Local and Target Exploration of Conglomerate-Hosted Gold Deposits Using Machine Learning Algorithms: A Case Study of the Witwatersrand Gold Ores, South Africa

Determining the gold grade and facies type in areas with little geological information and sparse exploration samples is fraught with uncertainties and often results in high operational costs. Point-wise gold grade data are commonly used to guide exploration and resource estimation with the application of spatial interpolation techniques such as kriging. Within this environment of data scarcity, the application of kriging leads to significant grade estimation errors, as high nugget thresholds reduce the effectiveness of kriging, a good example being the gold deposits in the Witwatersrand Basin of South Africa. To reduce the impact of subjective grade interpolation and geological interpretation, as well as to exploit currently unused geological descriptions, we present a novel machine learning-based algorithm called GS-Pred. It combines both sedimentological and gold assay data for point-wise gold grade prediction and automated facies identification in a conglomerate-hosted gold deposit. For this application, GS-Pred requires an input database of sedimentological descriptions, spatial information and gold grades and makes predictions of gold grades at any point within the spatial coverage of the input database, provided that it has appropriate sedimentological descriptions. In essence, GS-Pred examines the spatial and non-spatial variability of metal grades and provides information of the estimated resource below the nugget threshold. This proposed algorithm has been validated on subsets of data on gold grade and sedimentological characteristics of conglomerates in the Witwatersrand Basin. Validation results suggest that GS-Pred is more accurate than current machine learning techniques and ordinary kriging. The clustering result shows that there are four or at most five facies which can be distinguished from the clustering results within the dataset, which maximises the contrast in the inter-cluster prediction behavior. These clusters have a good spatial correspondence with the known geology, and the method, combined with gold grade predictions, was able to identify probable mineralization patterns, thus assisting in target exploration. This novel machine learning algorithm is entirely data driven. We have shown its successful application in a complex geological setting as the Witwatersrand Basin.