Professional Documents
Culture Documents
Conditional Similation PDF
Conditional Similation PDF
[W]hile the real surface z0 (x) is known only at a limited number of locations xi , . . . the simulated surface can be
known at almost every point x of the deposit. It is then possible to apply to the simulation the various processes
of extraction, hauling, stockpiling, etc., to study their technical and economic consequences and by feedback to
correct these processes. (Journel, 1974, p. 673).
284
Downloaded from https://www.cambridge.org/core. University of Warwick, on 22 Oct 2018 at 02:25:17, subject to the Cambridge Core terms of use, available at
Cambridge Books Online © Cambridge University Press, 2009
https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9780511545993.015
AN INTRODUCTION TO CONDITIONAL SIMULATION 285
25
(vii) Any combination of the purposes listed here.
20
Here, emphasis is placed on the use of conditional
Conditioning data simulations for characterizing grade continuity and
15 as a means of evaluating estimation methods.
Simulations produce values at the nodes of an ex-
Distance
tremely fine grid (i.e., very closely spaced nodes rel-
Figure 14.1: A comparison of idealized profiles of true grade
ative to the distance separating the conditioning data)
(solid line), simulated grade (dashed line), and kriged es-
timated grade (dotted line) on a common profile. Note the such that the character of a simulated deposit or do-
similar variability of the true and simulated grades in con- main is almost perfectly known by a large set of punc-
trast to the smoothed pattern of estimated grades. All pro- tual values. These are the values used for the various
files pass through the known data points. Redrawn from purposes listed previously. Because so many punctual
Journel (1975). values can be simulated within blocks of a practical
size (e.g., selective mining unit [SMU]), the average
application in a mining context because they do of all simulated values within a block can be taken as
not take into account the spatial characteristics of a close estimate of the true average value of the block.
the variable under study. Simulated data arrays that This opens the way for a number of practical appli-
retain the same density distribution (histogram) and cations of simulations (e.g., a mining sequence can
autocorrelation character and that are linked spatially be imposed on the blocks based on estimated grades
to reproduce the existing data are called conditional to determine the true impact of production on such
simulations. Conditional simulations are the type procedures as ore/waste selection, haulage schedul-
normally used for applications involving ore/waste ing, stockpiling, blending, and mill efficiency. Here,
grades. A conditional simulation and reality can be emphasis is directed to simulations as they relate di-
considered realizations of the same random function. rectly to the problem of obtaining high-quality block
Journel (1975) states, “Reality and simulation can be estimates in a production scenario.
considered as two variants of the same mineralized Various estimation methods can be applied to
phenomenon” (p. 8). blocks whose simulated true grades are known. Es-
timates by each method can be compared with true
block grades and the best method selected for use
14.2: AIMS OF SIMULATION in practice. When complex data distributions occur, it
Simulations serve a variety of purposes in the mineral has become common to use multiple indicator kriging
industry, including (see Journel, 1979) the following: to obtain grade estimates. One of the necessary steps
in this procedure involves a change of support opera-
(i) Study of grade continuity (e.g., Nowak et al., tion (see Chapter 12) that derives a block-grade distri-
1993) bution from a sample-grade distribution. Simulations
(ii) Optimizing sampling plans for advanced explo- can provide examples of the types of block-grade dis-
ration tributions that are encountered in practice that can be
(iii) Evaluation of resource/reserve estimation meth- compared with distributions obtained by change of
ods (e.g., Dowd and David, 1976) support procedures; simulations can also be used to
(iv) Mine planning (e.g., Blackwell et al., 1999) verify change of support procedures (e.g., Rossi and
Downloaded from https://www.cambridge.org/core. University of Warwick, on 22 Oct 2018 at 02:25:17, subject to the Cambridge Core terms of use, available at
Cambridge Books Online © Cambridge University Press, 2009
https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9780511545993.015
286 APPLIED MINERAL INVENTOR Y ESTIMATION
Parker, 1994). An important but little-used applica- The expected value of this kriging error is zero. Now
tion of simulation involves an understanding of the consider a random function Z 0 (x) independent of z(x)
spatial distribution of grade subpopulations (i.e., sim- but isomorphic (having similar structure) to it. That
ulations of the pattern of physical continuity in space is, Z s (x) has the same covariance C(h), as Z (x).
of various grade populations). Then the same kriging procedure leads to a com-
parable result (i.e., or any sampling of the random
14.3: CONDITIONAL SIMULATION AS AN function):
ESTIMATION PROCEDURE
∗ ∗
Z s (x) = Z sk (x) + [Z s (x) − Z sk (x)].
In certain respects, a simulated value is an estimate
of the true value at the site of the simulation. How-
ever, the simulated value is not a best estimate, and Because Eqs. 14.1 and 14.2 are isomorphic equiva-
the related estimation error is twice that of the cor- lents (realizations of the same random function), it is
responding kriging error. Consequently, conditional possible to mix components on their right sides to pro-
simulation has not been viewed as a reliable or ac- duce a new realization of the same random function,
cepted estimation procedure. Of course, simulation as follows:
and estimation have different purposes. Simulation
allows local variations in values of a variable to be Z c (x) = z k∗ (x) + [Z s (x) − Z sk
∗
(x)]
examined, particularly with regard to what impact
these local variations have on sampling plans, estima- where Z c (x) is constructed by combining the residuals
tion procedures, mine and mill planning, and financial of the second random function Z s (x) with the kriging
matters. Estimates, however, are made ultimately for Z k∗ (x) of the initial random function.
the purpose of identifying ore and waste on a scale Thus, simulation can be seen to be the addition of a
that physical separation can be achieved. simulated-error factor to a kriged value. Hence, krig-
There has been a trend toward the use of condi- ing is a product of the overall simulation procedure.
tional simulations as a realistic approach to estima- Equation 14.3 also demonstrates that the simulation is
tion, particularly for grade control during production conditional on the original data (i.e., reproduces the
(e.g., Blackwell et al., 1999). The general procedure original data) because, for the conditioning sample
involves the production of n simulations (for exam- sites, kriging reproduces the sample value, and the
ple, 10) of the same locations in space. The 10 values error term disappears.
simulated at each point define the probability distri-
bution for the grade at that point. These distributions
can be used in a variety of ways (e.g., to estimate the
probability that grade is above a cutoff at a point). Al- 14.5: SEQUENTIAL GAUSSIAN
ternatively, all such distributions in a specified block SIMULATION
can be combined to estimate the grade of the block Of the several simulation procedures in use (e.g., turn-
or to estimate the probability that the block grade is ing bands, LU decomposition, sequential Gaussian),
above a cutoff grade. only the sequential Gaussian procedure is considered
here. The method is based on the principal that an ap-
14.4: A GEOSTATISTICAL PERSPECTIVE propriate simulation of a point is a value drawn from
its conditional distribution given the values at some
Consider a single point, x, among many in space. This nearest points. The sequential nature rests on the fact
point has a real value of z 0 (x) and a kriged value of that subsequent points that are simulated make use
z k∗ (x) such that not only of the nearby original conditioning data, but
z 0 (x) = z k∗ (x) + [z 0 (x) − z k∗ (x)]. also the nearby previously simulated values. Software
(kriging error) for this purpose is readily available in Deutsch and
Downloaded from https://www.cambridge.org/core. University of Warwick, on 22 Oct 2018 at 02:25:17, subject to the Cambridge Core terms of use, available at
Cambridge Books Online © Cambridge University Press, 2009
https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9780511545993.015
AN INTRODUCTION TO CONDITIONAL SIMULATION 287
Journel (1998). The general procedure is as follows: This general situation (widely spaced data sites
relative to the average physical continuity of local
(i) Select a set of conditioning data. ore-bearing structures) is common to many deposits;
(ii) Transform conditioning data to equivalent nor- in particular, shear-controlled deposits such as many
mal scores. gold-bearing shear veins. When the exploration of
(iii) Develop a semivariogram model for the trans- such deposits has advanced to a stage that includes
formed data. underground workings (raises and drifts within the
(iv) Check transformed data for bivariate normal- structure) in addition to widely spaced piercement
ity by comparing sample indicator semivari- by diamond-drill holes, information can be collected
ograms for different thresholds (lower quartile, to construct a semivariogram model for gold grade
median, upper quartile) to a theoretical bivariate or accumulation and thickness. Underground access
model. provides the database with which to construct a semi-
(v) Proceed with sequential Gaussian simulation variogram model; diamond-drill intersections pro-
routine (from Deutsch and Journel, 1998). vide the conditioning data with which to develop
(vi) Conduct several checks to demonstrate that sim- simulations.
ulation has honored the desired constraints.
Downloaded from https://www.cambridge.org/core. University of Warwick, on 22 Oct 2018 at 02:25:17, subject to the Cambridge Core terms of use, available at
Cambridge Books Online © Cambridge University Press, 2009
https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9780511545993.015
288 APPLIED MINERAL INVENTOR Y ESTIMATION
Easting (ft) 15
≥ 0.951 0.195–0.361 35
0.639–0.951 0.090–0.195
t
0.361–0.639 < 0.090 50
0f
16
Figure 14.2: Contours for 302 blasthole Cu values (Sim-
ilkameen deposit). Data locations are omitted for clarity
but values are more or less uniformly distributed over the
field, with a general spacing of about 20 to 25 ft. Contours
demonstrate a preferential trend in a roughly northeasterly 80
40
direction.
25
125
(i) Select a specific area for which the simulation
will be performed (Fig. 14.2). 140
Downloaded from https://www.cambridge.org/core. University of Warwick, on 22 Oct 2018 at 02:25:17, subject to the Cambridge Core terms of use, available at
Cambridge Books Online © Cambridge University Press, 2009
https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9780511545993.015
AN INTRODUCTION TO CONDITIONAL SIMULATION 289
Columbia 0.12
Frequency (%)
Structure Range (ft)a C Anisotropy ratio
0.08
Nugget 0 0.3 1
Spherical 1 40 0.4 0.625
Spherical 2 160 0.3 0.688 0.04
Cu (%)
the same general character as the contoured condi-
b
tioning data (Fig. 14.2), except, as anticipated, the 0.16
N = 18000
relatively few and more widely spaced conditioning Mean = 0.4720
Std. Dev. = 0.3868
data are incapable of showing the local detail con-
0.12
tained in the simulated values. A more formal test Frequency (%)
of the similarity of spatial character is illustrated by
a comparison of the experimental semivariogram of 0.08
the simulated values with the semivariogram model
that characterizes the conditioning data. Figure 14.7
clearly demonstrates that the model for the condition- 0.04
Cu (%)
0.3
Figure 14.5: Histograms for (a) 302 blasthole Cu values
(conditioning data) and (b) 18,000 simulated Cu values,
Similkameen porphyry copper deposit. Note the similarity
0.2 of the two histograms.
γ (h)
Downloaded from https://www.cambridge.org/core. University of Warwick, on 22 Oct 2018 at 02:25:17, subject to the Cambridge Core terms of use, available at
Cambridge Books Online © Cambridge University Press, 2009
https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9780511545993.015
290 APPLIED MINERAL INVENTOR Y ESTIMATION
Table 14.2 Parameters for traditional least-squares linear models describing estimates versus true block
values, Similkameen test
a
Bracketed number is one standard deviation error.
cell (block) of the resulting array of 80 cells (8 × 10) against these true values on an x–y plot. Because the
is then estimated by several methods; in this case, expectation is for unbiased results with each estima-
nearest neighbor, various inverse-distance weighting tion technique, on average, producing the correct re-
procedures, and ordinary kriging using only the origi- sult, the data on x–y plots can be approximated by a
nal data. The true value of each block is taken to be the linear model. Three examples are shown in Fig. 14.8
average of all simulated points within the block. In- for ordinary kriging, inverse-distance weighting –
dividual estimation techniques can then be compared exponent 2; and inverse-distance weighting – expo-
nent 3. Statistical and linear model parameters are
summarized in Table 14.2.
This test demonstrates the relative quality of
various estimation methods as applied to an ideal-
3600
ized array of block values generated from a parent
population with statistical and spatial characteristics
3500 1.2
Northing (ft)
1.0
3400 0.8
γ (h)
0.6
3300 0.4
0.2
0.0
4400 4500 4600 4700 0 50 100 150 200
Downloaded from https://www.cambridge.org/core. University of Warwick, on 22 Oct 2018 at 02:25:17, subject to the Cambridge Core terms of use, available at
Cambridge Books Online © Cambridge University Press, 2009
https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9780511545993.015
AN INTRODUCTION TO CONDITIONAL SIMULATION 291
a
b
1.2 1.2
1.0 1.0
0.8 0.8
Cu (%)
Cu (%)
0.6 0.6
0.4 0.4
0.2 0.2
0.2 0.4 0.6 0.8 1.0 1.2 0.2 0.4 0.6 0.8 1.0 1.2
Cu (%) Cu (%)
1.50
1.25
1.00
Cu (%)
0.75
0.50
0.25
Cu (%)
Figure 14.8: Examples of block estimates by various estimation methods versus true block grades. (a) Ordinary kriging
estimates versus true block grades (x axis) (b) Inverse squared distance estimates versus true block grades. (c) Inverse
cubed distance estimates versus true block grades. Filled squares are individual 45 × 45 × 30 ft3 blocks. Linear models
are traditional least-squares fits, with error entirely in the ordinate (i.e., the y axis, which represents the estimation method).
Parameters for the linear models are listed in Table 14.2.
comparable to those of a part of the Similkameen bias (slope) and scatter of values (error) are greatest
porphyry copper deposit. In particular, kriging and for 1/d 3 . The choice between kriging and 1/d 2 is not
1/d 2 are clearly better than 1/d 3 , as indicated by the so clear, although, statistics aside, the kriging results
statistics in Table 14.2. Specifically, the conditional show much less scatter than do the1/d 2 results.
Downloaded from https://www.cambridge.org/core. University of Warwick, on 22 Oct 2018 at 02:25:17, subject to the Cambridge Core terms of use, available at
Cambridge Books Online © Cambridge University Press, 2009
https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9780511545993.015
292 APPLIED MINERAL INVENTOR Y ESTIMATION
14.7.5: Sequential Indicator Simulation literature would suggest. Most reported practical
applications relate to production scenarios, but
Another method also available in Deutsch and Journel
substantial scope for the use of conditional simu-
(1998) is sequential (multiple) indicator simulation
lations exists prior to exploitation at the advanc-
(SISIM). This does not require a Gaussian distribu-
ed exploration and feasibility stages of deposit
tion of the data and uses multiple indicator kriging,
development.
as described in Section 10.11.2, to model complex
4. When detailed semivariogram models are ob-
grade distribution patterns. The general procedure is
tained confidently by the judicious collecting of
as follows:
data early in an exploration program, conditional
(i) Select a set of conditioning data. simulation can provide remarkable insight into
(ii) Select a set of cutoff (threshold) grades and trans- grade continuity and how continuity can influ-
form to a set of indicator (0 or 1) data. ence the development of an adequate sampling
(iii) Develop semivariogram models for the various program, especially as regards sample spacing.
sets of indicator-transformed data.
(iv) Proceed with the simulation routine (from
Deutsch and Journel, 1998). 14.9: SELECTED READING
(v) Conduct several checks to demonstrate that the Deutsch, C., and A. Journel, 1998, GSLIB, Geosta-
simulation has honored the desired constraints. tistical software library and user’s guide; Oxford
In this case, the simulation routine randomly se- University Press, New York, 369 pp.
lects blocks for multiple indicator kriging (MIK) us- Journel, A. G., 1979, Geostatistical simulation: meth-
ing the original data and any nearby blocks already ods for exploration and mine planning; Eng. Min.
simulated. A random number (0 to 1) is selected and Jour., December, pp. 86–91.
applied to the MIK probability of being between two Journel, A. G., and E. H. Izaaks, 1984, Conditional in-
of the specified thresholds. The grade is interpreted dicator simulation: application to a Saskatchewan
linearly between the two threshold grades defined by uranium deposit; Math. Geol., v. 16, no. 7,
the random number selected. pp. 685–718.
Downloaded from https://www.cambridge.org/core. University of Warwick, on 22 Oct 2018 at 02:25:17, subject to the Cambridge Core terms of use, available at
Cambridge Books Online © Cambridge University Press, 2009
https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9780511545993.015
AN INTRODUCTION TO CONDITIONAL SIMULATION 293
at sites 0 to 19, inclusive. Note that this pro- at these sites are 1.2 and 4.1, respectively. Sup-
cedure produces a profile of a variable with a pose further that a random value at Site 2 is 1.1.
pure nugget effect. The conditioned value at Site 2 is determined as
follows. A line based on the two random points
2. Assume that at sites 0, 9, and 19 in Question 1, the has the equation y = 1.8 + (3.1 − 1.8) × /9.
values are known (and are to be used to condition For site x = 2, the corresponding line value of y is
the remaining values). The deposit is known to 2.1 (1.0 higher than the random value of 1.1). This
have alternating high- and low-grade zones on a difference must be maintained relative to a line
scale of 40 to 50 m. In this example, conditioning through the two known conditioning points. The
means that the position of intermediate samples equation for the conditioning line is y = 1.2 +
relative to a line joining two random end points 2.9 × /9 which, for site x = 2, gives y = 1.84.
that coincide with conditioning sites is maintained The conditioned value for Site 2 must be 1.0 lower
relative to a line joining the two conditioning than this point on the conditioning line; hence, the
points. conditioned value at Site 2 is 1.84 − 1.0 = 0.84.
Suppose the random values for Sites 0 and 9 Produce a profile of conditioned values from the
are 1.8 and 3.1, respectively, and the known values randomly selected values of Question 1.
Downloaded from https://www.cambridge.org/core. University of Warwick, on 22 Oct 2018 at 02:25:17, subject to the Cambridge Core terms of use, available at
Cambridge Books Online © Cambridge University Press, 2009
https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9780511545993.015