Discovering more than 2 million new crystal structures, Google's AI shortens 800 years of research for humanity

Researchers at Google's DeepMind artificial intelligence division have discovered 2.2 different crystal structures , opening up the potential for a range of new discoveries in renewable energy and next-generation computing. At the same time, these findings also show the power of artificial intelligence systems when used to search for new materials.

Of the more than 2 million new crystal structures, about 380,000 are stable enough to be developed into next-generation technologies, ranging from solar cells and electric vehicle batteries to superconductors for next-generation supercomputers. The new discoveries also show how powerful artificial intelligence systems can help humans shorten the time it takes to experiment and find new materials.

Google's DeepMind has discovered 2.2 different crystal structures.

' For me, materials science is essentially abstract thinking meeting the physical universe ,' said Ekin Dogus Cubuk, co-author of the study. ' It is hard to imagine any technology that would not be improved with better (newly discovered) materials .'

To discover these new crystals, DeepMind researchers developed a new advanced artificial intelligence network called GnoME (Graph Networks for Materials Exploration) . Scientists first built data from 48,000 materials known for thousands of years, including copper, iron, and recently discovered materials.

Google's GNoME AI discovered 421,000 new stable material crystals, outperforming human calculations and experiments.

From the data, the machine learning network first generates new material structures and then evaluates their stability. Over time, GNoME generates hundreds of thousands of these new crystals. DeepMind estimates that the number of materials identified by AI is equivalent to the empirical knowledge gained in 800 years of human development – ​​based on the 28,000 new materials identified in the past decade.

GNoME operating model.

" From microchips to batteries and photovoltaics, the search for new crystals is often hampered by expensive trial-and-error experiments ," the paper in Nature said. " Our work represents a major expansion of the range of stable materials known to mankind ."

6 examples of materials discovered by Google's AI algorithms and validated by outside research teams with their own research

Two potential applications for the new compounds include creating new types of flexible layered materials and developing forms of neuromorphic computing, using microchips to replicate human brain activity, Cubuk said.

Researchers at the University of California, Berkeley, and Lawrence Berkeley National Laboratory also used DeepMind's findings to experiment with new materials. Called the A-lab, the team of researchers deployed a computational effort that used historical data and machine learning to guide an automated lab and created 41 new compounds from a list of 58 discovered by the AI—a success rate of more than 70 percent.

That's why the researchers behind DeepMind's new AI tool have called GNoME the "ChatGPT of chemistry ," referring to the AI ​​tool's huge impact on materials science.

'Scientific discovery is going to be the new frontier of AI,' said Carla Gomes, co-director of the Cornell University AI Science Institute. 'That's why I find this so exciting.'