Advancing Materials Science through AI: GNoME’s Revolutionary Predictions
We are familiar with approximately 48,000 inorganic crystal structures, each conferring unique material properties. Now, Google DeepMind’s AI has ventured into uncharted territory, predicting over 2 million additional possibilities.
Exploring Over 2 Million Potential Inorganic Crystal Structures
Google DeepMind’s latest AI breakthrough promises to transform materials science, offering novel avenues for enhancing batteries, solar panels, computer chips, and other critical technologies. This development unfolds as noteworthy Chemical News for professionals immersed in the field of chemistry.
According to Ekin Dogus Cubuk at DeepMind, “Any endeavor to enhance technology invariably involves improving the materials—our aim was to broaden those possibilities.”
The AI model, Graph Networks for Materials Exploration (GNoME), specializes in forecasting inorganic crystal structures. These structures, characterized by repeating atomic arrangements, determine a material’s distinctive properties; for example, the hexagonal symmetry of a snowflake arises from ice’s crystal structure.
While organic crystals with carbon-hydrogen bonds are well-explored, the knowledge of inorganic crystals was limited to about 48,000. GNoME has exponentially increased this figure to over 2 million. Although some structures may prove impractical or decay into more stable forms, Chemical News has already circulated regarding GNoME’s predictions being validated through real-world experiments.
Also read : Scientists Found a New Material, Possibly Harder Than Diamond
GNoME, a graph neural network, learns relationships between objects like atoms and chemical bonds. Trained on an existing database of inorganic crystals, it generates new possibilities by altering elements or adjusting crystal symmetries. It also predicts the energies of these new crystals, indicating their stability.
Quantum mechanics simulations were employed to validate energy predictions, iteratively refining GNoME’s structure predictions over six rounds. “With each round, the model’s predictions consistently improved for generalizing to novel stable crystals,” notes Cubuk.
Among the 2.2 million predictions, 400,000 crystals were identified as being in their most stable form, with no lower-energy alternative. Some less stable crystals, referred to as metastable, still hold potential uses. For instance, diamond, a metastable form of carbon, showcases the practicality of such structures.
Read more article on chemistry breakthrough
In a post-GNoME development literature scan, the team found over 700 predicted crystals subsequently produced by other researchers. Notable examples include a diamond-like lithium and magnesium crystal for high-powered lasers and a low-temperature molybdenum superconductor.
Collaborating with Yan Zeng at Lawrence Berkeley National Laboratory, whose team is developing a robotic lab for crystal synthesis, DeepMind shared predictions. Of the 58 crystal structures independently predicted by Berkeley, 41 were successfully created by the automated lab. GNoME’s predictions, aligning with Berkeley’s results, suggest at least 70 percent accuracy.
The researchers have made the entire dataset of predicted crystal structures publicly accessible, anticipating accelerated material discovery. “This represents a significant leap compared to existing databases, enabling a tenfold scaling up of material exploration,” remarks Graeme Day at the University of Southampton, UK.
Potential applications abound, from enhanced alloys for cars to increased energy density in solid-state batteries and improved energy harvesting in solar panels. This development signifies a notable convergence of materials science and artificial intelligence, heralding groundbreaking advancements in the realm of chemistry and materials. Exciting Chemical News awaits the world of chemical research as technology and innovation progress.