Applications of Multivariate Statistics and Pattern Recognition to Geophysical Logging at Noranda

Abstract

In mining, waste rock dilution occurs when uneconomic rock is mined and processed with economic mineralized ore. A survey of selected stopes from various Noranda Group mines estimated that aggregate dilution and oreloss stood at 20% and 8% respectively in 1994. Various technologies are being investigated by Noranda Group mines in an ongoing effort to minimize dilution and oreloss. The following three potential applications of geophysical logging have been identified for dilution and oreloss reduction through improved orebody delineation.

Increasing Orebody Sampling at the Delineation Stage to Reduce Geological Dilution

Diamond core drilling, with geological core-logging, is the most commonly used method for orebody delineation. Unfortunately, the many task-intensive steps and expense of this method constrain the amount of samples one can obtain. However, by allotting an optimum percentage of the delineation drilling budget for geophysical logging of percussion-drilled holes, increased sampling would occur. Since percussion drilling is significantly cheaper than diamond drilling, more delineation holes could be drilled to better define the ore boundary geometry and, ultimately reduce geological dilution and oreloss.

Controlling Geological Dilution at the Production Stage

In production, thousands of meters of percussion holes are drilled annually for blasting purposes. However, since no core is recovered in this drilling process, information about the ore-waste contact is seldom acquired. Thus, there is the potential to geophysically log production holes to identify the ore-waste contact for optimal blast design with a resulting reduction in dilution and oreloss.

Interpreting Crosshole Geophysics Tomograms

By geophysically logging boreholes that were used in a crosshole geophysics survey, relationships can be developed to correlate logged geology and geophysical properties along the boreholes. These relationships could then be used in the interpretation of crosshole tomograms.

Using commercially-available geophysical logging equipment, combined with multivariate statistics and pattern-recognition techniques, we have analyzed data acquired at selected Noranda Group minesites to provide answers to the following questions:

1. can lithology be indirectly determined by geophysical logging?

2. to what accuracy can geophysical logging indirectly measure grade?

3. can geophysical logging aid in the interpretation of crosshole geophysics tomograms?

Geophysical Logging for Lithology

Based on our test data, interpretation of geophysical logs for lithology was very much site-dependent. In simple cases a single log or a cross-plot sufficed, while several logs and pattern recognition techniques were necessary in others.

The lithology of the Louvicourt deposit was found to be complicated from a geophysical point of view, and pattern recognition techniques were required to interpret the data. Two models, one using neural networks and one using discriminant analysis were used. Ore boundaries could be successfully mapped, however, the models could not distinguish between Zn- and Cu sulphide. Zones of pyrite were, in some cases, mis-interpreted as Zn- sulphides, especially by the discriminant model. The neural-network model did manage to locate selected zones of altered rock. Most waste rocks were interpreted as tuff by both models.

At Heath Steele Mine, it was found that gamma-gamma, natural gamma, gamma ratio, and magnetic susceptibility data could be successfully used to isolate waste (i.e. sedimentary & volcanic rocks and, to some extent pyrite/pyrrhotite) and ore (i.e. massive and semi-massive Pb-Zn) units. At Isle Dieu Mine, crossplots of gamma-gamma versus natural gamma clearly identified four main groupings of data, corresponding to the four main lithological units.

Geophysical Logging for Grade

Using multiple linear regression techniques, it was found that three of the metal assays at Louvicourt could be estimated -- in order of certainty: Fe, Cu, and Zn. Although these techniques tended to underestimate very high Zn grades, neural-network techniques were found to improve Zn-grade estimates (i.e. R2 increased from 0.45 to 0.70). Similarly, the Cu-equivalent grade could be estimated by applying neural-network techniques. Data from three of test holes showed that the correlation between the neural-network model and the Cu-equivalent assays was quite high with R2 = 0.77.

At Heath Steele, there was a good correlation between Pb-grade, Zn-grade, and the geophysical logs within the Pb-Zn sulphides. However, in boreholes containing many kinds of sulphides, these relationships were much more difficult to define. Cu grade could not be estimated with the available data using linear regression. In fact, the Cu grade was difficult to estimate using other methods as well, since the conductivity and susceptibility logs behaved peculiarly in three of the boreholes. However, reasonable estimates of Pb grade were obtained using neural networks. At Isle Dieu, simple linear regression of various geophysical logs could be used to estimate four different metal grades (i.e. Zn, Ag, Pb, & Cu). Estimates of the Zn, Ag, and Pb grades seemed to fit the assays fairly well.

Geophysical Logging for Interpretation of Crosshole Tomograms

Sonic logging results from the Louvicourt deposit showed that, of all of the minerals and metals assayed at Louvicourt, seismic velocity is most sensitive to iron content, followed by pyrite content. Therefore, as zones of higher seismic velocities may indicate zones of higher iron and pyrite content, this result was used in the interpretation of seismic tomography data acquired at the Louvicourt deposit.

1.0 Introduction

1.1 Dilution & Oreloss

In mining, waste rock dilution occurs when waste rock is mined and processed with together with ore. Puhakka (1991) has noted that waste rock dilution and oreloss exist in all phases of exploitation of an orebody, including geological modeling and evaluation, decisions regarding cut-off grade, design of the mining method, and mining and ore concentrating (see Figure 1).

Geological and Geostatistical Modeling:

In the exploration and underground delineation phases, only an estimated model of the orebody can be made. "Under-sampling" of the orebody requires some assumptions about its shape and size. This is the first source of geological dilution (see Figure 2).

Cut-off Grade Decision:

Decisions regarding cut-off grades are based on different numerical methods and estimates which can be affected by errors. This is a second source of geological dilution.

Mining Method Design:

In the phase of mining method design, some ore is often left in place in order to create protective pillars. These pillars may be lost if they cannot be extracted later in the mining process. In the case of narrow orebodies, it is often necessary to widen mining blocks and include waste rock in stope designs. Planned dilution occurs when the orebody is so complex that stopes are designed to include waste rock outside the ore contacts in the hangingwall and/or footwall. Planned dilution may also occur when zones of waste rock exist within the orebody itself, and it is impossible to design a stope without including these zones of waste rock.

Mining Inaccuracies:

Due to a lack of flexibility and precision of drilling equipment and blasting procedures, the ore contact cannot be followed in detail. Ore may be lost, and material below the cut-off grade may be drilled, blasted, loaded, and transported to the concentrator. This dilution is commonly referred to as unplanned dilution. Furthermore, cavings, seismic activity, and other geomechanical problems may cause unplanned dilution.

Ore Recovery in the Concentrator:

The feed to the concentrator should be as homogeneous as possible. However, all phases listed above affect grade variation of the feed. Furthermore, waste rock may disturb the process in the plant and lower the recovery.

Figure 1. The "Mining Cycle" and corresponding sources of dilution and oreloss.

Figure 2. Schematic Diagram (sectional view) for Geological Dilution and Oreloss. Insufficient sampling causes waste rock to be included in the geological model, as well as ore to not be included in the model. These problems decrease with decreasing borehole spacings (i.e. increased sampling).

1.2 Dilution/Oreloss Control and Geophysical Logging at Noranda

During 1994, selected stopes from various mines operating within the Noranda Group of companies were surveyed using the Cavity Monitoring System developed by the Noranda Technology Centre (NTC). Aggregate dilution and oreloss within the Noranda Group was estimated to be 20% and 8% respectively during 1994. As the control of dilution and oreloss represents large potential savings and increased revenue for the Noranda Group, various technologies are being investigated by Noranda Group mines in an ongoing effort to minimize dilution and oreloss. The following three potential applications of geophysical logging have been identified for dilution and oreloss reduction through improved orebody delineation.

Increasing Orebody Sampling at the Underground Delineation Stage to Reduce Geological Dilution:

To effectively and efficiently delineate an orebody, one would like to have as many samples of an orebody as possible. Diamond core drilling, with geological core-logging, is the most commonly used method in the sampling and delineation of ore deposits. Unfortunately, the many task-intensive steps and expense of this method constrain the amount of samples one can ultimately obtain.

However, by allotting an optimum percentage of the entire delineation drilling budget for percussion-drilled holes surveyed by geophysical logging, one could increase the sampling. Since percussion drilling is significantly cheaper than diamond drilling, more delineation holes could be drilled to better define the ore boundary geometry and, ultimately, reduce geological dilution and oreloss. Furthermore, increased logging productivity could be achieved since the logging speed is faster, assaying is not required (once calibrated), and since measurements are made in situ and core does not need to be recovered, related drilling decisions could be made "on the spot". Unfortunately, some of the advantages of geophysical logging in place of coring and assaying are mitigated by uncertainty in the interpretation of lithology and grade from geophysical data.

Using actual drilling and logging costs, quoted by the staff from five different Noranda operations, a cost-benefit analysis was performed to determine the economic benefits of geophysical logging for delineation. Results showed that the added costs of single-hole geophysical logging equipment would be covered by the savings from lower drilling costs and higher logging productivity if 10 to 20 % of the drilling budget was allotted for geophysical logging of percussion-drilled ore-delineation holes. By allotting as little as 30% of the entire underground drilling budget for drilling and geophysical logging of percussion-drilled delineation holes, the total drilling footage would increase by 45 to 90%, and the distance between ore-intersections would decrease by 17 to 28 %. This corresponds to an increase in the number of intersections with an orebody by a factor of 1.5 to 1.9.

Controlling Geological Dilution at the Production Stage:

In production, thousands of meters of percussion holes are drilled annually for blasting purposes. However, since no core is recovered in this drilling process, information about the ore-waste contact is seldom acquired. Although some sites have tried drill-cutting analysis, this technique is task-intensive, prone to high sampling inaccuracies, and the turn-around time for the data-analysis is often too slow for practical use.

To alleviate this problem, there is the potential to use geophysical logging to survey production holes. In so doing, the ore-waste contact could be rapidly identified, and the loading of explosives could be optimized to reduce dilution. Furthermore, in those cases where grade quantification is possible from the geophysical logging measurements, there are also possible downstream benefits in quantifying the mill feed metal content.

Interpreting Crosshole Geophysics Tomograms:

By geophysically logging boreholes that were used in a crosshole geophysics survey, relationships can developed to correlate logged geology and geophysical properties along the boreholes. These relationships could then be used in the interpretation of crosshole tomograms, which are maps of geophysical properties between boreholes. The objectives are detection of undiscovered lenses and defining "connectivity" by mapping ore contours between delineation drillholes.

1.3 Goals and Approach

Our goals are to provide answers to the following questions:

1. Can lithology be indirectly determined by geophysical logging at selected Noranda Group minesites, and, can geophysical logging differentiate ore sulphides from waste sulphides?
2. To what accuracy can geophysical logging estimate grade at selected Noranda Group minesites?
3. Can geophysical logging be used to aid in the interpretation of crosshole geophysics tomograms from selected Noranda Group minesites?

Our approach is to use commercially available geophysical logging equipment, combined with multivariate statistics and/or pattern recognition techniques to provide answers to the questions listed above.

2.0 Analysis Techniques

Multivariate statistical techniques, which have been successfully used for many years in oil and gas exploration to relate geophysical properties to stratigraphy, have been analogously applied to mineral exploration and mining problems in recent years (Urbancic , 1985; Mwenifumbo & Killeen, 1987; Wänstedt & Sundin, 1991; Wänstedt, 1992; Mwenifumbo, 1993; Mwenifumbo et al., 1993; Emilsson & Wänstedt, 1993; Wänstedt, 1993; Pflug et al., 1994). Cinq-Mars et al. (1992) describe how univariate, bivariate, and multivariate statistics -- including distribution analysis, histograms, scatter diagrams, multiple regression, and principal component analysis -- can be used in the interpretation of geophysical logs.

2.1 Discriminant Analysis

With discriminant analysis one attempts to find the linear function of independent variables, that produces the maximum differentiation between groups of different samples, so that the two groups are statistically distinct. This is achieved by forming a linear function of discriminating variables that takes the form:

where: Discriminant scoreDiscriminant weightsIndependent variables

The scores represent the position of each individual along a line in the multidimensional space defined by the independent variables (Figure 3). The discriminant weights , which define the direction of the line, are chosen to achieve a correlation that is as strong as possible between the computed and the desired classification. If a linear function successfully separates two different classes, e.g. rock types, it can then be used for classifying data from other boreholes with similar lithology.

Figure 3. Graphic illustration of two-group discriminant analysis (Hair et al., 1992).

In order to classify new individuals into groups, it is necessary to construct a discriminant function based on known data. In this case all the geophysical variables are included in the discriminant function. If the number of variables is large, a stepwise variable selection can be used to delete variables that do not contribute significantly to the result. Wilks' lambda is a measure of how well the independent and the dependent variables are correlated, and is in the univariate case equal to . A more detailed description of discriminant analysis can be found in Hair et al. (1992) and Seber (1984).

2.2 Neural Networks

A neural network is a self-adjusting matrix mechanism which has the ability to learn by pattern matching. The structural basis for artificial intelligence neural networks is the human brain. General applications of neural networks have been pattern detection, signal filtering, data segmentation, data compression & sensor data fusion, adaptive control, optimization of scheduling & routing, complex mapping, and adaptive interfaces for man/machine systems. An application specific to the interpretation of EM geophysics data is presented by Cisar et al. (1993). Neural networks can be seen as intermediate between statistical and structural pattern recognition methods, though they resemble the former more than the latter. The ability of learning in neural networks provides an interesting alternative to the Bayesian classifier (Bischof et al., 1992). It is especially interesting that no assumptions about the probabilistic model need to be made.

The back propagation and forward propagation operations of neural networks are analogous to the psychological Test-Operate-Test-Exit (TOTE) model of the brain's functional use of feed-back and feed-forward. Neural networks replicate these brain processes by information layering. In this paper, back propagation networks are used. Back propagation neural networks, generally have at least three layers: an input layer, a hidden layer, and an output layer. In each layer, there are a series of nodes that define the transfer of information.

Figure 4. Structure of a three-layer neural network.

A unit or processing element in a neural network can be considered as a simple processor with many different input connections from other units in the network, and one unit output which is sent to many other units (see Figure 4). Every connection in the network has a numerical value attached to it which is called weight, wij which denotes the weight of the connection from i to j). A unit i computes a net input (neti) from the outputs oj of the other units and from the weights wij of the connections (usually by a weighted sum). In most models a numerical value, called a bias (bi), is added to the net input. An activation function f is applied to this value yielding the output oi of the unit. This can be stated as:

The activation function, usually a sigmoid type function, is used to constrain the output value from a node to a range of 0 and 1. Therefore, it is essential to normalize the target signals in the same range before training. There are no restrictions on how to normalize the data. The test data set should, however, be normalized in the same manner as the training data set.

Back propagation is designed to reduce an error between the actual and the desired output of a network in a gradient manner. The usual error measure, the summed square error (SSE), is defined as:

where p ranges over all vectors in the training set and k denotes the output unit. When the input vector p is applied to the network, opk indicates the output of unit k, and tpk is the corresponding target output.

To determine the neural network efficiency, a known data set is run through the network. The back propagation step calculates the error vector by comparing the actual and target outputs. New sets of weights are iteratively calculated based on these error values until a minimum overall error, or global error, is obtained.

After the training stage, the weights on all connections are fixed and the network is passed through a testing stage. This is done by using data not included in the training set. Performance of the trained network can be evaluated by some simple statistical functions such as recognition ratio, defined as the percentage of the total number of correctly classified outcomes over the number of samples.

Typical neural network methodology involves:

1. choosing a set of historical data with well-defined input parameters (e.g. geophysical logging results) and desired output results (e.g. lithology, grade);

2. training the neural network using two thirds of the data-set (e.g. the geophysical and geological data);

3. test the prediction accuracy on the remaining data;

3.0 Geophysical Logging for Lithology & Grade

3.1 Introduction

In late 1993, Outokumpu's OMS-logg geophysical logging system was tested at three Noranda Group mines (Louvicourt, Isle Dieu, Heath Steele). The system consists of six main components:

· four different probes (apparent resistivity/conductivity by EM induction, natural gamma, gamma-gamma density, magnetic susceptibility);

· a depth transducer;

· a fiberglass-reinforced, 5-conductor, 150-m cable contained within a portable winch designed specifically for in-mine ore-delineation applications (i.e. logging of both up- and down-holes);

· drill hole interface (KTP-DHI) with rechargeable batteries;

· hand-held computer/data-logger (KTP-84);

· data-processing software (PORA) for a DOS-based PC.

Four holes were surveyed at Louvicourt, two holes at Isle Dieu, and four at Heath Steele measuring all four geophysical properties in each hole. Although the OMS-logg system is designed for use in up-holes (as well as down-holes), it was found that 50 m was the limit for upholes at +25°, and 70 m at +10°.

3.2 Quantitative Results

3.2.1 Louvicourt

Lithology:

On the basis of the geological mapping of the core from the four holes logged at Louvicourt, 12 to 13 different lithological zones can be selected, depending on how the geological core is interpreted. Waste rock types include coarse ash tuff, fine ash tuff, cherty tuff, fine and coarse ash tuff, intermediate dykes and rhyodacite. We have assumed that the various tuffs have similar compositions, and thus, all tuffs are grouped as one rock class. The various rock types are affected, to some degree, by alteration. Since the chemical alterations are not quantified (i.e. by chemical analyses), it is difficult to differ between the degree of alteration. Thus, we have grouped alterations as one rock class. The most important mineralized zones, as mapped by site geologists, are the pyrite disseminations (waste), massive Zn-sulphides and massive Cu-sulphides. These have been grouped as separate rock classes. To enable further, more detailed, classification of the sulphides, we have also included semi-massive sulphides as a separate rock class. The resulting geological model, consisting of nine main rock classes, is summarized in Table 1. The complexity of the model is quite obvious with four different geophysical parameters to describe nine different rock classes.

Rock Class	Lithological Unit
1	tuff
2	pyrite disseminations
3	fracture zones / faults
4	massive Zn ore
5	massive Cu ore
6	intermediate dykes
7	rhyodacite
8	alterations
9	semi-massive sulphide ore

Figure 5 is a plot of natural gamma versus gamma-gamma (i.e. a crossplot) for all the data from Louvicourt. The massive Zn and Cu sulphides cluster in the bottom left corner of the figure, corresponding to gamma-gamma and natural gamma values of less than 8000 CPS and 20 CPS respectively. Although not immediately apparent in the figure, the massive Cu sulphides can be further differentiated from the massive Zn sulphides. Pyrite disseminations, intermediate dykes, rhyodacite, and semi-massive sulphide ore cluster in the middle of Figure 5 (8000<GG<16000, 10<NG<40). Tuff, fracture zones and faults, and alterations appear at the top of Figure 5 (GG>16000, 10<NG<55).

Thus, by using the crossplot of gamma-gamma and natural gamma, it is relatively simple to locate zones of massive Zn and Cu ore sulphides. Although waste pyrite disseminations cluster in a region different to that of the massive sulphides, they cannot be differentiated from semi-massive sulphides. Similarly, the other lithological units (waste) can be differentiated from the massive ore sulphides, however, they cannot be uniquely identified among themselves.

An attempt has been made to utilize pattern recognition techniques to improve the ability to identify lithological units from the logging of the four different geophysical properties. The training data was taken from two of the logged boreholes (i.e. 565-023 and 565-025). Each sample consisted of an average of between 5 to 50 data points. Training of a neural network can take a relatively long time (several hours), however, once the network is trained, it takes a couple of seconds to obtain the results from a borehole. During training, some rock types were omitted since there seemed to be no possibility for the network to "learn" to recognize them. The massive Cu sulphides were not included in the training, since there were too few samples. Likewise, the rock type "fracture zones / faults" was excluded from the training since the numerical models were incapable of correctly classifying this "rock type". None of the geophysical methods used here are really affected by fractures, unless the fractures are clustered such as in a significant fault.

The resulting model was tested on the logdata from all the holes (see Figure 6 through 9). In the figures there are 4 "core logs". The first show the location of the rock types according to the core, the second and third according to the discriminant and neural models and finally the fourth shows to what extent the models manage to classify the rocks (see legend Figure 6).

In Figure 6, three kinds of lithology are plotted along with the geophysical logs for Borehole 565-023. The first is the geological core-log, the second is a log calculated using discriminant analysis, and finally, the third is a log calculated using neural networks. Both models show the most important feature, that is where the sulphide zones begin -- however, there are some important differences. Firstly, in hole 565-023 there is a zone with altered rock and pyrite disseminations showing a clear difference between the two numerical models. The neural model (NM) correctly classifies the bleached zone, while the discriminant model (DM) shows nothing. The zone with pyrite disseminations is classified as pyrite, ore, and alteration by the NM while the DM characterizes the whole zone as ore (class 4). Notice the low susceptibility value in the lower part of the zone. This may be caused by oxidation of the iron within the rock.

Secondly, the following zone of tuff is misclassified as pyrite dissemination by the NM. There is, however, a difference between this "tuff" and the one before the pyrite zone. When looking at the gamma-gamma log there is a slight but clear change in the log after the pyrite zone.

At the contact between waste and the massive sulphide zones, there is a short zone of altered rock. This zone is correctly classified by the NM, but not by the DM. The predicted sulphides start at almost exactly the same place for both models and the core. A fault just after 50 m is detected by the numerical models, but none of them classifies this zone correctly since this rock type is not included in the training file.

Therefore, as can be seen in Figures 6 through 9, not all units were classified correctly. The massive Zn sulphides were correctly classified in general, except were the model decided that the grade was too low. Pyrite disseminations were however incorrectly classified as massive sulphide ore in some occasions. The NM detects alterations in most cases although the lateral extent does not always coincide with core. A reason for the discrepancy, not counting the lack of quantification, is that altered zones near the ore contain varying amounts of pyrite, e.g. in 565-026. Rhyodacite is misclassified by both models . The dykes in 565-025 are not correctly classified but they are present in both models as "tuff". Furthermore, the CU-ore in 565-023 is classified as semi-massive sulphides by the NM.

Grade:

Correlation coefficients of the metal assays and calculated means of the geophysical measurements for the corresponding assay intervals are presented in Table 2. Another way to quantify correlation is by using the squared correlation coefficient (R2). Conductivity and Cu are well-connected (R2 = 0.83). Also, Zn and gamma-gamma show some correlation, but not nearly as high (R2 = -0.56). Fe and gamma-gamma, on the other hand correlate quite well (R2 = -0.86). Sums of metal grades such as Zn+Cu and Cu+Zn+Fe correlate with the geophysical logs to a high degree (R2 = -0.62 and -0.89 respectively, with gamma-gamma).

	cu	zn	cu+zn	ag	au	fe	fe+cu+zn
sg	0.149	0.566	0.567	0.352	0.352	0.969	0.946
sus	0.121	-0.076	0.034	0.147	0.106	0.268	0.211
cond	0.835	-0.215	0.485	0.654	0.340	0.232	0.373
ng	-0.262	-0.494	-0.598	-0.386	-0.354	-0.796	-0.832
gg	-0.223	-0.556	-0.617	-0.390	-0.380	-0.858	-0.886
gr	-0.292	-0.394	-0.543	-0.356	-0.298	-0.666	-0.715

From these relationships only two grades can be estimated with fair accuracy: Cu and Fe. Multiple linear regression analysis indicates that Cu estimated from conductivity has a R2 of 0.7 and a standard error of 2.5%. Using multiple linear regression with all geophysical logs as input to estimate Zn grade, the correlation is clearly lower (R2 = 0.45) than for Cu, resulting in a quite high standard error of 3.4%. The result of a regression with Fe and the geophysical logs gives a R2 of 0.7, and a standard error of 5.3%.

To improve Zn-grade estimates, Figure 10 shows Zn assays together with a Zn-grade calculated using neural networks. Even though this probably is not the optimum network there is still a clear resemblance between the two curves. The Zn-grade is calculated without any assumptions, such as the location of the ore boundary, etc. The correlation between assays and neural-network calculated values is 0.7. Notice that the neural model shows Zn only in the areas were assays were performed, hence, there seems to be little risk of misinterpreting waste rock for Zn. Moreover, Zn-grades are under-estimated in average.

An important grade calculation at Louvicourt, is "copper-equivalent" grade, which is equal to Cu grade plus one half of the Zn grade (i.e. Cu + 0.5 Zn). Neural networks have been used to estimate copper-equivalent grades from geophysical logs, resulting in a high correlation of R2 = 0.77 (see Figures 11 through 15).

In all of the holes the models fit the assayed grades fairly well. Notice that in hole 655-023 the model suggests ore around 20 m. There is no assay here so this is an assumed error, or misclassification. This is the same zone that is misclassified by the rock type models presented earlier.

3.2.2 Heath Steele

Lithology:

The lithological units mapped by site geologists during core-logging of holes C-232, C-233, and C-234 were grouped into six main rock classes (see Table 3). Hole C-207 was not included in the analysis of these three AQ-diameter holes since its diameter is larger (i.e. BQ), and as a result, its geophysical response is different.

Rock Class	Lithological Unit
1	massive Pb-Zn sulphides
2	semi-massive sulphides
3	tuff sulphides
4	massive pyrite
5	chlorite tuff
6	altered tuff

Figure 16 is a plot of natural gamma versus gamma-gamma for the data from the three holes:

Figure 16. Crossplot of natural gamma and gamma-gamma for holes C-232, C-233, and C-234 at Heath Steele. (For the key, see Table 3).

The sulphides have gamma-gamma and natural gamma count-rates less than 6000 CPS and 40 CPS respectively. Count-rates greater than these values correspond to tuffs (i.e. waste). The sulphides can be further segmented based on gamma ratio (i.e. natural gamma divided by gamma-gamma) and magnetic susceptibility (see Figures 17, 18, and 19).

If the gamma ratio of the isolated sulphides is greater than 0.0065, the zone corresponds to massive Pb-Zn sulphides (i.e. ore). For a ratio less than 0.0065, magnetic susceptibility values between -2000 and 10000 correspond to semi-massive ore, and values outside these limits correspond to zones of pyrite/pyrrhotite. The accuracy of the delineation relationships are presented in Table 4 and in Figures 20 to 22.

Figure 18. Gamma-gamma versus Magnetic Susceptibility for holes C-232, C-233, and C-234 at Heath Steele. (For the key, see Table 3).

Figure 17. Gamma Ratio versus rock-type for holes C-232, C-233, and C-234 at Heath Steele. (For the key, see Table 3).

Heath Steele
gamma-gamma < 6000 CPS & natural gamma < 40 CPS	gamma-gamma < 6000 CPS & natural gamma < 40 CPS	gamma-gamma < 6000 CPS & natural gamma < 40 CPS	gamma-gamma > 6000 CPS & natural gamma > 40 CPS
gamma ratio > 0.0065	gamma ratio < 0.0065	gamma ratio < 0.0065
	-2000 < magnetic susceptibility < 10000	-2000 > magnetic susceptibility > 10000
massive Pb-Zn sulphide (ore)	semi-massive sulphides (ore)	pyrite/pyrrhotite (waste)	tuff (waste)

Table 4. Delineation Accuracy: Ore /Waste for Heath Steele.	C-232	C-233	C-234
Ore/Waste contact (entire hole)	89 %	89 %	88 %
Ore/Waste contact (sulphide zone)	69 %	75 %	65 %

Grade:

A correlation analysis of assay intervals with averages for the geophysical measurements over the corresponding intervals for holes C-232, C-233, and C-234 is summarized below:

Metal Grade	Geophysical Property	Correlation Coefficient (R)
Pb	gamma-gamma	-0.438
Zn	gamma-gamma	-0.489
Pb+Zn	gamma-gamma	-0.477
Cu	inductive conductivity	0.484
Ag	gamma-gamma	-0.546

Figure 24. Estimated and assayed Ag grades for hole C-232 using sulphide-zone isolation and regression techniques.

Figure 23. Estimated and assayed Zn grades for hole C-232 using sulphide-zone isolation and regression techniques.

The correlation coefficients indicate grade estimation through multiple linear regression analysis is difficult for Heath Steele. In an attempt to improve grade estimation, sulphide zones were first isolated using the relationships presented in Figure 19, and then regression analyses were used to estimate grade for selected holes (see Figures 23 through 26).

Figure 25. Estimated and assayed Pb grades for hole C-234 using sulphide-zone isolation and regression techniques.

Figure 26. Estimated and assayed Zn grades for hole C-234 using sulphide-zone isolation and regression techniques.

In an attempt to improve the ability to estimate Pb grade from the four geophysical logs, a Pb neural network was developed for Heath Steele, and shows promise (see Figures 27 and 28).

Figure 27. Pb-assays and grade calculated using neural networks for hole C-233 at Heath Steele. Bold line denotes assays, sectioned line represents averaged estimates.

Figure 28. Pb-assays and grade calculated using neural networks for hole C-234 at Heath Steele.

3.2.3 Isle Dieu

Lithology:

Lithological units, as mapped by Isle Dieu geologists, were grouped into four main rock classes. From Figure 29 -- a plot of natural gamma versus gamma-gamma for all of the data -- simple relationships can be developed to map lithological units (see Table 6).

Rock Class	Lithological Unit	Properties
1	sulphides	GG < 8000; NG <30
2	basalt	8000 < GG < 14000; NG <30
3	intermediate dykes	8000 < GG < 14000; NG > 30
3	felsic dykes	8000 < GG < 14000; NG > 30
4	faults	GG > 14000

Figure 29. Natural gamma versus gamma-gamma for Isle Dieu. (See Table 6 for key.)

Using these simple relationships, Figure 30 is comparison of the geological logging with the geophysical interpretation for hole ID-797. Note the accuracy of the interpretation.

Figure 30. Geological logging and geophysical interpretation of Hole ID-797.

Grade:

A correlation analysis of assay intervals with averages for the geophysical measurements over the corresponding intervals for holes ID-797 and ID-799 is summarized below:

Metal Grade	Geophysical Property	Correlation Coefficient (R)
Zn	natural gamma	0.67
Cu	inductive conductivity	0.93
Ag	inductive conductivity	0.48
Pb	gamma-gamma	0.56

Using linear regression techniques, it is possible to estimate the grades of Zn and Ag. Zn grade can be estimated with an error of 9.4 % due to a squared correlation coefficient of 0.55 between natural gamma and Zn grade. Similarly, the squared correlation coefficient between conductivity and Ag is 0.87. Figures 31 and 32 show the accuracy of estimating Pb and Cu grades using gamma-gamma and inductive conductivity data respectively. Veins with high grades within the assayed intervals, clearly seen in the figures as anomalies in the estimate, increase the average grades over the zone. The relatively long intervals possibly have a negative effect on the correlations between grades and geophysical logs. Geophysical logging gives a higher lateral resolution than normal assaying procedures, in this case.

Figure 31. Estimated and assayed Pb grades for hole ID-797 using regression techniques.

Figure 32. Estimated and assayed Cu grades for hole ID-797 using regression techniques.

4.0 Geophysical Logging for Interpretation of Crosshole Tomograms

4.1 Introduction

During the Fall of 1994, the Borehole Geophysics Section of the Geological Survey of Canada (GSC) was contracted to acquire geophysical logging data in three vertical holes at Louvicourt Mine (holes 655-304, 655-302, and 655-284). Crosshole seismic data was acquired by the Noranda Technology Centre between holes 655-304 and 655-302, as well as holes 655-302 and 655-284. Figure 33 is a long section showing the location of the holes, as well as an interpretation of the geology based on geological core data.

Figure 33 Long section from Louvicourt Mine showing the location of the holes, as well as an interpretation of the geology based on geological logging.

4.2 Quantitative Results -- Sonic Logging

Seismic velocities (Vp) for the three logged holes were entered into a database such that average seismic velocities could be extracted and compared to mineral and metal assays over corresponding assay intervals. Table 8 shows correlation coefficients between logged seismic velocity and assayed copper, zinc, silver, gold, chalcopyrite, sphalerite, pyrite, and iron grades. Included are correlation coefficients for density ( = specific gravity) and acoustic impedence (the product of seismic velocity and density, i.e. Z = Vp x ).

	velocity	Cu%	Zn%	Ag(g/t)	Au(g/t)	Cpy%	Sph%	Py%	Fe%	SG(g/cc)	Z
velocity	1.0000
Cu%	-0.0372	1.0000
Zn%	-0.0653	-0.1475	1.0000
Ag(g/t)	0.0498	0.5946	0.0645	1.0000
Au(g/t)	0.0920	0.3446	0.1279	0.7351	1.0000
Cpy%	-0.0639	0.9058	-0.1650	0.3211	0.1660	1.0000
Sph%	-0.0490	-0.1238	0.8211	-0.0044	0.0119	-0.1279	1.0000
Py%	0.5455	0.0296	0.2342	0.2777	0.3647	-0.0403	0.0566	1.0000
Fe%	0.6362	0.2575	0.1901	0.3893	0.3658	0.1339	0.0908	0.8827	1.0000
SG(g/cc)	0.6150	0.2353	0.2579	0.3705	0.3536	0.1207	0.1339	0.9048	0.9851	1.0000
Z	0.7827	0.1661	0.1754	0.2993	0.2993	0.0697	0.0862	0.8785	0.9650	0.9714	1.0000

Table 8 shows that correlation of logged seismic velocity with mineral and metal grades is highest for percent-iron (Fe%), followed by percent-pyrite (Py%). This shows that there is a relatively strong relationship between logged seismic velocity and iron and pyrite content at Louvicourt (see Figure 33).

Figure 33. Plot of logged seismic velocity (Vp) versus assayed percent-iron for holes 655-284, -302, and -304 at Louvicourt.

A regression analysis indicates that the coefficient of determination (r2) for logged seismic velocity versus Fe% is 40.5% indicating that 40.5% of the variability of logged seismic velocity is explained by the variability of Fe%. Iron content can be predicted from seismic velocity logging with a 95% confidence interval of ±12.76%Fe (see Figure 34).

Figure 34. Prediction of Fe% from logged seismic velocity at Louvicourt.

4.3 Quantitative Results -- Seismic Tomography

Figure 35 shows tomographic images for seismic tomography data collected between holes 655-284 and 655-302 ("Plane 284-302"), and between holes 655-304 and 655-302 ("Plane 304-302"), after 10 iterations of the SIRT algorithm. In each image, stringer-sulphide and massive-sulphide zones as interpreted by site geologists from geological logging data prior to the tomographic survey, are outlined in red.

Figure 35 shows higher seismic velocities within the "upper" stringer zone for Plane 284-302. Based on the relationships observed in the analysis of the seismic logging data, a higher iron/pyrite content is suggested within this high-velocity zone. Furthermore, the images suggest that the upper stringer zone may not extend as far up and to the left as previously interpreted from the geological data in the long-section. The velocity distribution in the "lower" stringer and massive sulphide zones suggests that the iron/pyrite content may not be as high as suggested in the upper stringer zone.

The image for plane 304-302 shows higher seismic velocities in the stringer zone suggesting higher iron/pyrite content within this zone. The lower seismic velocities within the massive-sulphide zone suggest lower iron/pyrite content within these massive sulphides. Very high velocities at the "nose" of the massive sulphide zone may indicate a zone of high iron/pyrite content. Note the higher velocities in the upper part of the massive-sulphide intersection within hole 304 on the left. The image suggests higher iron/pyrite content in this area.

Figure 35. Tomographic images of seismic velocity for both planes together after 10 iterations of the SIRT algorithm. Note that data for each plane was processed separately.

5.0 Discussion & Conclusions

Using commercially-available geophysical logging equipment, combined with multivariate statistics and pattern-recognition techniques, we have analyzed data acquired at selected Noranda Group minesites to provide answers to the following:

1) Can lithology be indirectly determined by geophysical logging at selected Noranda Group minesites, and, can geophysical logging differentiate ore sulphides from waste sulphides?

Based on our test data, interpretation of geophysical logs for lithology was very much site-dependent. In simple cases a single log or a cross-plot sufficed, while several logs and pattern recognition techniques were necessary in others.

Louvicourt:

The lithology of the Louvicourt deposit was found to be complicated from a geophysical point of view, and pattern recognition techniques were required to interpret the data. Two models, one using neural networks and one using discriminant analysis were used.

Most of the major rock classes were detected, but not always correctly classified. Despite this, there seems to be potential for lithological predictions with a high degree of accuracy at Louvicourt.

The ore boundaries were located almost exactly at the same place using geophysics and core, with few exceptions. However, the models could not distinguish between Zn- and Cu-sulphides due to an insufficient number of data-points for neural-network training. In borehole 565-025, massive sulphides were classified as semi-massive by the neural model. Furthermore, pyrite-classification was a problem, especially for the discriminant model. The neural model managed to classify some occurrences of pyrite correctly.

Alterations were usually, in part, classified correctly by the neural model while the discriminant model never classified alterations correctly. A problem with the alteration is that it often occurs with pyrite disseminations (e.g. borehole 565-026), where the altered zone before the ore is disseminated with varying grades of pyrite. The first part of the zone was classified as alteration, while the second part was classified as semi-massive sulphides according to the neural model.

Finally, rhyodacite was always classified as tuff by both models. By looking at the geophysical logs it is clear that there is little difference between tuff and rhyodacite.

Heath Steele:

At Heath Steele Mine, it was found that gamma-gamma, natural gamma, gamma ratio, and magnetic susceptibility data could be successfully used to isolate waste (i.e. sedimentary & volcanic rocks and, to some extent pyrite/pyrrhotite) and ore (i.e. massive and semi-massive Pb-Zn) units.

Isle Dieu:

At Isle Dieu Mine, crossplots of gamma-gamma versus natural gamma clearly identified four main groupings of data, corresponding to the four main lithological units.

To summarize, there can be differences between the core logs and the different models. However, these differences are usually small, and from a dilution-control point-of-view, almost negligible. Even though all of the geophysical tools attempted to measure lithology indirectly, most of the models show indications for future potential improvements. Since core-logging is assumed to be correct, all the modeling is directed towards finding a model that resembles the core-logs as close as possible. A drawback with this approach is that optical examination of core emphasizes different properties than those recorded by geophysical logging. Potential methods to minimize such errors would be to classify rock-units based on logging data, or on chemical & physical properties derived from lab tests of core samples. Furthermore, when training the models, the levels of pyrite-dissemination and alteration need to be quantified for correct classification of such zones (i.e. through chemical analysis).

2) To what accuracy can geophysical logging estimate grade at selected Noranda Group minesites?

Louvicourt:

Using multiple linear regression techniques, it was found that three of the metal assays at Louvicourt could be estimated -- in order of certainty: Fe, Cu, and Zn. Although these techniques tended to underestimate very high Zn grades, neural-network techniques were found to improve Zn-grade estimates (i.e. R2 increased from 0.45 to 0.70). Similarly, the Cu-equivalent grade could be estimated by applying neural-network techniques. Data from three of test holes showed that the correlation between the neural-network model and the Cu-equivalent assays was quite high with R2 = 0.77.

Heath Steele:

At Heath Steele, there was a good correlation between Pb-grade, Zn-grade, and the geophysical logs within the Pb-Zn sulphides. However, in boreholes containing many kinds of sulphides, these relationships were much more difficult to define. Cu grade could not be estimated with the available data using linear regression. In fact, the Cu grade was difficult to estimate using other methods as well, since the conductivity and susceptibility logs behaved peculiarly in three of the boreholes. However, reasonable estimates of Pb grade were obtained using neural networks.

Isle Dieu:

At Isle Dieu, simple linear regression of various geophysical logs could be used to estimate four different metal grades (i.e. Zn, Ag, Pb, & Cu). Estimates of the Zn, Ag, and Pb grades seemed to fit the assays fairly well.

3) Can geophysical logging be used to aid in the interpretation of crosshole geophysics tomograms from selected Noranda Group minesites?

Sonic logging results from the Louvicourt deposit showed that, of all of the minerals and metals assayed at Louvicourt, seismic velocity is most sensitive to iron content, followed by pyrite content. Therefore, as zones of higher seismic velocities may indicate zones of higher iron and pyrite content, this result was used in the interpretation of seismic tomography data acquired at the Louvicourt deposit.

This work has demonstrated the potential utility of geophysical-logging for dilution- and oreloss-control at Noranda Group mines. As part of our implementation strategy, Noranda Technology Centre has purchased and OMS-logg geophysical logging system to pursue the following:

1. Collect extensive single-hole geophysical logging data at various Noranda Group minesites to determine the practicality of single-hole geophysical logging for delineation and dilution control;
2. Develop a physical-property/geology database to be used to investigate relationships between single-hole geophysical logging data and geology/grade;
3. Select, refine, and implement appropriate analysis techniques for deriving geological parameters from geophysical logging data.

Acknowledgments

The authors would like to thank Noranda Mining & Exploration for funding this work, and allowing its publication. We are grateful for the enthusiastic support and assistance of Ed Stuart (Louvicourt Mine), Art Hamilton (Heath Steele Mines), and André Bonenfant (Matagami Division).

References

Bischof, H. Schnieder, W., & Pinz, A.J. (1992). Multispectral classification of landsat-images using neural networks. IEEE Transactions on Geoscience and Remote Sensing, Vol. 30 , No. 3, pp. 482 - 490.

Cinq-Mars, A., Mwenifumbo, C.J., Killeen, P.G., & Chouteau, M. (1992). Multiparameter geophysical logging at the Yava lead deposit: a statistical approach. Geophysical Prospecting, v. 40, pp. 829-848.

Cisar, D., Dickerson, J., & Novotny, T. (1993). Electromagnetic data evaluation using a neural network. Proceedings of the Symposium on the Application of Geophysics to Engineering and Environmental Problems; April 18-22, 1993; San Diego, California. pp. 599-612.

Emilsson, J. & Wänstedt, S. (1993). Geophysical multiparameter logging techniques applied to ore exploration in the Skellefte Field. Proceedings of the 5th International Symposium on Geophysics for Minerals, Geotechnical, and Environmental Application; 24-28 October 1993, Tulsa, Oklahoma.

Hair, JR. J. F., Anderson, R. E., Tatham, R. L. & Black, W. C. (1992). Multivariate data analysis with readings. McMillan, New York.

Miller, F., Jacob, D., & Potvin, Y. (1992). Cavity monitoring system - update & applications. Proceedings of the 94th Annual General Meeting of the Canadian Institute of Mining & Metallurgy; April 26-30, 1992; Montréal, Québec, Canada.

Mwenifumbo, C.J. & Killeen, P.G. (1987). Natural gamma ray logging in volcanic rocks: the Mudhole and Clementine base metal prospects. in Buchands Geology, Newfoundland, ed. R.V. Kirkham, Geological Survey of Canada, Paper 86-24, pp. 263-272.

Mwenifumbo, C.J. (1993). Kernel density estimation in the analysis and presentation of borehole geophysical data. The Log Analyst; September-October, 1993; pp. 34-45.

Mwenifumbo, C.J., Killeen, P.G., & Elliot, B.E. (1993). Classic examples from the Geological Survey of Canada data files illustrating the utility of borehole geophysics. Proceedings of the 5th International Symposium on Geophysics for Minerals, Geotechnical, and Environmental Application; 24-28 October 1993, Tulsa, Oklahoma.

Pflug, K.A., Killeen, P.G., & Mwenifumbo, C.J. (1994). Acoustic velocity logging at the McConnel nickel deposit, Sudbury are, Ontario: preliminary in-situ measurements. in Current Research 1994-C, Geological Survey of Canada, pp. 279-286.

Puhakka, R. (1991). Geological Waste Rock Dilution. The Finish Association of Mining & Metallurgical Engineers, Espoo, Finland. p. 30.

Seber, G. A. F. (1984). Multivariate observations. John Wiley and Sons, Inc., New York.

Urbancic, T.I. (1985). Multiparameter Geophysical Logging for Gold: A Statistical Approach. M.Sc.A. Thesis, University of Toronto. p. 274.

Wänstedt, S. & Sundin, N-O. (1991). Geophysical logging applied to mine planning in Swedish mines. International Symposium on Mine Mechanization & Automation; June 10-13, 1991; Golden, Colorado.

Wänstedt, S. (1992). Geophysical Borehole Logging Applied to the Location and Characterization of Base Metal Ores. Masters Thesis, University of Luleå, Luleå, Sweden. p. 120.

Wänstedt, S. (1993). An overview of recent research on geophysical logging in underground mines in Sweden. Proceedings of the 5th International Symposium on Geophysics for Minerals, Geotechnical, and Environmental Application; 24-28 October 1993, Tulsa, Oklahoma.