UMA - A Family of Universal Models for Atoms

Explore a groundbreaking family of universal atomic models designed to revolutionize computational chemistry and materials science in this 58-minute conference talk. Learn about UMA (Universal Models for Atoms), Meta FAIR's innovative approach to creating AI models that can quickly and accurately compute properties from atomic simulations across diverse applications including drug discovery, energy storage, and semiconductor manufacturing. Discover how these models were trained on an unprecedented dataset of half a billion unique 3D atomic structures, compiled from multiple chemical domains including molecules, materials, and catalysts. Understand the development of empirical scaling laws that guide optimal model capacity scaling alongside dataset size for maximum accuracy. Examine the novel "mixture of linear experts" architectural design that enables the UMA-medium model to achieve 1.4 billion parameters while maintaining only ~50 million active parameters per atomic structure, ensuring both capability and computational efficiency. Review comprehensive evaluations demonstrating how a single UMA model can perform comparably or superior to specialized models across diverse tasks without requiring fine-tuning. Gain insights into the open-source release of UMA code, weights, and associated data, designed to accelerate computational workflows and enable the community to build increasingly capable AI models for atomic-scale simulations.