rdkit integration

[scott-materials]

I remember that there were plans to integrate rdkit into Simmate. Anything moving on this front? I did a quick search in the code and didn't find any significant mention of rdkit.

I ask because we may want to use some rdkit features to describe topology / connectivity in some of our solid state materials.

Thanks @jacksund !!

[jacksund]

A bunch has been done!! However, all of the code using rdkit is still private & closed-sourced. I am going through an approval/review process to make it available in this repo -- but I can't make any promises on when (or even whether) things will be added. I'll say I'm very optimistic based on initial internal discussions though 😄.

Which rdkit feature in particular are you after? Just the database cartridge and it's methods? Or more so toolkit fingerprints?

[scott-materials]

More on the fingerprint side. While we like CrystalNN for getting the coordination shell, we are missing longer-range information. There are a variety of ways to provide this longer-range information, such as adjacency matrices and radial distribution functions, as you know. But I'm interested to explore multiple long-range fingerprints and see what works best. My suspicion is that most fingerprints in rdkit are made for molecules rather than periodic solids, but we can easily take a small fragment of a periodic solid and apply the same methods to it.

[jacksund]

More on the fingerprint side.

In that case, I'd just install rdkit separately & alongside simmate (for my internal apps, I list it as an extra dependency). Then just play around.

My suspicion is that most fingerprints in rdkit are made for molecules rather than periodic solids, but we can easily take a small fragment of a periodic solid and apply the same methods to it.

Rdkit is 100% for molecules, and keep in mind that the majority of their fingerprints operate in 2D space (i.e. apply to lewis dot structures). Also these 2D fingerprints depend on bond-information being explicitly given. So yes, rdkit's fingerprints can be "long-range", but that "range" is based on 2D connectivity -- not 3D space.

Just something to keep in mind! Best of luck 😄

[scott-materials]

Thank you, this is helpful!

Related to these fingerprints, I recall you developing an USPEX-like partial radial distribution function several years ago. (I believe you were simply attempting to replicate the USPEX partial RDF.) Do you have that code available somewhere? Not sure if it made it into Simmate. After doing some brainstorming and discussion, we're thinking that the partial RDF would be a nice complement to CrystalNN.

[jacksund]

I have it in simmate, but it will likely be removed because I already added it to matminer!

In simmate: (it's buried since I was planning on removing it eventually anyways)

• https://github.com/jacksund/simmate/blob/main/src/simmate/toolkit/validators/fingerprint/pcrystalnn.py

In matminer:

• the PR: add partial site stats fingerprint hackingmaterials/matminer#809
• the pssf code I contributed: https://github.com/jacksund/matminer/blob/main/matminer/featurizers/structure/sites.py#L248

[scott-materials]

Wow... this is great. Thank you!