How did AI discover 2.2 million new materials in one go?

The answer

In November 2023, Google DeepMind unveiled GNoME (Graph Networks for Materials Exploration), an AI tool that predicted the stability of 2.2 million new crystal structures - equivalent to roughly 800 years of traditional experimental discovery. Of these, 380,000 were identified as the most stable and promising candidates for real-world synthesis, including 52,000 new layered compounds similar to graphene with potential applications in superconductors and advanced electronics. GNoME uses deep learning with an 'active learning' training loop: it generates structure predictions, validates them against density functional theory calculations, and feeds the results back into training. The system boosted materials stability prediction accuracy from 50% to 80%. Remarkably, 736 of GNoME's computational predictions were independently synthesized by labs worldwide, and an autonomous laboratory successfully created over 41 novel materials using GNoME's recipes.