Domain DiscoveryInstitutional🔗 source linked

GNoME

DeepMind's graph neural network for materials discovery — has identified 2.2 million new crystal structures, with hundreds already experimentally synthesized.

Last verified: July 2026

What it does

GNoME (Graph Networks for Materials Exploration) is a DeepMind model that applies graph neural networks to predict the stability of novel crystal structures. Rather than evaluating one candidate at a time experimentally, GNoME can screen vast combinatorial spaces of potential materials. Its published results include identifying 2.2 million new stable crystal structures, including 52,000 novel lithium-ion conductors — hundreds of which have since been experimentally synthesized by independent labs.

Access is primarily through DeepMind’s associated databases and published datasets, not a consumer product.

Best for

Materials science and chemistry researchers focused on crystal structure prediction, battery materials, and solid-state chemistry. Particularly relevant if you’re working on energy storage, lithium-ion conductors, or novel functional materials. The experimentally synthesized results give this model a track record beyond pure prediction.

Pricing

Research access via DeepMind and associated databases. Not a commercial SaaS product — access paths include the published datasets and, in some cases, collaboration with DeepMind.

Strengths

  • Scale of discovery is unprecedented: 2.2 million predicted stable crystal structures
  • Hundreds of predicted structures experimentally validated, giving it a real-world track record
  • Focused specifically on materials stability — a well-defined task with clear evaluation metrics
  • Results published in Nature with accompanying dataset releases

Limitations

  • Not a self-service tool — access paths are research-oriented (datasets, published APIs), not consumer-grade
  • Predicts thermodynamic stability, not functional performance — knowing a crystal is stable doesn’t tell you its conductivity, optical properties, or synthesis difficulty directly
  • Computational validation with DFT (density functional theory) is recommended before investing experimental resources in any specific predicted structure
  • Primarily useful for inorganic/solid-state materials; does not cover organic or biological materials domains

How it compares

vs. Key difference
AlphaFold AlphaFold predicts protein structure; GNoME predicts crystal material structure — different domains, similar AI approach
Traditional DFT screening GNoME dramatically accelerates structure screening; DFT remains the validation step after GNoME narrows the candidate space

Field Guide: Materials Science (coming soon)