Machine learning could significantly accelerate the search for new metals

It's a big step forward that could be used everywhere from outer space to the bottom of the ocean

According to new research, machine learning could aid in the development of new types of metals with useful properties such as resistance to extreme temperatures and rust. Metals that perform well at lower temperatures, for example, could improve spacecraft, whereas metals that resist corrosion could be used for boats and submarines.

Currently, scientists conduct experiments in the lab to discover new ways to combine metals. Typically, they begin with one well-known element, such as iron, which is inexpensive and malleable, and then add one or two others to see how they affect the original material. It's a time-consuming trial-and-error process that inevitably produces more duds than useful results.

However, a new paper published earlier this month in Science suggests that using AI, researchers can far more precisely predict which metal combinations will show promise.

Using this method, Max Planck Institute researchers identified 17 promising new metals. The team was looking for metals with low "invar," or how much materials expand or contract when exposed to high or low temperatures. 

Metals with low invar do not change size at high temperatures. They are often used in industries where that property is useful, like transporting and storing natural gas, says Ziyuan Rao, a researcher in materials science at the Max Planck Institute and co-author of the paper.

The team discovered these new metals using a combination of AI and lab experiments. They had to overcome a significant challenge first: a lack of existing data on which to train the machine-learning models. They trained the models using the available data, which consisted of several hundred data points describing the properties of existing metal alloys. The data was used by the AI system to make predictions for new metals with low invar.

The metals were then created in a lab, measured, and fed back into the machine-learning model. The model suggested combinations of metals, the researchers tested them, and the results were fed back into the model until 17 promising new metals stood out.

The findings could pave the way for more machine learning applications in materials science, a field that still heavily relies on laboratory experimentation. Also, experts in materials science think that the way machine learning is used to make predictions that are then tested in the lab could be used to find new things in other fields, like chemistry and physics.

To understand why this is such a significant breakthrough, consider how new compounds are typically created, Michael Titus, an assistant professor of materials engineering at Purdue University who was not part of the research, says that this is the case. Tinkering in the lab is a time-consuming and inefficient process.

"Finding materials that exhibit a special property is truly like finding a needle in a haystack," Titus says. He frequently tells his new graduate students that there are potentially a million new materials just waiting to be discovered. Machine learning could assist researchers in deciding which paths to take.

Easo George, a materials science and engineering professor at the University of Tennessee who was not involved in this study, was taken aback by what the team was able to accomplish with the new technique.

"It's quite impressive," he remarks.

The team hopes to use machine learning in the future to help discover new alloys with more than one desirable property. Computational methods, according to George, will be critical to the future of materials science.

"The machine-learning approach is likely to be dominant," he says, "because people have tried to scan very large spaces experimentally, but that is very time-consuming and expensive." "The question is, are you discovering anything useful?"