The search for the optimal sensor

1‌ Sensor-based sorting – state of the art

Sensor-based sorting is used in the mining industry to separate 10- to 100-mm-large rock particles of a heterogeneous ore into a concentrate with a high content of valuable minerals and waste with a low content of valuable minerals [6]. The sorting process can be divided into four basic steps, as shown in Fig. 1 based on the example of a belt sorter.

Before sensor-based sorting can be implemented as part of the processing operation, a sensor must be found that is capable of recognizing or even quantifying a certain property of the material. The detected material property must be suitable for separation of the ore into a product with a high content of valuable minerals and one with a low content of valuable minerals. Often, it is not possible to directly identify the recoverable mineral itself as no suitable sensors are available. If this is the case, a measurable property of the ore has to be found that correlates with the recoverable mineral/element – an auxiliary value, often referred to as a proxy. Examples of frequently used proxies are atomic density measured with a dual-energy X-ray transmission detector (DE-XRT), colour, fluorescence or transparency (optical sensor) or also the emission or absorption of short-wave infrared or near-infrared radiation (SWIR/NIR).

To find a suitable sensor, according to the state of the art, a multistage process is applied with trial and error testing (Fig. 2, classical approach). First, a sorting parameter (e.g. colour) and a suitable sensor (→ optical sensor) are selected. This step is followed by tests on laboratory scale with around 10 to 100 individual particles. Based on the results, a suitable sensor is identified, and the test work continues on pilot scale. Geochemical analyses of the sorted products are used to assess the sorting performance in respect of recovery, grade and mass pull. On the basis of these values, and after weighing up of the cost and benefit, it is assessed whether the application of sensor-based sorting is worth it.

This iterative and empirical approach has the important advantage that it can be applied without any prior knowledge of the composition of the raw material. It also has, however, a number of disadvantages:

It is not clear why a group of ore particles is sorted out and vice versa, why the other group is yielded into the concentrate.

It is also unclear whether the tested sorting mechanism has delivered its full potential or whether potential for optimization of the sorter has remained undiscovered.

There are no indications whether parameters other than those selected for empirical testing could be used to concentrate the recoverable element or mineral more efficiently.

It can be expected that a comprehensive understanding of the mineralogical and structural properties of the raw material to be sorted enables an estimation of the efficiency of a sorting process. The required data can be provided, for example, by means of automated mineralogy (AM). It is precisely these data generated with AM, in combination with ML, which are used in a completely novel approach for the simulation of sorting efficiency.

This new approach has been developed at the Helmholtz Institute Freiberg for Resource Technology (HIF) within the framework of the r4 programme funded by Germany’s Federal Ministry of Education and Research in the projects AFK (registration number 033R128) and ResErVar (registration number 033R129). Quantitative mineralogical and microstructural data collected from a series of polished thin sections provide the suitable basis to simulate the performance of sensor-based sorting using different separation parameters. Empirical tests with commercially available sensor systems were conducted to validate the simulations experimentally. A flowchart of this simulation-based approach is shown in Fig. 2 (right). In this paper, the procedure is shown based on the example of a complex skarn ore from Hämmerlein (Ore Mountains) in Germany. Other examples have already been published in international trade journals [3; 9].

2 Methods

In preparation for AM analyses, polished thin sections of a suitable selection of samples of the raw material to be sorted are produced. The selection of the samples and the number of the sections to be prepared is derived from the inherent complexity of the raw material. The more complex and variable the raw material is in terms of its composition, the more sections are needed. In preparation of the sections, a sawn-off remnant always remains which is very similar to the section in its composition. These remnants can be used for sorting tests.

On the polished thin sections, high-resolution quantitative mineralogical analyses were conducted with AM (e.g. with a Mineral Liberation Analyzer, MLA for short) with a spatial resolution of a few micrometres. General explanations with regard to the measuring instrument and the method can be read in [2]. The result of an AM dataset is a high-resolution false colour image, from which, amongst other things, mineral content, particle size distribution and mineral associations can be calculated. The above-mentioned remnants from thin section preparation are measured with sensors (e.g. SWIR, DE-XRT, XRF) that can be used in sensor-based sorting. The spatial resolution of the validation measurements typically lies in the range of a few millimetres.

For the application of ML, in a first step, the high-resolution AM false colour images are adapted to the coarser spatial resolution of the sensors. Fig. 3 shows an example of this with the three phases (chlorite, quartz and cassiterite) and information from SWIR and AM. The AM image is changed in its size, and in every coarse pixel, the surface area share of every mineral and from this the percentage of the value mineral are calculated. In the second step, a threshold for the value mineral is determined, and based on this a two-class image (valuable vs. non-valuable) is compiled. The simplified two-class images are used for training and testing a suitable classification process from ML (e.g. Random Forest). The classification process is trained to predict the probability from the SWIR data co-registered with the AM data that an individual sample exceeds the threshold value and should therefore be transferred into the high-value product during sorting. With this procedure, which is explained schematically in Fig. 4, an ML-optimized prediction on the value of the sample is generated.

3 Case study: Complex skarn ore from Hämmerlein

The polymetallic Sn-Zn-In deposit Hämmerlein is located in the central part of Saxony’s Ore Mountains near the border to the Czech Republic (Fig. 5). Saxore Bergbau GmbH has been the owner of the exploration licence for Hämmerlein and the neighbouring ore bodies in the Tellerhäuser deposit area since 2012. Hämmerlein consists of two lithologically different parts: a skarn with Sn-In-Zn-mineralization (skarn ore) and a greisenized mica schist known as Schiefererz [7]. Of primary economic interest is the tin contained in recoverable concentrations in cassiterite (SnO₂). As a result of a late metasomatic overprint of the deposits, cassiterite was precipitated together with chlorite, fluorite, diverse sulphides and often quartz, too, in veinlets and lenses (Fig. 6). The samples used in this study come from the Hämmerlein skarn deposit (+ 590 m level), which is accessible via the Zinnkammern Pöhla visitor and show mine.

3.1 Raw material characterization

The Skarn ore consist predominantly of iron oxides (magnetite, hematite), silicates (amphibole, garnet, epidote and chlorite) and sulphide minerals. The cassiterite concentrations vary between 0.0 und 19.2 wt.% (average content: 0.6 wt.%). The sample variability is visualized with eight selected samples from various lithounits with false colour images (Fig. 7). Table 1 contains a listing of all analysed samples from the various samples. For differentiation between the ore and gangue rock, a threshold value of 0.1 wt.% Sn in cassiterite per sample was chosen. This value is exceeded by 35 % of the samples (28 of 81). The threshold value lies below the economic cut-off grade for mining (0.5 wt.% Sn; s. [8]), is regarded, however, as reasonable in the sorting phase in which the ore has already been blasted, transported and comminuted (personal communication from Dr Marco Roscher, Saxore Bergbau GmbH).

3.2 Simulations

AM data enable the simulation of a theoretically unlimited number of sorting tests. In the following section, upgrading and sorting curves on the basis of the cassiterite content, the chlorite content and a ML-optimized prediction of the sample value are discussed. The results of the simulations are compared with data from an SWIR sensor as can also be used in sorting systems.

Fig. 8 A shows an upgrading curve in the Fürstenau-II diagram [4] with cassiterite recovery (y axis) versus gangue recovery (x axis) for the case that the samples are sorted according to the property of cassiterite content (AM). From the diagram, it can be seen that it would be possible to produce a concentrate consisting of 80 % of the cassiterite and 15 % of the gangue. It would also be possible to produce a product consisting of 95 % of the cassiterite and 30 % of the gangue. Fig. 8 B shows an upgrading curve with technical efficiency of the separation (difference of cassiterite recovery and gangue recovery, y axis) as a function of the cassiterite recovery (x axis) [4]. The efficiency of the separation is a parameter to evaluate the separation performance. To achieve a recovery of more than 80 % cassiterite, a sensor would be necessary that quickly and precisely identifies cassiterite (or tin) in concentrations between 1 und 2 wt.%. To our knowledge, this is not possible even with state-of-the-art sensor technology. Nevertheless, this upgrading curve delivers important information as it shows the best possible sorting result for the existing particle size. This result can only be surpassed with further comminution of the material. For the purpose of orientation, this best possible sorting result is shown in Fig. 8 C and Fig. 8 D. An efficient and optimized sorting curve approaches the best possible result.

Fig. 8 C and Fig. 8 D compare the measured surface area of the mineral chlorite from the AM and the SWIR dataset. From the upgrading curves, a good agreement is evident. Separation based on chlorite enables a separation efficiency of 45 % with a cassiterite recovery of 70 % – while 75 % of the gangue is separated. With SWIR, it is possible to identify chlorite contents of 1 vol.% and less. For this reason, the sensor is suitable to reliably achieve more than 90 % recovery. The ML-optimized prediction of the value of the sample is suitable for a much more efficient separation. In this way, it is possible to concentrate around 90 % of the cassiterite together with less than 30 % of the gangue. This result too can be realized in an operational environment with the use of the right sensor.

3.3 Properties of the sorted products

Before the decision is taken for or against the installation of a sensor-based sorting system in a mineral processing operation, a feasibility study must be undertaken. Taken into consideration in this are primarily the recovery of the valuable mineral and separation efficiency, information made available by the AM data. With separation of low-value material, however, important parameters of the preconcentrated product (e.g. modal mineralogy, mass pull, particle sizes) change, which influences the downstream beneficiation processes (e.g. comminution, flotation, density separation and magnetic separation). Here, too, the available data enable a range of evaluation possibilities, which influence the feasibility study for the sorting system.

This is illustrated in the following with reference to the example of ML-optimized value prediction. With the ML-optimized separation characteristic, mineralogical composition, cassiterite content and the average cassiterite particle size of the concentrate and waste are modelled for 60 %, 70 %, 80 %, 90 %, 95 % and 100 % cassiterite recovery (Fig. 9). The properties of the sorted products are compared, with a focus on a cassiterite recovery of 90 % and 95 %.

With increasing cassiterite recovery to 90 %, the mineral content of the concentrate is dominated by quartz, chlorite and iron oxide minerals. The fluorite and sulphide minerals, too, are collected predominantly in the concentrate (Fig. 9 A). Garnet, epidote and amphibole are collected predominantly in the tailings of the sorting process (Fig. 9 B). For this target value, the cassiterite content lies at 1.9 wt.% and the average particle size measures almost 300 µm (Fig. 9 C and Fig. 9 D).

The mass of the tailings in the low-value product at 90 % cassiterite recovery amounts to 73 wt.% and the average size of the cassiterite particles is below 25 µm. To reach a high degree of liberation with such small particle sizes, it would be necessary to grind the material very finely with a high energy input. At a recovery of 95 % cassiterite, the high-value product is increasingly diluted with tailings, as a result of which the mass pull increases to 86 wt.%. To achieve 5 % more cassiterite recovery, a processing plant would be necessary that can cope with a three-times-higher throughput rate. The lower cassiterite content (0.7 wt.%) and the smaller average cassiterite particle size (< 250 µm) are other properties of the concentrate that would deteriorate as a result of a higher cassiterite recovery.  With consideration of all parameters described, the target value for an as efficient as possible separation of the skarn ore based on the value prediction optimized with ML lies at around 90 % cassiterite recovery.

4 Experimental validation

The results show excellent agreement between the empirical test results and simulations of the separation performance on the basis of AM data. The chlorite content measured with AM could be used to predict the success of the skarn ore separation. As chlorite has characteristic absorption properties in short-wave infrared, an SWIR sensor was chosen. The data obtained could also be used to optimize the sorting process by means of ML. The properties of the minerals in the simulated concentrate can be used for planning and prediction of the functionality of downstream processes such as flotation, density separation and magnetic separation.

The sorting of skarn ore with ML-optimized prediction is a good example to illustrate the interaction between sensor-based sorting and downstream processing steps: The unavoidably high content of chlorite in the concentrate is a disadvantage for further processing and chlorite reacts to reagents typically used for the flotation of cassiterite. This makes the separation of these two minerals more difficult. Moreover, the changed composition of the ore has an effect on the subsequent density separation and magnetic separation, which must be taken into consideration in models that predict the separation behaviour of the ore.

The simulation data presented in this study are based on analysis with MLA, but datasets of equal quality can be generated with similar instrument platforms like TIMA-X (TESCAN, Brno/Czech Republic), Mineralogic (Zeiss, Oberkochen/Germany) and QEMSCAN (FEI Company, Hillsboro, OR/USA). Common to all these systems is that the analyses are time- and cost-intensive, especially with the high spatial resolution used for this study. In current research projects, several possibilities are being investigated in Freiberg to drastically reduce the cost and the time for the necessary mineralogical tests. This is aimed at making the simulation-based approach even more interesting for industrial users.

5 Conclusion

Automated mineralogy (AM) data can be used in an approach developed at the HIF for the selection of sensors for sorting complex raw materials, to conduct an unlimited number of sorting simulations. In this way, suitable proxies can be found that permit efficient separation of coarse-grained particles, and accordingly enable the selection of a suitable sensor. With reference to the example of a skarn ore from Hämmerlein, it was possible to show impressively that SWIR data are suitable to optimize the sensor settings with machine learning to deliver an optimal solution for the sorting application. Moreover, from AM data, parameters such as modal mineralogy, mineral associations and particle size distributions in the simulated products (concentrate and waste) can be derived. This information is of considerable significance for downstream beneficiation processes. The novel approach can be adapted to many other types of raw material; it therefore has the potential to become a key technology for the optimization of sorting processes.