Abstract: The coal-measure tight sandstone in the DJ block of the Ordos Basin is an ultra-low porosity and permeability reservoir, and strong heterogeneity is a key problem in reservoir prediction and evaluation. Based on 1 880 rock cores plug porosity and permeability analysis data and 33 sets of mercury intrusion analysis data, the probability cumulative distribution function of the flow interval index and pore throat radius was calculated, and the three flow units classification scheme with FZI was optimized. Tosolve the problem of the imbalance of the number of different types of sandstone samples, 734 sets of macro-rock equivalent type data and numerical logging data were subjected to sub-stratified sampling and homogenization sampling processing, and the data subset was reconstructed;Through the interactive verification method of parameter setting-training-modeling-testing-performance evaluation, the modeling parameters were optimized, and machine learning models of 7 algorithms in two categories, unsupervised and supervised, were constructed. They are K-means clustering, naive Bayes, decision tree, support vector machine, deep learning, random forest and gradient tree boosting respectively, and the two optimal classifications of gradient boosting tree learning and deep learning are determined according to the f1 score value. The research results show thatthe sandstone classification based on FZI flow unit can significantly reduce the influence of heterogeneity on the accuracy of permeability prediction; compared with unsupervised method, the supervised and active machine learning methods are more suitable for the classification principle of flow unit. Gradient boosting tree learning algorithm can use lithofacies, diagenetic facies, sedimentary micro-equal type data modeling, and can be used for basic geological research such as the genetic characteristics of flow units, while deep learning has stronger processing capabilities for numerical data such as well loggingand is suitable for engineering applications such as permeability prediction. The machine learning model with efficient and reasonable classification for flow units is not unique and can be constructed independently according to research purpose and data typesto optimize the algorithm.