Characterising Heavy Mineral Concentrate Grain Morphology and Mineralogy with Computer Vision

Traditional geological methods for quantifying grain morphology and mineralogy are often manual, time-consuming, and rely on subjective classification criteria. This case study delineates and characterises individual heavy mineral concentrate grains ranging from 300 to 100 microns in diameter within sample sizes ranging from hundreds to thousands of grains. We compare two segmentation algorithms designed to enhance the quantification of granule morphology: a) a Euclidean-based watershed (EBW) technique and a deep convolutional neural network, and b) the Segment Anything (SAM) model from Meta, for their grain segmentation accuracy. Empirical results indicate a marked improvement with the SAM model, showcasing a Dice Similarity Coefficient (DICE) of 0.974 compared to 0.665 for the watershed method, a Jaccard index of 0.949 versus 0.526, a Grain Count Accuracy (GCA) of 0.939 as opposed to 0.393, and an F-index of 0.974 compared to 0.665. We then integrate the SAM method with micro-X-ray Fluorescence (XRF) analysis to correlate mineralogical and morphological data on a grain-by-grain basis. This approach has been tested at scale in a robot-enabled commercial laboratory. It is a significant improvement in analytical precision and operational efficiency in joint characterisation of mineralogy and morphology of samples for geologists, geometallurgists, and process engineers in the environmental and mining industries.