Machine learning potentials: MIT method improves alloy models

MIT researchers have developed a training-data strategy that makes machine learning potentials more accurate at predicting the behavior of chemically disordered metal alloys, the team reported June 19, 2026 in a paper in Sciences Advances. The technique builds datasets that expose models to a wider variety of local atomic environments, and the researchers say models trained on those datasets outperform models trained by random sampling and other popular approaches.

What did the team build and how does it work?

The team produced training datasets that prioritize chemical diversity in local atomic environments by identifying and replacing redundant atomic motifs, using information-theory measures to guide sampling. This front-loads the models with non-redundant examples so each training sample adds new chemical information, rather than repeating the same local environments many times. Rodrigo Freitas, MIT’s TDK Career Development Professor in Materials Science and Engineering and the paper’s senior author, said the field’s core obstacle is clear: "The real challenge in our field is modelling these chemically disordered phases." The group applied the approach to chemically diverse metal alloys and trained machine-learning potentials that capture the subtle energetic biases toward certain local chemical configurations the authors describe in the paper.

How does this compare with existing dataset and modeling methods?

Models trained on the researchers’ motif-based datasets predicted material properties more accurately than models trained using random sampling or another popular sampling method, and the team reports their models were more accurate than much larger models created by companies like Google and Microsoft. The paper contrasts the new dataset construction with the leading brute-force approach, which the authors say can require more than 100,000 hours of computation to create the training data for a single material and still does not transfer well when material composition changes. The MIT method avoids those expensive brute-force procedures by optimising the informational content of the training set, swapping out repeated examples for previously unseen local environments.

What specific results did they show?

The researchers tested their models across different alloys and properties and used University of Sheffield Senior Lecturer Lewis R. Owen’s experimental data to compare simulations against real measurements of atomic ordering. Daniel Xiao led simulations showing the team’s models could predict phase diagrams that closely matched experimental data. The paper’s title is "Machine learning potentials for modeling alloys across compositions," and first author Killian Sheriff PhD ’26 worked with coauthors Daniel Xiao and Yifan Cao to test the approach broadly. The work was supported by the U.S. Air Force Office of Scientific Research.

Why it matters

The approach tackles a practical bottleneck in materials discovery: simulation fidelity for chemically disordered solids. Industry workflows depend on phase diagrams and accurate predictions of which phases will form during casting, welding, or heat treatment. By making machine-learning potentials both more accurate and less reliant on enormous brute-force datasets, the method can reduce the gap between simulation and experiment and lower the cost of exploring new alloys for aerospace, energy, and computing applications.

What to watch

Follow whether the group or outside teams integrate the motif-based sampling into standard materials-engineering toolchains and whether the method scales to nonmetal systems; the paper notes the approach could be adapted to semiconductors and other materials. A concrete next milestone will be independent demonstrations that these models can predict mechanical performance or radiation tolerance in alloys targeted for industrial or defense use.