GNoME and new materials
DeepMind predicted the structures of 2.2 million new crystals, 380 thousand of them judged stable, roughly eight hundred years of ordinary work.
Why it matters
After proteins the same scheme worked for materials: the model proposes and experiment checks.
The model trained on known structures and proposed new ones, scoring formation energy. External laboratories later synthesised some of the predicted compounds. Critics noted that a large share of the predictions are trivial variations on known compounds, and the argument about the real contribution continued. Either way the prediction database became a public resource for materials science.