zer0contextlost/crystal-lab-data

Crystal Lab open data

Crystal structures proposed by an AI model, screened with a machine-learned potential, and the most promising ones re-checked with density functional theory (DFT). Everything is published, including the failures.

AI generators now propose crystals much faster than anyone checks them with physics. This dataset is a small, open record of what happens when you do check: most candidates that look stable to a machine-learned model turn out to be known compounds, or less stable than their real competitors, once the competitors are taken seriously.

Built 2026-11-10. Rebuilt nightly while the pipeline runs. Mirrors: Hugging Face; raw Quantum ESPRESSO calculations parsed and searchable on NOMAD (snapshot of 2026-11-01), plus the #1856 SrCaCdP2 phonon runs (upload of 2026-11-06) and all #2571 BaCdGePd2 runs (upload of 2026-11-10).

Cite: the three NOMAD uploads (102 calculations) form a frozen snapshot with DOI 10.17172/nomad.5gmq-24hf. This repository keeps changing; the DOI does not.

2026-11-10: first candidate through every check

#2571 BaCdGePd2 (P4/nmm, generated 2026-10-03 from a surprise prompt, where the model picks the formula itself) is the first candidate to pass every step we run:

• DFT (PBE): 0.003 eV/atom above the hull built from 1,281 Alexandria entries in Ba-Cd-Ge-Pd; the nearest competing mix is BaGePd + CdPd. That is inside normal PBE error, so read it as "on the hull, within noise", not "stable".
• Phonons: no imaginary modes (2x2x1 supercell, 40 atoms, 10 displacements, 78 h).
• Prior art: no report of the composition in 27 databases and paper indexes that answered on 2026-11-10 (AFLOW, CMR, MPDD, MPOD and Matterverse did not respond).

What this does not show: that BaCdGePd2 can be made, that this is its lowest-energy structure (only the generated structure was tested), or anything about behaviour above 0 K. It is a prediction worth a synthesis attempt, nothing more. #1856 SrCaCdP2 also passed DFT and phonons, but OQMD, GNoME and NOMAD already list it, so it is an independent check of an existing prediction.

2026-11-03: stricter novelty rule

A candidate now counts as new only if its composition has not been reported anywhere we check, in any structure: Materials Project, Alexandria, OQMD, COD, MPDS, Materials Cloud, GNoME, NOMAD and about 20 more OPTIMADE databases, plus paper titles and abstracts (OpenAlex, arXiv). Before, novelty meant "no matching structure in Materials Project or Alexandria". The stricter check found that our first DFT pass, #1856 SrCaCdP2, was already predicted with the same structure in OQMD (entry 5879242), and that all five DFT jobs then in the queue were known compositions; those five were taken out of the queue. Every catalogued record was re-checked the same way and the already-reported ones re-labelled known composition (their old catalogue number and name are kept in the novelty record). Known formulas are now also skipped before the generator spends any compute on them. Details in METHODS.md §3 and §7, and in the prior_reports column of dft_results.csv.

What's in it

	count
Structures generated	5837
Kept as "discoveries" (stable by ML, and the composition is not reported in any database or paper we check)	895
ML hull check finished	644
ML hull check at or below 0 eV/atom	1
DFT jobs (passes / fails / still running or queued)	12 (3 / 2 / 7)

How the prompts were chosen, and how each way does

On 2026-11-01 the autopilot switched from random formulas to mostly substitution-based prompts (see METHODS.md). This table compares the sources; "within 0.05" counts ML hull results at most 0.05 eV/atom above the hull. "Skipped as already known" (from 2026-11-03) counts formulas dropped before any ML work because the composition was already reported; control rows are known compounds requested on purpose as a calibration check. New rows need a few nights of hull checks before their numbers mean much.

prompt mode	skipped as already known	structures	new	ML hull checked	median e_hull (eV/atom)	within 0.05	DFT passes
random formula (before 11-01)	0	3713	696	494	0.389	1	0
surprise (before 11-01)	0	1681	101	71	0.115	13	3
substitution	199	265	69	61	0.211	2	0
random	93	78	21	14	0.407	0	0
surprise	148	63	8	4	0.158	0	0
control	0	27	0	0	-	0	0
human	0	10	0	0	-	0	0

ML verdicts: unstable 2108, rock solid 1463, plausible 1307, borderline 776, falls apart 183. Novelty at generation time: skipped 3067, known 977, new composition 895, known composition 893, rediscovered 5.

Files

• structures.jsonl.gz: one JSON object per structure: formula, space group, CIF, how the prompt was chosen (prompt_mode, from 2026-11-01), ML energies and verdict, novelty check (with the matching database IDs when it was already known), ML hull result, DFT summary.
• structures.csv: the same without CIFs, for spreadsheets.
• dft_results.csv: one row per DFT job.
• dft/<id>/: every DFT job: candidate.cif, competitor structures (comp/), Quantum ESPRESSO inputs and outputs for the candidate and each competitor (runs/), result.json, and for passes the properties (props/: band gap, DOS, elastic tensor) and phonon runs (phonon/). .out files are gzipped.
• stats.json: the counts above. build.log: files left out of this build and why.
• METHODS.md: how every number was produced, and what it does and doesn't mean.

How to read the results

• An ML "rock solid" verdict means a low formation energy with MACE. It says nothing about competing phases.
• The ML hull compares against Alexandria near-hull phases computed with the same ML model. pending rows carry only previous_e_hull_eV_atom, a superseded value from before a stricter re-check; don't use it as a result.
• A DFT pass means: with PBE, the candidate is at most 0.025 eV/atom above the convex hull built from its Alexandria competitors. It does not mean the material has been made, or that it can be made. PBE errors of a few tens of meV/atom are normal, and the hull only contains the competitors we checked.
• Phonons are a screen on a small supercell (shortest side ~7 Å) for imaginary modes, not a full study.

Licence and credit

Our data (generated structures, ML and DFT results): CC BY 4.0, credit "Crystal Lab open data, zer0contextlost". Competitor structures come from the Alexandria database (Schmidt et al., CC BY 4.0); Alexandria and Materials Project IDs are given wherever a structure matched them. Pseudopotentials are SSSP 1.1.2 efficiency (Prandini et al.) and are not redistributed here. See ATTRIBUTION.md.

Structures from licensed databases (e.g. CCDC/ICSD) used in any check are never included; such checks are cited by deposition number only.

Corrections and questions: open an issue.

Mirror of https://github.com/zer0contextlost/crystal-lab-data (rebuilt nightly).