Meta's Universal Model for Atoms (UMA) Demo
This is the UMA! It is a large mixture-of-linear-experts graph network model trained on billions of atoms across five open-science simulation datasets released by the FAIR Chemistry team over the past 5 years. If you give it an input structure and which task you're interested in modeling in, it will output the energy, forces, and stress which you can use for a molecular simulation! Try one of these examples to see what it can do.
When you've run your first UMA simulation, click on the next tab above to explore the UMA model in more detail and see how it works across many different domains/examples!
These next examples are designed to show how the UMA model can be applied to different domains and types of structures, and how different model inputs can impact the results! Each UMA task corresponds to a different domain of chemistry, and a different Density Functional Theory (DFT) code and level of theory that was used for the training data.
As you try each one, look at how the inputs change below, and the simulation outputs change on the right. Feel free to try changing some of the settings below and re-run the simulations to see how the results can vary.
Once you understand how the UMA model can be applied to different types of molecules and materials, the final tab above will help you try it out with your own structures!
As the final step of the demo, try running your own structure through the UMA model!
To use a custom input structure with this demo:
- Request gated model access. Requests for model access are typically processed within a matter of minutes.
- Login to Hugging Face using the "Sign in with Hugging Face button" in the inputs section.
- Then upload a structure file below and click run!
- Note that uploaded structure will be stored by this demo to analyze model usage and identify domains where model accuracy can be improved.
- If you get a redirect error when logging in, please try visiting the direct demo url in a new tab (https://facebook-fairchem-uma-demo.hf.space/) and try again
- Your structure should be in a format supported by ASE 3.25, including .xyz, .cif, .pdb, ASE .traj, INCAR, or POSCAR.
- Your structure should either have periodic boundary conditions (PBC) all True, or all False. Support for mixed PBC may be added in the future.
Learn more about UMA
- UMA models predict motion and behavior at the atomic scale, ultimately reducing the development cycle in molecular and materials discovery and unlocking new possibilities for innovation and impact.
- UMA models are based on Density Functional Theory (DFT) training datasets. DFT simulations are a commonly used quantum chemistry method to simulate and understand behavior at the atomic scale.
- UMA models are large mixture-of-linear-experts graph networks models trained on billions of atoms across five open-science simulation datasets released by the FAIR Chemistry team over the past 5 years. This demo uses the small UMA model with 146M total parameters, 32 experts, and 6M active parameters at any time to predict across all of these domains.
Read the UMA paper for details or download the UMA model and FAIR Chemistry repository to use this yourself!
- The UMA model paper contains rigorous accuracy benchmarks on a number of validation sets across chemistry and materials science. As of model release the UMA model was at or near the state-of-the-art for generalization machine learning potentials. Read the UMA paper for details.
- Rigorously predicting when AI/ML models will extrapolate (or not) to new domains is an ongoing research area. The best approach is to find or build benchmarks that are similar to the questions you are studying, or be prepared to run some DFT simulations on predictions to validate results on a sample of structures that are relevant to your research problem.
- Many important technological challenges, including developing new molecules to accelerate industrial progress and discovering new materials for energy storage and climate change mitigation, require scientists and engineers to design at the atomic scale.
- Traditional experimental discovery and design processes are extremely time consuming and often take decades from ideation to scaled manufacturing.
- Meta's Fundamental AI Research Lab (FAIR) is drastically accelerating this process by developing accurate and generalizable machine learning models, building on work by academic, industrial, and national lab collaborators.
- Each simulation you see would take days or weeks using a traditional quantum chemistry simulation, but UMA can do it in seconds or minutes!
- Examples in the demo are cached ahead of time so they should load right away, but if you run a custom simulation you'll see a progress bar while the simulation runs.'
- While UMA represents a step forward in terms of having a single model that works across chemistry and materials science, we know the model has limitations and weaknesses and there will be cases where the model fails to produce an accurate simulation.
- Ab-initio calculations are not perfect. You should always consider the limitations of the level of theory, the code, and the pseudopotentials.
- The model code is available on github at FAIR chemistry repo
- This demo builds on a number of great open source packages like gradio_molecule3d, 3dmol.js, ASE, and many others!
Debugging
- Every calculation uses a pool of GPUs to process simulations for all current users. You can achieve much higher performance with a dedicated GPU and MD-mode enabled.
- Most simulation should finish within a few minutes. Example results are cached, and if you are running a custom simulation you can follow the progress bar
- if you don't see progress or the simulation takes more than ~5min, probably there was an error and please try submitting again.
- If you notice any issues please submit them as issues on the FAIR Chemistry GitHub.
- If you notice a redirect error when clicking the login to Huggingface button, open a new tab and go to the direct demo url (https://facebook-fairchem-uma-demo.hf.space/) and try again!
Simulation inputs
1. Input structure (example or upload your own!)
To use your own structures, you need access to the gated UMA model repository and you need to login with the button above. See the final tab above '3. Try UMA with your own structures!' for more details and debugging steps!
Note that uploaded structure will be stored by this demo to analyze model usage and identify domains where model accuracy can be improved.
2. Choose the UMA Model Task
OMol25 comprises over 100 million calculations covering small molecules, biomolecules, metal complexes, and electrolytes.
Relevant applications: Biology, organic chemistry, protein folding, small-molecule pharmaceuticals, organic liquid properties, homogeneous catalysis
Level of theory: wB97M-V/def2-TZVPD as implemented in ORCA6, including non-local dispersion. All solvation should be explicit.
Additional inputs: total charge and spin multiplicity. If you don't know what these are, you should be very careful if modeling charged or open-shell systems. This can be used to study radical chemistry or understand the impact of magnetic states on the structure of a molecule.
Caveats: All training data is aperiodic, so any periodic systems should be treated with some caution. Probably won't work well for inorganic materials.
OMC25 comprises ~25 million calculations of organic molecular crystals from random packing of OE62 structures into various 3D unit cells.
Relevant applications: Pharmaceutical packaging, bio-inspired materials, organic electronics, organic LEDs
Level of theory: PBE+D3 as implemented in VASP.
Additional inputs: UMA has not seen varying charge or spin multiplicity for the OMC task, and expects total_charge=0 and spin multiplicity=0 as model inputs.
OMat24 comprises >100 million calculations or inorganic materials collected from many open databases like Materials Project and Alexandria, and randomly sampled far from equilibria.
Relevant applications: Inorganic materials discovery, solar photovoltaics, advanced alloys, superconductors, electronic materials, optical materials
Level of theory: PBE/PBE+U as implemented in VASP using Materials Project suggested settings, except with VASP 54 pseudopotentials. No dispersion.
Additional inputs: UMA has not seen varying charge or spin multiplicity for the OMat task, and expects total_charge=0 and spin multiplicity=0 as model inputs.
Caveats: Spin polarization effects are included, but you can't select the magnetic state. Further, OMat24 did not fully sample possible spin states in the training data.
OC20 comprises >100 million calculations of small molecules adsorbed on catalyst surfaces formed from materials in the Materials Project. It was updated to total energy predictions for the UMA release.
Relevant applications: Renewable energy, catalysis, fuel cells, energy conversion, sustainable fertilizer production, chemical refining, plastics synthesis/upcycling
Level of theory: RPBE as implemented in VASP, with VASP5.4 pseudopotentials. No dispersion.
Additional inputs: UMA has not seen varying charge or spin multiplicity for the OC20 task, and expects total_charge=0 and spin multiplicity=0 as model inputs.
Caveats: No oxides or explicit solvents are included in OC20. The model works surprisingly well for transition state searches given the nature of the training data, but you should be careful. RPBE works well for small molecules, but dispersion will be important for larger molecules on surfaces.
ODAC23 comprises >10 million calculations of CO2/H2O molecules adsorbed in Metal Organic Frameworks sampled from various open databases like CoreMOF. It was updated to total energy predictions for the UMA release.
Relevant applications: Direct air capture, carbon capture and storage, CO2 conversion, catalysis
Level of theory: PBE+D3 as implemented in VASP, with VASP5.4 pseudopotentials.
Additional inputs: UMA has not seen varying charge or spin multiplicity for the ODAC task, and expects total_charge=0 and spin multiplicity=0 as model inputs.
Caveats: The ODAC23 dataset only contains CO2/H2O water absorption, so anything more than might be inaccurate (e.g. hydrocarbons in MOFs). Further, there is a limited number of bare-MOF structures in the training data, so you should be careful if you are using a new MOF structure.