision on keeping the State out of this package? Or did we say we will discuss it again? I keep coming back to it and am wondering whether what we are writing here is essentially ACE3. The best parts of ACE.jl should then be integrated here. And in my mind, State was a very successful idea.

[zhanglw0521]

No, we didn't have a firm decision on this but just said that it is not the most urgent thing.

And does it mean that this package will eventually serve as something like ACE3? I am somewhat for this idea, so will have totally no problem with it if we finally want State to lie here.

[CheukHinHoJerry]

I think it is time to discuss/iterate over this? I advocate for the State idea. And if we need more functionality maybe we can make it a tiny package so that P4ML can also depends on?

cc @cortner @zhanglw0521 @YangshuaiWang

[cortner]

We can try that. If I understand you right you want to take the ACE.State and put this into a separate package so we can use it from anywhere we want? Some ideas:

• ParticleStates.jl
• AbstractParticles.jl
• ParticleSystems.jl
• AbstractParticleSystems.jl

??

[CheukHinHoJerry]

What in my mind is that we have a tiny package for States and then write an interface for it from P4ML to make sure all the rrules work as expected even with State input, so that we can easily build any models from it. Not sure whether this would be ideal...

If we decide to try this I think AbstractParticles.jl makes sense.

[zhanglw0521]

If we decide to try this I think AbstractParticles.jl makes sense.

vote for this name too

[cortner]

Ok, I’ll try to get something started tomorrow

[cortner]

I am liking DecoratedParticles.jl -- but we can change this later if we want.

[cortner]

closed by creation of DecoratedParticles.jl