`eSCNMDBackbone` no longer using discrete sphere grid

[MJordahn]

What would you like to report?

My issue relates to the current UMA backbone in fairchem. The backbone of the network is called eSCNMDBackbone, yet to the best of my knowledge, this backbone is actually the eSEN GNN backbone. Perhaps it is only me, but I found it rather confusing that when reading the eSCN work, there is a requirement for discrete sphere grid projects (and integration over the sphere when computing energies and forces), yet this is no longer done in the eSCNMDBackbone.

I would just like to confirm that the UMA architecture, and the current eSCNMDBackbone indeed follows the best performing findings in the eSEN work? And if so, if there might be interest in renaming the eSCNMDBackbone to perhaps match the updated architecture? Alternatively, one could explicitly write in the backbone definition that this backbone has been updated to match the eSEN findings?

[wood-b]

Hi @MJordahn, indeed the naming is a bit confusing. It has to do with the history of the code and changing it is messy because we have to deal with backwards compatibility of released checkpoints. To answer your question, eSCNMDBackbone began as eSEN and has been built on for UMA. Hopefully having the file in src/fairchem/core/models/uma and referencing in configs/uma are helpful.

[github-actions]

This issue has been marked as stale because it has been open for 30 days with no activity.