A fast and accurate surrogate model screens over 10,000 possible metal-oxide supports for a platinum nanocatalyst to prevent sintering under high temperatures.
(Nanowerk News) Metal nanoparticles catalyze reactions to synthesize chemicals and fuels, but they tend to sinter—or clump together—which reduces their surface area and effectiveness under high temperatures. A new computational framework efficiently sifts through tens of thousands of possibilities to find optimal support materials that prevent nanoparticles from sticking together, according to a University of Michigan study published in Nature Catalysis (“Predictive model for the discovery of sinter-resistant supports for metallic nanoparticle catalysts by interpretable machine learning”).
“As particles grow in size, they hide these expensive atoms like platinum in the bulk instead of on the surface where they are able to do the reaction. Getting nanoparticles to stick to a support material spaces them out from one another to prevent this,” said Suljo Linic, the Martin Lewis Perl Collegiate Professor of Chemical Engineering at U-M and co-corresponding author of the study.
With so many factors influencing sintering, it would be impossible to identify the ideal metal nanoparticle-support combinations through experimentation. To sidestep experimentation, past researchers have leveraged a machine learning approach called neural network molecular dynamics to simulate how atoms move during sintering.
Taking this a step further, the U-M research team integrated interpretable machine learning into the process, which identifies the physical properties that prevent sintering. The interpretable machine learning step allowed the research team to quickly screen over 10,000 candidate materials, a scale that would be impossible with full simulations.
“Neural network molecular dynamics is like rendering a high-definition video while interpretable machine learning allows us to identify critical features of the high-definition video and use those features. The simpler model allows us to handle a large number of candidates with fewer computational resources,” said Chenggong Jiang, a doctoral student in chemical engineering at U-M and lead author of the study.
Researchers are searching for optimal support materials that prevent metal nanoparticles from sticking together while catalyzing reactions to make fuels or chemicals. Like rendering a high-definition video, neural network molecular dynamics simulate the movement of metal nanoparticles during high-heat reactions. In this simulation, five platinum nanoparticles stick to a cerium oxide surface during a high-heat reaction, helping them stay apart from one another. Each sphere represents an atom. (Video: Goldsmith & Linic Labs, University of Michigan Chemical Engineering)
The steps the researchers took to develop their predictive model are much like quick ways to identify common sicknesses like strep throat. In the case of a strep throat outbreak on a college campus, a highly accurate way to diagnose the community would be to swab the throats of everyone on campus, run tests and determine who is positive for the bacteria.
As running that many throat cultures would be expensive and time consuming, a less precise tactic could be to examine everyone’s throat and diagnose strep based on who has red or swollen tonsils. An even faster and less invasive route could be to simply use a forehead thermometer to find who has a temperature. Although slightly less accurate, this method could quickly weed out people without an infection, saving a more time-intensive throat examination as a follow up.
Similarly, the researchers moved from the most to least precise method. A few quantum chemical calculations—highly accurate but expensive like throat cultures—modeled how atoms move during sintering down to the location of electrons. Those calculations trained neural network molecular dynamics simulations, which are less precise and more efficient like the throat examinations.
“This neural-network-driven molecular dynamics and interpretable machine learning approach enables nanocatalyst simulations that are many thousands of times faster than traditional quantum-mechanical methods, while keeping high accuracy—turning months of computing into days or even hours,” said Bryan Goldsmith, an associate professor of chemical engineering at U-M and co-corresponding author of the study.
The research team simulated 203 nanoparticle-support pairs to determine the level of sintering and metal-support interactions under high temperatures. The results trained an interpretable machine learning model to predict simulation results based on 12 specific physical properties of the support material.
“We call it interpretable machine learning, because the inputs that we provide to understand these trends are based on physical insights as to what should govern these interactions,” said Linic.
Of the physical properties tested, surface energy—how tightly nanoparticles stick to the support—was the most influential feature to prevent sintering. This, along with several other important electronic and structural features, was validated through case studies using quantum chemical calculations.
After verifying the interpretable machine learning model through those case studies, the team used it to screen 10,662 metal-oxide supports for platinum nanoparticles. Similar to the forehead thermometer, this simple model thinned the group to 148 candidates. With a much more manageable candidate group, further simulations and experiments resulted in a sinter-resistant barium oxide (BaO) support.
“This framework is broadly applicable to numerous catalytic applications. If somebody at a chemical plant has some oxide material that they want to use as a support, they can start with the simplest interpretable model and move up from there,” said Linic.
