Explainable Artificial Intelligence Elucidates Synthesis-Structure-Property-Function Relationships in Nanostructured Catalysts

Abstract

Designing high-performance catalysts based on supported metal atoms and nanoparticles necessitates rigorous assembly control, which affects reactivity. Yet, attaining synthesis precision to discern catalyst speciation and properties remains challenging. This limitation is tackled by pioneering an eXplainable Artificial Intelligence (XAI) methodology that combines a decision tree classifier and a random forest regressor sequentially to elucidate synthesis-structure-property-function relationships in nanostructured catalysts, exemplified for the oxygen evolution (OER) and hydrogen evolution (HER) reactions. The decision tree accurately predicts the formation of single atoms versus nanoparticles for 37 metals anchored on nitrogen-doped carbon, identifying the metal's standard reduction potential and cohesive energy as determinants of speciation. The random forest regressor correlates the electrocatalytic performance of synthesized single-atom catalysts (SACs) to intrinsic properties, revealing a volcano-like relationship of current density with the electronegativity of the active site and metal-support interaction, providing insights beyond traditional adsorption energy descriptors. The integrated models are validated experimentally with an overall accuracy of over 80%, establishing user confidence. This XAI framework, adaptable to diverse synthetic protocols and material classes is a powerful tool to understand synthesis-structure-property-function relationships and promote data-informed experiments with reduced characterization efforts.