possible to apply for per-atom descriptors?

[thangckt]

Dear Developers.

Do you plan to update MLCVS that can allow to apply for per-atoms descriptors?
I mean the descriptors are computed as per-atom vectors than a global vector for a whole system

Thank you so much

[EnricoTrizio]

Hey Thang,

Right now we are working on a general reorganization of the library structure to make it both easier to use and to develop, so we are quite busy with that.
We have some projects in that sense but they will still require some time.

What did you have in mind? Something like Behler-Parrinello symmetry functions or stuff like that or something closer to graph neural networks?

[thangckt]

Dear @EnricoTrizio

I don't know Behler-Parrinello symmetry functions
I just think about the case MLCVS can be applied to compute a per-atom quantity (may be close to the concept of MULTICOLVAR in PLUMED.

Since you already have a PyTorch interface for PLUMED, this new feature will have a huge benefit, for example, it can mitigate other attempts to build a multicolvar, since we can define any arbitrary function as a pytorch model (not necessary an ML model), then just use it in Plumed in Inference mode.

For example: use mlcvs model to differentiate an atom is in FCC, BCC, or HCP structures. In this case, we define peratom desciptors and apply mlcvs model for each atom in the system.

Is it possible?

[EnricoTrizio]

Ciao @thangckt,
I'm so sorry. I lost track of your issue along the way. Please forgive me.
Luckily, at that time, I would have been of little help, I think 😅

The good news is that, in the meantime, we have been working on extending the transform module of the library (see #124), which is meant to provide preprocessing tools meant to operate directly on the atomic positions.
For example, we have already implemented the computation of several descriptors there.
I don't know if it directly solves your problem, but it can be a good starting point.

In principle, as the transforms are torch.nn.Modules, you could also use them as a standalone tool (i.e., without any NN model attached) to implement your own multicolvar-like functions, script them and import them in PLUMED.

Hope it helps!