Designing new compounds or alloys whose surfaces can be used as catalysts in chemical reactions can be a complex process that relies heavily on the intuition of experienced chemists. A team of researchers at MIT has devised a new approach using machine learning that removes the need for intuition and provides more detailed information than conventional methods can practically achieve.
For example, by applying the new system to a material that has already been studied for 30 years by conventional means, the team discovered that the surface of the compound could form two new atomic configurations that had not previously been recognized, and that another configuration had been seen in previous work is likely unstable.
The findings are described this week in the journal Nature Computational Sciencein a paper by MIT graduate student Xiaochen Du, professors Rafael Gómez-Bombarelli and Bilge Yildiz, MIT Lincoln Laboratory technical staff member Lin Li, and three others.
The surfaces of materials often interact with their environment in ways that depend on the precise configuration of atoms on the surface, which can vary depending on which parts of the material’s atomic structure are exposed. Consider a layer cake with raisins and nuts: Depending on exactly how you cut the cake, different amounts and compositions of the layers and fruit will be exposed at the edge of your slice. The environment is also important. The surface of the cake will look different if it is soaked in syrup, making it wet and sticky, or if you put it in the oven, crisping and darkening the surface. This is similar to how the surfaces of materials react when immersed in liquid or exposed to different temperatures.
The methods commonly used to characterize material surfaces are static, looking at a specific configuration out of millions of possibilities. The new method allows an estimation of all variations, based on only a few first-principles calculations that are automatically selected by an iterative machine learning process, in order to find those materials with the desired properties.
Furthermore, unlike today’s standard methods, the new system can be extended to provide dynamic information on how surface properties change over time under operating conditions, for example while a catalyst is actively promoting a chemical reaction or while charging a battery electrode or discharging.
The researchers’ method, which they call an automatic surface reconstruction framework, avoids the need to use selected examples of surfaces to train the neural network used in the simulation. Instead, it starts with a single instance of a pristine cutting surface, then uses active learning combined with a type of Monte-Carlo algorithm to select locations to sample on that surface, evaluating the results of each instance location to guide the selection of the next locations. Using fewer than 5,000 first-principles calculations, out of the millions of possible chemical compositions and configurations, the system can obtain accurate predictions of surface energies at various chemical or electrical potentials, the team reports.
“We’re looking at thermodynamics,” says Du, “which means that, under different kinds of external conditions like pressure, temperature, and chemical potential, which might be related to the concentration of a particular element, [we can investigate] what is the most stable structure for the surface?’
In principle, determining the thermodynamic properties of a material’s surface requires knowing the surface energies in a particular atomic arrangement and then determining these energies millions of times to include all possible variations and capture the dynamics of the processes they take country. While it’s theoretically possible to do this computationally, “it’s just not affordable” on a typical lab scale, Gómez-Bombarelli says. The researchers were able to get good results by looking at just a few specific cases, but these are not enough to provide a true statistical picture of the dynamic properties involved, he says.
Using their method, Du says, “we have new features that allow us to sample the thermodynamics of different compositions and configurations. We also show that we are able to achieve them at a lower cost, with fewer expensive quantum mechanical energy evaluations. And we can also do this for harder materials,” including three-component materials.
“What’s traditionally been done in the field,” he says, “is researchers, based on their intuition and knowledge, will only try a few guess surfaces. But we do comprehensive sampling and it’s done automatically.” He says that “we have transformed a process that was once impossible or extremely challenging due to the need for human intuition. Now, we need minimal human input. We just provide the pristine surface and our tool handles the rest.”
This tool, or set of computer algorithms, called AutoSurfRecon, has been made available free of charge by the researchers so that it can be downloaded and used by any researcher in the world to help, for example, develop new materials for catalysts such as to produce of “green” hydrogen as an emission-free alternative fuel or for new batteries or fuel cell components.
For example, says Gómez-Bombarelli, in developing catalysts for hydrogen production, “part of the problem is that it’s not really understood how their surface area is different from their volume as the catalytic cycle happens. So there’s this disconnect between what the hardware looks like when it’s in use and what it looks like when it’s being prepared before it goes into service.”
He adds that “at the end of the day, in catalysis, the entity that’s responsible for the catalyst doing something is a couple of atoms exposed on the surface, so it really matters a lot what the surface looks like right now.”
Another potential application is to study the dynamics of chemical reactions used to remove carbon dioxide from the air or from power plant emissions. These reactions often work using a material that acts as a kind of sponge to absorb oxygen, thus removing oxygen atoms from carbon dioxide molecules, leaving behind carbon monoxide, which can be a useful fuel or chemical feedstock. . Developing such materials “requires understanding what the surface does with oxygen and how it’s structured,” Gómez-Bombarelli says.
Using their tool, the researchers studied the surface atomic arrangement of the perovskite material titanium strontium oxide, or SrTiO3, which had already been analyzed by others using conventional methods for more than three decades, but was still not fully understood. They discovered two new arrangements of atoms on its surface that had not been reported before, and they predict that one arrangement that had been reported is actually unlikely to occur at all.
“This highlights that the method works without intuitions,” says Gómez-Bombarelli. “And that’s a good thing because sometimes intuition is wrong and what people thought was the case turns out not to be.” This new tool, he said, will allow researchers to be more exploratory, testing a wider range of possibilities.
Now that their code has been released to the wider community, he says, “we hope it will inspire very quick improvements” by other users.
The team included James Damewood, PhD at MIT, Jaclyn Lunger PhD ’23, now at Flagship Pioneering, and Reisel Millan, a former postdoc now at the Institute of Chemical Technology in Spain. The project was supported by the US Air Force, the US Department of Defense and the US National Science Foundation.