MIT Develops AI Method to Accurately Model Complex Metal Alloys
A team at MIT has created a method to accurately model the behavior of metals with complex chemical arrangements, using machine learning models trained on datasets that capture diverse atomic environments. The work, published in Science Advances, was led by Rodrigo Freitas and includes co-authors Killian Sheriff, Daniel Xiao, Yifan Cao, and Lewis R. Owen.
The Challenge
Traditional simulation struggles with disordered materials, and current machine learning approaches require extensive computational resources (over 100,000 hours per material) and do not transfer well to new compositions. The MIT team's approach reduces computation time by optimizing training data.
Key Innovations
The approach addresses the challenge of modeling chemically disordered materials, which are common in practical metals. The researchers used information theory to build training datasets that include a wide variety of local chemical environments, replacing redundant examples with new ones to improve model learning.
Models trained on these datasets predicted material properties more accurately than those using random sampling or other methods, and outperformed larger models from companies like Google and Microsoft in specific tests.
The method was validated against experimental data for multiple metal alloys, including phase diagram predictions.
Practical Applications
The technique is designed for atom-by-atom simulations and can predict properties such as phase stability, mechanical properties, and radiation tolerance. The researchers aim to make the method compatible with existing industrial workflows for materials design, including for sustainable steels and aerospace materials.
Support
The research was supported by the U.S. Air Force Office of Scientific Research.