**1. Introduction**

The mineralogical characterization of ore deposits using scanning electron microscopy (SEM)-based automated mineralogy systems such as mineral liberation analyzer (MLA) or QEMSCAN® are of major importance for mineral processing. Declining ore grades, the reconnaissance of new targets, and demands for improved energy efficiency require fast, reliable, and high-quality information on mineral resources. Highly variable and complex mineralogy affects mineral processing and extractive metallurgy, as has been shown by authors in the past [1–6]. Collecting quantitative data of mineral distributions, associations and textures are one of its key abilities. Using this technique, information gained by traditional bulk mineralogical analysis, chemical assay, and optical microscopy can be supported and expanded. Automated mineralogy is widely used in the mining business as it allows rapid identification and characterization of key minerals in ore and rock samples. Furthermore, it can be applied to all sorts of different mineral deposit types [1,7–11]. Mineral liberation analysis (MLA) software is one of the widely applied SEM-based applications offering a rapid quantitative characterization of mineral species and their relations in polished samples [12]. It is commonly used in conjunction with other micro-compositional or microstructural techniques such as electron-probe microanalyses (EPMA) or X-ray diffraction (XRD). However, SEM-based techniques cannot precisely distinguish mineral

polymorphs (e.g., rutile-anatase-brookite) or discriminate between minerals having a similar elemental composition (e.g., magnetite and hematite). Especially the discrimination of magnetite and hematite becomes an important factor in future processing stages at the Luossavaara-Kiirunavaara Aktiebolag (LKAB) enterprise. Almost pure magnetite mineralization is only prevalent in the Kiirunavaara deposit. Increasing hematite contents in the other deposits in the Kiruna district (Malmberget and Leveäniemi), and most predominantly in Per Geijer will significantly influence the beneficiation and pelletizing process. In order to sustain e fficient comminution and magnetic separation processes evaluation by automated mineralogy or similar methods will add significant value in the future.

The e ffect that similar average atomic numbers result in similar BSE intensities (BSE grey values) requires modification of the MLA technique, especially in the data processing stage. A technique for distinguishing between hematite and magnetite by MLA has been established by Figueroa et al. [13]. However, stable beam currents and even sample surfaces are obligatory for maintaining constant BSE grey values. As this often proves problematic, there is demand for additional analytic solutions.

In this study, Raman imaging is used as an alternative tool to highlight advantages and disadvantages of both methods with the focus on magnetite-hematite mineralization of the Per Geijer iron ore deposits close to the well-known Kiirunavaara deposit. This comparative approach for ore characterization provides detailed information by mapping mineral distributions of a potential new iron ore resource.

#### **2. Geology of the Per Geijer Deposits**

The iron oxide apatite ores (IOA) in the Kiruna area are hosted by the Svecofennian Kiirunavaara Group. Svecofennian rocks are represented by volcanic and sedimentary units generated by reworking of older crust and subduction and accretion of several volcanic arc complexes [14–17]. Three formations account for that group: (1) Hopukka Formation, (2) Luossavaara Formation, and (3) Matojärvi Formation. The giant Kiirunavaara ore body with pre-mining resources of more than 2000 Mt [18], is a massive tabular almost purely magnetite body with generally low phosphorus content. The smaller, magnetite-rich Luossavaara deposit north of Kiirunavaara is situated in the same stratigraphy. Closely associated with the east and northeast reside the Per Geijer orebodies at the stratigraphically upper contact of the Luossavaara Formation or within the lower part of the Matojärvi Formation [19]. These deposits are located north of the town of Kiruna and consist of five orebodies that have been mined in intervals during the 20th century as open pits. The Nukutus, Henry, and Rektorn mineralization occur stratigraphically at the upper contact of the Luossavaara Formation, while the Haukivaara and Lappmalmen deposits, at least partly, are located within the overlying Matojärvi Formation. Lappmalmen is a blind and, so far, unexploited ore body only known from exploration drilling. The ores exhibit large variations in texture, mineral composition, and relation to wall rocks [18]. Depending on the orebody and the vertical position of the mineralization in the deposit, magnetite and hematite occur in varying textures and proportions leading to di fferent ore types (e.g., hematite-dominated, magnetite-dominated, magnetite/hematite-mixed), as proposed by [20]. Generally, hematite occurs together with magnetite often replacing it, especially close to the upper part of the deposit. The overall phosphorus content is high and occurs as apatite. Other main gangue minerals are carbonate and quartz. The Per Geijer deposits have an average composition of 40–50% Fe and 3–5% P but higher and lower iron and phosphorus contents can locally occur.

#### **3. Materials and Methods**

#### *3.1. Samples and Preparation*

Sixty-five samples were selected from drill cores representing all five Per Geijer ore bodies after specific pre-defined ore type sections and macroscopic petrographic evaluation. For this study, 18 polished thin sections were prepared by the preparation laboratory of the Helmholtz Institute Freiberg for Resource Technology. Sample HA-81155-2 was selected for comparative SEM-MLA and

Raman measurements because of the occurrence of magnetite, hematite, and Ti-bearing hematite in satisfying abundances.

#### *3.2. Scanning Electron Microscope (SEM) and Mineral Liberation Analyzer (MLA)*

Samples were studied with the scanning electron microscope (SEM) including mineral liberation analyzer (MLA) (FEI, Brisbane, Australia) software on a Quanta 650 FEG-MLA650F (FEI©, Brno, Czech Republic). The analyses were carried out at the geometallurgical laboratory of the Department of Mineralogy at TU Bergakademie Freiberg, Germany. Energy-dispersive X-ray spectroscopy (EDS) was performed with two Bruker Nano Dual X-Flash 5030 detectors (Bruker, Berlin, Germany), using an accelerating voltage of 25 kV and a working distance of 12 mm. The workflow of the SEM-MLA analysis described in the following sections is shown in Figure 1.

**Figure 1.** Workflow of the SEM-MLA and Raman analysis.

#### 3.2.1. Measurement Mode and Setup

The grain-based X-ray mapping (GXMAP) mode was selected for this study. The GXMAP technique analyses selected grains (in this case, the whole thin section) with a closely spaced grid of X-ray measurements resulting in false-colored mineral maps [12,21]. Here, the whole sample was continuously mapped with an EDS analysis every 12 μm (slower mapping) or 36 μm (faster mapping). In both measurements, the frame resolution was 500 × 500 pixel with a magnification of either 3 or 2 μm/pixel, respectively. Based on backscattered images, the minerals were first separated according to their grey values and then mapped in a definite grid of EDS analysis. The basic setup included the definition of a BSE grey value range (BSE trigger). As element contents are responsible for the BSE contrast, chemically similar minerals have identical BSE grey levels. Thus, for hematite and magnetite classification, reference BSE values had to be manually defined. Furthermore, they were compared with optical reflected light images of the sample in which the two iron oxides have characteristic reflection properties. The details of the SEM image acquisition are listed in Table 1. A detailed description of the functionality of MLA and the measuring modes can be found in the literature [12,22–24].


**Table 1.** Details of the scanning electron microscope (SEM) + mineral liberation analyzer (MLA) and Raman measurements.

> (s)—slow measurement, (f)—fast measurement, (ts)—thin section.

#### 3.2.2. Mineral Reference List

A preliminary mineral reference list was generated by collecting mineral standards manually. Chemical compositions were obtained with the Bruker Esprit 1.9 software (Bruker, Berlin, Germany), mineral names, and formulae, as well as physical properties (e.g., density), were added.

#### 3.2.3. Online Data Processing

The preliminary mineral list had to be adjusted after first classification due to larger unknowns. Unknowns are detected when the comparison between the spectra of the mineral reference list and the measured X-ray spectra fails. This error in consistency occurs when the limits of a spectrum matching threshold (80%) and/or a low count limit (800) exceed the predetermined values. Thus, an assignation to a suitable mineral phase was not possible. However, some unknowns can be the result of mixed spectra of two different phases.

#### 3.2.4. Offline Data Processing

For off-line image and data processing collected X-ray spectra were classified. As for online data processing, comparison with the spectra of the mineral reference list is necessary. Therefore, the spectrum matching threshold and low count limit were set to 80% and 800 counts, respectively. However, the X-ray spectra and the BSE grey level of the different iron oxides (e.g., hematite and magnetite) are very similar. Consequently, the results from MLA were checked against the results from optical microscopy and manual correction was applied. For most samples, BSE grey value ranges were set at 150 to 190 for hematite and 191 to 250 for magnetite. These values were checked for every sample after classification and adjusted, if necessary. However, single frames in some samples yielded wrong classification of magnetite and hematite based on local BSE grey value variations. Manual correction was obtained by using the edit selected particles mode. By re-defining wrongly classified areas manually, the validity for the samples was improved.

To remove small unknowns and other noise from the image, the touch-up tool was used. All unknowns with a size smaller than 50 pixels were turned into the containing host (unknown to any host), which mainly affected pixels with mixed spectra. The remaining unknowns generated by holes and brackets of the thin section holder were turned into background (size < 2,000,000 pixels). Minerals, which feature a pixel size smaller than 50 or 70 that are located in another mineral, were transformed to the containing host mineral (mineral to any host). Finally, databases were created comprising information about modal mineralogy, mineral associations, as well as grain properties. For this study, only information of the modal mineral abundance was taken into account.

#### *3.3. Raman Spectroscopy and Imaging*

Raman maps were collected using an alpha300 M+ Raman Microscope System (WITec GmbH, Ulm, Germany) or an inVia Qontor Raman Confocal Microscope (Renishaw, Pliezhausen, Germany). The workflow of the Raman analysis described in the following sections is shown in Figure 1.

#### 3.3.1. Measurement Mode and Setup

The measurements were carried out at Papiertechnische Stiftung (PTS) Heidenau and Center for Molecular Bioengineering at TU Dresden (BCUBE). Both systems were equipped with an upright microscope using a 532 nm laser excitation and a Charge Coupled Device (CCD) detector with 1024 pixel. The Renishaw system was used with a 100× objective (Leica Objective N plan EPI 100×/NA 0.85) using 10% laser power (corresponding to approximately 5 mW) and a 1800 gr/mm grating (spectral range: 60–1837 cm<sup>−</sup>1). The Raman map of Figure 2 (2.127 cm × 2.232 cm) was collected with 30 μm step size and 275 ms exposure time using the Renishaw system.

**Figure 2.** Mineral distribution maps from reflected light (optical microscopy) and processed SEM-MLA and Raman analysis.

The WITec confocal Raman system is based on a Zeiss microscope and was used with a 100× objective (Zeiss EC Epiplan NA 0.9) using 20 mW laser power and a 600 gr/mm grating (spectral range: −123–3777 cm<sup>−</sup>1). The Raman maps of Figure 3 were collected using the WITec system. The overview map (2.38 cm × 1.3 mm) was collected with 12 μm step size, the detailed map with 1 μm step size (160 μm × 120 μm). In both cases, spectra were collected using 200 ms exposure time.

Before measurement, we optimized spectral acquisition parameters (especially laser power). Due to the large mapping areas (up to 2 cm × 2 cm), we decided to use the highest laser power while reducing exposure time as much as possible to achieve a high mapping speed. We started with point measurements on all minerals (acquisition time: 0.2 s) using increasing laser power to check for thermal decomposition or oxidation. This way, the maximum laser power for each mineral can be determined. In our samples, sulfides were found to be the most sensitive to laser damage resulting in significant visible and spectroscopical changes. The most sensitive minerals determined the laser power that was used. We did not observe any changes in the spectra of magnetite and hematite. In comparison, the literature reporting on hematite transformation [25], used exposure times from at least several seconds

up to several minutes (10 s–5 min). Samples were checked for oxidation effects after measurements with optical reflected light microscopy.

**Figure 3.** Reflected light, backscattered electron (BSE), false-colored MLA, and Raman images from the same area of sample HA-81155-2. Detailed comparison of magnetite, Ti-bearing hematite, and Ti-poor hematite is shown in zoomed-in maps.

#### 3.3.2. Mineral Reference List

Reference spectra were collected either by spot measurements, isolated by non-negative matrix factorization (NMF), or extracted by averaging mineral-specific areas of Raman maps based on the EDS maps obtained for MLA. Before linear combination, the spectra were background-corrected and normalized. All reference spectra can be found in the Supporting Information (Figure S1).

#### 3.3.3. Offline Data Processing-Data Preprocessing

Spectral datasets were imported into the in-house SpectralImaging software (TU Dresden, Dresden, Germany) on the basis of Matlab provided and developed by Matthias Finger (spectralimaging@outlook.com). First, cosmic ray removal was performed to eliminate spikes. Spectral datasets were typically cut to a wavelength of 100–1900 cm<sup>−</sup><sup>1</sup> where significant spectral differences

were found. Baseline correction was performed by asymmetric least squares correction (smoothness 5, asymmetry 2). If necessary, smoothing (Savitzky-Golay, seven points, degree 2) was applied.

#### 3.3.4. O ffline Data Processing-Multivariate Analysis

Multivariate analysis was used to identify the number of principal components and to image their distribution. We used two data analyses methods, principal component analysis (PCA) and non-negative matrix factorization (NMF). Both techniques were conducted using the above-mentioned in-house SpectralImaging software based on Matlab.

PCA and NMF are both unsupervised statistical methods to reduce the dimensionality of large datasets. Using PCA, original variables of the dataset are summarized in new variables, the so-called principal components. These components are calculated as the solution of an eigenvalue problem responsible for the whole variance of the data set. As a result, PCA weights a set of principal components with positive and negative values to represent the full spectral information.

NMF is similar to PCA but the weights and factors are constrained to be positive. Therefore, it is well-suited to represent data with non-negative features. We found that the resulting spectral factors are easier to inspect. That is why, if necessary, non-negative matrix factorization (NMF) was used to extract spectral signatures of individual minerals from the total set. To validate these unsupervised methods, a supervised linear combination approach was used to map the distribution of selected components. All maps shown were prepared using linear combination of a set of measured reference spectra (see above).
