Companies working at the frontier of aerospace, energy, and computing are constantly searching for new materials to improve performance. However, understanding how these materials will behave inside rockets or on computer chips typically requires physically making and testing them first. This is because even the most powerful simulation techniques struggle to model the complex chemical arrangements found in most solid materials, adding significant costs and time to materials innovation.
Now, a team of MIT researchers has developed a way to accurately model the behavior of metals, regardless of the complexity of their chemical arrangement. The core of the approach involves machine-learning models that accelerate and improve material simulations. The researchers enhanced these models by building training datasets that capture the diversity of atomic environments in chemically disordered materials.
In a new paper published in Sciences Advances, the team demonstrated that their approach can accurately predict material properties for a diverse group of metal alloys under various conditions. They also showed how the method could be used to develop new materials, especially in scenarios where experimentation is expensive.
“The focus of the paper is metallic alloys, but this could be adapted to other types of materials, like semiconductors,” says senior author Rodrigo Freitas, MIT’s TDK Career Development Professor in Materials Science and Engineering. “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.”
Modeling Metals
Material properties are largely determined by the internal arrangement of their chemical elements. Even two materials with the same chemical composition can differ drastically — one might be brittle while the other deforms without breaking — depending on the atomic arrangement.
Capturing that distinction requires simulating materials atom by atom. Researchers rely on models that describe how atoms interact. Over the last two decades, machine learning has become the most accurate way to build these models. Such models work well when chemical arrangements follow highly ordered patterns, but most solid materials are chemically disordered, with varying arrangements from one region to another.
“The real challenge in our field is modelling these chemically disordered phases,” Freitas says. “Chemical disorder means there’s a huge variety of local chemical environments, which is hard for the machine-learning model to learn. This is a problem because every single metal we use in practice is chemically disordered.”
The issue stems from a lack of representative training data for atom-by-atom simulations. The current leading approach uses brute force, often requiring over 100,000 hours of computation to create training data for a single material, and it does not transfer well when the material’s composition changes.
Information Theory Approach
In previous work, Freitas’ group developed a way to measure the chemical complexity of solid materials by analyzing the frequency and spacing of tiny groups of atoms. For this study, they used that capability to build better training datasets. They applied a mathematical approach known as information theory to generate datasets that capture a wider variety of local chemical environments inside disordered materials. The method works by swapping out atoms from samples to reduce repetition and expose the model to chemical environments it might otherwise miss.
“We kept optimizing the training set so it captured as many different local environments as possible,” Freitas says. “If the same kind of environment showed up many times, we replaced redundant examples with ones the model hadn’t seen before. That 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 another popular sampling method.
“The starting point for all these atom-by-atom simulations is: Are you able to accurately describe the chemical bond between atoms?” Freitas explains. “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 applied their technique to create machine-learning training datasets for a group of chemically diverse metal alloys. Using a set of machine-learning models, they showed that models trained on their datasets are more accurate than much larger models created by companies like Google and Microsoft.
From Lab to Industry
The method works, in part, by capturing hidden patterns in the sample data — described as “subtle energetic biases toward certain local chemical configurations.” These small energetic differences determine which phases form in an alloy, how those phases change with temperature and composition, and ultimately what properties the material will have.
As one test, simulations led by MIT PhD student Daniel Xiao showed that the team’s models could predict phase diagrams that closely matched experimental data. Phase diagrams map which phases are stable across different temperatures and chemical compositions, and they are a central tool for designing and processing alloys.
“Phase diagrams are one of the main ways people connect materials modeling to real processing decisions,” Freitas says. “If you are welding, casting, or heat-treating an alloy, you need to know which phases are likely to form under different conditions. Our goal is to make these kinds of predictions accurate enough, and accessible enough, that they become part of how people design materials.”
The researchers are now using the approach to study how changing an alloy’s composition affects mechanical properties and radiation tolerance, aiming to design materials that remain strong and damage-tolerant in harsh environments. They are also working to make the method easier to integrate with the tools and workflows materials engineers already rely on.
“Industry isn’t going to change the way they do things if what you’re creating doesn’t fit into their existing operating procedures,” Freitas says. “The goal is to make these predictions useful in the places where materials decisions are actually made.”
The research was supported by the U.S. Air Force Office of Scientific Research.
