Machine learning could change the empirical side of catalyst development

Greetings dear friends. Many times in research, combinations are made that are difficult to generalize, and in the field of catalysis, many catalysts have been fortuitous discoveries, so the empirical aspect of research in this area of chemistry dominates the literature, basically trial and error tests. But one group of researchers is changing the way catalysts are developed by combining random sampling, experimentation and machine learning to identify synergistic combinations of catalysts.

Machine learning could change the way catalysts are developed. Image edited by Hive account@emiliomoron, original from pixabay.com.

With this method, the researchers hope to shorten evaluation times and save the resources that are employed in this type of research, which relies heavily on random combinations.

Catalysts are chemical species capable of accelerating a chemical reaction, but not any chemical element turns out to be a catalyst, and a good catalyst for one reaction is not good for another, so there is a lot of trial and error in this type of research. On the other hand, there are chemical species that by themselves do not modify any parameter of chemical reactions, but when combined with others, one component complements the other, and the synergy between the two results in a good catalyst. So in the field of combined catalyst development, the synergy between the elements is the key, so it is very important to eliminate any type of combination that is not effective.

So far, in the field of catalysis this process of discarding inefficient combinations is done by experimentation, there are no equations or chemical laws that predict whether, for example, the Pt-Ni combination will be more efficient than the W-Ni combination in the hydrogenation reaction of a hydrocarbon. So before testing a combination, it is essential to gather all the information available in the literature, which is usually also biased by data from accidentally found combinations.

But this could change thanks to a new study recently published in the journal ACS Catalysis. This study details the identification of potentially effective combinations using a protocol based on a high-throughput screening instrument and software analysis; random samples of 300 solid catalysts from a universe of more than 36,000 catalysts for oxidative coupling of methane were evaluated using this procedure. Evaluating such a large number is almost impossible for humans, so the team designed this procedure to facilitate the study of the reaction. And with the obtained data set, free of bias, was used to design the new protocol, which serves as a guide for the development of new catalysts.

A form of decision tree classification was implemented in the software, which is widely used for the machine to understand how the selected combinations influence the performance of the catalysts, which helped to obtain the necessary guidelines for the design of the new catalysts. With random sampling, 51 catalysts out of 300 provided sufficiently superior C2 reaction performance to the non-oxidative non-catalytic process of the reaction.

Catalyst design guideline scheme. Source: Image designed by Hive account@emiliomoron, contains public domain image.

Decision tree classification was successfully implemented, facilitating efficient sampling of catalysts toward improved reaction performance. It demonstrates the importance of tools that help researchers to find synergistic combinations without bias, in the study and design of new catalysts, allowing to approach these studies with a less empirical approach, thus allowing to perform such demanding studies in more realistic time frames and optimizing resources.

Well friends, I hope you found the information interesting, which shows us how advances in autonomous learning are impacting various areas of scientific research. see you next time!

---