Multi-Modal Foundation · facebook
UMA
A family of universal atomic graph neural networks for chemistry and materials science simulations.
What UMA is
UMA (Universal Models for Atoms) is a family of atomistic neural network models developed by Facebook (Meta FAIR Chemistry). The project aims to provide unified machine learning representations for atomic systems spanning multiple domains in chemistry, physics, and materials engineering.
Trained across five open-science simulation datasets containing billions of atoms, UMA models serve as universal foundational representations for atomic structures. They are accessible via the FAIRChem framework as calculator interfaces for atomistic simulations.
How it works
UMA is built on a mixture-of-linear-experts graph network architecture designed to scale across diverse atomic environments. The models are trained on the combined data from five simulation datasets: OC20, ODAC23, OMat24, OMC25, and OMol25.
Key facts
- Architecture
- Mixture-of-linear-experts graph network
- Training data
- OC20, ODAC23, OMat24, OMC25, and OMol25
- Software interface
- FAIRChemCalculator (fairchem)
- ArXiv ID
- 2506.23971
What you can use it for
- Atomistic simulations
- Driving molecular and materials simulations through the FAIRChemCalculator interface across diverse atomic compositions.
- Materials and molecular property modelling
- Evaluating atomic structures spanning domains covered by OC20, ODAC23, OMat24, OMC25, and OMol25 datasets.
Run it here
We list this model but cannot run it yet: excluded (gated): Gated with manual approval; FAIR Chemistry Acceptable Use Policy must flow down..
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Trained on billions of atoms across five large-scale open simulation benchmarks covering inorganic materials and molecules.
- Offers multiple model scales and checkpoints, including small (uma-s) and medium (uma-m) variants.
- Integrates into the open-source fairchem ecosystem via standard calculator interfaces.
Watch out for
- The initial uma-s-1 checkpoint contains a known extensivity bug and is retained only for reference.
- Requires using the fairchem library and designated calculator classes for inference.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · other
This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Questions
Can I use UMA commercially?
Check the licence before commercial use. This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Which checkpoint should be used instead of uma-s-1?
Because uma-s-1 has a known extensivity issue, newer checkpoints such as uma-s-1.1 or uma-s-1.2.1 should be used.
What datasets were used to train UMA?
UMA models were trained on the entirety of five datasets: OC20, ODAC23, OMat24, OMC25, and OMol25.
How can UMA be run in atomistic simulation workflows?
UMA can be integrated directly into simulation workflows using the FAIRChemCalculator provided in the fairchem repository.
How to cite it
Brandon M. Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso-Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, John R. Kitchin, Daniel S. Levine, Kyle Michel, Anuroop Sriram, Taco Cohen, Abhishek Das, Ammar Rizvi, Sushree Jagriti Sahoo, Zachary W. Ulissi, C. Lawrence Zitnick. UMA: A Family of Universal Models for Atoms. arXiv:2506.23971 [cs.LG], 2025.
Written from UMA's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.