Fast prediction of adsorption isotherms is of great importance in the structural characterization and property prediction of zeolites prior to the synthesis of the target zeolite. Here, we employ the feature learning (FL) method to simulate the adsorption isotherms through density functional theory data generation of binding strength of nitrogen molecule adsorption in zeolites. Three features, that is, the size of adsorption cavities, the geometry of the pore apertures, and the local geometric distortion, are identified to control the binding strength, qualitatively in terms of a polynomial generative model. The local distortion of the Si–O–Si linkage is correlated with the electrostatic polarization of the nitrogen molecule upon adsorption in the zeolite cage. The electrostatic polarization from local zeolite environment can be further adjusted by the Si/Al ratios and thus can enhance the binding strength even twice as the silicalite zeolites when Si/Al = 47. The predicted adsorption isotherms of MFI and MWW zeolites from the feature-learned binding energies are in qualitative agreement with the experimental Brunauer–Emmett–Teller detection data. For 200,429 zeolites, including 248 known zeolites reported in the International Zeolite Association (IZA), the predicted adsorption isotherms are classified into three groups, among which the zeolites with the medium-sized three-dimensional channel architectures are favorable for nitrogen adsorption. The proposed FL scheme is a practical tool to give a quick prediction of adsorption isotherms and high-throughput screening of porous materials for adsorption-related applications.