Deciphering Key Descriptors for Scaling Relationships in Graphene-Supported Pt_n Clusters via Machine Learning

Subnano clusters have emerged as a promising class of electrocatalysts, enabling efficient utilization of noble metals and superior activity for key electrochemical reactions. However, their fluxional nature and complex structure-activity relationships give rise to non-monotonic catalytic behavior, making activity evaluation highly challenging. Here, we develop a machine learning (ML) framework to assess the size-dependent activity of graphene-supported Pt_n(n  =  7 - 13) subnano clusters for the oxygen reduction reaction (ORR). Using non-ab initio descriptors, the trained ML model accurately predicts site-specific adsorption energies of key ORR intermediates across heterogeneous active sites. Feature-driven uncertainty quantification highlights the irregular catalytic activity and the importance of ensemble representation, while identifying the key geometric descriptors responsible for breaking the scaling relationship in the subnanometer regime. Furthermore, we uncover critical insights into the dynamics of active electrocatalysts under ambient ORR conditions on supports, showing their strong oxidative tendencies that drive monolayer formation of intermediates. We also derive reaction networks for the rate-determining steps, pinpointing the saturation onset point that signals changes in binding sites. Unlike bare clusters, this study provides an atomistic understanding of fluxional dynamics and structural evolution on supports, offering guidelines for the rational design of highly efficient electrocatalysts for fuel cell applications.