Multi-element (ME) datasets provide comprehensive geochemical signatures of an orebody and are commonly used to gain insight into the mineralogy, lithology, alteration patterns and to identify target-pathfinders. However, little effort is made in using these data to explain comminution or recovery characteristics. This paper describes an agglomerative hierarchical clustering approach applied to ME data from the Tropicana Gold Mine, Australia, and investigates the relationship between the resultant classes and run-of-mine comminution and recovery parameters. First, it is demonstrated how an industry scale ME dataset is prepared for clustering. The preparation consists of verifying the absence of interlaboratory and intralaboratory bias between measurements, centred log-ratio transformation (clr), normalisation and principal component analysis (PCA). Afterwards, the first case study indicate that the clustering separation is primarily driven by geochemical differences caused by major rock-forming mineral signatures (felsic vs mafic, alteration vs no alteration, chert or quartz lithologies, unmineralised vs mineralised material). This case study separates the ME dataset into five unmineralised and two Au-mineralised material classes. The second case study continues with the two identified mineralised material classes and further separates these samples into five new classes. These classes are explored geochemically and by using the spatial context (within domains) better matched with metallurgical test results. It is found that domain-related material class proportions assist in interpreting different processing proxies such as the Equotip hardness (Leeb), Bond Work index (BWi), Axb, and processing recovery and reagent consumption. Knowledge of the processing parameters per domain and class composition can be used to infer such characteristics in the absence of standard metallurgical tests. This new approach of gaining insights into comminution and recovery parameters through geochemical analysis demonstrates the benefit of the conceptualised material fingerprinting concept.

Geochemical and mineralogical datasets from Tropicana Gold Mine, Australia, have been used to define Au-mineralised fingerprints. VNIR-SWIR spectral data were represented by four normalised wavelength regions and were clustered to form spectral classes. Sequentially, these spectral class proportions within a block and co-located pXRF data were clustered to discriminate material types (fingerprints). The hardness of each type was further explored using collocated BWi, Axb, Equotip rebound hardness and penetration rate datasets, but also by considering spatial contextual relationships and the within material type variability. The Tropicana orebody example gave a good illustration of how a phengitic-epidote K-feldspar rich domain (schistosity and softer, ∼15–18 kWh/t) separated from a harder (>20 kWh/t), shorter wavelength phengitic plagioclase-rich feldspar dominated domain. Exploring the within material type differences using the white mica composition (wAlOH) and a new w605 spectral feature demonstrated how the effects of shearing were captured within material types. Such findings will ultimately improve the understanding of the constitutive material hardness and have significance for process optimisation and blending strategy design.

Material attributes (e.g., chemical composition, mineralogy, texture) are identified as the causative source of variations in the behaviour of mineral processing. That makes them suitable to act as key characteristics to characterise and classify material. Therefore, vast quantities of collected data describing material attributes could help to forecast the behaviour of mineral processing. This paper proposes a conceptual framework that creates a data-driven link between ore and the processing behaviour through the creation of material “fingerprints”. A fingerprint is a machine learning-based classification of measured material attributes compared to the range of attributes found within the mine’s mineral reserves. The outcome of the classification acts as a label for a machine learning model and contains relevant information, which may identify the root cause of measured differences in processing behaviour. Therefore, this class label can forecast the associated behaviour of mineral processing. Furthermore, insight is given into the confidence of available data originating from different analytical techniques. Taken together, this enhances the understanding of how differences in geology impact metallurgical plant performance. Targeted measurements at low-confidence unit processes and for specific attributes would upgrade the confidence in fingerprints and capabilities to predict plant performance.