Real-Time Automated Fracture Detection in Geothermal Wells Using Low-Cost Drilling Data

Geothermal reservoirs present significant potential for sustainable energy extraction, with natural fractures enhancing reservoir permeability and fluid flow. However, despite their benefits for production, these fractures pose significant drilling challenges, including wellbore instability, loss of circulation, and drill string vibration. Although advanced techniques such as logging while drilling provide real-time formation evaluation and fracture identification, their high costs limit their application. This study proposes a cost-effective alternative using low-cost drilling data for real-time fracture identification. Various drilling parameters from the Utah FORGE injection well (16A(78)-32) were analyzed to identify key indicators of natural fractures. The continuous wavelet transform (CWT) was then applied using multiple wavelet types as a flexible mathematical tool to extract the most accurate signature of fracture variations in the drilling parameters. Normalized CWT scalogram characteristics were then correlated with fracture density distribution to extract meaningful signal features. Various machine learning (ML) classification algorithms were trained and tested to determine the best model for predicting fracture density and location. Finally, the developed model was validated using data from a separate well, FORGE production well 16B(78)-32, to assess its reliability and accuracy. The results indicate that the rate of penetration (ROP), mechanical specific energy (MSE), and drilling torque are the most significant indicators of the presence of natural and induced fractures. Specifically, fractures are associated with higher ROP, lower MSE, and lower-frequency fluctuations in drilling torque. Among the tested wavelet models, the Shannon wavelet provided the most accurate pattern recognition using a scalogram. The physical relationship between ROP, MSE, drilling torque, and fracture frequency distribution was analyzed, and their normalized CWT scalogram features were integrated with fracture density to enhance the accuracy of the ML models in predicting fractured regions. Five ML models were trained and tested, with XGBoost demonstrating the highest performance among the evaluated models, achieving an accuracy of 95.65%, a precision of 95.68%, and a recall of 95.65%. The high Matthews Correlation Coefficient (MCC) and Cohen's Kappa values (both approximately 0.95) further validate the robustness of the classification results, indicating a strong agreement between predicted and actual fracture locations. Finally, the proposed approach, when applied to the production well 16B(78)-32, predicted the fracture density and location with a precision of 95.68%. This study presents an automated model integrating CWT with ML for real-time fracture detection in geothermal wells. By uncovering the hidden patterns between drilling parameters and natural fracture density and location, it can pinpoint fracture zones prone to loss of circulation, drillstring vibration, and wellbore instability. Thus, it enhances drilling efficiency, facilitates the selection of appropriate fluid loss materials (type and size), and ultimately reduces drilling costs. Additionally, the proposed approach could be applied to other subsurface applications.