Machine learning algorithm optimization for intelligent prediction of rock thermal conductivity: A case study from a whole-cored scientific drilling borehole

• Published in: Geothermics 2023-06, Vol.111, p.102711, Article 102711
• Main Authors: Pang, Yumao, Shi, Bingbing, Guo, Xingwei, Zhang, Xunhua, Wen, Yonghang, Yang, Guoxin, Sun, Xudong
• Format: Article
• Language: English
• Subjects: convolutional neural network; machine learning algorithm; particle swarm optimization; random forest; support vector regression; thermal conductivity
• ISSN: 0375-6505; 1879-3576
• DOI: 10.1016/j.geothermics.2023.102711

•Machine learning algorithms are an effective way to obtain the thermal conductivity of rocks•Random forest model shows stronger applicability and higher accuracy in thermal conductivity prediction•P-wave velocity parameter can effectively improve the accuracy of thermal conductivity prediction Rock thermal conductivity (TC) is a key parameter in geothermal, petroleum geology, and basin research. Although experimental analysis relying on core samples is currently an effective way to obtain the TC of rocks, it is not always feasible as most boreholes have no or limited cores, making it difficult to establish TC model of geological body efficiently and accurately. This study sheds light on how machine learning algorithms can be used to accurately predict rock TC with easily accessible and high-resolution logging data. Based on 295 measured TC data, logging data, and vertical seismic profile (VSP) data collected from the whole-cored CSDP-2 borehole, the models including random forest (RF), convolutional neural network (CNN), support vector regression (SVR), and particle swarm optimization-SVR (PSO_SVR), were applied to predict the TC of the entire borehole section. Using a confusion matrix analysis, the primary-wave velocity (PWV) from the VSP survey, measured density, and logging data, including shallow lateral resistivity (LLS), compensated neutron log (CNL), density (DEN), gamma rays (GR), spontaneous potential (SP), and acoustic transit time (AC), were used as the input variables for training models and TC prediction. The results showed that the geophysical parameters reflecting those properties related to mineral composition, porosity and reservoir fluids of geological body can be well used to predict TC through machine learning algorithms. Regardless of unconsolidated sediments or rocks, the RF model showed stronger applicability and higher accuracy in TC prediction compared to the PSO_SVR and CNN models, while the SVR model showed poor applicability in this case study. The PWV data can effectively improve the accuracy of TC prediction of all models, and the RF model finally yielded the excellent performance with a correlation coefficient (> 0.86) and root mean squared error (8%) between the predicted and measured values. Future research should focus on studying the weighted analysis for the correlation between geophysical parameters and TC and developing more accurate predictive models for wider applications. [Display omitted]