GNoME
GoogleGNoME is Google DeepMind's Graph Networks for Materials Exploration effort for inorganic-crystal discovery, released as a database of more than 520,000 materials within 1 meV per atom of the convex hull, alongside model definitions for the GNoME and NequIP architectures and Colab notebooks for exploring the data. The repository publishes architectures and discovered structures rather than trained weights.
The Apache-2.0 label covers the notebooks and code; the database is CC BY-NC 4.0 and no trained checkpoint is distributed. Verified 2026-09-10 via the GitHub API and the repository README.
Openness
1 high confidence- weights
- closed(model definitions and architecture configuration only
- data
- closed(the models were trained on a 2018 Materials Project snapshot that is not released as a corpus
- code
- partial(Apache-2.0 model definitions and exploration Colabs
- license
- proprietary(no weights are distributed, so no weights licence exists
The Apache-2.0 repository label is a code licence, not a weights licence. What GNoME released is the discovered-materials database and the architectures behind it; there is no downloadable trained model, so the first rung of the ladder applies whatever else was published.
- https://raw.githubusercontent.com/google-deepmind/materials_discovery/main/README.md recorded 2026-09-10
Offers the dataset and Colab notebooks for download, says the repository provides model definitions for the GNoME and NequIP architectures with basic configuration parameters, lists model training colabs and configs as upcoming, and splits the licence between Apache-2.0 for code and CC BY-NC 4.0 for the database.
Adoption
2 medium confidenceNo Hugging Face or package artifact exists — nothing is distributed to band — so 1,235 GitHub stars are the last-resort signal.
- https://api.github.com/repos/google-deepmind/materials_discovery recorded 2026-09-10
stargazers_count: 1235; archived: false; fork: false.
Capability
1 high confidenceNothing here can be adapted, because no trained model ships and the training colabs are still upcoming. Discovery in the paper came from a substitution-and-filter pipeline around the model rather than from the model generating structures, so the generative attribute is not counted.
- https://raw.githubusercontent.com/google-deepmind/materials_discovery/main/README.md recorded 2026-09-10
Describes GNoME as a simple message-passing architecture trained to predict energies, provides model definitions and exploration Colabs only, and lists model training colabs and configs as upcoming.
Verified 2026-09-10