Roberto Mentzingen Rolo et al.
Mining
Mineração
Signed distance function
[Link] implicit geologic modeling
Roberto Mentzingen Rolo Abstract
Mestrando
Universidade Federal do Rio Grande do Sul - UFRS Prior to every geostatistical estimation or simulation study there is a need for de-
Departamento de Engenharia de Minas limiting the geologic domains of the deposit, which is traditionally done manually by
Porto Alegre – Rio Grande do Sul – Brasil a geomodeler in a laborious, time consuming and subjective process. For this reason,
robertorolo@[Link] novel techniques referred to as implicit modelling have appeared. These techniques
provide algorithms that replace the manual digitization process of the traditional
Ricardo Radtke methods by some form of automatic procedure. This paper covers a few well estab-
Mestrando lished implicit methods currently available with special attention to the signed distance
Universidade Federal do Rio Grande do Sul - UFRS function methodology. A case study based on a real dataset was performed and its
Departamento de Engenharia de Minas applicability discussed. Although it did not replace an experienced geomodeler, the
Porto Alegre – Rio Grande do Sul – Brasil method proved to be capable in creating semi-automatic geological models from the
ricardo.radtke10@[Link] sampling data, especially in the early stages of exploration.
João Felipe Coimbra Leite Costa Keywords: implicit geologic modeling, signed distance function, domaining.
Professor titular
Universidade Federal do Rio Grande do Sul - UFRS
Departamento de Engenharia de Minas
Porto Alegre – Rio Grande do Sul – Brasil
jfelipe@[Link]
1. Introduction
The necessary investment to start a for geological and statistical homogene- and laborious.
mine is in the order of tens to hundreds of ity within chosen domains (McLennan, For most mines, only a single work-
millions of dollars. For the investment to 2007). Generating accurate boundary ing model is maintained because of the
be profitable, the potential product in the models is clearly necessary, as the qual- time constraint. Rarely is there an oppor-
subsurface must be present in adequate ity of the models will influence various tunity to model alternative interpreta-
quantity and quality to justify a decision downstream mine practices, including tions and compare resource estimations
to invest. estimations/simulations and mining based on the alternative models; thus, the
All technological and financial de- planning (Cowan et al., 2003). These assessment of mining risks is avoided,
cisions are built on the knowledge about boundaries, separating different station- even though it is inherent in geological
the mineral deposit. Thus, the estimation ary domains, are traditionally modeled modeling (Cowan et al., 2003).
of grade and location of material in the explicitly by manual digitizing from Mining software packages have
ground (in situ resources) must be known sample data. provided computational tools to dis-
with an acceptable degree of confidence. First, two-dimensional poly-lines play drillhole data and to speed up the
A small difference between planned (es- are manually drawn on cross sections manual digitization of cross sections.
timated) and realized production have honoring the sample data, and then Despite these advances, explicit model-
a large impact on mine profitability linked by tie lines, these tied poly-lines ing still suffers from the disadvantages
(Sinclair and Blackwell, 2002). are triangulated to create a wire-frame presented. Recent novel techniques
The resource evaluation of a min- that represents a geologic unit, this referred to as implicit modeling provide
ing deposit is composed of two steps process has some disadvantages (Cowan algorithms that reduce the level of sub-
(Chilès et al., 2004): Delimitation of et al., 2003): its time consuming and jectivity by replacing the digitization
the boundaries of the units correspond- requires an experienced geomodeler process with some form of automatic
ing to the various geological formations to construct complex geometries; the procedure (Silva, 2015).
and estimation or simulation of grades model produced is unique to each in- The aim of this article is to inves-
within each unit. dividual geomodeler, and cannot be tigate the applicability of the signed dis-
Therefore, prior to every geosta- replicated, making auditing by external tance function methodology (Deutsch
tistical estimation or simulation, there personnel a hard task; it is inflexible, and Wilde, 2013; Silva, 2015) as a
is a need for delimiting the geologic since modifying the model as new data substitute or ancillary method for
units, implying a stationarity decision becomes available is time consuming traditional geologic modeling. The
REM, Int. Eng. J., Ouro Preto, 70(2), 221-229, apr. jun. | 2017 221
Signed distance function implicit geologic modeling
algorithm was applied in a real gold de- compared to an explicit model created interpreted geological structures were
posit, results discussed and the implicit by a geomodeler, in order to check if reproduced by the algorithm.
2. Geologic modeling
There are two groups of geologic form of domain delimitation. Implicit herein to illustrate the methodology.
modeling techniques available: methods are straightforward and com- 2) Stochastic techniques require
1) Deterministic techniques, which putationally fast. Some well-established more computational effort, and are
comprise explicit methods where the implicit methodologies are reviewed in not so straightforward. The established
boundary that separates the geological this paper: The Leapfrog© methodology algorithms are: sequential indicator
domains is defined explicitly by the (Cowan et al., 2003), potential field simulation (Journel, 1989b), trun-
geomodeler, in vertical and horizontal (Chilès et al., 2004) and signed distance cated gaussian simulation (Journel and
sections, from the drillhole data and function technique (Deutsch and Wilde, Isaaks, 1984), plurigaussian simulation
implicit methods which replace the 2013; Silva, 2015). For the latter, a (Galli et al., 1994) and multi-point
manual digitization of the explicit meth- special plug-in was developed to run at based methods (Guardiano and Sriv-
ods by an automatic or semi-automatic SGeMS. A real case study is presented astava, 1993).
3. Implicit methods
3.1 Leapfrog© methodology
The subject was introduced in the (McLennan, 2007). The volume func- porating geological morphology; (4)
geology field by Cowan et al., (2003), tion must be interpolated in order to interpolating the geological morphol-
based on the work of Savchenko et al., define the boundary interface, by imple- ogy; (5) morphologically constrained
(1995) for modeling objects interpolat- menting a fast scattered interpolator interpolation.
ing volume functions. The definition of method know as radial basis function RBF functions do not derive the
a volume function is attached to the no- (RBF) (Hardy, 1990). The interpolator covariance functions from the data;
tion of distance to an interface where the is represented as the linear combination they correspond to a simple isotropic
interface is defined as the surface sepa- of basic functions similar to dual kriging linear covariance function. Therefore,
rating two distinct domains. Distance is (Journel, 1989a). there is no possibility of incorporating
measured to the nearest interface, and it Leapfrog© provides one of the first anisotropy into the boundaries through
can be negative or positive depending on implicit boundary modeling implemen- RBF interpolation. Instead it is injected
whether the location is inside or outside tations within a commercial software manually in the form of deterministic
of the domain. The bounding interface package. There are five major steps in morphological constraints (McLen-
of interest is the surface corresponding the methodology (McLennan, 2007): nan and Deutsch, 2006). Besides this,
to a particular iso-value of the volume (1) data validation and composting; (2) dealing with multiple domains is not
function, usually the iso-surface zero interpolation and meshing; (3) incor- straightforward (McLennan, 2007).
3.2 Potential field
Presented by Chiles et al., (2004), the form of the locally varying drift; cokriging to optimally account for
this is an implicit 3D scalar field from (3) infer the potential field covariance both intersection and structural
which a geological interface is extract- function; (4) interpolate the potential dip data. The procedure is not sim-
ed as a particular iso-surface. There field with the universal cokriging ap- ple, the covariance of the potential
are five major steps to the methodology proach; (5) visualize the uncertainty in field is particularly difficult to infer,
(McLennan and Deutsch, 2006): (1) the boundary surface placement. since there are no hard potential
collect surface intersection and struc- A key feature of the potential field data available (McLennan and
tural orientation data; (2) determine field method is the use of universal Deutsch, 2006).
3.3 Distance function modeling
Signed distance function meth- interesting modifications were made The interface that separates the regions
odology for multiple categories (Silva, (Silva, 2015). in space is determined by the sign of the
2015) is a deterministic implicit model- For each sample, an anisotropic (or estimated signed distance values. The
ing technique where the implicit func- isotropic) distance between itself and the algorithm works as follows.
tion is constructed by interpolating a nearest sample belonging to an opposite The set of data z(u α),α = 1,...,n
distance measure based on conditioning domain is computed and assigned. Nega- is coded in indicators to specify the
data. This method is based on the vol- tive values represent the distance to the samples that belong inside or outside of
ume function technique, however, some boundary for samples inside the domain. the domains (Equation 1).
1, if z (u ) is inside
i (u ) = (1)
0, if z (u ) is outside
222 REM, Int. Eng. J., Ouro Preto, 70(2), 221-229, apr. jun. | 2017
Roberto Mentzingen Rolo et al.
A simple two-dimensional example (Figure 1) proposed by Silva (2015) is presented to illustrate the process.
Figure 1
Simple two dimensional example illustra-
ting Signed distance function modeling
workflow, adapted from Silva (2015).
Signed distance values are then calcu- the sample is inside the domain, the distance main, the distance is positive. The euclidean
lated for each sample using Equation 2. If is negative; if the sample is outside the do- norm is used to measure the distance.
||u u ||, if i (u ) = 1
d (u ) = (2)
+||u u ||, if i (u ) = 0
The location uβ corresponds to the x, y and z must be rotated and dilated (or Anisotropic distances should be used when
closest sample of the opposite domain to uα. contracted), using the Equation 3 trans- it is known that the geologic body extends
The distances can also be anisotro- formation, to x'', y'', z'', and then, the eu- more in a specific direction than in the oth-
pic, in this case, the set of data coordinates clidean distances are calculated normally. ers, as in the case of a tabular body.
x '' 1
a max 0 0 cos cos sin sin sin sin cos cos sin sin cos sin x
y '' = 0 1
0 sin cos cos cos sin y (3)
a min
z '' 0 0 1
a vert cos sin sin sin cos sin sin cos sin cos cos cos z
Where α, β and φ are azimuth, dip tion is then interpolated to all locations the spatial configuration of data. Others
and rake respectively, and amax, amin and of interest (Equation 4). Ordinary krig- techniques, such as inverse distance can
avert are the anisotropy ranges. ing will be used due to its ability to ac- also be used.
The signed distance implicit func- count for directions of continuity and
n
d * (u ) = OK
(u )d (u ) (4)
=1
Deutsch and Wilde, (2013) rec- always use all samples. and variations in the range influence
ommend the use of global kriging The signed distances have a non- the shape and spatial extension of the
(Neufeld and Wilde, 2005), a smooth stationary behavior, namely, the var- geologic domains.
interpolator, to avoid artifacts in the iograms do not have a sill. Also, the At the end, the domains are classi-
implicit model. However the use of linearity of the distance makes the fied at the unsampled location u in func-
global kriging is restricted to the num- origin behave close to a quadratic tion of the sign of the signed distance
ber of samples due to computational form. Thus, the Gaussian model is a estimates (Equation 5). If the estimate
efficiency. Instead of global kriging, well-suited structure for modeling this is negative, the location is classified as
ordinary kriging can be used consider- type of variogram. Known geological inside the boundary; otherwise, the lo-
ing a large neighborhood and a large trends can be incorporated into the cation is considered as outside. Models
amount of data. Kriging allows us to model by the variogram, whilst RBF can be created in any resolution at the
control the run-time by changing the based methods are not able to incor- cost of computer resources and time,
number of samples retained in the es- porate anisotropy. An increment in the the resolution is controlled by modify-
timates, while methods based on RBF nugget effect disconnects the domains ing the dimensions of the grid blocks.
1, if d * (u ) 0
i * (u ) = (5)
0, otherwise
REM, Int. Eng. J., Ouro Preto, 70(2), 221-229, apr. jun. | 2017 223
Signed distance function implicit geologic modeling
3.3.1 Distance function modeling for multiple geologic domains
The described signed distance gorithm developed by Silva (2015) domains in the deposit. For all sample
implicit geologic modeling methodol- to handle multiple domains will locations z(u α),α = 1,...,n, an indicator
ogy may only be applied for binary be presented. vector of K elements is coded according
cases. In that which follows, the al- Suppose there exists K multiple to Equation 6:
1, if z (u ) = k
i k (u ) = k = 1,...,K (6)
0, if z (u ) otherwise
Thus the k element of the vector is In the same manner as in Equation for each k element of the vector (Equation
one while the remainder K-1 elements are 6, the signed distance value to the closest 7). The signs of the distances remain equal
set to code 0. opposite domain is computed individually as in the binary case.
||u u ||, if i k (u ) = 1
d k (u ) = k = 1,...,K (7)
+||u u ||, if i k (u ) = 0
Interpolation is performed individually for each k. Ordinary kriging is applied multiple times (Equation 8).
n
d k* (u ) = OK
(u )d k (u ) k = 1,...,K
=1
(8)
Then the final rock type model is determined by the following equation.
i *(u ) = k ' such that d k*' = min d k* (u ) k =1{ }K
(9)
The estimated distance provides a site domain. In this sense, the minimum es- as the most probable domain to be found
measure of proximity to the closest oppo- timated signed distance value may be seen at an unsampled location (Silva, 2015).
3.3.2 Measure of uncertainty (softmax transformation)
The proposed algorithm does from a random function as in sto- The central idea is to transform
not characterize uncertainty. There- chastic methods; Instead, it is based the estimated signed distance into
fore, a heuristic measure of uncer- on a post-processing transformation values that can be interpreted as
tainty was proposed by Silva (2015), method, a widely used approach for posterior probabilities (Equation 10).
this measure of uncertainty is not multiple class classification (McCul- The transformed values lie between 0
based on multiple realizations drawn lagh and Nelder, 1989). and 1 and sum to 1 for all elements k .
d k* (u )
e
P (i (u ) = k ) = k d k* (u )
(10)
k '=1
e
P(i(u))=k represents the probabil- K categories, a large parameter leads to rock types on a particular block and on
ity at location u to belong to category greater uncertainties. the right, the probabilities of this block
k, d*k (u) is the estimated distance for To illustrate how the transforma- to belong to each of the five categories
category k and ω is a parameter that tion works Figure 2 shows, on the left, obtained from the distances by the
controls the uncertainty bandwidth for estimated distances for five different softmax transformation.
Figure 2
Estimated distances and probabilities
obtained by softmax transformation
for five rock types on a particular block.
The shorter the estimated distance location u, the greater is the probability
for a given category K at an unknown of that location to belong to category k.
224 REM, Int. Eng. J., Ouro Preto, 70(2), 221-229, apr. jun. | 2017
Roberto Mentzingen Rolo et al.
3.3.3 Reproducing target proportions (servo system)
To avoid bias introduced by pref- match the representative proportions on of updating the probabilities of the do-
erential sampling in geologic models, the declustered data. mains based on the difference between
Silva (2015) proposed the use of some The algorithm is based on the the target and current marginal distribu-
sample declustering techniques and the probabilities obtained after the distance tions. The nodes visited must follow a
application of a servo system (Strebelle, transformation (Equation 10). The core random path to avoid the introduction
2002), to generate geologic models that of the algorithm (Equation 11) consists of artifacts.
P (i (u ) = k )update = P (i (u ) = k ) + μ ( p (k ) Pkc (u )) (11)
P(i(u)=k)update represents the updated defined as μ=λ/(1-λ) , p(k) and pck (u) are correction when λ is closer to one, namely,
probability at location u to belong to respectively, the target proportion for the proportions gets closer to the target.
category k, P(i(u)=k) is the probability category k and the marginal proportion The blocks need to be reclassified,
obtained by the softmax transformation, for the category k calculated from the now, based on the updated probabilities.
μ is the servo system correction parameter blocks already visited. There is more of a The classifier becomes Equation 12.
i * (u ) = argmax P(i(u) = k )updated (12)
Sometimes, when using a high a moving window system, similar to tional. After all of the post processing
µ factor, the servo system generates resources classification methodology and corrections, the algorithm assigns
structures that do not make physical (Deutsch et al., 1998). conditioning data to the corresponding
sense to reach the target proportions. The determination of uncertainty grid node.
Hence the maps must be corrected by and use of the servo system are op-
4. Case study
This case study was conducted on a Furthermore, in cases where we do not the same range (1500m) and the same
major gold mine operation. The dataset know any geologic anisotropic pattern proportions between structures contribu-
presents 9140 data representing 5 differ- a priori, the same variogram model can tion (10% for nugget effect and 90% for
ent lithologies. The area of the deposit is be used for all categories. Figure 3 shows a Gaussian structure). It can be seen in
approximately 10km2 with 1300m dip. the experimental variograms and their Figure 3 that the non-stationarity behavior
First, signed distances were calcu- respective fitted models for all five calcu- of the distances makes the modelled range
lated for all 5 lithologies, in practice users lated signed distances. The distances of somehow arbitrary.
must only input the categorical dataset. the experimental variograms never show As the absolute value of each var-
Then the distances calculated were vario- the nugget effect but it can be arbitrarily iogram structure contribution does not
graphed. It may seem a laborious process, added by the user to control the con- affect kriging results, only the propor-
but the extremely continuous behavior of nection between the lithologies. In this tion between the contributions, the same
the distances makes the variograms easy case a 10% nugget effect was added on variogram model (Equation 13) could be
to model, and are similar to each other. all variogram models. All models have assumed for all five signed distances.
h
(h ) = 0.1 + 0.9gauss (13)
1500 m
Now, distances calculated for each rameters and variogram models for each must be used to interpolate all categories.
category must be interpolated to each grid lithology. Resolution is the same as the A maximum of 40 samples per estimation
node, and the category responsible for the explicit model provided. Grid properties makes the run-time a few minutes, the
most negative estimated distance retained are in Table 2 and kriging parameters are output model differs a little compared to
and assigned. Users must input kriging pa- in Table 1; the same kriging parameters models created using 100 or 200 samples.
Neighborhood Number of samples
Radius (X) Radius (Y) Radius (Z) Min. Samples Max. Samples
Table 1
Lithology (1-5) 3000m 3000m 3000m 4 40
Kriging parameters.
Number of blocks Block dimensions
Num. X Num. Y Num. Z Dim. X Dim. Y Dim. Z
70 60 57 50m 50m 25m
Table 2
Grid properties.
REM, Int. Eng. J., Ouro Preto, 70(2), 221-229, apr. jun. | 2017 225
Signed distance function implicit geologic modeling
Figure 3
Signed distances experimental
variograms and their respective models.
A categorical scatterplot, pair- algorithm, on the Y axis, was constructed estimated blocks are consistent in both
ing each block category defined by the (Figure 4) showing a linear correlation models; these are the points that fall in
geomodeler, on the X axis, and by the coefficient 0.93, namely, 95% of the the 45º line.
226 REM, Int. Eng. J., Ouro Preto, 70(2), 221-229, apr. jun. | 2017
Roberto Mentzingen Rolo et al.
Figure 4
Scatterplot of distance
function and explicit models,
comparing the rock types assigned
to each block. Numbers adjacent to
the dots refer to quantify of blocks paired.
A high linear correlation coefficient is not indicative of a good compared in order to check if interpreted geological structures were
model. More simplistic methods such as nearest neighbor also produce reproduced by the algorithm. Slices are vertical along X and Y direc-
highly correlated scattterplots similar to Figure 4; however, produce tions, the block model has 70 sections along the X direction and 60
unrealistic models. In Figure 5, both of the model’s sections were sections in the Y direction. Four sections along each axis were chosen.
Figure 5
Comparison between explicit and
implicit three dimensional geologic
models, as well as sections along X and
Y axis showing the categorical models
and the probabilities for each rock type.
REM, Int. Eng. J., Ouro Preto, 70(2), 221-229, apr. jun. | 2017 227
Signed distance function implicit geologic modeling
Some complex structures inter- created models will rarely be final models, servo system can also be used to control
preted by geomodeler were reproduced as they must be refined manually. The the extent of this kind of structure.
satisfactorily by the algorithm as can be higher the number of samples, the more The implicit geologic modeling
seen highlighted by green lines (Figure the implicit model approximates the ex- algorithm formulation, description,
5). Conversely, the model mathematically plicit model, and in this sense, there is no and implementation should be relatively
generated misses or does not accurately need for much interference. In the case straightforward. Too complex algorithms
reproduce some other structures, as seen where samples are scarce, the implicit will be difficult to implement, explain,
highlighted by red lines (Figure 5), usu- model serves as a proto model, saving time and justify in practical settings. McLen-
ally, because there were not enough and effort in the early stages of modeling. nan (2007) proposed six different criteria
samples in that region, in those cases the Uncertainty was calculated by soft- to evaluate implicit modeling algorithms:
geomodeler counted with his/her experi- max transformation, as the probability (1) Simplicity: The signed distance al-
ence and/or additional information for of each block belonging to each category gorithm is extremely simple, where the
creating the explicit model. Discordant in each Figure 5 section. The plug-in only tough step is variogram modeling
blocks are shown in red and are located at default value for the factor that regulates that can be simplified using the same
lithological transitions. Scattered blocks the interrelation among the probabilities isotropic variogram for all categories.
belonging to lithology 5, displayed at is ω=175. Lower ω values create more Speed: It is quite fast. For the case study
the upper part of implicit model, are the conservative models. example where for a dataset with 9140
blocks assigned conditioning to data lo- The Servo system and the magnitude samples, the algorithm took a few min-
cations, where probably the geomodeler of µ factor are dependent on how much utes to run. (3) Subjectivity: Using the
arbitrarily increased its influence in the you trust your declustered data propor- same parameters, the model is perfectly
explicit model. This feature is not ex- tions. In this case study, applying the servo replicated. (4) Flexibility: Incorporating
pected to be reproduced by the algorithm, system makes the model more discordant incremental geological data is easy and
as there are few samples from lithology 5 with the model provided for comparison fast. (5) Uncertainty: The transforma-
which are surrounded by many lithology (not necessarily wrong). Border effects tion of distances in probabilities creates
3 and 4 samples. often occur when there is extrapolation in a heuristic measure of uncertainty. (6)
A trained professional will never be estimates, extending structures exaggerat- Realistic: The algorithm generates geo-
replaced by an algorithm and implicitly edly beyond the limits of the samples. The logically plausible boundary models.
5. Conclusions
As could be observed in the case mentioned, making the variogram model- measure of uncertainty is satisfactory, but
study, the algorithm does not replace a ing arbitrary and questionable. Moreover it is not an assessment of uncertainty that
trained professional, but its simplicity and the method is based on kriging, so only result from multiple realizations of a sto-
speed justify its usage, especially in the linear relationships between domains are chastic model. Knowing this, future work
early stages of modeling. modeled, unless a large amount of data may be focusrd on combining some kind
As disadvantages of the method, the is available. of boundary simulation with the straight-
non-stationarity of the distances can be For practical purposes, the heuristic forwardness of the deterministic method.
References
CHILÈS, J et al. Modelling the geometry of geological units and its uncertainty in 3D
from structural data: the potential-field method. In: PROCEEDINGS OF INTER-
NATIONAL SYMPOSIUM ON OREBODY MODELLING AND STRATEGIC
MINE PLANNING, Perth, Australia, 2004. p. 24.
COWAN, E. J. et al. Practical implicit geological modelling. In: INTERNATIONAL
MINING GEOLOGY CONFERENCE, 5. 2003. p. 17-19.
DEUTSCH, C V. et al. Geostatistical software library and user’s guide. New York:
Oxford University Press, 1998.
DEUTSCH, C V., WILDE, J. Modeling multiple coal seams using signed distance
functions and global kriging. International Journal of Coal Geology, v. 112, p.
87-93, 2013.
GALLI, A. et al. The pros and cons of the truncated Gaussian method. In: GEOSTA-
TISTICAL SIMULATIONS. Springer Netherlands, 1994. p. 217-233.
GUARDIANO, F. B., SRIVASTAVA, R. M. Multivariate geostatistics: beyond biva-
riate moments. In: Geostatistics Troia’92. Springer Netherlands, 1993. p. 133-144.
HARDY, R. L. Theory and applications of the multiquadric-biharmonic method 20
years of discovery 1968–1988. Computers & Mathematics with Applications, v.
19, n. 8-9, p. 163-208, 1990.
JOURNEL, A. G. Imaging of spatial uncertainty: a non-Gaussian approach. In: GE-
OSTATISTICAL, SENSITIVITY, AND UNCERTAINTY METHODS FOR
GROUND-WATER FLOW AND RADIONUCLIDE TRANSPORT MODE-
228 REM, Int. Eng. J., Ouro Preto, 70(2), 221-229, apr. jun. | 2017
Roberto Mentzingen Rolo et al.
LING. Proceedings. 1989a.
JOURNEL, A. G., ISAAKS, E. H. Conditional indicator simulation: application to a
Saskatchewan uranium deposit. Journal of the International Association for Ma-
thematical Geology, v. 16, n. 7, p. 685-718, 1984.
JOURNEL, A. G. Fundamentals of geostatistics in five lessons. Washington, DC:
American Geophysical Union, 1989b.
MCCULLOUGH, P., NELDER, J. A. Generalized linear models. 1989.
MCLENNAN, J. A. The decision of stationarity. University of Alberta, 2007. (PhD
Thesis).
MCLENNAN, J. A., DEUTSCH, C. V. Implicit boundary modeling (boundsim). Ed-
monton: Centre for Computational Geostatistics, 2006.
NEUFELD, C., WILDE, B. A global kriging program for artifact-free maps. CCG
Annual Report, v. 7, p. 403-1, 2005.
SAVCHENKO, V. V. et al. Function representation of solids reconstructed from scatte-
red surface points and contours. In: Computer Graphics Forum. Blackwell Science
Ltd, 1995. p. 181-188.
SILVA, D. M. A. Enhanced geologic modeling with data-driven training images for
improved resources and recoverable reserves. University of Alberta, 2015. (PhD
Thesis).
SINCLAIR, A. J., BLACKWELL, G. H. Applied mineral inventory estimation.
Cambridge University Press, 2002.
STREBELLE, S. Conditional simulation of complex geological structures using multi-
ple-point statistics. Mathematical Geology, v. 34, n. 1, p. 1-21, 2002.
Received: 13 October 2016 - Accepted: 10 January 2017.
REM, Int. Eng. J., Ouro Preto, 70(2), 221-229, apr. jun. | 2017 229