Mechanism-informed graph-assisted screening of PtM local motifs for CO-tolerant methanol oxidation.

Efficient methanol oxidation reaction (MOR) catalysts require weakened CO binding at Pt and favorable formation of oxygenated species at adjacent metal sites. The many top-layer and subsurface arrangements of PtM alloys, however, create a local-motif space that is costly to enumerate with density functional theory (DFT). We develop a mechanism-informed, graph-assisted active-learning framework for adsorption-descriptor-based prioritization of CO-tolerance-relevant PtM motifs. A curated DFT dataset contains CO@Pt, OH@Pt, and OH@M adsorption energies. Interpretable element-motif descriptors reveal target-dependent behavior: Pt-site adsorption depends on coupled local-environment effects, whereas OH@M is strongly metal-specific. Under identical structure-grouped validation, descriptor-guided MEGNet-style, SchNet, and DimeNet++ encoders were compared on bare-surface graphs. The MEGNet-style encoder gave the most faithful standalone motif representation, with an auxiliary-descriptor reconstruction mean absolute error of 0.048 and a motif-family accuracy of 0.998, while its descriptor-latent Gaussian-process model matched DimeNet++ for Pt-site adsorption at substantially lower complexity. Retrospective uncertainty-aware and multi-objective acquisition accelerated the recovery of high-ranked motifs. Mo-containing families dominated the three-descriptor-qualified set; representative CO@M calculations showed that Ru- and Fe-containing motifs favor CO adsorption at M, whereas validated Mo-containing motifs retained a strong OH preference for M with nearly degenerate Pt/M CO adsorption. The workflow, therefore, provides interpretable, small-data-compatible motif prioritization before full MOR pathway, kinetic, and electrochemical validation.