Searching for superior materials

The process of developing a new material sounds simple: Just combine different ratios of different elements until you come up with something that meets your specifications. In reality, the process is much more tedious and complex—that’s why scientists at Los Alamos National Laboratory are turning to automation and artificial intelligence (AI) for a hand.

“Traditionally, humans mix and match elements to try and make materials, but it’s too much for a human to do manually,” says Saryu Fensin, leader of the Lab’s Quasi-static and Dynamic Behavior of Materials team. “If you pick just nine elements on the periodic table and think about the combinations to generate binary (two-element), ternary (three-element), quaternary (four-element) alloys (mixtures of two or more elements), there are 300 unique compositions.”

Through robotic automation and AI, however, Fensin aims to reduce the time it takes to manufacture and characterize new alloys  from hours to minutes.

Fensin is developing a continuous automated production cycle for alloys that starts with melting specific amounts of two elements together. The cooled combination is shaped into a thimble-sized pellet. From there, a robotic arm takes the pellet and places it into a load frame—a device used to test the mechanical properties of materials, including their strength, stiffness, and displacement. 

The data generated from the load frame goes into a database that trains an AI model. Each new alloy is a new data point. The goal is a model that can distinguish patterns across those data points, extrapolate patterns to unknown alloys, and ultimately propose a ranked list of known and unknown alloys optimized for specified properties. 

However, developing such a model requires a significant amount of data. “The biggest issue we have is we are data poor, and the way AI works is you train it on data where variables are systematically changed until it learns the patterns,” Fensin says. “So, we’re going to try brute force and generate lots of data to find trends.”

Fensin says that testing hundreds of millions of possible combinations is unrealistic even with an automated process. More realistically, she expects to produce and evaluate hundreds of thousands of combinations. “Tweaking compositions by a single percentage point will be too much, so in the first round we might change elemental compositions by 20 percent and see if properties change,” she says. “If we see changes, then we can zoom in and do a smaller change.” 

Fensin is quick to note, however, that “there’s always an exception to the rule, and there could be some composition that doesn’t follow the patterns. So, the challenge is going to be gathering enough data and validating predictions. If the model predicts a composition in a microstructure that would give us certain properties, we need to go make it. It can’t just be a computer exercise.”

This gets at another point Fensin is passionate about: the role of humans in this process. She believes that repetitive work should be automated so that people are free to support, utilize, and expand on that automated work. She refers to this as “collaborative robotics,” in which humans and automatons work together to maximize what each does efficiently.

Fensin sees three primary use cases for her team’s work: providing data to customers about existing alloys, fundamental research and development of new alloys in general, and designing new alloys for specific applications, including for stockpile stewardship—the maintenance of America’s nuclear weapons.

“This investment in fundamental research will be beneficial for stockpile stewardship,” Fensin says. “We can build a scientific foundation and start doing research so that we can be responsive to any issues that arise.” ★