Issue Date: December 7, 2015 | Web Date: December 3, 2015
Computers “Learn,” Then Predict Materials Properties
Some materials researchers believe that computers are the future: Teach them well, and they’ll help lead the way.
Scientists led by Logan Ward and Chris Wolverton at Northwestern University recently developed a machine-learning algorithm that mines existing data to predict the properties of uncharacterized solid-state materials. The technique could help researchers discover new materials for a range of applications, including energy generation and storage.
The approach can predict basic properties, such as a material’s electronic band gap, roughly 1 million times as fast as conventional computational methods, specifically density functional theory (DFT), according to Ward.
Furthermore, the machine-learning technique makes predictions that DFT cannot, such as alloy compositions that could form metallic glasses, Ward announced last week at the Materials Research Society Fall Meeting & Exhibit in Boston.
But DFT and machine learning aren’t exactly competitors. In fact, DFT is helping teach machine-learning algorithms, Ward told C&EN.
DFT predicts materials properties using calculations based on the electron density within a substance. Over the past five years, Wolverton’s team has developed high-throughput DFT methods to populate a large database of materials and their computed properties.
Information from this and other materials databases provides “training” sets that teach a machine learner. These training sets contain computed or measured properties, as well as fundamental attributes of the materials, such as the radii of their atoms, Ward explained.
The algorithm then looks for relationships between these attributes and materials properties to predict the behavior of unexplored materials.
“When it comes to machine learning, what you want is leads. You want hits,” said Jason Hattrick-Simpers, a chemical engineer at the University of South Carolina who was not involved with the project.
At the Boston meeting, Ward unveiled predictions for eight unstudied materials that could form metallic glasses, which impressed Hattrick-Simpers. If researchers analyze these alloys and find the materials don’t behave as predicted, the results still provide the algorithm with better data to refine its forecasting. “Win or lose, it’s a gain for the system,” he said.
The new machine-learning code is free to download, Northwestern’s Wolverton said.
- Chemical & Engineering News
- ISSN 0009-2347
- Copyright © American Chemical Society