Newswise — Upton, NY — Scientists at the U.S. Department of Energy’s (DOE) Brookhaven National Laboratory have developed a new machine learning framework that can accelerate the search for better catalysts (materials that accelerate chemical reactions) and provide more reliable results.
Finding high-performance catalysts used to accelerate processes from chemical manufacturing to energy production can be a time-consuming and expensive process, often relying on years of trial and error and vast computational resources. To make matters more difficult, ideal catalyst candidates are rare.
“Imagine driving to a new location without using GPS,” says Wenji Liao, a chemist at Brookhaven Research Institute. “Eventually you’ll get there, but you’ll take a big detour and waste time. Maybe that’s what discovering catalysts is all about.”
Researchers in Brookhaven Laboratory’s Department of Chemistry’s Catalyst Reactivity and Structure Group approached these discovery challenges using a novel multilayer machine learning approach that screens catalysts step-by-step, mimicking the way scientists assess performance in real experiments.
The team was able to exploit the chemical transformation of carbon dioxide (CO).2) Using methanol (a type of alcohol that can be used as a fuel) as a case study for the new approach, it outperformed traditional models. The study also shed new light on how scientists can control key chemical reaction steps to tune activity and selectivity, two key characteristics that make an effective catalyst in a chemical reaction process.
A paper describing their research was recently published in a journal. chemical catalyst.
The best catalysts must be active enough to drive the reaction efficiently, but selective enough to favor the desired products over undesired byproducts.
“Highly active and highly selective catalysts save energy and cost,” said Ping Liu, a chemist at Brookhaven Laboratory and adjunct professor at Stony Brook University. “Active catalysts mean they don’t require high pressures or temperatures to accelerate the reaction, and selective catalysts mean they don’t require potentially costly purification to get the desired product.”
Machine learning models promise faster catalyst discovery, but they face hurdles that Brookhaven scientists are trying to overcome in their research. According to the researchers, existing single-layer models were limited by the high cost of generating the large databases needed for analysis, poor data quality, and uneven data spread. Additionally, traditional models were not trained on chemical understanding to make accurate predictions about catalysts.
“Simple one-layer models lack the expertise needed to reliably predict good catalysts,” Liu says. “Based on all these limitations, we developed a multilayer binary machine learning approach that targets the complex reaction networks of real catalysis, which have never been considered before in models of this kind.”
Case study: CO rotation2 to methanol
Rather than having a single model predict catalyst performance all at once, the Brookhaven team’s approach breaks down the problem into a series of simple decisions. To test their approach, the researchers studied the performance of copper-based catalysts used to convert CO.2 to methanol.
According to the study, the researchers trained multiple models using synthetic datasets generated from dynamic Monte Carlo simulations, resulting in lower computational costs. These simulations capture how chemical reactions unfold over time, including the competition between multiple reaction pathways. This is an important feature that is often missing in simpler models.
“This helps improve the accuracy and reliability of the model,” said Anh Nguyen, a visiting graduate student at Stony Brook University. “Each layer has to do with how we think about catalysts as chemists and how we classify catalysts into different categories based on our understanding of chemistry or catalysis.”
In their case study, the researchers’ multilayered approach asked whether catalysts could accelerate reactions that convert CO.2 We will produce the desired product, methanol, and investigate whether it performs as well as or better than copper-based catalysts widely used in industrial and academic applications.
By applying the new framework, the team was able to screen catalyst designs that were both more active and more selective than copper catalysts. This method consistently outperformed traditional single-layer machine learning models, which struggled to find rare and high-performing candidates.
This framework also revealed which reaction steps are most important. This analysis showed that transitions between competing reaction pathways, rather than individual steps alone, play an important role in controlling both activity and selectivity.
“The multi-layered approach allows us to dig deeper into the understanding between what we have identified as important characteristics and reactive behaviors,” said Liu. “We have identified key steps that control both the activity and selectivity of CO.2 into methanol, bringing new insights into this process. ”
The process of converting CO2 Conversion to methanol, known as hydrogenation, is already a commercial process. Researchers said the initiative could be a step toward improving workflows for industry partners. This framework can also be applied to other processes.
To develop the new framework, the researchers used computational resources at the Center for Functional Nanomaterials, a Department of Energy Office of Science user facility in Brookhaven. Brookhaven Scientific Computing and Data Facility. and SeaWulf, Stony Brook University’s high-performance computing cluster.
This research was supported by the DOE Office of Science.
Brookhaven National Laboratory is supported by the U.S. Department of Energy, Office of Science. The Office of Science is the largest supporter of basic research in the physical sciences in the United States and works to address some of the most pressing challenges of our time. For more information, please visit: science energy government.
Follow @BrookhavenLab on social media. find us at Instagram, linkedin, ×and facebook.
