Why so many generated structures with hydrogen and/or helium?

[sgbaird]

'Ca2He2H1,volume=23,uid=3533.cif'
'Cr1Ag1H2He1,volume=31,uid=3d40.cif'
'Cs1Li1Ca1H2,volume=62,uid=8e9d.cif'
'Cs1Rb1La1Sn1Sb1Pd1H3I1Ne1He1,volume=177,uid=5495.cif'
'Cu1H3I1He1,volume=37,uid=3870.cif'
'K1Ca1Ac1Mg1Ti1Mn1Al1Cr1In2Ga1Co1Tc1Cu1Ag1Hg1Ge1Te2As1Pd1H3Rh1Se1C1Xe1I1Kr1He1,volume=1358,uid=eccb.cif'
'Li1Ti1H2Xe1Cl1,volume=37,uid=070b.cif'
'Mo1H1C1Kr1He1,volume=24,uid=879d.cif'
'Na1Li1H1Kr1,volume=18,uid=5905.cif'
'Na1Sr1Sc1H1C1Br1He3,volume=102,uid=7fcc.cif'
'Rb1Na1Li1Sc1Ti1Ni1Ru2H2Cl1,volume=123,uid=5fa0.cif'
'Rb1Sc1He1,volume=18,uid=90e7.cif'
'Rb1Ti1V1Ga1Ni1Ge1H2,volume=83,uid=cf25.cif'
'Sc1Cr1F1Ne1H1He1,volume=47,uid=5b5d.cif'
'Sr1Gd1Dy1Y1Ni1Pd1He1H2,volume=108,uid=cf05.cif'
'Zr1Tc1Cu1H4Se1Br1F1,volume=69,uid=4e8d.cif'

See also #79

[sgbaird]

xref: txie-93/cdvae#11

[sgbaird]

Doesn't seem to be an issue with noble gases being represented very frequently, which of course isn't expected anyway since it's experimentally verified compounds. Maybe $He$ is "supposed to be" $H$ and more so has to do with the RGB unscaling.

• consider kwarg for mapping noble gases to a "nearby" element or (better) using different elemental featurizer #82
• Consider using RGB image instead of grayscale to allow for averaging of atomic numbers #81
• consider composition matrix in place of top-left zeros #76

from pymatviz.elements import ptable_heatmap_plotly

equimolar_compositions = train_inputs.apply(
    lambda s: Composition(re.sub(r"\d", "", s.formula))
)
fig = ptable_heatmap_plotly(equimolar_compositions)
fig.show()

[sgbaird]

Also strange that there's not a single instance of oxygen in the generated structures, despite its prevalence. I don't see "adjacent" elements nitrogen or flourine either. Could be an issue with the scaling/unscaling, not enough data, bad model, bad representation, or biased sample. Last one seems unlikely.

Probably not an issue with data shuffling #84

[sgbaird]

from #79

The formulas:

which seem somewhat better.

Interested to see how it goes with imagen-pytorch instead of denoising_diffusion_pytorch.