Source-linked AI summary
Probing the limits of forward screening and inverse design of crystal materials
Zhilong Song, Zijun Pang, Lixue Cheng
TL;DR
The limits of crystal screening and inverse design remain unmeasured because public databases identify the same crystals differently and leave many crystal–property pairs empty. The paper structurally integrates 55 databases and uses CGTNet to complete missing properties. The resulting reference shows screening can find many unmeasured candidates, while much of a generator's apparent novelty disappears against the full database.
Problem
The reach of screening and the novelty of inverse design have not been measured against a crystal-level reference spanning databases with different identifiers and incomplete property records.
Source: §1 · p. 2Method
The authors matched crystals structurally across 55 databases, retained value provenance, and used task-trained CGTNet models to complete missing properties.
Source: §1 · p. 3, §5.1 · p. 18, §5.3 · p. 18Results
Against the integrated database, screening revealed tens of thousands of candidates with uncomputed target properties, while most stiffness-generator proposals had recorded compositions.
Source: §3 · p. 16Takeaways & Limitations
An integrated, completed database both widens the known-crystal search space for screening and provides the reference for judging generated crystals as new.
Source: §4 · p. 17On 64 of 73 tasks with measurable between-source spread, CGTNet's error exceeded the median spread, so completed values are not uniformly reliable.
Source: §3 · p. 16
Abstract
from arXiv · showhide
Crystal materials are designed either by screening known crystals for a target property or by generating new crystals with that property. Both routes are judged against public databases of known crystals, but these databases name the same crystal differently. Several materials announced as new by recent generative and autonomous discovery campaigns have since proved to be known phases or close variants of them. The limits of both routes remain unmeasured. Here we probe them by matching the crystals of 55 databases by structure into one self-extending database of 14.6 million crystals and 131.8 million property values, each traced to its source. Only 4.5% of its crystal-property pairs carry a value. To fill the empty pairs, we developed CGTNet, an equivariant graph transformer with state-of-the-art accuracy on several tasks of two public benchmarks. Trained on each task, it forms a library of property models whose predictions we release. The reach of forward screening has been underestimated. Known crystals hold tens of thousands of ultra-incompressible candidates without a recorded bulk modulus, and a post hoc audit of synthesis evidence finds among them ReN2, whose bulk modulus was measured at 428 GPa. The novelty claimed for inverse design has been overestimated. At an equal verification budget, a stiffness-conditioned generator returned more stiff hits than screening (100 versus 55), but all contained osmium and 82 repeated recorded compositions. Integrating and completing the databases of known crystals therefore moves the limits of both routes: it widens what screening can reach and sets the reference against which generated crystals are judged new.
1. INTRODUCTION
The limits of screening and inverse design had not been tested against a unified crystal-level reference: databases lack shared crystal identifiers, and incomplete references can distort novelty assessments. The paper builds and completes such a database to measure what both routes can reach.
1. INTRODUCTION
- Public databases rarely share crystal-level keys, while formula matching confuses polymorphs and property columns may represent different physical quantities.Consequently, screening one database misses known crystals and properties, and labels cannot be checked across sources.Source: §1 · p. 2
- Prior novelty assessments found known modifications among GNoME predictions [31], while MatterGen's flagship compound was a 1972 phase already in its training set [34].Source: §1 · p. 2
- 14,590,485 crystals and 131.8 million values were assembled by matching structures across 55 databases while retaining each value's source and method.The database grows when an incoming entry matches no crystal already held.Source: §1 · p. 3
- CGTNet's error on each of four JARVIS-DFT benchmark properties is below the best published value, with a denser neighbour graph for one band gap.The task-trained models are released with declared domains and held-out accuracy.Source: §1 · p. 3
- 53,813 ultra-incompressible candidates lacked recorded bulk moduli; an audit retained 62, including ReN2, measured at 428 GPa, while a generator yielded 100 stiff hits versus screening's 55.All generated hits contained osmium and 82 had recorded compositions; the best screening hit was stiffer than the best generated one.Source: §1 · p. 3
2. RESULTS
The results examine how integrating known-crystal databases and completing sparse property labels with CGTNet affect screening candidates and assessments of generated-crystal novelty. At equal verification budgets, generation produced more stiff hits, while screening found a higher maximum modulus and broader chemical coverage among synthesized crystals; novelty estimates varied with the comparison method.
2.1. Integration of known crystals
- The integrated database contains 14,590,485 distinct crystal structures and 131,824,881 property values; admitting unmatched entries added 6,365,347 structures, a 77.4% increase over the starting pool.Each entry was matched by authoritative identifiers or structural fingerprints, not chemical formula, and retained its source and method provenance.Source: §2.1 · pp. 3–4
- Only 4.52% of the integrated crystal–property matrix is populated, and 22.9% of crystals carrying values draw them from multiple databases, so no single source supplies their full labeled pool.Of 151 properties with values, 97 cover less than 1% of the pool and 85 are supplied by only one database.Source: §2.1 · p. 6
2.2. Property completion with CGTNet
- On JARVIS-DFT, CGTNet beat the best of at least 41 published results on all four properties (formation-energy error 23.7 vs 25.0 meV atom−1); on Matbench it led exfoliation energy and ranked second in logarithmic shear modulus.The reported JARVIS errors were 23.7 vs 25.0 meV atom−1 for formation energy, 23.4 vs 24.0 meV atom−1 for total energy, and 106.2 vs 109.3 and 223.3 vs 236.5 meV for OPT and MBJ band gaps. Matbench settings and scoring differed from published protocols, and the exfoliation-energy margin was within fold-to-fold spread.Source: Table I · p. 9, §2.2 · p. 9, Table II · p. 10
- Across 91 trained tasks, CGTNet added 694,947,742 predictions; median held-out R2 was 0.668, and semilocal bulk-modulus coverage rose from 0.20% to 93.2%.The holdout selected each checkpoint, making reported R2 values upper bounds; for 64 of 73 completed tasks with measurable cross-source spread, held-out mean absolute error exceeded the median spread.Source: §2.2 · p. 10
2.3. Limits of forward screening
- The audited screen found 62, 139 and 379 synthesized candidates across three targets, including ReN2 with measured bulk modulus 428 GPa versus a 349 GPa prediction; no synthesized candidate remained for T2.The T2 barrier criterion did not rank measured room-temperature conductivity among 48 crystals (Spearman ρ = −0.03).Source: §2.3 · pp. 11, 13
2.4. Limits of inverse design
- Against all 55 sources, the share of generated crystals unmatched by reduced composition ranged from 0.2% for CrystaLLM to 91.1% for GNoME, but GNoME’s high share was qualified by its element-rich compositions.For GNoME, 87.4% of its released set carried four or five elements, compared with 29.2% of the database; the authors argue novelty claims should use reduced-composition matching rather than looser criteria.Source: §2.4 · pp. 13, 15
- 70.8% of CrystaLLM structures versus 89.4% of random-enumeration structures were absent by fingerprint, but 499 of 500 versus 410 of 492 reused database compositions.The authors recommend reporting novelty under each criterion against a reference set that includes aggregating and specialized databases.Source: §2.4 · p. 15
- 100 versus 55 hits favored generation, while screening reached 433.8 GPa versus 375.4 GPa and, among made crystals, spanned 39 versus five chemical systems.The comparison used the top 100 from each route, verified with MatterSim; none of the four pre-registered outcomes fit. Route A’s five chemical systems all contained Os, and 82 generated structures had compositions already in the database.Source: §2.4 · p. 15
- The completion model predicted all 1,074 generated structures above 250 GPa, and its top 100 were verified hits; whether its ranking improved on generator conditioning was untested.Lower-ranked generated structures were not verified; the database still determines which generated crystals are new.Source: §2.4 · p. 15
3. DISCUSSION
Integrating and completing crystal databases shifts screening and inverse-design limits: screening finds candidates without recorded properties, while novelty claims depend on the reference set and identity criterion. Model quality, incomplete task coverage, and machine-learned-potential verification bound these findings.
3. DISCUSSION
- Screening reached tens of thousands of candidates without recorded target properties, including ReN2, whose measured bulk modulus confirms it; generation novelty shrinks against the full database.The conclusions depend on a completed database, while hit verification used a machine-learned potential and DFT verification remains follow-up work.Source: §3 · pp. 16–17
- The crystal–property matrix is 4.52% full, and CGTNet’s held-out error exceeds the median source spread for 64 of 73 tasks with measurable spread.Identity and synthesis audits changed which candidates counted as never computed or made; the Li-ion target retained no synthesized candidate, while the other three retained tens to hundreds, including ReN2.Source: §3 · p. 16
- At an equal verification budget, generation produced more hits per verification, but screening reached higher bulk modulus and, among synthesis-confirmed crystals, covered 39 chemical systems.Generated hits came from five osmium-containing chemical systems and were mostly of recorded composition.Source: §3 · p. 16
- The model library supports screening and scoring generated crystals, but chemistry coverage must be checked separately and the reported 91 models are a snapshot of 140 tasks.The library is expected to grow as remaining tasks are trained and as admitted sources enlarge training sets.Source: §3 · p. 17
4. CONCLUSION
The paper integrates 55 databases by crystal structure, completes missing property values with CGTNet, and builds a reusable property-model library. This extends screening’s reach while providing a broader reference for judging generated-crystal novelty.
4. CONCLUSION
- An integrated, completed database moves the limits of both screening and inverse design by extending what screening can find and setting a reference for novelty.The database’s property models can also score generative outputs.Source: §4 · p. 17
- A post hoc synthesis-evidence audit identified synthesized ultra-incompressible ReN2 among known crystals lacking a recorded value.Its measured bulk modulus is 428 GPa.Source: §4 · p. 17
- At an equal budget, the generator returned more hits, but they were osmium compounds of mostly recorded composition; screening’s best hit was stiffer.Source: §4 · p. 17
5. METHODS
The methods resolve crystal identity across sources by identifiers and structural fingerprints, preserve value provenance, complete properties with task-specific CGTNets, and compare screening and generation under fixed criteria. The release includes an open database and associated predictions, with some data and model weights restricted by source licences.
5.1. Database construction and crystal deduplication
- Each source adapter retains native identifiers and values with units and methods; properties map to a registry only when they denote the same physical quantity.Structures are kept where licences permit.Source: §5.1 · p. 17
- Entries match first by authoritative cross-database identifier, then by fingerprint, never formula, so polymorphs remain distinct.Fingerprinting follows primitive-cell reduction; unmatched entries are added, while enumerated framework libraries use a composite key including volume per site.Source: §5.1 · pp. 17–18
5.2. CGTNet and its training
- CGTNet benchmark runs used a standard split and 500 epochs, while Matbench used separate task runs and official folds.The architecture and selected settings were chosen using test error, and the evaluated JARVIS-DFT model was the final epoch.Source: §5.2 · p. 18
5.3. Property completion and the model library
- Completion trains a separate CGTNet for each of 176 selected training sets, with held-out splits by chemical system or, for 23 sets, by formula.The held-out tenth selects checkpoints, making reported held-out accuracy an upper bound; domain gates select structures for prediction, and the library records each task’s settings and metrics.Source: §5.3 · p. 18
5.4. Screening
- Screening applies pre-fixed targets and thresholds, using recorded values at the task’s theory level when available and predictions otherwise.The post hoc audited screen adds rules for synthesis evidence, missing recorded values, training-chemistry coverage, and redox activity.Source: §5.4 · pp. 18–19
5.5. Novelty audit and equal-budget comparison
- For the equal-budget comparison, MatterSim verified the top 100 ranked candidates from 1,074 valid unique MatterGen structures and 13,521,404 pool structures.A hit required bulk modulus of at least 250 GPa and a positive-definite elastic tensor.Source: §5.5 · p. 19
Data Availability
- The open release contains 10,903,588 crystals permitted for redistribution and commercial use, with structures, licences, provenance, and model predictions, but not source property values.Restricted-source checkpoints await preparation of a licence-filtered training subset; figure source data are provided as CSV files.Source: Data Availability · p. 19
Code Availability
- Acquisition adapters, matching and loading pipelines, audits, training and analysis scripts, and figure scripts are publicly available.Source: Code Availability · p. 19