Google DeepMind researchers used AI to find 2mn materials

Google DeepMind researchers discovered 2.2mn new crystal structures, which could lead to advancements in renewable energy and advanced computation. They also show how artificial intelligence can be used to discover new materials.
According to a Nature paper published on Wednesday, the number of theoretically stable combinations that have not been experimentally realized using the AI GNoME tool is 45 times greater than any other substance of this type ever discovered.

Researchers plan to share 381,000 of their most promising structures with other scientists so they can make them and test the viability for fields ranging from solar cells to ultra-high conductivity. This project shows how AI is able to reduce the time it takes for scientists to experiment and produce better products.
Ekin Dogus Cuuk, co-author of this paper, said: “Materials Science is to me where abstract thinking meets the physical world.” It’s difficult to imagine a technology that would not improve with improved materials.

Researchers set out to discover new crystals, to add to 48,000 that they had calculated to have been previously identified. These substances include those that have been known for centuries, like bronze and iron, as well as more recent discoveries.
The DeepMind identified novel materials using machine learning. First, the team generated candidate structures to gauge their stability. DeepMind estimates that the number of substances discovered is equivalent to nearly 800 years’ worth of experimentally acquired knowledge. This estimate was based on 28000 stable materials being found in the last decade.

Nature says that expensive trial and error approaches have slowed down the discovery of inorganic crystalline materials, from microchips to photovoltaics. Our work represents a massive expansion of stable materials that are known to mankind.
Cubuk stated that the new compounds could be used to create layered materials with a variety of applications, and also for neuromorphic computing. This involves using chips which mimic the brain’s workings.

According to a second paper published on Wednesday in Nature, researchers from the University of California Berkeley and Lawrence Berkeley National Laboratory used the findings to experiment with new materials.
The team used computation, historical data, and machine learning to guide the A-lab autonomous laboratory to create 41 new compounds from a list of 58 targets — a success percentage of over 70%.

Gerbrand Ceder is a professor and co-author at the University. He said that the high success rate was surprising, and it could be further improved. He said that the key to improvements was the combination of AI techniques with existing sources, such as a huge data set on past synthesis reactions.
He said that while the A-lab’s robotics is impressive, the real innovation lies in the integration of different sources of data and knowledge with A-lab to intelligently drive the synthesis.

The two Nature papers will allow for the identification of new materials “at the speed necessary to meet the grand challenges of our time”, according to Bilge Yildiz. A Massachusetts Institute of Technology Professor who was not involved with either research.

This database should be filled with a wealth of ‘gems,’ which can be uncovered to help solve the clean energy and environment challenges, said Yildiz. She works at MIT in departments of Materials Science and Engineering, and Nuclear Science and Engineering.
She added that the papers were a “very exciting” advance in her quest to “obtain material at speeds far exceeding traditional empirical synthesis methods”.