Chemistry & materials · 29 November 2023
An order-of-magnitude expansion in stable crystals
GNoME discovered 2.2 million crystal structures below the previous convex hull, of which 381,000 are new stable materials. Of the stable structures, 736 had already been independently realized in experiment.
- Field
- Chemistry & materials
- Date
- 29 November 2023
- AI system
- GNoME
- Institution
- Google DeepMind
- Published in
- Nature
- Identifier
- 10.1038/s41586-023-06735-9
Discovery of inorganic crystals has long been bottlenecked by expensive trial-and-error, whether in the laboratory or in first-principles calculation. Graph Networks for Materials Exploration (GNoME), published in Nature on 29 November 2023, trains graph neural networks at scale and uses them, in rounds of active learning, to propose candidate crystals whose energies are then checked with density functional theory.
Starting from about 48,000 stable crystals identified in earlier studies, the models discovered more than 2.2 million structures stable with respect to previous catalogues. Because newly found materials compete with one another for stability, the updated convex hull contains 381,000 new entries, for a total of 421,000 stable crystals — an order-of-magnitude expansion, the authors write, in stable materials known to humanity.
The paper also reports a direct experimental check: of the stable structures, 736 had already been independently realized in the Inorganic Crystal Structure Database. Final models predicted energies to 11 meV per atom, with a hit rate above 80% when structure was known. The structures have been shared with the research community, including through the Materials Project.