GNoME
DeepMind's graph neural network for materials discovery — has identified 2.2 million new crystal structures, with hundreds already experimentally synthesized.
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 |
Related content
Field Guide: Materials Science (coming soon)