FAIRChem v2 Unifies Atomistic Simulation With UMA

Density functional theory is central to predicting atomic energies, forces and material properties, but its computational cost limits the size and number of practical simulations. Meta FAIR Chemistry developed FAIRChem v2 and the Universal Model for Atoms, or UMA, to replace separate machine-learning potentials for molecules, catalysts and inorganic materials with one cross-domain framework. The approach matters because faster atomistic calculations could shorten research cycles in drug discovery, energy storage and semiconductor manufacturing.

Meta FAIR released UMA, along with code, model weights and associated data, on May 14, 2025. The family was trained on roughly 500 million unique three-dimensional atomic structures; UMA-medium has 1.4 billion parameters but activates about 50 million for each structure. Through FAIRChem v2’s Atomic Simulation Environment integration and GPU execution, researchers can use the same model for single-point energy and force predictions, geometry optimization, vibrational calculations and molecular dynamics by selecting the relevant domain task.