Physics-Informed Graph Neural Networks for Universal Catalyst Representation and Accelerated Discovery of ORR/OER Dual-Atom Catalysts

Abstract

Diatomic catalysts (DACs) have demonstrated great potential in the field of electrocatalysis due to their highly tunable bimetallic synergistic centers. However, the vast configurational space and high computational cost of density functional theory (DFT) calculations severely limit the discovery of candidate materials for complex, large-scale systems. This paper proposes an Active Site Physics-inspired Graph Neural Network (ASP-GNN). By deeply integrating underlying physical inductive biases (such as orbital overlap and electronegativity) and key-atom broadcasting mechanisms into variablearchitecture models like SE(3), the method successfully achieves high-precision predictions of adsorption energies with an extremely lightweight parameter set (1.7M). In the Open Catalyst 2020 (OC20) benchmark, ASP-GNN demonstrated predictive performance surpassing that of complex high-order geometric models. To address the challenges of highthroughput screening under computational resource constraints and in complex coordination environments, we synthesized a copper foil-supported carbon-based rare earth-transition metal diatomic catalyst (Cu@Gra_RE-TMNxC6-x) and innovatively introduced surface hydroxylation (OH-modification) to calculate 510 sets of high-quality datasets covering the adsorption energies of surface single/ dihydroxyl groups (OH * /2OH * ) in adsorption energy. Leveraging ASP-GNN's precise mapping of the key descriptor ΔEOH* (5-fold cross-validation MAE of only 0.2 eV), we implemented a funnel-shaped hierarchical sampling screening and successfully identified star candidates at the peak of the volcano plot. Full-reaction-path DFT validation demonstrated that the screened configurations, such as Eu-Ag-011001_OH (η ORR = 0.36 V) and Eu-Rh-001111 (η OER = 0.41 V), exhibit outstanding catalytic performance. This study not only provides a new physical regularization paradigm to overcome overfitting in small-sample datasets but also establishes a universal framework for the discovery of complex bifunctional electrocatalysts.