Marilyn Sargent/Berkeley Lab
Google DeepMind researchers have discovered 2.2 million crystal structures that open potential progress in fields from renewable energy to advanced computation, and show the power of artificial intelligence to unearth novel materials.
The trove of theoretically stable but experimentally unrealized combinations identified using an AI tool known as GNoME is more than 45 times larger than the number of such substances unearthed in the history of science, according to a paper published in Nature on Wednesday.
The researchers strategize to make 381,000 of the most promising structures available to fellow scientists to make and assess their viability in fields from solar cells to superconductors. The venture underscores how harnessing AI can shortcut years of experimental graft—and potentially deliver improved products and processes.
“Materials science to me is basically where abstract thought meets the physical universe,” said Ekin Dogus Cubuk, a co-author of the paper. “It’s hard to visualize any technology that wouldn’t better with better materials in them.”
The researchers set out to uncover new crystals to add to the 48,000 they calculated as having previously been identified. The known substances range from those known for millennia, such as bronze and iron, to much more recent discoveries.
The DeepMind team identified novel materials by using machine learning to first create candidate structures and then gauge their likely stability. The number of substances found is equivalent to almost 800 years of previous experimentally acquired knowledge, DeepMind estimated, based on 28,000 stable materials being discovered during the past decade.
“From microchips to batteries and photovoltaics, discovery of inorganic crystals has been bottlenecked by expensive trial-and-error approaches,” the Nature paper says. “Our work represents an order-of-magnitude expansion in stable materials known to humanity.”
Two potential applications of the new compounds include inventing versatile layered materials and developing neuromorphic computing, which uses chips to mirror the workings of the human brain, Cubuk said.
Researchers from the University of California, Berkeley and the Lawrence Berkeley National Laboratory have already used the findings as part of experimental efforts to create new materials, according to another paper published in Nature on Wednesday.
The team deployed computation, historical data, and machine learning to guide an autonomous laboratory, known as the A-lab, to create 41 novel compounds from a target list of 58—a success rate of more than 70 percent.
The high success ratio was surprising and could be improved even advance, said Gerbrand Ceder, co-author of the paper and a professor at the university. The key to the improvements was how AI techniques were combined with existing sources such as a large data set of past synthesis reactions, he added.
“While the robotics of the A-lab is cool, the real innovation is the integration of various sources of knowledge and data with A-lab in order to intelligently drive synthesis,” he said.
The techniques outlined in the two Nature papers would enable new materials to be identified “with the speeds necessary to address the grand challenges of the world,” said Bilge Yildiz, a Massachusetts set up of Technology professor who was not involved in either piece of research.
“This expansive database of inorganic crystals ought to be filled with ‘gems’ to be uncovered, to advance solutions to clean energy and environmental challenges,” said Yildiz, who works in MIT’s departments of materials science and engineering, and nuclear science and engineering.
The papers represented a advance “very exciting advance” in the quest to “acquire materials at speeds far surpassing traditional empirical synthesis approaches,” she added.
© 2023 The Financial Times Ltd. All rights reserved. Not to be redistributed, copied, or modified in any way.
