Accelerated discovery of CO2-to-C3-hydrocarbon electrocatalysts with human-in-the-loop

Abstract

Automated high-throughput experimentation combined with artificial intelligence holds the potential to accelerate materials discovery; however, utilizing this approach in heterogeneous electrocatalytic materials has been challenging. Here, we pursue the discovery of multi-element CO2 electrocatalysts by employing a machine learning algorithm that integrates human domain knowledge to enable on-the-fly editing of feature contributions. By combining this approach with an accelerated experimental platform, we navigate a 15-element space for CO2-to-C3 hydrocarbon electrosynthesis and achieve a'165x acceleration compared with a conventional screening approach-of which '33x comes from the new experimentation platform and a further '5x from incorporating human domain knowledge. We identify Cu0.98In0.02 as an effective catalyst for propylene electrosynthesis, achieving a production rate of 42 mmol gcat-1 h-1 in a 25 cm2 electrolyzer. Data mining on the 300-composition dataset reveals two distinct C-C coupling pathways toward C3 hydrocarbons-*CO dimerization and *CHx-mediated coupling-with composition-dependent factors governing each pathway.