Nature publishes UAB analysis of deep neural networks to discover new materials

“Many of the problems our society faces could be solved with new materials, which makes these types of neural networks a game changer in the field of material science,” according to Adam Smith.

An analysis of the use of deep neural networks to accelerate the discovery of new materials that was written by two professors and a graduate student at The University of Alabama at Birmingham (UAB), a partner institution of the Future Technologies & enabling Plasma Processes (FTPP) program, has been published as a Nature Computational Science Comment.

“The purpose of our paper was to review the state-of-the-art methods for crystal structure prediction using generative deep neural networks and give our comments on what are the strengths and weaknesses of each approach, as well as contribute to the discussion with our knowledge of promising routes of improvement,” says Adam D. Smith, a UAB graduate student in physics who coauthored the paper with Dr. Da Yan, an assistant professor in UAB’s Department of Computer Science, and Dr. Cheng-Chien Chen, an associate professor in UAB’s Department of Physics.

FTPP sponsored

The work was sponsored by FTPP, a $20 million effort managed at The University of Alabama in Huntsville to transition plasma research into commercial applications and establish a plasma workforce in Alabama. FTPP is funded by the National Science Foundation’s Established Program to Stimulate Competitive Research.

“The purpose of these generative deep neural networks is to accelerate how fast we can discover new materials with tailored properties for specific applications in the industrial, energy, medical and aerospace sectors using data already available to us,” says Smith, who is from Naperville, Ill.

“When a property of interest is defined and then a large quantity of materials that have higher or lower values for that property are used as training examples for the networks, we conclude that the performance of these neural networks is adequate,” he says. “But we stress that further research into generative neural networks that are able to learn from fewer data points is of utmost importance.”

The comment article is intended to communicate that there are people who value this work and to give new ideas from the authors’ perspective, Smith says.

“Our team at UAB, headed by Dr. Cheng-Chien Chen, mostly uses high-performance computing and first-principles calculations to uncover the physics which lead to material properties, but we too would like to see future generative models give us new directions for where to find interesting physics in materials,” he says. 

finding new materials

The dream of material science is to discover materials that have properties specifically tailored to various types of applications, Smith says.

“When we train a generative neural network model with a specific material property in mind, the model examines the types of atoms and arrangements of those atoms in a crystal structure that lead to the maximization of the property,” he says. “Generative models can learn from the data collected from material scientists around the world and generate new crystals which have maximized properties, thus accelerating the development of new materials.”

Generative models can be useful in any industry that benefits from specialized materials, like ultrahard materials for machining and mining, oxidation-resistant coatings for the nosecones of rockets for aerospace industries, or materials for lasing which have tailored wavelength and power for surgery, or detection of disease.

“Many of the problems our society faces could be solved with new materials, which makes these types of neural networks a game changer in the field of material science,” Smith says.

“Science is a collaborative process and we take every opportunity we can to contribute to the discussion, in this case by doing literature review around the subject of generative models for crystal structure prediction and giving our thoughts about where these types of models need to improve in order to be more broadly successful in the search for new materials.”