天然气勘探与开发2025，Vol.48Issue(5)：48-57,10.DOI:10.12055/gaskk.issn.1673-3177.2025.05.005
基于集成学习的复杂碳酸盐岩矿物组分含量测井分析
Ensemble-learning-based logging analysis of mineral composition contents in complex carbonates
摘要
Abstract
The conventional dual-mineral volume model is not accurate enough in calculating the mineral composition of complex carbonates in the Middle Permian Maokou Formation in the MX-PL area of the Sichuan Basin.To precisely quantify the mineral composition,using the training samples learned from conventional logging data together with mineral lithology scanning logging results,a Natural Gradient Boosting(NGBoost)model was constructed to predict the carbonate mineral contents,and coupled with the SHapley Additive exPlanations(SHAP)for analyzing the logging curves.A dual-driven technological system of"intelligent algorithm+feature attribution"was formed for complex lithology identification and quantitative mineral content evaluation.Research results are obtained as follows:(i)Through the optimization of ensemble learning model,the accuracy of lithology identification for the Maokou Formation has increased from 55%to over 85%;the predicted determination coefficients of compositions including calcite,dolomite and quartz achieve 0.91,0.87 and 0.86,respectively,with the quantitative accuracy for quartz improving sevenfold.(ii)SHAP global attribution indicates that shear-wave slowness contributes most to quartz content prediction,whereas neutron porosity serves as the key factor of sensitive curve for calcite content prediction.(iii)The feature coupling effect suggests that shear-wave slowness and neutron porosity synergistically enhance the positive effect for calcite content prediction,and the positive correlation between compressional-and shear-wave slowness determines the calculation model of quartz content.(iv)NGBoost provides a confidence interval of 80%which covers 78%of measured values,with the mean square error(MSE)lower than 0.003,significantly superior to XGBoot and random forest(RF)algorithms.It is concluded that the SHAP-empowered NGBoost model provides an accurate and interpretable solution for quantitative evaluation of complex minerals,and reveals the coupling patterns of logging curves to guide the optimization of petrophysical models for similar reservoirs.关键词
二叠系中统茅口组/矿物组分/自然梯度提升算法/SHAP模型/特征耦合/集成学习/岩性识别Key words
Middle Permian Maokou Formation/Mineral composition/NGBoost/SHAP model/Feature coupling/Ensemble learning/Lithology identification分类
能源科技引用本文复制引用
赵艾琳,谢冰,吴煜宇,王华,吴毓琼,黄宏,肖雪薇..基于集成学习的复杂碳酸盐岩矿物组分含量测井分析[J].天然气勘探与开发,2025,48(5):48-57,10.基金项目
中国石油天然气股份有限公司重大科技专项"复杂碳酸盐岩地球物理评价关键技术研究"(编号:2023ZZ16YJ04). （编号:2023ZZ16YJ04）
