Before Forge

Sandia National Laboratories
Matthew Barry at computer
Sandia National Laboratories scientist Matthew Barry explores how machine learning can help translate engineering requirements into recipes for advanced materials. (Photo by Craig Fritz) Click on the thumbnail for a high-resolution image.

Inventions often depend on what's available. Edison's light bulb did not work until he could make durable filaments; LED lights languished for years until researchers discovered the right combination of semiconductors. But what if engineers could invent new technologies without waiting for the right materials?

A small research project at Sandia National Laboratories might someday help make this possible.

Aided by machine learning, Sandia computational materials scientist Matthew Barry and his team are scratching the surface of what could eventually become a universal guide to identifying advanced metal alloys. Such a resource could help engineers create virtually anything they can imagine; they could simply look up how to design the materials with the properties they need. Today, by contrast, engineers typically constrain their designs based on the materials available to them.

Although relatively few alloys have been evaluated so far, early results have been tantalizing. The Sandia team has shown their model can quickly and accurately predict the performance of elaborate combinations of metals. In June, the team reported findings in the journal npj Computational Materials that identified useful, practical guidelines and trade-offs their model had uncovered.

The research has special importance for Sandia's nuclear weapons work. Theoretically, designers could develop exotic materials for future stockpile needs.

Research focuses on developing advanced alloys

Barry developed the framework for his model as a graduate student at Georgia Tech, where he earned a doctorate degree in 2023.

Since coming to Sandia, he has applied his concept to modeling a class of material called high entropy alloys, which combine complex and sometimes unusual mixes of metals. These alloys can be remarkable. Some can stay strong at high temperatures instead of weakening like most metals do.

"High entropy alloys originally gained interest because they have high-temperature mechanical properties that exceed traditional high-performance superalloys," Barry said. "We need these advanced materials to make certain applications viable."

But evaluating them can be difficult and costly. Computer simulations are faster than making and testing physical samples, but they can still be too slow.

Machine learning models that help researchers discover new materials already exist, but they tend to lose accuracy with high entropy alloys. Generating enough training data for these advanced materials has been a major challenge.

But Sandia's model shines with small data sets. It has accurately predicted material properties after being trained on just tens of samples, not hundreds or thousands. That efficiency makes it more practical and cost-effective to explore new alloy combinations, and is raising some excitement among the research team that it could scale up easily.

"It should theoretically be able to generalize across the periodic table, we just haven't tried yet," Barry said.

Efficiency driven by a clever training decision

The initial success of the model might raise the question: Why aren't all models this efficient? One reason it works so well is a clever decision the team made about how to teach it chemistry.

This part takes a little explaining.

An atom is surrounded by electrons buzzing around it like bees around a flower. When atoms combine to form molecules or metal crystals, their electrons combine, too, to form a kind of swarm around a garden. Electrons gravitate toward certain parts of the crystal structure more than others, like bees might prefer certain flowers. If you were to dig up some titanium lilies and replace them with niobium daisies, the overall shape and activity of the swarm would shift in response.

<strong>National security engineering often requires new, innovative materials. Computational materials scientists like Matthew Barry, shown here, develop the tools to design them at Sandia National Laboratories. </strong>(Photo by Craig Fritz) Click on the thumbnail for a high-resolution image.
National security engineering often requires new, innovative materials. Computational materials scientists like Matthew Barry, shown here, develop the tools to design them at Sandia National Laboratories. (Photo by Craig Fritz) Click on the thumbnail for a high-resolution image.

Sandia researchers, in collaboration with scientists at Georgia Tech and NASA, trained their machine learning model, which they named VASt - short for Voxelized Atomic Structure - to recognize an alloy by its unique electron fingerprint. This is fundamentally different from how alloys are normally defined, by their elements.

That decision turned out to be smart. The model discovered relationships between electron distribution and material properties, sidestepping calculations it would have needed to learn if it had been trained on individual elements.

For instance, the model found that more vanadium increases ductility, a material's ability to bend without breaking, while tungsten increases strength. So in their June paper, the authors proposed that part of a future alloy recipe might be to add just as much tungsten as needed to make a material strong enough for a particular application, then pack in as much vanadium as possible to keep it from cracking.

In that most recent study, it is important to note, Barry focused on thermo-mechanical properties, or how materials respond to heat and physical stress. It was funded, in part, by Sandia's Laboratory Directed Research and Development program.

The researchers believe that by training the model on electrons instead of elements, they were able to find trends more easily and could extrapolate results to metals containing new elements not included in its training data.

However, the researchers need to better understand whether there are nuances that make this approach work for high entropy alloys but not necessarily for all alloys. They are also working to validate predictions experimentally and find their model's limitations.

Still, it is a promising start that could someday help scientists create new blends of materials they had never thought of before, enabling engineers to build machines they had only dreamed of.

/Public Release. This material from the originating organization/author(s) might be of the point-in-time nature, and edited for clarity, style and length. Mirage.News does not take institutional positions or sides, and all views, positions, and conclusions expressed herein are solely those of the author(s).View in full here.