NIR-Spectroscopy and Machine Learning Models to Pre-concentrate Copper Hosted Within Sedimentary Rocks

Abstract

The western part of the Moroccan Anti-Atlas comprises numerous copper occurrences hosted within various sedimentary rocks, all containing low-grade copper concentrations. This study aims to assess the feasibility of using a near-infrared (NIR) sorting system to efficiently process these low-grade resources. In essence, it involves evaluating the potential of short-wave infrared (SWIR) spectroscopy and machine learning models to classify ore fragments into waste or concentrate based on their SWIR spectral characteristics. In order to conduct this study, the SWIR reflectance of 475 rock samples from the Tizert deposit was measured. Mineralogical analysis was performed, using X-ray diffraction and scanning electron microscopy, to understand the mineralogy of the samples and its relationship to SWIR spectra. Chemical analysis was also performed to categorize samples based on their copper content. Several machine learning models, including partial least squares discriminant analysis (PLS-DA), random forest (RF), and support vector machine (SVM) were evaluated based on both lithology and copper content characteristics. Among these, PLS-DA yielded the most favorable results, achieving an 84% accuracy in lithologies classification and 90% accuracy in classifying samples based on their copper content, utilizing a 0.2% cutoff grade. This laboratory-scale study validates the effectiveness of SWIR spectroscopy as a prominent tool for pre-concentrating sedimentary copper deposits. It enables the production of a concentrate with a copper content of 1.49% and waste with 0.12%, resulting in an upgrading rate of 43% from the feed, which originally has a copper grade of 1.04%.