You are on page 1of 11

Comput Geosci (2009) 13:79–89

DOI 10.1007/s10596-008-9106-x

ORIGINAL PAPER

Simulation of mineral grades with hard and soft


conditioning data: application to a porphyry copper deposit
Xavier Emery · Lucía N. Robles

Received: 19 November 2007 / Accepted: 25 August 2008 / Published online: 4 October 2008
© Springer Science + Business Media B.V. 2008

Abstract This work deals with the geostatistical sim- 1 Introduction and scope of the work
ulation of mineral grades whose distribution exhibits
spatial trends within the ore deposit. It is suggested that Geostatistics is widely applied to characterize the spa-
these trends can be reproduced by using a stationary tial distribution of mineral grades in ore deposits. In
random field model and by conditioning the realiza- this respect, kriging techniques are used to predict the
tions to data that incorporate the available information grade at any location of the deposit by accounting for
on the local grade distribution. These can be hard data the available neighboring data (e.g., drill hole or blast
(e.g., assays on samples) or soft data (e.g., rock-type hole data) and their spatial correlation, described by
information) that account for expert geological knowl- a covariance function or a semivariogram [4, 17, 21].
edge and supply the lack of hard data in scarcely sam- However, since it relies on a weighted average of the
pled areas. Two algorithms are proposed, depending data values, kriging yields an oversmoothed image of
on the kind of soft data under consideration: interval the true grade distribution in the deposit. This is an
constraints or local moment constraints. An application important limitation when the quantity of interest is
to a porphyry copper deposit is presented, in which it is defined in relation to a threshold, e.g., the resources
shown that the incorporation of soft conditioning data (mean grade and tonnage) that can be recovered above
associated with the prevailing rock type improves the a given cutoff, or when the main task is the characteri-
modeling of the uncertainty in the actual copper grades. zation of the uncertainty in the actual grades.
To avoid these limitations, one can use conditional
simulation, which consists in constructing alternative
Keywords Geostatistics · Conditional simulation · realizations of a random field representing the mineral
Gibbs sampler · Soft conditioning data · Spatial trends grade. The realizations are conditioned to the data
values and reflect the spatial variability of the actual
grades [4, 15]. So far, most geostatistical models used
for simulation assume that the random field is station-
ary, i.e., that the finite-dimensional distributions are
invariant under a translation in space. In particular,
X. Emery (B) stationarity implies that the expected mean grade is
Department of Mining Engineering, University of Chile,
Avenida Tupper 2069, Santiago 837 0451, Chile
constant in space, an assumption that is questionable
e-mail: xemery@ing.uchile.cl when the grade distribution presents systematic vari-
ations or “trends.” For example, one can think of
L. N. Robles deposits in which the mean grades and patterns of
Department of Mining, Metals and Materials Engineering,
McGill University, 3450 University St., Montreal,
spatial continuity depend upon the rock type, porphyry
Quebec H3A 2A7, Canada copper disseminated deposits with a high grade core
e-mail: lucia.robles@mail.mcgill.ca surrounded by lower grade zones, or gold vein deposits
80 Comput Geosci (2009) 13:79–89

with swarms of mineralized veins included in barren To include the data in the simulation process, a
rocks. random field model is required for Z . One option is to
To account for these spatial variations of grades, use the indicator approach, in which the model is based
one approach consists in dividing the ore deposit into on the semivariograms of indicator variables [1, 16]. In
subdomains (geological units), in each of which the this study, we use another approach and assume that
grades are assumed stationary and independent of the Z is (up to a monotonic transformation) a stationary
grades in other subdomains. This division is particularly Gaussian random field, i.e., all its finite-dimensional
interesting if the subdomains so defined correspond to distributions are multivariate Gaussian.
geological events with different geneses. However, it
may be problematical when the boundaries between
subdomains are not hard, i.e., if the grade transition 2 Proposed methodology
when crossing a boundary is gradual (in which case the
independence assumption is violated), or when their 2.1 First approach: interval constraints
exact position is uncertain [9, 20, 22].
Another option is to use nonstationary random field In this subsection, we consider the case when the con-
models, i.e., models whose parameters (mean, variance, ditioning information consists of a set of hard data
semivariogram, etc.) vary in space. Commonly, the ran- located at {xα , α ∈ A} and another set of interval
dom field is split into two components: a deterministic constraints located at {xα , α ∈ B} [24]:
drift that accounts for the spatial trend and a stationary 
random residual. Possible approaches are simple krig- ∀α ∈ A, Z (xα ) = zα
(1)
ing with a local mean and universal kriging, depending ∀α ∈ B, Z (xα ) ∈ Iα
on whether the drift is considered known or not [7].
However, the dichotomy into drift and residual is often In such a case, one can use an iterative algorithm
tricky: a model with known drift may lack realism known as the “Gibbs sampler” to convert the interval
because it does not account for any uncertainty in the constraints into a set of exact data that honor these
drift, while the inference of the model parameters (in constraints ([10, 11, 14]; Appendix). These new data
particular, the histogram, and semivariogram) becomes (that will be referred to as soft data) have the same
difficult if the drift is unknown [4, 21]. semivariogram as the original hard data, i.e., the semi-
A third option, which will be considered hereafter, variogram of the Gaussian random field Z . The hard
is to reproduce the trend via the definition of condition- and soft data are then used for conditioning the realiza-
ing data. Indeed, although a random field is stationary tions of Z over domain D. At this stage, any algorithm
(thus, with a constant prior mean), the conditioning for simulating stationary Gaussian random fields can be
process forces its realizations to honor the data and, considered [4, 7].
therefore, to reproduce their local properties, in par- Note that the algorithm described in Appendix must
ticular, the local mean (trend) and local dispersion be run as many times as there are realizations of Z
(proportional effect; [17]). (using different random number seeds) so that the soft
In the following, we will denote by x a vector of data generated by the algorithm are different from
spatial coordinates and we assume that grades can be one realization to another. The realizations so obtained
represented by a stationary random field Z = {Z (x), x are independent samples from the distribution of the
∈ D} defined over a domain D (the ore deposit or a random field Z conditional to the data in Eq. 1.
portion of it). To reproduce spatial trends, two types of
conditioning data will be considered: 2.2 Second approach: local moments approximated
by co-kriging

• Hard data, which provide the exact grade values at In some cases, the use of interval constraints may be
sampling locations, for instance drill hole or blast considered restrictive since it prevents the simulated
hole assays. grades to be outside the assumed intervals at the soft
• Soft data, which consist of qualitative information data locations. In this section, an alternative approach is
derived from expert knowledge of the mining geol- presented, in which the interval constraints are replaced
ogist (e.g., expected rock types at given locations). by conditions on the local mean value and covariance
In scarcely sampled areas, these data will supply the function of the random field Z .
lack of hard information and ensure the reproduc- Denote by {xα , α ∈ A} the set of hard data locations
tion of the trend. and Z 1 = [Z (xα )]α∈A the vector of hard data values.
Comput Geosci (2009) 13:79–89 81

Also let {xα , α ∈ B} be the set of soft data locations 3 Application to a mining dataset
and pose Z 2 = [Z (xα )]α∈B . In the following, Z 2 will
be determined in order to incorporate geological infor- The proposed approaches are now illustrated with an
mation (e.g., rock-type information) codified through application to an exploration drill hole dataset from a
indicator variables. Chilean copper deposit.
The method consists of the following steps:
3.1 Presentation of the data
1. Model the simple and cross semivariograms of the
Gaussian random field Z and rock-type indica- The Río Blanco–Los Bronces porphyry copper deposit
tors. To obtain mathematically consistent semivar- is a breccia complex located in the Chilean Central
iograms, a linear model of coregionalization can be Andes at an elevation of between 3,500 and 4,100 m
used [28]. above mean sea level. It is composed of multiple min-
2. Perform co-kriging of the Gaussian random vector eralized breccias of biotite, anhydrite, tourmaline, and
Z 2 by using the available hard data (Z 1 ) and rock- specularite with fragments of host rocks (granodioritic
type indicator data. In general, the latter data are pluton and andesitic lavas). Other breccias of this zone
known at the hard data locations (drill holes), but have matrices dominated by tourmaline, anhydrite,
additional indicator data can be defined according and/or specularite with sulfides [12, 25–27].
to the geological knowledge of the deposit. At this The available dataset consists of 1,333 exploration
stage, any type of co-kriging can be used, e.g., drill-hole samples collected in and around the Don Luis
simple co-kriging (assuming known mean values open pit, situated in the central part of the deposit and
for the Gaussian and indicator data) or ordinary owned by Codelco Chile. The extension of the area
co-kriging (with unknown mean values). The co- under study is about 550 × 550 × 230 m. Each sample
kriging results consist of: (composited at 16-m length) contains information on
the assayed total copper grade and the rock type that
• the vector of prediction Z 2∗ ;
predominates in the sample (Fig. 1). Three main rock
• the covariance–variance matrix of the predic-
types can be distinguished:
tion error Z 2 − Z 2∗ . This matrix can be com-
puted once the semivariograms models and
• Granitoid. Located in the eastern part of the sam-
data locations (hard and indicator data used as
pled area; this is one of the host rocks of the breccia
co-kriging input, target soft data) are known. It
complex.
is a positive semidefinite matrix, provided that
• Tourmaline breccia. It is located in the central part
consistent semivariograms models have been
of the sampled area and is composed of granodior-
defined at step 1.
ite clasts surrounded by matrix cement dominated
3. Simulate Z 2 as a Gaussian random vector with by tourmaline and sulfides (chalcopyrite, pyrite,
mean equal to the co-kriging prediction and molybdenite, and minor bornite).
covariance–variance matrix equal to that of the • Porphyry. It consists of low-graded bodies that
prediction errors. The simulation can be done by intruded in the central and western parts of the
using the matrix decomposition algorithm [2, 5]. sampled area.
4. Simulate Z over the domain of interest, condi-
tionally to Z 1 and Z 2 . Again, to obtain a set of The main statistics of the copper grades for each rock
independent realizations of the random field Z type and for the whole dataset are indicated in Table 1.
conditioned to the hard and geological (rock type)
information, step 3 must be run as many times as
3.2 Exploratory and semivariogram analyses
there are realizations of Z , so that each realization
uses a different outcome of Z 2 .
We are interested in reproducing the trends in the
copper grade distribution from the center (high-graded
The identification of the co-kriging prediction and tourmaline breccia) to the eastern and western borders
the covariance–variance matrix of co-kriging errors of the area under study (medium-graded granitoid and
with the conditional moments of Z 2 (step 3) is approxi- low-graded porphyry). The study is achieved as follows:
mate. It would be justified only if the conditioning data
had multivariate Gaussian distributions, which is not 1. Cell declustering of the data histogram in order to
the case of the indicator data (binary data). account for irregularities in the sampling design [7]
82 Comput Geosci (2009) 13:79–89

Fig. 1 Copper grades and rock types for the drill-hole data located between elevations 3700 and 3750 m

2. Transformation of the grade data into normal 3.3 Location of soft conditioning data
scores by using the declustering weights calculated
in the previous step To define the soft conditioning constraints, we will use a
3. Semivariogram analysis of the normal scores data; rock-type model provided by Codelco Chile and based
the experimental semivariogram is fitted by a on the geological interpretation of the ore body [27].
nested model composed of a nugget effect, an ex- We consider a set of control points at the nodes of
ponential, and two spherical structures: a regular grid with mesh 50 × 50 × 32 m and, from
the geological model, we determine to which rock type
γCu = 0.19 Nugget + 0.54 Exp (20 m, 115 m) these points are expected to belong (Fig. 2).
+ 0.26 Sph (200 m, 200 m)
+ 0.07 Sph(200 m,∞) (2) 3.4 Interval constraints

In the above expression, the distances into brack- Interval constraints at the soft data locations (control
ets correspond to the practical ranges along the points) are defined in accordance with the local copper
horizontal and vertical directions (main anisot- grade statistics in each rock type for the area under
ropy directions), respectively. The semivariogram consideration (Table 1). It is found that, for the points
model indicates that the spatial continuity of cop- located in porphyry, copper grades should be less than
per grades is more pronounced along the vertical 0.8%. In contrast, no restriction apply to the points
direction. located in granitoid or in breccia, as the grades in both
4. Definition of soft conditioning data and simulation rock types can reach an upper bound of about 2.3%,
of copper grades (see next subsections). which is the overall maximum grade.

Table 1 Statistics on original Granitoid Breccia Porphyry Total


(grade) data
Number of data 449 672 212 1333
Minimum (% Cu) 0.04 0.03 0.02 0.02
Lower quartile (% Cu) 0.51 0.65 0.11 0.47
Median (% Cu) 0.66 0.83 0.19 0.70
Upper quartile (% Cu) 0.87 1.03 0.31 0.93
Maximum (% Cu) 2.30 2.23 0.77 2.30
Mean (% Cu) 0.720 0.863 0.219 0.712
Standard deviation (% Cu) 0.291 0.318 0.143 0.364
Comput Geosci (2009) 13:79–89 83

Fig. 2 Rock-type model for the bench with elevation 3,700 m and locations of soft conditioning data

3.5 Moment constraints 3.6 Results

Local moment constraints are defined by co-kriging In each case (interval or moment constraints), the hard
the Gaussian random field representing the copper and soft data are used to construct 500 realizations of
grade at the soft data locations, using as input data the copper grades over the area under study, by using
the normal score transform of copper grades known at the methodology described in Section 2.
the hard data locations, as well as the breccia (B) and Figure 4b and c show the average simulated grades
porphyry (P) indicators known at the hard data and soft over the bench with elevation 3,700 m. For compari-
data locations. Here, ordinary co-kriging is used, with son, Fig. 4a displays the average simulated grades of
the following simple and cross semivariogram models realizations conditioned only to the hard data (drill
(Fig. 3): hole grade data) for the same bench. One observes
that accounting for soft conditioning data yields a lower
⎛ ⎞
γCu γCu−B γCu−P average simulated grade in the southwestern sector of
⎝ γCu−B γB γ B−P ⎠ the area under study, in agreement with the geological
γCu−P γ B−P γP knowledge on the rock type distribution (presence of
⎛ ⎞ low-graded porphyry, as seen in Fig. 2), whereas the
0.19 0.03 −0.01 grade model obtained by using only the hard data is
= ⎝ 0.03 0.027 −0.01 ⎠ Nugget likely to overstate the actual grades in this sector.
−0.01 −0.01 0.004 Because interval constraints are not active in breccia
⎛ ⎞ and granitoid, the grade models obtained by accounting
0.54 0.16 −0.19
+ ⎝ 0.16 0.082 −0.09 ⎠ Exp (20 m, 115 m) for such constraints (Fig. 4b) and by using only the
−0.19 −0.09 0.10 hard data (Fig. 4a) show little difference in the central
⎛ ⎞ and eastern sectors, where one expects the presence
0.26 0.033 −0.038 of breccia or granitoid. More contrasts are observed in
+ ⎝ 0.033 0.06 0.004 ⎠ Sph (200 m, 200 m) the grade model obtained by using moment constraints
−0.038 0.004 0.007 (Fig. 4c), as the recourse to local means allows dis-
⎛ ⎞ tinguishing between high-graded breccia and medium-
0.07 0.018 −0.007
graded granitoid, even if the range of copper grades in
+ ⎝ 0.018 0.06 0.002 ⎠ Sph (200 m, ∞) (3)
both rock types is the same (between 0% and 2.3%).
−0.007 0.002 0.015
This suggests that the local moment constraints convey
richer information than interval constraints. Indeed,
Equation 3 yields a valid model of coregionalization, unlike the latter constraints, the moment constraints
insofar as the sill matrices are positive semidefinite (the are “regionalized,” as they take into account the neigh-
eigenvalues are nonnegative). boring hard and soft data and the coregionalization
84 Comput Geosci (2009) 13:79–89

Fig. 3 Simple and cross semivariograms for the normal scores data and rock-type indicators along the main directions of anisotropy
(horizontal and vertical). Experimental semivariograms are indicated by dashed lines, models are indicated by solid lines

model between copper grades and rock-type indicators. interval constraints, this conditional variance decreases
To corroborate these statements, let us compare the by 5% with respect to the case when these constraints
conditional variance of the simulated Gaussian random are ignored and only the hard data are accounted for
field at the soft data locations. When incorporating (average variance of 0.87 vs. 0.91); the decrease mainly
Comput Geosci (2009) 13:79–89 85

a b c
Fig. 4 a–c Average simulated copper grades (500 conditional realizations), depending on whether or not soft data (interval constraints
or moment constraints) are considered. Grayscale representation of a planar section (bench with elevation 3,700 m)

affects the locations in porphyry rock type (Fig. 5a). the original dataset (1,333 data), we define a prediction
In contrast, when incorporating local moment con- subset containing n1 = 729 data and a validation subset
straints, the average conditional variance decreases by containing the remaining n2 = 604 data (Fig. 6).
half (from 0.91 to 0.45), which indicates a much more The exercise consists in simulating the copper grades
significant reduction of the uncertainty in the actual at the validation subset by taking into account the
grades (Fig. 5b). grades at the prediction subset as well as the soft data
defined in Fig. 2. The validation is performed through
3.7 Split-sample validation the following steps:

A validation exercise is now proposed to assess the 1. Draw a set of 500 conditional realizations of the
accuracy of the grade models displayed in Fig. 4. From copper grades at the locations of the validation

a b
Fig. 5 a, b Conditional variances of the simulated Gaussian random field at the soft data locations, depending on whether or not soft
data are considered
86 Comput Geosci (2009) 13:79–89

only (Fig. 7 and Table 2). In this last model, the errors
for the first seven points ( p = 0.1 to p = 0.7) are much
greater than the aforementioned standard deviations,
so that the uncertainty model is deemed inaccurate.
This statement is corroborated by calculating the good-
ness statistics [6]:
1

G=1− a ( p) p̂ − p dp (4)
0

with a( p) = 1 if p̂ > p, −2 otherwise. Maximum good-


ness corresponds to G = 1, while G = 0 reflects a
totally inaccurate model. In the present case, the good-
ness statistics is significantly closer to 1 when using soft
data (interval or moment constraints) than when using
the hard data only (Table 2).
For comparison, two indicator kriging approaches
are also put to the test. The first approach uses the
grade data at the hard data locations, codified into a set
of indicators (0/1) associated with given thresholds (the
deciles of the global grade distribution). The second ap-
Fig. 6 Prediction and validation subsets, defined by splitting the proach uses the same hard data, as well as the interval
area under study into cells with side length 150 m and by assigning constraints at the soft data locations; the latter are cod-
one of two cells (alternatively) to each subset ified into 0, 1 or missing values, depending on whether
the threshold is above, below or within the interval [16].
Again, it is observed that the incorporation of interval
data. Three models are put to the test, depending
on whether or not soft data are incorporated in the
conditioning data and whether these soft data con-
sist of interval constraints or moment constraints.
2. From the simulated grades, construct a set of sym-
metric intervals with probabilities ranging from 0.1
to 0.9. The interval
 with probability
 p is bounded
by the (1 − p) 2 and (1 + p) 2 percentiles of the
distribution of simulated grades.
3. For each probability p, determine the actual pro-
portion p̂ of validation data that belong to the
corresponding probability intervals, and plot this
proportion versus p (accuracy plot) [6].
4. Check that the use of soft data improves the un-
certainty model (accuracy plot closer to the first
bisector). If there were no spatial correlation, the
standard
 deviation of the error p̂ − p would be

p (1 − p) n2 (with n2 the number of data in the
validation subset), i.e., 0.012 for p = 0.1 or p = 0.9,
and 0.020 for p = 0.5. These figures give a rough
idea of the expected deviation between p̂ and p in
the accuracy plot.
In the present case study, the models based on hard Fig. 7 Accuracy plots for the copper grades simulated at the
and soft data (either interval or moment constraints) validation data subset, depending on whether or not soft data
yield better results than that based on the drill hole data (interval constraints or moment constraints) are considered
Comput Geosci (2009) 13:79–89 87

Table 2 Theoretical probabilities and actual proportions of data within probability intervals, depending on whether soft data (interval
constraints or moment constraints) are considered or not
Theoretical Actual proportion of data within interval
probability Hard data only Hard data + interval constraints Hard data + moment constraints
MG IK MG IK MG
0.1 0.149 0.126 0.098 0.109 0.089
0.2 0.258 0.222 0.204 0.202 0.202
0.3 0.383 0.318 0.298 0.282 0.305
0.4 0.495 0.422 0.397 0.359 0.396
0.5 0.589 0.548 0.515 0.483 0.508
0.6 0.681 0.656 0.609 0.584 0.628
0.7 0.768 0.753 0.715 0.704 0.707
0.8 0.844 0.838 0.828 0.803 0.815
0.9 0.902 0.934 0.916 0.924 0.896
Goodness statistics 0.937 0.965 0.989 0.975 0.989
MG Multivariate Gaussian Model, IK Indicator Kriging

constraints improves the uncertainty model obtained 4 Discussion and conclusions


by using only the hard data (Fig. 8 and Table 2). Here,
the proposed simulation approach based on interval The proposed methodology, based on the incorpora-
constraints (Fig. 7) gives slightly better results than the tion of soft conditioning data, is simple to implement
indicator approach, which may be explained because and avoids splitting the random field to simulate into
the multivariate Gaussian model is well suited to the a drift and a residual, as it is often assumed in geosta-
description of grades in disseminated deposits like the tistical trend models. In particular, the prior (noncon-
one under study. ditional) random field model remains stationary, which
yields a convenient framework for inferring the para-
meters (Gaussian transformation function and semivar-
iogram of normal scores data). The reproduction of the
spatial trends is forced by the conditioning to data.
The soft data locations must be defined in accor-
dance with the expert judgment of the mining geologist,
e.g., expected rock type, mineralization, or alteration.
Note that soft data are needless in areas where the hard
data are abundant, as the latter data will be sufficient to
reproduce the local trends [18]. Instead, the definition
of soft data is crucial in areas where the drill-hole
spacing is large or in extrapolation situations in order to
supply the lack of hard conditioning data. The spacing
of the soft data locations can be chosen according to
the range of correlation: a large spacing may be enough
in case of a strong spatial continuity, while a smaller
spacing should be considered if the grades are highly
variable in space. Care must also be taken with the
definition of the soft conditioning information (in par-
ticular, with the interval constraints), as it may lead to
biased predictions or inaccurate uncertainty models if
this information is inappropriate or is in disagreement
with the available hard information. As suggested in
this work, the soft data suitability can be checked by
Fig. 8 Accuracy plots for indicator kriging estimates at the
validation data subset, depending on whether or not interval using split-sample validation or leave-one-out cross-
constraints are considered validation.
88 Comput Geosci (2009) 13:79–89

Acknowledgements The first author acknowledges the fund- References


ing from the Chilean Fund for Science and Technology Devel-
opment, through research project FONDECYT 1061103. The
1. Alabert, F.: Stochastic imaging of spatial distributions us-
manuscript benefited from the comments by two anonymous
ing hard and soft information. Master’s thesis, Stanford
reviewers.
University, Stanford (1987)
2. Alabert, F.: The practice of fast conditional simulations
through the LU decomposition of the covariance matrix.
Math. Geol. 19(5), 369–386 (1987). doi:10.1007/BF00897191
3. Armstrong, M., Galli, A., Le Loc’h, G., Geffroy, F., Eschard,
Appendix R.: Plurigaussian Simulations in Geosciences. Springer,
Berlin (2003)
Consider a stationary Gaussian random field Z = 4. Chilès, J.P., Delfiner, P.: Geostatistics: Modeling Spatial Un-
{Z (x), x ∈ D} and the conditioning constraints given certainty. Wiley, New York (1999)
5. Davis, M.W.: Production of conditional simulations via the
in Eq. 1. The Gibbs sampler allows replacing the in- LU triangular decomposition of the covariance matrix. Math.
terval constraints by a set of values that comply with Geol. 19(2), 91–98 (1987)
these constraints and with the spatial continuity of the 6. Deutsch, C.V.: Direct assessment of local accuracy and pre-
Gaussian random field. The algorithm is described in cision. In: Baafi, E.Y., Schofield, N.A. (eds.) Geostatistics
Wollongong’96, pp. 115–125. Kluwer Academic, Dordrecht
the following: (1997)
7. Deutsch, C.V., Journel, A.G.: GSLIB: Geostatistical Soft-
ware Library and User’s Guide, 2nd edn. Oxford University
1. For α ∈ B, assign to Z (xα ) any value zα within Press, New York (1998)
interval Iα . 8. Emery, X.: Using the Gibbs sampler for conditional simula-
2. Select an index β at random (uniformly) in B. tion of Gaussian-based random fields. Comput. Geosci. 33(4),
522–537 (2007). doi:10.1016/j.cageo.2006.08.003
3. Determine the distribution of Z (xβ ) conditioned to 9. Emery, X., González, K.E.: Incorporating the uncertainty in
all the other data {Z (xα ), α ∈ A ∪ B, α = β}. Since geological boundaries into mineral resources evaluation. J.
Z is a stationary Gaussian random field, this dis- Geol. Soc. India. 69(1), 29–38 (2007)
tribution is normal, with expectation and variance 10. Freulon, X.: Conditional simulation of a Gaussian ran-
dom vector with nonlinear and/or noisy observations. In:
equal to the simple kriging prediction and simple Armstrong, M., Dowd, P.A. (eds.) Geostatistical Simulations,
kriging variance of Z (xβ ) by {Z (xα ), α ∈ A ∪ B, pp. 57–71. Kluwer Academic, Dordrecht (1994)
α = β}, respectively. 11. Freulon, X., de Fouquet, C.: Conditioning a Gaussian model
4. Generate a value zβ according to the previous con- with inequalities. In: Soares, A. (ed.) Geostatistics Tróia’92,
pp. 201–212. Kluwer Academic, Dordrecht (1993)
ditional distribution. 12. Frikken, P.H., Cooke, D.R., Walshe, J.L., Archibald, D.,
5. If zβ belongs to interval Iβ , then set Z (xβ ) = zβ . Skarmeta, J., Serrano, L., et al: Mineralogical and isotopic
6. Go back to step 2 and loop until a large number of zonation in the Sur-Sur tourmaline breccia, Río Blanco - Los
iterations are done. Bronces Cu-Mo deposit, Chile: implications for ore genesis.
Econ. Geol. 100(5), 935–961 (2005). doi:10.2113/100.5.935
13. Galli, A., Gao, H.: Rate of convergence of the Gibbs sam-
pler in the Gaussian case. Math. Geol. 33(6), 653–677 (2001).
It can be shown [10] that the set of simulated values
doi:10.1023/A:1011094131273
{zα , α ∈ B} converges to a realization of the Gaussian 14. Geman, S., Geman, D.: Stochastic relaxation, Gibbs distrib-
random vector {Z (xα ), α ∈ B} conditioned to the con- utions, and the Bayesian restoration of images. IEEE Trans.
straints given in Eq. 1. The number of iterations needed Pattern Anal. Mach. Intell. 6(6), 721–741 (1984)
15. Journel, A.G.: Geostatistics for conditional simulation of ore-
to ensure such a convergence depends on the semivar-
bodies. Econ. Geol. 69(5), 673–687 (1974)
iogram of the Gaussian random field and on the data 16. Journel, A.G.: Constrained interpolation and qualitative in-
locations and is still unknown to a great extent. For formation: the soft kriging approach. Math. Geol. 18(3), 269–
some partial results on the rate of convergence of the 286 (1986). doi:10.1007/BF00898032
17. Journel, A.G., Huijbregts, C.J.: Mining Geostatistics. Acad-
Gibbs sampler, the reader is referred to [3, 13, 19, 23].
emic Press, London (1978)
The convergence can also be checked empirically, by 18. Journel, A.G., Rossi, M.E.: When do we need a trend model
calculating the sample semivariogram of the simulated in kriging? Math. Geol. 21(7), 715–739 (1989). doi:10.1007/
values over several realizations and comparing it to the BF00893318
19. Lantuéjoul, C.: Geostatistical Simulation: Models and Algo-
semivariogram model [8]. As a rule of thumbs, it is
rithms. Springer, Berlin (2002)
advisable to update each soft datum several hundreds 20. Larrondo, P., Leuangthong, O., Deutsch, C.V.: Grade estima-
or thousands of times. tion in multiple rock types using a linear model of coregion-
Comput Geosci (2009) 13:79–89 89

alization for soft boundaries. In: Magri, E., Ortiz, J., 25. Serrano, L., Vargas, R., Stambuk, V., Aguilar, C., Galeb, M.,
Knights, P., Henríquez, F., Vera, M., Barahona, C. (eds.) Holmgren, C., et al: The late Miocene to early Pliocene Río
International Conference on Mining Innovation Minin 2004, Blanco-Los Bronces copper deposit, Central Chilean Andes.
pp. 187–196. Gecamin, Santiago (2004) In: Camus, F., Sillitoe, R.H., Petersen, R. (eds.) Andean Cop-
21. Matheron, G.: The Theory of Regionalized Variables and per Deposits: New Discoveries, Mineralizations, Styles and
its Applications. Cahiers du Centre de Morphologie Math- Metallogeny, pp. 119–130. Society of Economic Geologists,
ématique de Fontainebleau, Ecole Nationale Supérieure des Littleton (1996), Special Publication no. 5
Mines de Paris, Paris (1971) 26. Skewes, M.A., Holmgren, C., Stern, C.R.: The Donoso
22. Ortiz, J.M., Emery, X.: Geostatistical estimation of min- copper-rich, tourmaline-bearing breccia pipe in central Chile:
eral resources with soft geological boundaries: a comparative petrologic, fluid inclusion and stable isotope evidence for
study. J. S. Afr. Inst. Min. Metall. 106(8), 577–584 (2006) an origin from magmatic fluids. Miner. Depos. 38(1), 2–21
23. Roberts, G.O., Sahu, S.K.: Updating scheme, correlation (2003). doi:10.1007/s00126-002-0264-9
structure, blocking and parametrization for the Gibbs sam- 27. Vargas, R., Gustafson, L.B., Vukasovic, M., Tidy, E., Skewes,
pler. J. R. Stat. Soc. B. 59(2), 291–317 (1997). doi:10.1111/ M.A.: Ore breccias in the Río Blanco-Los Bronces porphyry
1467-9868.00070 copper deposit, Chile. In: Skinner, B.J. (ed.) Geology and
24. Robles, L., Emery, X., Ortiz, J.M.: Geostatistical simulation Ore Deposits of the Central Andes, pp. 281–297. Society of
of mineral grades in the presence of spatial trends. In: Magri, Economic Geologists, Littleton (1999), Special Publication
E.J. (ed.) 33rd International Symposium on Applications of no. 7
Computers and Operations Research in the Mineral Industry 28. Wackernagel, H.: Multivariate Geostatistics: An Introduction
APCOM 2007, pp. 87–93. Gecamin, Santiago (2007) with Applications, 3rd edn. Springer, Berlin (2003)

You might also like