Abstract
Lithology is a key parameter in reservoir fine description and evaluation. It is difficult to directly identify reservoir lithology using a single logging curve or conventional cross-plot method due to the mud-gravel mixing in complex reservoirs. The accurate identification of conglomerate reservoir lithology has always been a high-profile issue in reservoir characterization. In this study, over 70 m of cores were observed in detail. And the manually identified lithology after depth correction is matched with five log curves, including GR (natural gamma ray log), DT (Delta-T Compressional log), RHOB (density log), TNPH (thermal neutron porosity log), and M2R1 (shallow high resolution array induced resistivity log). With the logging data as input, three machine learning models were built separately, and the prediction results were compared through different methods, including accuracy analysis parameters and ROC curves. The results show that the machine learning model based on logging data has excellent performance in the lithology prediction of conglomerate reservoir, and the XGBoost model shows the best prediction results with the highest prediction accuracy of 0.902. In addition, the optimal model is interpreted by SHAP method. On the whole, TNPH curve plays the most important role in lithology prediction. This study offers valuable insights into lithology prediction for complex reservoirs.