The Application of Machine Learning Methods to Portable X-Ray Fluorescence Data: A Case Study from Zwartfontein Farm, Northern Limb of the Bushveld Complex, South Africa

Geological characterisation during mineral exploration and mining operations currently relies on costly laboratory analyses with slow turnaround times and subjective visual logging. These create bottlenecks and bias in decision-making. This thesis demonstrates how a workflow combining compositional data analysis, wavelet tessellation and machine learning can rapidly transform low-cost portable X-ray fluorescence (pXRF) data into interpretable geological information. The development of methods was tested, as a case study, on the Northern Limb of the Bushveld Complex of South Africa – a complex multi-commodity deposit (Ni-Cu-(Co)-PGE) comprising intricately layered mafic-ultramafic cumulate rocks overprinted by variable alteration minerals. Although pXRF provides rapid multi-elemental analysis, individual measurements on whole or half core often lack precision, due to the small measurement window on the instrument (relative to grain size of core) and limitations for lighter elements inherent in pXRF technology. Crucially however, the relative compositions recorded by pXRF down-core show broadly similar trends as in laboratory-derived assays, enabling the use of machine learning techniques on these large-volumes, noisy datasets. Three objectives were addressed using over 200,000 pXRF measurements from the Zwartfontein Farm area of the Northern Limb: First, a lithology prediction workflow was developed using supervised machine learning with a novel application of the tessellated wavelet transform for noise reduction. Secondly a machine learning approach to determine mineralogy was established from pXRF and hyperspectral datasets, addressing the limitations of traditional normative calculations in hydrothermally altered rocks. This was alongside the development of a web-based application (webNORM) for accessible normative mineralogy calculations. Thirdly, the prediction of intervals of high platinum-group element (PGE) grade is developed via application of decision tree-based algorithms to pXRF-derived proxies for the PGE. The integrated workflow enables prediction of lithology, mineralogy, and grade from a single pXRF dataset, providing immediate operational feedback during drilling campaigns. For best applicability, models trained on the Zwartfontein Farm area most likely require retraining for adjacent areas along the Northern Limb of the Bushveld, due to the supervised learning approach adopted here. Nonetheless, the approaches used here have broad applicability to other mineral systems. This thesis fundamentally demonstrates how machine learning can unlock the potential of large volumes of low-cost, noisy data, such as pXRF, to provide rapid geological characterisation in an exploration campaign, with the potential to provide significant savings in time, resources and costs. It is envisaged that similar workflows could be developed in a mining and production environment to the same end.<p></p>

Author dc:creator

• Tom Buckle (21040781)

• machine learning
• Bushveld Complex
• platinum group elements
• pXRF

Statement dc:rights

• All rights reserved
• Open Access after 2031-04-27

Identifier

10779/exe.32229300.v1

OAI identifier oai:identifier

oai:figshare.com:article/32229300