Open Materials 2024 (OMat24)

Open Materials 2024 (OMat24) is a large-scale inorganic materials dataset released by Meta's FAIR Chemistry team. It contains more than 110 million single-point density functional theory (DFT) calculations, built around non-equilibrium structures sampled deliberately for structural and compositional diversity: the training split holds roughly 101 million structures drawn from about 3.2 million distinct bulk materials, with a separate validation split of around 1 million. Each entry carries total energy, atomic forces and stress, which makes it suited to training machine-learning interatomic potentials rather than to text retrieval.

Access is through the Hugging Face dataset hub, where the data is served as ASE-compatible LMDB files (.aselmdb) packaged in tar.gz archives, with a 1 million subsplit provided for quick iteration. The fairchem library (MIT licensed, on GitHub) handles loading and training workflows, and the accompanying EquiformerV2 checkpoints trained on OMat24 top the MatBench-Discovery leaderboard for predicting ground-state stability and formation energies.

For RAG and AI work this is a specialist scientific resource, best used to train or evaluate foundation models and interatomic potentials for materials discovery, or to supply verifiable computed properties to a chemistry knowledge graph or structured retrieval layer rather than as chunked prose. On licensing: the dataset is released under CC BY 4.0, so commercial use is permitted with attribution and there is no territory exclusion. The important caveat is that the accompanying pretrained models ship under a separate Meta FAIR research licence that restricts commercial use, so do not assume the model weights share the open data terms. Compared with our sibling entries, OMat24 covers inorganic bulk materials where OMol25 covers molecular systems, and it complements LeMaterial's harmonised merge of existing databases by contributing fresh, diversity-sampled DFT calculations rather than re-packaging prior ones.