MIT researchers have developed a machine-learning framework that can accurately predict the behavior of metal alloys regardless of the complexity of their chemical arrangement, solving a long-standing bottleneck in materials science that has forced companies to rely on physical testing rather than simulation. The approach, detailed in a new paper published in Science Advances, uses information theory to build training datasets that capture the full diversity of atomic environments inside chemically disordered materials, enabling simulations that match real-world experimental data with unprecedented fidelity.
The challenge facing materials engineers has been deceptively simple: even the most powerful simulation techniques struggle to model the complex, disordered chemical arrangements found in virtually all practical metals. Every alloy used in aerospace engines, semiconductor manufacturing, and energy infrastructure is chemically disordered at the atomic level, meaning its local atomic environments vary from one region to another. Traditional machine-learning models for atom-by-atom simulation require enormous computational resources — often exceeding 100,000 hours of computation to create training data for a single material — and still fail to transfer reliably when the material’s composition changes.
Rodrigo Freitas, MIT’s TDK Career Development Professor in Materials Science and Engineering and senior author of the study, noted that the work addresses a fundamental limitation in the field. “Chemical disorder means there’s a huge variety of local chemical environments, which is hard for the machine-learning model to learn,” Freitas stated. “This is a problem because every single metal we use in practice is chemically disordered.”
The Core Innovation: Information Theory Meets Materials Science
The MIT team’s breakthrough centers on how training data is constructed. Rather than using brute-force sampling that produces redundant examples of common atomic configurations while missing rare but important ones, the researchers applied a mathematical framework from information theory to systematically construct datasets that capture maximum chemical diversity.
- Redundancy elimination: The method identifies when the same type of atomic environment appears repeatedly in the training data and replaces those redundant examples with configurations the model has not yet encountered.
- Active optimization: The training set is iteratively refined to ensure each example contributes new information about local chemical environments that would otherwise be overlooked.
- Transferability: Models trained on these optimized datasets generalize across different alloy compositions without requiring the massive recomputation typically needed.
Killian Sheriff PhD ’26, the paper’s first author, worked with MIT PhD students Daniel Xiao and Yifan Cao, along with University of Sheffield Senior Lecturer Lewis R. Owen, to test the approach across multiple alloy systems and material properties. The team drew on Owen’s experimental data to validate their simulations against real measurements of atomic ordering in alloys.
How Does the MIT Machine-Learning Model Predict Metal Alloy Behavior?
The model works by enabling atom-by-atom simulations that accurately describe chemical bonds between atoms in disordered materials. Traditional models become unreliable when the chemical arrangement is not highly ordered, because they lack training data representative of the full range of local atomic environments. The MIT approach uses information theory to generate training datasets that systematically capture that diversity, then feeds those datasets into machine-learning models that predict material properties such as phase stability, mechanical behavior, and response to temperature changes. The result is a simulation that reflects real-world chemistry rather than an idealized approximation.
Phase Diagrams and Real-World Validation
The team demonstrated that their approach could produce phase diagrams — maps showing which material phases are stable across different temperatures and chemical compositions — that closely matched experimental measurements. Phase diagrams are essential tools for industrial processes such as welding, casting, and heat-treating alloys, where knowing which phases will form under specific conditions directly dictates processing decisions.
Daniel Xiao led the phase diagram validation efforts, showing that the machine-learning models trained on the MIT datasets could reproduce experimental results with accuracy comparable to or better than much larger models developed by major technology companies. “We got to a point where we were convinced it worked without using these expensive brute-force methods,” Freitas explained. “I told Killian, ‘This is a good paper. But if you can show that simulations with these models can now accurately predict useful materials properties, then it becomes a very good paper.’ Killian took that to heart and tested this as widely as he could.”
Broader Applicability Beyond Metallic Alloys
While the current research focuses on metallic alloys — the field in which Freitas works — the methodology is not limited to metals. The same information-theoretic approach to training dataset construction could be adapted for other classes of materials, including semiconductors. This opens the door to applications in electronics design, where accurate simulation of chemically disordered semiconductor materials could reduce the need for costly physical prototyping in chip manufacturing and development.
Industry Implications and Next Steps
The researchers are now applying their approach to study how changing an alloy’s composition affects mechanical properties and radiation tolerance, specifically targeting the design of materials that remain strong and damage-tolerant in harsh environments such as nuclear reactors and space vehicles. A parallel effort focuses on integrating the method into the existing tools and workflows that materials engineers already rely on in industrial settings.
“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 said. “The goal is to make these predictions useful in the places where materials decisions are actually made.”
What This Means for Engineers and Researchers
For materials scientists and engineers working in aerospace, energy, computing, and manufacturing, the most immediate implication is the potential to reduce the expensive and time-consuming cycle of physical synthesis and testing that currently dominates materials development. Rather than having to make a material and then test it to understand its properties, engineers could use these machine-learning models to screen candidate compositions computationally, reserving physical experimentation for the most promising candidates. The method is particularly valuable in scenarios where experimentation is prohibitively expensive, such as with rare or hazardous materials. Researchers interested in adopting the approach can start by evaluating their current training datasets for redundancy and applying information-theoretic principles to maximize chemical diversity — the core technique that the MIT team has now validated across multiple alloy systems. As the team works to integrate the method into standard materials engineering tools, the path from academic research to industrial application appears increasingly direct.
