A group of MIT researchers has developed a machine learning (ML) approach that could revolutionize the way we design catalysts for chemical reactions. The method simplifies the intricate process of designing new compounds or alloys, traditionally dependent on the intuition of experienced chemists, by using ML to provide more detailed information than conventional techniques can.
The research, presented this week in Nature Computational Science, explains how the team applied their system to a material studied for 30 years using conventional methods. They discovered two new atomic configurations on the compound’s surface that were previously unidentified, and identified another arrangement that appears to be unstable.
The surfaces of materials interact with their environment, and these interactions can vary depending on the specific configuration of atoms at the surface. Traditional methods for characterizing these surfaces are static, focusing on specific configurations out of millions of possibilities. But the new approach can estimate all the variations based on a few calculations chosen by an iterative ML process. This system seeks out materials with specific desired properties.
Unlike current methods, the new system can also provide dynamic information about how surface properties change over time under operating conditions, like during a chemical reaction or when a battery electrode is charging or discharging.
The team’s Automatic Surface Reconstruction framework starts with a single clean-cut surface example, then combines active learning with a Monte-Carlo algorithm to select sites to sample on that surface. By analyzing the results of each example, the system guides the selection of the next sites. The system can obtain accurate predictions of surface energies across different chemical or electrical potentials, using fewer than 5,000 first-principle calculations.
An additional benefit of the method is the lower cost, allowing for fewer expensive quantum mechanical energy evaluations. The method is applicable to more complex or “harder” materials, including three-component materials. The researchers have made the computer algorithms, known as AutoSurfRecon, freely available online, which can assist in developing new materials for catalysts, battery or fuel cell components.
Potential applications also include studying the dynamics of chemical reactions used to remove carbon dioxide from the air or power plant emissions. These reactions often work by using a material that absorbs oxygen, which strips oxygen atoms from carbon dioxide molecules, leaving behind carbon monoxide, a useful fuel or a chemical feedstock.
Using their tool, the researchers scrutinized the surface atomic arrangement of perovskite material strontium titanium oxide, which despite being analyzed using conventional methods for over thirty years was still not fully understood. They discovered two new atomic arrangements on its surface that were not previously reported, further highlighting the benefits of their intuition-free approach.
With their code now available to the wider research community, the team hopes for swift enhancements by other users. The research was supported by the U.S. Air Force, the U.S. Department of Defense, and the U.S. National Science Foundation.
