Welcome to atomistica.online

Articles & Tutorials

Open Molecules 2025 and Universal Model for Atoms

OMol25 and UMA: A New Era for Atomistic Modeling

When people think of Meta, social media likely comes to mind. But behind the scenes, Meta’s Fundamental AI Research (FAIR) team has been reshaping scientific research, including computational chemistry and materials science. In 2025, this shift became undeniable with the open release of OMol25 and UMA: two groundbreaking contributions to molecular modeling now available through Atomistica.online 2025 as well.

These two tools tackle some of the biggest challenges in the field, massive, high-quality datasets and general-purpose machine learning potential, and make them accessible to researchers, educators, and students alike.

OMol25: A Dataset on an Unprecedented Scale

OMol25 is more than just a dataset, it’s a massive scientific effort. It includes results from over 100 million quantum-mechanical calculations, all performed at the high-fidelity ωB97M-V/def2-TZVPD level of theory. The dataset spans 83 elements and required 6 billion core-hours to complete, making it one of the largest DFT-based datasets ever assembled.

Unlike older datasets such as QM9 or SPICE, OMol25 covers:

  • Large biomolecules, including protein-ligand and nucleic acid complexes.
  • Electrolyte systems, including ILs and redox-active clusters.

  • Transition metal complexes, generated with GFN2-xTB and AFIR sampling.

Even more impressively, OMol25 recalculates and unifies existing benchmarks like ANI-2X, SPICE, and Transition-1x at a consistent DFT level. The result? A dataset that acts as a “molecular ImageNet”—a foundation for training general-purpose, high-accuracy models.

UMA: Foundation Models for Molecular Simulation

Alongside OMol25, Meta released the UMA family of potentials, fast, general-purpose models for predicting molecular and material properties. UMA models are trained on OMol25 plus other FAIR datasets like OC20, ODAC23, and OMat24. This was made possible by a clever Mixture of Linear Experts (MoLE) architecture, enabling the model to generalize across chemically diverse systems while keeping inference fast. Each UMA model undergoes two training stages: Direct-force training, followed by Conservative-force fine-tuning, ensuring energies and forces are physically sound—ideal for use in molecular dynamics and energy-sensitive tasks.

Even the UMA-small model, now publicly available, achieves <1 kcal/mol MAE, rivaling DFT accuracy while operating at a fraction of the computational cost.

UMA models are not just tools, they’re starting points, pre-trained “backbones” ready to be fine-tuned for specialized tasks like reactivity prediction, solvation energy, or conformer search. Think of them as the GPT of molecular simulation.

Try It Now on Atomistica.online 2025

We’re proud to announce that OMol25 and UMA models are now available on Atomistica.online through a simplified browser-based interface. No installations, no coding, just powerful modeling at your fingertips.

Here’s what you can do:

  • Explore OMol25-based predictions in seconds.
  • Use UMA for quick and efficient geometrical optimizations.
  • Combine with our Ionic Liquids Generator or ML predictors for targeted material design.

  • Access g-xTB for fast quantum chemistry, built directly into the platform.

Atomistica.online 2025 also features:

  • ORCA batch input file generator
  • Multiwfn-based surface property analyzer
  • Live analytics dashboard for real-time usage stats
  • Curated models for ionic liquid density and viscosity prediction (IonIL-IM series)
  • 3Dmol.js-powered visualization for direct structure interaction
  • and many more tools…

And it’s 100% free for academic and educational users.