The complex diagenetic facies of the tight sandstone reservoir in the Shaximiao Formation, located in the Jinqiu gas field to Tianfu gas area in the central Sichuan region, pose significant challenges to reservoir evaluation and natural gas exploration and development. Traditional diagenetic facies identification methods are often low in accuracy, heavily reliant on specialized personnel, and time-consuming. There is an urgent need for a diagenetic facies identification method that is highly accurate, cost-effective, and fast. Firstly, based on cast thin section identification data, the lithology of the tight sandstone was determined using a ternary plot of components. Image processing techniques were then used to identify the types and proportions of pores and cements, and the diagenetic facies of the tight sandstone were classified. Secondly, the corresponding 1 019 depth-based well log data for core-divided diagenetic facies were analyzed in terms of distribution range, median, uniformity, and skewness. These 6 types of well log data were standardized to a 0-1 range, and data imbalance was addressed using synthetic minority over-sampling technique (SMOTE). Finally, 10 traditional machine learning algorithms and ensemble learning algorithms were selected for model training and performance comparison. The study found that ensemble learning algorithms, especially the extreme randomized trees (ET) algorithm, performs best in diagenetic facies identification, achieving higher accuracy and F₁ scores than traditional machine learning algorithms. This significantly improved identification accuracy and stability. The ET model was then used to predict the diagenetic facies of the JQ8 well, validating the feasibility of the method. This study provides effective technical methods and references for diagenetic facies research in tight sandstones.