Emory and I also wrote a higher level summary of the work available here: rdcu.be/eUVIw

@berkeleylab.lbl.gov
@molecularfoundry.lbl.gov @mitcheme.bsky.social

We can't wait to see what the community does with OMol! Don't hesitate to reach out with feedback on the data, models, or paper - we aren't going to submit to a journal until the leaderboard goes up, which means we have time to incorporate community feedback (within reason) 10/10

A special shout out to co-first authors Daniel Levine and Muhammed Shuaibi who moved mountains making OMol a reality. I also want to recognize the substantial and critical contributions of @ewcspottesmith.bsky.social, Michael Taylor, Muhammad Hasyim, and Kyle Michel 9/N

Co-leading OMol with Brandon and Larry was a joy and an honor - as was assembling a world-leading team of scientists from 2 companies, 2 national labs, and 6 universities who were excited to help build an open-source, revolutionary molecular DFT dataset to push science forward 8/N

Right now, OMol data has energy, forces, partial charges, partial spins, and HOMO/LUMO. But we have far more info that we still need to parse and hope to do a battery of GBW postprocessing. Plus we have 10 petabytes of electron densities. Lots more to come! 7/N

And check out the UMA demo (facebook-fairchem-uma-demo.hf.space UMA is trained on OMol + other FAIR Chemistry datasets) - metal complexes at +1 vs +2 correctly optimize to tetrahedral/planar and reduced ethylene carbonate correctly ring-opens while a neutral EC remains stable 6/N

Data, models, & paper are available to download now! 5/N
Paper: arxiv.org/abs/2505.08762
Data + models: huggingface.co/facebook/OMo...

We're also releasing baseline models trained on OMol. To guide future MLIP development, we built novel evaluations on intermolecular interactions, conformers, and charge/spin. We hope to include frequency, ΔG, and TSopt tasks when we put up a public leaderboard in the summer 4/N

OMol was constructed via an unprecedented diversity of methods: MD, ML-MD, RPMD, rattling, Architector, rxn path interpolation, AFIR, optimization, and scaled separation. We also recalculated some previous datasets and did additional sampling/structure generation atop others 3/N

OMol covers 83 elements, a wide range of intra and intermolecular interactions, explicit solvation, reactive structures, conformers, charges -10 to 10, 0-10 unpaired electrons, and 2-350 atoms per snapshot. It required >6B CPU hrs, 10x more than any prev MLIP training dataset 2/N