Intelligent Lithology Identification of Conglomerate Based on K-Means-RBF Clustering Method Using Well Logging Data

Jing Zhang^1,2Jun Zhao^3*Jianhua Qin²Yingwei Wang²Chao Zheng³Lin Liu²Daiyan Zhang²

• ¹China University of Petroleum Beijing, Changping, China
• ²Xinjiang Oilfield Company, Karamay, China
• ³Southwest Petroleum University School of Geoscience and Technology, Chengdu, China

Abstract: Lithology identification serves as the foundation of reservoir evaluation and is a critical step in both reservoir parameter estimation and reservoir development assessment. Well logging data, which contain abundant subsurface information, are commonly used as the primary data source for lithology identification. However, due to the influence of multiple factors during the interpretation process, logging responses often exhibit non-uniqueness, resulting in ambiguous outcomes. Consequently, traditional methods often fail to meet the accuracy requirements of practical lithology identification. To address this issue, this study proposes an intelligent lithology identification method that integrates K-Means clustering with a Radial Basis Function (RBF) neural network, leveraging the advantages of machine learning in data analysis and modeling. Using well logging data from conglomeratic reservoirs as the research object, a correlation analysis was conducted to select optimal logging curves. Five logging parameters—acoustic travel time (AC), compensated neutron porosity (CNL), bulk density (DEN), natural gamma ray (GR), and true formation resistivity (RT)—were identified as input features. The method aims to classify six lithology types: siltstone, fine sandstone, medium to coarse sandstone, gravelly coarse sandstone, conglomeratic sandstone, and conglomerate. To evaluate the performance of the proposed method, three widely used machine learning models—Backpropagation Neural Network (BP), Random Forest (RF), and Support Vector Machine (SVM)—were employed as benchmarks. Under identical training–testing data splits and parameter optimization settings, the K-Means-RBF model outperformed the BP, RF, and SVM models, with improvements in overall classification accuracy of 7.0%, 4.0%, and 8.0%, respectively. These results demonstrate that the proposed K-Means-RBF-based method exhibits strong robustness and promising application potential in well log-based lithology identification.