Catalyst chemistry is a lot like cooking in a fine dining kitchen. Just as chefs spend hours adjusting recipes for palate perfection, chemists likewise spend time painstakingly tweaking catalysts (chemical reaction boosters) for optimal results.
“Figuring out an optimised catalyst is a formidable task when there’s a vast space of possibilities to explore,” said Joyjit Chattoraj, a Research Scientist from A*STAR’s Institute of High Performance Computing (IHPC). “You need to play around with a wide array of parameters, ranging from preparation methods and operating conditions, to the presence and concentration of base and promoter metals.”
The good news is that catalyst chemists now have a secret weapon in their kitchens: machine learning, or ML. Through ML, computational models can help automate the trial-and-error nature of catalyst optimisation, which Chattoraj said speeds things up tremendously.
However, existing ML models are only as good as the data they’re trained with. Just as even a good kitchen appliance with bad ingredients makes poor dishes, a good ML model with bad data makes bad predictions. For example, without incorporating a basic knowledge of physics, some of today’s purely data-driven models sometimes start bending the laws of thermodynamics by predicting catalytic conversion efficiencies over 100 percent.
To figure out a model with more realistic predictions, Chattoraj and Teck Leong Tan led a team of A*STAR scientists from IHPC, in collaboration with the Institute of Sustainability for Chemical, Energy and Environment (ISCE2) who aim to build better ML frameworks for hydrogen fuel technology, a greener energy alternative to fossil fuels. They focused on ML-driven catalyst optimisation for the water-gas shift (WGS) reaction, a key process that eliminates toxic carbon monoxide (CO) produced from hydrogen fuel cells.
They started with a cleanup of an open-source WGS dataset used in previous models to ensure all data points in it adhered to chemical reaction principles. Next, they added physical and chemical data tags (known as ‘fingerprints’) to the dataset to facilitate new catalyst discovery. They leveraged this spruced-up data to develop a new-and-improved physics-guided ML framework more likely to obey the laws of thermodynamics.
“Our state-of-the-art model can establish complex relationships between catalyst synthesis parameters through forward modelling: a training process more accurate and reliable than its conventional counterparts in predicting CO conversion,” said Chattoraj. “This paves the way for reliable inverse modelling, allowing us to search for potential catalyst candidates based on optimal compositions and reaction conditions.”
With this ML-powered catalyst cookbook under their belts, the researchers are now working on applying their model to discover next-generation WGS catalysts that stay effective even at low reaction temperatures, which could benefit many industries with stakes in fuel cell technology.
Again, getting the ingredients right is part of the challenge; the current training dataset for their model has limited experimental data on low-temperature reactions, said Chattoraj. “Overcoming this would require more data and a better understanding of how catalyst properties affect CO conversion, which will only be possible if knowledge is shared across different research domains."
The A*STAR-affiliated researchers contributing to this research are from the Institute of High Performance Computing (IHPC) and the Institute of Sustainability for Chemical, Energy and Environment (ISCE2).
