Deep learning-assisted borehole image analysis for enhanced geothermal reservoir evaluation

Abstract

Evaluating the geothermal reservoir potential often requires fracture analysis, as fractures serve as key pathways for fluid flow in subsurface formations. Borehole images (BHIs) are essential for this analysis, providing 2D representations of boreholes with millimetre-scale resolution. However, their interpretation is highly subjective, leading to uncertainties in the results and the subsequent quantitative assessment of the fracture networks. In the West Netherlands Basin (WNB), accurate fracture characterization is critical for assessing the geothermal viability. However, the traditional manual interpretation of BHIs has shown inconsistencies. This study introduces a supervised deep learning (DL) approach to support fracture analysis using high-resolution formation micro-imager (FMI) data from the Naaldwijk well (NLW-GT-01). The proposed DL-based system integrates a U-Net model (PickNet) for segmentation and a fully connected convolutional network (FitNet) for automated feature extraction. Initially trained on synthetic low-resolution BHIs, the model has been adapted for FMI data using two approaches: (1) transfer learning and (2) a simplified adaptation method that involves resizing the FMI input, leading to some resolution loss. A comparison of these approaches has revealed that the simplified adaptation produces better results, closely aligning with conservative manual interpretations calibrated with core samples while enabling more detailed fracture detection. To enhance reliability, we propose a semi-automated human–machine collaboration framework, where experts validate or refine the automatically detected features. This approach leverages human expertise to improve interpretation accuracy while addressing challenges related to robustness and redundancy in the supervised learning model.