The Search for New Metals is Now Easy with Machine Learning

Machine learning has the potential to significantly speed up the search for new metals.

According to a recent study, machine learning could aid in the creation of new metal types with advantageous characteristics like resistance to rust and high temperatures. A variety of industries could benefit from this; for instance, spacecraft could be improved with metals that function well at lower temperatures, while boats and submarines could benefit from corrosion-resistant metals. Currently, attempts to produce new metals are mostly conducted in laboratories by scientists. Typically, they begin with one well-known element, such as iron, which is readily available and malleable, and then add one or two more to examine how it affects the base material. Trial & error is a hard process that invariably produces more failures than successful outcomes.

However, the latest report, which was just published in Science, implies that researchers can much more precisely forecast which metal combinations will show potential using AI. Using this technique, researchers at the Max Planck Institute were able to uncover 17 interesting new metals. Invar, which describes how much materials expand or contract when subjected to high or low temperatures, was something the team was looking for in metals. Low invar metals do not expand or contract in size at high temperatures. Ziyuan Rao, a materials science researcher at the Max Planck Institute and a co-author of the article, adds that they are frequently employed in industries where that quality is advantageous, such as the transportation and storage of natural gas.

Through a combination of AI and laboratory tests, the team was able to discover these new metals. First, they had to overcome a major obstacle: a dearth of available data for the machine-learning models to be trained. Several hundred data points representing the characteristics of current metal alloys were used to train the models. This information was used by the AI system to forecast the appearance of new metals with low invar. The findings of the measurements were then given back into the machine-learning model by the researchers when they produced those metals in a lab. The researchers tested the suggested metal combinations, fed the results back into the model, and so on until the 17 potential new metals emerged.

The results may open the door for more machine learning applications in the field of materials science, which currently mainly depends on laboratory testing. Additionally, according to specialists in materials science, the method of employing machine learning to produce predictions that are subsequently verified in the lab might be modified for discovery in other domains like chemistry and physics. According to Michael Titus, an assistant professor of materials engineering at Purdue University who was not involved in the research, it is worthwhile to examine the conventional method by which novel compounds are typically produced in order to comprehend why it is an important development. The experimentation in the lab is laborious and ineffective.

Finding materials with a unique quality is genuinely like finding a needle in a haystack, according to Titus. He frequently informs his brand-new graduate students that there are probably a million potential brand-new materials out there just waiting to be found. Machine learning might aid in the choice of study directions. What the team was able to accomplish with the novel technique shocked Easo George, a professor of materials science and engineering at the University of Tennessee who was not involved in this study. He remarks, "It's pretty stunning."

The team hopes to find new alloys with several desirable properties in the future with the aid of machine learning. George concurs that the future of materials research will depend heavily on computational approaches. People have tried to scan very big spaces experimentally, but that is highly time and money consuming. So, he predicts that the machine-learning approach will prevail. "Are you discovering something useful? That's the test", he says.