“Scaling deep learning for materials discovery,” by Amil Merchant, Simon Batzner, Samuel Schoenholz, Muratahan Aykol, Gowoon Cheon and Ekin Dogus Cubuk at Google DeepMind, was published in Nature in late 2023. It described GNoME, short for Graph Networks for Materials Exploration, a graph neural network trained to predict whether a hypothetical crystal would be stable.
The work combined the model with an active-learning loop: GNoME proposed candidate structures, a fraction were checked with the slower but accurate method of density functional theory, and those results were fed back to improve the model. Run at scale, the protocol identified 2.2 million stable crystal structures, of which roughly 380,000 were new predictions beyond what was already known, expanding the catalogue of candidate stable materials by close to an order of magnitude. A companion Nature paper from Lawrence Berkeley National Laboratory used GNoME predictions to drive an autonomous laboratory that synthesized some of the proposed compounds.
Discovering new stable materials underpins better batteries, solar cells, superconductors and catalysts, and the conventional path of trial-and-error synthesis is slow and expensive. GNoME showed that a learned model could enormously widen the pool of candidates worth testing, which is why it became one of the most-cited examples of AI accelerating the physical sciences.