Using motif-centric AI to explore new inorganic materials

Two College of Science and Technology (CST) professors have developed an advanced artificial intelligence approach able to accurately predict physical properties of inorganic materials.

Qimin Yan, assistant professor of physics, and Slobodan Vucetic, professor of computer and information sciences and director of Temple’s Center for Hybrid Intelligence, co-authored a paper outlining their research that was published this spring in the prestigious Science Advances journal. The paper, “Structure motif-centric learning framework for inorganic crystalline systems,” is the product of a collaboration between Vucetic and Yan that began five years ago.

“By demonstrating that structure motifs in crystals, similar to those building blocks in a LEGO playset, can be incorporated into and greatly improve the prediction accuracy of a machine learning framework, our work represents a fundamental new step towards the continued development of AI for inorganic material systems,” says Yan.

“Our approach is significant because it allows scientists to rapidly screen a large number of candidate inorganic materials in search for the ones that have desired chemical and physical properties. Ability to screen thousands of candidate materials per hour compares favorably to the traditional approach that requires days or weeks to computationally or experimentally characterize a single material,” says Vucetic.

“It also showcases the potential for interdisciplinary collaborations within CST,” adds Yan, whose research utilizes computational methods to accelerate the discovery, understanding and development of advanced materials. “Dr. Vucetic understands how artificial intelligence can be utilized to solve material science problems.”

Vucetic describes Yan as having “a rare ability to explain complex physics concepts to nonexperts and this really made our collaboration possible.”

The researchers concluded that structural motifs found in inorganic crystals can form the basis for a machine-learning framework that is more accurate in predicting the electronic structures of metal oxides. The research, Yan says, could have potential implications for a variety of industrial applications, including the development of new photovoltaic materials in solar cells and photocatalysts for solar fuel generation.

Read the article in Science.

-Bruce E. Beans

August 4, 2021