Crystal-hunting DeepMind AI could help discover new wonder materials

An artificial intelligence created by Google DeepMind may help revolutionise materials science, providing new ways to make better batteries, solar panels, computer chips and many more vital technologies.

“Anytime somebody wants to improve their technology, it inevitably includes improving the materials,” says Ekin Dogus Cubuk at DeepMind. “We just wanted them to have more options.”

The AI model, called Graph Networks for Materials Exploration, or GNoME, is designed to predict inorganic crystal structures, which are repeating arrangements of atoms that provide materials with particular properties – for example, the six-fold symmetry of a snowflake is a result of the crystal structure of ice.

Organic crystals, which include carbon-hydrogen bonds, are well understood because of numerous examples in biological systems, but until now we only knew of around 48,000 possible inorganic crystals. GNoME has massively expanded that figure to more than 2 million, and while some of these new structures might decay into more stable forms or be impossible to create altogether, more than 700 of the predictions have already been made in the lab.

GNoME is a graph neural network, a kind of AI that can learn the relationships between objects, such as atoms and their chemical bonds. Cubuk and his team trained GNoME on an existing database of known inorganic crystals and used it to generate new possible crystals by changing the elements or playing with the known crystals’ symmetries. It also predicted the energies of the new crystals, a measure of their stability.

The researchers used quantum mechanics simulations to assess the accuracy of these energy predictions, then fed these results back into GNoME’s structure predictions, for a total of six rounds. “What we saw is that, every round, the model’s predictions got better and better for generalising to novel stable crystals,” says Cubuk.

Of the 2.2 million predictions, there were 400,000 crystals that were in their most stable form, with no lower-energy form possible. Some less stable crystals could still be useful, however – these are known as metastable crystals. Diamond, a metastable form of carbon, is one example.

By scanning the scientific literature published after GNoME was developed, ensuring any results found weren’t present in its training data, the team discovered that GNoME had predicted more than 700 crystals since produced by other researchers. These include a diamond-like lithium and magnesium crystal that could be used in high-powered lasers and a low-temperature molybdenum superconductor.

DeepMind shared its predictions with Yan Zeng at Lawrence Berkeley National Laboratory in California and her colleagues, who are developing a robotic lab capable of autonomously synthesising crystals. The Berkeley team independently predicted 58 crystal structures, of which the automated lab was able to create 41. As GNoME had also predicted these structures, this external verification suggests that its predictions are at least 70 per cent accurate, says Cubuk.

The researchers have now made the entire data set of predicted crystal structures available to others. “It is going to accelerate discovery of new materials,” says Graeme Day at the University of Southampton, UK. “That’s the big deal of it – compared to what’s been in these databases before, you’re able to scale up by an order of magnitude.”

These materials could include things like better alloys for cars, improved energy density for solid state batteries and more effective energy harvesting for solar panels, says Cubuk.

Andy Cooper at the University of Liverpool, UK, also says it will speed up the process of discovering new materials, but knowing a material’s properties, such as its conductivity or ability to store energy, is important too. “Property calculations tend to often be quite expensive,” says Cooper. “You know the structure might exist, but if you don’t know what it does, then it’s not clear whether to make it or not.”

For now, the best way to understand a material’s properties is to synthesise it, which is a major bottleneck for chemistry, even with the assistance of robotic labs such as those used by Zeng and her team. “The world’s capability to predict and calculate things, and use machine learning to extrapolate further, is evolving faster than the robotic capability to look for the materials,” says Cooper.