In 1905, a young American scientist named William Coolidge set to work on an ambitious assignment for his new job at General Electric: creating a better light bulb. The carbon filament used in Thomas Edison’s world-changing invention wasn’t very energy efficient. Scientists were experimenting with tungsten, which had the advantage of a higher melting point, but the metal wire was too brittle and broke easily.
Several years later, after numerous trials, errors, and accidents—by some accounts, a breakthrough occurred when the tungsten rod fell into liquid mercury pooling from the heating system—Coolidge created “ductile tungsten,” a more bendable form of the metal that ultimately became the wiring still found inside incandescent light bulbs today.
Ekin Dogus Cubuk, a research scientist at Google DeepMind, points to ductile tungsten as an example of how new materials can revolutionize existing technologies—often after careful, laborious efforts riddled with unexpected twists and turns. More than 100 years after that landmark innovation, his team has created an AI tool that could transform the way scientists discover new materials. Much like ductile tungsten, any of these compounds could improve upon current materials and lead to next-generation electric vehicle (EV) batteries or better solar panels. “Many materials discoveries involve a lot of trial and error, luck, and serendipity,” Cubuk says. “With AI we hope to reduce the reliance on luck. We have the potential to accelerate each stage of the process, from discovery to development, with more accurate and scalable predictions.”
GNoME catapults materials discovery
The AI tool that Cubuk’s team built is called GNoME (Graph Networks for Materials Exploration), and it has used deep learning to determine the structure of 2.2 million inorganic crystals. Of those, GNoME identified 380,000 materials that are stable at low temperatures, according to simulations. “We hope that having this information will help the community to catapult further breakthroughs in materials discovery and design,” Cubuk explains.
Since the public release of these crystals in 2023, scientists have been experimenting with AI-identified compounds in labs worldwide—including three that were recently synthesized in Japan by researchers at Hokkaido University.
“There is a wealth of AI knowledge, but currently there’s a gap between translating those insights to material synthesis in the lab. This early research shows how we can bridge the gap to explore new materials more efficiently,” says Akira Miura, an associate professor at Hokkaido University, who led the study.
Meanwhile, GNoME independently identified hundreds of crystals that scientists had already discovered on their own as stable, including “exciting” superconductor candidates, according to Cubuk. Furthermore, hundreds of these crystals were independently synthesized by labs around the world. That’s a strong signal that the new material candidates in GNoME’s findings have real-life viability.
GNoME was initially trained with data on crystal structures that is available through the Materials Project, an open-source reference that has been central to new materials discovery. Identifying the structure of a new material is just the beginning of what can be a long, intensive development process. The original idea of the Materials Project was to reduce invention time by focusing experiments on compounds with the most potential for success, according to founder and UC Berkeley professor Kristin Persson. That itself was a game changer from the days when Persson was a graduate student in the mid-’90s and had to comb through oversize tomes to look up a material’s chemical structure and properties. In the laboratory, researchers would combine intuition with past results to inform their experimentation. “There were no recipes for it except in the researcher’s brain,” Persson says. “They would use their experience to make new materials, and it would take a long time.”
Not that long ago, it was relatively rare to be able to calculate a new material on the computer that was thermodynamically stable, explains Chris Wolverton, a professor of Materials Science and Engineering at Northwestern University. For one, the sheer magnitude of the required computational power was a hurdle. “Google is particularly rich when it comes to computational time and power,” he says. “They were able to—in a very short order—calculate the properties of several million compounds.”
The availability of this vast data trove is opening up new possibilities to materials researchers, says Persson. “They’re giving this body of work to the community and saying, ‘Do what you want with it,’” she says. “It’s the new kid on the block in terms of the tools we have. If you’re a materials company and not thinking of machine learning, you’re behind, and Google put a brilliant spotlight on that.”
That doesn’t mean researchers are suddenly going to have piles of new materials cluttering their lab benches. The tool has identified crystals that computations predict will be stable under absolute zero, or about -273C—an ideal condition in which particle motions come to a stop. A key question persists: Can these crystals be created in a stable form that won’t change or decompose at room temperature? Myriad factors, including temperature and pressure, can influence whether a theoretical compound can exist in a form that would become, say, the next EV battery. Scientists will have to figure out the right experimental methods to create materials that will be useful in innovation. “That’s the next giant challenge in this field: predicting synthesis recipes,” Wolverton says.
Materials scientists explore next steps
Indeed, the tremendous scale of the AI findings has spurred important conversations about the best way to identify and synthesize new materials. A researcher may spend an entire career conducting experiments on just a few compounds, a painstakingly slow and labor-intensive process. The key is finding a balance between creating AI-led possibilities and developing the known science, explains University of Liverpool professor Andy Cooper, whose research group focuses on ways to accelerate the discovery of functional materials: “I think that’s a challenge for the future. What is the right place to be on that spectrum?”
New materials improve the foundations of modern life
At Google DeepMind, Cubuk envisions a future when new materials could reshape today’s most exciting technologies—from semiconductors to new power sources for supercomputers. As AI tools evolve, a key challenge is determining which potential materials best correspond to technologies and products, Cubuk explains. For example, knowing what criteria makes a good lithium-ion conductor helps researchers better identify the most promising candidates. Already, the GNoME team has identified 528 potential materials for batteries—more than 25 times the number identified in a 2017 study, which was led by Austin Sendek, the CEO of Aionics, a technology company dedicated to designing high-performance batteries with AI.
AI’s contributions might also open up researchers’ minds to unforeseen possibilities, says Jakoah Brgoch, an associate professor of chemistry at the University of Houston, whose work explores functional inorganic materials. Maybe the AI-suggested compounds aren’t viable under current lab temperatures and pressure, for example; but what happens if you alter those variables? “Science is often a slog,” Brgoch says. “The hope and the hype is we don’t have to go through this slog. Generative AI can predict the next big thing. As we exhaust what’s known, we have to develop new ways to push boundaries, and this is one of those paths. Of course, experimentally validating these predictions is easier said than done.”
According to current industry wisdom, new material discovery requires, on average, at least 10 years of work and north of $10 million to $100 million. While the cost to build a factory is the same, what used to take weeks to simulate on a computer can now be done in seconds with AI, says Sendek.
“AI-based materials discovery methods will unlock unorthodox solutions that would have otherwise eluded human intuition, while also leading to fewer ‘dead ends’ that cannot be commercialized,” says Sendek, pointing to the recent discovery by Aionics’s AI system that adding a luxury perfume molecule to a battery’s electrolyte could improve the battery’s overall performance. Aionics is now actively testing this molecule in real-life batteries.
“Making meaningful progress in materials innovation requires us to discover better materials faster and then get them to market in a fraction of the time.”
