Keywords: 
• Key words: Carbonate sedimentary facies /
• Random forest algorithm /
• Changxing Formation /
• Sichuan Basin

Abstract: Abstract: [Objective] Accurate identification of carbonate sedimentary facies forms the basis for lithofacies paleogeographic reconstruction. However, challenges arise from the ambiguity of facies characteristics and the subjectivity of manual interpretation, while traditional well-log facies interpretation methods suffer from inefficiency and experience-dependent limitations. This study aims to construct a reusable intelligent interpretation model using machine learning algorithms to overcome the limitations in well-log facies interpretation and enhance the accuracy and efficiency of carbonate facies identification. [Methods] Five boreholes from the reef-shoal facies belt of the Changxing Formation in the Yuanba area of northeastern Sichuan Basin were selected as the study objects. Feature parameters, including acoustic log, gamma-ray well log , and other well-log curves, were used to establish a random forest algorithm-based model for carbonate well-log facies identification. To address class imbalance, the SMOTE-Nearmiss-1 resampling algorithm was implemented, while optimal hyperparameters were determined through grid search coupled with K-fold cross-validation. [Results] The improved model demonstrated significantly enhanced recognition capability for minority classes, achieving an accuracy of 0.87 in single well testing. This case highlights the significant potential of machine learning models in accurately identifying carbonate sedimentary facies.