GNoME: Graph Networks for Materials Exploration

License: Paper copyright held by authors (Nature); released data at Materials Project · Updated 2026-03-04

GNoME trains graph networks on public crystal databases (Materials Project, OQMD and others), generates candidate structures through structural and compositional pipelines, verifies them with DFT calculations, and feeds the results back into training - an active-learning flywheel.

Reported results: 2.2M predicted structures, 381k new stable entries added to the convex hull, raising known stable materials from ~48k to ~421k. External labs have since synthesized hundreds of the predicted compounds.

Why it matters for CAE engineers: It is the clearest demonstration that learned interatomic potentials plus first-principles verification can search design spaces far beyond brute-force enumeration - the same pattern applies to alloy, catalyst and battery material screening.

Tags: Materials DiscoveryGNNDFTCrystalActive Learning
Related entries