Precursors predicted by artificial neural networks for mass balance calculations: Quantifying hydrothermal alteration in volcanic rocks. (April 2016)

• Record Type: Journal Article
• Authors: Trépanier, Sylvain; Mathieu, Lucie; Daigneault, Réal; Faure, Stéphane
• Is Part Of: Computers & geosciences. Volume 89(2016)
• Page Start: 32
• Page End: 43
• Publication Date: 2016-04
• Subjects: Hydrothermal alteration -- Immobile elements -- Mineral exploration -- Neural networks -- Multi-layer perceptron -- Mass balance calculations; Environmental policy -- Periodicals; 550.5
• DOI: 10.1016/j.cageo.2016.01.003
• Languages: English
• ISSNs: 0098-3004

Abstract

Abstract: This study proposes an artificial neural networks-based method for predicting the unaltered (precursor) chemical compositions of hydrothermally altered volcanic rock. The method aims at predicting precursor's major components contents (SiO2, FeO T, MgO, CaO, Na2 O, and K2 O). The prediction is based on ratios of elements generally immobile during alteration processes; i.e. Zr, TiO2, Al2 O3, Y, Nb, Th, and Cr, which are provided as inputs to the neural networks. Multi-layer perceptron neural networks were trained on a large dataset of least-altered volcanic rock samples that document a wide range of volcanic rock types, tectonic settings and ages. The precursors thus predicted are then used to perform mass balance calculations. Various statistics were calculated to validate the predictions of precursors' major components, which indicate that, overall, the predictions are precise and accurate. For example, rank-based correlation coefficients were calculated to compare predicted and analysed values from a least-altered test dataset that had not been used to train the networks. Coefficients over 0.87 were obtained for all components, except for Na2 O (0.77), indicating that predictions for alkali might be less performant. Also, predictions are performant for most volcanic rock compositions, except for ultra-K rocks. The proposed method provides an easy and rapid solution to the often difficult task of determining appropriate volcanic precursor compositions to rocks modified by hydrothermal alteration. It is intended for large volcanic rock databases and is most useful, for example, to mineral exploration performed in complex or poorly known volcanic settings. The method is implemented as a simple C++ console program. Highlights: Method for predicting the fresh precursor to hydrothermally altered volcanic rocks. Precursors are predicted using multi-layer perceptron neural networks. Precursor compositions (major oxides) are predicted from immobile elements ratios. Chemical method for quantifying alteration from large datasets of volcanic rocks. New chemical method developed for grassroot or regional mineral exploration.