GNoME finds 2.2 million candidate new materials

Google DeepMind published GNoME (Graph Networks for Materials Exploration), a deep-learning system trained to predict which hypothetical inorganic crystal structures would be thermodynamically stable, alongside two papers in Nature. The company reported that GNoME had predicted 2.2 million new candidate structures, of which it flagged 380,000 as the most stable and therefore the most promising targets for experimental synthesis — an expansion, DeepMind said, of the roughly 48,000 stable inorganic materials known to science before the project, built on around 28,000 prior computational discoveries.

DeepMind said external laboratories, working independently and in parallel, had already synthesised 736 of GNoME’s predicted structures by the time of publication, and that an autonomous robotic laboratory using GNoME’s predictions had synthesised a further 41. The company framed candidate materials among the predictions as potentially useful for superconductors, batteries and other technologies, though it stopped short of demonstrating specific functional applications itself.

The scale of the claim drew scrutiny after publication. In April 2024, materials scientists Anthony Cheetham and Ram Seshadri, both at UC Santa Barbara, published a critique in Chemistry of Materials arguing that GNoME’s outputs were “solely crystalline inorganic compounds” rather than the broader class of “materials” the announcement implied, and that the paper offered “scant evidence” that the predictions met the three criteria that would make a discovery useful to experimentalists: credibility, novelty and demonstrated utility. Their objection centred less on any specific error rate than on the claim that raw structural stability, without evidence a material could be made or was good for anything, did not by itself constitute a discovery.

GNoME nonetheless became one of the most widely cited examples of AI-accelerated science, and its dataset was released for other researchers to build on.