Hard machine learning can predict hard materials

Superhard materials are in high demand by industry, for use in applications ranging from energy production to aerospace, but finding suitable new materials has largely been a matter of trial and error, based on classical hard materials such as diamonds. Until now.

In a paper in Advanced Materials, researchers from the University of Houston (UH) and Manhattan College report a machine-learning model that can accurately predict the hardness of new materials, allowing scientists to more readily find compounds suitable for use in a variety of applications.

Materials that are superhard – defined as those with a hardness value exceeding 40 gigapascals on the Vickers scale, meaning it would take more than 40 gigapascals of pressure to leave an indentation on the material's surface – are rare.

"That makes identifying new materials challenging," said Jakoah Brgoch, associate professor of chemistry at UH and corresponding author of the paper. "That is why materials like synthetic diamond are still used even though they are challenging and expensive to make."

One of the complicating factors is that the hardness of a material may vary depending on the amount of pressure exerted, known as load dependence. That makes testing a material experimentally complex and ensures that computational modeling is currently almost impossible.

The model reported by the researchers overcomes these difficulties by predicting the load-dependent Vickers hardness based solely on the chemical composition of the material. The researchers report finding more than 10 new stable borocarbide phases that show promise; work is now underway to design and produce the materials so they can be tested in the lab.

Based on the model's reported accuracy, the odds are good. Researchers reported its accuracy at 97%.

First author Ziyan Zhang, a doctoral student at UH, said the database built to train the algorithm is based on data involving 560 different compounds, each yielding several data points. Finding the data to build a representative dataset required poring over hundreds of published academic papers.

"All good machine learning projects start with a good dataset," said Brgoch, who is also a principal investigator with the Texas Center for Superconductivity at UH. "The true success is largely the development of this dataset."

Researchers have traditionally used machine learning to predict a single variable of hardness, Brgoch said, but that doesn't account for the complexities of a property like load dependence, which he said still isn't well understood. That makes machine learning a good tool, despite earlier limitations.

"A machine learning system doesn't need to understand the physics," he said. "It just analyzes the training data and makes new predictions based on statistics."

Machine learning does have limitations, though. "The idea of using machine learning isn't to say, 'Here is the next greatest material', but to help guide our experimental search," Brgoch said. "It tells you where you should look."

This story is adapted from material from the University of Houston, with editorial changes made by Materials Today. The views expressed in this article do not necessarily represent those of Elsevier. Link to original source.