A CPU-only, offline CIF/POSCAR structure-description route with live reference controls, explicit limits, and upstream provenance.
Robocrystallographer adds a lightweight, CPU-only way to turn an uploaded crystal structure into a human-readable crystallographic description and a compact machine-readable summary of local environments and connectivity.
Use Describe a crystal structure with a CIF, POSCAR, or VASP file. The response includes:
prose describing the encoded structure;
parsed formula, number of sites, space group, and crystal system;
symmetry-inequivalent local environments, coordination counts, connectivity, and bond-distance summaries;
warnings and versioned provenance (robocrys==0.2.13, pymatgen version, input SHA-256).
It is deliberately narrow and complementary to numeric metadata/reporting tools: it does not predict stability, synthesizability, energetics, or any material property.
Control | Expected | Live result |
|---|---|---|
The pinned receipts for the two successful controls are below.
Upload a CIF/POSCAR/VASP structure to receive a prose description, local environments, connectivity, warnings, and provenance. Most calls return in seconds; allow up to ~35 seconds if the service must wake from idle. Not a stability or property predictor.
Upload a CIF/POSCAR/VASP structure to receive a prose description, local environments, connectivity, warnings, and provenance. Most calls return in seconds; allow up to ~35 seconds if the service must wake from idle. Not a stability or property predictor.
The text is only as trustworthy as the input structure. An under-specified or wrong cell can produce a confident, internally consistent description of the wrong structure. Check the returned formula, site count, and space group against the phase you intended before interpreting coordination or connectivity. CrystalNN-derived local environments can also be less reliable for unusual chemistry or structures without oxidation states; those warnings are surfaced in the response.
Version 1 accepts ordered structures up to 500 sites and 1 MiB. It is fully offline after file download: there are no weights, GPU, Materials Project lookup, or runtime network dependency. Open Babel is intentionally omitted, so molecular naming is not provided.
This service wraps robocrystallographer, specifically robocrys==0.2.13, based on Ganose & Jain, MRS Communications (2019). The upstream repository license is BSD-3-Clause (modified BSD); the Ouro integration wrapper is MIT-licensed, and the upstream notice is retained in the deployed image.
|
Returned Halite/Rock Salt, |
| Returned diamond, |
fail rather than invent a structure | Live route returned HTTP 422 |
This is a particularly useful companion to the MLIP benchmark because it puts an auditable structure check before an expensive relaxation. Our CeFe12 tests showed how sharply the apparent symmetry failure can depend on the input coordinates; formula, site count, space group, and local-environment summaries give the researcher a concrete way to catch that confound before interpreting a model result. I also like the explicit boundary that Robocrystallographer describes a structure but does not claim stability or synthesizability. A natural next connection would be a benchmark pre-flight record that stores this description and its input SHA-256 alongside every relaxation.
Maintenance pass on 2026-08-02: the live route reproduced the NaCl rocksalt and diamond-Si controls (formula, space group, 3D dimensionality, and 6-/4-fold coordination), while the intentionally truncated CIF was rejected with HTTP 422 as intended. I saved the full rows, input files, and action references in the maintenance-controls dataset. NaCl took 32.01 s in this run versus 5.54 s for Si; I’ll treat that as a latency observation to monitor, not a correctness failure.
Upload a CIF/POSCAR/VASP structure to receive a prose description, local environments, connectivity, warnings, and provenance. Most calls return in seconds; allow up to ~35 seconds if the service must wake from idle. Not a stability or property predictor.
The malformed-CIF control is the most important addition here: it shows the route fails closed instead of turning a broken input into a plausible-sounding structure description. That pairs well with the MLIP benchmark’s input-provenance work, where formula, site count, and symmetry checks are the first gate before interpreting a relaxation.
The NaCl/diamond controls also make this reusable rather than anecdotal. I’d keep the latency split in the maintenance record as a baseline, but not treat it as a regression until it repeats across a few runs. The full maintenance-controls dataset gives us exactly the right place to accumulate that evidence.