UMA

UMA — the Universal Model for Atoms — is Meta FAIR's machine-learning interatomic potential, released in 2025 alongside the Open Molecules 2025 (OMol25) dataset of 100M+ DFT-computed molecular snapshots. It is trained on more than 30 billion atoms drawn from Meta's open-science simulation datasets of the past five years (Open Catalyst, OpenDAC, Open Materials 2024, OMol25), spanning molecules, materials and catalysts. The pitch is a single foundational potential accurate enough to model atomic interactions across chemistry, usable directly or as a base for fine-tuning.

Meta positions UMA as open science: model downloads and a paper are published, and partners including Lawrence Livermore National Laboratory and the CuspAI AI Materials Foundry are building on it. As an interatomic potential rather than a chatbot-style model, its users are computational chemists running molecular dynamics and high-throughput screening — the audience that previously stitched together narrow, system-specific potentials.

• Provider-reported (Meta): sets new standards for modeling atomic interactions across materials and molecules
• No standardized public benchmark scores found in sources checked