A unified crystal database changes screening and novelty claims
Probing the limits of forward screening and inverse design of crystal materials · arXiv:2610.32678
In plain words
Researchers combined crystal records from 55 databases so they could tell when different entries describe the same structure. They trained models to estimate missing properties for crystals whose structures were already known.
The combined records were sparse, but the models let the team screen for stiff materials. One candidate, ReN₂, had a measured bulk modulus of 428 GPa even though the source databases did not record that value.
This shows how broader records can uncover overlooked candidates, not that every prediction is confirmed.
The research
A screening result depends on which known crystals and property measurements are available. A generated structure can also appear new when the comparison database simply does not contain it.
Can one reference resolve crystal identity across sources and fill enough missing measurements to judge both routes?
The authors match crystals across 55 databases, estimate missing properties, and use the shared reference to reassess screening and generated-material novelty.
They match records by shared identifiers or structural fingerprints, keeping each property’s source and method, then train separate models to estimate missing crystal–property values. Only 4.52% of pairs across the 151 populated properties had recorded values.
The completed reference exposed known crystals without recorded stiffness, including synthesized ReN₂ measured at 428 GPa. It also showed that two ternary samples close in fingerprint-based novelty differed more than eightyfold in composition absence.
The evidence supports broader reference sets and explicit novelty criteria, not a blanket claim that generation or screening wins.
ReN₂ is one audited example, and the screen predicted a bulk modulus 18% below its measured value. Screening and generation hits were checked with MatterSim, a machine-learned potential; verification with density-functional theory was left for follow-up.
What the evidence shows
- Build a shared crystal reference
The paper introduces an integrated crystal database that matches structures across 55 databases and preserves where each property value came from. This shared reference addresses a problem that a chemical formula cannot solve: different crystal structures can have the same formula.
It also lets the authors compare screening of known crystals and claims that generated crystals are new against one resolved collection.
- Estimate missing measurements
Combining databases does not fill in their missing measurements. Of the crystal–property pairs for 151 properties with at least one recorded value, only 4.52% had a value.
The authors developed CGTNet, a model trained separately for each property task, to estimate eligible missing values. Its task-specific domain gate limits predictions to crystals where the property is defined.
- Find overlooked stiff crystals
Using the completed database, the authors screened known crystals for ultra-incompressibility, requiring bulk modulus of at least 250 GPa and shear modulus of at least 150 GPa. They found 53,813 candidates without a recorded bulk modulus.
A later audit found synthesis evidence for 62, including ReN₂, measured at 428 GPa. Its screen prediction was 18% lower, so this is evidence of an overlooked candidate, not exact predictive accuracy.
- Compare novelty tests
The authors also checked how novelty depends on the test used. In two samples consisting entirely of ternary compounds, fingerprint-based novelty differed by 18.6 percentage points.
But 499 of 500 CrystaLLM structures had compositions already in the database, compared with 410 of 492 from random enumeration. A structure can therefore seem absent by one test while its composition is already represented.
- Keep route results in scope
With the same verification budget, generation produced more hits among the top 100 candidates than screening, 100 versus 55. Screening’s best hit was stiffer, however, and all generated hits contained osmium while 82 had recorded compositions.
The verification used MatterSim, with density-functional theory checks still pending. A separate lithium-conductivity check also found that the chosen migration-barrier measure did not rank conductivity among 48 eligible crystals.
Deep
- Original paper authors
- Zhilong Song, Zijun Pang, Lixue Cheng
- Submitted to arXiv
- License
- CC BY 4.0