The polished surface of each of the four epoxy-mounted rock chips was scanned by AM using a Zeiss Sigma 300 VP with two EDS detectors and the Mineralogic software at Luleå University of Technology (Table 1). A pixel size (beam sampling distance) of 30 μm was chosen; a trade-off between resolution and time spent on analysis. Two regions of interest (4 mm diameter) in the rock chips of samples 2 and 3, subsequently drilled for extraction of the mini cores, were additionally scanned at 1 μm pixel size (Table 1). The same custom-made mineral library in the Mineralogic software was used for initial mineral classification in all scans. Element filters for Au, Ag, and Bi were placed at the top of the mineral library to identify sub-pixel sized grains of trace metal minerals recorded as mixed phase spectra (Warlo et al., 2019). Results were processed in the Mineralogic Explorer software. Here, mineral maps were stitched and montaged and quantitative data on bulk mineralogy, mineral association, metal deportment, were extracted. The accuracy and precision of the mineral library was assessed by manual spot-checking the BSE image and respective EDS spectra of the identified minerals. For Au and Ag minerals, every pixel found was reviewed. Based on this assessment, the mineral library was then fine-tuned for each scan and data reclassified.

For comparison with XCT, the detailed mineral maps were subsequently simplified based on the limiting number of segmented classes in the XCT scans. The Mineral X-Ray Linear Attenuation Coefficient Tool (MXLAC) by Bam et al. (2020) was used to calculate the mineral x-ray linear attenuation coefficient of each mineral for the respective beam energy used in the XCT analysis. All minerals identified by AM where ranked accordingly. For certain mineral groups (e.g., amphibole, epidote), the attenuation coefficient of only one mineral from the group was calculated to represent the whole group. For broader-defined groups (e.g., silicates, Fe-oxides/carbonates), no values were calculated but the approximate position in the ranking estimated. The coefficients were then used in conjunction with visual comparisons of AM and XCT images to assign the minerals to the classes segmented in the XCT scans (Table 2).

Since Mineralogic Explorer does not offer an option for analysis of subregions within a scan, the simplified quantitative maps were converted into grey scale images, imported into ORS Dragonfly, and re-segmented based on their histograms. This allowed overlay of the AM data with the XCT data and quantitative analysis of the same region. For the epoxy-mounted rock chip scans, a circular/cylindrical (AM/XCT) region was chosen, which maximized area/volume while excluding the border between rock and epoxy. For the high-resolution scans, an area/volume that maximized the overlap between the 1 μm pixel size mineral maps and the XCT scans of the mini cores was chosen.

The AM data of the 1 μm pixel size scan of sample 2 became corrupted after reclassification with the updated mineral library. Results were still accessible but further reclassification was no longer possible. This caused issues with simplification of the mineralogy for comparison with XCT. Some minerals initially colored the same could no longer be separated. This was the case for Fe-oxides (ilmenite, magnetite, hematite, FeO, Fe-oxides/carbonates) and tin minerals (cassiterite, tin minerals). They were therefore treated together according to the mineral of highest concentration (Fe-oxides/carbonates and tin minerals, respectively). Quantitatively, the miss-classified Fe-oxides covered ∼0.2 area % and cassiterite <0.01 area % in the whole mineral map. Since no equivalent quantitative data could be inferred for the region compared with XCT, no correction was performed. Regardless, impact on the quantitative comparison was minimal.

The half drill core piece was scanned by XCT at 45 μm voxel size using a GE phoenix v|tome|x s at the Geological Survey of Finland, Espoo, Finland (Table 1). The four epoxy-mounted rock chips were scanned at 30 μm voxel size using a Zeiss Xradia 510 Versa at Luleå University of Technology (Table 1). Two rock chips each (samples 1 and 2, and 3 and 4) were stacked with their polished surfaces face to face, fixed in place with tape, and analyzed together (see Supplementary Figure S1). This was to minimize the cone-beam effect on the polished surfaces of the samples. The two mini drill cores were also analyzed with the Zeiss Xradia 510 Versa at 4 μm voxel size. In addition, a cylindrical volume of interest of 1 mm in both diameter and height within each 4 mm drill core was scanned at 1 μm voxel size. The Zeiss Xradia system has multiple detector objectives enabling sample imaging at a number of resolution and field of view pairings. Here, the rock chip samples were scanned using a ×0.4-objective, while the mini drill core samples were scanned using a ×4-objective, specified more in detail in Table 1.

The tomographic reconstruction was carried out using filtered back projection, and beam hardening correction applied where necessary. However, due to the complex mineralogy within the samples (large range of x-ray linear attenuation coefficients) optimization for all minerals was not possible. Hence, corrections and filters commonly were applied in favor of the higher x-ray linear attenuation coefficients (denser minerals) as these were in focus in this study.

The XCT data were processed and analyzed using the software ORS Dragonfly (version 2021.1). Images of the raw XCT data are shown in Supplementary Figures S1,S2. Image processing was performed on each data set prior to mineral segmentation. The ‘Unsharp’ filter was applied to increase edge contrast of grains, followed by the ‘Mean Shift’ filter to smoothen the noise introduced by the ‘Unsharp’ filter. Segmentation was performed on a cylindrical subvolume in each data set, so that: 1) the epoxy mount was excluded 2) depth-related grey scale gradients (e.g., cone-beam effect), did not affect segmentation significantly, 3) each horizontal slice was of equal area, and 4) the top slice overlapped with the respective mineral map from AM. The only exception was the half drill core piece where segmentation was performed on a half cylinder partly including the contact with the surrounding air. The ‘Histographic Segmentation’ plug-in in the ORS Dragonfly software was used for mineral segmentation. This plug-in includes a combination of grey scale thresholding and the sobel filter to define classes (groups of minerals of similar x-ray linear attenuation coefficient) and determine markers in the sobel image for the watershed algorithm (Guntoro et al., 2019a). Individual grains were determined based on a 6-point particle connection (shared voxel faces). Heavy trace minerals in the samples of mineralization style 2 were of too low volume to be identified accurately by ‘Histographic Segmentation’. Hence, the trace minerals were segmented by manual grey scale thresholding. The ‘Histographic Segmentation’ (and grey scale thresholding) were performed six times for each sample to get an estimate of the standard deviation for the various segmented phases.

The segmented XCT data were used for quantitative analysis. The mean bulk volumes of the segmented mineral classes and their respective area percentage in each horizontal slice were calculated for the six segmentations of each scan. The mean data of the slices overlapping with the mineral maps were used for quantitative comparison between XCT and AM. In addition, the grey scale and segmented XCT images were visually inspected concerning texture and structure of the ore samples.

The AM scan shows the contact between biotite schist and an intruding aplite dyke (Figure 2). The aplite dyke contains several up to ∼4.5 mm sized scheelite grains together with chalcopyrite and pyrrhotite. Several sub-millimeter inclusions of Bi-tellurides occur in the pyrrhotite. Some ambiguity on pixel classification exists in very fine-grained regions, e.g., the biotite schist, due to mixed EDS spectra. Overall, mineral classification and resolution of textures agree with previous observations from reflected light microscopy and SEM-EDS. In the XCT scan only four classes (mineral groups) can be visually separated, 1) silicates and carbonates, 2) sulfides and oxides, 3) scheelite, and 4) heavy minerals (mostly Bi-tellurides; Figure 2, Table 3). Comparison of the top slice of the segmented XCT data with the same area in the AM mineral map shows considerable quantitative differences (Table 3). Although total difference in area between the techniques are largest for the silicates/carbonates (∼11%), the relative difference is much larger for the heavy minerals (∼70%) due to their overall low abundance (Table 3). The sulphides/oxides are high in both total and relative area difference. The differences can be explained mainly by the effect of streak artefacts in the XCT data, caused by the scheelite and heavy mineral grains, on the histographic segmentation. This results in spurious aureoles of sulfides/oxides around scheelite and heavy mineral grains (Figures 2, 3). Additionally, some scheelite grains have brighter rims than cores, which causes segmentation of the rims into the class of heavy minerals (Figures 2, 3). Segmentation of regions without scheelite grains appear satisfactory. Texturally, the unaided eye and AM indicate a preferred planar orientation of scheelite and sulfides within the quartz vein (Figure 2). XCT confirms this and shows continuity of the scheelite vein deep into the sample. Scheelite grains are relatively large (∼7 mm) and elongated towards depth but strongly fractured (Figure 3). Heavy minerals mainly form inclusions within the sulfides. The XCT images clearly display the dissemination of heavy mineral inclusions and the problem with grain size estimations based on 2D sections. The largest Bi-telluride grain observed on the polished surface by AM (∼0.5 mm) is intersected at an edge and grows considerably larger with depth in the sample (∼2.5 mm) (Figure 3). This also exceeds previously reported maximal grain sizes of Bi-minerals (0.5 mm) at the Liikavaara Östra Cu-(W-Au) deposit (Warlo et al., 2020).

The XCT scan of the half drill core shows a complex intergrowth of sulfide and gangue minerals (Figure 4). The undifferentiated sulfides mainly occur to one side of the core. The gangue can be differentiated into a dark grey fine-grained matrix replacing larger fractured grains of even darker grey. Very bright heavy mineral grains occur scattered within the gangue. These bright phases are likely galena and/or scheelite, based on the relatively large grain sizes compared with other heavy minerals in the Liikavaara Östra Cu-(W-Au) deposit (Warlo et al., 2020). To target the Au mineralization, the rock chip was taken from the sulfide-dominated part of the drill core. Although the XCT scan of the half core shows only few bright phases in this region (Figure 4), this is to be expected considering the generally small grain size of Au minerals (Warlo et al., 2020).

AM of the surface of the rock chip resolves the mineralogy in better detail compared with XCT (Figure 5). A fine-grained matrix of mainly biotite, chlorite, ankerite, quartz, and Fe-oxides/carbonates, with ubiquitous brecciated subhedral tourmaline, replaces large (>15 mm) quartz grains. Sub-to euhedral pyrite crystals grow within and replace massive chalcopyrite, which permeates the fine-grained silicate matrix. Thin calcite-filled cracks occur throughout the rock. Very fine mineral grains of <30 microns of precious metals Au and Ag, and Bi form scattered inclusions mainly in pyrite and quartz. In the XCT data of the rock chip, four mineral classes (similarly to the half drill core scan) are segmented (Figure 5, Table 3): 1) quartz and albite, 2) silicates and carbonates, 3) sulfides and oxides, and 4) heavy minerals. The high beam energy (160 kV) necessary to penetrate the sample sufficiently, does not allow visual separation of pyrite and chalcopyrite due to their overlapping x-ray attenuation coefficient at that energy (Figure 5; Godel 2013; Kyle and Ketcham 2015; Reyes et al., 2017). Similarly, no differentiation between the dense minerals, including Au-, Ag- and Bi-minerals, is possible. On the other hand, quartz and albite are somewhat distinguishable from other silicates and carbonates unlike in sample 1. The fine-grained character of large parts of the sample cause a pronounced partial volume effect and significant grey scale overlap between the classes. This lowers image resolution compared to AM (both have a 30 micron pixel/voxel size), and causes over-segmentation of particularly silicates/carbonates at the expense of quartz/albite and sulfides/oxides (Figure 5, Table 3). Further, fine textures like carbonate-filled cracks are hardly resolved by the segmentation. Concerning detection of heavy trace metal minerals, both techniques are well suited. For AM, simple element filters enable effective detection and classification of trace metal mineral grains even below pixel size due to their unique element composition (Figure 5, Table 3; Warlo et al., 2019). For XCT, the much higher x-ray linear attenuation coefficient of the heavy trace metal minerals compared with the other minerals in the sample allow rapid segmentation via grey scale thresholding, although not between the different heavy trace metal minerals. In comparison, AM is more effective than XCT in the detection of heavy minerals. The area classified as heavy minerals is ∼70% less in the XCT slice compared to the corresponding AM mineral map (Table 3). Yet, due to the benefit of XCT to analyze volumes, the overall number of heavy mineral grains detected by XCT is much larger than for AM (Figure 4). For example, the amount of heavy minerals in the top XCT slice is only 0.5% of the segmented volume of heavy minerals across all slices. To study the effect of voxel size on quantification, the AM mineral map was also compared with the corresponding XCT slice of the half drill core scan (Table 3). The discrepancy of sulfides/oxides is larger than between AM and the 30 micron XCT scan, whereas it is lower for quartz/albite. Notably, no heavy minerals are classified for the half drill core XCT slice (Table 3).

In a small area within the AM mineral map, two pixels are identified as Au: one in pyrite, the other in adjacent quartz. This region of interest was subsequently scanned by AM at 1 μm pixel size (Figures 5, 6). It reveals numerous <30 μm Au grains (and Ag, Bi) within the pyrite and quartz, but also tourmaline, ankerite, and chlorite (Figure 6). Similarly, the XCT scan at 4 μm voxel size of the mini core, drilled through this region of interest, shows many more dense particles not seen in the coarser scans (Figure 4). Yet, the Au grains identified by AM are not resolved. The smaller size of the mini core compared with the rock chip allows XCT analysis at lower beam energy (80 kV), which makes visual separation of chalcopyrite and pyrite possible (Figure 6). Yet, grey levels are still too close for meaningful histographic segmentation. The change in beam energy also causes tourmaline to dominantly be classified together with quartz rather than the group of silicates/carbonates. Table 3 summarizes the quantitative comparison of this region of interest between the 1 μm pixel size AM scan and all other scans. Expectedly, but still noteworthy, the detection of heavy minerals is higher for AM than XCT and increases with scan resolution for both techniques. The detected amounts for silicates/carbonates and quartz/albite vary significantly, partly due to the change in beam energy (160–80 kV) for XCT analysis and in consequence ambiguous classification of some minerals (e.g., tourmaline). Yet again, overall detection of heavy minerals (all slices) is much higher for XCT than AM (Figure 4). For the 4 μm voxel size XCT scan, the portion of the class of heavy minerals in the top XCT slice is only 0.01% of the volume across all slices.

The 1 μm voxel size XCT scan of a cylindrical volume (1 mm height, 1 mm diameter) within the mini core was centered around the largest Au grain in pyrite (identified by AM). This scan resolves the Au grain and many more dense particles with depth, undetected by the 4 μm voxel size scan (Figure 4). Conversely, visual separation between chalcopyrite and pyrite is not possible. The position of the cylinder within the mini core causes the mini core to act as a shield against the x-rays. Hence, a high beam energy (160 kV) is necessary for proper x-ray penetration.

The study of this sample demonstrates how information on the occurrence and distribution of particularly fine-grained minerals, such as heavy trace metal minerals, is highly reliant on the technique and resolution of analysis.

The AM mineral map shows the aplite dyke to consist mainly of K-feldspar, plagioclase, and quartz with minor biotite and muscovite (Figure 7). Sulfides (pyrite, pyrrhotite, and chalcopyrite), tourmaline, and calcite occur disseminated. In the sample, three quartz vein connect at an angle (Figure 7). Molybdenite mineralization occurs in one vein but not in the others. Conversely, scheelite occurs in one of the other veins but not in the molybdenite-filled vein. Chalcopyrite, pyrite, and pyrrhotite are found in all three veins. In the XCT data, differentiation of the aplite dyke is hardly possible due to beam hardening correction optimized for the highly attenuating scheelite and molybdenite rather than the silicates and carbonates (similar to sample 1). Only the disseminated sulfides can be segmented from the silicates. Yet, comparison between the mineral map and the top slice of the XCT data show a considerable under-segmentation of sulfides and over-segmentation of silicates (Figure 7, Table 4). Nevertheless, the veins and their orientation in the sample are clearly visible. XCT data confirms the separation of scheelite and molybdenite into two veins and shows the extent of the mineralization with depth in the sample (Figure 8). XCT also shows that the veins are dipping in the same direction. In addition, individual segmentation of both scheelite and molybdenite is possible and the quantitative difference between AM and XCT is similar for both minerals (Table 4). The main errors for scheelite come from streak artefacts and a pronounced partial volume effect at grain boundaries (despite the use of a Sobel filter) in the XCT data. For molybdenite, errors are especially miss-segmentation of fine grains as silicates and sulfides. This is because blurring is more pronounced in fine molybdenite grains (Hanna and Ketcham 2017), which therefore appear darker than larger molybdenite grains in the XCT images.

The 1 μm pixel size scan by AM was focused on an area of molybdenite mineralization at the intersection of the molybdenite-bearing quartz vein with the scheelite-bearing quartz vein. In several molybdenite grains, bright inclusions (a few micron large) are visible in the BSE image (Figure 9). Some grains are identified as hessite associated with native Bi, others escape detection. The topography of many deformed and fractured molybdenite grains is particularly problematic. Uneven surfaces cause low and spurious x-ray counts in the EDS detector (Newbury and Ritchie 2015). To limit misidentification a cut-off value was set at 1,000 x-ray counts per pixel. As a result, many molybdenite grains contain unclassified areas (Figure 9). At the intersection between the quartz veins, three (∼40–80 μm) large scheelite grains were detected, but well separated from the molybdenite grains. The XCT data reveal the contact between scheelite and molybdenite mineralization at depth in the sample, but are unsuccessful in resolving heavy mineral inclusions in molybdenite (Figure 8). Streak artefacts and blurring of the scheelite and molybdenite grains affect segmentation of the XCT data. Particularly, grain edges and space between grains in clusters of molybdenite grains are segmented as sulfides due to blurring of bright molybdenite with dark quartz. This causes an extreme over-segmentation of sulfides compared to AM and shows why for example AM is needed to correctly interpret XCT data (Table 4).

The XCT scan (cylinder with 1 mm height and diameter) of 1 μm voxel size of a molybdenite cluster is still hardly of good enough quality to resolve heavy mineral inclusions. Yet their presence in molybdenite of the Liikavaara Östra Cu-(W-Au) deposit is known from AM and prior studies by synchrotron radiation x-ray fluorescence mapping (Warlo et al. submitted). The study of such small-scale features might be difficult to achieve in a lab and could require e.g., synchrotron-based XCT (Wang and Miller 2020). Nevertheless, the high-resolution XCT scan successfully visualizes the flaky texture of molybdenite well (Figure 8). All three XCT scans show molybdenite grains to be predominantly aligned with the vein (Figure 8). For that reason, many molybdenite grains in the surface maps appear as thin elongated grains (cut more or less perpendicular to the crystallographic c-axis) rather than tabular flakes.

AM analysis shows gradually increasing biotitization of the K-feldspar and plagioclase matrix of the biotite schist towards the contact with the quartz vein (Figure 10). It also reveals ubiquitous epidote and tourmaline alteration within the biotite schist. The quartz vein contains disseminated calcite, which extends into the biotite schist. A thin straight calcite veinlet cuts through the quartz vein and continues as a zigzagging chlorite veinlet through the biotite schist. It also cuts disseminated pyrite associated with chalcopyrite (Figure 10). Mineralization of pyrite, chalcopyrite, pyrrhotite, and sphalerite within the quartz vein is strongly associated with calcite, which appears to replace the sulfides. Molybdenite mainly occurs along the contact of the quartz vein with the biotite schist and shows association with calcite but no signs of replacement like the other sulfides. Finally, a single rounded 100 μm large grain of scheelite is found in the quartz vein. In the XCT data, low or missing contrast in the grey values between many minerals results in a very simplified segmentation of five mineral classes (Figure 10, Table 4): 1) quartz and albite, 2) silicates and carbonates, 3) sulfides and oxides, 4) molybdenite, and 5) scheelite. This means that most of the features observed by AM concerning the biotite schist cannot be differentiated by XCT (Figure 10). In addition, calcite dissemination within the quartz vein is poorly resolved. However, it is possible to follow the molybdenite mineralization towards depth in the sample along the contact between the quartz vein and biotite schist (Figure 11). Additionally, larger scheelite grains are revealed at depth in the sample (Figure 11). Comparison of the top sample slice with AM shows a relatively good fit of segmented molybdenite, whereas sulfides/oxides are over-segmented (Table 4). Under-segmentation of silicates/carbonates appears to compensate mainly for this difference (Table 4). The scheelite grain observed by AM is too fine-grained for proper resolution and segmented as quartz/albite.

Linking AM and XCT has a strong potential in (ore) geology to improve quantitative and textural information on a rock sample (Butcher 2020). XCT is already applied in combination with x-ray diffraction microtomography (e.g., Artioli et al., 2010; Valentini et al., 2011; Mürer et al., 2018; Takahashi and Sugiyama 2019) and x-ray fluorescence microtomography (e.g., Wildenschild and Sheppard 2013; Laforce et al., 2017; Suuronen and Sayab 2018), although mainly constrained to synchrotron facilities and very small sample sizes. In contrast, the recently developed Oreexplore GeoCore X10 merges XCT with x-ray fluorescence analysis for drill core pieces up to 4 × 1 m but limited to a maximal voxel size of 200 μm (Bergqvist et al., 2019). The combination of these analytical techniques in a single system is possible since all techniques utilize an x-ray beam. In contrast, AM is an electron beam-based technique. Thus, a combination of XCT and AM in a single system is impossible with the current technological knowledge. Instead, research and development have to focus on software solutions to improve cross-system integration. So far commercially available integration of XCT and AM only concerns geo-referencing of samples for easier visual correlation and presentation (e.g., Zeiss Atlas 5). This is mainly helpful for qualitative comparison, e.g., for the XCT data to aid in 3D interpretations of AM data. In this study, XCT scans of ore samples show the extent of mineralization with depth in the samples, the angle and orientation of veins, and shapes and sizes of well-separated mineral grains of high contrast to the surrounding phases (e.g., molybdenite, scheelite, heavy minerals). Yet, the segmentation and study of gangue minerals like silicates and carbonates in XCT datasets is challenging, due to too low grey value contrast between the minerals. While in this study AM analysis was mainly performed prior to XCT analysis, results suggest XCT also to be useful for targeted sampling for 2D analysis. In none of the XCT analyses, the topmost slice (the polished surface, which was also analyzed by AM) is the slice with the highest content of the minerals of interest (heavy minerals, molybdenite, and scheelite; Tables 3, 4). Hence, if the goal was to maximize the amount of these minerals in a sample for AM, the XCT scan data should be used to guide the sub-sampling of the sample. This would be especially helpful to better quantify trace metal minerals, as the required number of grains for statistical robustness could be achieved with fewer samples (Goodall and Scales 2007; Godel 2013).

On the other hand, integrating 2-dimensional AM data and 3-dimensional XCT data for quantitative analysis is not straightforward. The reliance on dedicated software to process 2-dimensional AM data is a challenge when merging with 3-dimensional XCT data. The AM software used in this study (Mineralogic Explorer) calculates valuable mineral parameters like mineral association, grain size distribution, metal deportment, etc. but only for the entire AM scan. Obtaining this information for a region of interest in the scan, e.g., for better fitting with the XCT data, is not possible. Further, the software is limited to the processing of AM data. AM and XCT images are easily overlaid in a graphical software but at the cost of losing the functionalities of the AM software. In addition, success in overlaying the images can depend on the quality of the stitching of the montaged AM mineral map. In this study, multiple steps were required to integrate the AM mineral maps into the XCT processing software (ORS Dragonfly) for geo-referencing and quantitative analysis. Currently, the best way to address data integration between AM and XCT seems to be through programming of dedicated algorithms in e.g., MatLab (Reyes et al., 2017; Guntoro et al., 2019b). However, this requires proficiency in programming. Furthermore, software is not the only challenge concerning quantitative correlation.

In this study, one issue of XCT analysis was x-ray attenuation artefacts at the top of the samples (cone-beam effect; Cnudde and Boone 2013). To prevent these artefacts, the epoxy mounted rock samples were stacked with their surfaces facing each other during XCT analysis. However, instead dense minerals caused artefacts into adjacent samples. Thus, to resolve this issue the topmost slices were excluded from segmentation. Consequently, comparison with AM of the same exact surface was hardly possible. To assess the impact of a potential slight vertical offset between the AM mineral maps and topmost segmented XCT slices on quantitative comparison, the relative variation in bulk mineralogy across the five topmost segmented XCT slices was calculated in all scans (Tables 3, 4). Results indicate that the variation in bulk mineralogy is generally smaller between XCT slices than between AM and XCT for major and minor phases. This is to say, segmentation plays a larger role in the quantitative differences between AM and XCT than a potential vertical offset. Yet, for trace phases, especially the heavy mineral phase in sample 2, variation between slices is larger than between AM and XCT. This is due to the small grain size of the heavy minerals, which are partly not detectable across more than one slice. To improve correlation a better approach may be to cut and polish a rock sample after XCT analysis, to perform AM on a surface corresponding to an inner slice in the XCT data, unaffected by the cone-beam effect. An even better alternative may be to try to remove the cone-beam effect altogether. This is achieved with a helical scanning trajectory (De Witte 2010; Cnudde and Boone 2013) or a multiscan with top and bottom scans centered on the edges of the sample.

A 30 μm step size was chosen for both AM and XCT analysis for better modal comparison. Yet, effective resolution differed significantly. For AM, a 30 μm pixel size was sufficiently large not to cause overlap of the beam interaction volume with more than one pixel. In contrast, in XCT analysis the impact of mineral grains on the calculated x-ray linear attenuation exceeded their own voxels due to instrument-related resolution limitations (Kyle and Ketcham 2015; Hanna and Ketcham 2017). This blurring resulted in a relatively worse resolution of XCT than AM. To achieve similar resolution, a better XCT instrument would need to be used and/or the step size of XCT analysis reduced. Yet, the latter would entail a reduction in scanning volume also (Kyle and Ketcham 2015). Moreover, other artefacts like beam hardening, particularly for high-Z mineral grains, would persist (Kyle and Ketcham 2015). Hence, a perfect correlation of XCT data with AM data is hardly possible.

The possibility to differentiate between individual minerals is much lower for XCT than AM due to the nature of the techniques (x-ray linear attenuation coefficients vs. energy dispersive x-ray spectra coupled with backscattered electron intensity). Thus, AM mineral maps require simplification for correlation with the segmented mineral classes of XCT. This simplification may be based on visual comparison of the data sets and calculations on the x-ray linear attenuation coefficients for the various minerals quantified by AM. The approach works well for coarse-grained minerals, easily identified in the data of both techniques, but is less reliable for fine-grained minerals, which are blurred and lack distinct grain boundaries in the XCT images. Furthermore, in this study, molybdenite grains are significantly darker than scheelite grains in all analyses of this study, regardless of size and shape, despite a higher calculated x-ray linear attenuation coefficient. This creates ambiguity regarding which segmented class some minerals belong to.

Reliable mineral segmentation of XCT data, particularly of complex geological samples, is challenging, as described in literature (e.g., Godel 2013; Guntoro et al., 2019a, 2019b; Wang and Miller 2020), but also demonstrated by the results of this study. The challenge starts with the generation of high-quality images used for segmentation: 1) the co-occurrence of minerals with largely different x-ray linear attenuation coefficients makes choosing a beam energy optimized to separation of all minerals impossible. The same applies to selecting of filters and beam hardening corrections post analysis (Herman 1979; Kyle and Ketcham 2015; Hanna and Ketcham 2017). Rocks abundant in heavy minerals require high beam energies for adequate x-ray penetration. This lowers the x-ray attenuation contrast between many minerals. In addition, analytical settings to minimize streak artefacts of highly attenuating minerals like scheelite lower overall beam intensity, thus requiring longer scan times (Gusenbauer et al., 2016). Post-analysis metal artefact reduction methods exist but are not yet common in industrial XCT analysis (Gusenbauer et al., 2016). Combining the results from multiple scans at varied beam energies may help to alleviate some of the issue concerning the x-ray linear attenuation coefficient but comes with a cost of increased scan time. 2) Choosing a voxel size to resolve features of interest adequately while minimizing scan time (and thereby cost of analysis) is difficult in rocks of heterogeneous grain size. Fine intergrowths of minerals with small differences in attenuation coefficient easily become visually homogenized through blurring. A small voxel size may not be cost effective due to long scan times, but too large a voxel size carries the risk of small grains to be unresolved. Additionally, scan volume limits the available scan resolution (Kyle and Ketcham 2015). 3) The edge contrast between mineral grains, especially for minerals of similar x-ray attenuation coefficient (e.g., silicates), is less pronounced than between e.g., grain and air, which makes segmentation more difficult in low-porosity rocks. 4) Minerals may exhibit a large range of grey scale values, even if chemical composition is uniform, based on grain size and associated minerals due to blurring (Kyle and Ketcham 2015). Other artificial grey scale variations (e.g., gradients, partial volumes, streaks, shadows, and rings) cause overlap in the grey scale value ranges of minerals of similar x-ray linear attenuation coefficient, which again constrains segmentation.

Recently, advancements have been made to improve segmentation of XCT data through matching of slices in the XCT data with corresponding AM quantitative maps via a MatLab algorithm, which are then used to determine global thresholds for segmentation (Reyes et al., 2017) or as training data for machine learning algorithms (Guntoro et al., 2019b; Guntoro et al., 2020). Results were shown to be superior to traditional segmentation techniques like grey scale thresholding or watershed segmentation (Reyes et al., 2017; Guntoro et al., 2019b). This integration of XCT and AM is promising and will perhaps help to routinize the use of XCT in ore geology and the mining industry in the future. However, the complexity of the method, the necessity of proficient programming skills, and a relatively long post-processing time per sample are still an obstacle to broad application. In this study, especially the segmentation of chalcopyrite and pyrite in the XCT scan of the mini core of sample 2 and molybdenite in the rock chips of samples 3 and 4 could benefit from the use of machine learning. Concerning chalcopyrite and pyrite, they are clearly differentiable with the naked eye in the 80 kV scan, yet histographic segmentation is unsatisfactory (Figure 6). The slight grey scale gradient from the center to the sidewalls of the cylinder causes overlap of grey scale values between the minerals (cupping effect; Cnudde and Boone 2013). Additionally, the noise in the images prevents the sobel filter from properly resolving grain boundaries between pyrite and chalcopyrite despite sharpening and denoising of the XCT images. Matching of the characteristic cubic shaped pyrite grains between the XCT data and AM may however overcome these hurdles. Similarly, molybdenite grains have rather distinct elongated shapes, which may support discrimination against other sulfides of similar x-ray linear attenuation coefficient. In this study, especially the decrease in brightness of molybdenite with grain size leads to false segmentation of small molybdenite grains as sulfides/oxides and silicates/carbonates.

For every type of microanalysis method, time-to-result (and therefore the cost), has to be weighed up against the resolution needed to resolve the textures and mineral chemistry, which in turn may impact the quantity and quality of information gained. As a consequence, a common approach is sequential analysis – where the resolution is increased but at the same time the area scanned is made progressively smaller. This approach of ‘zooming in’ on a region of interest was also used in this study.

In the 30 μm pixel size AM mineral map of the rock chip of sample 2, several Au and Ag minerals are detected, but generally not in sizes larger than 1 or 2 pixels (Figure 5). The subsequent 1 μm pixel size scan on a much smaller area reveals numerous grains of Au, Ag, and Bi that are undetected in the 30 μm resolution scan (Figure 6). Thus, the 1 μm resolution provides much better information on parameters like grain size distribution, mineral association, etc. Obviously, a 1 μm pixel size scan of the entire polished surface of the rock chip would be ideal, yet it would consume an unreasonable amount of time (several days) to complete. In addition, if and to what extend identification of trace metal minerals could be expected would be purely based on observations of e.g., petrographic microscopy (often too low resolution) or manual SEM-EDS (mainly spot-checking, possibly time-consuming). Hence, the coarser but faster 30 μm scan serves as a good overview and allows targeted high-resolution analysis. Similar results are obtained by XCT, where with each scan at a smaller voxel size heavy mineral grains that the prior scans fail to resolve are detected (Figures 4, 6). Generally, coarser grains account for the majority of the bulk metal in an ore deposit and fine grains, even in high numbers, accumulate to only a small overall volume. In mining, loss of the finest-grained fractions is commonly accepted, partly due to missing tools and added costs for their recovery. However, for economic trace metals like Au, the fine-grained fractions can be substantial and loss to tailings a significant loss in potential profit (Vaughan 2004). Hence, these high-resolution analyses are needed to provide the necessary knowledge to develop production strategies.

Similarly, the study of molybdenite grains in the rock chip and mini core of sample 3 shows how high-resolution analyses are necessary to quantify micro- and nano-inclusions in molybdenite grains. Even the 1 μm resolution scans of AM and XCT only managed to resolve a small portion of the inclusions. Although the effect of these inclusions on processing of molybdenite requires more research, a study by Triffett et al. (2008) suggests foreign mineral grains trapped in molybdenite have potentially a negative effect on flotation.