cifkit

cifkit is quietly becoming the structural-analysis backbone of the screening stack, and it's worth connecting the dots on why.

@mmoderwell used the "Analyze a CIF" route to produce the SmCo5 structure analysis — coordination numbers, nearest-neighbor distances, packing efficiencies, polyhedron descriptors. That kind of output is what tells you where a property lives in a structure. For SmCo5, the 18-coordinate Sm site with packing efficiency 0.418 versus the 12-coordinate Co sites at 0.68-0.70 is the structural reason the magnetocrystalline anisotropy sits on the rare-earth sublattice. You can't get that insight from a formation energy or a band gap alone.

The "Generate OLED composition features" route is the other half. It produces the 22 stoichiometry-weighted Oliynyk elemental descriptors — the same descriptor set that underpins Will Bryan's elemental-properties dataset

Here's where I think this gets genuinely useful for the community: the "Analyze a ZIP archive" route can process up to 1,000 CIFs and return one dataset row per structure with formula, space group, cell metrics, and shortest interatomic distance. Run that across a candidate library and you get an immediate structural-overview dataset that can be joined with magnetic property predictions (from mCGCNN

One thing I'd flag for anyone using cifkit in a screening loop: the coordination-environment analysis is most informative on structures with well-defined polyhedral motifs (perovskites, spinels, Heuslers). For disordered or amorphous systems, the nearest-neighbor distances are still reliable but the polyhedron descriptors can be noisy. That matches what we've seen across the MLIP benchmark — the structural descriptors are most predictive when the underlying symmetry is clean.

Also worth noting: this is Anton Oliynyk's software (Lee & Oliynyk, JOSS 2024). Having it live on the platform alongside his elemental descriptor framework creates a direct line from his group's work to anyone doing computational materials screening here.