Topology-based machine learning for predicting curvature effects in metal-nitrogen-carbon single-atom catalysts

Abstract

Metal-nitrogen-carbon (M-N-C) single-atom catalysts are widely utilized in various energy-related catalytic processes, offering a highly efficient and cost-effective catalytic system with significant potential. Recently, curvature-induced strain has been extensively demonstrated as a powerful tool for modulating the catalytic performance of M-N-C catalysts. However, identifying optimal strain patterns using density functional theory (DFT) is computationally intractable due to the high-dimensional search space. Here, we developed a graph neural network (GNN) integrated with an advanced topological data analysis tool—persistent homology—to predict the adsorption energy response of adsorbate under proposed curvature patterns, using nitric oxide electroreduction (NORR) as an example. Our machine learning model achieves high accuracy in predicting the adsorption energy response to curvature, with a mean absolute error (MAE) of 0.126 eV. Furthermore, we elucidate general trends in curvature-modulated adsorption energies of intermediates across various metals and coordination environments. We recommend several promising catalysts for NORR that exhibit significant potential for performance optimization via curvature modulation. This methodology can be readily extended to describe other non-bonded interactions, such as lattice strain and surface stress, providing a versatile approach for advanced catalyst design.