Open data and universal models for atomic systems¶
FAIR Chemistry develops open datasets and machine-learning models for molecules, materials, and catalysts. UMA brings these domains together in one pretrained model while preserving each dataset’s level of theory.
pip install fairchem-coreFrom open datasets to UMA¶
UMA learns from more than 500 million density functional theory calculations. At inference time, you choose a task that matches your chemistry domain and the corresponding level of theory.
Surface reactions, adsorption, and catalyst design.
Bulk materials, phonons, and elastic properties.
Conformers, reactions, and electronic properties.
Packed organic molecules in crystal structures.
CO₂ and H₂O adsorption in metal-organic frameworks.
UMA uses a learned task embedding to apply one pretrained model across these
domains without mixing their reference methods. Start with
uma-s-1p2p1, the fastest current UMA model with
state-of-the-art accuracy on most supported benchmarks.
See what UMA can do¶
Choose your next step¶
Set up fairchem-core and request access to UMA checkpoints.
Run a molecular calculation and relax an inorganic crystal.
Find the right task and workflow for your scientific problem.
Understand UMA tasks, inputs, architecture, and limitations.
Scale from ASE calculations to training and batched inference.
Watch introductory videos and technical presentations.
Copyright © Meta Platforms, Inc | Terms of Use | Privacy Policy