(659a) Knowledge Extraction Via Machine Learning from High-Throughput Catalytic Experiments

High-throughput experimentation on catalytic reactions produce a wealth of information. This quickly becomes problematic for researchers, who are tasked with examining this wealth of information and extracting fundamental chemical information to use for catalyst optimization. Recent advances in artificial intelligence, specifically in the field of machine learning, have produced automated methods of data analysis to quickly extract important relationships from complex datasets, allowing the researcher to focus on applying expert knowledge rather than crunching numbers.

This machine learning methodology has been applied to analyze a high-throughput screening of ammonia decomposition catalysts. Ammonia decomposition has been extensively studied in a fairly narrow design space predominately focused on Ru catalysts supported by single alkali metals¹. In an effort to expand the design space, high-throughput experimentation was employed to study a wide variety of catalysts with differing elements (Ru-M-K, where M is a supporting metal) and weight loadings. They were tested for activity towards ammonia decomposition under varying reaction conditions, such as inlet ammonia concentration and reactor temperature. Several hundred potential catalyst features were chosen and a machine learning algorithm was utilized to extract catalyst features that caused the greatest change in catalytic activity. This knowledge was then used to guide further catalyst synthesis.

[1] S. Mukherjee, S. V. Devaguptapu, A. Sviripa, C. R. F. Lund, and G. Wu, “Low-temperature ammonia decomposition catalysts for hydrogen generation,” Appl. Catal. B Environ., vol. 226, no. December 2017, pp. 162–181, 2018.