Applicability of Near-Infrared (NIR) spectroscopy for sensor based sorting of a porphyry copper ore. OCM conference presentation 07-03-2013 TU Delft, Resource Engineering section M.Dalm MSc. 1
Table of contents Introduction Test work methods Analysis of samples Regression analysis Applicability of NIR spectroscopy for sorting Economic feasibility analysis Conclusions Slide 2 of 26 2
Introduction Sensor based sorting Sensor based sorting represents all applications where singular particles are mechanically separated on certain physical properties after determining these properties by a sensor. Applicability to processing porphyry Cu ore: CPU Can be implemented as a pre-concentration step to remove sub-economic material prior to the conventional concentrating methods. Conventional concentrating methods require mineral liberation processes. Potential to reduce ore processing costs. Slide 3 of 26 3
Introduction Sensor based sorting The feasibility of sensor based sorting is completely dependent on an appropriate sensor that can distinguish between economic and sub-economic material. No sensor is known that allows a direct detection of Cu grades for porphyry copper ores. Why test the Near-Infrared sensor? Near-Infrared spectroscopy is proven to be a valuable tool in mapping distributions of mineral alterations in hydrothermal ore deposits. The value in this results from the geology of porphyry Cu systems. Slide 4 of 26 4
Introduction Geology of porphyry Cu systems A porphyry system results from intrusion of oxidised magma, saturated with sulphide and metal rich aqueous fluids. Hydrothermal alteration results in the addition, removal or redistribution of preexisting rock components. The hydrothermal alteration zones are conformable to the ore mineralisation zones! Slide 5 of 26 5
Introduction Aim of the test work Aim of the test work Define the possibilities and limitations of a Near-Infrared sensor to distinguish between Cu grades of a porphyry Cu ore. Slide 6 of 26 6
Test work methods Dataset 150 samples were collected from the oversize output of milling operations at a South American mining operation that exploits a porphyry Cu deposit. The samples were cut in halves to provide a clean surface for the Near-Infrared spectroscopic measurements. One half of each sample could be used for petrographic & geochemical analysis (XRF & XRD). 7 Slide 7 of 26
Test work methods Classification on visual interpretation A B C F E M H G N O D1 I D2 J2 K3 K1 L1 J1 K2 L2 Rest 8 Slide 8 of 26
Test work methods Sample selection for geochemical & petrographic analysis A B C F E M H G O N D1 I D2 J2 K3 K1 L1 J1 K2 L2 Rest XRD & XRF XRF Petrography 30 samples 35 samples 30 samples 9 Slide 9 of 26
Test work methods Near-Infrared spectroscopic measurements Near-Infrared reflectance spectra were measured with an ASD Fieldspec3 portable spectroradiometer. Locations of measurements were documented so that NIR spectra could be compared with the visible characteristics of the samples. Slide 10 of 26 10
Analysis of samples Near-Infrared spectroscopic measurements The measured NIR reflectance spectra were analysed by comparing them to characteristic mineral spectra from spectral libraries. The NIR spectra showed absorptions by 1, 2 or 3 different NIR active minerals. In all the measured spectra, absorptions by muscovite and/or illite are dominant. Mica NIR spectra Illite, Low crystallinity Mica mixed NIR spectra Tourmaline Muscovite and illite are both mica minerals that mainly differ in crystallinity. Muscovite, High crystallinity Chlorite Other minerals showing absorptions in the NIR spectra include tourmaline, chlorite and kaolinite. Kaolinite Slide 11 of 26 11
Analysis of samples Near-Infrared spectroscopic measurements The following Cu-bearing minerals were determined from the sample set by the petrographic analysis; Chalcopyrite (CuFeS 2 ) Chalcocite (Cu 2 S) Fe Fe Chalcopyrite a Covellite (CuS) Digenite (Cu 9 S 5 ) Chalcocite b Bornite (Cu 5 FeS 4 ) Covellite a These minerals are all sulphides that do not produce strong deterministic absorptions in the Near-Infrared. Fe Bornite b Fe ion absorptions may be observed, but at the samples these can also be caused by other Fe-bearing minerals such as pyrite, hematite or magnetite. a USGS spectral library by Clark et al. (2007) b JPL spectral library by Grove et al. (1992) Slide 12 of 26 12
Analysis of samples Classification on mineralogy Samples were classified based on all gathered information on the mineralogy: Mineral determinations from NIR reflectance spectra X-ray diffraction Petrography Classification on visual interpretation Slide 13 of 26 13
Analysis of samples Classification on mineralogy A Comp2 E B F C D2 D1 Comp3 Comp1 G O Comp4 Comp5 M N H I J2 K3 K1 L1 J1 K2 L2 Rest Comp6 14 Slide 14 of 26
Cu grade (%) Analysis of samples Classification on mineralogy Analysis of the Cu grades of the groups resulting from classification on the mineralogy showed interesting results. 3.0 2.5 2.0 All samples with NIR spectra that show absorptions by illite & chlorite. 1.5 1.0 An economic cut-off grade of 0.4% Cu was assumed for further processing of the samples. 0.5 0.0 0 A 1 B 2 C 3 D 4 E 5 F 6 7 (68) (45) (14) (7) (9) (7) Group from comprehensive classification Slide 15 of 26 15
Regression analysis Introduction Statistical technique used to analyse the relationships between NIR response and Cu grade of the samples. Reason: The NIR reflectance spectra may contain additional features relating to the Cu grade that were not discovered from visual interpretation of the spectra. Types of analysis performed: Partial least squares regression (PLSR) Multivariate logistic regression (MLR) Slide 16 of 26 16
Regression analysis Partial least squares regression - Introduction Used to predict Cu grades from the NIR reflection spectra. Multivariate linear regression model: ˆ 0 p xp y PLSR models the structure of the dataset by calculating the principle components of the predictor matrix. As predictors, the reflection values at all measured wavelengths from the NIR reflection spectra were used. Slide 17 of 26 17
Regression analysis Partial least squares regression - Results The dataset was divided into a calibration set (2/3 of all samples) and a validation set (1/3 of all samples). No linear relation between any NIR spectral variable and the Cu grade. No decrease in MSE for the Validation set Slide 18 of 26 18
Regression analysis Multivariate logistic regression - Introduction Similar to linear regression, but handles outcome variables that are binary. Multivariate logistic regression model: gp ( x ) e Pn ( xn, p ), g ( ) 1 p x e g ( x ) x x p n, p 0 1 n,1 p n, p Used to estimate the probability that a sample is below or above the assumed economic cut-off grade of 0.4% Cu. As predictors, a set of calculated properties of several characteristic mineral absorptions were used. Slide 19 of 26 19
Regression analysis Multivariate logistic regression Calculation of predictors Selection of calculated absorption feature properties was based on the characteristic absorptions of the determined NIR active mineralogy. Calculated properties include: Influence of secondary mineral absorptions Absorption minimum locations. Absorption depths. Impact of secondary absorptions on dominant absorption slopes. Fe ion absorption. Fe ion absorption Absorption depth Minimum location Slide 20 of 26 20
Regression analysis Multivariate logistic regression - Results Results of logistic regression at an economic cut-off grade of 0.4% Cu. The dataset was divided into a calibration set (2/3 of all samples) and a validation set (1/3 of all samples). Waste 1 & Ore 1 samples: All samples showing absorptions by illite and chlorite in the NIR reflection spectra. Waste 2 & Ore 2 samples: All other samples. Slide 21 of 26 21
Regression analysis Multivariate logistic regression - Results Results of logistic regression at an economic cut-off grade of 0.4% Cu. Waste 1 & Ore 1 samples: All samples showing absorptions by illite and chlorite in the NIR reflection spectra. Ore 2 samples: All other samples. Slide 22 of 26 22
Applicability of NIR spectroscopy for sorting Introduction Based on the results of logistic regression a sensor based sorting solution was evaluated. Slide 23 of 26 23
Applicability of NIR spectroscopy for sorting Economic evaluation of a sensor based sorting solution The total costs of a NIR sorter that is 24/7 in operation are around 1.24 $/t feed material c. 19% of material rejected -> 6.42 $/t reject material. @ Assumed cut-off grade = 0.4% Cu, Cu price = 8000 $/t, Assumed recovery = 85% -> Processing costs = 27.20 $/t Grade reject material = 0.18% Cu -> Loss saleable Cu = 12.24 $/t Estimated profit = 27.20 $/t 6.42 $/t 12.24 $/t = 8.54 $/t (reject material) @ Typical production of a large porphyry Cu mine: 100 000 (t/d) x 0.19 x 8.54 = 164 848 $/d Uncertainties: Representation of test work results for the entire ore deposit. Degree of surface contamination of ore particles feeding the sorter. Influence of water on the discrimination capabilities of the NIR sensor. Sorting efficiency of sensor based sorting equipment. c Bergmann, J.M., Sensor-based sorting. Industrial minerals, 2011, July. Slide 24 of 26 24
Conclusions (1/2) A relation is present between the Cu grade and the hydrothermal alteration mineralogy for the 150 mill pebble samples. Predicting a Cu grade from the NIR response of the 150 ore samples can only be based on certain NIR active mineral assemblages that represent a specific type of hydrothermal alteration, constituting a certain range of Cu grades. The only Cu grade discrimination that can be made on the set of 150 ore samples by NIR spectroscopy is to separate the lower grade samples that contain illite and chlorite as the dominant NIR active mineralogy. Slide 25 of 26 25
Conclusions (2/2) If sensor based sorting can be applied with the results from the performed test work, there is good potential that such an application can have significant economic benefits. Uncertainties are; Representation of test work results for the entire ore deposit. Degree of surface contamination of the ore particles feeding the sorter. Influence of water on the discrimination capabilities of the NIR sensor. Sorting efficiency of sensor based sorting equipment. Logistic regression is a good method to create algorithms in order to classify the ore material on a specific NIR active mineralogy that represents a certain range of Cu grades. Slide 26 of 26 26
Applicability of Near-Infrared (NIR) spectroscopy for sensor based sorting of a porphyry copper ore. OCM conference presentation 07-03-2013 TU Delft Resource Engineering section M.Dalm MSc. 27
Recommendations For further studies on ore characterisation with NIR spectroscopy Use a sample set with a known origin of samples relative to the geology of the ore deposit. Better investigation on the relations between hydrothermal alteration, Cu mineralisation and NIR active mineralogy. Better determination on the significance of the outcomes for possible sensor based sorting solutions. NIR hyperspectral imaging might have great advantages over NIR point spectra. More detailed information about the NIR active mineralogy of samples. Slide 28 of 26 28
Introduction Geology of porphyry Cu systems Relations between hydrothermal alteration zones and ore mineralisation zones can be complicated in practice; Telescoping of the porphyry system. Superimposition of breccia & diatreme intrusions. Slide 29 of 26 29
Analysis of samples Classification on mineralogy - Results Slide 30 of 26 30
Regression analysis Logistic regression Calculation of predictors 0.0025 0.0035 0.0075 0.0070 0.0020 0.0030 0.0065 0.0060 0.0025 0.0055 0.0015 0.0020 0.0050 0.0045 0.0040 0.0010 0.0015 0.0035 0.0030 0.0005 0.0010 0.0025 0.0020 0.0005 0.0015 0.0000 0.0000 0.0010 0.0005 0.0000-0.0005-0.0005-0.0005 Slide 31 of 26 31
Regression analysis Multivariate logistic regression - Results Results of logistic regression at an economic cut-off grade of 0.4% Cu. 2 types of p-values for testing the significance of predictors: p1: t-test statistic p2: model deviance to χ 2 distribution p-value < 0.05 was considered significant. Slide 32 of 26 32
Regression analysis Logistic regression Calculation of predictors Width at half depth (nm) Depth (-) Minimum location (nm) 1350 1375 1400 1425 1450 1475 Wavelength (nm) Slide 33 of 26 33