[FEATURE] Using and Predicting Atomic Charges for Defect Systems

[BorisWasBezet]

Hello,

I am currently using NequIP to model systems with defects, and I have noticed that the force predictions near these defect sites are often poor. This suggests that the model is missing some important information needed to make accurate predictions in these regions. In particular, ionic bonding and local charge redistribution appear to play a significant role in these systems, and explicitly incorporating atomic charges might be necessary to capture these effects.

While reading through the code, I noticed that NequIP retrieves atomic charges from ASE, but I have not been able to determine whether these charges are actually used anywhere in the model or during training. I would therefore like to ask whether NequIP currently incorporates atomic charges in any way. If not, what would be the recommended approach for adding them as per-atom input features?

Additionally, I am interested not only in using atomic charges as inputs, but also in having NequIP predict per-atom charges as an additional output. Is this something the model architecture already supports, or could support with reasonable modifications? If this is not currently available, I would really appreciate any guidance, suggestions, or pointers on how I might implement this myself.

Any clarification or advice would be greatly appreciated!

Kind regards,
Boris

[kavanase]

Hi Boris.
Yes poor predictions at defect sites are currently a major issue for MLIPs.

Currently no, atomic charges are not used anywhere in NequIP.

There are some options for this, but I'm not sure how useful they will be in this case. There is the NequIP-LES extension package that includes long-range electrostatics via latent ewald summations (see their repo/paper for details). Additionally, we have incorporated charge/spin embeddings to handle non-zero charge/spin systems, which will be released soon, which could be useful (but this is currently for total charge/spin).
It is also possible to assign charges to atoms, and denote them as different 'species' when handled by NequIP (i.e. H^0+, H^1+, H^1- etc), which was previously trialled with the SPICE dataset (see https://arxiv.org/abs/2504.16068).

Additionally, I am interested not only in using atomic charges as inputs, but also in having NequIP predict per-atom charges as an additional output. Is this something the model architecture already supports, or could support with reasonable modifications?

I will defer to @cw-tan as the expert for this. I think the NequIP-LES approach outputs predicted effective atomic charges, but I'm not totally sure

[cw-tan]

On the develop branch, we can now embed discrete node fields: https://github.com/mir-group/nequip/blob/develop/nequip/nn/embedding/node.py. Shouldn't be too much extra work if you want to take in continuous variables (but worth thinking about normalization properties).

[BorisWasBezet]

Thank you both for the tips! I will give the LES extension a try first to see if it helps.