Abstract: Shale oil and gas accumulations in Hongxing area are closely related to lithofacies. However, traditional methods of lithofacies identification based on geologic and core data suffer from low accuracy and limited applicability. To address the bottleneck issue of low identification accuracy, we construct a heterogeneous ensemble learning model that integrates three machine learning methods: Bayesian classification, back propagation neural network (BP neural network), and random forest. The study adopts a 3-endmember 4-component scheme based on mineral composition (siliceous, argillaceous, and calcareous) and total organic carbon content, which classifies Hongxing shales into five lithofacies types: carbon-rich hybrid shales, carbon-rich siliceous shales, high-carbon hybrid shales, high-carbon siliceous shales, and medium-carbon calcareous shales. An adaptive probability fusion mechanism is employed to integrate the learning strengths of these three methods. Experimental results demonstrate that the new model achieves a comprehensive identification accuracy of 93.19% in the test set, representing an improvement of 14.85% over single-method approaches on the average. High robustness could be perfectly maintained in some extreme conditions: insufficient sample size, strong data noise, and class imbalance. In view of its spatial generalization capacity confirmed by cross-regional validation (in Fuxing area), this model provides a high-precision lithofacies identification tool for shale gas sweet spotting.