Multi-domain based mineral resource estimation in narrow stratabound tabular deposits

Abstract

The application of geostatistical methods in mineral resource estimation has been at the core of the mining industry for decades. These methods have been extensively studied and their use in quantifying resources and reserves have resulted in many successful techno-economic valuation of mines. The models generated from these methods influence grade control protocols and ore extraction optimisation to realise economic benefits for the mine. Recently, there has been a rise in the integration of machine learning (ML) approaches for mineral resources estimation. This has been influenced by the increase in computational power, the production of large datasets from exploration and mining operations that have proven difficult to handle, and the growth in digitisation and automation applications in the mining and minerals industry. ML methods have shown in many studies that they provide an opportunity to produce high-accuracy models and the ability to decipher hidden patterns within the data that might have been difficult to model using traditional analytical methods. This research aimed to investigate the application of ML algorithms in generating mineral resource estimation domains used in the estimation of in-situ gold grades of the Carbon Leader Reef in the Carletonville Goldfield of the Witwatersrand Basin (South Africa). The ML algorithms used are Random Forest (RF), k Nearest Neighbour (kNN), AdaBoost (AB) and Support Vector Machine (SVM). The total number of primary samples before study selection were 5611 and after narrowing down the study area, they were 2017. The dataset used in this research comprises (a) sample spatial coordinates (X, Y, Z), (b) gold (Au) grade in grams per tonne (g/t), (c) gold accumulation expressed as Au (cm.g/t), (d) sedimentological properties (i.e., vertical thickness of the orebody commonly referred to as channel width (CW) measured in centimetres (cm), percentage conglomerate within the ore zone, pebble size, packing, roundness, sorting, pebble assemblages), and (e) mineralisation descriptions (i.e., carbon content, colour and basal contact). The integration of ML algorithms with traditional geostatistical methods in mineral resource evaluation workflows has shown positive and promising results. The optimal pointwise domain generation was done using traditional a k means clustering and k-means coupled with signed distance functions and produced two mineral resource estimation domains. The traditional k-means clustering resulted in domain mixing which is unfavourable for resource estimation because it may lead to underestimation or overestimation of grades at these locations and subsequent misclassification of ore and waste. It also may complicate the generation of domain boundaries which is the next step after the pointwise domaining process, the boundary lines may appear obscure and incomprehensible. The signed distance-based k-means clustering model was selected as the best performing model because it resulted in unmixed and clear mineral resource estimation domains which is desirable for grade estimation. These signed distance-based k-means domains were selected as the base of the domain boundaries. The RF model had the highest performance matrices scores, where both the accuracy and F1 scores were 99.9% while the kNN model had the lowest scores with accuracy and F1 scores were 99.4% each. Despite the RF performing better, the kNN model resulted in a more geologically reasonable domain boundary which is consistent with the geological contacts of the ore body under study. The domain boundary displayed an uneven geometry, typical of Carbon Leader Reef’s in variable (e.g. high-to-low energy) depositional environments. These abrupt shifts in energy conditions affect how conglomerates are arranged, which in turn influences grade distributions. Following the definition of estimation domains and subsequent variography, the kNN boundaries were used to guide block ordinary Kriging for grade estimation. Post-estimation metrics showed that both the Kriging Efficiency (KEFF) and the Slope of Regression (SLOR) exceeded 50%, reflecting a favourable sample configuration and confirming the appropriateness of the selected domains and resultant boundaries for resource estimation.