Integration of Machine Learning Algorithms with Gompertz Curves and Kriging to Estimate Resources in Gold Deposits

• DOI: 10.1007/s11053-020-09750-z
• Publication Date: 2020-09-27T16:02:20Z
• Authors: Steven E. Zhang; Glen T. Nwaila; Leon Tolmay; Hartwig E. Frimmel; Julie E. Bourdeau
• Keywords: 0101 mathematics; 01 natural sciences

Resource estimation on gold (Au) deposits usually requires costly Au assays and is often characterized by high degree of uncertainty especially in areas with limited number of samples. This paper reports a refinement of a novel machine learning approach (GS-Pred) that incorporates network analysis for geology-based anomalous data detection and outlier removal, and adopts feature weighting using a spatial self-similarity model inspired by kriging to enhance prediction performance for in situ Au-grade estimation. In this application, machine learning algorithms are integrated with sequential-kriging block modeling for high resolution in situ grade estimation. This process is fully automatable, and it utilizes both geological data and Au assays, making it possible to also estimate Au grade in areas that only have geological descriptions. The results of our expanded GS-Pred and block modeling, using data from the auriferous conglomerates of the Witwatersrand Basin (South Africa), demonstrate improvements in the GS-Pred performance and flexibility relative to the original algorithms. Additionally, our results provide further evidence of strong sedimentological control on Au concentration within the Witwatersrand Basin, which is suitable for quantitative predictions. Our algorithms feature fast data processing, geology- and assay-based outlier detection, visualization of complex geospatial data, and they open new avenues for intelligent and automated in situ Au-grade prediction. We demonstrate that GS-Pred target predictions are feasible substitutes for assays for the purpose of block modeling under suitable deployment conditions.