New materials are discovered slowly, one painstaking experiment at a time
The old claim turned out to be wrong and was rejected by evidence.
TaughtDiscovering a new stable material, the kind that could make a better battery or a superconductor, is slow, hard lab work. Across all of history, chemists had characterized only a few tens of thousands of stable inorganic crystals, each the product of careful synthesis and testing.
NowIn November 2023, Google DeepMind reported in Nature that its GNoME AI had predicted 2.2 million new crystal structures, about 380,000 of them stable: roughly a tenfold jump over all the stable inorganic crystals humanity had catalogued before, generated in a matter of weeks.
What actually happened
GNoME (Graph Networks for Materials Exploration) learns the patterns of what makes a crystal stable, proposes vast numbers of candidates, and filters them by predicted stability far faster than a lab could. DeepMind likened the haul to nearly 800 years of prior knowledge, and hundreds of the predictions were subsequently made and confirmed by experimenters.
The honest caveat is important and became a live debate: a prediction of stability is not a material in your hand, and some researchers questioned how many of the candidates are genuinely novel or synthesizable. So this is a reversal of the discovery *rate* and *scale*, not a claim that 2.2 million new substances now sit on shelves. What changed for good is the method; materials discovery is no longer gated by one experiment at a time.
Newsletter
Get new weekly revisions
Sources
Who was taught this
Still standard through 2023, so anyone who finished school between 1950 and 2023 learned the earlier version.