search

hackingmaterials / matminer

Data mining for materials science
https://hackingmaterials.github.io/matminer/
Other
472 stars 194 forks source link

BranchPointEnergy featurizer slow + recommendation for weights #323

Open mkhorton opened 5 years ago

mkhorton commented 5 years ago

I'm getting times ~15 seconds per band structure, which seems a little slow for what this is. Perhaps worth making heavier use of numpy for this.

Would also recommend an additional variant citing the recent Shapera and Schleife (DOI: 10.1002/adts.201800075) which has optimized values for n_vb and n_cb kwargs -- the defaults of n_vb=1 and n_cb=1 are not ideal for general use.

computron commented 5 years ago

The relevant section of the paper is below:

screen shot 2018-12-07 at 11 21 33 am
computron commented 5 years ago

N_e = sum of s and p valence electrons per formula unit of the material. I imagine the best is we do sum of s and p valence electrons per unit cell in the band structure calculation (bs.structure).

Probably the way to get sum of s and p valence electrons is to call "full_electronic_structure" on an element and subtract the "full_electronic_structure" of the corresponding noble gas. One way to get the noble gas is to call the regular "electronic_structure" method on the Element and do a regex