GNoME might be described as AlphaFold for materials discovery, based on Ju Li, a materials science and engineering professor on the Massachusetts Institute of Technology. AlphaFold, a DeepMind AI system announced in 2020, predicts the structures of proteins with high accuracy and has since advanced biological research and drug discovery. Due to GNoME, the variety of known stable materials has grown almost tenfold, to 421,000.

“While materials play a really critical role in almost any technology, we as humanity know only a couple of tens of 1000’s of stable materials,” said Dogus Cubuk, materials discovery lead at Google DeepMind, at a press briefing. 

To find latest materials, scientists mix elements across the periodic table. But because there are such a lot of combos, it’s inefficient to do that process blindly. As a substitute, researchers construct upon existing structures, making small tweaks within the hope of discovering latest combos that hold potential. Nonetheless, this painstaking process continues to be very time consuming. Also, since it builds on existing structures, it limits the potential for unexpected discoveries. 

To beat these limitations, DeepMind combines two different deep-learning models. The primary generates greater than a billion structures by making modifications to elements in existing materials. The second, nevertheless, ignores existing structures and predicts the soundness of latest materials purely on the idea of chemical formulas. The mixture of those two models allows for a wider range of possibilities. 

Once the candidate structures are generated, they’re filtered through DeepMind’s GNoME models. The models predict the decomposition energy of a given structure, which is a crucial indicator of how stable the fabric might be. “Stable” materials don’t easily decompose, which is essential for engineering purposes. GNoME selects probably the most promising candidates, which undergo further evaluation based on known theoretical frameworks.

This process is then repeated multiple times, with each discovery incorporated into the subsequent round of coaching.

In its first round, GNoME predicted different materials’ stability with a precision of around 5%, but it surely increased quickly throughout the iterative learning process. The ultimate results showed GNoME managed to predict the soundness of structures over 80% of the time for the primary model and 33% for the second.