Automated sampling system

Abstract

An automated sampling system configured to determine the modal mineralogy of a slurry stream in real-time, the system comprising: a sampling collection and preparation unit configured to separate a slurry sample into a first sample fraction comprising particles having a first size in a first range and a second sample fraction comprising particles having a second size in a second different range; a robotic optical microscopy unit configured to obtain optical image data from the first sample fraction and the second sample fraction; and a modal mineralogy analyser configured to analyse the optical image data using a Machine Learning ("ML") algorithm in order to determine the modal mineralogy of the slurry sample.

Description

[0001] AUTOMATED SAMPLING SYSTEM

[0002] FIELD OF THE INVENTION

[0003] The present invention relates to an automated sampling system, a method of determining the modal mineralogy of a slurry stream in real-time, a robotic optical microscopy unit, a method of robotic optical microscopy, a computer implemented method of determining the modal mineralogy of a slurry stream in real-time, a computer- implemented method of training a neural network to determine the modal mineralogy of a slurry, a method of extracting one or more substances from an ore and a system configured to extract one or more substances from an ore.

[0004] BACKGROUND

[0005] It is known to perform mining operations and to seek to extract minerals and metals of interest from an ore. The extraction process may vary dependent on factors such as the depth of the mineral deposits, the nature of the surrounding rock and the desired output.

[0006] One method of mining involves open-pit mining which is particularly effective for minerals that are located close to the surface. This method entails the removal of layers of rock and soil to expose the mineral-bearing rock for extraction. It is suitable for the extraction of minerals such as copper, iron and gold.

[0007] In cases where mineral deposits are located deeper below the surface underground mining is employed. This method involves the construction of tunnels and shafts to reach the mineral deposits. Common underground mining techniques include room and pillar mining, longwall mining, and cut and fill mining. This method is often used for extracting coal, gold, and uranium.

[0008] Another method for mineral extraction is placer mining, which is effective for retrieving minerals from riverbeds or stream sediments. In placer mining, techniques such as panning, sluicing, and dredging are used to separate heavier minerals from lighter sediment through water-based processes. This method is often used for gold and gemstone extraction.

[0009] Heap leaching is utilised for extracting metals from low-grade ores. In this process, ore is stacked in large heaps, and a chemical solution is applied to dissolve the target minerals. The solution percolates through the heap, dissolving minerals such as gold, silver, copper, and uranium, which are then recovered from the collected solution. Froth flotation is another method used to separate valuable minerals from nonvaluable rock. This process involves mixing ground ore with water and chemicals to form a slurry, introducing air bubbles, which attach to the desired mineral particles, and floating them to the surface for collection. Froth flotation is commonly used for the extraction of copper, lead and zinc.

[0010] Other separation methods are known including tank leaching and gravity separation.

[0011] Hydraulic mining uses high-pressure waterjets to dislodge rock and extract minerals typically from alluvial deposits. Historically employed for gold extraction, this method is now less commonly used due to environmental issues.

[0012] In-situ leaching, or solution mining, is a method where a chemical solution is pumped directly into the ore body to dissolve the mineral in place. The solution is then pumped back to the surface and processed to recover the mineral. This method is often used for uranium and copper extraction.

[0013] Magnetic separation and gravity separation are also used for mineral extraction. Magnetic separation utilises magnetic properties to separate minerals such as iron ore from surrounding materials. Gravity separation exploits the differences in mineral density to separate heavier minerals from lighter materials, often using shaking tables or jigs.

[0014] Bioleaching is a method that uses bacteria to break down minerals and extract metals from low-grade ores. This technique is employed to extract metals such as copper and gold, offering a more environmentally friendly alternative to traditional chemical methods.

[0015] Finally, smelting is a thermal process used to extract metals by heating the ore to high temperatures causing the metal to separate from impurities. This method is used for extracting metals such as iron, copper and lead.

[0016] Each of these methods can be applied depending on the specific characteristics of the mineral deposits and geological formations. They can also be combined or adapted to improve efficiency and resource recovery in mineral extraction operations.

[0017] Of particular interest is the use of a concentrator to extract minerals and metals of interest. In order to extract minerals or metals of interest it is known to utilise one or more processes as mentioned above. For example, froth flotation is a commonly used separation process especially for sulphide minerals whereas leaching is more suitable for oxides. Sulphide concentrators are known which are utilised to process sulphide ore which often contains valuable metals such as copper, nickel, zine, lead or gold bound with sulphur in the form of a sulphide minerals such as chalcopyrite, galena, sphalerite and pentlandite. It is also known to recover molybdenum from molybdenite using flotation techniques wherein molybdenum is commonly recovered with copper.

[0018] It is known to crush sulphide ore into smaller particles and then to grind the crushed particles into a fine powder which liberates the minerals of the ore. The ore is then mixed with water to form a slurry or slurry and is then treated with chemicals. The chemicals may be referred to as reagents. It will be understood that there are various different type of reagents may be added such as collectors, frothers, pH modifiers, depressants and others.

[0019] The reagents which are added to the slurry are intended to make the sulphide minerals hydrophobic i.e. water-repellent. Air is then introduced into the slurry creating bubbles which causes a froth to form. The hydrophobic sulphide minerals attach to the bubbles and rise to the surface forming a froth which is then skimmed off. At the same time material which is not of interest (e.g. gangue) sinks to the bottom and may be discarded as tailings.

[0020] Forth flotation is usually performed in stages (e.g. rougher, scavenger and cleaners) in order to achieve target recovery and grade. It will be understood therefore by those skilled in the art that it would be beneficial to understand the slurry concentration / grade at each stage of processing.

[0021] After froth flotation excess water is then removed from the concentrate and the concentrate from flotation is thickened by sedimentation to remove much of the water and is then filtered to produce a solid concentrate.

[0022] The waste material from flotation (tailings) are also thickened and dewatered before being disposed of in tailings storage facilities.

[0023] It is common for copper, nickel and zinc to be recovered from sulphide ores such as chalcopyrite (copper), pentlandite (nickel) and sphalerite (zinc). Gold and silver may also be recovered from sulphide minerals and lead may be extracted from sulphide ores such as galena.

[0024] It will be understood that conventional approaches are capable of achieving relatively high recovery rates. However, there are numerous operational parameters to control and the recovery process could be made further efficient. Furthermore, it is also desired for environmental reasons to reduce the usage of water and chemicals during the refining process. It will be understood that improvements in the refining and recovery process can both increase the recovery rates yet further and also at the same time reduce the use of water and chemicals per unit mass of the product by optimising the separation process. To achieve that, understanding mineralogy of the ore in the process is desirable in order to find the most appropriate set points for the process.

[0025] Furthermore, it will be understood by those skilled in the art that the refining and recovery process can be energy intensive. For example, comminution energy can account for the majority of concentrator energy use and cost.

[0026] It is known to use a Scanning Electron Microscope (“SEM”) with energy dispersive X-ray (“EDS”) like QEMSCAN and Mineralogy Liberation Analysis (“MLA”) to analyse mineral samples. In particular, various automated analysers are known including ZEISS Mineralogic and TESCAN Integrated Mineral Analyzer (“TIMA”).

[0027] A system known as the Advanced Mineral Identification and Characterization System (“AMICS”) is also known which comprises a software package for the automated identification and quantification of minerals and synthetic phases. The software package utilises imaging and analysis capabilities. An arrangement is known comprising an energy dispersive X-ray spectrometry (“EDS”) system utilising a scanning electron microscope (“SEMs”) and providing a fully automated Mineral Liberation Analyzer (“MLA”).

[0028] However, conventional approaches require sample preparation for polished sections and the turnaround for analysis are typically two shifts or longer.

[0029] Known concentrators adopt online elemental analysers such as X-Ray Fluorescence (“XRF”) or Laser Induced Breakdown Spectroscopy (“LIBS). However, it will be understood that conventional analysers are expensive and are unable to analyse a slurry in real-time.

[0030] It will be understood that in a separation process particles and minerals are separated rather than elements. Elements of interest appear in different mineral formats which behave differently in the process and which therefore require different process settings and treatments in order to optimise separation.

[0031] There are various problems associated with the use of conventional analysers. For example, online XRF has poor accuracy for elements below 20 atomic number (Ca) as these elements have low-energy X-rays that get occluded through the slurry. Because of this, one limitation of XRF is its inability to characterise clays and differentiate copper sulphide species as they contain elements like S, Si, Mg, Al that have low atomic numbers which greatly affects flotation performance. Whilst LIBS does not suffer from this limitation it is significantly more expensive and has less accuracy than XRF for heavier elements like copper.

[0032] It will be apparent, therefore, that conventional analysers suffer from various problems and that conventional arrangements are incapable of detecting minerals in a stream slurry in real-time.

[0033] SUMMARY

[0034] According to an aspect there is provided an automated sampling system as claimed in claim 1.

[0035] According to an aspect there is provided an automated sampling system configured to determine the modal mineralogy of a slurry stream in real-time, the system comprising: a sampling collection and preparation unit configured to separate a slurry sample into a first sample fraction comprising particles having a first size in a first range and a second sample fraction comprising particles having a second size in a second different range; a robotic optical microscopy unit configured to obtain optical image data from the first sample fraction and from the second sample fraction; and a modal mineralogy analyser configured to analyse the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

[0036] Other embodiments are contemplated wherein a slurry sample may be sampled without necessarily separating the sample fraction into a first sample fraction (which may comprise fine particles) and a second sample fraction (which may comprise coarser particles).

[0037] According to various embodiments the sampling collection and preparation unit may comprise a particle size sorter which may be configured to separate finer particles from the slurry. For example, a cyclonic separator may be provided to separate a slurry sample into a fine sample comprising particles having a maximum size < D3pm and a coarser sample comprising particles having a size > D3pm. For example, D3 may be 38 pm which corresponds with the minimum sieve size which is commonly available in testing laboratories. However, it will be understood that D3 may take other values.

[0038] It should be noted that various embodiments relate to the determination of the modal mineralogy of a slurry in real-time, something which is not known in the art. It will be understood that the ability to determine the modal mineralogy of a slurry stream in real-time represents a significant advance in the art as it allows real-time operational parameters of a mineral processing system to be adjusted in real-time in order, for example, to increase the efficiency of the extraction process of minerals from the slurry because the modal mineralogy of the slurry at any point in time can be determined. Furthermore, it will be appreciated that current mineral processing systems utilise various reagents and operating set points and hence knowing the precise mineralogy of the slurry in real-time enables the processes involving adding water (or dewatering) and adding reagents to be optimised and the total volume of water and chemicals can be reduced. Furthermore, knowing the real-time modal mineralogy of the slurry enables operational parameters, such as the final grind size to be optimised thereby making the overall process more energy efficient.

[0039] Furthermore, it will be understood that according to various embodiments whilst the total energy usage may not be reduced, embodiments are contemplated wherein the recovery efficiency (i.e. metal production per unit of energy) may be improved (i.e. increased).

[0040] Various embodiments are concerned with improvements in flotation process modelling and control which are particularly beneficial for concentrators used for mineral extraction, especially sulphide concentrators. However, the present invention is not limited to froth flotation processes. For example, other forms of separators can benefit from on-line modal mineralogy analysis according to various embodiments. For example, according to various embodiments separators such as leaching and gravity separators may utilise the automated sampling system and method as disclosed herein. In particular, the system and method according to various embodiments are particularly suitable for wet separation techniques which conventional methods are unsuited to determine modal mineralogy in-line.

[0041] According to various embodiments high-resolution and accurate data about the separation system may be collected and determined in real-time. For example, it will be understood that the performance of a flotation system performance depends upon several components namely: (i) the equipment e.g. cell design, agitation, air flow, bank configuration and control; (ii) the chemistry e.g. reagent choices, dosages and pH; and (iii) the operation e.g. feed rate, mineralogy, feed size distribution, solids density and temperature.

[0042] In particular, embodiments of the present invention are concerned with determining the model mineralogy of a slurry stream in real-time and utilising this understanding in order to control and vary various operational parameters or conditions such as, for example, the level of agitation, air flow, control and reagent dosages in order to optimise the extraction process and make the process more efficient and less energy intensive. Also, according to various embodiments other aspects such as the pH of the slurry, the feed rate, the feed particle size distribution, and the solids density in the slurry may be monitored using sensors and may also according to various embodiments be adjusted or otherwise controlled in response to the real-time determination and analysis of the modal mineralogy of a slurry stream.

[0043] It will be recognised, therefore, that the present invention represents a significant advance in the art and enables significant efficiencies to be achieved particularly in the use of water, chemicals, heat and energy in the process of refining ores.

[0044] It will be appreciated that mining operations and the recovery of minerals / metals from ore are operated at large scale especially in countries such as Australia and in locations such as North America and South America especially in respect of processing copper deposits. Accordingly, the ability to improve current processes in terms of efficiency and energy savings represents a significant advance in the art due to the scale of such an impact upon mining operations.

[0045] Although various embodiments are focussed upon the commercial extraction of minerals / metals from ore which has obvious industrial application and utility, the present invention is also advantageous in terms of helping geologists to analyse ore as it is extracted in real-time and to be able to link the analysis of ore to a specific location and time of extraction.

[0046] According to various embodiments an automated sampling system for analysing a slurry in-real time are disclosed together with an associated method.

[0047] In the context of the present application a “slurry” should be understood as comprising a mixture of ground ore particles which are suspended in water. It will be understood that during a separation minerals are separated from gangue i.e. waste material. The slurry density is the ratio of solids to liquid and is an important ratio to optimise during e.g. a flotation separation process.

[0048] Furthermore, in the present application reference is made to determining the “modal mineralogy” of a sample. It will be understood that modal mineralogy refers to the volume or mass proportions of minerals in a rock or ore and is concerned with making a quantitative determination of the mineral composition of a sample thereby providing insights into its geological formation and economic value.

[0049] According to various embodiments a modal mineralogy analyser is provided which may be configured to use an unsupervised clustering algorithm in order to analyse the first sample fraction. The algorithm may comprise a SLIC superpixels algorithm. However, according to other embodiments a different algorithm may be used. The modal mineralogy analyser may be configured to use a supervised Deep Learning (“DL”) algorithm in order to analyse the second sample fraction. The supervised Deep Learning (“DL”) algorithm may comprise one or more Deep Learning (“DL”) instance segmentation algorithms. For example, the one or more Deep Learning (“DL”) instance segmentation algorithms may comprise a first Deep Learning (“DL”) instance segmentation algorithm to identify particles and a second Deep Learning (“DL”) instance segmentation algorithm to classify minerals in the particles.

[0050] However, other embodiments are contemplated wherein different algorithms may be utilised either to analyse the first sample fraction and / or the second sample fraction. For example, a supervised learning algorithm may be utilised. It will be understood that a supervised learning algorithm differs from an unsupervised clustering algorithm in that the supervised learning algorithm relies upon labelled training data to make predictions. In supervised learning, algorithms such as decision trees, support vector machines (“SVM”) and neural networks can be used to classify data into predefined categories. These algorithms learn from input-output pairs during training, enabling them to make informed predictions when presented with new, unseen data. Supervised learning is especially useful when labelled data is available, and the task involves specific target outcomes.

[0051] The modal mineralogy analyser may be configured to determine the size of particles in the slurry sample.

[0052] The modal mineralogy analyser may be configured to determine the particle size distribution (“PSD”) of the slurry sample.

[0053] The modal mineralogy analyser may be configured to determine the percentage of solids in respect of particles in the slurry sample.

[0054] According to another aspect there is provided a method of determining the modal mineralogy of a slurry stream in real-time, the method comprising: separating a slurry sample into a first sample fraction comprising particles having a first size in a first range and a second sample fraction comprising particles having a second size in a second different range; obtaining optical image data from the first sample fraction and the second sample fraction; and analysing the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

[0055] The first sample fraction may comprise particles having a maximum particle size selected from the group consisting of: (i) < 10 pm; (ii) 10-20 pm; (iii) 20-30 pm; (iv) 30-40 pm; (v) 40-50 pm; (vi) 50-60 pm; (vii) 60-70 pm; (viii) 70-80 pm; (ix)0-90 pm; (x) 90-100 pm; and (xi) > 100 pm. The second sample fraction may comprise particles having a minimum particle size selected from the group consisting of: (i) < 10 pm; (ii) 10-20 pm; (iii) 20-30 pm; (iv) 30-40 pm; (v) 40-50 pm; (vi) 50-60 pm; (vii) 60-70 pm; (viii) 70-80 pm; (ix)0-90 pm; (x) 90-100 pm; (xi) 100-110 pm; (xii) 110-120 pm; (xiii) 120-130 pm; (xiv) 130-140 pm; (xv) 140-150 pm; (xvi) 150-160 pm; (xvii) 160-170 pm; (xviii) 170-180 pm; (xix) 180-190 pm; (xx) 190-200 pm; and (xxi) > 200 pm.

[0056] The method may further comprise using an unsupervised clustering algorithm to analyse the first sample fraction. The algorithm may comprise a SLIC superpixels algorithm. The modal mineralogy analyser may be configured to use a supervised Deep Learning (“DL”) algorithm in order to analyse the second sample fraction. The supervised Deep Learning (“DL”) algorithm may comprise one or more Deep Learning (“DL”) instance segmentation algorithms. The one or more Deep Learning (“DL”) instance segmentation algorithms may comprise a first Deep Learning (“DL”) instance segmentation algorithm to identify particles and a second Deep Learning (“DL”) instance segmentation algorithm to classify minerals in the particles.

[0057] According to another embodiment a semi-supervised learning algorithm may be utilised which combines elements of both supervised and unsupervised learning. According to such an embodiment an algorithm may be utilised that makes use of a small amount of labelled data combined with a larger amount of unlabelled data. By leveraging the information from the labelled set, the semi-supervised learning algorithm can produce more accurate predictions than a purely unsupervised method. This approach is particularly beneficial when obtaining labelled data is expensive or timeconsuming.

[0058] According to another embodiment a reinforcement learning algorithm may be utilised wherein an agent learns to make decisions by interacting with an environment and receiving feedback in the form of rewards or penalties. Through trial and error, the agent optimises its actions to maximise cumulative rewards over time. Reinforcement learning is particularly useful in dynamic or interactive environments where the agent can learn optimal strategies through exploration.

[0059] According to other embodiments a dimensionality reduction algorithm such as Principal Component Analysis (“PCA”) ort-Distributed Stochastic Neighbour Embedding (“t-SNE”) may also be utilised in place of an unsupervised clustering algorithm. These techniques reduce the number of variables in a dataset while preserving its structure, making it easier to identify patterns or clusters within the data. Although not a clustering algorithm per se, dimensionality reduction methods may be used in conjunction with other algorithms to improve the performance of clustering or classification tasks. According to yet further embodiments a genetic algorithm or evolutionary algorithm may be employed as an optimisation techniques in cases where the objective is to find the best solution from a set of possible solutions. These algorithms simulate the process of natural evolution, using mechanisms such as selection, mutation and crossover to evolve solutions overtime. Genetic algorithms are particularly useful for solving complex optimisation problems where traditional methods may struggle to find optimal solutions efficiently.

[0060] The method may further comprise determining the size of particles in the slurry sample.

[0061] According to various embodiments the method may further comprise determining the particle size distribution (“PSD”) of the slurry sample.

[0062] The method may further comprise determining the percentage solids of particles in the slurry sample.

[0063] According to another aspect there is provided a robotic optical microscopy unit configured to determine the modal mineralogy of a slurry stream in real-time, wherein the robotic optical microscopy unit is configured to obtain optical image data from a first sample fraction and from a second different sample fraction; and wherein the robotic optical microscopy unit is configured to analyse the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

[0064] According to another aspect there is provided a method of optical microscopy comprising: determining the modal mineralogy of a slurry stream in real-time by obtaining optical image data from a first sample fraction and from a second different sample fraction; and analysing the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

[0065] According to another aspect there is provided a computer implemented method of determining the modal mineralogy of a slurry stream in real-time, the method comprising: providing a trained deep neural network that is executed by software using one or more processors of a computing device, the trained deep neural network having been trained with a training set of images comprising microscopy images; separating a slurry sample into a first sample fraction comprising particles having a first size in a first range and a second sample fraction comprising particles having a second size in a second different range; obtaining optical image data from the first sample fraction and from the second sample fraction; inputting the optical image data to the trained deep neural network; and outputting modal mineralogy data related to the slurry, wherein the modal mineralogy data includes data relates to the size, particle size distribution (“PSD”) and percentage solids of particles in the slurry.

[0066] According to another aspect there is provided a computer-implemented method of training a neural network to determine modal mineralogy in a slurry, the method comprising: obtaining a first set of optical images of a slurry as a first training set; training a neural network as a first stage using the first training set; creating a second training set comprising the first training set and modal mineralogy data associated with the first training set; and training the neural network as a second stage using the second training set.

[0067] According to another aspect there is provided a method of extracting one or more substances from an ore comprising: processing an ore to form a slurry; controlling one or more first operational parameters or conditions of the slurry; obtaining optical image data from the particles in the slurry; analysing the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry’s particles in real time; and altering or varying one or more of the first operational parameters or conditions of the slurry dependent upon the determined modal mineralogy of the slurry s particles.

[0068] According to various embodiments the one or more first operational parameters or conditions may be selected from the group comprising: (i) feed flow rate; (ii) mineralogy; (iii) particle or feed size distribution; (iv) slurry density; (v) reagent type; (vi) reagent dosage; (vii) air flow rate; (viii) slurry level; (ix) slurry chemistry; (x) liberation size; (xi) residence time; (xii) valuable and gangue mineral distribution; (xiii) minerals liberation degree; (xiv) mineral association; (xv) mineral grain size; (xvi) pH of the slurry; (xvii) pressure; and (xviii) temperature.

[0069] According to another aspect there is provided a system configured to extract one or more substances from an ore comprising: a processor for processing an ore to form a slurry; a controller for controlling one or more first operational parameters or conditions of the slurry; an optical imager for obtaining optical image data from the slurry; and an analyser for analysing the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry in real time; wherein the controller is configured to alter or vary one or more of the first operational parameters or conditions of the slurry dependent upon the determined modal mineralogy of the slurry.

[0070] The one or more first operational parameters or conditions may be selected from the group comprising: (i) feed flow rate; (ii) mineralogy; (iii) particle or feed size distribution; (iv) slurry density; (v) reagent type; (vi) reagent dosage; (vii) air flow rate; (viii) slurry level; (ix) slurry chemistry; (x) liberation size; (xi) residence time; (xii) valuable and gangue mineral distribution; (xiii) minerals liberation degree; (xiv) mineral association; (xv) mineral grain size; (xvi) pH of the slurry; (xvii) pressure; and (xviii) temperature.

[0071] According to an aspect there is provided an automated sampling system configured to determine the modal mineralogy of a slurry stream’s particles in real-time, the system comprising: a cyclonic separator configured to separate a slurry sample into a first sample portion comprising particles having an average size Di and a second sample portion comprising particles having an average size D2, wherein D2> Di; a robotic optical microscopy unit configured to obtain optical image data from the first sample portion and from the second sample portion; and a modal mineralogy analyser configured to analyse the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

[0072] According to an aspect there is provided a computer-implemented method of determining the modal mineralogy of a slurry stream in real-time, the method comprising: cyclonically separating a slurry sample into a first sample portion comprising particles having an average size Di and a second sample portion comprising particles having an average size D2, wherein D2> Di; using a robotically operated optical microscope to obtain optical image data from the first sample portion and from the second sample portion; and analysing the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

[0073] BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Various embodiments of the present invention will now be described, by way of example only, and with reference to the accompanying drawings in which:

[0075] Fig. 1 shows a XRPD spectra obtained for a concentrate sample;

[0076] Fig. 2 shows a XRPD spectra obtained for a tailings sample; Fig. 3 shows sample images of the major minerals under optical microscopy as obtained according to various embodiments and a brief description of their appearance;

[0077] Fig. 4A shows a prototype optical microscopy apparatus which was built out of steel slotted racks and which is included for illustrative purposes and Fig. 4B shows cuvette samples which were placed into a cuvette holder;

[0078] Fig. 5 shows example images collected from the fines and coarse samples at varying concentrate blends according to various embodiments;

[0079] Fig. 6 shows the sample output of a coarse algorithm according to various embodiments;

[0080] Fig. 7 shows instance segmentation algorithm steps on a sample image of a fine fraction according to various embodiments;

[0081] Fig. 8 shows a parity plot of one pass of P50 measurements relating to 30 samples according to various embodiments;

[0082] Fig. 9 shows parity plots comparing one pass (20 pictures) and five pass (100 pictures) measurements on three samples from a validation dataset;

[0083] Fig. 10 shows parity plots of one pass mineralogy grades for 30 samples according to various embodiments;

[0084] Fig. 11 shows parity plots comparing one pass (20 pictures) and five pass (100 pictures) measurements on three samples from a validation dataset;

[0085] Fig. 12 shows an illustration of a hypothetical cuboid bounding a particle and a sieve diameter definition;

[0086] Fig. 13 illustrates two possible sieve diameters for a cuboid;

[0087] Fig. 14 shows image density calibration results for a fines fraction;

[0088] Fig. 15 shows image density calibration results for a coarse fraction;

[0089] Fig. 16 shows various influential parameters on floatation processes;

[0090] Fig. 17 shows the impact of NaHS dosage and pH level on chalcocite, bornite and chalcopyrite floatation recovery; Fig. 18 shows a history of mineral detection and composition analysis;

[0091] Fig. 19 shows a robotic sampling and scanning unit according to various embodiments;

[0092] Fig. 20 shows a screenshot of a machine learning software and user interface under operation according to various embodiments;

[0093] Fig. 21 shows the methodology according to various embodiments which combines SEM and EDS map scan with image data according to various embodiments to identify locked pyrite and iron oxide bearing gangue;

[0094] Fig. 22 shows the conversion of images to modal mineralogy by size;

[0095] Fig. 23 shows elemental assays calculated according to various embodiments by reading of 20 images versus ICP-MS results;

[0096] Fig. 24 shows elemental assays calculated according to various embodiments from reading 100 images (five passes) versus ICP-MS results;

[0097] Figs. 25A and 25B show a comparison of fully liberated mineral appearances under SEM, MinDet (i.e. according to embodiments of the present invention) and EDS;

[0098] Figs. 26A and 26B show locked sulphide particles scanned by SEM and MinDet (i.e. according to embodiments of the present invention), wherein only sulphide particles had one point EDS scan which is shown by a circle; and

[0099] Fig. 27 shows a flow chart illustrating various aspects of embodiments according to the present invention.

[0100] DESCRIPTION

[0101] Various embodiments will now be described in further detail. In particular, an automated sampling system will be described below which is capable of analysing a slurry sample in real-time. In particular, as will be described in more detail according to various embodiments a system and method for determining the modal mineralogy of a slurry in real time and on-line is disclosed. It will be understood that modal mineralogy is concerned with determining the liberation, minerals grain size, mineral distribution and optionally other parameters in respect of particles e.g. suspended in a liquid such as water. In the context of the present application a “slurry” should be understood as comprising a mixture of ground ore particles suspended in water. It will be understood that during a separation process minerals are separated from gangue i.e. waste material. According to various embodiments one or more chemicals or reagents may be added to the slurry. It will be understood that multiple different types of reagents may be added to the slurry including one or more collectors, frothers, pH modifiers or depressants. A “slurry” should also be understood as comprising a mixture of crushed ore and water during the initial stages of mineral processing including stages such as transportation, grinding and also other stages such as dewatering.

[0102] In the context of flotation separation a reagent may be added to the slurry in order to render minerals of interest especially sulphides hydrophobic. As a result, when a froth is created by pumping air into the slurry then hydrophobic mineral substances become entrained in the bubbles which form on the surface of the slurry as forth and conversely waste material (gangue) sinks to the bottom. Accordingly, the process of froth separation enables minerals / metals of interest to be separated from waste material.

[0103] It should, however, be understood that the present invention is not limited to flotation separation techniques but has wider applicability to other types of separation techniques. For example, other forms of separators can benefit from on-line modal mineralogy analysis according to various embodiments. For example, according to various embodiments separators such as leaching and gravity separators may utilise the automated sampling system and method as disclosed herein. In particular, the system and method according to various embodiments are particularly suitable for wet separation techniques and determining the modal mineralogy in real time during processing of the slurry - something that conventional systems are not capable of doing.

[0104] It will be understood that the slurry density is the ratio of solids to liquid and is an important ratio to optimise during e.g. a flotation separation process. According to various embodiments, a determination may be made of the modal mineralogy of a slurry in real-time and on-line which enables operational parameters of a separation process (such as a leach, gravity, flotation or other separation process) to be optimised in order to increase the efficiency of extraction of minerals / metals of interest and / or reduce waste and / or reduce the use of chemicals and / or to make the process more energy efficient.

[0105] Accordingly, the present invention represents an important advance in the art and has potential very significant applications in the field of mineral extraction which is a large scale industry with very significant energy demands. It will be understood that mineral extraction is a major industry in Australia, North America and South America and that in particular the extraction of metals, minerals and rare metals has significant commercial importance and utility. In the following disclosure a prototype apparatus and a method for analysing a slurry is first described. The apparatus was further developed and a further apparatus and method are disclosed may be utilised in the real-time and on-line analysis of a slurry. In particular, according to various embodiments the results of the real-time and on-line analysis of the slurry may be utilised to control and / or optimise various processing stages in mineral and metal extraction. According to various embodiments the illumination set up along with the lens and camera sensor was further developed and improved in order to provide improved detection and differentiation of shiny minerals such as sulphides and naturally free metals. The camera sensor was further developed and changed from a roller to a global shutter in order to address microscope automated movements and shakes. The resolution and optical set up was also modified. According to various embodiments the further apparatus utilises software which uses different libraries and structures for learning and mineral detection which results in improved operating times. According to various embodiments 2D robotic arms and software were used rather than an electrical actuator. Advantageously, the robotic arms eliminate any need to make manual adjustments of the samples. Also, an algorithm for the robotic arms was developed which makes auto scanning possible.

[0106] An in-slurry measurement system has been developed which utilises machine vision and optical microscopy for the determination of modal mineralogy of the slurry thereby enabling improvements to be made in respect of froth flotation performance modelling, control and operation In particular, according to various embodiments of the present invention the modal mineralogy of a slurry may be determined in real-time and various operational parameters of the extraction and refining process may be controlled and optimised in real-time resulting in a significant improvement in performance and reduction of resources and energy.

[0107] When reference is made to the optical analysis of a slurry then it should be understood that such a term may be used to indicate that the wavelength is measured in the visible spectrum. However, it should also be understood the wavelength spectrum may be extended to include the near infrared and / or ultraviolet.

[0108] As will be explained in more detail below various tests and trials were initially performed using an apparatus to demonstrate the feasibility of imaging particles in-slurry as optical reflective properties can be occluded by a slurry medium and slimes. The tests and trials detailed below demonstrate the effectiveness of utilising machine vision algorithms to identify the boundaries of particles (e.g. for sizing) and for classifying the minerals (e.g. for surface grade calculation) within optical images.

[0109] In addition, estimates of the number of images required for a representative sample to obtain an acceptable accuracy for particle size distribution and mineralogy are presented below together with a comparison of the speed of the present invention relative to existing mineralogy characterisation methods.

[0110] Reference is made below at various points to instruments which were tested by imaging particles in-slurry. In contrast, it will be understood that existing commercial methods such as XRF and LIBS are performed in-slurry which indicates that these commercial machines can sample from a process stream, perform measurements and return the sample to the process stream.

[0111] In particular, an important aspect of various embodiments is that slurries may be analysed in real-time and their modal minerology may be determined by sampling a slurry. Then having determined the modal minerology of a slurry in real-time various operational parameters relating to the processing of the slurry may be altered in real-time in order to make the extraction of minerals / metals from the slurry more efficient in terms of overall recovery rate, speed, chemical and water use, and energy use.

[0112] Instance

[0113] As will be explained in more detail below, a modal mineralogy analyser is disclosed which utilises a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of a slurry sample in real-time.

[0114] As will be explained in more detail below a slurry sample may be separated into a fine sample comprising particles which have a size < 38 pm and a coarse sample comprising particles which have a size > 38 pm. However, it should be understood that the boundary between referring to particles as being either fine or coarse is not fixed by a 38 pm size limitation and that a 38 pm size was selected as it is commonly the smallest standard sieve size found in a commercial laboratory which was suitable for the ore being tested in the trials. In accordance with other embodiments a fine sample may for example comprise particles having a different size e.g. < 20 pm, < 40 pm, < 60 pm, < 80 pm, < 100 pm or < 150 pm and the particles coarser that the cut size (D3) are categorised as the coarse sample.

[0115] It will be understood that instance segmentation is a machine vision task which can detect and segment boundaries and identify objects from the background and each other. The term instance indicates that images can contain multiple instances of the same class and requires distinction. Thus, instance segmentation is particularly suitable according to various embodiments as optical microscopy images of samples have multiple particles of multiple minerals and require segmentation of boundaries of each particle in order to evaluate sizes and modal mineralogy.

[0116] Deep Learning (“DL”) is a subset of Machine Learning (“ML”) which utilises deep neural networks to accomplish tasks such as machine vision, autonomous driving and natural language processing. The large model size composing of many neural network layers (in some cases up to millions or billions of parameters) allows DL models to learn many features from input data. DL can learn complex features to distinguish minerals and boundaries but typically requires a large and clean labelled dataset via supervised learning. However, it is possible to overcome any large dataset requirement by utilising pre-trained models on other datasets and tweaking the model parameters for a study with a smaller dataset. This practice is known as transfer learning. and classifier instance

[0117] When collection of a labelled dataset is not feasible, it is still possible to segment images through unsupervised learning. Such an example is superpixel technique which groups pixels according to common characteristics. For example, a known superpixel method is to use the Simple Linear Iterative Clustering (“SLIC”) algorithm. This algorithm first creates initial region centres based on pre-defined number of superpixels and image dimension. Then, pixels are aggregated to regions according to the similarity criteria to the region clusters. The clusters are iteratively updated, and the reconstruction error is evaluated. This step repeats until the error reaches a certain threshold.

[0118] It will be understood that the SLIC algorithm is a special case of the k-means algorithm for generating superpixels. The benefit of superpixels over DL models is that they do not require a labelled dataset for training at the cost of a simpler segmentation algorithm. According to various embodiments the SLIC superpixel algorithm was used to process fine particles and slimes (< 38 pm). The trade-off is that the superpixels could only be classified through low-level features like pixel colour and brightness.

[0119] The methodology used to develop and validate in-slurry optical microscopy mineralogy characterisation technique will now be described.

[0120] Data is presented below related to the analysis of a sulphide polymetallic ore sample. The samples were processed in a concentrator in order to produce copper concentrate. The copper concentrate and tails were then analysed utilising X-ray Powder Diffraction (“XRPD”) and Inductively Coupled Plasma-Mass Spectrometry (“ICP-MS”) in order to determine the mineralogy and elemental constituents.

[0121] Once the mineralogy was determined, the minerals were then identified and labelled under an optical microscope. Then, flotation samples were blended with precalculated ratios of concentrate and tailings with three particle size distributions in order to simulate a range of possible slurry sample which might be expected to be present in the concentrator. These samples were then screened to fine fraction (< 38 pm) and coarse fraction (> 38 pm) prior to imaging with an optical microscope. The fine fraction images were labelled using a superpixel + classifier algorithm whilst the coarse fraction images were labelled using a SOLO v2 DL instance segmentation algorithm. The predicted particle size distribution and grade were then compared with the laboratory analysis. Details of each step are outlined in more detail below.

[0122] According to further embodiments other instance segmentation algorithms may be used. For example, according to an embodiment a YOLOv5 instance segmentation architecture may be utilised which is a modification of the detection architecture. In addition to a YOLOv5 object detection head a fully connected neural network such as ProtoNet may be utilised. The object detection head combined with the ProtoNet makes up a YOLOv5 instance segmentation architecture. The mineral detection software is always upgraded with the latest libraries and algorithms in the market.

[0123] Bulk ore sample and flotation reagents provided by a poly metal ore operation located in central NSW, Australia were processed in order to test the apparatus. The deposit contained copper, iron, lead, and zinc sulphides along with gold. A concentrator was utilised which had the flexibility to produce gold dore (i.e. a semi-pure concentrate of gold and silver) and separate copper, lead, and zinc concentrates depending on the stockpile blending. The chosen processing route was to produce a copper concentrate.

[0124] In order to demonstrate the applicability of the present invention to a range of different slurry compositions, a variety of particle size distribution and grades were generated with known mineralogy and particle size distributions (“PSD”). Instead of generating n samples that required n ICP analyses, low- and high-grade stockpiles were obtained for several size fractions and analysed. Then, by blending pre-calculated ratios of the known high- and low-grade flotation samples, intermediate samples were created as required for demonstrating the effectiveness of the present invention across a range of different slurry compositions.

[0125] Thus, the high- and low- grade samples were taken from the highest and lowest possible grade obtained in the process that were: (i) high-grade sample taken from flotation concentrate collected at 20 s residence time and (ii) low-grade sample taken from flotation tailings after 3 min of flotation. This was determined to be the time where true flotation stops because the froth bubbles were transparent and not loaded with minerals.

[0126] To determine the mineralogy grade, two bags of each size fraction of the flotation concentrate and tailings were randomly selected, pulverised and sent for XRPD analysis to determine the corresponding mineralogy and ICP-MS analysis to determine the Cu, Fe, Pb, S and Zn elemental contents.

[0127] Flotation procedure and sample list

[0128] The SAG feed sample (< 150 mm) was first crushed with a jaw crusher and recirculated with a screen until all particles were smaller than 3.35 mm. The particles were then blended, homogenised and split into 2 kg sub-samples for storage. Then, a grind calibration curve was conducted for the rod mill to determine the grind time required to achieve approximately P80 = 150 pm. The samples were grinded for several different times at 50% solids and screened through a 150 pm sieve until roughly 80% passing was achieved. This corresponds to about 8 min 30 s with a P80 of 145 pm.

[0129] Each 2 kg sub-sample was grinded for 8 min 30 s at 50% solids and then diluted to 35% solids for flotation in a 5 L Metso Denver Cell with the following conditions as recommended by Aurelia Metals: (i) 1 .8 mL / t mixture of isopropyl ethyl thiocarbamate (“IET”) and sodium diisobutyl dithiophosphate (“SDD”) - chalcopyrite and galena collector; (ii) 300 g / t zinc sulphate (ZnSO4) - sphalerite depressant; (iii) 100 g / t sodium metabisulfite (“SMBS”) - pyrite depressant; and (iv) 10 g / t methyl isobutyl carbonyl (“MIBC”) - frother.

[0130] The typical yield for the flotation concentrate at 20 s residence time was about 82.1 g and this was repeated ten times in order to obtain the required stockpile mass for each size fraction for sample blending. After sufficient samples of the concentrate and tailings were generated, they were combined and screened into < 38, 38-53, 53-75, 75- 106, 106-150 and >150 pm size fractions. Samples in each size fraction were riffled into 5 g bags to ensure they were homogenous.

[0131] In order to generate realistic samples for development of the technique, the samples’ cumulative particle size distribution (“PSD”) should have a characteristic “S” shape as if sampled from a concentrator stream. Therefore, PSD 1 (8 min 30 s) and 3 (11 min 15 s) were taken from the grind calibration test in Section 0, while PSD 2 (~9 min 53 s) was interpolated between using the Rosin-Rammler equation as an intermediate.

[0132] Table 1 below details the three PSDs which were used in order to simulate flotation samples. Size Fraction (pm) PSD 1 (%) PSD 2 (%) PSD 3 (%)

[0133] < 38 46.10 40.58 37.65

[0134] 38-53 11.67 11.04 10.61

[0135] 53-106 12.70 12.59 12.38

[0136] 106-150 11.69 12.38 12.59

[0137] 150-212 9.12 10.60 11.28

[0138] > 212 8.72 12.81 15.49

[0139] Table 1 Table 2 below shows the concentrate and tailings blend for each sample used. In particular, Table 2 shows sample IDs and their corresponding concentrate content and PSD e.g. sample 28 has 35% concentrate, 65% tailings and PSD 2.

[0140] Concentrate Development Dataset Validation Dataset

[0141] Content (%) PSD 1 PSD 2 PSD 3

[0142] 0 1 21 42

[0143] 30 6 - 48

[0144] 35 7 28 49

[0145] 40 8 29 50

[0146] 45 9 30 51

[0147] 50 10 31 52

[0148] 55 11 32 53

[0149] 60 12 33 54

[0150] 65 13 34

[0151] 70 14 35

[0152] 100 20 41 62

[0153] Total 11 10 9

[0154] Table 2 Table 3 below details the particle size distribution of the flotation concentrate and tailings. In particular, Table 3 shows particle size distribution of the flotation concentrate and tailings.

[0155] Size Fraction Flotation Concentrate Flotation Tailings

[0156] Mass Discrete Count Mass Discrete Count

[0157] (g) (%) (g) (%)

[0158] < 38 35.3 43.00 784.6 40.20

[0159] 38-53 19.8 24.12 159.6 8.18

[0160] 53-75 4.8 5.89 179.9 9.22

[0161] 75-106 6.8 8.24 235.3 12.06

[0162] 106-150 9.8 11.94 240.9 12.34

[0163] >150 5.6 6.82 351 .3 18.00

[0164] Sum 82.1 100 1951.6 100

[0165] Table 3

[0166] X-ray powder diffraction analysis

[0167] XRPD analysis of the samples was performed in order to identify the minerals present in a sample. It will be understood that XRPD involves analysing the diffraction of X-rays from the pulverised sample that depends on the mineral’s crystalline structure. According to various embodiments an instrument such as a XRD instrument with a Cu Ka source may be used.

[0168] Fig. 1 shows an example of XRPD spectra obtained for a concentrate sample with mineral peaks highlighted from Table 3 above.

[0169] Fig. 2 shows XRPD spectra obtained for the tailings sample with the mineral peaks highlighted from Table 2. The spectra is cropped for clarity purposes since the peaks at 13° and 27° went up to 10,000 counts. The bulk occurring minerals in the concentrate were chalcopyrite (CuFeS2), sphalerite (ZnxFe(i.X)S), galena (PbS), pyrite (FeS2), quartz (SiO2), and pennanite (Mn2+5AI(AISi30io)(OH)8). There were also traces of cubanite (CuFe2S3) and felbertalite (Cu2Pb6Bi8Si9).

[0170] For reasons of simplification cubanite and felbertalite were excluded from further consideration as they had low amounts. Also, for sphalerite it was assumed that x = 1 i.e. ZnS. Furthermore, all gangue minerals including quartz, pennanite and other gangue minerals were grouped which were not characterised in the tailings XRD sample. For mass calculations it was assumed that the density was 2700 kg / m3as most gangue minerals have similar density.

[0171] Inductively Coupled Plasma-Mass Spectrometry analysis

[0172] The elemental assay results of ICP-MS analysis is presented below. The calculations and assumptions used to convert the elemental assay results to mineral grades are also detailed below. In particular, Table 4 below shows the mineral balance of the concentrate and tails calculated from the mineral balance of the ICP results.

[0173] Sample Size Fraction Chalcopyrite Pyrite Galena Sphalerite Gangue (pm) (%) (%) (%) (%) (%)

[0174] Concentrate < 38 41.9 4.5 15.5 4.4 33.6

[0175] 38-53 57.5 10.3 15.1 5.5 11.5

[0176] 53-75 59.3 11.8 13.0 5.4 10.4

[0177] 75-106 57.2 10.7 10.2 5.0 16.9

[0178] 106-150 57.8 0.0 6.0 4.3 31.9

[0179] > 150 52.7 1.3 2.6 2.6 40.8

[0180] Tailings < 38 1.6 1.6 0.7 0.4 95.7

[0181] 38-53 1.3 2.9 0.5 0.5 94.8

[0182] 53-75 0.9 2.7 0.3 0.4 95.7

[0183] 75-106 0.7 2.5 0.3 0.3 96.1

[0184] 106-150 0.5 1.9 0.2 0.3 97.2

[0185] > 150 0.3 0.9 0.1 0.1 98.7

[0186] Table 4 Optical microscopy labelling

[0187] Once the mineralogy of the samples was known, samples were then observed under AN optical microscope and labelled to collect a dataset for development of the DL instance segmentation algorithm. Fig. 3 shows sample images of the major minerals under optical microscopy and a brief description of their appearance.

[0188] According to various embodiments flotation slurry samples were fitted into a

[0189] 10 mm x 10 mm x 45 mm cuvette. The total mass of each sample was about 3 g with the samples split to < 38 pm (fines) and > 38 pm (coarse) as the fines and slimes obscured the optical microscopy images when combined together with the coarse fraction. Details of the methodology used to collect and label the optical microscopy images of the fines and coarse fractions of the flotation sample are presented below.

[0190] Software and hardware

[0191] An algorithm was developed using Python 3.7.9 with a i7-9700 Central Processing Unit (CPU) and NVIDIA GeForce RTX 2080 Mobile 8 GB Graphics Processing Unit (GPU). A DL algorithm was used which was the SOLO v2 implementation based on MMdetection (v1.0.0), PyTorch v1.4.0 and NVIDIA CUDA 10.1 made available on GitHub. A superpixel algorithm used was a CUDA implementation of the SLIC superpixel algorithm (cuda-slic) on Python.

[0192] As detailed above, it will be understood that the algorithms disclosed above are examples of suitable algorithms which may be utilised. However, the present invention is not limited to the specific algorithms identified above and may be implemented using other algorithms as also discussed above.

[0193] Fig. 4A shows an optical microscopy apparatus which was built out of steel slotted racks. According to various embodiments a motorised platform is provided which holds a microscopic objective connected to an active pixel sensor capable of capturing 1920 x 1080-pixel resolution images and video at 30 image / s. The microscopic objective used was 10X Compact Adjustable Objective. This type of microscopy is known as reflected light microscopy wherein the illumination is from the side of the objective rather than opposing in transmitted light microscopy.

[0194] As can be seen from Fig. 4B cuvette samples were placed into a cuvette holder and held in place while the sensor captures digital images as the platform moves from bottom to top using a control board. The pixel sensor is connected to a computer where the images were saved and processed later. Fig. 5 shows example images collected from the fines and coarse samples at varying concentrate blends. The chalcopyrite concentration can be seen from barely present at 0% concentrate (1.05% chalcopyrite) to representing half the minerals present in the images at 100% concentrate (51.1% chalcopyrite) for both fines and coarse fractions. This demonstrates that according to various embodiments minerals may be identified in-slurry. For the 30 samples, there was a total of 604 and 637 images for the coarse and fines respectively. According to various embodiments the slurry images were labelled for instance segmentation algorithms and the subsequent methodology used to calculate the particle size distribution and grades from the algorithm outputs will be described in more detail below.

[0195] Coarse dataset (> 38 urn) and associated algorithm

[0196] For the coarse samples, DL instance segmentation was used as the boundaries and minerals are easier to distinguish and label. 117 images from 6 coarse samples were labelled fortraining purposes. An additional 24 photos were randomly selected as a validation and test set from other coarse samples to validate and quantify the algorithm performance.

[0197] Table 5 below summarises the distribution of mineral classes for each dataset used to develop the coarse algorithm. By comparing the ground truth and predictions of the trained algorithm on unseen (test) images, we can comment on the generalisation ability of the algorithm. Full details of the evaluation metric, Average Precision, APcoco and how it is calculated can be found in Koh et al., (2021) and Everingham et al. (2010). These images were labelled using VIA annotation software (Dutta & Zisserman, 2019). In particular, Table 5 below shows the breakdown of number of instances by class in the datasets.

[0198] Dataset Quartz Chalcopyrite Pyrite Sphalerite Galena Average

[0199] (instances / image)

[0200] Train (117 images) 3,024 1 ,071 255 109 236 40.1

[0201] Validation (12 223 89 11 4 30 29.8 images)

[0202] Test (12 images) 176 108 21 5 29 28.3

[0203] Total (141 images) 3,243 1 ,268 287 118 295 37.0

[0204] Table 5

[0205] Details of the coarse algorithm development and results are detailed below. Coarse DL algorithm details

[0206] A DL instance segmentation algorithm was used for transfer learning. In particular, the DL algorithm used was the pre-trained SOLOv2_Light_512_DCN_R50 with the default configuration. It will be understood that other algorithms may equally be used.

[0207] According to various embodiments the task may be separated into two separate tasks wherein two different instance segmentation algorithms may be utilised as this has been found to result in improved performance. One algorithm may be utilised to identify particles and the other algorithm may be utilised to classify particles consequently.

[0208] However, embodiments are also contemplated wherein instance segmentation may be performed using a single algorithm. It will be understood, however, by those skilled in the art that using a single algorithm might result in fewer particle detections when all classes have low confidence. This could also be due to having a low number of training examples. As a result, it is preferred to adopt a two-algorithm approach. In particular, with a two-algorithm approach, the first algorithm may be arranged to focus upon segmenting particles regardless of mineral class which may yield a better accuracy for size distribution estimation. Similarly, with this algorithm the unknown particles may be classified as gangue instead of not detected at all. According to various embodiments images of detected particles may be rescaled into an array of 192 x 192 pixels which may then be classified by the second algorithm. It will be understood that utilising a 192 x 192 image size represents a trade-off between speed and accuracy.

[0209] The cross-sectional area predictions were used to estimate the sieve diameter which can be compared to the laboratory size analysis. The volume was calculated as a sphere of the equivalent cross-sectional area and then multiplied by the density of the mineral to obtain the mass for particle size distribution and grade calculations.

[0210] Table 6 below summarises the accuracy metric of the two SOLO v2 algorithms used for the coarse images. Note that these reflect accuracy in replicating labelling of an expert mineralogist, it is impossible to know for certain whether the ground truth label was correct as it is not feasible to find and measure the submerged particles for each image. Table 6 below shows the average precision of the coarse algorithms for particle detection and consequent particle identification.

[0211] Coarse Algorithm 1 . Particle Detection

[0212] Class Particle

[0213] APcoco 60.1% Coarse Algorithm 2. Particle Identification

[0214] Class Gangue Chalcopyrite Pyrite Sphalerite Galena

[0215] APcoco 89.4% 83.1 % 44.3% 30.2% 30.5%

[0216] Table 6

[0217] For context, the published SOLO v2 algorithm achieved 35.0% APcoco for the COCO dataset. The benchmark COCO dataset had a total of 80 classes and averaged 7.7 instances / image. Although the dataset which was tested according to various embodiments had considerably fewer classes, the average number of instances per image was still significantly higher for Particle Detection at 37.0 instances / image. Therefore, 60.1% APcoco for the Particle Detection algorithm demonstrates the effectiveness of the approach according to various embodiments.

[0218] The outputs of the instance segmentation algorithm were particle boundaries (and hence size in pixels) and classification of each particle.

[0219] Fig. 6 shows the sample output of the coarse algorithm.

[0220] The cross-sectional area predictions were used to estimate the sieve diameter which can be compared to the laboratory size analysis. The volume was calculated as a sphere of the equivalent cross-sectional area and then multiplied by the density of the mineral to obtain the mass for particle size distribution and grade calculations.

[0221] Fines dataset (< 38 urn) and associated algorithm

[0222] The particle boundaries for the fine images could not be labelled for developing a DL algorithm in a feasible amount of time as the particles were too small with the current resolution of the microscope. Furthermore, the translucent fine quartz formed a grey matrix that made it difficult to distinguish boundaries unlike in the coarse fraction. To overcome this, the SLIC superpixels algorithm was used which is an unsupervised clustering algorithm to cluster similar groups of pixels (superpixels) together. The superpixels are then classified with a separate algorithm according to each superpixels’ spectra.

[0223] Fine algorithm details

[0224] The SLIC superpixel clustering algorithm requires two inputs, the number of segments n and the compactness parameter k. For n, high n was favoured (oversegmenting) as neighbouring instances could be combined again later, if the boundaries were correct. On the other hand, low n (under-segmenting) caused superpixels to not segment boundaries correctly because there were insufficient superpixels. It was found that n = 10,000 was effective i.e., for 1920 x 1080 = 2,073,600 each superpixel would have = 207 pixels (14.4 x 14) or roughly 3.5 x 3.5 pm. The compactness parameter controls the shape superpixels, where higher values yield more regularly shaped or square superpixels. It was found that k = 10 yielded acceptable results. Once the image was segmented, each superpixel was classified according to their pixel colour value. If the superpixel does not fit into any of the thresholds, it was classified as gangue.

[0225] Fig. 7 shows instance segmentation algorithm steps on a single image of the fine fraction.

[0226] As the coarse and fines fraction were imaged separately in different cuvettes, the algorithms needed to be able to estimate the mass of the samples from the image to determine the full particle size distribution. To achieve this, the concept of image densities was used i.e., the number of images for each sample multiplied by a calibrated image density (grams / image) to obtain the mass contributed by the fine / coarse fraction. The image density is a function of the gangue grade that was calibrated with the development dataset.

[0227] Validation of methodology and sampling error

[0228] The model development and calibration utilised images of samples from PSD 1 and PSD 2 only. For an unbiased test, the final algorithm was used on PSD 3 to validate the entire methodology. Each of the samples was measured once (one pass measurement) and compared with the ground truth from the ICP-MS analysis. Then, to estimate the uncertainty of the PSD and mineralogy from sampling, three samples of PSD 3 (0%, 50% and 100% concentrate) were measured an additional four times (total of five times, five pass measurement). These samples were shaken, imaged, and measured with the algorithm for each repeat. By comparing the predictions of these three samples, we have quantified the sampling error for each individual measurement (one pass) compared to the error of the average of five measurements (five pass).

[0229] Results

[0230] The average time taken for the entire algorithm to process each sample was determined as being 5.36 min (see Table 7 below). Table 7 below shows the time taken for the methodology to measure the sample once. Step Average Time Taken Algorithm Speed (minutes) (seconds / image)

[0231] Coarse Algorithm 0.75 2.16

[0232] Fine Algorithm 4.56 13.7

[0233] Processing Algorithm 0.05 0.075

[0234] Outputs

[0235] Total 5.36 N / A

[0236] Table 7

[0237] Particle size measurement results

[0238] The models were calibrated to measure P50. The actual P50 was known as the cuvette samples were created by weighing and combining masses of screened flotation concentrate and tailings.

[0239] Fig. 8 shows a parity plot of the one pass P50 measurements of 30 samples. The true P50 value for the development dataset was PSD 1 = 56.1 pm and PSD 2 = 50.1 pm, while the validation dataset was PSD 3 = 43.0 pm (all with very minor spread on the x-axis due to sample preparation of the masses). As expected, there is a spread of measurement results for each sample (y-axis) like any measurement technique. The average error for one pass measurement across the 30 samples was ± 2.32 pm for the 3 different PSD samples. The P50 measurement results on the validation dataset has a low spread, indicating that the methodology is feasible even for samples outside of the development dataset range.

[0240] Fig. 9 shows the comparison of one pass measurements and five pass (five repeated measurements) on three samples in the validation dataset. These corresponded to the 0%, 50% and 100% concentrate ratio. The results show that the average error is ± 2.84 pm across each individual measurement due to sampling variances. However, if the average of five measurements is taken, the error reduces to ± 2.14 pm. As expected, there is a trend of decreasing error as more measurement samples are taken to reduce the sampling variance. This would be a trade-off of time required per sample and the desired measurement accuracy.

[0241] Mineralogy characterisation results

[0242] The accuracy of the mineralogy characterisation measurement is determined by comparing measured mineralogy to the blended ratio of ICP-MS assay values. Fig. 10 shows the comparison of the measured chalcopyrite, gangue and other sulphides (e.g. pyrite, sphalerite and galena) against the assay grades for the 30 samples. Note the assay grades are a mass grade, whereas the measured mineralogy in this methodology is a surface measurement calibrated to the ICP-MS mass grade in the development dataset. The average error of the mineralogy characterisation is 5.21%, 5.37% and 3.01% for gangue, chalcopyrite and other sulphides respectively. The validation dataset results are all very close to the assay grades, indicating that the methodology is successful on unseen samples. For the other sulphides grades, the algorithm tends to under-predict which is possibly due to the lower accuracy of the coarse algorithm as the low confidence classes were defaulted to gangue.

[0243] Fig. 11 summarises the effect of repeated measurement on the same three samples in the validation dataset. In particular, Fig. 11 shows parity plots comparing the one pass and five pass measurements on three samples from the validation dataset. Through five measurements, the mean absolute error for the mineralogy characterisation significantly decreases from 5.56% to 1 .67% for gangue, 4.20% to 1 .33% for chalcopyrite and 2.86% to 2.13% for other sulphides. As the errors look normally distributed around the parity it is unlikely there is an error in the calibration but from various sources in the methodology.

[0244] Sources of error in methodology

[0245] When comparing the results from the methodology and the assay results, the total error, ETcan be broken down into five sources of error:

[0246] ET = Es + E2D + EML + EOM + Ep + EA (Eqn. 1) wherein Es is the sampling error, E2D is the 2D to 3D conversion error, EML is the Machine Learning algorithm error, EOM is the optical microscopy presentation error, Ep is the laboratory sample preparation error and EA is the assumptions error.

[0247] The various different sources of error will now be considered in further detail below.

[0248] Es sampling error

[0249] The sampling rate of the methodology is the amount of surface that is measured by optical microscopy in relation to the total surface area of particles in the cuvette. This rate can be approximated by assuming all particles are perfect spheres with diameters equal to the mean size for any particle size fraction. Then, the density is the weighted average according to the ICP-MS analysis to determine the number of spheres, and consequently total surface area of those spheres. This represents the approximation of the total surface area of sample that can theoretically be measured forthat cuvette. The sampling area would be the field of view multiplied by the number of images taken for that sample. This yields a sampling ratio of about 1 in 50,000. The sampling error represents the error in estimating the population from the given sample. This explains the high variance in one pass sampling. However, our results show that averaging 5 measurements (effectively improving the sampling ratio to 1 in 10,000) significantly reduces the errors for P50 and mineralogy characterisation.

[0250] E2D 2D to 3D conversion error

[0251] The 2D to 3D conversion error describes any part of the methodology of estimating the 3D properties of the particle when only measuring the 2D surface of the particle. This includes: (i) the estimation of mass grade to compare with the assay results; (ii) the estimation of the particle’s mineralogy given only the measurement of a sub-section of the total surface area of the particle; (iii) the estimation of the particle’s shape and hence mass by observing the surface; and (iv) the estimation of sieve diameter.

[0252] All the above requires some calibration parameter to convert the 2D measurement to 3D which is a limitation of the optical microscopy methodology. However, surface grade measurements should in theory be more favourable for flotation modelling or monitoring as it measures mineral liberation compared to assay grade. However, for a ground truth of surface liberation would require comparison measurements of QEMSCAN and MLA.

[0253] 2D sieve diameter

[0254] A conjecture is proposed to estimate the bounds of the sieve diameter from any 2D projection of a particle. The sieve diameter of a particle dsieve is the size of the smallest sieve which that particle may pass through. More formally, suppose if the coordinates of boundaries of any 3-dimensional particle (x, y, z) can be defined in an orthogonal 3-dimensional space (X, Y, Z), then the minimum bounding box (cuboid in this case) of the 3-dimensional point set which completely encompasses the particle has dimensional lengths { / = max(x) - min(x), j = max(y) - min(y), k = max(z) - min(z)} such that I < j < k.

[0255] Fig. 12 is an illustration of a hypothetical cuboid bounding a particle and the sieve diameter definition.

[0256] As the orientation of the particle changes, the dimensions of the cuboid do not change but the two-dimensional rectangular projection of the cuboid on the sieve changes. Therefore, there exists an orientation of the particle such that j, the larger side of the two-dimensional rectangular projection is minimised. This is equal to the sieve diameter. To illustrate this concept, consider a cuboid with dimensions dx, dy, dz along the X, Y and Z axes respectively such that dx <dy « dz. Hence, dsieve would be found along the X-Y plane.

[0257] Fig. 13 also illustrates two possible sieve diameters for a cuboid.

[0258] It can be proven that: and:

[0259] Theorem 1

[0260] For a 3-dimensional object bound by a cuboid dx<dy«dz, dsieve does not necessarily equal dy and requires rotation of the object along the Z axis on the centre of the object to determine dsieve. This may be referred to as the 2nd minimum diameter.

[0261] However, it is desired to estimate dsieve just by observing a 2-dimensional projection of the 3-dimensional object. Consider a 2-dimensional projection area of a 3- dimensional object onto any plane. If the unobserved cuboid dimension is the longest, i.e. dz Theorem 1 deduces that dsieve can at most be dy if dx < dy « dz. However, from the definition of dsieve, this is still true even if the bounding cuboid has length dz = dy. This provides the upper bound of dsieve just by observing a 2-dimensional projection. Then, we consider the case where the non-observed bounding cuboid dimension is the shortest dimension, dx. It will be understood that dy and dz can be obtained through rotation. Then, from the definition of dsieve it will be understood that no matter how small dx is, dsieve = dy. Therefore, conjecture 1 as detailed below is proposed.

[0262] Conjecture 1

[0263] By observing any 2-dimensional projection of a 3-dimensional object, it is possible to evaluate the bounds of dsieve. If the shortest diameter d1 , and the 2nd minimum diameter, d2 of the 2-dimensional projection are known, the dsieve is bound by these diameters, d1 < dsieve d2.

[0264] This is a conjecture based on observation rather than rigorous proof. It can be shown to be true as above for regular polygons like a rectangular cuboid not a generalised proof for any volume. According to various embodiments d1 and d2 may be obtained by rotating the cross-sectional area in 5° increments (Theorem 1). Once these bounds are known, a calibration factor f that is between 0 and 1 may be used where dsieve = fd2 - d1 + d1 . This is similar to a shape factor and is calibrated based on the dataset.

[0265] EML Machine Learning algorithm error

[0266] The ML algorithm error is related to the errors from utilising algorithms to replicate the ground truth labels by the expert mineralogist. A perfect algorithm would have APcoco = 100% for all classes. According to various embodiments a high target AP that is consistent for all classes. Although not done here, it should be possible to compare the measurements of samples labelled by an expert mineralogist, i.e. EH - Human Labels:

[0267] EH=Es + E2D + Eo + Ep + EA (Egn. 4) with the error arising from the methodology ETto guantify EML through subtraction. From various results the coarse algorithm has lower accuracy for the pyrite, sphalerite and galena minerals due to having less training instances compared to gangue and chalcopyrite. The most time-consuming part of a DL implementation is typically the collection of a large, and relatively clean dataset for model development.

[0268] Eo optical microscopy presentation error

[0269] The optical microscopy presentation error encompasses errors arising from limitations of segmenting and distinguishing minerals through the optical microscope. This includes: (i) when particles are stacked, especially guartz (translucent) being at the foreground, making it difficult to distinguish particle boundaries; (ii) limitation of detecting minerals by optical reflectance images. Some minerals can be confused or indistinguishable in the optical spectra. Example, the highly luminous surfaces that reflect a lot of light could be pyrite or chalcopyrite. Distinguishing mineral grains would reguire higher resolution or more spectral wavelengths; (iii) when particles lie on the bounds of the camera image, making it impossible to estimate the size of the particle. This can be minimised by having a field of view much greater than the top size; (iv) detecting only one mineral class per particle; and (v) motion blur.

[0270] EPlaboratory sample preparation error

[0271] The laboratory sample preparation error describes errors that could affect the ground truth mineralogy in the submitted sample and the sample in the measured cuvettes. This includes: (i) the homogenisation of the samples which were ultimately sent for characterisation and cuvette sample creation; and (ii) the amount of sample used for ICP-MS analysis relative to the bulk sample. In ICP-MS, there is a dilution factor involved during sample preparation.

[0272] To quantify this effect, two random samples from the homogenised bags were sent for ICP-MS analysis, and the assay results had minimal deviation. Out of all the sources of error, this is assumed to be least or insignificant relative to others.

[0273] E assumptions error

[0274] This error includes all the errors contributed by assumptions which otherwise would have made it unfeasible or require significantly better hardware and more time. Among them are: (i) limiting the mineral detections in this study to gangue, chalcopyrite, pyrite, sphalerite, and galena. Without limiting the detections for the scope of this study, it would be impossible to collect a suitable dataset size for each of the minerals present in the sample. As most gangue minerals like clay and chlorites have density close to 2700 kg / m3, the error contributing to size and modal mineralogy is minimal.

[0275] Embodiments are contemplated wherein a large library of a labelled mineral images dataset is obtained in order to reduce this issue and increase the DL accuracy of each class; (ii) assuming the fines fraction consisted of spheres with a diameter of 19 pm to evaluate the mineralogy. Without this assumption, it would not be possible to use the density of minerals that certainly influences the particle size distribution and mineralogy grade. Comparing ratios of areas would not provide an accurate measurement of modal mineralogy. Various embodiments are wherein the concept of image density may be used to combine the mass of fine and coarse fractions. As the samples were measured separately, this was necessary to provide the full particle size distribution and grades.

[0276] According to various embodiments it is apparent that deep learning instance segmentation can be utilised on optical reflectance microscopy RGB images as a tool for in-slurry characterisation of mineralogy in copper ore samples.

[0277] It will be understood that a large deep learning model with millions of parameters can be trained on a smaller minerals processing image dataset through transfer learning. According to various embodiments reagents which were used for chalcopyrite beneficiation naturally led to an imbalanced dataset for the other sulphides like pyrite, galena and sphalerite. Nevertheless, embodiments are contemplated wherein a standardised approach to the imaging procedure is provided in order to develop a library of labelled images of known minerals and degree of liberation to further improve the accuracy of the proposed methodology. A person skilled in the art will understand that, for example, images of galena particles as disclosed in the present application fragment into cubic crystals and exhibit rough surfaces. The colour appearance is also similar with their study and how pyrite / chalcopyrite is distinguished, where pyrite exhibits the highest reflection of blue wavelength and chalcopyrite less, appearing green-ish yellow.

[0278] Although the experimental design covered a range of % Cu between 0.1% and 20%, further embodiments are contemplated wherein a full range of possible grades may be analysed in a chalcopyrite concentrator e.g. from final concentrate (-29% Cu) to final tailings (<0.1% Cu).

[0279] According to various embodiments different light intensity beams may be used to identify voidage directly from images rather than prior knowledge of the sample. Such an approach is beneficial in terms of improving the accuracy on translucent quartz minerals.

[0280] It will be apparent that a methodology for characterising modal mineralogy of multiple particles in-slurry has been demonstrated using machine vision and optical microscopy for a copper sulphide ore. According to various embodiments particles in the flotation slurry sample can be measured using an optical microscope with Red Green Blue (RGB) images as long as the fine (< 38 pm) particles were separated from the coarse (> 38 pm) particles and measured separately. Through these images, a DL model and a SLIC superpixels + pixel classifier model were utilised to perform instance segmentation on the coarse and fine particles, respectively. The methodology achieved the following measurement accuracy on unseen flotation slurry samples with Cu grades between 0.1% and 20% through one pass: (i) P50 at 43.0, 50.1 and 56.1 pm with ± 2.32 pm error; (ii) gangue grade with ± 5.21 % error; (iii) chalcopyrite grade with ± 5.37% error; (iv) Other sulphides grade with ± 3.01% error.

[0281] Through five repeated measurements of the same samples, the methodology error improves: (i) P50 at 43.0 pm from ± 2.84 pm to ± 2.14 pm; (ii) Gangue grade from ± 5.56% to 1 .67%; (iii) Chalcopyrite grade from ± 4.20% to 1 .33%; and (iv) other sulphides grade from ± 2.86% to 2.13%.

[0282] At this current state, the methodology takes about 5 min for one pass measurement.

[0283] Calibrating image densities for coarse and fine fractions

[0284] Figs. 14 and 15 show the polynomial equation calibration for calculating image density given measured gangue grade for fines and coarse respectively. For comparison, the image density at 0% concentrate (100% tailings) 0.056 g / image and 0.085 g / image for fine and coarse respectively yields a bulk density of about 1 ,200 kg / m3to 1 ,800 kg / m3for the fine and coarse respectively. This indicates that the packing voidage is about 0.55 for fines and 0.68 for coarse, which is within the possible values which is 0.40-0.85 for fine sand. In particular, Fig. 14 shows image density calibration results for the fines fraction and Fig. 15 shows image density calibration results for the coarse fraction.

[0285] XRPD analysis

[0286] In order to identify the minerals present in a sample, the peaks of the spectra may be compared with known mineral peaks from a database like the International Centre for Diffraction Data (ICDD). Only the three highest intensity diffraction angles are shown for each mineral (see Table 8 below). The method is semi-quantitative as the higher peaks indicate a higher grade and can only be compared to the counts of each other. However, to quantify the grades require comparison with peaks from a known grade sample.

[0287] Table 8 below shows 20 Diffraction Angles of Bulk Minerals detected from XRPD in this study.

[0288] Mineral Formula 20 20 20

[0289] Name Diffraction Diffraction Diffraction

[0290] Angle 1 (°) Angle 2 (°) Angle 3 (°)

[0291] Quartz SiO226.65 20.85 50.14

[0292] Chalcopyrite CuFeS229.45 49.10 57.91

[0293] Galena PbS 30.07 25.96 43.06

[0294] Sphalerite ZnxFe(i.X)S 28.56 47.51 56.29

[0295] Cubanite CuFe2S327.74 25.50 27.59

[0296] Chlorite Mn2+5AI(AISi3Oio)(OH)812.39 6.18 59.94

[0297] (Pennanite)

[0298] Felbertalite Cu2Pb6Bi8Si950.95 52.42 44.58

[0299] Pyrite FeS256.29 33.04 37.07

[0300] Table 8

[0301] ICP-MS analysis results

[0302] Table 9 below summarises the average values obtained for the ICP-MS analysis. Based on the assumption made in Section 2.1 .2, the chalcopyrite grade can be calculated by assuming all Cu comes from chalcopyrite only. The same can be done for galena and sphalerite using Pb and Zn respectively (assuming Sphalerite has no Fe, i.e., ZnS). Then, the pyrite grade can be calculated based on the S grade and subtracting the Fe from chalcopyrite. The rest of the mass is assumed to be gangue. Due to the accuracy of the ICP-MS analysis, this methodology yielded negative mass for pyrite in the 106-150 pm for the concentrate. Hence, the pyrite grade was assumed to be 0 in this size fraction.

[0303] Table 9 below shows a summary of the elemental grades by size obtained from the ICP Analysis.

[0304] Sample Size Fraction (pm) Cu (%) Fe (%) Pb (%) S (%) Zn (%)

[0305] Concentrate <38 14.500 22.450 13.450 20.600 2.980

[0306] 38-53 19.900 27.950 13.100 29.450 3.720

[0307] 53-75 20.550 29.350 11.250 30.600 3.650

[0308] 75-106 19.800 28.650 8.845 28.700 3.350

[0309] 106-150 20.000 28.650 5.180 21.650 2.895

[0310] +150 18.250 24.850 2.275 20.300 1.755

[0311] Tails <38 0.570 8.715 0.636 1.635 0.272

[0312] 38-53 0.455 6.365 0.441 2.215 0.316

[0313] 53-75 0.319 6.110 0.303 1.920 0.262

[0314] 75-106 0.232 6.030 0.264 1.745 0.224

[0315] 106-150 0.167 5.590 0.152 1.290 0.170

[0316] +150 0.088 5.710 0.077 0.613 0.084

[0317] Table 9

[0318] According to various embodiments a module is disclosed which uses a Machine Learning (“ML”) algorithm and a sampling device to estimate the modal mineralogy of a slurry stream in real-time. According to embodiments a system is disclosed which automatically prepares samples taken from a slurry stream, such as a conventional flotation feed and routes them to the measurement chambers. According to various embodiments the ML algorithm continuously analyses samples from the slurry stream to report mineralogy by size, particle size distribution, and percent solids in that stream. The module may, for example, be used to test ore samples taken from copper operations in countries such as Australia.

[0319] Test data was obtained and the outcomes generated by the device during this trial were then compared with the results of chemical assays conducted on the representative samples taken before the trial. There was a strong correlation between the assay results and mineral compositions reported by the module.

[0320] Mineral separation and concentration usually occur after an ore is processed through a grinding circuit and minerals are almost liberated. There are several parameters which impact the separation efficiency of the process. One of the most used separation processes in the industry for base and precious metals is froth flotation.

[0321] Fig. 16 shows influential parameters on flotation process. In froth flotation, there are many parameters which affect the process performance and can be grouped into chemistry, operation and equipment.

[0322] Equipment and chemistry components are selected or controlled according to the operations conditions such as feed flow rate, mineralogy, particle size distribution, and density in a separation process. Process operators typically adjust operating conditions like the reagent type and dosage, air flow rate, and slurry level to achieve the highest metal production at the target grade. Because of this, mineralogy plays the most important role as it defines the slurry chemistry, liberation size, and residence time of the process. Mineralogy of an ore includes, valuable and gangue mineral distribution, minerals liberation degree, mineral association, and mineral grain size which all directly impact the flotation performance. Other influential components like pH, feed rate, feed size distribution, slurry density, air flowrate, and slurry chemistry can be measured and controlled online. However, choosing appropriate set points forthose parameters to maximise separation and recovery strongly depends on the feed ore mineralogy. The impact of slurry chemistry on three major industrial copper minerals (chalcocite, bornite, and chalcopyrite) flotation kinetics and recovery has been investigated and it has been demonstrated that even within major copper minerals in a concentrator, process optimisation could result in enhanced metal yield in an operation.

[0323] Fig. 17 shows the impact of NaHS dosage and pH level on chalcocite, bornite and chalcopyrite flotation recovery.

[0324] It is known that galena particles in the middle size fractions which are fully liberated float faster than particles in the coarser or finer size fractions. Flotation kinetics was even slower for the galena particles which has association with gangue minerals. It is known that gangue minerals such as clays and pyrite can impact the flotation performance. Identifying those minerals and their grades can be beneficial in determining the optimal reagent suite (depressant or dispersant type and dosage) for maximising recovery and separation efficiency in a flotation process.

[0325] Several methodologies have been developed and adapted to detect and determine the composition of minerals since 1669. Fig. 18 shows a history of mineral detection and composition analysis.

[0326] In the early days, contact and reflecting goniometers were invented to observe and study mineral crystals and their uniform geometrical shapes using telescopes. At the end of the 19th century, X-ray machines were invented which was also used in mineralogy studies. The solid-state X-ray detector or energy-dispersive spectrometer (EDS) was developed in the late 1960s and rapidly found use as electron-beam instruments because of its speed in collecting and simultaneously displaying x-ray data from a wide energy range. Scanning Electron Microscopes (SEM) with energy dispersive X-ray (EDS) was also wildly deployed in 80s and 90s and still in use until today. Historically, the mineral detection process was a qualitative method due to the nature of the measurement methods mentioned here. Therefore, qualitative methods known as modal mineralogy such as QEMSCAN and Mineralogy Liberation Analysis (MLA), and more recently automated solutions like ZEISS Mineralogic and TESCAN Integrated Mineral Analyzer (“TIMA”) were developed in the last few decades.

[0327] However, the measurement techniques mentioned above require special samples to be prepared, such as polished sections, resulting in relatively lengthy turnaround time between sampling and analysis results. Because of this, they cannot be implemented as an online measurement method in their current format.

[0328] Table 10 below shows the time required to measure key sample attributes after collecting a sample. Drying and preparation of a representative sub sample is not included in the timing. Table 10

[0329] Among these methodologies, only elemental analysers like X-Ray Fluorescence (“XRF”) and Laser Induced Breakdown Spectroscopy (“LIBS”) are suitable for online process monitoring and control. These methodologies provide an estimation of bulk mineral distribution through a mass balance of the elemental mass measurement through calibration of known composites.

[0330] However, the elemental analysis lacks mineral association and liberation degree which are also crucial information in flotation performance. Furthermore, the online XRF has poor accuracies for lighter elements such as Ca and S which makes it difficult to identify clay minerals and differentiate the copper sulphide minerals from each other. In contrast, LIBS does not have this limitation but is less accurate for heavier elements like copper.

[0331] According to various embodiments a sampling and measurement apparatus is provided which is capable of conducting modal mineralogy analysis on a sample collected from a slurry stream or tank. The apparatus and method are particularly applicable in terms of optimising recovery and yield of separation processes in real time and on-line. In particular, the apparatus may be located on site at a processing facility so that a slurry can be analysed whilst being processed rather than a sample being sent to a laboratory for analysis as in conventionally performed.

[0332] As is discussed in more detail below, according to various embodiments the illumination system was further developed and improved in order to provide improved detection and differentiation of shiny minerals such as sulphides. The camera sensor was further developed and changed from a roller to a global shutter in order to address microscope automated movements and shakes. The resolution and optical set up was also modified. According to various embodiments the apparatus utilises software which uses different libraries and structures for learning and mineral detection which results in improved operating times. According to various embodiments 2D robotic arms and software were used rather than an electrical actuator. Advantageously, the robotic arms eliminate any need to make manual adjustments. Also, an algorithm for the robotic arms was developed which makes auto scanning possible.

[0333] Three integrated components are utilised in order to determine the modal mineralogy of slurry streams in real time i.e. a sample collection and preparation unit, a robotic optical microscopy scanning unit and modal mineralogy software for sample analysis. The methodology was developed and validated with a sulphide polymetallic ore sample obtained from a poly-metal mine in New South Wales, Australia. The samples were processed in a laboratory scale flotation cell to generate high-grade concentrate and low-grade tail. Then, flotation samples were blended with pre-calculated ratios of concentrate and tailings with three particle size distributions to create a range of possible samples in the concentrator.

[0334] To collect a dataset for the mineral recognition software, random samples of the homogenised concentrate and tailings were analysed with X-ray Powder Diffraction (“XRPD”) and SEM + EDS to determine the ground truth mineralogy and elemental constituents. The bulk occurring minerals in the concentrate were chalcopyrite (CuFeS2), sphalerite (ZnxFe(i.X)S), galena (PbS), pyrite (FeS2), quartz (SiO2) and pennanite (Mns+Al(AlSi3O10)(OH)8). There were also traces of cubanite (CuFe2S3) and felbertalite (Cu2Pb6Bi8Si9) which were neglected in the analysis. Once the ground truth mineralogy was determined, a representative sample for each mineral were cross-referenced, identified, and labelled under the optical microscope.

[0335] Finally, to validate the modal mineralogy methodology, tailing and concentrate samples were combined in different portions from low grade (1 .05% chalcopyrite) to 100% concentrate (51.1% chalcopyrite) for calibration and testing. The samples were sent for ICP-MS analysis to determine the Cu, Fe, Pb, S, and Zn elemental contents. The system and method according to various embodiments which may be known as “MinDet” is configured to identify minerals in each image which was converted to elemental assay using their mineral composition, the surface of the image covered by the mineral, and the mineral density. Then, the quantified results are compared with the ICP-MS results which is quantitative as well.

[0336] As discussed above according to various embodiment a slurry sample such as flotation feed and products may be separated into fine particles (< 38 pm) and coarse particles (> 38 pm) prior to imaging with an optical microscope in-slurry. To achieve this in continuous mode, a desliming cyclone may be provided.

[0337] According to various embodiments the underflow and overflow of the cyclone may be arranged to flow through observation towers which may consist of a chamber with a viewing window. The sampling unit and observation tower may be designed and printed on a 3D printer. The apparatus may be run in either a batch or continuous mode, wherein the continuous mode is for use in the operation plant and the batch mode / bench top version is for use in metallurgical laboratories.

[0338] The apparatus may comprise a viewing chamber that exposes the slurry sample for photography with a motorised or robotic optical microscope. This may be coupled with an adequate illumination set up designed for scanning the particles near the surface of the viewing chamber. Depending on the particle and mineral grain sizes, the adequate lens specification and digital sensor resolution is selected for the application. For this ore type, a 10X magnification lens and a high-definition (1920 x 1080-pixel at 472 pm x 266 pm) video at 30 frames / second rate sensor was used. This analysis provides mineral mix, elemental grade and stream particle size distribution (“PSD”) in near real time for use in concentrator set point decisions and control systems.

[0339] Fig. 19 shows a sampling and scanning unit according to various embodiment wherein the compartments comprise a control unit 1 , a sampling unit 2, a digital microscope 3, a 3D traverse 4, automatic valve(s) 5 and input / output streams 6.

[0340] Modal

[0341] Two different methodologies were applied to analyse the fine and coarse fractions. The fine fraction images were labelled with the superpixel + classifier algorithm, while the coarse fraction images were labelled with the instance segmentation algorithm. Fig. 20 shows the software user interface and the information extracted from the sample during the optical microscopy scanning process. The apparatus according to various embodiments can detect particles in-slurry using optical reflective properties, identify the boundaries of particles (for sizing) and classify the minerals (for surface grade calculation) within slurry images, estimate the number of images required for a representative sample to obtain an acceptable accuracy for particle size distribution and mineralogy.

[0342] Fig. 20 shows a screenshot of the machine learning software and user interface under operation according to various embodiments.

[0343] Deep Learning (DL) is a subset of Machine Learning (ML) that utilises deep neural networks to accomplish tasks like machine vision, autonomous driving and natural language processing. DL can learn complex features to distinguish minerals and boundaries but typically requires large and clean labelled dataset known as supervised learning. However, it is possible overcome this requirement by utilising pre-trained models on other datasets and tweaking the model parameters for a study with a smaller dataset (also known as transfer learning).

[0344] When collection of a labelled dataset of suitable size is not feasible, it is still possible to segment images through unsupervised learning. Such an example would be superpixel techniques which groups pixels according to common characteristics. One of the most popular superpixel methods is the Simple Linear Iterative Clustering (“SLIC”) algorithm. The trade-off is that the superpixels could only be subsequently classified through low-level features like pixel colour and brightness.

[0345] To reduce the effects of slurry sampling on the results, the validation of PSDs and modal mineralogy was done with a batch mode as the sample quality can be controlled while being prepared. This focuses the accuracy of the methodology instead of the variance from slurry sampling which can be estimated based on the sampled mass or area for a given slurry stream. As the result, a more adequate comparison and error estimation is obtained when the system reading is compared to the ground truth. The mineral and elemental grade of the tested samples were calculated knowing the tail and concentrate portions and their ICP-MS results in the sample mixture. The same analysis was done to calculate the PSDs from manual sieving of the concentrate and the tail samples in a metallurgy laboratory.

[0346] Modal mineralogy - learning and training

[0347] The algorithm according to various embodiments may be trained to estimate the modal mineralogy of particles in slurries. To train the Deep Learning algorithm, a dataset of labelled particles by their mineralogy may be collected for possible minerals encountered in the slurry sample.

[0348] According to various embodiments several samples were pulverized and analysed with XRD to determine the mineral composition present in this ore, which the bulk sulphides were chalcopyrite, pyrite, galena, sphalerite with traces of bornite. A representative sample of each mineral was obtained and imaged under the Scanning Electron Microscopy (“SEM”) and Energy Dispersive X-Ray Spectroscopy (“EDS” or “EDX”). The instrument used in this study was the HITACHI (RTM) TM3030 Scanning Electron Microscopy (“SEM”). The analyses provide the ground truth mineralogy necessary to label the MinDet microscopy image of any particle. The SEM images provide a grayscale image where contrast (different atomic number) and texture can be used to identify regions of different minerals. Then, the EDS uses a focused beam of X- Rays at any small point which excites electrons and measures the X-ray emissions which are characteristic to the known elements. However, EDS is generally classed as a semi- quantitative analysis method as the accuracy depends on the calibration of the machine on known samples and sometimes incorrect automatic element identification.

[0349] Fig. 21 shows the methodology used according to various embodiments which combines SEM and EDS map scans with the MinDet image to identify locked pyrite and iron oxide bearing gangue. The learnings from comparing the MinDet images (i.e. obtained according to various embodiments) with the SEM and EDS results was then used to identify and label 5211 particles. The information was used to train the DL algorithm for quartz, chalcopyrite, pyrite, sphalerite and galena which are the major sulphide and gangue minerals in the ore.

[0350] Modal mineralogy - validation

[0351] The detailed methodology for validating the modal mineralogy methodology of MinDet is discussed above. In the validation process, a new set of samples were made by mixing concentrate and tail samples at different ratios. After scanning those samples by MinDet, the images were run through DL and superpixel algorithms simultaneously to get the sample’s modal mineralogy by size. An example of the image and results reported by MinDet is shown in Fig. 22. In this image the assumption is all the particles contain one mineral only which is not usually the case for coarser size fractions. Liberation degree of minerals reduces as particle size increases until the particle size become way larger than the mineral grain size.

[0352] Fig. 22 shows conversion of images to modal mineralogy by size.

[0353] MinDet software detects the particles and their mineralogy in an image and calculates the mineral content by surface area in that image. The surface content then is converted to elemental by mass assays using the density of the detected minerals and their mineral composition. To validate the MinDet results only 20 images were taken from each sample and only the sulphide minerals were detected in the first attempt. The non detected minerals assume to be non sulphide gangue at specific gravity of 2.7.

[0354] In the first run of validation with 20 images, nine samples were selected to cover a wide range of mineral grades. The elemental assay reported by MinDet were compared to the actual ICP-MS assay results for those samples. The error values for copper, lead, zinc, and sulphur were 1.3%, 1.5%, 0.26% and 1.74% respectively. Error for iron was higher (7.18%) due to its existence in some of the gangue minerals which were not detected.

[0355] Fig. 23 shows elemental assays calculated from MinDet reading of 20 images versus ICP-MS results.

[0356] Fig. 24 shows elemental assays calculated from MinDet reading of 100 images (five passes) versus ICP-MS results. It was hypothesised the more images analysed by MinDet can improve the accuracy of the results. Therefore, three samples were selected which to be scanned five times with MinDet which are shown as five pass sample results in Fig. 24. The error values for copper, lead, zinc and sulphur were significantly reduced to 0.41 %, 0.66%, 0.15% and 1.03% respectively for those samples. Error for iron is still in the same range (7.05%) due to the existence of the undetected iron bearing gangue minerals in the ore body.

[0357] Figs. 25A and 25B shows several examples of fully liberated minerals, their images, and their scans used to develop the mineral detection algorithm. These minerals are fully liberated because the SEM images do not show grain boundaries and have homogenous texture. Galena, PbS (86.6% Pb, 13.4% S) is a light grey and silvery mineral known to form cubic crystal structures which can also be observed in the particles imaged with right angled corners. The EDS scan also shows about 90% Pb and around 5% S which is close to the elemental composition of galena. There are also traces of oxygen which could be due to surface oxidation of the particles.

[0358] Pure sphalerite, ZnS (67.1% Zn, 32.9% S) can form red colour shades where increasing Fe content forms darker crystals. In these examples, the EDS scan shows about 37-53% Zn content with 4-9% Fe and 20-28% S indicating that it is likely sphalerite. There are also traces of oxygen for surface oxidation. There was also a significant number of other elements which could either be contamination or incorrect elemental identification. Chalcopyrite, CuFeS2(34.5% Cu, 30.5% Fe, 35.0% S) is a brassy yellow colour with greenish-black streaks. The EDS scan was consistently reading about 32% Fe, 34% Cu and 32% S which matches extremely well with the true elemental composition. The other two gangue minerals can be seen to be amorphous rather than crystalline. It is not certain what the minerals are since there were little to no S content, they are unlikely to be sulphide. Although not shown here, the most abundant gangue mineral is quartz which is transparent / translucent and easily recognisable. However, it is important to note that all these minerals do not look alike in the microscopy image and can be easily distinguished from each other.

[0359] It will be understood that chalcopyrite regions can be easily distinguished from other minerals as they exhibited the characteristic green-ish black colour. Similarly, Sphalerite is not as shiny as the other sulphides and exhibits a dark purple to red colour. Pyrite can also be distinguished from the darker blue grains. On the other hand, Quartz has a brighter blue and semi-transparent appearance.

[0360] Figs. 26A and 26B show locked sulphide particles scanned by SEM and MinDet (i.e. according to various embodiments). Only sulphide particles had one point EDS scan which is shown by the red circle.

[0361] The EDS has a map scan mode which repeatedly conducts point scans in a raster scan pattern of the specified region and colour codes the presence of the elements. One of the particles from Figs. 26A and 26B was map-scanned under EDS with the results shown in Fig. 21 . This particle was selected to demonstrate and validate the mineral association detected by MinDet and SEM. The EDS map scan indicated that there is an iron sulphide in the blue reflective region of the images collected by MinDet, and from the XRD analysis we know the only iron sulphide present is pyrite. The area is shown by the red line in the image. At the same time, the lack of sulphide in the brown regions, surrounded by yellow line, indicates that it is an iron oxide bearing gangue mineral.

[0362] Particle Size Distribution Estimation

[0363] A Particle Size Distribution (PSD) algorithm was developed by training a ML algorithm to identify the particle boundaries and calculate the surface area by counting the number of pixels in the particles image. P50 measurement was used for the comparison. A total of 100 images were collected to reduce the error in the measurements from a sample containing 50% tail and 50% of concentrate. Three samples sub samples were scanned by MinDet at P50 of 43.0 pm. The average error for those three samples was ± 2.14 pm.

[0364] The development and building of an apparatus to analyse mineral distribution and particle size in real-time is particularly beneficial for mineral processing and separation process plants. The apparatus according to various embodiments integrates slurry sampling and separation, optical microscopy scanning, and a deep learning algorithm to achieve real-time estimation of modal mineralogy of a slurry.

[0365] To test the efficacy of this methodology, a validation study was conducted on a polymetallic sulphide ore sample which where the ground truth was measured by ICP- MS and laboratory screening analysis which are the gold standard methodologies in industry used to measure assay and particle size distribution results. In the validation, the MinDet apparatus was shown to estimate the P50 and mineralogy with high degree of accuracy and was found to be reliable for online measurement of the poly-metal ore sample.

[0366] Through real-time estimation of modal mineralogy of process plant slurries, operators and process control algorithms can be more proactive and optimise separation efficiency by adjusting the relevant operation set points. For example in flotation, these could be reagent selection and dosage control, optimum grind size determination, airflow and tank slurry level controls. The flexibility of the MinDet apparatus to operate between batch and continuous mode also opens the opportunity to be used for rapid assaying in commercial labs with a minimum amount of sample preparation, increasing the number of tests possible.

[0367] Fig. 27 shows a flow diagram illustrating the process of real-time analysis of a slurry according to various embodiments. As shown in Fig. 27, initially one or more valves open and a slurry is arranged to flush and clean a tube. The one or more valves - M - may then be closed with the effect of cut or freezing the slurry stream. A microscope may then be moved into position and the microscope video system may be switched ON The microscope may then be arranged to acquire images. In particular, the microscope may step until the highest focus available is achieved. A determination is then made as to whether or not a FFT autofocus image in focus is obtained. If an in focus image is obtained then the highest focus image is sent for processing. The image may then be appended to a list by sample and subjected to DL CNN image segmentation. Mineral classes and regions may be determined and a 2D-3D size estimation may be performed e.g. utilising a minimum bounding box. A determination may then be made of the mineralogy of the slurry in real-time and element by size calculations may be performed.

[0368] It will be apparent that various embodiments of the present invention enable slurries to be analysed in real-time and a determination be made as to the mineralogy of the slurry. This enables mining processes to be optimised in real-time based upon the actual mineralogy of the slurry which is being produced in real time.

[0369] While particular embodiments have been illustrated and described, it would be obvious to those skilled in the art that various changes and modifications can be made without departing from the scope of the present invention.

Claims

Claims1 . An automated sampling system configured to determine the modal mineralogy of a slurry stream in real-time, the system comprising: a sampling collection and preparation unit configured to separate a slurry sample into a first sample fraction comprising particles having a first size in a first range and a second sample fraction comprising particles having a second size in a second different range; a robotic optical microscopy unit configured to obtain optical image data from the first sample fraction and the second sample fraction; and a modal mineralogy analyser configured to analyse the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

2. An automated sampling system as claimed in claim 1 , wherein the modal mineralogy analyser is configured to use an unsupervised clustering algorithm to analyse the first sample fraction.

3. An automated sampling system as claimed in claim 2, wherein the algorithm comprises a SLIC superpixels algorithm.

4. An automated sampling system as claimed in of claims 1 , 2 or 3, wherein the modal mineralogy analyser is configured to use a supervised Deep Learning (“DL”) algorithm to analyse the second sample fraction.

5. An automated sampling system as claimed in claim 4, wherein the supervised Deep Learning (“DL”) algorithm comprises one or more Deep Learning (“DL”) instance segmentation algorithms.

6. An automated sampling system as claimed in claim 5, wherein the one or more Deep Learning (“DL”) instance segmentation algorithms comprises a first Deep Learning (“DL”) instance segmentation algorithm to identify particles and a second Deep Learning (“DL”) instance segmentation algorithm to classify particles.

7. An automated sampling system as claimed in any preceding claim, wherein the modal mineralogy analyser is configured to determine the size of particles in the slurry sample.

8. An automated sampling system as claimed in any preceding claim, wherein the modal mineralogy analyser is configured to determine the particle size distribution (“PSD”) of the slurry sample.

9. An automated sampling system as claimed in any preceding claim, wherein the modal mineralogy analyser is configured to determine the percentage solids of particles in the slurry sample.

10. A method of determining the modal mineralogy of a slurry stream in real-time, the method comprising: separating a slurry sample into a first sample fraction comprising particles having a first size in a first range and a second sample fraction comprising particles having a second size in a second different range; obtaining optical image data from the first sample fraction and the second sample fraction; and analysing the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

11. A method as claimed in claim 10, further comprising using an unsupervised clustering algorithm to analyse the first sample fraction.

12. A method as claimed in claim 11 , wherein the algorithm comprises a SLIC superpixels algorithm.

13. A method as claimed in any of claims 10, 11 or 12, wherein the modal mineralogy analyser is configured to use a supervised Deep Learning (“DL”) algorithm to analyse the second sample fraction.

14. A method as claimed in claim 13, wherein the supervised Deep Learning (“DL”) algorithm comprises one or more Deep Learning (“DL”) instance segmentation algorithms.

15. A method as claimed in claim 14, wherein the one or more Deep Learning (“DL”) instance segmentation algorithms comprises a first Deep Learning (“DL”) instance segmentation algorithm to identify particles and a second Deep Learning (“DL”) instance segmentation algorithm to classify particles.

16. A method as claimed in any of claims 10-15, further comprising determining the size of particles in the slurry sample.

17. A method as claimed in any of claims 10-16, further comprising determining the particle size distribution (“PSD”) of the slurry sample.

18. A method as claimed in any of claims 10-17, further comprising determining the percentage solids of particles in the slurry sample.

19. A robotic optical microscopy unit configured to determine the modal mineralogy of a slurry stream in real-time, wherein the optical microscopy unit is configured to obtain optical image data from a first sample fraction and from a second different sample fraction; and wherein the robotic optical microscopy unit is configured to analyse the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

20. A method of robotic optical microscopy comprising: determining the modal mineralogy of a slurry stream in real-time by obtaining optical image data from a first sample fraction and from a second different sample fraction; and analysing the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.21 . A computer implemented method of determining the modal mineralogy of a slurry stream in real-time, the method comprising: providing a trained deep neural network that is executed by software using one or more processors of a computing device, the trained deep neural network having been trained with a training set of images comprising microscopy images; separating a slurry sample into a first sample fraction comprising particles having a first size in a first range and a second sample fraction comprising particles having a second size in a second different range; obtaining optical image data from the first sample fraction and the second sample fraction; inputting the optical image data to the trained deep neural network; and outputting modal mineralogy data related to the slurry, wherein the modal mineralogy data includes data related to the size, particle size distribution (“PSD”) and percentage solids of particles in the slurry.

22. A computer-implemented method of training a neural network to determine modal mineralogy in a slurry, the method comprising: obtaining a first set of optical images of a slurry as a first training set; training a neural network as a first stage using the first training set; creating a second training set comprising the first training set and modal mineralogy data associated with the first training set; and training the neural network as a second stage using the second training set.

23. A method of extracting one or more substances from an ore comprising: processing an ore to form a slurry; controlling one or more first operational parameters or conditions of the slurry; obtaining optical image data from the slurry;analysing the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry in real time; and altering or varying one or more of the first operational parameters or conditions of the slurry dependent upon the determined modal mineralogy of the slurry.

24. A method as claimed in claim 23, wherein the one or more first operational parameters or conditions are selected from the group comprising: (i) feed flow rate; (ii) mineralogy; (iii) particle or feed size distribution; (iv) slurry density; (v) reagent type; (vi) reagent dosage; (vii) air flow rate; (viii) slurry level; (ix) slurry chemistry; (x) liberation size; (xi) residence time; (xii) valuable and gangue mineral distribution; (xiii) minerals liberation degree; (xiv) mineral association; (xv) mineral grain size; (xvi) pH of the slurry; (xvii) pressure; and (xviii) temperature.

25. A system configured to extract one or more substances from an ore comprising: a processor for processing an ore to form a slurry; a controller for controlling one or more first operational parameters or conditions of the slurry; an optical imager for obtaining optical image data from the slurry; and an analyser for analysing the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry in real time; and wherein the controller is configured to alter or vary one or more of the first operational parameters or conditions of the slurry dependent upon the determined modal mineralogy of the slurry.

26. A system as claimed in claim 25, wherein the one or more first operational parameters or conditions are selected from the group comprising: (i) feed flow rate; (ii) mineralogy; (iii) particle or feed size distribution; (iv) slurry density; (v) reagent type; (vi) reagent dosage; (vii) air flow rate; (viii) slurry level; (ix) slurry chemistry; (x) liberation size; (xi) residence time; (xii) valuable and gangue mineral distribution; (xiii) minerals liberation degree; (xiv) mineral association; (xv) mineral grain size; (xvi) pH of the slurry; (xvii) pressure; and (xviii) temperature.

27. An automated sampling system configured to determine the modal mineralogy of a slurry stream in real-time, the system comprising: a separator configured to separate a slurry sample into a first sample portion comprising particles having an average size Di and a second sample portion comprising particles having an average size D2, wherein D2> Di; a robotic optical microscopy unit configured to obtain optical image data from the first sample portion and from the second sample portion; and a modal mineralogy analyser configured to analyse the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

28. A computer-implemented method of determining the modal mineralogy of a slurry stream in real-time, the method comprising: separating a slurry sample into a first sample portion comprising particles having an average size Di and a second sample portion comprising particles having an average size D2, wherein D2> Di; using a robotically operated optical microscope to obtain optical image data from the first sample portion and from the second sample portion; and analysing the optical image data using a Machine Learning (“ML”) algorithm in order to determine the modal mineralogy of the slurry sample.

Images