Proposing and making a new inorganic material · Contested (since 2023)

In November 2023 Google DeepMind published GNoME, 2.2 million predicted crystal structures of which 380,000 were judged stable, alongside an autonomous laboratory at Berkeley that reported synthesising more than forty of the predicted compounds without human intervention.

Chemists answered within months, and the answer is the entry. Cheetham and Seshadri, reading the GNoME set, found scant evidence of compounds that were at once novel, credible and useful. Leeman and colleagues re-examined the autonomous laboratory and concluded that about two thirds of its claimed successes were probably compositionally disordered versions of already known compounds, misread by the automatic interpretation of the diffraction patterns. Generating candidates at scale is machine work now, and so is a good deal of the synthesis. Judging that a thing is actually new, and that the powder in the crucible is what the model asked for, was not ceded, and it was the experimental community that caught the error by looking at the data itself.