Lithological classification of drill cores using XCT and XRF data: a geometallurgical perspective

• Integrated XCT-XRF framework for automated core lithology classification. • Percentile-based XCT features introduced for texture and density characterization. • Reduced spatial bias using drill hole wise validation and best result achieved by XGBoost. • Key mineralogical and metallurgical drivers for lithology prediction revealed by SHAP analysis. • XCT and XRF fusion advances predictive geometallurgical ore modelling. This study presents an integrated framework for the automated lithological classification of drill cores by combining X-ray computed tomography (XCT) and X-ray fluorescence (XRF) data within a machine learning setting. A total of 293 m of drill cores from the Kalkaroo copper–gold deposit in South Australia were classified into five lithological categories at 5 cm intervals. XCT voxel attenuation values were summarised using percentile-based descriptors to capture density and textural variability, while XRF provided complementary geochemical information. Random Forest (RF) and Extreme Gradient Boosting (XGBoost) classifiers were applied using drill hole-wise leave-one-out cross-validation to mitigate spatial autocorrelation. The integrated XCT-XRF dataset outperformed single-modality inputs, with the XGBoost method achieving the highest predictive performance (Macro-F1 = 0.86). SHapley Additive exPlanations (SHAP) revealed that Cu, Fe, S, Ca, and K were the dominant discriminators, while high-end XCT percentiles enhanced the resolution of dense sulphide phases. This proved critical for distinguishing chalcopyrite-rich intervals, which govern copper recovery and concentrate grade, from pyrite-dominant horizons, which increase lime consumption and depress flotation selectivity. Downhole analysis confirmed the geological plausibility of predictions and highlighted transitions with implications for blending, comminution energy and metallurgical forecasting. The results demonstrate that XCT-XRF integration provides robust, interpretable, high-resolution lithology classification with geometallurgical relevance. Quantitative calibration to comminution indices or flotation performance requires additional datasets and is outside the scope of the present study.