The identification of the “sweet spot” of low-permeability sandstone reservoirs is a basic research topic in the exploration and development of oil and gas fields. Lithology identification, reservoir classification based on the pore structure and physical properties, and petrophysical facies classification are common methods for low-permeability reservoir classification, but their classification effect needs to be improved. The low-permeability reservoir is characterized by low rock physical properties, small porosity and permeability distribution range, and strong heterogeneity between layers. The seepage capacity and productivity of the reservoir vary considerably. Moreover, the logging response characteristics and resistivity value are similar for low-permeability reservoirs. In addition to physical properties and oil bearing, they are also affected by factors such as complex lithology, pore structure, and other factors, making it difficult for division of reservoir petrophysical facies and “sweet spot” identification. In this study, the logging values between low-porosity and -permeability reservoirs in the Paleozoic Es3 reservoir in the M field of the Bohai Sea, and between natural gamma rays and triple porosity reservoirs are similar. Resistivity is strongly influenced by physical properties, oil content, pore structure, and clay content, and the productivity difference is obvious. In order to improve the identification accuracy of “sweet spot,” a semi-supervised learning model for petrophysical facies division is proposed. The influence of lithology and physical properties on resistivity was removed by using an artificial neural network to predict resistivity R0 saturated with pure water. Based on the logging data, the automatic clustering MRGC algorithm was used to optimize the sensitive parameters and divide the logging facies to establish the unsupervised clustering model. Then using the divided results of mercury injection data, core cast thin layers, and logging faces, the characteristics of diagenetic types, pore structure, and logging response were integrated to identify rock petrophysical facies and establish a supervised identification model. A semi-supervised learning model based on the combination of “unsupervised supervised” was extended to the whole region training prediction for “sweet spot” identification, and the prediction results of the model were in good agreement with the actual results.