Killian Sheriff was staring at a screen full of atomic configurations, each one a slightly different jumble of nickel, chromium, and iron atoms, when the breakthrough clicked: what if the key to predicting how alloys behave isn’t in perfect order, but in embracing the chaos? At MIT’s Department of Materials Science and Engineering, Sheriff, a Ph.D. graduate, and his advisor Rodrigo Freitas had been wrestling with a decades-old bottleneck in materials science — the inability to accurately simulate how real-world metal alloys behave at the atomic level. Most alloys, from the steel in bridges to the superalloys in jet engines, are chemically disordered, meaning their atoms are arranged in unpredictable, ever-varying patterns. Traditional simulation methods, even advanced machine-learning models, fail here because they’re trained on overly simplistic or repetitive data. But now, Sheriff, Freitas, and their team have cracked the code by designing smarter training data that captures the true diversity of atomic environments.
The implications are profound. Companies in aerospace, energy, and computing spend years — and millions — developing and testing new materials through trial and error. A single new turbine alloy can take over a decade to certify. But with this new technique, published in Science Advances, researchers can simulate material behavior with unprecedented accuracy, slashing both time and cost. The team used information theory — a mathematical framework originally developed for data compression — to systematically generate training sets that expose machine-learning models to the full spectrum of local atomic arrangements in disordered alloys. Instead of relying on random sampling or brute-force computation, they optimized each training example to maximize informational value, replacing redundant atomic configurations with rare or underrepresented ones.
The results speak for themselves. Models trained on these enhanced data sets outperformed conventional approaches in predicting key properties like elastic stiffness and thermal expansion across a wide range of alloys, including nickel-based superalloys and high-entropy systems. What once required over 100,000 hours of computation for a single material can now be achieved with far greater accuracy and efficiency. And because the method is generalizable, it’s not limited to metals — it could accelerate the design of semiconductors, battery materials, and sustainable steels. “This is not specific to any one application — you could use this approach to create new sustainable steels, new materials for aerospace, and more,” says Freitas, MIT’s TDK Career Development Professor. “That’s what makes this exciting.”
With collaborators including MIT Ph.D. students Daniel Xiao and Yifan Cao, and University of Sheffield’s Lewis R. Owen, the team has opened a new pathway in materials discovery. Where once scientists had to make and test materials physically, they can now simulate them with confidence. The future of innovation isn’t just faster — it’s smarter, guided by algorithms that finally understand the beautiful mess of atoms.