Google DeepMind says it has developed an AI model capable of predicting millions of inorganic crystal structures that could potentially be used to form next-gen microprocessors, electric batteries, solar panels, and the like.

Crystalline structures are made up of atoms arranged in a repeating pattern. They often exhibit particular characteristics that allow them to conduct electricity, light, or magnetism under specific conditions that can make them useful for electronic engineering. 

Silicon forms a diamond cubic crystal structure, for example, and as we’re sure you know is used as the basis of the minute electronic circuits in computer chips. Graphene is made up of layers of carbon atoms in a hexagonal lattice, and its thermal properties have been tested to make heat pipes and thermal straps for spacecraft [PDF].

Scientists working to discover as-yet-unknown crystalline structures with desirable properties used to experiment on known materials – getting them to react with different elements and molecules – in hopes of something cool happening. The trial and error method is obviously complex and time-consuming, and frequently leads nowhere.

Things have improved with the use of computer simulations that model whether a new structure might be chemically stable or not, and worth trying to make in a lab.

The computational approach has reportedly led to the discovery of tens of thousands of potential crystal structures. Researchers at Google DeepMind this week said they have developed an AI model that has come up with 2.2 million possible crystal structures, and believe that 380,000 of the candidate materials could be stable enough to use in future technologies. 

This software, nicknamed GNoME, is based on a graph neural network that has been trained on data taken from 69,000 known crystals. Given the atomic structure or chemical formula of a new material, GNoME predicts its potential properties so that scientists can then calculate how stable it might be. If it all looks promising, it might be worth pursuing as a viable material.

“To assess our model’s predictive power during progressive training cycles, we repeatedly checked its performance using established computational techniques known as Density Functional Theory (DFT) – used in physics, chemistry and materials science to understand structures of atoms, which is important to assess the stability of crystals,” Amil Merchant and Ekin Dogus Cubuk, co-authors of a Nature paper on GNoME and researchers at Google DeepMind, explained.

“We used a training process called ‘active learning’ that dramatically boosted GNoME’s performance. GNoME would generate predictions for the structures of novel, stable crystals, which were then tested using DFT. The resulting high-quality training data was then fed back into our model training.”

Using this approach, GNoME reportedly generated 52,000 compounds with a similar structure to graphene, and 528 lithium ion conductors, as well as 15 lithium transition-metal oxides that could potentially be used to craft materials for superconductors and rechargeable batteries.

It’s not quite clear how accurate GNoME is exactly with its predictions, but it does seem like it could be a promising tool. We’re not totally sold on it but it’s an interesting direction to take machine learning.

Scientists essentially could use something like GNoME to figure out whether a new compound is worth making in a lab and developing into something that has a potential commercial angle. Out of the millions of structures it predicted, 736 crystals matched structures that have been chemically verified in separate previous experiments, so it’s perhaps onto something, at least. 184 of those structures were only recently discovered, too.

Combining robotics with AI

It’s a big step from predicting novel crystalline structures to creating materials for future hardware, however.

In another project, collaborating with researchers from the Lawrence Berkeley National Laboratory working on the Materials Project, some of GNoMe’s outputs were used to select 58 inorganic powders for a robotic arm to cook up. They chose targets that were predicted to not react with oxygen, carbon dioxide, and water – so they would be stable in normal conditions. 

The suggested chemical recipe to make the inorganic powders was generated by a large language model trained on text data scraped from academic papers. These were then fed as a series of instructions for a robotic arm to carry out. The robot selects the ingredients, mixes them up and heats them to create the compound. The whole system, named A-Lab, has been described in a second paper published in Nature. 

Working non-stop over 17 days, A-Lab managed to produce 41 of the 58 targets. “We had this staggering 71 [percent] success rate, and we already have a few ways to improve it,” enthused Gerd Ceder, the principal investigator for A-Lab and a scientist at the Berkeley Lab and the University of California, Berkeley. 

Out of the 41 materials successfully synthesized by the robot, 35 were made from recipes generated by A-Lab. The system doesn’t usually get it right on the first go, however, and spits out hundreds of potential recipes that have to be tested. “Despite 71 [percent] of targets eventually being obtained, only 37 [percent] of the 355 synthesis recipes tested by the A-Lab produced their targets,” the researchers wrote in their paper.

They found that slow reaction times, volatile materials, and the system’s inaccuracies led to failures. Sometimes the chemical compound was too complicated for the robot to make, or wasn’t stable enough. In some cases, the crystalline structure of a powder ended up as a gooey mess. “This success rate could be improved to 74 [percent] with only minor modifications to the lab’s decision-making algorithm, and further to 78 [percent] if the computational techniques were also improved,” they suggested.

Ceder declared that the A-Lab experiments proved that advances in AI and robotics can help scientists make and test materials faster than before. 

“We have to create new materials if we are going to address the global environmental and climate challenges,” argued Kristin Persson, the founder and director of the Materials Project at Berkeley Lab and a professor at UC Berkeley. “With innovation in materials, we can potentially develop recyclable plastics, harness waste energy, make better batteries, and build cheaper solar panels that last longer, among many other things.” ®

• SEO Powered Content & PR Distribution. Get Amplified Today.
• PlatoData.Network Vertical Generative Ai. Empower Yourself. Access Here.
• PlatoAiStream. Web3 Intelligence. Knowledge Amplified. Access Here.
• PlatoESG. Carbon, CleanTech, Energy, Environment, Solar, Waste Management. Access Here.
• PlatoHealth. Biotech and Clinical Trials Intelligence. Access Here.
• Source: https://go.theregister.com/feed/www.theregister.com/2023/11/30/ai_predicts_novel_materials/