Developing new materials for aerospace, energy, electronics, and other advanced industries is often a slow and expensive process. Before engineers can deploy a new alloy, they must manufacture samples and conduct extensive testing because existing simulation methods struggle to accurately predict how complex materials will behave in real-world conditions. Researchers at MIT have now developed a machine-learning-based approach that could significantly accelerate this process by improving the accuracy of alloy simulations.
The challenge lies in the chemical disorder found in most practical metals. While traditional machine-learning models can accurately simulate materials with highly ordered atomic structures, real-world alloys contain a wide variety of local atomic arrangements. These variations strongly influence material properties, including strength, durability, and resistance to failure. Capturing this complexity has historically required massive computational resources, often involving more than 100,000 hours of calculations for a single material system.
Led by Rodrigo Freitas and colleagues, the MIT team addressed this problem by creating more informative training datasets for machine-learning models. Using concepts from information theory, the researchers developed a method to identify and maximize the diversity of local atomic environments represented in training data. Instead of repeatedly exposing models to similar configurations, the technique replaces redundant examples with previously unseen atomic arrangements, allowing the models to learn a broader range of chemical behaviors.
The results were striking. Models trained using the new datasets predicted material properties more accurately than those trained with conventional sampling methods and even outperformed much larger machine-learning models developed using resource-intensive approaches. The researchers successfully applied the technique to a diverse range of metal alloys and demonstrated accurate predictions of phase diagrams, which are critical tools for understanding how materials behave under different temperatures and compositions.
The work could have significant industrial implications. Accurate phase diagrams help engineers make informed decisions during welding, casting, and heat treatment processes. The team is now extending the approach to study mechanical performance and radiation tolerance, with the goal of designing stronger, more resilient materials for harsh environments.
By making high-fidelity materials simulations more accessible and computationally efficient, the research offers a pathway to faster innovation and reduced reliance on costly experimental trial and error.