Herein we proposed a data-driven high-throughput principle to screen high-performance single-atom materials for hydrogen evolution reaction (HER) and hydrogen sensing by combing the theoretical computations and a topology-based multi-scale convolution kernel machine learning algorithm. After the rational training by 25 groups of data and prediction of all 168 groups of single-atom materials for HER and sensing, respectively, a high prediction accuracy (> 0.931 R² score) was achieved by our model. Results show that the promising HER catalysts include Pt atoms in C₄ and Sc atoms in C₁N₃ coordination environment. Moreover, Y atoms in C₄ coordination environment and Cd atoms in C₂N₂-ortho coordination environment were predicted with great potential as hydrogen sensing materials. This method provides a way to accelerate the discovery of innovative materials by avoiding the time-consuming empirical principles in experiments.