LOCAL: A Graph-Based Active Learning Approach for Stability Analysis of DAC@NG Catalysts (2503.19445v1)

Abstract: Dual atomic catalysts supported by nitrogen-doped graphene (DAC@NG) offer significant potential in catalytic applications by overcoming intrinsic limitations associated with single atomic catalysts. However, accurately determining their stability and atomic-scale configurations remains computationally challenging due to extensive structural variability. In this study, we present the LOCalization and Active Learning (LOCAL) framework, an innovative, scalable approach employing two graph convolutional network (GCN) models (POS2COHP and Graph2E) to predict stability energies directly from initial DAC@NG structures. Leveraging an extensive dataset of 611,648 DAC@NG structures, encompassing 38 metal elements, six distinct graphene quadra-vacancy patterns, and diverse carbon/nitrogen coordination environments, LOCAL achieved a remarkable validation mean absolute error of just 0.145 eV. Utilizing this framework, we systematically analyzed stability trends across various metal pairs, successfully generating phase diagrams for experimentally validated bimetallic systems (Co-Ni, Fe-Ni, Fe-Mn, and Ag-Ni). These results underscore LOCAL's capability for rapidly evaluating structural stability, significantly accelerating the discovery and optimization of high-performance catalysts. The developed dataset and LOCAL framework are publicly available, offering a valuable resource for future catalyst design and broader exploration of catalytic materials.