Application of sedimentological and gold assay data for resource estimation: a case study of the Middelvlei Reef, Witwatersrand Basin (South Africa)

Abstract

Block modelling applied to point data is common practice in estimation of resources in the mining industry. This study compares the results of geostatistical block-modelling techniques applied to two different approaches to obtaining the point data. In the first instance, a novel approach to generating point grade data using a machine learning algorithm to predict metal grades that are based on a support of assay, reef width and additional descriptive geological data is considered. This is in contrast to the conventional approach, which uses the assay data, reef widths only and interpreting geozones for estimation. The relevance and importance of this study centres on the need to efficiently use the vast amount of additional available historical data collected from the Witwatersrand Supergroup, at a time when most of the primary reef horizons are being depleted and information on secondary reef intersections is generally limited. Several studies of the Middelvlei Reef have indicated a notable relationship between its sedimentological and depositional characteristics and the distribution of gold. The Middelvlei Reef is a medium- to large-pebbled oligomictic, auriferous conglomerate with interbedded quartzite and a thickness of two to four metres, along with pebble lags as thin as 10 cm. It is mined selectively, primarily for its gold content as a secondary reef at the study site. This study compares ordinary and simple kriged block model estimates using the same omni-directional spherical variogram, based on observed assay data and enhanced quality grade data predicted by GS-Pred using the assay data and accompanying geological descriptions. The comparisons are conducted on the swath plots generated from the estimations on a more local scale. The swath plots indicate that the estimates for the data predicted by GS-pred produces estimates that are aligned to the sample points, particularly where the number of samples in the swaths is higher and where declustering has been carried out. The results indicate that data derived from GS-Pred can be incorporated into existing resource estimation procedures relevant to the operational mining environment. As GS-Pred is sensitive to the quality of the geological data collected, this study recommends that the limitations of the reliance on variable quality and often qualitative historical data be addressed to enhance GS-Pred performance and consistency. The recommendation also holds for other similar iii machine learning-based approaches, as it pertains to data quality rather than algorithm formulation. For example, much of the geological descriptions, such as matrix colour and percentage mineralisation, are subjective and inconsistently interpreted within and between operations. Also, additional validation steps on assumptions and theory in the GS-Pred process should enhance this novel approach