Predicting Activation Energy of Hydrocarbon Dehydrogenation on Au(111) via Machine Learning

Activation of C-H bonds in hydrocarbons is fundamental for synthesizing organic functional materials, yet traditional density functional theory (DFT) methods for determining reaction energy barriers are computationally intensive and often suffer from convergence challenges. We report here the construction of a comprehensive DFT-based data set of hydrocarbon dehydrogenation reactions on the Au(111) surface and propose a feature-enhanced graph neural network (F-GNN) that integrates eight chemically informed descriptors with molecular graph representations. This F-GNN model accurately predicts reaction activation energies, outperforming conventional approaches such as the Brønsted-Evans-Polanyi relationship and standalone machine learning models. Our findings demonstrate that combining chemical prior knowledge with data-driven features enables efficient and precise energy barrier prediction, offering a promising strategy to accelerate reaction path screening and mechanistic understanding in surface-catalyzed hydrocarbon transformations.