Companies operating at the intersection of aviation, energy and IT are constantly looking for fresh materials to improve performance. But to understand how these materials will actually behave once they’re in rockets or on computer chips, companies must first make the material and then test it. This is because even the most powerful simulation techniques have difficulty modeling the elaborate chemical systems in most current solid materials. The problem increases the costs and time of material innovation.
Now a team of MIT researchers has created a way to accurately model the behavior of metals, regardless of the complexity of their chemical systems. At the heart of this approach are machine learning models that make material simulations faster and more precise. Scientists have improved these models by building training datasets that capture the diversity of atomic environments in chemically disordered materials.
In new paper in researchers demonstrated that their approach could be used to accurately predict the material properties of a diverse group of metal alloys under various conditions. They also showed how this approach can be used to develop fresh materials, especially in scenarios where experimentation is pricey.
“The paper focuses on metal alloys, which is the field I work in, but could be adapted to other types of materials, such as semiconductors,” says senior author Rodrigo Freitas, MIT’s TDK Career Development Professor in Materials Science and Engineering. “It’s not specific to one application – you can use this approach to create new sustainable steels, new materials for the aerospace industry, and more. That’s what makes it exciting.”
Joining Freitas on the article were first author Killian Sheriff PhD ’26; MIT PhD students Daniel Xiao and Yifan Cao; and University of Sheffield Senior Lecturer Lewis R. Owen.
Metal modeling
The properties of materials are determined primarily by the internal arrangement of their chemical elements. Even if two materials have the same mix of chemical elements, different chemical arrangements can make the difference between a brittle material and one that deforms without breaking.
Capturing this distinction requires atom-by-atom simulation of materials. For this purpose, researchers rely on models describing the interactions of atoms. Over the past two decades, machine learning has emerged as the most precise method for building these models. Such models work well when the chemical systems inside materials have highly ordered patterns, but this is not the case for most solid materials, whose atomic chemical arrangements are disordered and vary from region to region.
“The real challenge in our field is modeling these chemically disordered phases,” says Freitas. “Chemical disorder means that there is a huge variety of local chemical environments, which is difficult to learn in a machine learning model. This is a problem because every metal we use in practice is chemically disordered.”
The problem comes down to the lack of representative training data for atom-by-atom simulations. The current leading approach to creating such data is brute force and often requires over 100,000 hours of computation to create training data for a single material. Even then, it doesn’t transfer well when researchers change the composition of the material.
In previous work, Freitas’ group developed a way to measure the chemical complexity of solid materials by analyzing the frequencies and spacing of miniature groups of atoms. In this study, researchers used this opportunity to create better training datasets. They used a mathematical approach known as information theory to generate training datasets that take into account a wider range of local chemical environments inside disordered materials. This method replaces atoms in samples to reduce repeatability and expose the model to a chemical environment that might otherwise be missed.
“We continuously optimized the training set to include as many different local environments as possible,” says Freitas. “If the same type of environment appeared multiple times, we replaced redundant examples with ones the model hadn’t seen before. This makes the training set much more informative because each example adds something new.”
When trained on the researchers’ datasets, the models predicted material properties more accurately than models trained using random sampling or other common sampling methods.
“The starting point for all atom-by-atom simulations is the question: Can you accurately describe the chemical bond between atoms?” explains Freitas. “If not, it can still teach you about materials in general, but it won’t tell you what will happen to specific materials in the real world. This approach makes the simulations very chemically faithful to better reflect what happens to the materials.”
The researchers used their technique to create machine learning training datasets for a group of chemically diverse metal alloys. Using a suite of machine learning models, they showed that the models trained on their datasets were more precise than the much larger models created by companies like Google and Microsoft.
“We got to the point where we were confident it would work without expensive brute force methods,” Freitas says. “I said to Killian, ‘That’s a good article.’ But if you can show that simulations using these models can now accurately predict the properties of useful materials, it will become a very good paper. ” Killian took this to heart and tested it as widely as he could.”
Sheriff worked with Xiao and Cao to test this approach for different alloys and properties. The team also used Owen’s experimental data to compare the simulations with actual measurements of the ordering of atoms in the alloys.
From laboratory to industry
This method relies, in part, on spotting hidden patterns in sample data. The researchers describe the patterns presented in the paper as “subtle energetic biases toward certain local chemical configurations.”
These small energy differences matter because they determine which phases form in the melt, how those phases change with temperature and composition, and ultimately what properties the material will have. In one test, Daniel Xiao ran simulations showing that the team’s models could predict phase diagrams that closely matched experimental data. Phase diagrams illustrate which phases are stable at different temperatures and chemical compositions, and are a major tool for alloy design and processing.
“Phase diagrams are one of the main ways to connect materials modeling with real processing decisions,” says Freitas. “If you’re welding, casting or heat treating an alloy, you need to know which phases are likely to form under different conditions. Our goal is to ensure that these kinds of predictions are precise and accessible enough to become part of the way people design materials.”
Scientists are now using this approach to study how changing alloy composition affects mechanical properties and radiation tolerance, with the goal of designing materials that remain powerful and resistant to damage in harsh environments. They are also working to make the method easier to employ, using tools and workflows that materials engineers already employ.
“The industry won’t change the way it works if what you’re building doesn’t fit into existing operating procedures,” Freitas says. “The goal is to make these predictions useful where decisions about materials are actually made.”
The research was supported by the United States Air Force Office of Scientific Research.