Ghost atoms & IBOs

Hi Team,

The IBO code currently fails for systems with ghost atoms (e.g., with a ‘X-’ prefix):

[...]
  File "/Users/janus/anaconda3/lib/python3.7/site-packages/pyscf/lo/ibo.py", line 135, in ibo_loc
    AtomOffsets = MakeAtomIbOffsets(Atoms)[0]
  File "/Users/janus/anaconda3/lib/python3.7/site-packages/pyscf/lo/ibo.py", line 402, in MakeAtomIbOffsets
    iBfAt.append(iBfAt[-1] + nAoX[Atom])
KeyError: 'XO'

Out of curiosity, what is the expected and correct behaviour here? To treat the ghost atoms merely as their native atoms (for instance, treating ‘X-O’ as ‘O’)? Isn’t this what’s done elsewhere?

Cheers, Janus

[cgk-x]

In case anyone else is interested: I think the correct way to deal with ghost atoms is:

• Do not put any minimal-basis (B2) functions on the ghost atoms. I.e., treat the ghost atoms as if they would not exist in any way, shape, or form as far as the B2 basis is concerned.

• Retain the ghost atom basis functions unchanged in the main computational basis B1. I.e., fully include them in the overlap matrices S11 and S12 (on the B1 side) which get supplied to the IAO construction routine.

The reason I think this is the correct approach is that, ultimately, the ghost atoms are only used to extend the computational basis set B1 in a specific way, and therefore provide more and targeted variation freedom to the original electronic structure computation. But that is the only thing they really do, and that is not very different from, say, using a plane wave basis as B1 to expand the main orbitals in. But as far as the chemical analysis is concerned, they do not exist and therefore need not be treated.[1]

If that is done, IBOs based on such an IAO construction should work automatically, and unchanged, I think. Haven’t looked into the PSI version yet.

[1] To retain correct atomic numbering, and simplify dealing with real- world existing code, the easiest way of dealing with ghost atoms might be to assign them empty minimal basis sets (i.e., basis sets without any functions). That would allow leaving them in, and might simplify code adjustments. Haven’t looked into how PySCF would handle those.

[januseriksen]

Sorry for being a nuisance, but would anyone ‘volunteer’ to chip in with help, pointers, etc.? Happy to help and discuss. (cf. comment from Gerald above)