News

MIT researchers created a technique that captures chemical arrangements across materials to improve predictions of how metal alloys and other complex materials will behave. This figure compares a random sampling approach to the researchers’ new motif-based sampling. Credit: MIT

Designing the materials that go into jet engines, computer chips, medical implants and other cutting-edge technologies could soon get faster and far less expensive — thanks to a new artificial intelligence approach developed by Massachusetts Institute of Technology (MIT) researchers using a powerful supercomputer in San Diego. The advance could accelerate progress on stronger aerospace components, more efficient semiconductors and other high-performance materials, all while reducing reliance on costly, time-consuming lab experiments. For consumers, that could eventually mean better batteries, more durable electronics and safer aircraft parts reaching the market sooner.

The team at MIT used National Science Foundation (NSF) ACCESS allocations on the Expanse system at the San Diego Supercomputer Center (SDSC), located at University of California San Diego Halıcıoğlu School of Data Science and Computing, to build machine-learning models that describe how atoms interact inside complex metal alloys, then use those models in atom-by-atom simulations. The findings were published in the Science Advances journal.

That might sound abstract, but atomic arrangement matters enormously. It determines whether a material is strong or brittle, conductive or insulating and how well it holds up under extreme heat or stress.

The challenge is that in many advanced materials, especially so-called "high-entropy alloys" made by blending several different metals together, atoms don't line up in neat, predictable patterns. Instead, they arrange themselves in a chaotic, almost random way that has long made it difficult for computer models to capture what's really happening at the atomic scale.

To tackle this, Rodrigo Freitas, an assistant professor of materials science and engineering at MIT, and his team developed a new method called motif-based sampling. Rather than trying to random atomic arrangements, the approach breaks a material's structure down into small, recognizable chemical "neighborhoods," then optimizes how often different arrangements appear. This gives the AI framework a much richer and more realistic picture of the material's chemistry to learn from.

“The challenge was not just building a machine-learning model,” Freitas said. “It was building the right training data for materials where atoms can arrange themselves in many different local environments. Motif-based sampling gives us a way to choose those examples more systematically, and the NSF ACCESS allocation on SDSC’s Expanse gave us the computing power to generate the data, train the models and test them in large atom-by-atom simulations.”

Once trained, the AI models were used inside atom-by-atom simulations. The researchers used Monte Carlo simulations to test whether the models could correctly predict phase diagrams, essentially maps that show how a material's structure changes with temperature and composition, for both simple two-metal alloys. The predictions lined up well with existing experimental data and established industry models.

The team then pushed further, using simulations powered by the machine-learning models to estimate melting points for complex alloys where real-world experimental data is scarce or difficult to obtain. The predictions held up well. Additional tests showed the AI could also accurately predict how materials expand when heated and how much heat they can store, both important properties for engineers designing components that operate in extreme environments.

Beyond matching known results, the framework opened the door to studying properties that have historically been very hard to predict, such as the material's resistance to a type of structural defect that affects strength and durability. According to Freitas, the simulations were able to capture how this property changes depending on a material's chemical makeup — insight that could help guide the design of tougher, more reliable alloys.

"Such computational accuracy can support materials design and development, particularly in scenarios where experimental exploration is prohibitively challenging or expensive," Freitas said. "We would not have been able to complete our work without the support of the NSF ACCESS allocations on SDSC's Expanse."

The time on SDSC’s Expanse was supported by NSF ACCESS allocation (no. MAT210005).