Finding a material that could improve a battery, solar cell, or computer chip has often meant months or years of experiments. Google DeepMind’s GNoME tool aims to make the search broader and faster: it predicted structures for 2.2 million materials, and more than 700 have since been created in a lab and are being tested.
Searching beyond familiar structures
Scientists can discover materials by combining elements from the periodic table. The number of possible combinations makes a blind search inefficient, so researchers often start with known structures and make small changes to them. That approach can be slow, and its reliance on existing examples may make unexpected discoveries harder to find.
GNoME, short for graphical networks for material exploration, uses two deep-learning models to search in different ways. One generates more than a billion structures by modifying elements in existing materials. The other starts from chemical formulas and predicts whether new materials would be stable, without relying on existing structures.
Combining the models gives researchers a wider pool of candidates. GNoME then estimates each structure’s decomposition energy, an indicator of stability. Materials that resist decomposing are more useful for engineering, so the models prioritize promising candidates for further evaluation against known theoretical frameworks.
The system repeats this process, feeding discoveries into later rounds of training. In its first round, it predicted stability with a precision of around 5%. By the final results, it predicted stability correctly more than 80% of the time for the first model and 33% for the second.
A much larger catalogue of candidates
The discoveries have helped raise the number of known stable materials almost tenfold, to 421,000. That is a substantial expansion of the options available for research, but it does not mean every predicted structure has been made or shown to be useful.
More than 700 of the predicted materials have been produced in the lab and are now being tested. DeepMind also identified 528 promising lithium-ion conductors. Conductors help electric current flow between battery components, so some of these candidates may support more efficient batteries.
Possible applications extend beyond batteries. Chris Bartel, an assistant professor of chemical engineering and materials science at the University of Minnesota, said the materials could be candidates for batteries, computer chips, ceramics, and electronics. Their potential still needs to be established through synthesis and testing.
AI has already been used in materials discovery. The Materials Project, led by Kristin Persson at Berkeley Lab, has used similar techniques to discover and improve the stability of 48,000 materials. Researchers cited GNoME’s scale and precision as distinguishing features: it was trained on at least an order of magnitude more data than any previous model, according to Bartel.
From predictions to physical samples
Finding a candidate is only part of the work. Researchers must also synthesize it and establish whether it performs as hoped. Berkeley Lab’s autonomous laboratory, called the A-Lab, combines machine learning and robotic arms to develop materials using information from the Materials Project, including some GNoME discoveries.
The lab chooses how to make a proposed material and creates up to five initial formulations. A machine-learning model trained on scientific literature generates the formulations. After experiments, the lab uses the results to adjust its recipes and try again.
Berkeley Lab researchers reported that the A-Lab performed 355 experiments over 17 days and synthesized 41 of 58 proposed compounds. They said that amounts to two successful syntheses a day. By comparison, a human-led effort can take months or years when experiments do not go well.
The A-Lab’s capacity to keep trying matters because a predicted structure is not yet a practical material. Each attempt can provide results that guide the next formulation. GNoME broadens the search for candidates, while the autonomous lab helps test ways to make them.
Faster discovery is only one part of innovation
Researchers at DeepMind and Berkeley Lab see these tools as a way to accelerate hardware innovation in energy, computing, and other fields. Materials innovation is one part of work on clean energy, but the discovery process alone does not deliver a finished technology.
Even after a material is discovered, it can take decades for industry to bring it to the commercial stage. Dogus Cubuk, materials discovery lead at Google DeepMind, said reducing that timeline to five years would be a major improvement. GNoME and the A-Lab address earlier steps—identifying candidates and testing how to make them—while their eventual industrial use remains a separate challenge.