Mineral classification of lithium-bearing pegmatites based on laser-induced breakdown spectroscopy: Application of semi-supervised learning to detect known minerals and unknown material
Abstract
Mineral exploration and active mining relies on extensive drilling campaigns that produce large numbers of drill cores. LIBS is ideally suited for their fast and effective measurement, but matrix effects complicate quantitative geological LIBS applications due to the extensive amount of different minerals, rock types, and lithologies, as well as all textural and optical parameters increasing physical matrix effects. This is challenging for the application of LIBS in geological exploration, since LIBS data processing highly depends on matrix-matched models. The fast acquisition of new data is in conflict with the large amount of existing minerals and lithologies. As a result, new appearances are common during ongoing drilling campaigns, resulting in incomplete train sets for supervised classification and quantification.
This paper presents a novel semi-supervised learning (SSL) classification model to resolve related issues by separating known minerals in geological drill cores based on a set of train samples, while also detecting unknown material, i.e. new lithologies and/or minerals not in the train set. Using a combination of supervised Linear Discriminant Analysis (LDA) and semi-supervised One-Class Support Vector Machines (OC-SVM), main minerals and known accessory minerals were effectively separated from unknown material in LIBS mappings of Spodumene and Muscovite pegmatite, as well as from Metagreywacke in drill cores from the Rapasaari lithium deposit in Finland. Self-learning was applied to automatically increase the number of train samples, which effectively decreased the number of unknowns due to physical matrix effects in coherent crystals. Validation with respect to the main minerals revealed an almost perfect classification of albite, spodumene, K-feldspar, quartz, and muscovite. Measurement points of Metagreywacke, which were only included in the test set, were correctly detected as unknown. Transferring the developed model onto LIBS mappings and drill core profile measurements displayed excellent classification results for main and accessory minerals included in the train set. Mixed spectra at mineral borders, as well as accessory minerals not in the train set were correctly identified as unknown.- Publication:
-
Spectrochimica Acta - Part B: Atomic Spectroscopy
- Pub Date:
- March 2022
- DOI:
- Bibcode:
- 2022AcSpB.18906370M
- Keywords:
-
- Laser-induced breakdown spectroscopy (LIBS);
- Linear discriminant analysis (LDA);
- One-class support vector machines (OC-SVM);
- Spodumene pegmatite;
- Unknown matrix