GNoME

Published in Nature on November 29, 2023, GNoME (Graph Networks for Materials Exploration) is DeepMind's deep-learning tool for materials discovery. Trained on Materials Project data and improved through active learning, it predicted the stability of 2.2 million new crystal structures — described as roughly 800 years of prior knowledge — with 380,000 predicted stable enough for experimental synthesis. Over 730 of the predicted structures were independently synthesized by external labs, and the predictions feed Berkeley Lab's autonomous A-Lab synthesis robot.

DeepMind released the GNoME predictions openly to the research community and contributed the stable structures to the Materials Project database. The caveat is that GNoME is a prediction pipeline, not downloadable model weights, and predicted stability is not a synthesis recipe — most candidates still await experimental validation, and the jump from a stable crystal to a manufacturable technology remains a decade-scale engineering problem.