Google DeepMind researchers have made a groundbreaking discovery, uncovering 2.2 million crystal structures that have the potential to revolutionize various fields, from renewable energy to advanced computation. This significant advancement demonstrates the capacity of artificial intelligence to identify novel materials.
The repository of theoretically stable but not yet realized combinations, identified through the AI tool called GNoME, surpasses the total number of such substances discovered throughout the history of science by over 45 times. A paper published in Nature on Wednesday highlights this remarkable feat.
To expedite progress in fields like solar cells and superconductors, the researchers plan to share 381,000 of the most promising structures with fellow scientists for testing and experimentation. This initiative underscores the ability of AI to streamline the experimental process, potentially leading to the development of superior products and processes.
Ekin Dogus Cubuk, a co-author of the paper, expressed the significance of materials science, describing it as the intersection of abstract thought and the physical universe. He emphasized the potential improvement in various technologies through the incorporation of superior materials.
The research team aimed to discover new crystals to augment the existing 48,000 identified. These known substances span millennia, from ancient bronze and iron to more recent discoveries. Using machine learning, the DeepMind team generated candidate structures and assessed their likely stability, resulting in a substantial expansion of knowledge equivalent to almost 800 years of previously experimentally acquired data.
The Nature paper notes the historical bottleneck in the discovery of inorganic crystals and highlights the order-of-magnitude expansion in stable materials. Potential applications of these new compounds include the creation of versatile layered materials and the advancement of neuromorphic computing, mimicking the human brain’s workings.
Researchers from the University of California, Berkeley, and the Lawrence Berkeley National Laboratory have already incorporated these findings into experimental efforts to create new materials. Their success rate exceeded 70%, showcasing the potential of combining AI techniques with existing knowledge sources.
Gerbrand Ceder, co-author of the paper and a professor at the University of California, Berkeley, emphasized the innovation in integrating various sources of knowledge and data with the A-lab, an autonomous laboratory. This approach resulted in creation 41 novel compounds out of a target list of 58.
According to Bilge Yildiz, a Massachusetts Institute of Technology professor, the technique outlined in the Nature papers enables the identification of new materials at the speeds necessary to address global challenges. Yildiz, who was not involved in the research, sees the expansive database of inorganic crystals as a source of “gems” that could advance solutions to clean energy and environmental challenges.