Mineralogy from µXRF Elemental Maps: Comparison of Supervised and Unsupervised Methods Open Access

This study evaluates the effectiveness of four methods for estimating mineralogy from micro-XRF (µXRF) elemental maps: two optimized/supervised methods (linear programming and Random Forests) and two unsupervised methods (k-means clustering and UMAP combined with HDBSCAN). Micro-XRF is a non-destructive, high-resolution imaging technique that provides spatially resolved elemental data, and bridges the scale gap between micro- and macro-scales in mineralogical studies. Analyzing samples from porphyry copper deposits (representing a range of textures and mineralogical complexities), each method was assessed based on its ability to accurately represent the mineralogy and texture of each sample relative to a petrographic image. Compared to traditional petrographic assessment, Random Forests yielded the lowest normalized root mean squared error (NRMSE) values (0.09–0.38), while k-means (0.06–0.44), linear programming (0.07–0.44), and UMAP-HDBSCAN (0.18–0.64) showed more variable and generally higher error values across the samples. K-means and linear programming produced results within minutes, Random Forests in ∼10 minutes, and UMAP-HDBSCAN in ∼20 minutes. The results demonstrate that supervised methods, particularly Random Forests, deliver better mineralogical estimates, especially in complex mineral assemblages. However, unsupervised methods (particularly k-means) offer a faster initial assessment, making them valuable for preliminary analysis. Combining different methods using known mineral compositions or training datasets could enhance cluster-to-mineral correlation. The integration of supervised, optimization, and unsupervised approaches for mineralogy from µXRF data can enhance the robustness and efficiency of mineralogical interpretations, adding significant value across mineral exploration and processing.