Machine learning filters out efficient electrocatalysts in the massive ternary alloy space for fuel cells

Despite their potential promise, multicomponent materials have not been actively considered as catalyst mate-rials to date, mainly due to the massive compositional space. Here, targeting ternary electrocatalysts for fuel cells, we present a machine learning (ML)-driven catalyst screening protocol with the criteria of structural sta-bility, catalytic performance, and cost-effectiveness. This process filters out only 10 and 37 candidates out of over three thousand test materials in the alloy core@shell (X3Y@Z) for each cathode and anode of fuel cells. These candidates are potentially synthesizable, lower-cost and higher-performance than conventional Pt. A thin film of Cu3Au@Pt, one of the final candidates for oxygen reduction reactions, was experimentally fabricated, which indeed outperformed a Pt film as confirmed by the approximately 2-fold increase in kinetic current density with the 2.7-fold reduction in the Pt usage. This demonstration supports that our ML-driven design strategy would be useful for exploring general multicomponent systems and catalysis problems.

Automated summary

Researchers have developed a machine learning-driven catalyst screening protocol to identify ternary electrocatalysts for fuel cells that are potentially synthesizable, lower-cost, and higher-performance than conventional Pt materials. Experimental results showed that one of the final candidates, Cu3Au@Pt, outperformed Pt materials in oxygen reduction reactions. This approach demonstrates the usefulness of machine learning in exploring multicomponent systems and catalysis problems.