Companies working on the frontiers of aerospace, energy, and computing are always looking for new things to improve performance. But to understand how those materials will behave once they’re inside rockets or chips, companies must first make the materials and test them. That’s because even the most powerful simulation techniques struggle to model the complex chemical structures in many of today’s solids. The problem adds cost and time to the creation of new materials.
Now a team of MIT researchers has developed a way to accurately show the behavior of metals, regardless of the complexity of their chemical composition. At the center of the approach are machine learning models that make simulations faster and more accurate. The researchers developed those models by creating training data sets that capture the variation in atomic positions in chemically inactive materials.
In a new paper on Advances in Sciencethe researchers show their method can be used to accurately predict the properties of the materials of a diverse group of metal alloys under many conditions. They also show how the method can be used to create new things, especially in situations where testing is expensive.
“The paper’s focus is metal alloys, which is the field I work in, but this can be adapted to other types of materials, such as semiconductors,” said senior author Rodrigo Freitas, TDK’s Career Development Professor in Materials Science and Engineering. “This isn’t specific to any one application — you can use this method to build new sustainable metals, new aerospace materials, and more. That’s what makes this exciting.”
Joining Freitas on the paper are co-first authors Killian Sheriff PhD ’26; MIT PhD students Daniel Xiao and Yifan Cao; and University of Sheffield Senior Lecturer Lewis R. Owen.
Modeling instruments
Physical materials are largely determined by the internal arrangement of their chemical properties. Even if two materials have the same chemical composition, different chemical compositions can make the difference between a material that breaks and one that deforms without breaking.
Capturing those differences requires simulating the material atom by atom. To do that, researchers rely on models that describe how atoms interact. Over the past two decades, machine learning has become the most accurate way to build those models. Such models work well when the chemical arrangement within the material follows highly ordered patterns, but this is not the case with many solid materials, whose atomic chemical systems are disordered and vary from region to region.
“The real challenge in our field is to model these chemically inefficient phases,” Freitas said. “Chemical disorder means that there is a wide variety of local chemical properties, which are difficult for a machine learning model to learn. This is a problem because all the metals we use are chemically free.”
The problem comes down to the lack of training data representing those atom-by-atom simulations. The current best way to create such data is brute force, often requiring more than 100,000 hours of computation to create a single object’s training data. However, it does not transfer well when researchers change the composition of the material.
In previous work, Freitas’ group had developed a method to measure the chemical complexity of solids by analyzing the frequency and splitting of small groups of atoms. In this study, the researchers used that ability to create better training datasets. They used a mathematical technique known as information theory to create training data sets that capture the wide range of local chemical properties within unstructured materials. The method works by swapping atoms from the samples to reduce duplication and expose the model to chemical properties that might otherwise be missed.
“We’ve continued to develop the training set to cover more local areas,” Freitas said. “If the same type of area is seen multiple times, we replace the obsolete examples with ones that the model has never seen before. That makes the training more instructive because each example adds something new.”
When trained on the researchers’ data sets, the models predicted material properties more accurately than models trained using random sampling or other popular sampling methods.
“The starting point for all these atom-by-atom simulations is: Can you accurately describe the chemical bond between the atoms?” Freitas explained. “Otherwise, it can still teach you about materials in general, but it doesn’t tell you what will happen to certain materials in the real world. This approach makes the simulations more reliable in terms of their chemistry, so they better reflect what’s happening in the materials.”
The researchers used their method to create machine learning training datasets for a group of chemically different metal alloys. Using a set of machine learning models, they showed the models trained on their dataset were more accurate than the larger models created by companies like Google and Microsoft.
“We got to the point where we were convinced that it works without using these cruel and expensive methods,” said Freitas. “I told Killian, ‘This is a good paper. But if you can show that simulations of these models can now accurately predict properties of useful materials, then it’s a very good paper.’ Killian took that to heart and tested it as much as possible.”
Sheriff worked with Xiao and Cao to explore how to use different alloys and structures. The team also mapped Owen’s experimental data to compare simulations against actual measurements of atomic ordering in alloys.
From lab to industry
The method works, in part, by capturing hidden patterns in sample data. The researchers describe the patterns in the paper as “subtle strong biases toward a particular configuration of local chemicals.”
That small difference in strength is important because it determines what phases form in the alloy, how those phases change with temperature and composition, and ultimately what properties they will have. As one experiment, Daniel Xiao led simulations showing that group models can predict phase diagrams that closely match experimental data. Phase diagrams map which phases are stable across different temperatures and chemical compositions, and is a central tool for designing and processing alloys.
“Category diagrams are one of the main ways people connect modeling to real-world processing decisions,” Freitas said. “When you’re welding, casting, or heating an alloy, you need to know what phases are likely to form under different conditions. Our goal is to make these types of predictions accurate enough, and accessible enough, to become part of the way people design things.”
Researchers are now using the method to study how changing the alloy’s composition affects mechanical properties and radiation tolerance, with the goal of designing materials that stay strong and withstand damage in harsh environments. And they’re working to make the approach easier to use with the kinds of tools and workflows developers already rely on.
“The industry will not change the way they do things if what they create does not fit into their existing work processes,” said Freitas. “The goal is to make these predictions useful in areas where material decisions are made.”
The research was supported by the US Air Force Office of Scientific Research.
