Mineralogical Analysis of Solid Sample Flame Emission Spectra by Machine Learning

Abstract

Solid pre-concentrated ore samples used in pyrometallurgical copper smelters are analyzed by flame emission spectroscopy using a specialized flame OES system. Over 8500 complex spectra are categorized using an artificial neural network, ANN, that was optimized to have ten hidden layers with 40 nodes per layer. The ANN was able to quantify the elemental content of all samples to within better than 1.5% w/w, and was able to identify the prevalent minerals to within better than 2.5%w/w. The flame temperature was obtained with an uncertainty of 3 K and the particle sizes to within 2 m. The results are found to be superior to those obtained to a non-linear partial least squares fit model, which is equivalent to an ANN having no hidden layers.