Using Machine Learning to Discover the Intricacies of Chemical Bonding

Industry Insights

Hongliang Xin, a professor of chemical engineering at the College of Engineering, and his colleagues have developed a novel artificial intelligence system that can speed up finding materials for critical technologies like fuel cells and carbon capture devices.

Their research in the journal Nature Communications, named “Infusing theory into deep learning for interpretable reactivity prediction,” presents a unique approach called TinNet—short for the theory-infused neural network—that blends machine-learning algorithms and concepts to uncover new catalysts. Catalysts are materials that initiate or accelerate chemical reactions.

TinNet is centered on deep learning, a subset of machine learning that uses algorithms to simulate how human brains function. Deep learning has lately played a crucial role in the development of technologies, including self-driving cars.

Xin and his research associates  intend to employ machine learning in the realm of catalysis to create new and improved energy technologies and products that will enhance people’s lives.

“Approximately 90% of the products you find today are simply the result of catalysis,” Xin remarked. The trick is to locate effective and sturdy catalysts for each application, which might be tough to find.

“Knowing how catalysts relate to different intermediates and controlling their bond strengths to be in the Goldilocks Zone is very critical to building effective catalytic processes,” Xin added. “And our research delivers just that.”

Machine-learning algorithms can be useful since they can detect complicated patterns in large data sets, which humans are not particularly good at, according to Xin. However, deep learning has limits, especially with regard to predicting very complicated chemical interactions, which is a critical step in discovering materials for a certain function. Deep learning can underperform in various applications for a variety of reasons, some of which are unknown.

“Most machine-learning models built for mechanical properties prediction or categorization are frequently called ‘black boxes’ and give limited physical insights,” said Hemanth Pillai, a chemical engineering grad student, and research co-author.

TinNet, a hybrid technique, integrates sophisticated catalysis theories with artificial intelligence to assist researchers in peering into this “black box” of material design to comprehend what is occurring and why, and it might help researchers break new ground in a range of domains.

“Hopefully, we can make this method more widely available to the community so that others can utilize it and truly further improve the technology for renewable energy and emission reductions technologies that are critical for society,” Xin added. “I believe this is the essential technology that might lead to certain breakthroughs.”

Leave a Reply