GB2303475A

GB2303475A - Locating mineral deposits

Info

Publication number: GB2303475A
Application number: GB9605227A
Authority: GB
Inventors: George Bolt; Nick Ryman-Tubb; Gavin Daneel
Original assignee: Neural Mining Solutions Pty Ltd; BARNARD ERIC EDWARD
Current assignee: Neural Mining Solutions Pty Ltd; BARNARD ERIC EDWARD
Priority date: 1996-03-12
Filing date: 1996-03-12
Publication date: 1997-02-19
Anticipated expiration: 2016-03-12
Also published as: GB2303475B; AU1933397A; WO1997034169A1; GB9605227D0

Description

Exploration for Minerals This invention relates to the exploration for minerals for example but not exclusively those containing metals.

Man has, of course, explored for minerals for thousands of years. Current techniques rely on the analysis of geological, geochemical and geophysical data by a geologist. Large amounts of multilayered data is visualised, often using image processing and the geologist seeks to locate anomalies which could indicate the presence of a mineral deposit. Different data sets are used to help confirm, or deny, the analysis by assessing distinctive characteristics for anomalous areas perhaps by looking at the data in a different way. This analysis is generally both tedious and costly. It requires assumptions and rigid rules.

According to a preferred embodiment of the invention there is provided a method of locating a mineral deposit the method comprising the steps of: obtaining a plurality of data sets of indicators of physical or chemical conditions at a plurality of sites in an area known to contain the mineral, at least one of the indicators being an indicator ofthe mineral; li training cluster identifying means to identify clusters and anomalies in the data sets; iii training an approximator to identify correlations in the data sets iv obtaining a plurality of the indicators at a plurality of sites in a search area thought to contain the mineral; v inputting the indicators obtained from the search area to the trained cluster identifying means and trained approximator means to obtain an indication of the location of the mineral deposit;; vi visualising the indications; Located deposits can be further explored and exploited for example in known manner.

In the invention a model, especially a neural model, is trained to locate anomalies or clusters with data from an area known to contain a mineral deposit. The model can thus reconcile particular features with the existence or absence of the mineral.

The mechanism involved between the indicator and mineral need not be understood. The data may involve a high degree of redundancy for example including data sets which gives no information about the sought mineral.

At least one set of data should provide a positive or negative indicator of the existence of mineral. Preferably a plurality of data sets provide an indication.

Having trained the model, data from an area thought to contain a mineral deposit is loaded into the model. The model searches for anomalies and preferably clusters and reconciles this with the information already learnt. The model can then point with a high degree of confidence to areas likely to contain the sought mineral.

Preferably neural anomaly detectors and approximators are used.

An SOM is preferably used as the anomaly and clusters identifying means but other means such as ART could be used. Similarly a preferred neural approximator is an MLP but others such as RBF's could be used. The precise means are not the subject of the invention. Those skilled in the art will have preferred means of their own or will develop suitable ones by reference to the text books hereinafter referred to or other reference works.

By using the invention it is not necessary to know what measured parameters are indicators of the sought mineral. This has several consequences. First one can search for a mineral by inputting data sets which might or might not be indicators of the mineral sought provided that at least one is. Hitherto unsuspected relations between a parameter and a mineral can be found. Secondly the information can be used several times over, to search for different minerals.

It will thus be apparent to the skilled worker that no single indicator must be used in accordance with the invention. Thus if one is looking for potassium or for uranium then a good parameter to measure could be radioactivity. If looking for gold then chemical analysis of the area for arsenic which commonly occurs with gold would be appropriate. It is not necessary for the data to have a positive correlation for the mineral sought. As will be noted for the examples herein the presence of copper and/or silver is an indicator that molybdenum levels will be low.

In order to illustrate the invention reference is made to the following examples.

Embodiments of the invention will be illustrated by reference to the accompanying figures of which Fig 1 a map showing the results of an assay using plots for nine minerals of a first survey area.

Fig 2 is a map showing regions of anomaly located in the map of Fig 1.

Fig 3 shows the map of Fig 2 together with a nickel survey map.

Fig 4 a contour plot of the area covered by Fig 1 where the data is restricted to a subject.

Fig 5 is a topological map resulting from cluster analysis of the first survey site.

Fig 6 is a geographical map showing the location of the clusters of Fig 5.

Fig 7 shows a plot of the existence of a cluster overlying the presence of silver.

Fig 8 is a schematic of a display for exploring relationship.

Fig 9 is a neural correlation of data from the survey site A.

Fig 10 is a chart showing factors which influence gold assay level at site A.

Fig 11 is a chart showing the results of neural sensitivity analysis of site B.

Fig 12 is a chart showing the factors which influence white core colour at site B.

Fig 13 is a map showing the results of a neural fUzzy search for silver at site A.

Fig 14 is a map showing the results of a neural fUzzy search for areas containing high levels of gold and low levels of copper and nickel.

Fig 15 is a map of the geographical locations of areas of high copper assay at site B.

EXAMPLE 1 Five sets of surface and aerial data were obtained for an area known to have mineral concentrations as follows: 1. Aerial comprising: 1 Magnetic field and radioactive emissions corresponding to ii Uranium iii Thorium iv 40K and v Total counts 2. Gravity 3. Induced polarity ('IP') 4. Chemical soil surveys for copper 5. Transient electromagnetic ('Sirotem') survey data taken on a 200m grid (EM) The EM data set contained up to twenty components for signal returns at 1 ms intervals.The EM data was pre-processed into seven components hereinafter referred to as 'EM special data': Simple sum ofEM 1 ms to 4 ms ii Simple sum of EM 5 ms to 8 ms iii Simple sum of EM 9 ms to 12 ms iv EM 1 ms signal: A measure ofthe overall initial EM response.

v Immediate decay of 1 ms signal i.e.

EM2002m5 EM2001ms vi Average decay over the period lms to 10 ms. This is the average of proportional decay in neighbouring signal returns over the first 10 ms of raw EM data. The proportional decay is given by the expression

where En is signal return at n ms.

vii Highllow signals over period 13 ms to 20 ms. This is the sum of abnormally high or low signal returns over the period 13 to 20 ms.

An abnormal signal differs from the mean by more than 1.5 standard deviations.

The various data sets were sampled at different locations within the survey area.

To allow cross-layer comparison, the data sets were interpolated. This was done using an inverse weighting algorithm, Data was interpolated onto three different scales: 1. 20x20grid 2. 50x50 grid 3. 100xlOO gnd The 50x50 grid represents objects of about 300m square. This window reflects the standard sized search employed by geologists.

Three groups of data were created and analysed.

1. AN Raw Data Sets 2. Data set 1 with EM data replaced by EM special data 3 Aerial & Soil Geochemistry Data Sets These lead to three sets of four different neural analyses being performed. The four neural analyses were: 1. Cluster Identification 2. Anomaly Detection 3. Correlation Analysis 4. Fuzzy Neural Searching Cluster identification identifies global trends and patterns across the entire survey region.

For example a plot could depict the output map of the Self-Organising Map (SOM) neural computer employed to perform cluster identification. An SOM is also known as a Kohonen network. Details may be found in 'Self Organisation and Associative Memory' by T.Kohonen Springer Verlag 3rd Edition 1989. On the basis of this document and known SOMs those skilled in the art will have little difficulty in devising a suitable SOM. This may for example be a 2-D grid (motto be confused with the 2-D survey region) of 10xlO units. This grid may be divided into a number of separate regions; each region corresponding to a particular cluster identified. The number of units for a particular region indicates the level of generality for that cluster. For example, a cluster 1 could be assigned 2 units for data group "All Data Sets".This would indicate that it represents fairly unusual types of deposit characteristics.

Subsequent plots would show which areas within the survey region match each cluster type.

Anomaly detection locates unusual patterns of survey results. This could be performed using both 10x10 SOM and a 15x15 SOM. Similar results are likely to be obtained from both.

For the 10x10 SOM, anomaly feature plots are given for the top ten anomalies.

The level of anomaly (R=value) and the location in the survey region would be indicated. The feature plot itself indicates the type of deposit within the anomaly.

Values for each survey component are conveniently normalised between 0 and 1 where 0 represents the lowest value and 1 the highest value in the data.

Neural correlation analysis using an approltimator shows how different survey components influence each other. The analysis may need only to be performed across a limited area of the survey region. This limitation could be defined by setting a small "Reality" value. In correlation analysis an interpolated data set could be used to train a neural model known as a Multi Layer Perceptron (MLP).

MLP's are discussed in "Parallel Distributed Processing" by DE Rumelhart and JL McClelland Viol. 1 MIT Press (1986). One of the layers in the interpolated data set may be designed as the variable for which correlations are to be found. This variable is the output target unit of the MLP. The remaining data layers in the data set provide input units. After training the correlation of each input on the output is determined directly from the MLP by measuring the gradient of the relationship between the input and output.

Fuzzy searching locates regions of strong correlation. In a neural fizzy search a search pattern may consist of three elements of the data sets. First the values of interest of, for example, magnetic field, radioactive decay due to uranium and copper concentration in the ground. The second is the flag indicating whether the value of interest is a maximum, minimum or equality. Thirdly a weighting to be accorded to the data set is required. A continuously valid degree or match between the search pattern and the grid points is obtained using a fizzy matching neural computer. The output may be presented in the form of a continuously valid degree of match across the region.

Fuzzy neural searching allows specific patterns of survey data to be located and a measure of how "close" regions match the search pattern.

The trained model may then be used to locate mineral.

EXAMPLE 2 Survey A A survey collected soil geochemical data from an approximately rectangular sample grid consisting of 1,542 locations. Chemical assays were given for the following nine constituents: 1. Cadmium 2. Gold 3. Copper 4. Lead 5. Molybdenum 6. Nickel 7. Silver 8. Zinc 9. Arsenic Assay results were given in ppm (parts per million).

The co-ordinates (northings and eastings) of the locations from where the soil samples were taken. For each co-ordinate, the elevation of the soil sample was given (all elevations in the data supplied were zero).

Figure 1 shows the results for each assay using plots. These contour plots would normally be overlaid, in various combinations by a geologist to determine by observation any correlations and anomalies.

Survey B A survey of a different area gave a dataset which contained results of a variety of analyses from 77 survey drill holes: a) Drill hole information b) Assay results for copper content (ppm) c) Lithological descriptions of the rocks.

Drill hole information gave northing, easting and elevation location, orientation, azimuth and maximum depth reached.

Assays and lithological descriptions were given for every two metre increment of a drill hole. A single assay for copper was Cu in ppm.

Lithological descriptions were divided into four sub-classes: Colours n Physical properties iii Rock types iv Mineral constituents The labels for each sub-class are given in Table 1. Each two metre interval could be given any number of descriptor labels. For example, Grey Muscovite, Pale Green Muscovite, Sericite, Schist, and Quartz Veins.

Once combined, Survey B's dataset consisted of a 90 component vector for each two metre intervals. This broke down into: 1 component for copper (Cu) assay 19 components for colour 18 components for physical properties 30 components for rock type 22 components for mineral constituents The copper assay component was a numeric figure in units of ppm. All other components were represented by a '1' if that lithological feature was present otherwise '0'.

Table 1

Colours Physical Properties Rock Types Minerals White, Silver, Altered, Band, Cavity, Alluvium, Azurite, Basalt, Silver-Grey, Chips, Cleavage, Breccia, Basic Biotite, Carbonate, Black, Black- Crenulated, Cubic, Fine, Calcite, Clay, Chalcocite, Green, Dark Fine Grained, Fractured, Dacite, Felsic, Chlorite, Green, Grey- Fragments, Grained, Gossan, Gravel, Chrysocola, Green, Green, Indurated, Interbedded, Ironstone, Chalcopyrite, Brown, Yellow, Interlaminated, Laminated, Limonite, Epidote, Iron, Pink, Grey, Dark Leached, Lithic, Massive, Mylonite, Mud, Feldspars, Graphite, Grey, Grey- Mottled, Particles, Pebbles, Psammite, Garphitic, Brown, Red Pieces, Platey, Schist, Rhyolite, Rock, Haematite, Golden, Cream, Stained, Trace, Uncleaved, Rubble, Scree, Malachite, Mica, Khaki, Darker, Veined, Vuggy, Weathered, Sediment, Slag, Muscovite, Pyrite, Pale Zone Slate, Soil, Quartz, Sericite, Volcanics Sulphide Detection of anomalous or unusual patterns within survey results is currently performed by an expert geologist employing various visualisation techniques.

Observing in this way is feasible for a small number of overlaid survey maps, but rapidly becomes increasingly difficult as more surveys are considered. In Survey A, soil assays result in a total of nine separate layers. Attempting to observe anomalies from some combination of a nine-layered visual display is impossible.

Useful information is missed by humans.

Neural analysis was performed on this data that very simply and quickly allowed anomalous regions within the Survey A to be identified. A continuous value defining the actual degree of anomaly was generated for each sample location.

Neural anomaly detection is more powerful than simply finding localised regions where a single component from a survey dataset is particularly prominent. For example, regions with higher levels of a certain mineral than ambient levels. It is actually able to detect regions where multiple combinations of survey components are together anomalous with respect to the entire survey region. It is also able to perform the simple single component anomaly detection as well.

The particular neural computing technique employed to perform anomaly detection was based upon a Self-Organising Map (SOM).

The SOM is an unsupervised learning neural model. In other words, it does not require to be told by an expert, during training, the output for example inputs.

Instead, it only takes a collection of example inputs. During training, these are shown; without any explanation, to the SOM as data pattern examples. For example, when training a SOM on the Survey A survey data, individual input patterns consisted of the nine chemical soil assay figures for each sample location.

The SOM automatically forms a special topologically organised map (such that nearby output units in the grid represent similar input patterns) of these inputs on its two dimensional output grid. By this means, a much larger range of possible input patterns can be represented by the SOM than would be possible if the output grid was not topologically organised. A SOM is efficient in representing large complex datasets, such as survey results. The final trained SOM actually represents a model of the structure of data seen during training.

A SOM was trained on the soil geochemistry assays from Survey A. The ninedimensional inputs (one input per chemical assay) were fed to the SOM as example inputs.

Once trained, the SOM was then used to determine regions of anomaly within the Survey A dataset. For every soil sample, a numeric figure can be determined using the SOM which reflects how anomalous its nine assays are with respect to the entire pattern of assays throughout the survey area. Figure 2 shows the pattern of anomaly in the area surveyed, displayed as a contour plot based on the same coordinates as previously used in Figure 1. The contours represent increasingly anomalous regions within the survey site.

It can be seen that the neural anomaly tool automatically discovered six main areas of anomaly within the survey data. These six regions are marked 'A' - F on Fig 2.

Other less unusual areas (but still anomalous) were also discovered and could perhaps provide lower priority leads to be later investigated by a geologist.

As well as finding anomalous regions within a survey, the tool can also provide information on the factors leading to the high degree of anomaly. Information on the six main anomalous regions given in Fig. 2 is given in Table 2. Figures are given as percentages over the range of the assayed values of each mineral, i.e. 0% equals its lowest value and 100% its highest value.

Table 2 Region A B C D E F Cadmium 63% 19% 90% 0% 50% 95% Gold 13% 64% 71% 0% 3% 100% Copper 44% 27% 53% 30% 45% 73% Lead 45% 48% 22% 52% 30% 37% Molybdenum 22% 35% 20% 50% 25% 50% Nickel 4% 3% 5% 2% 90% 3% Silver 80% 11% 45% 77% OO/o 0% Zinc 95% 42% 69% 41% 50% 55% Arsenic 25% 97% 57% 12% 5% 36% Regions A, C and D located in accordance with the invention corresponded to known areas of silver anomaly.

Regions of anomaly can be compared with original survey maps. For example, region E has an anomalous proportion of nickel. Figure 3 shows the anomaly map, as given above, together with the nickel survey map. It can be seen how anomaly region E strongly correlates with the anomalous deposit of nickel.

The previous anomaly analysis was performed on all the nine geochemical assays from the survey A data. It can also be carried out on some subset of constituents, for example, just gold, nickel, silver and arsenic. In this case, a SOM was trained showing it only these four constituents.

After training, the SOM was analysed as previously and anomaly details extracted.

The results of this analysis is presented below.

Figure 4 shows a contour plot mapping the anomaly level over the survey site A for this particular selection of minerals.

Table 3 below shows anomaly descriptions for each of the regions labelled on the anomaly map (A-E). It can be seen that the anomalous nickel plug (region D) and a known silver anomaly (A and to a lesser extent C) are clearly indicated. Gold anomalies are indicated at B and E.

Table 3

Region A B C D E Gold 11% 64% 35% 10% 100% Nickel 4% 3% 5% 90% 3% Silver 90% 11% 45% OO/o OO/o Arsenic 25% 97% 45% 3% 36% Clustering is another analysis method which will be of great use to a geologist in analysing data from either a single survey or multiple surveys. It aids the geologist by extracting common types of geological deposits and their form. The tool automatically determines clustering across survey readings by using commercial neural computer techniques. It also supplies a numeric figure of importance for each cluster found.A geologist can prioritise further detailed analysis within survey data using these important figures.

Clustering is performed using a Self-Organising Map, as was used to determine anomalous regions. However, it is employed in a different way. As before, the SOM is trained by showing it survey data from various sampling points without explanation. Again, training is totally automated without need for user guidance.

After training, the SOM is then analysed to determine clustering structures as formed in its topologically organised two-dimensional output map.

An automatic procedure analyses the trained SOM and extracts from its internal weights the survey data descriptions most representative of each cluster. Each cluster can be viewed as describing typical types of deposit. A numeric figure is obtained which indicates the "importance" of each cluster. Further, the relationship between different clusters is determined.

The SOM used previously for anomaly detection in the survey site A was analysed for geological deposit clustering. The topological map of the SOM resulting from the cluster analysis is shown in Figure 5. It can be seen that six clusters were detected. Note that this two-dimensional map does not relate to the geographical map of the survey site A. The architecture of a SOM's output is arranged as a two-dimensional grid of neural processing units. Each cluster corresponds to a specific class of geological deposit.

The importance of a cluster is directly related to the area covered by that cluster on the two-dimensional topographic map. More important clusters cover a larger area. Cluster 4 is clearly the most important.

Clusters which are, to some degree, similar in terms of geological deposit share nearby regions on the topographic map. Clusters which are very dissimilar are assigned regions spaced far apart. In Figure 5, clusters I and 5 are similar, but 1 and 6 are very dissimilar.

Table 4 shows the typical soil geochemistry assays for each of the six clusters detected. Percentage values range from the minimum assayed value (0%) to the largest assay (100%).

Table 4

Mineral Cad- Gold Copper Lead Molyb- Nickel Silver Zinc Arsenic Cluster mium denum 1 27% 27% 43% 34% 26% 7% 3% 60% 31% 2 45% 18% 41% 56% 38% 5% 22% 70% 34% 3 55% 10% 46% 48% 43% 5% 28% 70% 31% 4 25% 39% 48% 46% 4% 3% 65Y 18% 5 21% 21% 47% 30% 26% 7% 3% 61% 21% 6 1% 1% 2% 2% 3% 0% 0% 3% 1% Neural analysis also provides a level of importance for each cluster. This will aid a geologist in visualising the relative merits of each cluster with respect to each other and their relevance in the survey. Important figures for the clusters are given in Table 5.

Table 5

Cluster | 1 | 2 3 4 5 6 Importance 0.10 0.02 0.04 0.72 0.08 ~ 0.04 This indicates that cluster '4' is the most important, followed by clusters '1' and '5' The geographical location of a specific type of deposit can be examined by choosing a cluster from the above table then displaying an overlaid survey map.

Figure 6 shows the geographical location of the six clusters in the survey site. The contours highlight.

Figure 7 shows an overlaid contour map of silver analysis with Cluster 3. It can be seen how this cluster typifies the deposition of silver within the survey area A.

Determining relationships between different components of survey results is another key objective for a geologist. Either all components or various groups of components, from possibly multiple surveys, can be chosen for correlation analysis.

Knowledge of the structure of these relationships (or correlations) will provide a geologist with leads for future analysis and exploration.

Not only is it possible to study the correlation between two particular components, but the interplay of correlations between multiple components can be considered and explored simultaneously.

For example, in the survey area A a geologist could consider how simultaneously increasing levels of copper and lead affect molybdenum. This allows for deeper understanding of survey data.

The neural computing technique used employed auto-associative neural models.

These are unsupervised learning systems which do not need prior information of example known correlations or given any external user guidance during training.

Instead, the auto-associative neural model merely takes survey datasets and automatically learns embedded relationships between the various survey components chose for correlation analysis.

Once the neural model has completed training, it is then analysed to determine any possible mixture of correlations. It can be thought that the neural computer becomes a complete model of the survey data.

After training, sensitivity analysis is performed to extract these learned relationships. It also allows complex exploration of the relationships between components of the survey datasets. Traditional correlation analysis allows only two components to be studied for their effect on each other. Neural sensitivity analysis allows the simultaneous study of inter-relationships between multiple components to be understood.

The effect of a single component, from the survey datasats, on all other components can be instantly analysed. Results identify whether it has a positive (supportive) or negative (inhibitive) influence on these other components and the strength of this influence. The detailed effect can be obtained, allowing non-linear relationships to be analysed and obtained.

The actual value of the relationship strength for one particular component should be considered independently from those for other components. Where many components influence the target component, strength values will all be fairly small.

However, in the case where the target component is only influenced by a few other components, strength values will be much larger. The actual numeric size of values is not particularly important. Instead, it is the relative value of strengths.

More detailed relationship analysis can be interactively explored. This process is best described comparing it to an audio graphic equaliser (see Fig 8). This visual object has sliders for every component of the survey dataset allowing the value of each to be altered by moving its slider up and down. Values are depicted by a bar chart. if one slider is moved, then all other bars which are correlated to it, by some degree, correspondingly move. By examining how bars move in respect to changes in various sliders, a detailed understanding of the relationships between multiple survey components can be obtained.

An auto-associative neural model was trained on the survey area A dataset. Once trained, it was analysed using sensitivity analysis as described above. Figure 9 shows the results of the sensitivity analysis for each of the nine soil chemical assays. Table 6 gives a matrix of numeric strengths for the nine soil chemical assays performed with respect to each other. The neural computer has determined a wide variety of correlations within the survey site. For example, zinc is strongly affected by both copper and lead. Copper is inhibited by molybdenum. Gold is supported by arsenic.

The correlations of the various assays with gold are shown in Figure 10. The high positive correlation with arsenic is clearly identified by the neural computer, with a strength of 0.40, Table 6

Cadmium Gold Copper Lead Moly- Nickel Silver Zinc Arsenic bdenum Cadmium 0 -0.01 -0.01 -0.04 0.14 0.19 0.17 0.08 0.04 Gold 0.07 - 0.17 -0.12 -0.04 -0.03 -0.05 -0.04 0.34 Copper -0.04 0.12 - 0.01 -0.18 0.20 -0.16 0.34 0.07 Lead -0.12 -0.16 0.11 , 0.19 -0.12 0.10 0.37. 0.06 Mo1y 0.19 -0.01 -0.18 0.23 - -0.08 -0.28 0.00 0.01 denum 0.15 -0.07 -0.10 0.06 -0.33 Nickel 0.25 -0.19 0.15 -0.18 -0.07 - -0.10 0.06 4.33 Silver 0.18 -0.08 -0.17 0.12 -0.30 -0.10 - 0.07 0.09 Zinc 0.07 -0.03 0.19 0.19 0.04 0.03 0.06 0 .05 Arsenic -0.08 0.40 0.02 0.09 0.04 -0 .25 0.08 -0.03 Summaries of the important correlations determined by the neural computer for each assay are given in Table 7.

Table 7

Mineral Positive Negative Cadmium Nickel, Molybdenum, Lead Silver Gold Arsenic Nickel, Lead Copper Zinc, Gold Nickel Mol Molybdenum,Silver Silver Lead Mol enum, Zinc Nickel, Gold Molybdenum Lead, Cadmium Silver Nickel Co er, Cadmium Arsenic Silver Cadmium Mo enum Zinc Copper er Arsenic Gold Nickel A similar neural correlation analysis was performed on the survey area B dataset.

This dataset contains many more components than area A.90 in total. This does not pose a problem for neural computing techniques. There is no reasonable limit to the size or complexity of survey datasets that can be analysed.

The area B survey was analysed for factors influencing the assayed copper level.

Figure 11 shows the results of the neural sensitivity analysis. It can be seen that two factors influence copper level. These are lithological descriptions having the colour label "Green" and mineral label "Quartz". Both of these factors strongly support high copper assays.

To gain an idea of the complexity of the task the neural computer has performed so easily, it should be considered that the dataset analysed consisted of 77 drill holes with 90 descriptions for each 2m interval. In total, 859 separate 90-component descriptions. This would be a hugely complex manual task for a geologist.

The relationships affecting any component of the survey can be determined by the neural computer. As an example, Figure 12 shows the factors influencing core samples which are coloured "White". In decreasing strength, the important positive factors are: Alteration (physical property), Quartz, Carbonate, Chrysocola, Calcite.

A geologist will often want to study the occurrence of specific combinations of minerals, rock types, physical properties, assayed levels etc. These interests will also arise from use of the previous neural analysis tools leading to interesting facts being determined about the survey data. These may prompt the geologist into performing some detailed searching through the survey datasets based on knowledge gained.

However, the geologist is unlikely to be able to specify exactly what is of interest.

For this reason, conventional searching techniques will not be particularly effective.

Instead, neural computers with their powerful pattern recognition capabilities are able to accept a "vague description" from the geologist and perform a neural fuzzy search.

The neural computing techniques employed to perform neural fUzzy searching are completely unsupervised and automatic. No user guidance is required once the search parameters are supplied by the geologist.

The geologist is able to search for regions specifying a range of requirements for individual survey components for example "Copper Assay" or Quartz: Approximate values (e.g. for assays) Existence (e.g. for specific lithological features) Upper bounds Lower bounds The result of the neural search is the degree of match between the vague search description supplied by the geologist and every location within the survey region.

This can be displayed using a contour plot showing the geographical location of matching regions within the survey site.

Neural searches were applied to datasets from both the survey sites A and B.

To illustrate simplistically neural fUzzy searching, the site A survey dataset was searched for regions having a assayed silver level of greater than 30%. (0% equates with the minimum assayed level of silver and 100% with the highest assayed level.) Figure 13 shows a contour map illustrating the distribution of high silver deposits within the geographical map of the site.

A more complex neural search is given in Figure 14. Here, it is assumed that in an imaginary gold extraction process, both copper and nickel act as contaminants and are hard to remove. The neural search performed looks for regions in which a sufficiently high level of gold exists without having high levels of contaminants. It can be seen that a few areas (1) were found which closely matched the search.

Also, a few areas (2) are indicated as being regions to avoid due to unwanted high levels of contaminants.

A searches was also performed on the area B drilling survey datasets. In the first, the known correlation of high levels of assayed copper and green coloured core samples was supplied as the neural search pattern. Fig 15 shows a plan view of the geographical location of matching regions. Again, contours indicate areas (1) with a high degree of match and area (2) with very poor matching. The plan view means that the actual degree of match represents the entire drill hole.

The neural search indicates that the bottom right corner of the region most closely matches the search pattern.

Claims

1. A method of locating a mineral deposit the method comprising the steps of: i obtaining a plurality of indicators of physical or chemical conditions at a plurality of sites in an area known to contain the mineral, at least one of the indicators being an indicator of the mineral; ii training neural anomaly identifying means to identify anomalies in the data sets; iii obtaining a plurality of the indicators at a plurality of sites in a search area thought to contain the mineral; iv inputting the indicators obtained from the search area to the trained anomaly identifying means to obtain an indication of the location of the mineral deposit; and v visualising the indication.

2. A method as claimed in claim 1 wherein the anomaly identifying means also identifies clusters.

3 A method as claimed in claim 1 fitrther comprising a cluster identifying means to identify clusters.

4. A method as claimed in any one of the preceding claims wherein an approximator is trained to identify correlations in the data sets.

5. A method as claimed in claim 3 wherein the cluster identifying means is a neural cluster identifying means.

6. A method as claimed in any one of the preceding claims wherein the neural anomaly detector comprises an SOM or an ART.

7. A method as claimed in any one of the preceding claims wherein the approximator means comprise neural approximating means.

8. A method as claimed in claim 7 wherein the neural approximating means comprises an MLP or an RBF.

9. A method as claimed in any one of the preceding claims wherein a data set is interpolated using an inverse weighting algorithm.

10. A method as claimed in any one of the preceding claims wherein the data set comprises electromagnetic (EM) survey data.

11. A method as claimed in claim 10 wherein the EM data is processed prior to input into the neural anomaly detector into at least one group of components selected from the group comprising: i simple sum OfEM1ms to EM4ms; ii simple sum of EM5ms to EM8ms; iii simple sum of EM9ms to EM12ms; iv EMlms EM2002t v 2nts EM200

vii Sum of signals over the period 13 to 20 ms differing from the mean by more than 1.5 standard deviations; wherein EM,, is the signal return at xms.

12. A method as claimed in any one of the preceding claims wherein the data set is selected from the group comprising: a) radioactive decay; b) gravity c) chemical analysis d) rock colour e) rock physical properties f) rock type

13. A method of winning a mineral from a deposit comprising the steps of: locating a deposit by a method as claimed in any one of the preceding claims; and ii recovering the mineral.

12. A method as claimed in any one of the preceding claims wherein the data set is selected from the group comprising: a) radioactive decay; b) gravity c) chemical analysis d) rock colour e) rock physical properties f) rock type 13. A method of winning a mineral from a deposit comprising the steps of: locating a deposit by a method as claimed in any one of the preceding claims; and ii recovering the mineral.

Amendments to the claims have been filed as follows Claims 1.A method of locating a mineral deposit the method comprising the steps of: i obtaining a plurality of data sets of indicators of physical or chemical conditions at a plurality of sites in an area known to contain the mineral, at least one of the indicators being an indicator of the mineral; ii training neural anomaly identifying rneans to identify anomalies in the data sets; iii obtaining a plurality of the indicators at a plurality of sites in a search area thought to contain the mineral; iv inputting the indicators obtained from the search area to the trained anomaly identifying means to obtain an indication of the location of the mineral deposit; and v visualising the indication.

2. A method as claimed in claim 1 wherein the anomaly identifying means also ideiitifies clusters.

3 A method as claimed in claim 1 flirtlier comprisinga cluster identiing means to identi clusters.

5. A method as claimed in claim 3 wherein the cluster identifSring means is a neural cluster identifying means.

11. A method as claimed in claim 10 wherein the EM data is processed prior to input into the neural anomaly detector into at least one group of components selected from the group comprising: simple sum ofEMims to EM4ms; ii simple sum of EMsms to EM8ms; iii simple sum of EM9m5 to EM 1 2ms; iv EMlms EM2002ms v EM2001ms

vii Sum of signals over the period 13 to 20 ms differing from the mean by more than 1.5 standard deviations; wherein EMxms is the signal return at xms.