Related Applications

The present application claims convention priority from Australian provisional patent application no. 2021901798 (filed on 16 June 2021) and Australian provisional patent application no. 2022900471 (filed on 28 February 2022). The entire contents of both is incorporated herein by reference in their entirety.

Field of the Invention

The present invention relates to a method and system for logging data for a mineral sample, such as but not limited to drill-hole logging data.

Background

Mining explorations typically involve obtaining mineral samples from a drill site and evaluating the composition of those samples to determine whether a resource is present at the site. One technique for obtaining mineral samples is reverse circulation (RC) drilling, where drill cuttings or chips are brought to the surface by a circulation of air through the drill. Samples of drill chips are typically collected for regular depth intervals during drilling (e.g. 2 metre intervals) to evaluate the mineral composition throughout a length of the drill-hole.

For each interval, a sample of drill chips may be logged and another sample may be sent to a laboratory for analysis, for example by X-ray fluorescent (XRF) analysis or Fourier Transform Infrared (FTIR) analysis. Field logging of drill-hole samples involves visually inspecting the samples and recording the material types present as well as other physical characteristics such as colour, shape and texture.

Field logging is a routine practice typically done by geologists. While compositional assay can reveal the elemental composition of a sample, field logging is necessary to determine geological material types present in a sample, such as hematite, goethite, shale etc.

Information regarding both the composition and material type of drill samples is necessary to better understand the structures and mineralogical compositions of an area. Such information can then be used for ore-body modelling and the development of mining plans.

The accuracy of the field logging data is therefore important for resource evaluation and planning in the minerals industry. However, inaccuracies in the material types logged may arise not only due to complexities and diversities in mineralisation and geology, but also due to subjective biases and human error. There may also be inconsistencies between the logging performed by different geologists. It is therefore common for the estimated composition of the logged chips to differ from the actual composition, and thus validation of the logging data is required to check and/or improve its accuracy. Validation can comprise steps to refine the logged material types and corresponding percentages in order to improve the consistency between the logging and laboratory-analysed chemical composition of the geological sample.

For iron ore exploration and mining in particular, incorrect drill-hole logging information can result in outcomes with significant financial implications. For example, material types such as ochreous goethite and shale are commonly confused due to similarities in colour and texture in chip samples obtained from RC drilling. However, chemically these material types are very different; for instance, shale (kaolinite) is high in silica and alumina and low in iron, whereas ochreous goethite has a high iron grade but is much lower in silica and alumina than shale. Furthermore, ochreous goethite tends to be sticky due to its water holding capacity, which can cause problems such as blocking screen decks and ore transfer chutes, leading to unplanned downtime. Validation during iron ore exploration can thus provide more accurate knowledge of the distribution of ochreous goethite, which can assist in planning blending strategies to manage risks in mining.

Logging and is also an extremely time consuming and labour- intensive task, given that on average each 2 metre interval of an RC drill-hole for example may take a number of minutes to log, and validate and there may be hundreds of kilometres of RC drill-holes drilled each year.

A system integrating methods for logging using objective measurements therefore presents advantages in terms of accuracy, speed and repeatability, and reductions in labour over the existing geologist-driven process.

The present disclosure develops on concepts disclosed in the Applicant's earlier PCT application no. PCT/AU2018/050046 (published as WO 2018/136998 A1). This earlier disclosure is incorporated herein in its entirety. This disclosure may be consulted to better understand aspects of this disclosure. However, it should be understood that the present disclosure supersedes that in the earlier PCT application in relation to any perceived contradiction . Summary of the Invention

Disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement: an infrared material type logger configured to: receive infrared spectra associated with the sample; process the infrared spectra using a pre-trained machine learning algorithm or a statistical algorithm to generate one or both of: an initial material type abundance estimate for the sample, wherein the initial material type abundance estimate is an estimate of the presence of one or more particular material types within the sample and an estimate of the relative abundance of each said material type, and an initial lump percentage estimate for the sample; and store the generated initial material type abundance estimate and/or initial lump percentage estimate in the memory.

The infrared spectra may be generated using Fourier transform infrared (FTIR) spectroscopy. The sample may be prepared for FTIR spectroscopy using one or both of: dehydration; and pulverisation, preferably wherein the pulverisation is performed in two stages, first to about 3 mm and second to about 150 microns.

Optionally, the infrared material type logger is further configured to: resample the received infrared spectra to a predefined common set of wavenumbers. The received infrared spectra may be resampled to a common set of 2966 integer wavenumbers within a predefined range. Optionally, the infrared material type logger is further configured to: clip the received infrared spectra to theoretical minimum and/or maximum values. The infrared spectra may be clipped according to a minimum value greater than 0, such as 0.1%.

Optionally, the infrared spectra are obtained from a known Fourier transform infrared (FTIR) spectrometer and are pre- processed for baseline removal, wherein the baseline removal uses a baseline removal algorithm defined by one or more predetermined parameters, and wherein said parameters are predetermined based on an optimisation parameter search and a comparison between FTIR spectra pairs sourced from the FTIR spectrometer and another FTIR spectrometer. The infrared spectra may be adjusted FTIR spectra, each generated by normalising its removed baseline and combining said normalised baseline and the FTIR spectrum after baseline removal.

The infrared spectra may be processed using a pre-trained machine learning algorithm, wherein the pre-training utilises a training set of infrared spectra, each labelled with logged material type compositions and/or lump percentages represented by the infrared spectra. The training set may comprise infrared spectra associated with known samples extracted from different project areas.

The infrared spectra may be processed using a ridge regression algorithm.

Optionally, the infrared material type logger is further configured to: perform a search on a generated initial material type abundance estimate and/or lump percentage estimate, the search configured to identify one or more groups, each of two or material types with correlated errors, based on a rule indicative of a threshold correlation of errors of members of the group. The search may be a greedy tree search. Optionally, the infrared material type logger is further configured to: remove individual constituents of identified group (s) from the initial material type abundance estimate and/or lump percentage estimate.

Also disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement: a photographic image logger configured to: receive one or more photographs of the sample; process the one or more photographs using a pre-trained machine learning algorithm to generate one or both of: an initial material type abundance estimate for the sample, wherein the initial material type abundance estimate is an estimate of the presence of one or more particular material types within the sample and an estimate of the relative abundance of each said material type, and an initial lump percentage estimate for the sample; and store the generated initial material type abundance estimate in the memory.

The pre-trained machine learning algorithm may comprise a pre-trained general image classification neural network.

The image classification neural network may be modified by removing a final general image classification layer of the network while preserving a prior lower-level training, and training a new final classification layer for classifying the presence of specific material types.

Optionally, the photographic image logger is further configured to: partition the one or more photographs into partitions; process each partition separately using the machine learning algorithm; and utilise a prediction model applied to each processed partition to generate the initial material type abundance. The prediction model may comprise a multiple linear regression algorithm treating each processed partition as a unique independent variable.

Also disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement: a photographic image logger configured to: receive one or more photographs of the sample; process the one or more photographs to generate one or more visual cue classifications for the sample; and store the generated visual cue classifications in the memory.

The one or more visual cues may be selected from: a primary colour of the sample; a secondary colour of the sample; a representative distribution of chip shapes; and textural cues.

Optionally, the photographic image logger is further configured to: determine a primary colour and/or a secondary colour of the sample by: an assessment of a histogram of the photograph (s); or determining an average colour within the photograph. In an alternative option, the photographic image logger is further configured to: determine a primary colour and/or a secondary colour of the sample by: utilising a neural net classifier pre-trained with a training set of images, each labelled with a primary colour.

The photograph (s) may be divided into image patches and each patch may have a primary colour and/or secondary colour determined, and a majority of a colour classification selected as the primary colour and/or a next biggest majority selected as the secondary colour.

Optionally, the photographic image logger is further configured to: determine a representative chip shape classification for the sample by: using a suitable pre trained neural net classifier; or by assessing individual chip shape outlines.

In an embodiment, one photograph is taken of a sample. In another embodiment, a plurality of photos are taken of a sample to enable a 3-dimensional reconstruction.

Optionally, the photographic image logger is further configured to: calibrate the photograph (s) by: determining calibration data associated with the camera(s) taking the photograph (s) by capturing an image(s) with said camera(s) of a known calibration target and comparing an appearance of the calibration target in the image(s) to the known appearance of the calibration target. The calibration data may be suitable for: accounting for a white balance applied by the camera(s); and/or accounting for distance distortions in the image(s).

Also disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement: a measurement data logger configured to: receive drilling measurement data corresponding to one or more measurements made in relation to, and during, extraction of the sample; process the measurement data using a pre trained machine learning algorithm to generate an estimate of the presence of particular material types of the sample; and store the estimate of the presence of particular material types in the memory.

The measurement data may include one of more of: holdback pressure; holdback force; pushdown pressure; pushdown force; penetration rate; torque pressure; torque force; weight on bit; drill string weight; water volume; air flow rate; air pressure; rotation rate; drill string vibration frequency; drill string vibration amplitude; and drill string vibration acceleration.

Optionally, the measurement data logger is configured to utilise a bag-level randomised tree algorithm and/or a Random Forests algorithm in order to produce an estimate of the presence of material types in the sample.

Also disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory configured to implement two or more of: the infrared material type logger of the logging system described above; either or both photographic image logger of the logging systems described above; the photographic image logger of the logging system described above; and the measurement data logger of the logging system described above.

Also disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement: receive an initial material type abundance estimate associated with the sample; receive assay data indicative of an actual composition of the sample or another mineral sample provided from the region of interest; receive a lump percentage estimate associated with the sample; modify the material type abundance estimate based on one or more optimisation criteria and the received assay data and/or lump percentage estimate, wherein the optimisation criteria include one or more of: material type addition criteria, wherein a material type is added to the material type abundance estimate according to the presence of one or more other material types in the material type abundance estimate and an associated predefined addition rule; and/or material type removal criteria, wherein a material type is removed from the material type abundance estimate according to a predefined removal rule.

Also disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement: receive an initial material type abundance estimate associated with the sample; receive assay data indicative of an actual composition of the sample or another mineral sample provided from the region of interest; receive a lump percentage estimate associated with the sample; modify the material type abundance estimate based on one or more optimisation criteria and the received assay data and/or lump percentage estimate, wherein the optimisation criteria include: defining one or more sets of material types, each set comprising at least two material types, such that a sum of the optimised percentages of material types of a set identified in the material type abundance estimate is within a tolerance of the sum of percentages of the material types of the set before optimisation.

The tolerance may be between 0% and 10%.

Also disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement: an infrared material type logger configured to: receive infrared spectra associated with the sample; process the infrared spectra using a pre-trained machine learning algorithm or a statistical algorithm to generate one or both of: an initial material type abundance estimate for the sample, wherein the initial material type abundance estimate is an estimate of the presence of one or more particular material types within the sample and an estimate of the relative abundance of each said material type, and an initial lump percentage estimate for the sample; and store the generated initial material type abundance estimate and/or initial lump percentage estimate in the memory; receive assay data indicative of an actual composition of the sample or another mineral sample provided from the region of interest; receive a lump percentage estimate associated with the sample; modify the material type abundance estimate based on one or more optimisation criteria and the received assay data and/or lump percentage estimate.

Also disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement: a photographic image logger configured to: receive one or more photographs of the sample; process the one or more photographs using a pre-trained machine learning algorithm to generate one or both of: an initial material type abundance estimate for the sample, wherein the initial material type abundance estimate is an estimate of the presence of one or more particular material types within the sample and an estimate of the relative abundance of each said material type, and an initial lump percentage estimate for the sample; and store the generated initial material type abundance estimate in the memory; and receive assay data indicative of an actual composition of the sample or another mineral sample provided from the region of interest; receive a lump percentage estimate associated with the sample; modify the material type abundance estimate based on one or more optimisation criteria and the received assay data and/or lump percentage estimate.

Also disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement: a photographic image logger configured to: receive one or more photographs of the sample; process the one or more photographs to generate one or more visual cue classifications for the sample; and store the generated visual cue classifications in the memory; and receive an initial material type abundance estimate associated with the sample; receive assay data indicative of an actual composition of the sample or another mineral sample provided from the region of interest; receive a lump percentage estimate associated with the sample; and modify the material type abundance estimate based on the visual cue classifications and one or more optimisation criteria and the received assay data and/or lump percentage estimate .

Also disclosed herein is a logging system for logging data obtained for a sample, comprising: a data input system configured to receive input logging data and assay data associated with the sample; and a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement: a measurement data logger configured to: receive drilling measurement data corresponding to one or more measurements made in relation to, and during, extraction of the sample; process the measurement data using a pre trained machine learning algorithm to generate an estimate of the presence of particular material types of the sample; and store the generated estimate of the presence of material types in the memory; and receive an initial material type abundance estimate associated with the sample; receive assay data indicative of an actual composition of the sample or another mineral sample provided from the region of interest; receive a lump percentage estimate associated with the sample; and modify the material type abundance estimate based on the estimate of the presence of material types and one or more optimisation criteria and the received assay data and/or lump percentage estimate.

Also disclosed herein is a logging method for logging data obtained for a sample implemented by a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement the steps of: receiving infrared spectra associated with the sample; processing the infrared spectra using a pre-trained machine learning algorithm or a statistical algorithm to generate one or both of: an initial material type abundance estimate for the sample, wherein the initial material type abundance estimate is an estimate of the presence of one or more particular material types within the sample and an estimate of the relative abundance of each said material type, and an initial lump percentage estimate for the sample; and storing the generated initial material type abundance estimate and/or initial lump percentage estimate in the memory.

Also disclosed herein is a logging method for logging data obtained for a sample implemented by a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement the steps of: receiving one or more photographs of the sample; processing the one or more photographs using a pre-trained machine learning algorithm to generate one or both of: an initial material type abundance estimate for the sample, wherein the initial material type abundance estimate is an estimate of the presence of one or more particular material types within the sample and an estimate of the relative abundance of each said material type, and an initial lump percentage estimate for the sample; and storing the generated initial material type abundance estimate in the memory.

Also disclosed herein is a logging method for logging data obtained for a sample implemented by a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement the steps of: receiving one or more photographs of the sample; processing the one or more photographs to generate one or more visual cue classifications for the sample; and storing the generated visual cue classifications in the memory. Also disclosed herein is a logging method for logging data obtained for a sample implemented by a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement the steps of: receiving drilling measurement data corresponding to one or more measurements made in relation to, and during, extraction of the sample; processing the measurement data using a pre-trained machine learning algorithm to generate an estimate of the presence of particular material types of the sample; and storing the estimate of the presence of particular material types in the memory.

Also disclosed herein is a logging method for logging data obtained for a sample implemented by a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement the steps of: receiving an initial material type abundance estimate associated with the sample; receiving assay data indicative of an actual composition of the sample or another mineral sample provided from the region of interest; receiving a lump percentage estimate associated with the sample; modifying the material type abundance estimate based on one or more optimisation criteria and the received assay data and/or lump percentage estimate, wherein the optimisation criteria include one or more of: material type addition criteria, wherein a material type is added to the material type abundance estimate according to the presence of one or more other material types in the material type abundance estimate and an associated predefined addition rule; and/or material type removal criteria, wherein a material type is removed from the material type abundance estimate according to a predefined removal rule. Also disclosed herein is a logging method for logging data obtained for a sample implemented by a data logging controller comprising a processor and a memory, the memory storing program instructions configured to cause the processor to implement the steps of: receiving an initial material type abundance estimate associated with the sample; receiving assay data indicative of an actual composition of the sample or another mineral sample provided from the region of interest; receiving a lump percentage estimate associated with the sample; modifying the material type abundance estimate based on one or more optimisation criteria and the received assay data and/or lump percentage estimate, wherein the optimisation criteria include: defining one or more sets of material types, each set comprising at least two material types, such that a sum of the optimised percentages of material types of a set identified in the material type abundance estimate is within a tolerance of the sum of percentages of the material types of the set before optimisation.

Also disclosed herein is a method for training an anomaly detector for anomaly detection in respect of Fourier transform infrared (FTIR) spectroscopy spectra, said anomaly detector implementing a neural network, comprising: obtaining a plurality of training samples comprising FTIR spectra associated with samples associated with a common class; undertaking unsupervised training of the neural network, such that the trained neural network is configured to determine one or more latent variables for spectrum reconstruction when provided with a FTIR spectrum as an input, such that the trained neural network generates a pseudospectrum in response to receiving an FTIR spectrum as an input.

Optionally, the method comprises: subsequently to training the neural network, for each training sample: generating an associated pseudospectrum using the trained neural network with the training sample as an input, and determining a score indicative of a similarity between the training sample and its associated pseudospectrum; comparing the scores against a distribution and removing training samples outside of the distribution; and retraining the neural network using the non-removed training samples. The distribution may be a normal distribution. The score may be based on a calculated reconstruction error for each training sample. The score may be based on a calculated reconstruction probability for each training sample. In this case, each score may be calculated by: calculating a reconstruction probability for each wavenumber of the FTIR spectrum; summing said reconstruction probabilities across the entire FTIR spectrum to thereby generate the score.

Optionally, the method comprises: undertaking further testing for neural network optimisation and hyperparameter testing .

The neural network may be based on a variational autoencoder .

Also disclosed herein is a method of anomaly detection in respect of Fourier transform infrared (FTIR) spectroscopy spectra, comprising: obtaining an FTIR spectrum for anomaly detection associated with a sample; calculating a score for said spectrum using an anomaly detector implementing a neural network trained according to the method previously disclosed, said anomaly detector associated with a same class as that of the sample; and comparing the score against a predefined threshold configured for identifying anomalies .

The anomaly detector may be selected from a plurality of anomaly detectors each associated with a different class. Optionally, in respect to the comparison between the score and the predefined threshold indicating an anomaly, the method comprises recording an anomaly flag in association with the FTIR spectrum.

Also disclosed herein is an anomaly detector for anomaly detection in respect of Fourier transform infrared (FTIR) spectroscopy spectra, said detector configured for applying the method disclosed above to provided FTIR spectra.

Also disclosed herein is a computer program or computer readable medium comprising a computer program, the program comprising code configured to cause a processor to implement at least one of the above methods.

Various aspects described above can be implemented together providing the combined benefits—for example, various loggers defined as implemented by the aspects of a logging system may be implemented within a common logging system.

Brief Description of Drawings

In order that the invention may be more clearly understood, embodiments will now be described, by way of example, with reference to the accompanying drawings, in which:

Figure 1 is a flow diagram of a method according to an embodiment .

Figure 2 is schematic diagram of a system according to an embodiment .

Figures 3A and 3B relate to a method for generating estimate of initial material type abundance of a sample using Fourier transform infrared spectrum analysis according to an embodiment.

Figures 3C to 3F relate to a method of baseline removal. Figure 3C shows a discrepancy between FTIR spectra for the same sample due to a difference in baseline, Figure 3D shows an absorbance difference distribution for a number of samples measured on two different FTIR spectrometers,

Figure 3E shows the effect of baseline removal on two FTIR spectra of the same sample, and Figure 3F shows a method for identifying baseline removal parameter(s).

Figures 4A to 4C relate to a technique of using photographs of a sample to determine visual cues and/or an estimate of initial material type abundance of a sample according to an embodiment .

Figure 5 shows an exemplary relationship between functional features of the system of Figure 2, according to an embodiment .

Figures 6A and 6B relate to an optimisation process implemented by an optimisation module of the system of Figure 2 according to an embodiment.

Figures 7A to 7F relate to an embodiment including a neural network implemented anomaly detector for FTIR spectra.

Detailed Description

Figure 1 is a flowchart of a method of logging data for a mineral sample according to an embodiment. The method 100 will herein be described in the context of iron ore mining exploration using reverse circulation (RC) drilling to obtain mineral samples. However, a person skilled in the art will appreciate that the disclosed method can be used in other applications and can involve other drilling techniques.

Throughout this specification, unless the context requires otherwise due to express language or necessary implication:

• The term "composition" and variants thereof refer to a chemical composition of a material, i.e. a set of chemical elements and/or compounds, such as but not limited to Fe, Si0 ₂ , A1 ₂ 0 ₃ , P, S, Mn, MgO, Ti0 ₂ , CaO, H ₂ 0, which might be present in a mineral sample. The term "composition" may also be used in a manner that refers to the amounts or proportions of these chemical elements and/or compounds present in a mineral sample.

• The term "material type" refers to a type of material characterised by its constituents, including various elements and/or compounds, and/or physical properties such as hardness, texture, colour and shape. Various material types may have a known theoretical composition. For example, ochreous goethite is a material type that has high iron (Fe) content, but is relatively low in silica (Si0 ₂₎ and alumina (A1 ₂ 0 ₃₎ . Some material types may have very similar chemical compositions, but different physical properties.

The method 100 comprises providing a mineral sample from a region of interest (step 110). The region of interest according to a specific embodiment is at a particular depth or depth range of a drill-hole. Samples of drill cuttings or chips brought to the surface are collected for each regular length intervals of the drill-hole. For example, if the intervals are chosen to be 2 metre intervals, drill chips may be collected for each of the ranges 18m-20m, 20m- 22m, 22m-24m etc. below the surface. The method further comprises obtaining analytical data associated with the chemical composition of the sample (step 120). In a specific embodiment, the assay utilises X-ray fluorescent (XRF) analysis to determine the presence of particular constituents, and amounts of those constituents, of the assayed sample. In particular, the XRF analysis can be arranged to measure Fe, S1O2, AI2O3, P, S, Mn, MgO, T1O2, and CaO. It will be appreciated that other analytical techniques can be used, for example, Loss on Ignition (LOI), such as Total LOI, LOI425 (measuring goethite-bound water) and LOI650 (kaolinite associated water) content, can be determined using a Thermogravimetric Analyser (TGA).

The method further comprises obtaining analytical data associated with the spectral reflectance of the sample (step 130), from which mineralogical characteristics of the sample are determined. According to a specific embodiment, Fourier transform infrared (FTIR) reflectance is used. The sample used for XRF analysis may be the same sample as that used for FTIR analysis, or they may be different samples obtained from the same region of interest (e.g. 2 metre drilling interval). Generally, the sample used for step 120 and the sample used for step 130 should be representative of the same region of interest.

In an existing method of manual geological logging, one sample from the 2-metre drilling interval is analysed by a geologist, who records field logging information. Such field logging includes visual inspection of the samples to estimate the percentages of various material types present, usually in increments of 5%. Material types may be identified at scales ranging from microscopic to macroscopic. These qualitative physical properties remain consistent across various sites, though minor changes in geochemistry may occur. For example, some material types that have been defined for iron ore explorations are provided in Table 1. Reference is made herein to the 3-letter codes of Table 1 when describing embodiments related to iron ore exploration—such reference should not be construed as limiting.

TABLE 1: Examples of Material Types defined for use during

Iron Ore Explorations

In addition to recording an estimate of material types present, in an existing method of manual geological logging, the geologist may also record an estimate of other physical characteristics of the logged sample. For example, such as the sample colour, chip shape, hardness, texture, and magnetic susceptibility can also be observed and noted during field logging.

Step 140 of the method comprises using a processor to automatically log the mineral sample according to at least one predetermined criterion and based on the analysis from of the comparison from step 120. In this specification, it will be understood that the term "processor" refers to any device capable of processing program instructions typically stored as program code in a memory, which can include a volatile memory (e.g. DRAM and/or SRAM) and/or a non volatile data storage device (e.g. a magnetic hard drive and/or a FLASH or EPROM-based memory). The processor may be a microprocessor, microcontroller, programmable logic device, a computing device, or any other suitable processing device.

In that regard, with reference to Figure 2, according to an embodiment, the step 140 of the method 100 is performed using a data logging system 200 for logging data obtained for a mineral sample. The system 200 comprises a data input system 210 arranged to receive logging data associated with the logged sample (examples shown include FTIR spectra, photographic images of samples, and measurement data associated with extraction of the sample) and compositional assay data obtained from chemical analysis of the sample. The system 200 further comprises a data logging controller 220 arranged to determine a value of a discrepancy between the assay data and the logging data, and modify or adjust the logging data based on the value and according to at least one predetermined criteria.

The data logging controller 220 includes a processor 222 and data storage 224 in which program instructions are stored to be executed by the processor 222. Therefore, the data processor 220 in this embodiment can perform the function of the processor used in step 140 of the method 100. Accordingly, for convenience, other method steps in further embodiments will be discussed in the context of implementation by the logging system 200. Figure 2 shows the logging controller 220 configured to implement at infrared material type (FTIR) logger 230, photographic image logger 240, a measurement data logger 250 as described herein. It should be understood that one or more of these loggers 230-250 may be excluded depending on the particular implementation requirements.

Fourier Transform Infrared Spectrum Analysis

Infrared spectroscopy measures a sample's response to incident radiation across the infrared band of wavelengths, e.g. the percentage of incident radiation which is reflected at each wavelength. Radiation within the infrared band can induce molecular or mineral bond vibrations, and so a sample's infrared spectra is sensitive to the sample's mineral composition.

Referring to Figure 3A, a method 300 for producing analytical data associated with the spectral reflectance of the sample is shown (i.e. corresponding to step 130 of Figure 1). The method may determine initial material type abundance estimates and/or lump percentage estimates using Fourier Transform Infrared (FTIR) Spectrum Analysis. The lump percentage estimates the breakdown of the ore represented by the sample into lump (particles >6.3mm or 0.25" in diameter) and fines product.

At step 310, infrared spectra of a sample are received by the FTIR logger 230. In an embodiment, infrared reflectance spectra are obtained by Fourier-transform infrared spectroscopy of a suitably prepared sample, which rapidly measures infrared spectra at high resolution. Sample preparation can include one or both of: dehydration to prevent unbound water affecting the infrared spectra; and pulverisation (which must be consistent between samples as FTIR spectra are often sensitive to particle sizes within the sample). In an embodiment, a sample is pulverised in two stages: first to 3mm (e.g. Boyd Crushed using the Rocklabs Boyd Crusher), and then to 150 micron (LM5 Pulverising using the Essa® LM5 Pulverising Mill).

The supplied infrared spectra are then pre-processed at step 320, for example using FTIR logger 230.

In an embodiment, infrared spectra are resampled to a common set of wavenumbers so that the method is advantageously not restricted to spectra measured by specific FTIR machines. In an embodiment, infrared spectra are resampled by linear interpolation to a common set of integer wavenumbers over a particular range, for example, as output by particular FTIR machine such as 2966 integer wavenumbers between 6000 cm ^A -l and 282 cm ^A -l or 1499 integer wavenumbers between 6001.5 cm ^A -l to 223.7 cm ^A -l. Next, optionally, as infrared spectra may exceed their theoretical minimum and maximum amplitudes due to measurement noise, the imported infrared spectra can be clipped to their theoretical minimum and maximum values. In an embodiment, input reflectance spectra are clipped to between 0.1% and 100%, where the lower bound is greater than zero to enable subsequent logarithmic transform. Then the infrared spectra are statistically transformed to aid subsequent prediction algorithms. In an embodiment, infrared reflectance spectra are converted to fractions (e.g. by dividing by 100) and logarithmically transformed using known techniques.

With reference to Figures 3C-3G, in an embodiment, the infrared spectra are pre-processed for baseline removal. Referring to Figure 3C, it has been found that there can be systematic differences between absorbance spectra from different testing facilities (or, in fact, at different FTIR spectrometers within the same facility)—in this case, spectrum 360a is noticeably different to spectrum 360b. Figure 3D shows an absorbance difference distribution for 3130 pairs of samples (each pair representing one spectrum of a sample tested at a first facility and a second spectrum of the same sample tested at a second facility).

As can be seen, there is a reasonable variation, especially at lower wavenumbers. It would be preferred if the difference distribution was centred around 0 with a small variance .

Baseline removal is expected to be beneficial as it can account for the variation in differences with wavenumber, on the assumption that the relative absorption peak heights within a particular spectrum are non-problematic. That is, the difficulty in comparing inter-facility FTIR spectra is due to a variable offset rather than variations in the ratios of peak heights.

For example, Figure 3E shows the difference between two spectra 362a, 362b before baseline removal and after baseline removal 364a, 362b (common suffix shows common spectra) . The calculated baselines are shown as 363a and 363b.

Figure 3G shows a method for determining baseline removal configurations for a pair of FTIR spectrometers (i.e. a first FTIR spectrometer and a second FTIR spectrometer).

The method generally involves selecting a baseline removal algorithm parameterised by one or more baseline parameters and determining suitable values for said one or more parameters . The method can be extended to include determining both a suitable baseline removal algorithm (usually from a finite group of possible algorithms) and its associated one or more baseline parameters. The baseline removal algorithm may be of a known type. At step 370, a sample set is created for testing by both FTIR spectrometers. The sample set comprises sample pairs, where each sample pairs comprises two samples from the same source (e.g. the sample pairs may be created simply by dividing a sample into two). The purpose of each sample pairs is to enable a comparison, on the basis that the samples of the pair are known to have the same composition, in the FTIR spectra generated by the two FTIR spectrometers. The sample pairs are used for statistical analysis and therefore, a suitable number should be provided based on a required statistical certainty. For example, the number of pairs can be 100 or more, and more preferably, 1000 or more. The number of pairs may depend on, for example, the capability of a particular facility for consistency in sample preparation and measurement. In the example of Figure 3E, 3130 sample pairs were utilised. Although the samples with a particular sample pair are from the same source (e.g. obtained by dividing into two an original geological specimen), the source of material for different sample pairs can differ.

At step 371, FTIR spectra pairs are generated for each sample pair. Each FTIR spectra pair comprises a first FTIR spectrum measured by the first FTIR spectrometer on one of the samples of its associated sample pair and second FTIR spectrum measured by the second FTIR spectrometer on the other sample of the sample pair. Relevantly, the first FTIR spectrum and second FTIR spectrum have a known association for step 372.

At step 372, a parameter search is performed over the FTIR spectra pairs. The parameter search is assessed based on a similarity between the first FTIR spectrum and second FTIR spectrum for each FTIR spectra pair after baseline removal is applied, as a function of the baseline parameters. In an embodiment, the number of baseline parameters searched is twice the number of baseline parameters associated with the baseline algorithm; that is, comprising a first parameter set comprising values for the one or more baseline parameters associated with the first FTIR spectrometer and a second parameter set comprising values for the one or more baseline parameters associated with the second FTIR spectrometer.

It should be noted that there can be a trivial solution for generating a closest similarity between the FTIR spectra after baseline removal, namely that where both spectra have zero absorbance. Therefore, the parameter search should be encouraged to find non-zero absorbance solutions.

In an example, the baseline removal algorithm can be asymmetric least squares (ALS) baseline correction having two parameters: A (baseline smoothness) and p (baseline overshoot allowance). This algorithm was used to generate the result of Figure 3E.

The parameter search can be based on a suitably configured metric which defines a similarity comparison. The metric provides, in effect, a means to compare the "quality" of the various combinations of baseline parameters. In an example embodiment, the metric is based on:

• Attempting to minimise a mismatch between baseline corrected FTIR spectra of the first FTIR spectrometer and the second FTIR spectrometer; for example, attempting to minimise the mean integral of absolute differences of corrected absorbance spectra. • Prevent over-smoothing by attempting to maximise the mean integral of corrected absorbances (which can include negative values after baseline removal).

The metric can include predefined (e.g. user settable) weightings in respect of the various functions being optimised during the parameter search.

In one example, a metric is calculated on the basis of an "integrated absolute difference" {IAD) and an "aggregate area under the curve" {AUC) . A smaller IAD is desired as it measures differences between the absorption peaks after baseline removal (smaller differences indicting a closer match). A larger AUC is desired as a lower value for this measure indicate that the baseline removal step removed a significant portion of the raw absorbance spectrum (i.e. the spectrum before baseline removal).

Accordingly, in this example, for a particular selection of baseline parameters, the quality of the baseline removal can be measured according to:

Here, a larger value for the metric reflects a better- quality baseline removal.

The two functions, AUC{a^ _n ,a _2n ) and MD(a _ln ,a _2n ) are calculated for each of the N sample pairs (indexed by n), the results for each pair summed for each function. In one example: r ^u 2

Ju

(2) auc( where u is the log-transform of the wavenumbers v (i.e. u = logv), is the log-transform of the minimum wavenumber and u ₂ is the log-transform of the maximum wavenumber. ^a i,n( ^u ) is the value of the baseline-removed absorbance peak (i.e., raw absorbance minus baseline) at u for the first FTIR spectrum of the n'th sample pair and a _2n (u) is the baseline-removed absorbance peak at u for the second FTIR spectrum of the n'th sample pair.

Once the parameter search is complete, the determined first parameter set and second parameter set allow can advantageously improve the accuracy of FTIR analysis independently of whether a sample is tested by the first FTIR spectrometer or the second FTIR spectrometry, on the basis that, after baseline removal, the relative peak heights of the FTIR spectra of either machine can be more reliably be assumed to be equivalent for the same sample.

However, the information removed through baseline removal can (in at least some cases) comprise important sample information. To enable FTIR spectrometer-independent assessment of the baseline information, for a particular FTIR spectrum, the removed baseline can be normalised to a common standard and then adjusted FTIR spectra created by adding together the baseline removed FTIR spectra and the normalised baseline. Advantageously, the "adjusted FTIR spectra" may then be assumed to be FTIR spectrometer- independent and therefore the adjusted FTIR spectra may be better suited for further analysis, for example, as per various embodiments described herein.

In an embodiment, the "common standard" is based, at least in part, on the parameter search of Figure 3G. The baseline removal algorithm based on the first parameter set can be understood as a first function and the baseline removal algorithm based on the second parameter set can be understood as a second function. The common standard can therefore be based on a transform linking the first function and second function.

Therefore, in effect, knowledge of the first parameter set and the second parameter set enables a normalisation function to be determined. In one example, the baseline of the first FTIR spectra is the common standard, in which case, the normalisation function corresponds to a transform of the baseline of the second FTIR spectra to make it consistent. In another example, the common standard requires transforms of both baselines, for example, a median representation between the two or some other representation .

It should be understood that the method of Figure 3G and/or the baseline normalisation can be extended to more than two FTIR spectrometers. For example, instead of "sample pairs" comprising two samples, there can be groups of N identical samples (where N is the number of FTIR spectrometers). In this case, the parameter search of step 372 is simply over N parameter sets.

In an embodiment, parameter sets for additional FTIR spectrometers can be determined after the first parameter set and second parameter set are determined. In this case, the first (or equivalently second) FTIR spectrometer can be assessed with respect to an additional FTIR spectrometer in a similar way to the method of Figure 3G. However, it may be preferred to fix the parameter set of the first FTIR spectrometer to its already determined values such that adjusted FTIR spectra of the first FTIR spectrometer remain comparable to adjusted FTIR spectra of the second FTIR spectrometer .

The pre-processed infrared spectra are then processed by the FTIR logger 230 at step 330. FTIR logger 230 produces initial material type abundance estimates and/or lump percentage estimates from infrared spectra using a machine learning algorithm or a statistical algorithm, based on a training set of (pre-processed) infrared spectra, each labelled with logged material type compositions and/or lump percentage (as appropriate). The training sets may advantageously comprise samples from separate project areas, as the same material types may have varying physical properties between different project areas. The training sets may advantageously comprise samples from a large number of project areas, in which case the material type estimates will be geochemically validated in a later optimisation step (see optimisation module Step 025).

In an embodiment, step 330 comprises using ridge regression to predict initial material type abundance estimates and/or lump percentage from the (pre-processed) infrared reflectance spectra. In an embodiment, the ridge regression algorithm uses efficient leave-one-out cross-validation to select an optimal regularisation coefficient from one of thirteen candidates: 10 ^L -6, 10 ^L -5, ..., 10 ^L 6. Predicted material type abundances can be independently clipped to lie within their theoretical range of between 0 and 1 for fractions, or 0 and 100 for percentages. Similarly, predicted lump percentages can be clipped to between 0 and 1 for fractions, or 0 and 100 for percentages. In an embodiment, a search is performed on the initial material type abundance estimates from step 330 in order to identify one or more groups, each of two or more material types with correlated errors at optional step 340.

Errors between different material type abundance predictions may exhibit correlation since there may exist groups of spectrally similar or identical material types. Therefore, knowledge of these groups can advantageously be useful for refining the material type predictions in future optimisation steps (see optimisation module 280), as a group's total abundance may be more reliable than the individual predicted abundances of members of the group.

The search of step 340 therefore attempts to identify groups of material types where the sum of a group's predicted abundances is more accurate than the independent material type abundance predictions.

Referring to Figure 3B, in an embodiment, this search is a greedy tree search. At step 341, a matrix (R) is generated and input in which a particular entry Ri _j is the residual for the i ^th sample's prediction of the j ^th material type abundance. The variable i is used to index rows in R and may be assigned any integer value between 1 and n inclusive, where n is the number of samples. The variable j is used to index columns in R and may be assigned any integer value between 1 and m inclusive. At the start of the search m is equal to the number of input material types, but actually represents the number of material types and/or material type that are current candidates for merging.

Then, at step 342, a candidate pair of material types and/or material type groups (71,72) for merging is generated, where j and j are indices for columns in R and may each correspond to individual material types or a group thereof. At step 343, pre-merge mean-squared-error (MSE) is calculated for the candidate pair of material types and/or material type groups over all 'h' samples, for example according to:

Then, at step 344, post-merge MSE is calculated for candidate material types and/or material type groups over all n samp1es:

Then, at step 345, a record is made (e.g. in storage 220) of the merge error ratio of post-merge MSE to pre-merge MSE, for example:

Post merge MSE

Merqe error ratio = —- ———

Pre merge MSE

(6)

A check is then made at 346 as to whether all possible material type and/or material type group pairs have been evaluated, if not the method returns to 342. Otherwise, the method proceeds to merge material types and/or material type groups with lowest merge error ratio to obtain a new larger group, at step 347. For example, this can be expressed as: R 1 ^R i+ ^R h

DELETE COLUMN [R. _j m m — 1

(7)

Next a check is made at step 348 as to whether 'm' is equal to or below 2. If not, then the method returns to step 341. Otherwise, the method outputs a full list of material type and/or material type group merges performed during the search and the corresponding merge error ratios at step 349.

Finally, referring back to Figure 3A, at step 350, the estimates of initial material type abundance (including, where applicable, groups of correlated abundances) and/or lump percentage are stored for later use, for example in data storage 224.

Measurement While Drilling Data

Typically, when drilling a hole for obtaining samples, various parameters of the drill rig are monitored by the drill rig's computer system. These include one or more of: holdback pressure; holdback force; pushdown pressure; pushdown force; penetration rate; torque pressure; torque force; weight on bit; drill string weight; water volume; air flow rate; air pressure; rotation rate; drill string vibration frequency; drill string vibration amplitude; and drill string vibration acceleration. Such drilling measurement data are typically sampled at a much higher resolution (e.g. one or more orders of magnitude) than the geology logging intervals. For example, the geology logging intervals may be 2 m while the drilling measurement data may be recorded approximately every 1 mm. According to an embodiment, a measurement data logger 250 (see Figure 2) is utilised for estimating properties of the subsurface from the recorded measurements of the drilling parameters. In an embodiment, a subset of parameters is used comprising one or more of (and preferably all of): penetration rate; rotation rate; torque; and weight on bit. Generally, material type presence/absence predictions are reported for geology logging intervals (e.g. 2-metre intervals) .

In an embodiment, a bag-level randomised trees (BLRT) algorithm (Komarek et al., 2019) is used to predict the presence of each material type within a geology logging interval, given the set of measurement while drilling samples recorded within that geology logging interval. The BLRT algorithm is suitable for producing predictions pertaining to the geology logging interval.

A Random Forests(TM) (RF) algorithm (Breiman, 2001; "Random Forests" is a trademark of Leo Breiman and Adele Cutler and is licensed exclusively to Salford Systems for the commercial release of the software) may be used to predict the presence of each material type within single measurement while drilling samples. These individual predictions are then grouped by their constituent geology logging interval, and each group of predictions is statistically aggregated to produce material type presence predictions that described the entire geology logging interval. In an embodiment, the statistical aggregation used is the arithmetic mean. In another embodiment, the statistical aggregation used is the maximum function. In yet another embodiment, the statistical aggregation used is the geometric mean. Photographic Imaging

According to an embodiment, with reference to Figure 4A, a method is provided for analysing photographs of chip samples (i.e. corresponding to a geological sample).

The photographs comprise visual information corresponding to visual cues which a geologist typically uses when logging a sample. The visual cues can include, for example, one or more of: a primary colour of the sample; a secondary colour of the sample; a representative distribution of chip shapes; and textural cues. Textural cues can assist in the identification of specific material types; for example, vitreous goethite is identifiable through its vitreous texture. This visual information is complementary to the geochemical assays and reflectance spectra described above. The photographic image logger 240 processes sample photographs to generate sample visual cue classifications for each sample. Herein, "primary colour" refers to the predominant colour of the sample and "secondary colour" refers to a next most predominant colour of the sample.

Figure 4B shows a collection of chip sample trays 490 each associated with a camera 491. Although not shown, an embodiment may utilise multiple cameras 491 per chip sample tray 490. For example, in an embodiment, the photograph is captured by a single overhead camera 491 providing an overhead view of the chips of a sample within the corresponding chip sample tray 490. In another embodiment, multiple photographs of the sample are taken, each from a different angle and optionally using different cameras 491 or a single moveable camera 491, which may advantageously enable 3D reconstruction of the sample from which virtual measurements can be taken. Referring back to Figure 4A, photographs of each (chip) sample are received at step 400. The chip sample can be prepared for photographing by pouring coarse retains into a sample tray cell and allowing them to settle such that the surface is relatively flat. In an embodiment, a fine water spray is applied in order to wash fines off the surface of the larger chips, and to increase the contrast of textural features in the image. The sample is typically held in a chip sample tray.

The photographs are calibrated at step 420. Calibration accounts for distortions introduced by the digital camera 491, for example a camera 491 may apply a white balance to a photograph to satisfy certain assumptions, such as the average colour of the image being grey (Ebner 2007)), or that the top 1% of image red, green and blue values represent the colour white (Ebner 2003). Additionally, distances between features in the photograph may be distorted with respect to actual distances between said features—for example, due to optical lens distortions.

Each camera 491 is calibrated for each chip sample tray 490; that is, calibration data is generated for each combination of chip sample tray 490 and camera 491. In an embodiment, calibration occurs before any samples are imaged. In another embodiment, the calibration data is generated contemporaneously with the imaging of a particular sample, for example, a photograph of a particular sample may also comprise calibration information.

In an embodiment, calibration data for a particular camera 491 is generated by photographing a calibration target comprising a number of patches with known colour and location characteristics. In an embodiment, the calibration target is an X-Rite ColorChecker® NANO, which fits within a cell of a sample tray. By identifying the locations and colour of each known colour patch in the photograph, a colour transformation can be estimated corresponding to each patch, and then reversed, thus correcting the camera's white balance assumptions. That is, the difference between the known colour value of each patch and that in the photograph can be used to effectively "undo" the effect of the white balancing.

A ColorChecker target can be identified by first identifying individual colour patches. In hue-saturation- value (HSV) space where the values of each channel range from 0 to 1, colour patches can be isolated by: applying thresholds in each dimension of the space such as 0.3 < H < 0.9, S > 0.1 and V > 0.5, with the intent of removing the black frame of the ColorChecker and producing a thresholded image; identifying connected components in the thresholded image and identifying a number of largest blobs, for example 15, with an aspect ratio such that the minor axis length is at least half of the major axis length to allow for misshapen blobs due to small variations in colour; forming putative matches between blobs and target colours where the difference in hue is less than 0.2; and using a robust estimation algorithm such as RANSAC to fit a homography (Hartley 2004) which transforms the coordinates of blob centroids in the thresholded image to known grid coordinates on the ColorChecker target.

The homography is invertible and the inverse transforms image coordinates to plane coordinates, allowing measurements in the image to be transformed to measurements on the calibration target of known size. This allows the measurement of objects within the image, such as the sizes of chips for estimating the particle size distribution. The homography can also account for optical distortions. Correspondences between the colours of blobs and known colours on the ColorChecker can be used to calibrate the colour, such as using the Chromatic Adaptation method.

Alternatively, the target may be identified using a neural network trained for this task (Fernandez 2019).

At step 430, the calibrated photographs are analysed to identify visual cues that may be present.

In an embodiment, a primary colour visual cue of the samples is determined for each photograph. The primary colour is retrieved through analysis of each sample photograph's histogram, in the hue-saturation-value colour space, in the red-green-blue colour space, and/or the CIELAB colourspace, or any other appropriate colour space. Alternatively, the average colour of the image can be used. In another embodiment, the primary colour is classified by a neural net classifier trained using an appropriately prepared training set comprising a number of sample photographs (e.g. more than 1000), each tagged with a primary colour represented within the training image.

In an embodiment, a similar process is utilised for identifying a secondary colour; the training images can be tagged indicating the presence of a specific secondary colour (or colours). The colours may be classified for a particular photograph as a whole, or through the subdivision of the photograph into image patches, which are independently classified, and then a majority colour class used as representative of the image's primary colour; the secondary colour can be classified as the second most commonly classified image.

In an embodiment, a representative classification of chip shape is determined as a visual cue. Generally, the representative distribution of chip shapes as logged by a geologist classifies may be sample as:

TABLE 2: Examples of Chip Shape Classification

Table 2 may be implemented according to the guidelines described in the Field Geologists' Manual (fourth edition).

In an embodiment, the representative chip shape classification is obtained using a neural net classifier trained on a number (e.g. more than 1000) of sample photographs, each labelled with a corresponding logged chip shape. In one embodiment, the chip shape of a sample may be classified from a single image. In another embodiment, the chip shape of a sample may be classified through a majority vote, where independent classifications were made from patches of the chip photograph.

In an embodiment, the representative chip shape classification is obtained by assessing individual chip outlines within the photograph. Then each chip outline is processed individually. A chip centre is estimated as the circumcentre of the chip outline. The chip outline is then converted to polar coordinates about the circumcentre, producing a transformed outline. Phase congruency (Kovesi 1999) is computed for the transformed outline. The feature types computed at points of phase congruency (Kovesi 2002) are used to classify the chip shape. The representative chip shape of the sample is then derived through a majority vote from all classified chips.

In an embodiment, the particle size distribution is estimated by analysing the sizes of the crushed chips present in a photograph. In an embodiment, the distribution is the proportion of material falling within specified size ranges, for example, each of these three ranges: 0-0.5mm in diameter, 0.5-lmm in diameter, and l-3mm in diameter. In a process of modelling lump percent from particle size distribution, the lump percent is estimated using a neural net that was trained to predict the lump percent from the particle size distribution, where the input training data was obtained using the particle size quantities described in the sample preparation step of step 310, and the predicted lump percent for each sample was calculated from logging for that sample from the known lump percentage for each material type. In practice, the step 430 calculates the particle size distribution from the photograph, which is provided to the neural net to estimate the lump percent for the sample.

Figure 4C shows a variation of Figure 4A in which for the photograph is analysed to predict the presence of specific material types, at step 440 (i.e. identifying an initial material type abundance estimate). This can be an alternative method of obtaining the initial material type abundance estimates to that described with respect to Figures 3A-3C. Alternatively, this method may be utilised as a complementary technique—for example, the results of the FTIR estimate and photograph estimate of the presence of specific material types can be combined, optionally weighted to favour a technique considered to be more reliable. Steps 410-430 are the same as for Figure 4A (although step 430 may be incorporated into step 440 or excluded).

In an embodiment, a pre-trained general image classification neural network is used to classify the chip sample photograph to identify the presence of specific material types. The pre-trained network may be modified by removing the final general image classification layer of the network while preserving the prior lower-level training, and training a new final classification layer for classifying the presence of specific material types. This method, known as "ablation", provides an advantage of exploiting the training of lower-level features from a large dataset, as a basis for classification of a smaller dataset. In one embodiment, a pre-trained VGG16 neural network (Simonyan and Zisserman 2015) with batch normalisation was used.

The chip sample photos can also be used for predicting material type composition

In an embodiment, partitions (i.e. contiguous patches) of the photograph were used as input for a pre-trained general image classification neural network. The outputs of each neural networks' layer are extracted (Garcia-Gasulla et al. 2018) and used as input into a second simpler prediction model. In an embodiment, a pre-trained VGG16 neural network with back propagation (Simonyan and Zisserman, 2015) is used as the pre-trained neural network, in one example this produced 12,416 outputs. In that embodiment, those 12,416 outputs can then be used as independent variables in multiple linear regression (the second simpler prediction model) to predict material type composition. Once predictions are computed for all image patches extracted from a single photograph, the predictions are statistically aggregated to give an estimate of material type composition for the entire photo. In an embodiment, the arithmetic mean is used as the statistical aggregation function. The physical size of the image patches affects performance and are optimised accordingly. In one embodiment, 3 mm square patches were used. In another, 1.25 cm square patches were used.

Optimisation Module

Figure 5 shows schematically the relationship between the optimisation module 280 and the FTIR logger 230, photographic image logger 240, measurement data logger 250, and assay module 270, according to an embodiment. The data types generated by the loggers 230-260 are shown in broken lines. The broken line arrow indicates that the photographic image logger 240 may provide initial material type abundance estimation alternatively or complementarily to the FTIR logger 230. Similarly, although both FTIR logger 230 and photographic image logger 240 are shown contributing to the lump percentage estimate, this may be provided by only one of these two.

In an embodiment, the optimisation module 280 receives initial material type abundance estimates 281, lump percent estimate 282, sample visual cues 283, and properties of the subsurface 284. Additionally, the laboratory assays via assay module 270 (i.e. analytical data determined at step 120 of Figure 1) are also provided to the optimisation module 280 The optimisation module 280 is configured to generate material type logging 286 for the sample that optimally satisfies the inputs 281-285.

More generally, not all datasets 281-284 are necessarily utilised (e.g. due to availability). In an example implementation, the optimisation module 280 requires a minimum of initial material type abundance estimates 281, lump percent estimate 282, and laboratory assays 285 as inputs. Other implementations may define different combinations of minimum datasets 281-284.

In an embodiment, an optional additional input module 260 is provided (shown in Figure 2, not shown in Figure 5). In this case, the input from the FTIR logger 230 is combined with an input from the additional input module 260. The additional logging input may be either be from known logging sources, or from other suggested inputs. An example of an additional logging source is logging from a previously created block model of the area containing the drillhole being logged. Another example of an additional logging source is from an adjacent sample, either from a preceding or following interval in the hole, or from a sample retrieved from a similar depth in a nearby drillhole. The combination may be a weighted average of the infrared material type logger input with the additional input module 260.

Optimisation is not just a goal-seeking exercise to minimise the discrepancy between the initial material type abundance estimate and the other inputs; several geological and physical constraints ought to be satisfied for the validated composition to be accurate and meaningful. Thus, the optimisation should be performed according to predetermined criteria or "optimisation criteria". For example, some material types such as pisolite are physically distinctive and their presence or absence should be obvious in a chip sample photograph. Therefore, an optimisation criterion may relate to prohibiting these materials from being removed or added if originally identified as present or absent within a photograph, respectively, in the logging data.

An analysis of historically logged samples may provide information concerning which material types were commonly logged together. This information may assist in understanding the geological context of the different material types. In this regard, the Apriori algorithm may be used to determine "association rules" from compositional data previously logged on the basis that geologists have previously identified geologically valid combinations of material types. The association rules determine when logging a material type X should lead to another material type Y being present in the logging data. Each association rule has a confidence value and a support value. The confidence value is the percentage of compositions containing material type X that also contain Y, while the support value is the percentage of all compositions containing both X and Y.

In this example, the data set used for the Apriori algorithm included over 60,000 logging compositions recorded by geologists. The Apriori algorithm can also be utilised independently for each of the three stratigraphic classes, since material types and/or association rules may vary according to stratigraphy. For example, where kaolinite is present, depending on the stratigraphic class, the kaolinite should be logged as either the clay type (detritals) or the shale type (bedded). As another example, banded iron formation should only be logged in bedded strata class, and therefore should not be logged elsewhere. Association rules can be developed with a minimum support value of 0.1% (per stratigraphic class), and a minimum confidence value of 0.1%, to identify only significant trends in compositions. From the association rules, a list of subsets of geologically valid material types can be developed for each stratigraphic class, and ranked according to the most common subsets, as shown in Table 3 below. Notably, the frequent presence of clay in the detritals class, high grade hematite and goethite types in mineralised bedded class, and shale in shale intervals, is expected.

TABLE 3: Commonly co-logged subsets of material types

The optimisation criteria discussed above, including the material type association rules, are stored as a database in the data storage 224 of the system 200 together with the logging and assay data. The logging controller 220 can thus refer to these rules when executing a logging process.

Applications of the logging criteria during operation of the system 200 according to embodiments will now be discussed .

In general terms, the association rules developed above are used to assist in finding compositions satisfying known combinations of material types. Other physical information logged during examination of the logged sample, such as colour and hardness, may also be used.

Figures 6A and 6B shows the logging process 600 for logging data the logged sample according to an embodiment. The process 600 comprises two main sub-processes for adjusting the logging data: a material type composition modification process 610; and an optimisation process 620.

The material type composition modification sub-process 610 comprises either or both of a material type addition step 611 and a material type removal step 612.

In the material type addition step 611, selected material types are added to the composition in order to complete a mineralogical-hardness spectrum of material types. This covers two aspects: the division of material types into hard, medium and friable hardness classes; and the mineralogy of specific groups of material types, namely goethite, hematite, and hematite-goethite material types (or alternatively, goethite, martite, and martite-goethite types). It should be noted that a goethite material type is predominantly - but not purely - goethite, and similarly, a hematite material type is predominantly - but not purely - hematite. A hematite-goethite material type is a matrix of both hematite and goethite though not necessarily a 50-50% mixture. The minerals of hematite and goethite have been presented here for illustration in the context of iron ore mineralogy and this system is not restricted to adding only these material types.

In one instance of addition, where a friable goethite material type and a hard goethite material type are specified in the current estimate of the material composition of a sample, a medium hardness goethite material type is added to the estimate, since in practice it is unlikely that a hard type and a friable would exist without a medium type present. Similar rules apply for hematite material types, and for hematite-goethite material types. In another instance of addition, where a goethite type and a hematite type of the same hardness (say, medium hardness) are specified in the composition, the hybrid hematite-goethite type is added to the composition, as it is unlikely that the predominantly hematite and predominantly goethite material types would exist without a hybrid type also present.

More generally, the logging controller 220 is provided with addition rules configured to add to an estimate of material composition one or more additional material types when defined conditions are met in the estimate of material composition—for example, as described above, when the presence of two materials A and B implies the presence of a third material C.

In the material type removal step 612, material types that have an initial proposed composition less than a given threshold, for example 2%, are removed unless they have been predetermined (i.e. a rule has been stored in the logging controller 220) to occur in such trace amounts.

Some examples of material types that are allowed to occur in trace amounts can include: pyrite, pyrolusite, dolomite. Derivative compositions can be formed for processing, where each derivative composition contains the material types from the original, but with each material type excluded in turn. Also, from the material types in the input composition each combination of possible pairs of material types are enumerated, and a derivative combination formed with that pair of material types removed. Each of these combinations are provided to the optimisation function to be considered in parallel. With particular reference to Figure 6B, the process 620 involves using an optimisation function (step 622), wherein the optimiser 280 calculates proposed optimum percentages for each material type by minimising a cost function 624 and applying constraints 626.

In general terms, the cost function provides an indication of a degree of variation between a theoretical logged composition and the provided analytical data such as the laboratory assays via assay module 270, and/or a variation from one of the initial estimates provided to the optimisation module 280 (e.g. initial material type abundance estimate and lump percentage estimate), in order to satisfy the other objective in the optimisation. Evaluation of the cost function is performed by a cost evaluating component of the optimiser 280.

According to this embodiment, the cost function is a function of three error components: assay error (E _as sa _y ), hardness change (Ehardness) and lump error (Ei _U m _P )· Thus, lower cost function values are better. Each component of the cost function will now be discussed in more detail.

Firstly, in relation to the assay error (E _as sa _y ), the cost function utilises an assay error tolerance factor, which is the absolute assay percentage error relative to a predetermined tolerance value for each component of the logged composition. An assay error tolerance factor of 1 represents the largest allowable absolute assay error for that component.

The assay error tolerance values are predetermined and set independently for each component, and may vary according to different requirements. These are also stored in the data storage 224. For example, a lower level of accuracy for logging of low-grade (waste) drilling intervals may acceptable. In one example, the following assay error tolerance values may be used:

Table 4: Example assay error tolerance values

The logging controller 220 then retrieves the predetermined error tolerance values from the data storage 224. All solutions of the cost function having theoretical assay error tolerance values within the respective tolerance of the laboratory assay value are considered equally valid.

Further, a minimum assay error tolerance factor of 0.5 is enforced during optimisation. This avoids unnecessarily optimising the compositions to fractions of a percent when compositions are generally presented to the user to the nearest integer percentage for simplicity.

For an element or compound 'a' (i.e. Fe, S1O2 etc.) _/ laboratory assay value '1/ , theoretical assay value 'T', error tolerance value 'e', and tolerance factor weighting 'f', the assay error component E _as sa _y is given by:

Errors in Fe, S1O2 and AI2O3 are more significant in terms of grade than for other elements which generally occur in trace amounts. Therefore, their respective tolerance factors may be doubled before summing the tolerance factors for all elements. Secondly, the mineral hardness change component (Ehardness) is taken into account to preserve information regarding the RC chip hardness recorded in the original logging data. In this regard, each material type has a theoretical or predefined hardness value. The theoretical hardness of a sample can thus be estimated using the percentages of material types in the initial logging, and the predefined hardness value for respective material types. Therefore, logged material types for a drilling interval (and their intermediate states) can also be divided into three categories: hard, medium and friable.

For each hardness category, the optimiser 280 calculates the differences in the hardness values between the original logging data and proposed optimised data, minus a grace change in hardness of 10%, to allow for minor changes in hardness without penalty. To calculate the hardness error component, a change in hardness A _h is computed as follows:

The (total) change in hardness A _h therefore comprises a sum of the max function calculation for each hardness category. The max function prevents negative values from being included after subtracting the grace change in hardness.

The hardness error component E _hardness is then provided using a Gaussian function:

D 2 ft

E hardness = exp(—

0.3 *0.25 ² )

(10) The constant value of 0.3 in Eq. 10 is used to adjust the weighting, and was determined empirically. The standard deviation value of 0.25 was derived from the training data. Thirdly, regarding lump error (Ei _UmP ), each material type also has a theoretical lump percentage. A lump percentage for each material type provides a breakdown of the ore into lump (particles >6.3mm or 0.25" in diameter) and fines product. In contrast to the hardness measure of the logging data, which is a qualitative material property, the lump percentage is a quantitative measure. Notably, for the same material type, the lump percentage (like other properties) may vary across different sites, and material type grades can also vary for the resulting lump and fines product at the same site. Typically, the Fe grade is higher for lump product.

Since lump and fines products are marketed separately, changes in the lump percentage as a result of a logged composition being modified may have significant commercial implications. Thus, the lump error is taken into account in an attempt to maintain similarity between the theoretical lump percentage for the proposed optimised data and an initial lump percentage estimate, for example derived from the FTIR spectrum or a sample photograph (as shown in Figure 5).

In this embodiment, a sigmoid function as shown below is used to calculate the lump error component Ei _ump from the change in the lump percentage Di:

E lump 0.5 The denominator of '50' in the squared term controls the rate of drop-off of the error value. The result ranges from 0.5 (due to the constant term of 0.5) to 1 (when Di = 0).

The cost function used in the optimisation process is then derived using E _as sa _y , Ehardness and Ei _um p as follows:

E assay * E hardness

E total * (1 + n)

E lump

(12)

In the above formula, 'n' is the number of components with theoretical values arising from the proposed optimised data varying from the assay values by more than the tolerance amount.

The optimisation function may be implemented using the ALGLIB™ optimisation package provided by the ALGLIB Project. The optimisation function uses the cost function and boundary and/or linear equality constraints.

The boundary constraint may ensure that the percentage for each material type lies between 0 and an upper bound, which is the percentage of that material type that would cause the theoretical value for any element to be exceeded by the error tolerance. In other words, this ensures that an error tolerance for any component cannot be exceeded by a single material type.

Further constraints may be applied to specific material types, for example, textural types such as pisolite, where only a small variation in material type percentage is allowed. Moreover, during logging, textural types are rarely confused with other material types and thus should not be removed. Such specific material type constraints may prevent the entire removal of textural material types, thus preserving accuracy. Finally, a linear constraint may also be used to ensure that the material types' percentages sum to 100%.

According to a specific embodiment, the optimisation module 280 applies further restrictions to optimisation of material type percentages such that the sum of the optimised percentages of a set of material types is within some tolerance percentage of the sum of the input percentages of the same set of material types. The tolerance may be, for example, 10% or 0%, or a threshold between. This allows for material type abundances to be transferred only within the set of material types, within the given tolerance. The set of material types may be based on mineralogical characteristics, or determined from similarities in the estimated FTIR spectra of material types, optionally allowing for reasonable confusion in the FTIR material type estimation step.

In one example, the sum of percentages of particular predefined material types are kept constant during optimisation. For example, in relation to iron ore explorations, the material types SHL and CLA are kept constant during optimisation. In another example in the same field, the sum of percentages of the material types HGF and HGM are kept constant during optimisation. In another example in the same field, the sum of percentages of the material types GOE and GOV are kept constant during optimisation. In another example in the same field, the sum of percentages of the material types CHT, BIF, BPO, GOE,

GOE and GOV are kept constant during optimisation. Other predetermined sets of material types are possible. In another example, the sum of percentages of material types with a goethite mineralogy are kept constant during optimisation. In another example, the sum of percentages of material types with a hematite mineralogy are kept constant during optimisation. In another example, the sum of percentages of material types with a kaolinite mineralogy are kept constant during optimisation. In another example, the sum of percentages of material types of other detectable mineralogy are kept constant during optimisation. In an example, the material type groupings determined by the search of step 340 are used as the particular predefined material types.

According to a specific embodiment, the optimisation function is an iterative function. In each iteration, the current state is formed from the material type percentages of the intermediate state, and the gradient of the cost function is estimated from the intermediate state at that iteration .

The dimensionality of the gradient of the cost function is equal to the number of material types being examined. The gradient in each dimension is estimated by:

• first, temporarily altering the intermediate state for this dimension by adding 1% to the corresponding material type value, while reducing the percentages of the other material types by 1/(N-l)%, where N is the number of material types, thus ensuring that the total composition remains at 100%;

• second, evaluating the cost function at the altered intermediate state for this dimension; and

• third, calculating the difference between the cost function value for the intermediate state, and the cost function value for the altered intermediate state for this dimension.

The gradient of the cost function is used to determine the proportions in which the material type percentages will be changed. In this example, the magnitude of these changes are controlled by a constant step length provided by the ALGLIB™ optimisation algorithm, and the supplied constraints are used to enforce bounds on the magnitude such that the percentages of each material type remain valid as described above.

In an embodiment, the optimisation function iterates until a condition is met, for example:

• when the magnitude of the gradient is less than a predetermined value (i.e. the cost function has reached a local minimum from where there is no clear direction for improvement); or

• when the change in the cost function in successive iterations is less than a predetermined value (i.e. the cost function has reached a local minimum); or

• where the change in composition in successive iterations is less than a predetermined value (i.e. there is negligible change in material type percentages); or

• a maximum number of iterations, e.g. 10, 20, 30, has been performed.

Using the cost function and constraints according to the embodiment described above, the optimiser 280 provides a single solution, for each intermediate state resulting from the material type composition modification process 610, regardless of the initial percentages of each material type. This produces optimised intermediate states (step 622). Moreover, when solved for a particular element in the logged composition, a resulting value of the cost function may be used to rank the intermediate states. This will be discussed in more detail below. Notably, when percentages of material types are modified according to the optimisation process 620, it is not necessary to compensate for the change in percentage since the optimisation process will find the appropriate percentages of material types of the intermediate states that best fits the laboratory assays, hardness distribution and the lump percentage.

After the optimisation process 620, the logging process 600 comprises executing an intermediate state penalty process 630.

In the penalty process 630, once the material type percentages have been optimised, the logging controller 220 determines whether a penalty applies according to the various geological conditions (step 632), and applies a corresponding penalty if applicable. In particular, an intermediate state penalty is applied to geologically unusual combinations of material types in the intermediate state.

The intermediate state penalty according to this embodiment is in the form of a numeric multiplier applied to the cost value of an intermediate state. Large penalty multipliers (e.g. 4-8) may be used so that a prospective match of an intermediate state with the assayed composition must be to a sufficient degree to counteract the penalty.

One or more geological conditions, such as stratigraphy, conflicting and prohibited material types, texture, hydration, and hematite-goethite continuity, are used as the basis for penalties. In this example, where the same condition is violated multiple times, penalties are applied repeatedly for each violation. Various penalty types according to specific embodiments are discussed in more detail below.

• Conflicting and prohibited material types. Some material type combinations are geologically inappropriate. For example, there are two kaolinite types: clay in hydrated and detritals intervals; and shale in unhydrated intervals. These two kaolinite types should not be logged together or logged in the wrong stratigraphy. In practice, doing so may lead to geological misunderstandings during modelling, thus a penalty is applied to prevent these situations. A penalty is also applied for combining material types predominantly comprising one element, e.g. gibbsite (alumina) and quartz (silica), in place of a kaolinite type which is high in both elements.

• Distinctive material types. Penalties are applied to prevent the complete removal of a material with distinctive texture, or addition of a material type with distinctive appearance if not originally logged, since the geologist is likely to have logged the material type if present.

• Hydration. Some material types such as vitreous goethite have characteristics arising from hydration. Therefore, these material types should only be included in a composition if the drilling interval associated with the logged data being validated is in a known hydrated zone. Accordingly, a penalty is applied if these material types are included in compositions from non-hydrated zones. Conversely, a penalty is applied if a material type, which never occurs in the hydrated zone, is included in a composition from a known hydrated zone. • Hematite-goethite continuity. Recall that in Table 1 above, various compositions of hematites and goethites are shown at different levels of hardness (friable, medium, and hard). In reality, such compositions occur naturally in a continuous spectrum of hardness. Thus, it is unusual for the hard H2H type to be logged with the friable H2F type without the medium H2M type also being logged. A penalty is applied when the continuous spectrum of hardness is broken in the logging data or intermediate states. Similarly, Table 1 also shows predominantly goethite material types of varying hardness (GOV, GOE, GOL) and intermediate hematite- goethite types (HGH, HGM, HGF). It is unusual for a predominantly goethite material types to be logged alongside a predominantly hematite material type if an intermediate type is not also logged; thus a penalty is also applied in that situation.

For each intermediate state, the penalties described above are accumulated to provide an intermediate state penalty (step 634). This total penalty is then multiplied by the respective cost function value calculated from cost function used in the optimisation process 620 (step 636). This product is used to rank the intermediate states (step 638). The logging controller 220 then preferably selects a predefined number of the highest ranked intermediate states (step 639), i.e. the intermediate states with the lowest product of their respective cost function values and intermediate state penalties. For example, between around 30-50 of the highest ranked intermediate states may be selected .

A penalty determiner executes a final selection process. Penalties may be applied for unlikely material type associations. For convenience, the modified logged composition immediately prior to the final selection process 640 may be referred to as the "penultimate states".

Recall that during logging, colours of each region of interest (or drilling interval) may also be logged. The final selection process 640 comprises a colour penalty process 642, which involves examining, for each material type, the logged colours provided as part of the training data or a colour derived from a photo sample. More specifically, the colour penalty process 642 comprises:

• For a particular material type IH _N in the penultimate composition, examine past logging data, identify those that also logged IH _N , and the colour logged for the associated drilling interval.

• Then, determine a percentage of the past data identified with IH _N that have also logged the same colour as the colour logged with m .

• Multiply the percentage from the preceding step with the percentage of the material type in the penultimate composition .

• Sum the value obtained from the preceding step for all material types to determine colour penalty values p _coi .

A minimum colour penalty value of 0.5 is used to avoid small values arising where little training data is available. Therefore, the colour penalty values p _C oi lie in the interval [0.5, 1].

Similarly, the frequencies of the logged chip shapes (angular, sub-angular, rounded, sub-rounded, or combinations thereof), and stratigraphic class for each material type are examined to determine other penalties, such as a chip shape penalty p _C hi _P and stratigraphic class penalty p _s trat.

For example, the logging controller 220 may be configured to determine p _C hi _P by executing the following steps:

• retrieve the logged chip shapes of past logging data and determine the historic distributions of logged chip shapes for each material type from past data;

• for each particular material type IH _N corresponding to the penultimate composition, where N is the number of material types, determine the percentage of the past data identified with the material type IH _N that also have logged the chip shape logged for the associated drilling interval;

• multiply the determined percentage with the percentage of the material type in the penultimate composition;

• sum the product of the percentages obtained from the previous step for all material types to determine the chip shape penalty p _ChiP .

In a further example, the logging controller 220 may be configured to determine the stratigraphic class penalty Pstratby executing the following steps:

• determine the historic distributions of material types for stratigraphic classes from past data;

• for each particular material type IH _N corresponding to the penultimate composition, determine the percentage of the past data identified with IH _N that lie in the same stratigraphic class as the associated drilling interval; • multiply the determined percentage with the percentage of the material type in the penultimate composition;

• sum the product of the percentages obtained from the previous step for all material types to determine the stratigraphy penalty p _s trat.

A minimum chip shape penalty value of 0.5 is used to avoid unduly small values. A minimum stratigraphy penalty value of 0.5 is also used to avoid unduly small values. Thus, the resulting chip shape penalty p _ChiP and stratigraphy penalty Pstrat also both lie in [0.5, 1].

Lastly, the selection process 640 comprises an association penalty determination step 648. In this step, the material types in each penultimate composition is examined by utilising the association rules to penalise combinations of material types not seen in the past data used to develop the association rules. A score is calculated based on the association rules and confidence values determined by the Apriori algorithm described above.

The score is computed for a set of N material types by first numbering all subsets of N-l material types. For a given subset S, if an association rule exists for the subset, the score is the highest confidence value between the individual material types rrg and m ₂ , where rrg e S and m ₂ ? S . If no such association rule exists, a similar process is performed for subsets of size N-2, and the score computed using the product of the two confidence values, each derived by taking into account one of the material types excluded from the calculation.

As an illustrative example, say there is a combination of four material types in the logging data: A, B, C, D. In this example, the association rule for the set {A,B,C,D} does not exist in the association rules database, but an association rule does exist for the set {A,B,C}. The confidence values between individual material types may then be calculated using the Apriori algorithm to link the combination {A,B,C} to the absent type D. In other words, the confidence values between individual material types A- D, B-D and C-D are calculated. The score mentioned above is designated as the maximum confidence value out of A-D, B-D and C-D. Therefore, m2 is considered as D and A, B and C are in turn considered as ml.

Further in this example, where there is no association rule for {A,B,C} or any other triplet, N-2 is considered. In this case, if there is an association rule for {A,B}, confidence values for the pairs A-C and A-D, or B-C or B-D, or A-C and B-D, or B-C and A-D, are calculated. The confidence values for each calculated pair is multiplied to obtain the score.

The confidence value, and therefore the association penalty P _assoc , are in the range (0,1] and is applied to the final penalty by dividing it by p _aSsoc

The final penalty p _finai is then determined (step 643) as follows:

Using the sum of the p _finai value and the cost function value derived during the optimisation process 620, the logging controller 220 ranks the penultimate states (step 645) and the top-ranked state selected (step 647) as the material type logging which is representative of the sample, which is stored in a material type logging database (either in data storage 224 or separately). Anomaly Detection

Figure 7A shows system 200 including an anomaly detector 290 for analysing FTIR spectra, for example, as received by the FTIR logger 230. The anomaly detector 290 is configured to identify potentially anomalous FTIR spectra, which can be labelled as such and/or communicated to a suitable user interface for inspection. The anomaly detector 290 can be implemented by the logging controller 220 (as shown) or can be implemented separately, for example, implemented by a physically or logically distinct server to the logging controller 220.

In an embodiment, the anomaly detector 290 implements a neural network suitably trained for anomaly detection. The anomaly detector 290 can be configured for anomaly detection for FTIR spectra of a particular class of samples, such that there can be, in effect, one or more anomaly detectors 290, each characterised by a particular class of FTIR spectra. In one example, a class corresponds to FTIR spectra of samples from a particular region.

For example, when considering samples taken from a region that is relatively well-known mineralogically, it may be surmised that the samples' mineralogical compositions lie within some expected ranges and so their corresponding FTIR spectra would be similarly constrained (and therefore, belong to a same class). In practice, of course, geological modelling is complex and heterogeneous, resulting in some samples' compositions lying outside of the expected range. Analysis of these samples' FTIR spectra, with the context of other FTIR spectra, may advantageously be used to identify anomalous geology. Figure 7B shows a method for training of the anomaly detector 290 for a nominal class of sample. At step 700, FTIR spectra are obtained for samples of the particular class ("training samples"). The FTIR spectra can be based on previous samples associated with the class or can be acquired for the purposes of training (or both).

In an example, samples used for anomaly detector 290 training were collected from reverse circulation drilling at 2-meter intervals as part of a region wide study (Pilbara)—therefore, the class in this case may be defined by the Pilbara region. The pulps were ground into powders at 150 pm. FTIR spectra were collected on a DRIFTS style Thermo Fisher Nicolet iS50 over wavenumbers (v) ranging from 232 — 6, 000 cm ^-1 using a series of scans with a caesium iodide source and boxcar apodization that was determined as optimal within the lab such that it that maximized their workflow. The initial FTIR spectra resolution was resolved at 8 cm ^-1 and subsequently resampled to ~3.85 cm ^-1 for a total of 1,453 wavenumbers. Wavenumbers < 400 cm ^-1 were filtered out due to excessive signal noise. Total reflectance (0-100%) was collected for each sample within this study on the drilling pulps of chips. These pulps constitute pulverized mixtures of the chips, which represent the lithological units over the 2-meter interval. Therefore, the each FTIR spectrum represents a texturally complex and mineralogically heterogeneous sample that is compositionally unique. The samples used in this example consist of a mix of an initial training set of waste banded iron formation ("BIF"; n=1579) for training and validation. Data was split during training into a testing and validation split of 80/20 for the samples.

At step 701, the neural network of the anomaly detector 290 is trained using the training samples. The neural network can be trained according to unsupervised methods. In an embodiment, the neural network utilises latent variables—for the purposes of this disclosure, the neural network is assumed to implement a variational autoencoder model (VAE). The VAE is a deep generative model that assumes that the original dataset follows an underlying probability distribution, which can be trained to create new data (Kingma and Welling, 2019; Pereira and Silveira, 2018). A VAE is similar in its architecture to a traditional autoencoder; however, it differs in that latent distribution parameters modelled per-sample rather than as point estimates. In a VAE, a latent variable z* is sampled from a prior Gaussian distribution Pe(z). For practicality, the true posterior (r _q (z\c) is approximated through a parametric inference model q ₍ p(z\x)« R _f (z\x).

In general terms, the VAE is applied to model the FTIR spectra. Latent representations are learned from a number of samples and in broad terms can compress the important features of the spectra into a lower dimensional representation (i.e. the set of latent variables Z _f ) that may be used to generate new spectra.

The encoder/decoder design of the VAE, alternatively called the inference/generative network, employs a multilayer perceptron architecture with a feed forward linear neural network with rectified linear units (ReLU; Nair and Hinton, 2010) and SoftPlus (Zheng et al., 2015) activations at the ends of both the encoder and decoder portions of the network (Figure 7F shows a general model of a VAE). The sizes examined in these models are two hidden layers and a latent layer ((1453 ® 800 ® 400 ® L/ _;aίbhί (m _z , )). The decoder is similar within the hidden layer structure and number of nodes per layer. The final layer of the decoder is altered from traditional VAEs (e.g., Kingma and Welling 2014) to predict a collection of independent Gaussian random variables (in practice an array of m and s ² components). This architecture setup is important since it can be leveraged to calculate a reconstruction probability (An and Cho, 2015; Pereira and Silveira, 2018; Xu et al., 2018).

This is contrasted to a point estimate reconstruction as is commonly done in known conventional VAEs (e.g., Kingma and Welling, 2014). In a conventional VAE, the primary objective is to obtain a reconstruction in a generative sense. The modified structure of the VAE described herein allows for computing the "reconstruction probability", so that the log-likelihood of an initial FTIR sample can be computed given that some approximate posterior distribution can be computed. Such an approach may advantageously provide improved performance compared to existing approaches. The approximate posterior parameters for the mean (m _z) and variance (s ² ) are calculated at the end of a final hidden layer ReLU (Nair and Hinton, 2010) and SoftPlus activations. For the purposes of this disclosure, the prior pg(z) is assumed to follow a Gaussian distribution (i.e. r _q (z)= J\f(0,1)) due to the continuous nature of the data.

According to an implementation, a VAE loss function is calculated through maximizing the evidence lower bound (ELBO). This is accomplished through a combination of minimizing the Kullback-Leibler ("KL") divergence between the posterior and prior distributions of the latent variables (Eq. 14; right hand term) and maximising the reconstruction probability between the (learned) decoded sample (pg(x|z); left hand term) and the input FTIR sample (x‘) (Eq. 14; left hand term): The KL divergence in Eq. 14 acts as a regularizer and is calculated from the entropy between the learned distribution in the inference network (<?f(z|c)) and a prior distribution ( Pe(z)) of the latent variables z. Minimising the KL divergence ensures that the latent posterior distributions ¾y(z|c) do not deviate significantly from their prior . A standard Gaussian (J\T(0,1)) can be chosen for the prior . Minimising the second loss term [logp0(x|z)]) maximises the log-likelihood between the reconstructed and original sample.

The VAE is therefore trained using the training samples such that a set of latent variables are determined from which an input FTIR spectrum can be reconstructed from the latent variables as a "pseudospectrum". In a general sense, anomaly detection is based on a difference between an input FTIR spectrum and its generated pseudospectrum.

In a general sense, anomaly detection within unsupervised generative models can be accomplished through the reconstruction error (RE) or a reconstruction probability (RP). Since no anomalous labels exist within the FTIR samples (the FTIR spectra are not annotated), it is assumed that the training sample represent an imbalanced dataset (e.g., W-non- anomalous ^ ^ anomalous)·

The anomaly detector 290 therefore implements an anomaly scorer 295 configured to calculate a "score" indicative of a difference between an input FTIR spectrum and the pseudospectrum generated by the VAE based on that FTIR spectrum. The anomaly scorer 295 can be configured to calculate a score for each FTIR spectrum analysed, the score indicative the similarity (or equivalently, difference) between an associated FTIR spectrum and its generated pseudospectrum. In an embodiment, the anomaly scorer 295 calculates a reconstruction error (RE), for example according to:

The reconstruction error determines the mean absolute difference between the mean of the reconstructed samples to that of the initial input (x) averaged over a set of randomly chosen samples (L). The error is calculated at each unique wavelength and summed over the entire length of data (i.e. the entire FTIR spectrum). Generally, larger reconstruction errors correspond to poorly reconstructed FTIR spectra (a low similarity between input FTIR spectrum and generated pseudospectrum) with the largest errors indicative of an anomalous FTIR spectrum. Note in this case, E(r _q (Ci|z _f ) represents the mean m of the output layer of the decoder. The reconstruction error can therefore be interpreted as the score for the particular FTIR spectrum.

In another embodiment, since the VAE effectively directly predicts a probabilistic FTIR spectrum through reflectance distribution parameters (m _h , , the anomaly scorer 295 can calculate a reconstruction probability from a single (probabilistic) reconstructed sample. An advantage of the reconstruction probability may be that it leverages the probabilistic nature of the VAE's reconstructions. The reconstruction probability is computed after the model is trained. The FTIR spectra are processed by the trained encoder of the VAE to produce latent variables, which are subsequently taken as input by the decoder of the VAE to produce probabilistic reconstructions (parameterized by m _h and s _h) . The reconstruction probability can be calculated through Monte Carlo sampling (L):

^ _z ~^ ₍ z _| x ₎ [logp

A reconstruction probability is calculated for each wavenumber and then summed across the entire FTIR spectrum. Therefore, the reconstruction probability is used as a cumulative sum in the threshold as the detection method across all wavenumbers to determine a total score for a particular FTIR spectrum.

The training samples themselves may comprise anomalous samples, however, this is not a priori known (i.e. the training is unsupervised). The presence of anomalous sample (s) can therefore adversely affect the training of the anomaly detector 290. Therefore, in an embodiment, at step 702, after initially training the neural network on all training samples (i.e. step 701), the anomaly scorer 295 calculates a score for each training sample. At step 703, the scores are assessed against an expected distribution (for example, as assumed herein, a normal distribution) . If the scores are not normally distributed, then at step 704, training samples are removed with a score outside of the distribution.

The VAE is then retrained using the reduced set of training samples, at step 705, which are expected to better reflect non-anomalous FTIR spectra for the particular class. Further training of the VAE can be undertaken at step 706, either after reducing the set of training samples at step 705 or if there are no training samples determined to be outside the normal distribution at step 703. The further training can be in respect of optimisation and hyperparameter testing.

In an embodiment, learning and updating of the weights of the VAE can be achieved through the Adam optimizer (Kingma and Ba, 2014). For example, different learning rates (10 ⁵ , 10 ⁴ , and 10 ³ ) and/or different latent dimensions (i.e. number of latent variables such as 2, 10, 20, 40, 80, and 160) can be tested. Testing can be based on sampling during the calculation of the score (e.g. reconstruction probability), in order to determine the optimum learning hyperparameters that maximise the reconstruction probabilities .

In an implementation, sampling of the prior during the reparameterization trick was done with one sample, (similar to Kingma and Welling (2014)). During the anomaly detection, data can be sampled at different rates (for example, 10 and 512) to ensure no sensitivity to the sampled set. Numerical computations in one experiment were performed on a workstation with an RTX 3090 with 24 GB of ram and Intel i9-10900KFU with 64 GB of ram. The particular VAE (for example, as defined by its hyperparameters) chosen for implementation for the anomaly detector 290 can be that which, during testing, produces the least overall reconstruction loss.

Figure 7C shows a method of anomaly detection according to an embodiment. The method utilises a trained anomaly detector 290 as described with reference to Figure 7B. At step 710, a FTIR spectrum is provided for which an anomaly detection is required. Similarly, at step 711, a suitable anomaly detector is selected based on the class of the FTIR spectrum (although, in implementations with one anomaly detector type, this will automatically be selected) . For example, if the FTIR spectrum is known to be for a sample taken within a particular geographic region or with particular geologic properties (which, as described above, can be a "class"), then a suitable anomaly detector 290 for that region is selected.

At step 712, the anomaly detector 290 analyses the FTIR sample to determine a score (e.g. in an embodiment, a reconstruction probability)—the score is dependent on the training sample used for training of the anomaly detector 290, as described herein.

At step 713, the score is assessed against a suitable predefined threshold, which can optionally be a user- settable parameter. The predefined threshold is selected such as to identify anomalies, where anomalies represent larger deviations between the FTIR spectrum and its generated pseudospectrum compared to non-anomalies. In one example, the predefined threshold is a reconstruction probability less than 5% (e.g. less than the 5 ^th percentile) .

If the assessment at step 713 indicates that the FTIR spectrum is anomalous, then it is flagged as such, at step 714. For example, by recording an anomaly flag in a suitable data structure in which the flag can be associated with the FTIR spectrum. In an implementation, the flag can constitute metadata associated with the FTIR spectrum.

If the assessment at step 713 indicates that the FTIR spectrum is non-anomalous, then it is flagged as such, at step 715. For example, by recording a non-anomaly flag in a suitable data structure in which the flag can be associated with the FTIR spectrum. In an implementation, the flag can constitute metadata associated with the FTIR spectrum.

Alternatively, an actual data record is only made if the FTIR spectrum is to be flagged as anomalous (or, alternatively, to be flagged as non-anomalous). The lack of a recorded flag can then be interpreted as implying the opposite (i.e. non-anomalous or anomalous, respectively).

Figure 7D shows an example of an original FTIR spectrum 750a and its generated pseudospectrum 751a. In this case, the two spectra are very similar with the largest deviations in the small wavenumber region. Figure 7E shows an example of an original FTIR spectrum 751a and its generated pseudospectrum 751b. In this case, the two spectra are quite dissimilar. The sample associated with Figure 7E can intuitively be understood to be more likely to be classified as anomalous than that of Figure 7D; note that this is due to the poorer fit between the original FTIR spectrum and that of is generated pseudospectrum, which is a reconstruction of the spectrum using the VAE trained on FTIR spectra assumed to be non-anomalous (for example, due to the filtering step 704.

Embodiments disclosed herein are based on the realisation that during the logging process, it is appropriate to utilise a processing machine for some aspects of the logging data; however, for other aspects it is also appropriate to preserve the initial input by a user, since such input is likely to be correct. For instance, while it is appropriate to utilise a machine for adjusting the estimated compositions based on the material types logged, the machine processes for adjusting the compositions ought to be guided by the physical properties of the mineral sample logged and other known geological factors of the region of interest. Therefore, according to embodiments herein described, it is desired that the proposed validated compositions are those that depart least from the original physical properties logged as a result of the adjustments made to the logging data.

Reference herein to background art is not an admission that the art forms a part of the common general knowledge in the art, in Australia or any other country.

In the claims which follow and in the preceding description of the invention, except where the context requires otherwise due to express language or necessary implication, the words "comprise" and "include" or variations such as "comprises", "comprising", "includes", or "including" are used in an inclusive sense, i.e. to specify the presence of the stated features but not to preclude the presence or addition of further features in various embodiments of the invention.

References

Komarek T., Somol P. (2019) "Multiple Instance Learning with Bag-Level Randomized Trees". In: Berlingerio M.,

Bonchi F., Gartner T., Hurley N., Ifrim G. (eds) Machine Learning and Knowledge Discovery in Databases. ECML PKDD 2018. Lecture Notes in Computer Science, vol 11051. Springer, Cham. https://doi.org/10.1007/97 8-3-030-10925-7_16

Breiman, L. Random Forests. "Machine Learning" 45, 5-32 (2001). https://doi .org/10.1023/A:1010933404324

Ebner, Marc. "The Gray World Assumption", Color Constancy. John Wiley & Sons, 2007. ISBN 978-0-470-05829-9.

Ebner M. (2003) Combining White-Patch Retinex and the Gray World Assumption to Achieve Color Constancy for Multiple Illuminants. In: Michaelis B., Krell G. (eds) Pattern Recognition. DAGM 2003. Lecture Notes in Computer Science, vol 2781. Springer, Berlin, Heidelberg, https://doi .org/10.1007/978-3-540-45243-0_9

Hartley and Zisserman, Multiple View Geometry in Computer Vision, Second Edition, Cambridge University Press, 2004.

Fernandez, P. D. M., F. A. Guerrero-Pena, T. I. Ren, and G. J. J. Leandro, "Fast and robust multiple ColorChecker detection using deep convolutional neural networks, " Image and Vision Computing, Volume 81, 2019, pp. 15-24.

Peter Kovesi, "Image Features From Phase Congruency", Videre: A Journal of Computer Vision Research. MIT Press. Volume 1, Number 3, Summer 1999. Peter Kovesi, "Edges Are Not Just Steps", Proceedings of ACCV2002 The Fifth Asian Conference on Computer Vision, Melbourne Jan 22-25, 2002. pp 822-827.

Simonyan, K., Zisserman, A., 2015. Very deep convolutional neural networks for large-scale image recognition, in: 3rd International Conference on Learning Representations, ICLR 2015, Conference Track Proceedings.

Garcia-Gasulla, D., Pares, F., Vilalta, A., Moreno, J., Ayguade, E., Labarta, J., Cortes, U., Suzumura, T., 2018.

On the Behavior of Convolutional Nets for Feature Extraction. Journal of Artificial Intelligence Research 61, 563-592 . https://doi.org/10.1613/jair.5756 ColorChecker by Bruce Lindbloom, see: http://www .brucelindbloom.com/index.html?Eqn_ChromAdapt.htm 1 (retrieved 4 Feb 2021).

An earlier version of this webpage also describing the Chromatic Adaptation method can be found at https://web .archive.org/web/20120107033440/http://www.bruce lindbloom.com/index .html?Eqn_ChromAdapt.html.

Berkman, D. A. & Australasian Institute of Mining and Metallurgy. (2011). "Field geologist's manual" 4 ^th Ed. Carlton, Vic: Australasian Institute of Mining and Metallurgy .

Kingma, D.P., Ba, J., 2014. Adam: A Method for Stochastic Optimization .

Kingma, D.P., Welling, M., 2019. An Introduction to Variational Autoencoders. Found. Trends® Mach. Learn. 12, 307-392 . https://doi.org/10.1561/2200000056 Kingma, D.P., Welling, M., 2014. Auto-encoding variational bayes. 2nd Int. Conf. Learn. Represent. ICLR 2014 - Conf. Track Proc.

Pereira, J., Silveira, M., 2018. Unsupervised Anomaly Detection in Energy Time Series Data Using Variational Recurrent Autoencoders with Attention, in: 201817th IEEE International Conference on Machine Learning and Applications (ICMLA). IEEE, pp. 1275-1282. https://doi.Org/10.l109/ ICMLA.2018.00207

Nair, V., Hinton, G.E., 2010. Rectified linear units improve Restricted Boltzmann machines, in: ICML 2010 - Proceedings, 27th International Conference on Machine Learning, pp. 807-814.

An, J., Cho, S., 2015. Variational Autoencoder based Anomaly Detection using Reconstruction Probability, Special Lecture on IE.

Xu, H., Feng, Y., Chen, J., Wang, Z., Qiao, H., Chen, W., Zhao, N., Li, Zeyan, Bu, J., Li, Zhihan, Liu, Y., Zhao, Y., Pei, D., 2018. Unsupervised Anomaly Detection via Variational Auto-Encoder for Seasonal KPIs in Web Applications, in: Proceedings of the 2018 World Wide Web Conference on World Wide Web - WWW '18. ACM Press, New York, New York, USA, pp. 187-196. https://doi .org/10.1145/3178876.3185996

Hao Zheng, Zhanlei Yang, Wenju Liu, Jizhong Liang and Yanpeng Li, "Improving deep neural networks using softplus units, " 2015 International Joint Conference on Neural Networks (IJCNN), 2015, pp. 1-4, doi:

10.1109/1JCNN.2015.7280459.