consider kwarg for mapping noble gases to a "nearby" element or (better) using different elemental featurizer

[sgbaird]

e.g. $He \rightarrow H\space\mathrm{or}\space Li$

Issue with generated structures having noble gases might be helped by:

• Consider using RGB image instead of grayscale to allow for averaging of atomic numbers #81

[kjappelbaum]

I do not know if you already investigated this, but sometimes it tends to help to not encode atomic numbers but something that is closely related to chemistry (and varying with a more consistent trend across the PTE) such as the electronegativity.

Maybe one could give the user the option to specify any Element property as the encoding of the "chemistry" of a given site. It could default to the atomic number but in this way, you can easily switch to X, electron_affinity, or whatever.

If you want me to make a PR for that, simply assign me to this issue.

[sgbaird]

That would be great! The modified Pettifor scale is one that I considered, mostly based on prior experience with ElMD (see e.g. mat_discover). I like the idea of using an Element property to keep it flexible.

[kjappelbaum]

Since I ended up writing the same code in different packages, I started collecting it in element-coder.

[sgbaird]

@kjappelbaum is element-coder in a state that it can be incorporated into xtal2png?

[kjappelbaum]

I think so - I use it already in one of my other projects. I was waiting for the fit but then got distracted by other things. Will give it a stab this WE!

[sgbaird]

Ok, cool! Not sure if you saw, fit is implemented now #91, though some of the decisions I made on how to set it up were somewhat arbitrary. Lmk if you think something should change with that approach.

Thank you! I appreciate you getting involved despite many other priorities I'm sure

[sgbaird]

See

xtal2png/src/xtal2png/core.py

Line 303 in 09e26f5

atomic_numbers = [s.atomic_numbers for s in S]

in particular (small refactor about to pull into main in #112 (EDIT) #113)

[kjappelbaum]

at least for regression, the encoding makes a huge difference: Here on a MOF dataset, showing top-100-in-top-100 (greater is better), contrasting the mod. Pettifor with electronegativity.

I'm not using the xtal2png for this plot but revised autocorrelation functions, but wouldn't be surprised to also see an impact here.

[sgbaird]

Awesome, looking forward to seeing how this impacts xtal2png. Thanks for sharing!

[sgbaird]

@kjappelbaum this weekend I'm planning to try doing a distributed hyperparameter optimization on my uni's HPC via RayTune. Getting RayTune and Slurm to play nicely together might take some time (my first time connecting the two), but my goal is to at least get a dummy run operating. Two hyperparameters that I think would be really important to have for the full run (which might not happen until next week) are the elemental encodings and lexicographic site sorting (#75).

[kjappelbaum]

ok, the PR is anyhow close to being done. I'll finish it tomorrow morning

[kjappelbaum]

But also looking forward to seeing the RayTune + Slurm code 👍🏽

[sgbaird]

ok, the PR is anyhow close to being done. I'll finish it tomorrow morning

Awesome, I just took a closer look at the changes and am seeing that now. Thank you!

But also looking forward to seeing the RayTune + Slurm code 👍🏽

Sounds good! I hope I don't disappoint haha