From logs to layers: Framework for interpreting subsurface facies in Iceland geothermal reservoirs

• Propose a Gaussian Mixture Model framework for lithological clustering using well log data (i.e., Gamma Ray, Neutron, Resistivity). • Identifies thin intervals and reclassifies missing cuttings in geothermal wells. • Improves geological models in high-temperature geothermal fields. • Highlights the utility of machine learning in challenging volcanic environments. Accurate identification of borehole rock units plays a critical role in understanding the subsurface configuration and fluid flow dynamics, which are essential inputs for developing robust reservoir models and designing wells to optimize steam production in geothermal fields. This study demonstrates the effectiveness of Gaussian Mixture Modeling (GMM), an unsupervised machine learning algorithm that clusters data based on similar characteristics, for characterizing rock unit properties—such as lithology, porosity, and alteration state—directly from well logs without relying on retrieved cuttings. Using Gaussian Mixture Model-based clustering, we analyze three geothermal wells from northern Iceland: two from the Krafla geothermal field (IDDP-1 and KJ-39) and one from the Theistareykir geothermal field (THG-13). A structured workflow incorporating gamma-ray, resistivity, and neutron log data enables the classification of rock units and the differentiation of compositional and textural variations, especially in intervals with missing cuttings. The clustering results reveal three primary groups of rock units—mafic, intermediate, and felsic—and distinct porosity and alteration characteristics. This methodology effectively assigns rock types and porosity categories to intervals with missing cuttings and corrects for intervals that are potentially misassigned due to misinterpretation arising from mixed cuttings. Gamma-ray values are found to effectively distinguish compositions ranging from mafic to felsic, while a matrix for interpreting resistivity and neutron aids in identifying porosity and alteration states for each rock unit. The developed matrix is calibrated for each well, enhancing its utility as a relative guide for interpreting geothermal facies across diverse volcanic lithologies. This study demonstrates the importance of integrating geophysical log-based clustering techniques with traditional geological analysis to refine geothermal reservoir characterization, offering a valuable approach for resolving facies in geothermal fields with limited cutting recovery.