Potential unit issue with StructureMatcher.get_rms_dist() in evaluation

[rongaoli]

https://github.com/facebookresearch/all-atom-diffusion-transformer/blob/b9ce505f170597a7c8ca50d13ce8e15df21cf8c9/src/eval/crystal_reconstruction.py#L98C13-L98C79

This method might return a volume-normalized dimensionless value rather than actual distances in Ångstroms. The normalization seems to be (volume/nsites)^(1/3).

However, in your paper, it appears to treat these values as Ångstrom units.

Could you clarify this, thank you!

[chaitjo]

Hi @rongaoli - thanks for raising the issue. If what you say is true, this was an error on my part!

Where did you find the information on "method might return a volume-normalized dimensionless value rather than actual distances in Ångstroms"?

[rongaoli]

Hi, @chaitjo

You can check this doc:
https://pymatgen.org/pymatgen.analysis.html#:~:text=get_rms_dist(struct1%2C%20struct2)

Hope it can be helpful :)

[chaitjo]

Thanks for pointing me to this. I will update the arXiv version of the paper with this information eventually.

I think this doesn't really impact the results/claims made in the paper, but is more of a detail regarding the evaluation of the Autoencoder part.

Until then, keeping this issue open...