This full spectrum of structural configurations is still difficult to capture and further limits the transferability of contemporary machine learning models on material behavior.
Now, a novel framework to model metallic behavior has been created by researchers at the Massachusetts Institute of Technology (MIT).
These refined models can reproduce the thermodynamic behavior of metallic alloys as functions of composition and structure over their full ranges.
Machine learning typically speeds up materials simulations, but the researchers enhanced reliability by training datasets that represent the wide range of atomic configurations in chemically disordered materials.
As a result, machine learning models trained on these datasets predicted material properties more accurately than those trained with random or standard sampling methods.
The chemical composition of a material, the relative amounts of its constituent elements, exerts a strong influence on its properties. Changes to this makeup can fundamentally restructure how the material is laid out internally, resulting in everything from perfectly ordered compounds to completely mixed solid solutions. This full spectrum of structural configurations is still difficult to capture and further limits the transferability of contemporary machine learning models on material behavior.
Now, a novel framework to model metallic behavior has been created by researchers at the Massachusetts Institute of Technology (MIT). Building on information theory, the team developed a new approach that samples chemical patterns and, in some cases, designs machine learning potentials (MLPs) more efficiently than state-of-the-art approaches. These refined models can reproduce the thermodynamic behavior of metallic alloys as functions of composition and structure over their full ranges.
Machine learning typically speeds up materials simulations, but the researchers enhanced reliability by training datasets that represent the wide range of atomic configurations in chemically disordered materials. As a result, the models are more representative of true structural complexity, resulting in improved physical predictions.
Senior author Rodrigo Freitas, MIT’s TDK Career Development Professor in Materials Science and Engineering, said, “The focus of the paper is metallic alloys, which is the field I work in, but this could be adapted to other types of materials, like semiconductors. This is not specific to any one application; you could use this approach to create new sustainable steels, new materials for aerospace, and more. That’s what makes this exciting.”
Researchers discovered 268 new high-performance metal alloys
Researchers previously developed a metric for the chemical complexity of solid materials based on the geometry of small atomic clusters, which are groups of atoms. The present study, building on that premise, applied this method to create optimized training datasets. Using information theory, they produced datasets that span a much larger variety of local atomic environments in disordered systems.
The approach works by swapping atoms to avoid repetition and expose the model to new chemical patterns. As a result, machine learning models trained on these datasets predicted material properties more accurately than those trained with random or standard sampling methods.
Freitas explains, “The starting point for all these atom-by-atom simulations is: Are you able to accurately describe the chemical bond between atoms? If not, it can still teach you about materials in general, but it doesn’t tell you what will happen to specific materials in the real world. This approach makes the simulations high fidelity in terms of their chemistry, to better reflect what’s happening to materials.”
The researchers tested their new technique on chemically diverse metal alloys. Machine‑learning models trained with their datasets proved more accurate than much larger models built by companies like Google and Microsoft.
This simple metal alloy is the toughest material on Earth
“We reached a point where we knew it worked without relying on expensive brute‑force methods,” Freitas said. “I told Killian, ‘This is a good paper. But if you can show that these models now predict useful material properties accurately, then it becomes a very good paper.’ Killian took that to heart and tested it as widely as possible.”
The team tested the approach across different alloys and properties, while the experimental data helped compare the simulations with real measurements of atomic ordering.
The method works by spotting hidden patterns in the data, which the researchers call ‘subtle energetic biases’ toward certain atomic arrangements. These small differences matter because they decide which phases form in an alloy, how those phases shift with temperature and composition, and ultimately what properties the material will have.
As one test, Daniel Xiao ran simulations showing that the models could predict phase diagrams that closely matched experimental results. Phase diagrams map which phases are stable under different conditions, making them a key tool for designing and processing alloys.
The researchers are now using their method to examine how changes in an alloy’s composition affect its strength and radiation resistance. Their goal is to design materials that stay tough in extreme environments. They’re also working to make the approach easier to use with the everyday tools and workflows that materials engineers already rely on.
Journal Reference:
Killian Sheriff, Daniel Xiao, Yifan Cao, Lewis Owen, and Rodrigo Freitas et al. Machine learning potentials for modeling alloys across compositions. Science Advances. DOI: 10.1126/sciadv.aea9951
