Question about the oxygen reference in the mace_omat hull split
I've been using the mace_omat split as a MACE-OMAT-consistent hull and encountered something I don't completely follow, so I wanted to check whether I'm misreading it.
In data/mace_omat-00000-of-00001.parquet, the only pure-oxygen entry with e_above_hull = 0 is agm003804182, a one-atom O cell. Its mace_omat_energy is -24.81 eV/atom, while the other columns in the same row are much lower in magnitude:
| column (eV/atom) | agm003804182 |
|---|---|
| true_energy (DFT) | -2.661 |
| orb_conserv_inf_energy | -2.660 |
| uma_energy | -2.685 |
| mace_mp_energy | -3.048 |
| mace_omat_energy | -24.809 |
The O2 crystal in the dft split (agm002119100) is -4.96 by DFT and -4.91 by MACE-OMAT, which looks like the value I'd expect. So there seems to be an issue with the energy MACE-OMAT computes for the one-atom O cell. I get the same -24.8 from the public mace-omat-0-medium checkpoint on that structure, so this seems to be the model's actual output on that cell. The N entry agm004462522 seems to have the same issue, but the effect seems mild (-10.5 eV/atom vs -6.25 eV/atom by DFT).
I only find 2 rows with O in the composition in MACE-OMAT split versus ~18k in the dft split, and I think it could be because the abnormal -24 eV value was used as the reference for O.
I just wanted to clarify if this is expected or if I'm missing something.
Thank you!