Deep Learning Helps U Missouri Researchers Predict New Material Behaviors
- By Dian Schaffhauser
- 04/18/19
Graphene is a big, almost-magical subject in material science. This 2D layer of carbon elements one atom thick isn't just the thinnest and lightest material known on earth, but it's also extraordinarily strong, highly conductive (even though it's not made of metal) and can be used to create tiny devices that pack an electrical wallop in a miniscule space. It can also be used in myriad innovative applications — uber-thin touchscreens, stronger tools, more durable running shoes, supercapacitors that could make batteries obsolete, impermeable armor and new kinds of waterproofing, to name just a few.
But to make those products possible, there's a lot of time-consuming experimentation that needs to be done to figure out what happens when the atoms in the graphene are replaced with non-carbon atoms, such as during the process of creating the material used in the soles for those running shoes.
A research project at the University of Missouri is using "deep learning" to develop those new materials and hopes high-performance computing can speed the process along. In this type of machine learning, a computer model is fed multiple examples of classification to teach it how to solve a problem.
In this case, researchers in the university's College of Engineering fed "a few thousand" combinations of graphene structures and their properties into deep learning models. From there, according to an article about the project, over the course of two days, the computer could take over to predict the properties of "billions of other possible structures of graphene" without having to go through testing of each one separately.
"If you put atoms in certain configurations, the material will behave differently," explained Jian Lin, an assistant professor of mechanical and aerospace engineering, involved in the research. "Structures determine the properties. How can you predict these properties without doing experiments? That's where computational principles come in."
"You can train a computer to do what it would take many years for people to otherwise do," added Yuan Dong, a research assistant professor of mechanical and aerospace engineering and lead researcher on the study. "This is a good starting point."
"Give an intelligent computer system any design, and it can predict the properties," noted Jianlin Cheng, a professor of electrical engineering and computer science and also on the research team. "This trend is emerging in the material science field. It's a great example of applying artificial intelligence to change the standard process of material design in this field."
A study on the findings was recently published in npj Computational Materials.
The project was supported by funding from a university startup fund and grants issued by the NASA-Missouri Space Grant Consortium, the National Energy Technology Laboratory, the National Science Foundation and the U. S. Department of Agriculture.
About the Author
Dian Schaffhauser is a former senior contributing editor for 1105 Media's education publications THE Journal, Campus Technology and Spaces4Learning.