Scientists use artificial neural networks to predict new stable materials

Diagram of an artificial neural network predicting a stable garnet crystal prototypeCREDIT: Weike Ye

Artificial neural networks – algorithms inspired by connections in the brain – have "learned" to perform various tasks, ranging from pedestrian detection in autonomous cars to medical image analysis and language translation. Now, researchers at the University of California at San Diego are forming artificial neural networks to predict new, stable materials.

"The prediction of material stability is a central problem in materials science, physics and chemistry," said Shyue Ping Ong, senior author and professor of nanoengineering at UC San Diego's Jacobs School of Engineering. "On the one hand, you have a traditional chemical intuition such as Linus Pauling 's five rules that describe the stability of crystals in terms of rays and compresses of ions.On the other hand, you have expensive quantum mechanical calculations to calculate the energy this must be done on supercomputers.We have used artificial neural networks to bridge these two worlds. "

By driving artificial neural networks to predict the crystal formation energy using only two inputs – the electronegativity and the ionic radius of the constituent atoms – Ong and his team at the Virtual Materials Laboratory have developed models capable of identify stable materials in two classes of crystals called garnets. and perovskites. These models are up to 10 times more accurate than the previous machine learning models and are fast enough to efficiently filter thousands of documents in a few hours on a laptop. The team details the work in an article published on September 18 at Nature Communications.

"Garnets and perovskites are used in LED lamps, rechargeable lithium-ion batteries and solar cells – these neural networks can significantly accelerate the discovery of new materials for these and other important applications." PhD student in the virtual laboratory of Ong Materials.

The team made its models available to the public via a web application on http: // crystals.have. This allows other people to use these neural networks to calculate the training energy of any garnet or perovskite composition on the fly.

The researchers plan to extend the application of neural networks to other crystal prototypes as well as to other material properties.

Like this article? Click here to subscribe to Lab Manager free newsletters