MIT researchers have developed a machine learning-based method for designing new compounds or alloys for use as catalysts in chemical reactions. Traditional methods of designing such materials rely on static observations of a single configuration, out of millions of possibilities, and the intuition of experienced chemists. However, the new method employs machine learning algorithms to dynamically estimate all possible configurations, with a major advantage being that it does not require human input for training the algorithm.
Literally termed the Automatic Surface Reconstruction Framework, this approach begins with a single pristine cut surface, uses active learning and a particular type of Monte-Carlo algorithm to select sites to sample, and evaluates the site results to guide the next set of samples. This helps in achieving accurate predictions of the surface energies across varying chemical or electrical potentials with fewer tests, saving considerable resources.
The researchers have demonstrated the method’s potential by applying it to the material, strontium titanium oxide, widely studied for thirty years. They discovered two new atomic configurations on the material’s surface, previously unidentified, and suggested that one previously observed configuration was likely unstable.
The method also dynamically tracks how the surface properties of a material change over time in operating conditions, such as during a chemical reaction or when a battery electrode is charging or discharging. Its application can extend to developing catalysts for cleaner hydrogen production, components for new batteries or fuel cells, and studying the dynamics of chemical reactions used to clean the air or power plant emissions.
This more efficient and detailed approach to surface characterization and prediction could revolutionize the fields of materials science and chemistry by enabling more accurate, comprehensive sampling and less reliance on human intuition. Furthermore, the researchers have made their tool, AutoSurfRecon, freely available for use by other researchers across disciplines.
The team’s findings were published in the journal Nature Computational Science. The research team comprised Xiaochen Du, a graduate student at MIT, professors Rafael Gómez-Bombarelli and Bilge Yildiz, staff member Lin Li from MIT Lincoln Laboratory, and three others. The study was supported by the U.S. Air Force, U.S. Department of Defense, and U.S. National Science Foundation.
