One example of this is Victor Fung, an assistant professor with Georgia Tech’s School of Computational Science and Engineering (CSE). Fung recently led a research team that developed a new, first-of-its-kind algorithm that can reconstruct atomic structure in generative models.

A significant application Fung focuses this research toward is in the field of materials science and engineering. The algorithm could be key in developing further AI tools and new materials to the benefit of individual researchers and entire communities alike.

“Structural representations are a well-known concept people have used in other machine learning applications for chemistry and materials, like training models to predict energies and forces,” Fung said. “But this is really the first time that anyone has used this in generative models.”

Structure is a key property in a material design. For example, structure plays a role in determining superconductivity within electronics, biological viability in drugs, and catalyzation of certain chemical reactions.

Fung explained that using generative models to study atomic structure, and to design new materials, could be vital in climate remediation. This may include developing greener catalysts for use in fuel cells, designing better material for carbon capture, and discovering new light-absorbent molecules for application in solar panels.

The algorithm can help engineers create new materials with targeted properties by building models atom-by-atom, a concept called inverse design. The algorithm is a progressive step forward in allowing computer models to create new materials tailor-made with specific functions and characteristics in mind by designers.

Specifically, the algorithm allows materials scientists to know the exact structure of materials that exhibit a desired property, potentially making proposed material designs a reality.

“If we know the structure of material, we can be sure of what properties it has, and we will have a clear goal to try to synthesize it and develop applications,” Fung said. “We basically have the key to defining the material in the chemical space.”

Fung’s paper is the first in a forthcoming series of studies to develop new generative models for atomic structure. He and his co-researchers think the series could result in new algorithms and models that yield commercial benefits, as well as solve large, scientific problems.

As part of this campaign to share his research, Fung is set to discuss the findings March 31 at 2023 Symposium on Materials Innovations, hosted by Georgia Tech’s Institute for Materials (IMat). 

School of CSE Ph.D. student Shuyi Jia worked with Fung to develop the algorithm and is a co-author on the paper. The pair partnered with Oak Ridge National Laboratory scientists Jiaxin Zhang, Junqi Yin, and Panchapakesan Ganesh through the study.

Along with AI tools like ChatGPT and DALL-E, generative models are popularly used today in images, text, audio, and other types of information. They are not as common in overcoming scientific challenges due to their data-intensive nature, an obstacle that Fung’s algorithm helps overcome.

In technical terms, the algorithm makes it possible for generative models to work with non-invertible structural representations, such as atom-centered symmetry functions.

Now that the group has learned how to use models to generate structure, they want to extend this to broader problems in materials design and discovery. This includes being able to generate structures with different chemical compositions as well.

Here, their algorithm becomes a tested, verified method using generative models to understand and overcome complex engineering problems.

“People who are interested in solving these kinds of problems in materials discovery, whether for specific applications, specific types of materials, or specific properties, can potentially use this approach, or at least take inspiration from it,” Fung said.