Importance of consistent and well-dispersed catalyst datasets for machine learning in oxidative coupling of methane

Abstract

The role of highly uniform, diverse experimental data in catalyst informatics is examined using an oxidative coupling of methane dataset measured by a single researcher under consistent devices and conditions. Broad compositional coverage and minimized experimental variability enable machine learning to capture composition–performance relationships using simple one-hot encoding. Inverse analysis of the compositional space identifies promising catalysts for experimental validation. These results demonstrate that carefully curated, well-distributed datasets, even if relatively small, enable machine learning to effectively capture composition–performance relationships.