| 老挝他曲钾盐矿床测井岩性的机器学习识别与模型对比 |
| Received:August 19, 2025 Revised:January 03, 2026 点此下载全文 |
| 引用本文：DING Jian,FENG ZhiBing,YUAN XingMin,WANG ChunLian,JIANG Li.2026.Machine learning for lithology identification and model comparison from well logging in Thakhek potash mine, Laos[J].Mineral Deposits,45(1):121~140 |
| Hits: 144 |
| Download times: 83 |
| Author Name | Affiliation | E-mail | | DING Jian | State Key Laboratory of Nuclear Resources and Environment, East China University of Technology, Nanchang 330013, Jiangxi, China | | | FENG ZhiBing | State Key Laboratory of Nuclear Resources and Environment, East China University of Technology, Nanchang 330013, Jiangxi, China | zbfengjl@163.com | | YUAN XingMin | Qinghai Geological Bureau of Nuclear Industry, Xining 810001, Qinghai, China | | | WANG ChunLian | MNR Key Laboratory of Metallogeny and Mineral Assessment, Institute of Mineral Resources, Chinese Academy of Geological Sciences, Beijing 100037, China | | | JIANG Li | State Key Laboratory of Nuclear Resources and Environment, East China University of Technology, Nanchang 330013, Jiangxi, China | |
|
| 基金项目:深地国家科技重大专项(编号:2024ZD1003300); 中国铀业有限公司-东华理工大学核资源与环境国家重点实验室联合创新基金项目(编号:2023NRE-LH-08); 江西省自然科学基金项目(编号:20252BAC240270)联合资助 |
|
| 中文摘要:精准识别地层岩性是钾盐矿层位厘定与资源量估算的重要地质依据。文章以老挝他曲钾盐矿区为研究对象，基于区内3口钻井的钾盐测井数据，划分训练集与验证集，并预留四口井作为盲井进行模型验证。采用超参数搜索策略优化模型，对比了随机森林、GBDT、XGBoost、CatBoost、SMOTE-CatBoost及Stacking集成算法在岩性识别中的应用效果，其中SMOTE技术用于改善样本不均衡问题。结果表明，Stacking集成模型泛化能力最优，其外部测试宏平均F1分数达81.35%，井间平均准确率为96.38%;SMOTE-CatBoost模型次之；GBDT模型效果最差，宏平均F1分数仅为70.12%，平均准确率为93.25%。Stacking集成模型通过融合随机森林、XGBoost和CatBoost等多类具有差异学习偏差的基模型，显著提升了蒸发岩系中薄互层岩性的综合识别能力，为深部钾盐矿勘探提供了有效技术支撑。 |
| 中文关键词:钾盐 测井数据 机器学习 岩性识别 集成学习 SMOTE Stacking模型 |
| |
| Machine learning for lithology identification and model comparison from well logging in Thakhek potash mine, Laos |
|
|
| Abstract:Accurately identifying formation lithology serves as a crucial geological basis for determining potash ore horizons and estimating resources. This study focuses on the Thakhek potash mining area in Laos. Using potash logging data from three boreholes in the region, the dataset was divided into training and validation sets, with an additional four boreholes reserved as blind wells for model validation. The model was optimized through a hyperparameter search strategy, and the performance of various algorithms—including Random Forest, GBDT, XGBoost,CatBoost, SMOTE-CatBoost, and Stacking ensemble method—was compared for lithology identification. The SMOTE technique was employed to address sample imbalance issues. Results indicate that the stacking ensemble model demonstrated the best generalization capability, achieving a macro-average F1 score of 81.35% on external testing and an average inter-well accuracy of 96.38%. The SMOTE-CatBoost model performed second best,while the GBDT model yielded the poorest results, with a macro-average F1 score of only 70.12% and an average accuracy of 93.25%. By integrating multiple base models such as Random Forest, XGBoost, and CatBoost, which exhibit diverse learning biases, the Stacking ensemble model significantly enhanced the comprehensive identification ability of thin interbedded lithologies in evaporite sequences, providing effective technical support for deep potash exploration. |
| keywords:sylvite logging data machine learning lithological identification ensemble learning SMOTE Stacking model |
| View Full Text View/Add Comment Download reader |
|
|
|
