Rational design of catalysts relies on a deep understanding of the active centers. The structure and activity distribution of active centers on a surface are two of the central issues in catalysis and important targets of theoretical and experimental investigations. Herein, we report a machine learning-driven adequate sampling (MLAS) framework for obtaining a statistical understanding of the chemical environment near catalyst sites. Combined strategies were implemented to achieve highly efficient sampling, including the decomposition of degrees of freedom, stratified sampling, Gaussian process regression, and well-designed constraint optimization. The MLAS framework was applied to the rate-determining step in NH3 synthesis, namely the N2 activation process. We calculated the produced population function, PA, which provides a comprehensive and intuitive understanding of active centers. The MLAS framework can be broadly applied to other more complicated catalyst materials and reaction networks.