Machine-learning approach for mineralogy prediction from drill core images in clay-rich and adjacent formations

High-resolution mineralogical characterisation of drill cores is an important aspect of subsurface exploration. Conventionally, this requires a geophysical logging campaign with subsequent interpretations based on statistical methods and laboratory-based X-ray diffraction (XRD) measurements on selected core samples for calibration. While these methods are proven to provide reliable mineralogical information, lab analyses rely on sample availability (i.e., cores or, less useful, cuttings), are time-consuming, costly and thus limited (i.e., sparse) along boreholes. In this study, we present a novel image-based approach that uses drill core photographs in combination with computer vision and machine learning to predict mineralogical composition at high frequency. A transfer learning approach is employed, in which a ConvNeXt-Tiny convolutional neural network model is first trained for lithological formation classification. Subsequently, this pre-trained model is fine-tuned for mineralogical regression. The data from XRD measurements serve as ground truth labels and are grouped into four mineral classes: total clay, quartz-feldspar silicates, carbonates and a class integrating the rest of the minerals. The dataset contains seven boreholes from Nagra drilling campaigns and is divided into training, validation and test subsets at the borehole level. The trained model is applied to the drill core images of each borehole to generate mineralogical depth profiles, which are evaluated against XRD measurements and MultiMin-derived mineralogy (Marnat & Becker 2020). The results demonstrate that ConvNeXt-based transfer learning enables robust, image-only prediction of mineralogical trends across boreholes, and highlight the potential as a complementary tool to conventional laboratory-based mineralogical analyses.