Enhancing Lithological Interpretation: An Integrated Machine Learning Approach for Automated Log Classification in Volcanic Geothermal Reservoirs, Japan

Abstract Japan's volcanic geothermal reservoirs present significant subsurface characterization challenges due to complex lithological heterogeneity, extensive faulting, and fracture-controlled permeability. Conventional lithological evaluation often relies on adequate wireline and logging-while-drilling (LWD) measurements, drilling cuttings and expert borehole image interpretation. However, the logging variety is often limited, and drilling cuttings recovery is frequently poor or absent within productive geothermal feed zones because of severe mud losses. As a result, critical information is unavailable where accurate reservoir characterization is most needed. This study presents an integrated machine learning workflow for automated lithofacies classification using borehole image data. The methodology combines unsupervised and supervised learning within a two-stage framework. In the first stage, key geological attributes, including high-frequency resistivity, boundary dip trend, and fracture density, are extracted from borehole images and used as inputs to an Index Probabilized Self-Organized Mapping (IPSOM) workflow (Leon et al., 2015). The resulting clusters are calibrated using geological interpretation and drilling cuttings information to establish lithofacies classes. In the second stage, these calibrated lithofacies labels are used to train a supervised deep learning model that directly predicts lithofacies from borehole images. In this study, the IPSOM workflow successfully differentiated major volcanic lithofacies, including fractured andesite tuff and highly fractured dacite tuff intervals. The supervised model achieved a validation F1 score of 79%, demonstrating its ability to learn lithofacies characteristics from borehole image data. Application to blind test intervals showed promising performance in identifying highly fractured dacite tuff associated with permeable fluid-flow pathways. By integrating geological expertise with machine learning, the workflow reduced interpretation time from approximately one month to 4–6 hours while maintaining geological consistency. The methodology provides a scalable framework for rapid lithofacies characterization and supports more efficient geothermal reservoir evaluation and development.