Boffins deem Google DeepMind's material discoveries rather shallow

Web titan rejects criticisms, insists AI-found compounds are legit

AI on its own may not be as useful for discovering new materials as Google's DeepMind team has suggested.

Two materials scientists affiliated with UC Santa Barbara have analyzed a Google paper published in Nature last November and conclude that it promises more than it delivers. Google's DeepMind team, however, disagrees and stands by their work, arguing that their UCSB critics mischaracterize their research and fail to appreciate their goals.

In November 2023, Google DeepMind boffins Amil Merchant and Ekin Dogus Cubuk announced the publication of a paper in the scientific journal Nature titled, "Scaling deep learning for materials discovery." Their co-authors include: Simon Batzner, Samuel Schoenholz, Muratahan Aykol, and Gowoon Cheon.

The paper describes Graph Networks for Materials Exploration (GNoME), "our new deep learning tool that dramatically increases the speed and efficiency of discovery by predicting the stability of new materials."

GNoMe has expanded the number of materials available to science by discovering crystalline structures, "many of which escaped previous human chemical intuition," the authors claim.

"With GNoME, we’ve multiplied the number of technologically viable materials known to humanity," declared Merchant and Cubuk.

"Of its 2.2 million predictions, 380,000 are the most stable, making them promising candidates for experimental synthesis. Among these candidates are materials that have the potential to develop future transformative technologies ranging from superconductors, powering supercomputers, and next-generation batteries to boost the efficiency of electric vehicles."

Well, maybe not so much it seems.

Novel, credible and useful?

In an article published Monday in the journal Chemistry of Materials, Anthony Cheetham and Ram Seshadri, both research professors in the Materials Research Lab at UC Santa Barbara, argue that the materials identified by GNoME aren't as useful as has been suggested.

Cheetham and Seshadri take issue with the expansive language of the DeepMind paper, which states, "Our work represents an order-of-magnitude expansion in stable materials known to humanity."

The two UC Santa Barbara boffins argue the DeepMind predictions "are solely of crystalline inorganic compounds and should be described as such, rather than using the more generic label 'material.'" They note that there are many more types of materials into which GNoME did not delve, such as polymers, glasses, metal−organic frameworks, heterostructures, and composites.

More significantly, they argue that meaningful predictions of new materials should be: credible – the structure and composition of matter should be something that can be realized in experiments; novel – not a trivial extension of known compounds; and useful – exhibiting enough evidence of utility to be recognized as materials.

And GNoME's addition to the canon of known stuff doesn't fit within this triangle, it's argued.

"We examine the claims of this work here, unfortunately finding scant evidence for compounds that fulfill the trifecta of novelty, credibility,and utility," explain Cheetham and Seshadri in their analysis of DeepMind's work. "While the methods adopted in this work appear to hold promise, there is clearly a great need to incorporate domain expertise in materials synthesis and crystallography."

The paper from Google DeepMind is not particularly useful to experimentalists

Cheetham elaborated to The Register. "We do believe that there is a lot of potential for the applications of AI to materials science (and indeed to chemistry, though that is a broader question).

"However, the paper from Google DeepMind is not particularly useful to experimentalists such as ourselves because it offers an overwhelming number of predictions (2.2 million, of which nearly 400,000 are believed to be stable), many of which do not appear to be very novel. These are chemical compounds rather than materials because they have no demonstrated functionality or utility at this point. "

Cheetham said it's premature to place limits on the potential utility of AI in materials science at this point. "However, as I hope our article makes clear, we do believe that AI has a great future in the area when it is combined with first-class domain expertise from materials scientists."

"We stand by all claims made in Google DeepMind’s GNoME paper," a Google DeepMind spokesperson told The Register.

"Our GNoME research represents orders of magnitude more candidate materials than were previously known to science, and hundreds of the materials we’ve predicted have already been independently synthesized by scientists around the world." ®

More about


Send us news

Other stories you might like